text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1708.00331v1 [cs.DS] 1 Aug 2017
Exact Approaches for the Travelling Thief
Problem
Junhua Wu
Markus Wagner
Sergey Polyakovskiy
Frank Neumann
Optimisation and Logistics
School of Computer Science
The University of Adelaide, Adelaide SA, Australia
Abstract
Many evolutionary and constructive heuristic approaches have been
introduced in order to solve the Traveling Thief Problem (TTP). However, the accuracy of such approaches is unknown due to their inability
to find global optima. In this paper, we propose three exact algorithms
and a hybrid approach to the TTP. We compare these with state-of-theart approaches to gather a comprehensive overview on the accuracy of
heuristic methods for solving small TTP instances.
1
Introduction
The travelling thief problem (TTP) [3] is a recent academic problem in which
two well-known combinatorial optimisation problems interact, namely the travelling salesperson problem (TSP) and the 0-1 knapsack problem (KP). It reflects
the complexity in real-world applications that contain more than one N P-hard
problem, which can be commonly observed in the areas of planning, scheduling
and routing. For example, delivery problems usually consist of a routing part
for the vehicle(s) and a packing part of the goods onto the vehicle(s).
Thus far, many approximate approaches have been introduced for addressing the TTP and most of them are evolutionary or heuristic [16]. Initially,
Polyakovskiy, Bonyadi, Wagner, Michalewicz, and Neumann [17] proposed two
iterative heuristics, namely the Random Local Search (RLS) and (1+1)-EA,
based on a general approach that solves the problem in two steps, one for
the TSP and one for the KP. Bonyadi et al. [4] introduced a similar twophased algorithm named Density-based Heuristic (DH) and a method inspired
by coevolution-based approaches named CoSolver. Mei et al. [11–13] also investigated the interdependency and proposed a cooperative coevolution based
approach similar to CoSolver, and a memetic algorithm called MATLS that
1
attempts to solve the problem as a whole. In 2015, Faulkner et al. [7] outperformed the existing approaches by their new operators and corresponding series
of heuristics (named S1–S5 and C1–C6). Recently, Wagner [21] investigated the
Max-Min Ant System (MMAS) [20] on the TTP, and El Yafrani and Ahiod [6]
proposed a memetic algorithm (MA2B) and a simulated annealing algorithm
(CS2SA). The results show that the new algorithms were competitive to the
state-of-the-art on a different range of TTP instances. Wagner et al. [22] found
in a study involving 21 approximate TTP algorithms that only a small subset
of them is actually necessary to form a well-performing algorithm portfolio.
However, due to the lack of exact methods, all of the above-mentioned approximate approaches cannot be evaluated with respect to their accuracy even
on small TTP instances. To address this issue, we propose three exact techniques and additional benchmark instances, which help to build a more comprehensive review of the approximate approaches.
In the remainder, we revisit the definition of the TTP in Section 2 and
introduce our exact approaches in Section 3. In Section 4, we elaborate on the
setup of our experiments and compare our exact and hybrid approaches with
the best approximate ones. The conclusions are drawn in Section 5.
2
Problem Statement
In this section, we outline the problem formulation. For a comprehensive description, we refer the interested reader to [17].
Given is a set of cities N = {1, . . . , n} and a set of items M = {1, . . . , m}.
City i, i = 2, . . . , n, contains a set of items Mi = {1, . . . , mi }, M = ∪ Mi .
i∈N
Item k positioned in the city i is characterised by its profit pik and weight
wik . The thief must visit each of the cities exactly once starting from the first
city and return back to it in the end. The distance dij between any pair of
cities i, j ∈ N is known. Any item may be selected as long as the total weight
of collected items does not exceed the capacity C. A renting rate R is to be
paid per each time unit taken to complete the tour. υmax and υmin denote
the maximal and minimum speeds of the thief. Assume that there is a binary
variable yik ∈ {0, 1} such that yik = 1 iff item k is chosen in city i. The
goal is to find a tour Π = (x1 , . . . , xn ), xi ∈ N , along with a packing plan
P = (y21 , . . . , ynmn ) such that their combination [Π, P ] maximises the reward
given in the form the following objective function.
!
mi
n−1
n X
X
X
dxi xi+1
dxn x1
+
(1)
Z([Π, P ]) =
pik yik − R
υmax − νWxn
υ
− νWxi
i=1 max
i=1
k=1
where ν = (υmax − υmin ) /C is a constant value defined by input parameters.
The minuend is the sum of all packed items’ profits and the subtrahend is the
amount that the thief pays for the knapsack’s rent equal to the total traveling
time along Π multiplied by R. In fact, the actual travel
the distance
Pspeed
Palong
mj
i
dxi xi+1 depends on the accumulated weight Wxi = j=1 k=1
wjk yjk of the
2
items collected in the preceding cities 1, . . . , i. This then slows down the thief
and has an impact on the overall benefit Z.
3
Exact Approaches to the TTP
In this section, we propose three exact approaches to the TTP.
As a simplified version of the TTP, Polyakovskiy and Neumann [16] have
recently introduced the packing while travelling problem (PWT), in which the
tour Π is predefined and only the packing plan P is variable. Furthermore,
Neumann et al. [14] prove that the PWT can be solved in pseudo-polynomial
time by dynamic programming taking into account the fact that the weights are
integer. The dynamic programming algorithm maps every possible weight w to
a packing plan P , i.e. f : w 7→ P , which guarantees a certain profit. Then the
optimal packing plan P ∗ is to be selected among all the plans that have been
obtained.
Here, we adopt these findings to derive two exact algorithms for the TTP.
Let · denote all possible weights for a given TTP instance. Let [Π, f (·)] designate the best solutions for the instance with tour Π obtained via the dynamic
programming for the PWT. As Π is to be variable, the optimum objective value
of the TTP is Z ∗ = arg max∀Π,w∈· Z([Π, f (w)]),. This yields the basis for two
of our approaches: dynamic programming (DP) and branch and bound search
(BnB). The following sections describe the two approaches as well as a constraint
programming (CP) technique adopted for the TTP.
3.1
Dynamic Programming
Our DP approach is based on the Held-Karp algorithm for the TSP [8] and
on the dynamic programming to the PWT [14]. Algorithm 1 depicts the pseudocode for our approach. Let Ṡ = N \ {1} be a subset of the cities and k ∈ N
refer to a particular city. Then [S, k] is a tour starting in city 1, visiting all
the cities in S ⊆ Ṡ exactly once, and ending in city k. The optimal solution of the TTP therefore can be described by [Ṡ, 1, fxn (Wxn )]∗ , where Wxn is
the total weight of the knapsack when leaving the last city xn and fxn results
from the dynamic programming algorithm for the PWT considering the tour
Π = (1, . . . , xn ). The following statement is valid with respect to the TTP’s
statement:
Z([Ṡ, 1, fxn (Wxn )]∗ ) = Z([Ṡ \ {xn }, xn , fxn−1 (Wxn − W xn )])
dxn x1
+ P xn + R
.
vmax − νWxn
Here, W xn and P xn are the total weight and the total profit of the items picked
in city xn . Clearly, Z([Ṡ \ {xn }, xn , fxn−1 (Wxn − W xn )]) is optimal for the
tour [Ṡ \ {xn }, xn ]. Furthermore, such a relationship exists for every pair of
[S, i, fj (Wj )] and [S \ {j}, j, fj−1 (Wj − W j )], where i ∈ Ṡ \ S and j ∈ S. In
3
Algorithm 1 Dynamic programming to the TTP
1: procedure Dynamic programming
2:
Ṡ = {2, ..., n}, S = ∅, w = 0
3:
for i = 2 to n do
4:
store [S, i, w] = Z([S, i, fi (w)]) in a mapping M
5:
for i = 3 to n do
6:
M0 = M
7:
for all [S, k, w] in M 0 do
8:
for all ṡ ∈ Ṡ \ S do
9:
for all s ∈ S do
10:
calculate Z([S ∪ {ṡ}, ṡ, fṡ (·)]) from Z([S, s, fs (·)])
11:
and store [S ∪ {ṡ}, ṡ, ·] = max Z([S ∪ {ṡ}, ṡ, fṡ (·)]) in M 0
12:
M = M0
13:
calculate Z([Ṡ, 1, f1 (Wxn )]∗ ) from Z([Ṡ \ {i}, i, fi (·)]∗ ), i ∈ {2, ..., n} in
M
fact, having an optimal solution for a given TTP instance, one can compute the
optimal solution for the instance that excludes the last city from the solution of
the original problem. Following this idea, we build our DP for the TTP.
The DP is costly in terms of the memory consumption, which reaches O(2n nC).
To reduce these cost, let EU define an upper bound on the value of a feasible
solution built on the partial solution [S, k, fk (·)] as follows:
EU ([S, k, fk (·)]) = max Z([S, k, fk (·)]) +
mi
X X
i∈Ṡ\S j=1
pij − R
dk1
vmax
It estimates the maximal profit that the thief may obtain by passing the remaining part of the tour with the maximal speed, that is generating the minimal
possible cost. Obviously, this guarantees that the complete optimal solution
can not exceed the bound. Therefore, if any incumbent solution is known, it is
valid to eliminate a partial solution [S, k, fk (w)], w ∈ · if EU ([S, k, fk (w)]) < ZL ,
where ZL is the objective value of the incumbent. In practice, one can obtain
an incumbent solution (and compute ZL ) in two stages. First, a feasible solution Π for the TSP part of the problem can be computed by solvers such
as Concorde [1] or by the Lin-Kernighan algorithm [10]. Second, the dynamic
programming applied for Π contributes the packing plan.
3.2
Branch and Bound Search
Now, we introduce a branch and bound search for the TTP employing the upper
bound EU defined in Section 3.1. Algorithm 2 depicts the pseudocode, where
Πi , i ∈ {1, ..., n} denotes a sub-permutation of Π with the cities 1 to i visited,
and fi is the mapping f : w 7→ P calculated for Π by the dynamic programming
for the PWT.
4
Algorithm 2 Branch and Bound Search for the TTP
1: procedure BnB Search
2:
Create an initial solution to gain the benefit best and an tour permutation Π
3:
Create an empty mapping M
4:
Set l = 0
5:
Search(Π, l, M, best)
6: function Search(Π, l, M, best)
7:
if l == n then
8:
calculate Z([Π, fn (·)]) from Z([Πn−1 , fn−1 (·)]) in M
9:
return max{max Z([Π, fn (·)]), best}
10:
else
11:
for i = l + 1 to n do
12:
Swap cities l + 1 and i in Π
13:
Set M’ = Calculate Z([Πl+1 , fl+1 (·)]) from Z([Πl , fl (·)]) in M
14:
if max EU ([Πl+1 , fl+1 (·)]) > best then
15:
best = max{best, Search(Π, l + 1, M 0 , best) }
16:
Swap cities l + 1 and i in Π
17:
return best
A way to tighten the upper bound EU is by providing a better estimation
of the remaining distance from the current city k to the last city of the tour.
Currently, the shortest distance from k to 1, i.e. dk1 , is used. The following two
ways can improve the estimation: (i) the use of distance df 1 from city f to city
1, where f is the farthest unvisited city from 1; (ii) the use of distance d∗ − dt ,
where d∗ is the shortest path that can be pre-calculated and dt is the distance
passed so far to achieve city k in the tour Π. These two ideas can be joined
together by using the max{df , (d∗ − dt )} to enhance the result.
3.3
Constraint Programming
Now, we present our third exact approach adopting the existing state-of-theart constraint programming (CP) paradigm [9]. Our model employs a simple permutation based representation of the tour that allows the use of the
AllDifferent filtering algorithm [2]. Similarly to the Section 2, a vector
W = (W1 , . . . , Wn ) is used to refer to the total weights accumulated in the
cities of tour Π. Specifically, Wi is the weight of the knapsack when the thief
departs from city i. The model bases the search on two types of decision variables:
• x denotes the particular positions of the cities in tour Π. Variable xi takes
the value of j ∈ N to indicate that j is the ith city to be visited. The
initial variable domain of x1 is D (x1 ) = {1} and it is D (xi ) = N \ {i} for
any subsequently visited city i = 2, . . . , n.
5
• y signals on the selection of an item in the packing plan P . Variable yik ,
i ∈ N , k ∈ Mi , is binary, therefore D (yik ) = {0, 1}.
Furthermore, an integer-valued vector d is used to express the distance matrix so
that its element n (xi − 1)+xi+1 equals the distance dxi xi+1 between two consecutive cities xi and xi+1 in Π. The model relies on the AllDifferent[x1 , . . . , xn ]
constraint, which ensures that the values of x1 , . . . , xn are distinct. It also involves the Element(g, h) expression, which returns the hth variable in the list
of variables g. In total, the model (CPTTP) consists of the following objective
function and constraints:
max
mi
n X
X
pij yij
i=1 j=1
−R
n−1
X
i=1
Element(d, n (xn − 1) + 1)
Element(d, n (xi − 1) + xi+1 )
+
vmax − νElement(W, xi )
vmax − νElement(W, xn )
AllDifferent[x1 , . . . , xn ]
X
Wi = Wi−1 +
wij yij , i ∈ {2, . . . , n}
!
(2)
(3)
(4)
j∈Mi
Wn ≤ C
(5)
Expression (2) calculates the objective value according to function (1). Constraint (3) verifies that all the cities are assigned to different positions, and thus
are visited exactly once. This is a sub-tour elimination constraint. Equation (4)
calculates the weight Wi of all the items collected in the cities 1, . . . , i. Equation (5) is a capacity constraint.
The performance of a CP model depends on its solver; specifically, on the
filtering algorithms and on the search strategies it applies. Here, we use IBM
ILOG CP Optimizer 12.6.2 with its searching algorithm set to the restart
mode. This mode adopts a general purpose search strategy [18] inspired from
integer programming techniques and is based on the concept of the impact of
a variable. The impact measures the importance of a variable in reducing the
search space. The impacts, which are learned from the observation of the domains’ reduction during the search, help the restart mode dramatically improve
the performance of the search. Within the search, the cities are assigned to the
positions first and then the items are decided on. Therefore, the solver instantiates x1 , . . . , xn prior to y21 , . . . , ynmn variables applying its default selection
strategy. Our extensive study shows that such an order gives the best results
fast.
4
Computational Experiments
In this section, we first compare the performance of the exact approaches to
TTP in order to find the best one for setting the baseline for the subsequent
comparison of the approximate approaches. Our experiments run on the CPU
6
cluster of the Phoenix HPC at the University of Adelaide, which contains 3072
Intel(R) Xeon(R) 2.30GHz CPU cores and 12TB of memory. We allocate one
CPU core and 32GB of memory to each individual experiment.
4.1
Computational Set Up
Instance
eil51 n05
eil51 n06
eil51 n07
eil51 n08
eil51 n09
eil51 n10
eil51 n11
eil51 n12
eil51 n13
eil51 n14
eil51 n15
eil51 n16
eil51 n17
eil51 n18
eil51 n19
eil51 n20
m4 uncorr 01
m5 uncorr 01
m6 uncorr 01
m7 uncorr 01
m8 uncorr 01
m9 uncorr 01
m10 uncorr 01
m11 uncorr 01
m12 uncorr 01
m13 uncorr 01
m14 uncorr 01
m15 uncorr 01
m16 uncorr 01
m17 uncorr 01
m18 uncorr 01
m19 uncorr 01
n
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Running time (in seconds)
DP
BnB
CP
0.018
0.023
0.222
0.07
0.079
0.24
0.143
0.195
0.497
0.343
0.505
4.594
0.633
1.492
63.838
0.933
5.188
776.55
2.414
23.106
12861.181
3.938
204.786
14.217
2007.074
13.408
36944.146
89.461
59.526
134.905
366.082
830.18
2456.873
-
m
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Table 1: Columns ‘n’ and ‘m’ denote the number of cities and the number of
items, respectively. Running times are given in seconds for DP, BnB and CP
for different numbers of cities and items. ‘-’ denotes the case when an approach
failed to achieve an optimal solution in the given time limit.
To run our experiments, we generate an additional set of small-sized instances following the way proposed in [17]1 . We use only a single instance of
the original TSP library [19] as the starting point for our new subset. It is
entitled as eil51 and contains 51 cities. Out of these cities, we select uniformly
at random cities that we removed in order to obtain smaller test problems with
n = 5, . . . , 20 cities. To set up the knapsack component of the problem, we
adopt the approach given in [15] and use the corresponding problem generator
available in [? ]. As one of the input parameters, the generator asks for the
range of coefficients, which we set to 1000. In total, we create knapsack test
problems containing k(n−1), k ∈ {1, 5, 10} items and which are characterised by
a knapsack capacity category Q ∈ {1, 6, 10}. Our experiments focus on uncorrelated (uncorr), uncorrelated with similar weights (uncorr-s-w), and multiple
1 All
instances are available online: http://cs.adelaide.edu.au/~optlog/research/ttp.
php
7
strongly correlated (m-s-corr) types of instances. At the stage of assigning the
items of a knapsack instance to the particular cities of a given TSP tour, we
sort the items in descending order of their profits and the second city obtains
k, k ∈ {1, 5, 10}, items of the largest profits, the third city then has the next k
items, and so on. All the instances use the “CEIL 2D” for intra-city distances,
which means that the Euclidean distances are rounded up to the nearest integer.
We set υmin and υmax to 0.1 and 1.
Tables 1 and 3 illustrate the results of the experiments. The test instances’
names should be read as follows. First, eil51 stays for the name of the original
TSP problem. The values succeeding n and m denote the actual number of
cities and the total number of items, respectively, which are further followed by
the generation type of a knapsack problem. Finally, the postfixes 1, 6 and, 10
in the instances’ names describe the knapsack’s capacity C.
4.2
Comparison of the exact approaches
We compare the three exact algorithms by allocating each instance a generous
24-hour time limit. Our aim is to analyse the running time of the approaches
influenced by the increasing number of cities. Table 1 shows the running time
of the approaches.
4.3
Comparison between DP and Approximate Approaches
With the exact approaches being introduced, approximate approaches can be
evaluated with respect to their accuracy to the optima. In the case of the TTP,
most state-of-the-art approximate approaches are evolutionary algorithms and
local searches, such as Memetic Algorithm with 2-OPT and Bit-flip (MA2B),
CoSolver-based with 2-OPT, and Simulated Annealing (CS2SA) in [6], CoSolverbased with 2-OPT and Bit-flip (CS2B) in [5], and S1, S5, and C5 in [7].
4.3.1
Hybrid Approaches.
In addition to existing heuristics, we introduce enhanced approaches of S1 and
S5, which are hybrids of the two and that one of dynamic programming for
the PWT [14]. The original S1 and S5 work as follows. First, a single TSP
tour is computed using the Chained Lin-Kernighan-Heuristic [10], then a fast
packing heuristic is applied. S1 performs these two steps only once and only in
this order, while S5 repeats S1 until the time budget is exhausted. Our hybrids
DP-S1 and DP-S5 are equivalent to these two algorithms, however, they use the
exact dynamic programming to the PWT as a packing solver. This provides
better results as we can now compute the optimal packing for the sampled TSP
tours.
4.3.2
Results.
We start by showing a performance summary of 10 algorithms on 432 instances
in Table 2. In addition, Table 3 shows detailed results for a subset of the best
8
gap
avg
stdev
#opt
#1%
#10%
MA2B
0.3%
2.2%
312
265
324
CS2B
15.3%
17.8%
70
100
161
CS2SA
11.5%
16.7%
117
132
206
S1
38.9%
29.4%
3
10
27
S5
15.7%
24.6%
42
160
203
C5
09.9%
18.8%
193
193
240
DP-S1
30.1%
20.1%
5
9
33
DP-S5
3.3%
8.5%
85
245
288
Table 2: Performance summary of heuristic TTP solvers across all instances for
which the optimal result has been obtained. #opt is the number of times when
the average of 10 independent repetitions is equal to the optimum. #1% and
#10% show the number of times the averages are within 1% and 10%.
approaches on a subset of instances. Figure 1 shows the results of the entire
comparison. We include trend lines2 for two selected approaches, which we will
explain in the following.
We would like to highlight the following observations:
1. S1 performs badly across a wide range of instances. Its restart variant
S5 performs better, however, its lack of a local search becomes apart in
its relatively bad performance (compared to other approaches) on small
instances.
2. C5 performs better than both S1 and S5, which is most likely due to its
local searches that differentiate it from S1 and S5. Still, we can see a
“hump” in its trend line for smaller instances, which flattens out quickly
for larger instances.
3. The dynamic programming variants DP-S1 and DP-S5 perform slightly
better than S1 and S5, which shows the difference in quality of the packing
strategy; however, this is at times balanced out by the faster packing which
allows more TSP tours to be sampled. For small instances, DP-S5 lacks a
local search on the tours, which is why its gap to the optimum is relatively
large, as shown by the respective trend lines.
4. MA2B dominates the field with outstanding performance across all instances, independent of number of cities and number of items. Remarkable
is the high reliability with which it reaches a global optimum.
Interestingly, all approaches seem to have difficulties solving instances with
the knapsack configuration multiple-strongly-corr 01 (see Table 3). Compared
to the other two knapsack types, TTP-DP takes the longest to solve the strongly
correlated ones. Also, these tend to be the only instances for which the heuristics
rarely find optimal solutions, if at all.
2 They
are fitted polynomials of degree six used only for visualisation purposes.
9
Instance
eil51 n05
eil51 n05
eil51 n05
eil51 n05
eil51 n05
eil51 n05
eil51 n10
eil51 n10
eil51 n10
eil51 n10
eil51 n10
eil51 n10
eil51 n12
eil51 n12
eil51 n12
eil51 n12
eil51 n12
eil51 n12
eil51 n15
eil51 n15
eil51 n15
eil51 n15
eil51 n15
eil51 n15
eil51 n16
eil51 n16
eil51 n16
eil51 n16
eil51 n16
eil51 n16
eil51 n17
eil51 n17
eil51 n17
eil51 n17
eil51 n17
eil51 n17
eil51 n18
eil51 n18
eil51 n18
eil51 n18
eil51 n18
eil51 n18
eil51 n19
eil51 n19
eil51 n19
eil51 n19
eil51 n19
eil51 n19
eil51 n20
eil51 n20
eil51 n20
eil51 n20
eil51 n20
eil51 n20
m4 multiple-strongly-corr 01
m4 uncorr 01
m4 uncorr-similar-weights 01
m20 multiple-strongly-corr 01
m20 uncorr 01
m20 uncorr-similar-weights 01
m9 multiple-strongly-corr 01
m9 uncorr 01
m9 uncorr-similar-weights 01
m45 multiple-strongly-corr 01
m45 uncorr 01
m45 uncorr-similar-weights 01
m11 multiple-strongly-corr 01
m11 uncorr 01
m11 uncorr-similar-weights 01
m55 multiple-strongly-corr 01
m55 uncorr 01
m55 uncorr-similar-weights 01
m14 multiple-strongly-corr 01
m14 uncorr 01
m14 uncorr-similar-weights 01
m70 multiple-strongly-corr 01
m70 uncorr 01
m70 uncorr-similar-weights 01
m15 multiple-strongly-corr 01
m15 multiple-strongly-corr 10
m15 uncorr 01
m15 uncorr 10
m15 uncorr-similar-weights 01
m15 uncorr-similar-weights 10
m16 multiple-strongly-corr 01
m16 multiple-strongly-corr 10
m16 uncorr 01
m16 uncorr 10
m16 uncorr-similar-weights 01
m16 uncorr-similar-weights 10
m17 multiple-strongly-corr 01
m17 multiple-strongly-corr 10
m17 uncorr 01
m17 uncorr 10
m17 uncorr-similar-weights 01
m17 uncorr-similar-weights 10
m18 multiple-strongly-corr 01
m18 multiple-strongly-corr 10
m18 uncorr 01
m18 uncorr 10
m18 uncorr-similar-weights 01
m18 uncorr-similar-weights 10
m19 multiple-strongly-corr 01
m19 multiple-strongly-corr 10
m19 uncorr 01
m19 uncorr 10
m19 uncorr-similar-weights 01
m19 uncorr-similar-weights 10
TTP-DP
OPT
RT
619.227
0.02
466.929
0.02
299.281
0.02
773.573
0.08
2144.796
0.07
269.015
0.04
573.897
1.21
1125.715
0.93
753.230
0.86
1091.127 14.89
6009.431
6.39
3009.553
8.87
648.546
4.58
1717.699
3.94
774.107
3.36
1251.780 117.99
8838.012 35.79
3734.895 38.36
547.419 39.82
2392.996 89.46
637.419 16.35
920.372 3984.29
9922.137 740.22
4659.623 867.78
794.745 105.5
4498.848 623.4
2490.889
59.5
3601.077 211.5
540.897
36.4
3948.211 245.4
685.565 248.6
3826.098 2190.4
2342.664 134.9
2275.279 554.5
556.851
70.8
2935.961 787.7
834.031 715.7
5531.373 6252.4
2644.491 366.1
3222.603 1462.7
532.906 148.3
4420.438 1929.3
910.229 1771.6
2604.844 830.2
4048.408 3884.3
472.186 412.3
5573.695 5878.8
518.189 4533.7
2092.673 2456.9
3044.391 12776.0
451.052 1007.7
4169.799 15075.7
MA2B
Gap Std RT
29.1 12.1 2.71
0.0 0.0 3.22
0.0 0.0 3.21
13.4 0.0 1.44
0.0 0.0 3.35
0.0 0.0 3.51
0.0 0.0 6.07
0.0 0.0 6.06
0.0 0.0 5.87
0.0 0.0 7.99
0.0 0.0 8.6
0.0 0.0 6.78
0.0 0.0 6.08
0.0 0.0 7.21
0.0 0.0 7.03
0.0 0.0 9.19
0.0 0.0 9.76
12.3 0.0 8.34
0.0 0.0 7.87
0.0 0.0 7.28
0.0 0.0 6.86
2.1 1.1 12.11
0.0 0.0 9.67
0.0 0.0 7.98
0.0 0.0 7.7
0.0 0.0 9.1
1.0 0.7 8.4
0.0 0.0 9.0
0.0 0.0 8.5
0.0 0.0 8.7
0.0 0.0 8.4
0.0 0.0 9.8
0.0 0.0 8.3
0.0 0.0 9.6
0.0 0.0 8.1
0.0 0.0 9.7
7.9 0.8 10.2
0.0 0.0 10.5
0.0 0.0 9.7
0.0 0.0 10.3
0.0 0.0 8.5
0.0 0.0 9.9
0.0 0.0 9.3
- 10.4
0.0 0.0 9.7
0.0 0.0 10.9
0.0 0.0 9.2
0.0 0.0 10.5
0.6 0.6 11.1
- 12.1
0.0 0.0 8.7
0.0 0.0 9.8
0.0 0.0 7.9
0.0 0.0 9.4
C5
Gap
35.5
0.0
7.8
14.3
7.4
0.0
0.0
0.0
0.0
0.0
6.6
0.0
4.6
0.0
0.0
0.0
0.0
12.3
14.1
3.8
0.0
0.0
7
0.0
18.9
12.9
1.6
7.1
0.0
5.8
0.2
0.0
0.0
0.0
0.0
0.0
9.2
0.4
0.2
0.0
0.0
0.0
20.1
0.0
0.0
0.0
0.0
14.1
0.0
0.0
0.0
0.0
Std
1.20e-6
2.20e-6
2.40e-6
0.0
0.0
2.30e-6
0.0
1.30e-6
0.0
0.0
2.30e-6
2.30e-6
2.20e-6
1.20e-6
2.30e-6
0.0
0.0
0.0
1.30e-6
0.0
1.60e-6
2.70e-6
1.20e-6
0.0
1.6e-6
0.0
2.3e-6
1.6e-6
3.0e-6
1.5e-6
1.5e-6
1.5e-6
0.0
0.0
0.0
0.0
0.0
1.5e-6
0.0
1.3e-6
1.3e-6
2.9e-6
1.6e-6
0.0
1.4e-6
1.5e-6
0.0
1.4e-6
0.0
0.0
0.0
0.0
DP-S5
Gap
Std
41.3
0.0
0.0 2.20e-6
7.8 1.20e-6
12.8
0.0
6.6 2.30e-6
0.0
0.0
0.0
0.0
0.0 1.30e-6
0.0
0.0
0.0
0.0
0.0
0.0
0.0 2.30e-6
4.6 2.20e-6
0.0 1.20e-6
0.0 2.30e-6
0.0
0.0
0.0
0.0
0.2
0.0
13.3 1.30e-6
3.8
0.0
0.0 1.60e-6
0.0 2.70e-6
1.9
0.0
0.0
0.0
18.9 1.6e-6
16.6 1.3e-6
1.6 2.3e-6
7.1 1.6e-6
0.0 3.0e-6
13.6
0.0
0.0 1.5e-6
0.0 1.5e-6
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
12.9 1.7e-6
0.4 1.5e-6
1.8
0.0
0.2
0.0
0.0 1.3e-6
0.3 1.8e-6
20.1 1.6e-6
0.0
0.0
0.0 1.4e-6
0.0 1.5e-6
0.0
0.0
12.3
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
Table 3: Comparison between DP and the approximate approaches running in
10 minutes limits. Each approximate algorithm runs 10 times for each instance
T −Obj
and use the average as the objective Obj. Gap is measured by OPOP
% and
T
runtime (RT) is in second. The results of C5 and DP-S5 are obtained when
they reach the time limit of 10 minutes per instance. Highlighted in blue are
the best approximate results. DP runs out of memory for the instances without
results.
10
Figure 1: Showing a gap to an optimal solution when one has been obtained
by an exact approach. From left to right: the 432 instances first sorted by the
number of cities, then by the total number of items.
5
Conclusion
The traveling thief problem (TTP) has attracted significant attention in recent
years within the evolutionary computation community. In this paper, we have
presented and evaluated exact approaches for the TTP based on dynamic programming, branch and bound, and constraint programming. We have used the
exact solutions provided by our DP approach to evaluate the performance of
current state-of-the-art TTP solvers. Our investigations show that they are obtaining in most cases (close to) optimal solutions. However, for a small fraction
of tested instances we obverse a gap to the optimal solution of more than 10%.
Acknowledgements
This work was supported by the Australian Research councils through grants
DP130104395 and DE160100850, and by the supercomputing resources provided
by the Phoenix HPC service at the University of Adelaide.
References
[1] D. Applegate, R. Bixby, V. Chvatal, and W. Cook. Concorde tsp solver.
http://www.math.uwaterloo.ca/tsp/concorde.html, 2006.
[2] P. Benchimol, W.-J. v. Hoeve, J.-C. Régin, L.-M. Rousseau, and M. Rueher.
11
Improved filtering for weighted circuit constraints. Constraints, 17(3):205–
233, Jul 2012. ISSN 1572-9354. doi: 10.1007/s10601-012-9119-x.
[3] M. Bonyadi, Z. Michalewicz, and L. Barone. The travelling thief problem:
The first step in the transition from theoretical problems to realistic problems. In Evolutionary Computation (CEC), 2013 IEEE Congress on, pages
1037–1044, 2013.
[4] M. R. Bonyadi, Z. Michalewicz, M. R. Przybylek, and A. Wierzbicki. Socially inspired algorithms for the travelling thief problem. In Proceedings
of the 2014 Annual Conference on Genetic and Evolutionary Computation,
GECCO ’14, pages 421–428. ACM, 2014.
[5] M. El Yafrani and B. Ahiod. Cosolver2b: An efficient local search heuristic
for the travelling thief problem. In Computer Systems and Applications
(AICCSA), 2015 IEEE/ACS 12th International Conference of, pages 1–5.
IEEE, 2015.
[6] M. El Yafrani and B. Ahiod. Population-based vs. single-solution heuristics
for the travelling thief problem. In Proceedings of the Genetic and Evolutionary Computation Conference 2016, GECCO ’16, pages 317–324. ACM,
2016.
[7] H. Faulkner, S. Polyakovskiy, T. Schultz, and M. Wagner. Approximate
approaches to the traveling thief problem. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation, GECCO ’15,
pages 385–392. ACM, 2015.
[8] M. Held and R. M. Karp. A dynamic programming approach to sequencing
problems. In Proceedings of the 1961 16th ACM National Meeting, ACM
’61, pages 71.201–71.204. ACM, 1961.
[9] J. N. Hooker. Logic, optimization, and constraint programming. INFORMS
Journal on Computing, 14(4):295 – 321, 2002.
[10] S. Lin and B. W. Kernighan. An effective heuristic algorithm for the
traveling-salesman problem. Operations research, 21(2):498–516, 1973.
[11] Y. Mei, X. Li, and X. Yao. Improving Efficiency of Heuristics for the
Large Scale Traveling Thief Problem, pages 631–643. Springer International Publishing, Cham, 2014. ISBN 978-3-319-13563-2. doi: 10.1007/
978-3-319-13563-2 53.
[12] Y. Mei, X. Li, F. Salim, and X. Yao. Heuristic evolution with genetic programming for traveling thief problem. In 2015 IEEE Congress
on Evolutionary Computation (CEC), pages 2753–2760, May 2015. doi:
10.1109/CEC.2015.7257230.
12
[13] Y. Mei, X. Li, and X. Yao. On investigation of interdependence between
sub-problems of the travelling thief problem. Soft Computing, 20(1):157–
172, 2016.
[14] F. Neumann, S. Polyakovskiy, M. Skutella, L. Stougie, and J. Wu. A
Fully Polynomial Time Approximation Scheme for Packing While Traveling. ArXiv e-prints, 2017.
[15] D. Pisinger. Where are the hard knapsack problems? Comput. Oper. Res.,
32(9):2271–2284, Sept. 2005. ISSN 0305-0548. doi: 10.1016/j.cor.2004.03.
002.
[16] S. Polyakovskiy and F. Neumann. The packing while traveling problem.
European Journal of Operational Research, 258(2):424 – 439, 2017.
[17] S. Polyakovskiy, M. R. Bonyadi, M. Wagner, Z. Michalewicz, and F. Neumann. A comprehensive benchmark set and heuristics for the traveling
thief problem. In Proceedings of the 2014 Annual Conference on Genetic
and Evolutionary Computation, GECCO ’14, pages 477–484. ACM, 2014.
[18] P. Refalo. Principles and Practice of Constraint Programming – CP
2004, chapter Impact-Based Search Strategies for Constraint Programming,
pages 557–571. Springer, 2004.
[19] G. Reinelt. TSPLIB- a traveling salesman problem library. ORSA Journal
of Computing, 3(4):376–384, 1991.
[20] T. Stützle and H. H. Hoos. MAX MIN Ant System. Future Generation
Computer Systems, 16(8):889 – 914, 2000.
[21] M. Wagner. Stealing items more efficiently with ants: A swarm intelligence approach to the travelling thief problem. In M. Dorigo, M. Birattari,
X. Li, M. López-Ibáñez, K. Ohkura, C. Pinciroli, and T. Stützle, editors,
Swarm Intelligence: 10th International Conference, ANTS 2016, Brussels,
Belgium, September 7-9, 2016, Proceedings, pages 273–281. Springer, 2016.
[22] M. Wagner, M. Lindauer, M. Mısır, S. Nallaperuma, and F. Hutter. A
case study of algorithm selection for the traveling thief problem. Journal
of Heuristics, pages 1–26, 2017.
13
| 8cs.DS
|
arXiv:1711.04323v1 [cs.CV] 12 Nov 2017
High-Order Attention Models for Visual Question
Answering
Idan Schwartz
Department of Computer Science
Technion
[email protected]
Alexander G. Schwing
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
[email protected]
Tamir Hazan
Department of Industrial Engineering & Management
Technion
[email protected]
Abstract
The quest for algorithms that enable cognitive abilities is an important part of
machine learning. A common trait in many recently investigated cognitive-like
tasks is that they take into account different data modalities, such as visual and
textual input. In this paper we propose a novel and generally applicable form
of attention mechanism that learns high-order correlations between various data
modalities. We show that high-order correlations effectively direct the appropriate
attention to the relevant elements in the different data modalities that are required
to solve the joint task. We demonstrate the effectiveness of our high-order attention
mechanism on the task of visual question answering (VQA), where we achieve
state-of-the-art performance on the standard VQA dataset.
1
Introduction
The quest for algorithms which enable cognitive abilities is an important part of machine learning
and appears in many facets, e.g., in visual question answering tasks [6], image captioning [26],
visual question generation [18, 10] and machine comprehension [8]. A common trait in these recent
cognitive-like tasks is that they take into account different data modalities, for example, visual and
textual data.
To address these tasks, recently, attention mechanisms have emerged as a powerful common theme,
which provides not only some form of interpretability if applied to deep net models, but also often
improves performance [8]. The latter effect is attributed to more expressive yet concise forms of the
various data modalities. Present day attention mechanisms, like for example [15, 26], are however
often lacking in two main aspects. First, the systems generally extract abstract representations of
data in an ad-hoc and entangled manner. Second, present day attention mechanisms are often geared
towards a specific form of input and therefore hand-crafted for a particular task.
To address both issues, we propose a novel and generally applicable form of attention mechanism
that learns high-order correlations between various data modalities. For example, second order
correlations can model interactions between two data modalities, e.g., an image and a question, and
more generally, k−th order correlations can model interactions between k modalities. Learning these
correlations effectively directs the appropriate attention to the relevant elements in the different data
modalities that are required to solve the joint task.
We demonstrate the effectiveness of our novel attention mechanism on the task of visual question
answering (VQA), where we achieve state-of-the-art performance on the VQA dataset [2]. Some
Original Image
Unary Potentials
Pairwise Potentials
Final Attention
What does the man
have on his head?
What does the man
have on his head?
What does the man
have on his head?
What does the man
have on his head?
How many cars
are in the picture?
How many cars are
in the picture?
How many cars are
in the picture?
How many cars are in
the picture?
Figure 1: Results of our multi-modal attention for one image and two different questions (1st
column). The unary image attention is identical by construction. The pairwise potentials differ
for both questions and images since both modalities are taken into account (3rd column). The final
attention is illustrated in the 4th column.
of our results are visualized in Fig. 1, where we show how the visual attention correlates with the
textual attention.
We begin by reviewing the related work. We subsequently provide details of our proposed technique,
focusing on the high-order nature of our attention models. We then conclude by presenting the
application of our high-order attention mechanism to VQA and compare it to the state-of-the-art.
2
Related work
Attention mechanisms have been investigated for both image and textual data. In the following we
review mechanisms for both.
Image attention mechanisms: Over the past few years, single image embeddings extracted from a
deep net (e.g., [17, 16]) have been extended to a variety of image attention modules, when considering
VQA. For example, a textual long short term memory net (LSTM) may be augmented with a spatial
attention [29]. Similarly, Andreas et al. [1] employ a language parser together with a series of neural
net modules, one of which attends to regions in an image. The language parser suggests which neural
net module to use. Stacking of attention units was also investigated by Yang et al. [27]. Their stacked
attention network predicts the answer successively. Dynamic memory network modules which capture
contextual information from neighboring image regions has been considered by Xiong et al. [24].
Shih et al. [23] use object proposals and and rank regions according to relevance. The multi-hop
attention scheme of Xu et al. [25] was proposed to extract fine-grained details. A joint attention
mechanism was discussed by Lu et al. [15] and Fukui et al. [7] suggest an efficient outer product
mechanism to combine visual representation and text representation before applying attention over
the combined representation. Additionally, they suggested the use of glimpses. Very recently, Kazemi
et al. [11] showed a similar approach using concatenation instead of outer product. Importantly, all of
these approaches model attention as a single network. The fact that multiple modalities are involved
is often not considered explicitly which contrasts the aforementioned approaches from the technique
we present.
Very recently Kim et al. [14] presented a technique that also interprets attention as a multi-variate
probabilistic model, to incorporate structural dependencies into the deep net. Other recent techniques
are work by Nam et al. [19] on dual attention mechanisms and work by Kim et al. [13] on bilinear
2
models. In contrast to the latter two models our approach is easy to extend to any number of data
modalities.
Textual attention mechanisms: We also want to provide a brief review of textual attention. To
address some of the challenges, e.g., long sentences, faced by translation models, Hermann et al. [8]
proposed RNNSearch. To address the challenges which arise by fixing the latent dimension of neural
nets processing text data, Bahdanau et al. [3] first encode a document and a query via a bidirectional
LSTM which are then used to compute attentions. This mechanism was later refined in [22] where
a word based technique reasons about sentence representations. Joint attention between two CNN
hierarchies is discussed by Yin et al. [28].
Among all those attention mechanisms, relevant to our approach is work by Lu et al. [15] and the
approach presented by Xu et al. [25]. Both discuss attention mechanisms which operate jointly over
two modalities. Xu et al. [25] use pairwise interactions in the form of a similarity matrix, but ignore
the attentions on individual data modalities. Lu et al. [15] suggest an alternating model, that directly
combines the features of the modalities before attending. Additionally, they suggested a parallel
model which uses a similarity matrix to map features for one modality to the other. It is hard to extend
this approach to more than two modalities. In contrast, our model develops a probabilistic model,
based on high order potentials and performs mean-field inference to obtain marginal probabilities.
This permits trivial extension of the model to any number of modalities.
Additionally, Jabri et al. [9] propose a model where answers are also used as inputs. Their approach
questions the need of attention mechanisms and develops an alternative solution based on binary
classification. In contrast, our approach captures high-order attention correlations, which we found to
improve performance significantly.
Overall, while there is early work that propose a combination of language and image attention for
VQA, e.g., [15, 25, 12], attention mechanism with several potentials haven’t been discussed in detail
yet. In the following we present our approach for joint attention over any number of modalities.
3
Higher order attention models
Attention modules are a crucial component for present day decision making systems. Particularly
when taking into account more and more data of different modalities, attention mechanisms are able
to provide insights into the inner workings of the oftentimes abstract and automatically extracted
representations of our systems.
An example of such a system that captured a lot of research efforts in recent years is Visual Question
Answering (VQA). Considering VQA as an example, we immediately note its dependence on two or
even three different data modalities, the visual input V , the question Q and the answer A, which get
processed simultaneously. More formally, we let
V ∈ Rnv ×d , Q ∈ Rnq ×d , A ∈ Rna ×d
denote a representation for the visual input, the question and the answer respectively. Hereby, nv , nq
and na are the number of pixels, the number of words in the question, and the number of possible
answers. We use d to denote the dimensionality of the data. For simplicity of the exposition we
assume d to be identical across all data modalities.
Due to this dependence on multiple data modalities, present day decision making systems can be
decomposed into three major parts: (i) the data embedding; (ii) attention mechanisms; and (iii) the
decision making. For a state-of-the-art VQA system such as the one we developed here, those three
parts are immediately apparent when considering the high-level system architecture outlined in Fig. 2.
3.1
Data embedding
Attention modules deliver to the decision making component a succinct representation of the relevant
data modalities. As such, their performance depends on how we represent the data modalities
themselves. Oftentimes, an attention module tends to use expressive yet concise data embedding
algorithms to better capture their correlations and consequently to improve the decision making
performance. For example, data embeddings based on convolutional deep nets which constitute the
state-of-the-art in many visual recognition and scene understanding tasks. Language embeddings
heavily rely on LSTM which are able to capture context in sequential data, such as words, phrases
and sentences. We give a detailed account to our data embedding architectures for VQA in Sec. 4.1.
3
Yes
MCB
MCB
Unary
Potential
Decision (Sec. 3.3)
MCB
Ternary
Potential
Pairwise
Potential
Concatenate
LSTM
LSTM
1D-Conv
Word
Embedding
Unary
Potential
Softmax
Softmax
Softmax
Pairwise
Potential
Unary
Potential
Pairwise
Potential
Word
Embedding
ResNet
1. Yes
2. Yellow
…
17. No
18. Food
Is the dog trying
to catch a frisbee?
Attention (Sec. 3.2)
Data embedding (Sec. 3.1)
Figure 2: Our state-of-the-art VQA system
3.2
Attention
As apparent from the aforementioned description, attention is the crucial component connecting data
embeddings with decision making modules.
Subsequently we denote attention over the nq words in the question via PQ (iq ), where iq ∈
{1, . . . , nq } is the word index. Similarly, attention over the image is referred to via PV (iv ),
where iv ∈ {1, . . . , nv }, and attention over the possible answers are denoted PA (ia ), where
ia ∈ {1, . . . , na }.
We consider the attention mechanism as a probability model, with each attention mechanism computing “potentials.” First, unary potentials θV , θQ , θA denote the importance of each feature (e.g.,
question word representations, multiple choice answers representations, and image patch features)
for the VQA task. Second, pairwise potentials, θV,Q , θV,A , θQ,A express correlations between two
modalities. Last, third-order potential, θV,Q,A captures dependencies between the three modalities.
To obtain marginal probabilities PQ , PV and PA from potentials, our model performs mean-field
inference. We combine the unary potential, the marginalized pairwise potential and the marginalized
third order potential linearly including a bias term:
PV (iv ) = smax(α1 θV (iv )+α2 θV,Q (iv )+α3 θA,V (iv )+α4 θV,Q,A (iv ) + α5),
PQ (iq ) = smax(β1 θQ (iq )+β2 θV,Q (iq )+β3 θA,Q (iq )+β4 θV,Q,A (iq ) + β5),
PA (ia ) = smax(γ1 θA (ia )+γ2 θA,V (ia )+γ3 θA,Q (ia )+γ4 θV,Q.A (ia ) + γ5).
(1)
Hereby αi , βi , and γi are learnable parameters and smax(·) refers to the soft-max operation over
iv ∈ {1, . . . , nv }, iq ∈ {1, . . . , nq } and ia ∈ {1, . . . , na } respectively. The soft-max converts the
combined potentials to probability distributions, which corresponds to a single mean-field iteration.
Such a linear combination of potentials provides extra flexibility for the model, since it can learn
the reliability of the potential from the data. For instance, we observe that question attention relies
more on the unary question potential and on pairwise question and answer potentials. In contrast, the
image attention relies more on the pairwise question and image potential.
Given the aforementioned probabilities PV , PQ , and PA , the attended image, question and answer
vectors are denoted by aV ∈ Rd , aQ ∈ Rd and aA ∈ Rd . The attended modalities are calculated
as the weighted sum of the image features V = [v1 , . . . , vnv ]T ∈ Rnv ×d , the question features
Q = [q1 , . . . , qnq ]T ∈ Rnq ×d , and the answer features A = [a1 , . . . , ana ]T ∈ Rna ×d , i.e.,
aV =
nv
X
iv =1
PV (iv )viv ,
aQ =
nq
X
PQ (iq )qiq ,
iq =1
and
aV =
na
X
ia =1
4
PA (ia )aia .
tanh
conv
tanh
θQ,V,A (iq )
conv
tanh
θQ,V,A (iv )
conv
tanh
tanh
conv
conv
θQ,V (iv ) A
conv
Corr3
conv
V
Unary Potential
conv
conv
conv
θQ,V (iq )
conv
conv
tanh
V
Ternary Potential
Unary Potential
tanh
conv
Corr2
θV
conv
conv
tanh
V
Pairwise Potential
tanh
conv
Q
Unary Potential
conv
tanh
conv
θQ,V,A (ia )
conv
tanh
conv
tanh
conv
tanh
Corr3
tanh
The attended modalities, which effectively focus on the data relevant for the task, are passed to a
classifier for decision making, e.g., the ones discussed in Sec. 3.3. In the following we now describe
the attention mechanisms for unary, pairwise and ternary potentials in more detail.
conv
conv
module (e.g., visual and question) marginalized over its two data modalities. (c) ternary attention module (e.g.,
visual, question and answer) marginalized over its three data modalities..
conv
tanh
(a)
(b)
(c) Unary Pairwise
PairwisePotential
Ternary
Unary
Threeway
Threeway
Potential
Potential
Potential
Potential
Figure 3: IllustrationPotential
of our k−order
attention.
(a) unary attention module (e.g.,
visual).
(b)Potential
pairwise attention
conv
tanh
conv
Corr3
conv
Ternary Potential
Q
conv
Corr2
tanh
conv
Pairwise Potential
tanh
conv
Corr2
conv
Pairwise Potential
3.2.1 Unary potentials
We illustrate the unary attention schematically in Fig. 3 (a). The input to the unary attention module
is a data representation, i.e., either the visual representation V , the question representation Q, or the
answer representation A. Using those representations, we obtain the ‘unary potentials’ θV , θQ and
θA using a convolution operation with kernel size 1 × 1 over the data representation as an additional
embedding step, followed by a non-linearity (tanh in our case), followed by another convolution
operation with kernel size 1 × 1 to reduce embedding dimensionality. Since convolutions with kernel
size 1 × 1 are identical to matrix multiplies we formally obtain the unary potentials via
θV (iv ) = tanh(V Wv2 )Wv1 ,
θQ (iq ) = tanh(QWq2 )Wq1 ,
θA (ia ) = tanh(AWa2 )Wa1 .
where Wv1 , Wq1 , Wa1 ∈ Rd×1 , and Wv2 , Wq2 , Wa2 ∈ Rd×d are trainable parameters.
3.2.2 Pairwise potentials
Besides the mentioned mechanisms to generate unary potentials, we specifically aim at taking
advantage of pairwise attention modules, which are able to capture the correlation between the
representation of different modalities. Our approach is illustrated in Fig. 3 (b). We use a similarity
matrix between image and question modalities C2 = QWq (V Wv )> . Alternatively, the (i, j)-th entry
is the correlation (inner-product) of the i-th column of QWq and the j-th column of V Wv :
(C2 )i,j = corr2 ((QWq ):,i , (V Wv ):,j ),
corr2 (q, v) =
d
X
q l vl .
l=1
where Wq , Wv ∈ Rd×d are trainable parameters. We consider (C2 )i,j as a pairwise potential that
represents the correlation of the i-th word in a question and the j-th patch in an image. Therefore, to
retrieve the attention for a specific word, we convolve the matrix along the visual dimension using a
1 × 1 dimensional kernel. Specifically,
!
nq
nv
X
X
θV,Q (iq ) = tanh
wiv (C2 )iv ,iq , and θV,Q (iv ) = tanh
wiq (C2 )iv ,iq .
iv =1
iq =1
Similarly, we obtain θA,V and θA,Q , which we omit due to space limitations. These potentials are
used to compute the attention probabilities as defined in Eq. (1).
3.2.3
Ternary Potentials
To capture the dependencies between all three modalities, we consider their high-order correlations.
(C3 )i,j,k = corr3 ((QWq ):,i , (V Wv ):,j , (AWa ):,k ),
corr3 (q, v, a) =
d
X
l=1
5
ql vl al .
Threeway
Potential
Unary
Potential
Pairwise
Potential
v
h
v
T
P
Threeway
Potential
aV
aA
Outer Product
Space
Outer Product
(b)
Space
(a)
Outer Product
Space
MCB
MCB
MCT
aV
MCT
aQ
aQ
Pairwise
Potential
Unary
Potential
Figure 4: Illustration of correlation units used for decision making. (a) MCB unit approximately sample from
outer product space of two attention vectors, (b) MCT unit approximately sample from outer product space of
three attention vectors.
Where Wq , Wv , Wa ∈ Rd×d are trainable parameters. Similarly to the pairwise potentials, we use
the C3 tensor to obtain correlated attention for each modality:
!
nq na
nv X
na
X
X
X
θV,Q,A (iq ) = tanh
wiv ,ia (C3 )iq ,iv ,ia , θV,Q,A (iv ) = tanh
wiq ,ia (C3 )iq ,iv ,ia
iv =1 ia =1
iq =1 ia =1
and
θV,Q,A (ia ) = tanh
nq
nv X
X
wiq ,ia (C3 )iq ,iv ,ia .
iv =1 iq =1
These potentials are used to compute the attention probabilities as defined in Eq. (1).
3.3
Decision making
The decision making component receives as input the attended modalities and predicts the desired
output. Each attended modality is a vector that consists of the relevant data for making the decision.
While the decision making component can consider the modalities independently, the nature of
the task usually requires to take into account correlations between the attended modalities. The
correlation of a set of attended modalities are represented by the outer product of their respective
vectors, e.g., the correlation of two attended modalities is represented by a matrix and the correlation
of k-attended modalities is represented by a k-dimensional tensor.
Ideally, the attended modalities and their high-order correlation tensors are fed into a deep net which
produces the final decision. The number of parameters in such a network grows exponentially in
the number of modalities, as seen in Fig. 4. To overcome this computational bottleneck, we follow
the tensor sketch algorithm of Pham and Pagh [21], which was recently applied to attention models
by Fukui et al. [7] via Multimodal Compact Bilinear Pooling (MCB) in the pairwise setting or
Multimodal Compact Trilinear Pooling (MCT), an extension of MCB that pools data from three
modalities. The tensor sketch algorithm enables us to reduce the dimension of any rank-one tensor
while referring to it implicitly. It relies on the count sketch technique [4] that randomly embeds an
attended vector a ∈ Rd1 into another Euclidean space Ψ(a) ∈ Rd2 . The tensor sketch algorithm
then projects the rank-one tensor ⊗ki=1 ai which consists of attention correlations of order k using
the convolution Ψ(⊗ki=1 ai ) = ∗ki=1 Ψ(ai ). For example, for two attention modalities, the correlation
d2
matrix a1 a>
by the convolution Ψ(a1 ⊗a2 ) = Ψ(a1 )∗Ψ(a2 ).
2 = a1 ⊗a2 is randomly projected to R
The attended modalities Ψ(ai ) and their high-order correlations Ψ(⊗ki=1 ai ) are fed into a fully
connected neural net to complete decision making.
4
Visual question answering
In the following we evaluate our approach qualitatively and quantitatively. Before doing so we
describe the data embeddings.
4.1
Data embedding
The attention module requires the question representation Q ∈ Rnq ×d , the image representation
V ∈ Rnv ×d , and the answer representation A ∈ Rna ×d , which are computed as follows.
Image embedding: To embed the image, we use pre-trained convolutional deep nets (i.e., VGG-19,
ResNet). We extract the last layer before the fully connected units. Its dimension in the VGG net
case is 512 × 14 × 14 and the dimension in the ResNet case is 2048 × 14 × 14. Hence we obtain
6
Table 1: Comparison of results on the Multiple-Choice VQA dataset for a variety of methods. We
observe the combination of all three unary, pairwise and ternary potentials to yield the best result.
test-dev
test-std
Method
Y/N
Num
Other
All
All
HieCoAtt (VGG) [15]
HieCoAtt (ResNet) [15]
RAU (ResNet) [20]
MCB (ResNet) [7]
DAN (VGG) [19]
DAN (ResNet) [19]
MLB (ResNet) [13]
79.7
79.7
81.9
-
40.1
40.0
41.1
-
57.9
59.8
61.5
-
64.9
65.8
67.7
68.6
67.0
69.1
-
66.1
67.3
69.0
68.9
Unary+Pairwis (ResNet)
80.9
Unary+Pairwise (ResNet)
82.0
Unary + Pairwise + Ternary (VGG)
81.2
Unary + Pairwise + Ternary (ResNet) 81.6
36.0
42.7
42.7
43.3
61.6
63.3
62.3
64.8
66.7
68.7
67.9
69.4
68.7
69.3
2-Modalities:
3-Modalities:
3-Modalities:
3-Modalities:
nv = 196 and we embed both the 196 VGG-19 or ResNet features into a d = 512 dimensional space
to obtain the image representation V .
Question embedding: To obtain a question representation, Q ∈ Rnq ×d , we first map a 1-hot encoding of each word in the question into a d-dimensional embedding space using a linear transformation
plus corresponding bias terms. To obtain a richer representation that accounts for neighboring words,
we use a 1-dimensional temporal convolution with filter of size 3. While a combination of multiple
sized filters is suggested in the literature [15], we didn’t find any benefit from using such an approach.
Subsequently, to capture long-term dependencies, we used a Long Short Term Memory (LSTM)
layer. To reduce overfitting caused by the LSTM units, we used two LSTM layers with d/2 hidden
dimension, one uses as input the word embedding representation, and the other one operates on
the 1D conv layer output. Their output is then concatenated to obtain Q. We also note that nq is a
constant hyperparameter, i.e., questions with more than nq words are cut, while questions with less
words are zero-padded.
Answer embedding: To embed the possible answers we use a regular word embedding. The
vocabulary is specified by taking only the most frequent answers in the training set. Answers that
are not included in the top answers are embedded to the same vector. Answers containing multiple
words are embedded as n-grams to a single vector. We assume there is no real dependency between
the answers, therefore there is no need of using additional 1D conv, or LSTM layers.
4.2
Decision making
For our VQA example we investigate two techniques to combine vectors from three modalities. First,
the attended feature representation for each modality, i.e., aV , aA and aQ , are combined using an
MCT unit. Each feature element is of the form ((aV )i · (aQ )j · (aA )k ). While this first solution
is most general, in some cases like VQA, our experiments show that it is better to use our second
approach, a 2-layer MCB unit combination. This permits greater expressiveness as we employ
features of the form ((aV )i · (aQ )j · (aQ )k · (aA )t ) therefore also allowing image features to interact
with themselves. Note that in terms of parameters both approaches are identical as neither MCB nor
MCT are parametric modules.
Beyond MCB, we tested several other techniques that were suggested in the literature, including
element-wise multiplication, element-wise addition and concatenation [13, 15, 11], optionally followed by another hidden fully connected layer. The tensor sketching units consistently performed
best.
4.3
Results
Experimental setup: We use the RMSProp optimizer with a base learning rate of 4e−4 and α = 0.99
as well as = 1e−8 . The batch size is set to 300. The dimension d of all hidden layers is set to 512.
The MCB unit feature dimension was set to d = 8192. We apply dropout with a rate of 0.5 after the
word embeddings, the LSTM layer, and the first conv layer in the unary potential units. Additionally,
for the last fully connected layer we use a dropout rate of 0.3. We use the top 3000 most frequent
7
How many glasses are on the table?
How many glasses are on the table?
How many glasses are on the table?
Is anyone in the scene wearing blue?
Is anyone in the scene wearing blue?
Is anyone in the scene wearing blue?
What kind of flooring is in the bathroom?
What kind of flooring is in the bathroom?
What kind of flooring is in the bathroom?
What room is this?
What room is this?
What room is this?
st
Figure 5: For each image (1 column) we show the attention generated for two different questions in columns
2-4 and columns 5-7 respectively. The attentions are ordered as unary attention, pairwise attention and combined
attention for both the image and the question. We observe the combined attention to significantly depend on the
question.
Is this animal drinking water?
What kind of animal is this?
no
... this
red
no
yes
white
forks
4
1
tomatoes
presidential
blue
3
13
green
2
fila
i ... don't
Is this animal drinking water?
0.00
blue
red
cutting ... cake
green
bear
1
white
objazd
3
elephant
4
giraffe
yes
reject
cow
2
spain
no
0.02
0.04
0.06
0.08
Attention
0.10
0.12
0.14
What kind of animal is this?
0.00
0.02
0.04
0.00
0.08 0.10
Attention
0.12
0.14
0.16
3
not
white
1
on ... boy's
red
yes
4
no
2
aspro
pimp
player
blue
pain
if ... you've
green
no ... image
yes
next ... to
blue
green
parka
3
1
pirates
gadzoom
picture ... of
2
picture
clock
no
4
white
photo
red
What is on the wall?
0.06
Is a light on?
What is on the wall?
0.02
0.04
0.06
0.08 0.10
Attention
0.12
0.14
0.16
Is a light on?
0.00
0.02
0.04
0.06
Attention
0.08
0.10
0.12
Figure 6: The attention generated for two different questions over three modalities. We find the attention over
multiple choice answers to emphasis the unusual answers.
answers as possible outputs, which covers 91% of all answers in the train set. We implemented our
models using the Torch framework1 [5].
As a comparison for our attention mechanism we use the approach of Lu et al. [15] and the technique
of Fukui et al. [7]. Their methods are based on a hierarchical attention mechanism and multi-modal
compact bilinear (MCB) pooling. In contrast to their approach we demonstrate a relatively simple
technique based on a probabilistic intuition grounded on potentials. For comparative reasons only,
the visualized attention is based on two modalities: image and question.
We evaluate our attention modules on the VQA real-image test-dev and test-std datasets [2]. The
dataset consists of 123, 287 training images and 81, 434 test set images. Each image comes with 3
questions along with 18 multiple choice answers.
Quantitative evaluation: We first evaluate the overall performance of our model and compare it to a
variety of baselines. Tab. 1 shows the performance of our model and the baselines on the test-dev and
the test-standard datasets for multiple choice (MC) questions. To obtain multiple choice results we
follow common practice and use the highest scoring answer among the provided ones. Our approach
(Fig. 2) for the multiple choice answering task achieved the reported result after 180,000 iterations,
which requires about 40 hours of training on the ‘train+val’ dataset using a TitanX GPU. Despite
the fact that our model has only 40 million parameters, while techniques like [7] use over 70 million
parameters, we observe state-of-the-art behavior. Additionally, we employ a 2-modality model having
a similar experimental setup. We observe a significant improvement for our 3-modality model, which
shows the importance of high-order attention models. Due to the fact that we use a lower embedding
dimension of 512 (similar to [15]) compared to 2048 of existing 2-modality models [13, 7], the
2-modality model achieves inferior performance. We believe that higher embedding dimension and
proper tuning can improve our 2-modality starting point.
Additionally, we compared our proposed decision units. MCT, which is a generic extension of MCB
for 3-modalities, and 2-layers MCB which has greater expressiveness (Sec. 4.2). Evaluating on the
’val’ dataset while training on the ’train’ part using the VGG features, the MCT setup yields 63.82%
1
https://github.com/idansc/HighOrderAtten
8
Is she using a battery-operated device?
Is this a boy or a girl?
Is she using a battery-operated device?
Is she using a
battery device?
Ours: yes
[15]: no
[7]: no
GT: yes
Is this a boy or a girl?
Is this boy
or a girl?
Ours: girl
[15]: boy
[7]: girl
GT: girl
Figure 7: Comparison of our attention results (2nd column) with attention provided by [15] (3rd column)
and [7] (4th column). The fourth column provides the question and the answer of the different techniques.
What color is
the table?
GT: brown
Ours: blue
What color is the table?
What color is the table?
What color is the table?
What color is
the umbrella?
GT: blue
Ours: blue
What color is the umbrella?
What color is the umbrella?
What color is the umbrella?
Figure 8: Failure cases: Unary, pairwise and combined attention of our approach. Our system
focuses on the colorful umbrella as opposed to the table in the first row.
where 2-layer MCB yields 64.57%. We also tested a different ordering of the input to the 2-modality
MCB and found them to yield inferior results.
Qualitative evaluation: Next, we evaluate our technique qualitatively. In Fig. 5 we illustrate the
unary, pairwise and combined attention of our approach based on the two modality architecture,
without the multiple choice as input. For each image we show multiple questions. We observe the
unary attention usually attends to strong features of the image, while pairwise potentials emphasize
areas that correlate with question words. Importantly, the combined result is dependent on the
provided question. For instance, in the first row we observe for the question “How many glasses are
on the table?,” that the pairwise potential reacts to the image area depicting the glass. In contrast, for
the question “Is anyone in the scene wearing blue?” the pairwise potentials reacts to the guy with the
blue shirt. In Fig. 6, we illustrate the attention for our 3-modality model. We find the attention over
multiple choice answers to favor the more unusual results.
In Fig. 7, we compare the final attention obtained from our approach to the results obtained with
techniques discussed in [15] and [7]. We observe that our approach attends to reasonable pixel and
question locations. For example, considering the first row in Fig. 7, the question refers to the battery
operated device. Compared to existing approaches, our technique attends to the laptop, which seems
to help in choosing the correct answer. In the second row, the question wonders “Is this a boy or a
girl?”. Both of the correct answers were produced when the attention focuses on the hair.
In Fig. 8, we illustrate a failure case, where the attention of our approach is identical, despite two
different input questions. Our system focuses on the colorful umbrella as opposed to the object
queried for in the question.
5
Conclusion
In this paper we investigated a series of techniques to design attention for multimodal input data.
Beyond demonstrating state-of-the-art performance using relatively simple models, we hope that this
work inspires researchers to work in this direction.
9
Acknowledgments: This research was supported in part by The Israel Science Foundation (grant
No. 948/15). This material is based upon work supported in part by the National Science Foundation
under Grant No. 1718221. We thank Nvidia for providing GPUs used in this research.
References
[1] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural networks
for question answering. arXiv preprint arXiv:1601.01705, 2016.
[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick,
and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015.
[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning
to align and translate. arXiv preprint arXiv:1409.0473, 2014.
[4] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data streams. In ICALP.
Springer, 2002.
[5] Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A matlab-like environment for
machine learning. In BigLearn, NIPS Workshop, number EPFL-CONF-192376, 2011.
[6] Abhishek Das, Harsh Agrawal, C Lawrence Zitnick, Devi Parikh, and Dhruv Batra. Human attention
in visual question answering: Do humans and deep networks look at the same regions? arXiv preprint
arXiv:1606.03556, 2016.
[7] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach.
Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint
arXiv:1606.01847, 2016.
[8] Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,
and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pages 1693–1701, 2015.
[9] Allan Jabri, Armand Joulin, and Laurens van der Maaten. Revisiting visual question answering baselines.
In ECCV. Springer, 2016.
[10] U. Jain∗ , Z. Zhang∗ , and A. G. Schwing. Creativity: Generating Diverse Questions using Variational
Autoencoders. In CVPR, 2017. ∗ equal contribution.
[11] Vahid Kazemi and Ali Elqursh. Show, ask, attend, and answer: A strong baseline for visual question
answering. arXiv preprint arXiv:1704.03162, 2017.
[12] Jin-Hwa Kim, Sang-Woo Lee, Donghyun Kwak, Min-Oh Heo, Jeonghee Kim, Jung-Woo Ha, and ByoungTak Zhang. Multimodal residual learning for visual qa. In NIPS, 2016.
[13] Jin-Hwa Kim, Kyoung-Woon On, Jeonghee Kim, Jung-Woo Ha, and Byoung-Tak Zhang. Hadamard
product for low-rank bilinear pooling. arXiv preprint arXiv:1610.04325, 2016.
[14] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M Rush. Structured attention networks. arXiv
preprint arXiv:1702.00887, 2017.
[15] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for
visual question answering. In NIPS, 2016.
[16] Lin Ma, Zhengdong Lu, and Hang Li. Learning to answer questions from image using convolutional neural
network. arXiv preprint arXiv:1506.00333, 2015.
[17] Mateusz Malinowski, Marcus Rohrbach, and Mario Fritz. Ask your neurons: A neural-based approach to
answering questions about images. In ICCV, 2015.
[18] Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende.
Generating natural questions about an image. arXiv preprint arXiv:1603.06059, 2016.
[19] Hyeonseob Nam, Jung-Woo Ha, and Jeonghee Kim. Dual attention networks for multimodal reasoning
and matching. arXiv preprint arXiv:1611.00471, 2016.
[20] Hyeonwoo Noh and Bohyung Han. Training recurrent answering units with joint loss minimization for
vqa. arXiv preprint arXiv:1606.03647, 2016.
[21] Ninh Pham and Rasmus Pagh. Fast and scalable polynomial kernels via explicit feature maps. In SIGKDD.
ACM, 2013.
10
[22] Tim Rocktäschel, Edward Grefenstette, Moritz Hermann, Karl, Tomáš Kočiskỳ, and Phil Blunsom.
Reasoning about entailment with neural attention. In ICLR, 2016.
[23] Kevin J Shih, Saurabh Singh, and Derek Hoiem. Where to look: Focus regions for visual question
answering. In CVPR, 2016.
[24] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual
question answering. arXiv preprint arXiv:1603.01417, 2016.
[25] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for
visual question answering. In ECCV, pages 451–466. Springer, 2016.
[26] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel,
and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML,
2015.
[27] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image
question answering. In CVPR, 2016.
[28] Wenpeng Yin, Hinrich Schütze, Bing Xiang, and Bowen Zhou. Abcnn: Attention-based convolutional
neural network for modeling sentence pairs. arXiv preprint arXiv:1512.05193, 2015.
[29] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in
images. In CVPR, 2016.
11
| 2cs.AI
|
Double Deep Machine Learning
Moshe BenBassat ([email protected])
Arison School of Business, Interdisciplinary Center (IDC), Herzliya, Israel
Abstract
Very important breakthroughs in data-centric machine learning algorithms led to impressive performance in ‘transactional’
point applications such as detecting anger in speech, alerts from a Face Recognition system, or EKG interpretation. Nontransactional applications, e.g. medical diagnosis beyond the EKG results, require AI algorithms that integrate deeper and
broader knowledge in their problem-solving capabilities, e.g. integrating knowledge about anatomy and physiology of the heart
with EKG results and additional patient’s findings. Similarly, for military aerial interpretation, where knowledge about enemy
doctrines on force composition and spread helps immensely in situation assessment beyond image recognition of individual
objects.
An initiative is proposed to build Wikipedia for Smart Machines, meaning target readers are not human, but rather smart
machines. Named ReKopedia, the goal is to develop methodologies, tools, and automatic algorithms to convert humanity
knowledge that we all learn in schools, universities and during our professional life into Reusable Knowledge structures that
smart machines can use in their inference algorithms. Ideally, ReKopedia would be an open source shared knowledge repository
similar to the well-known shared open source software code repositories.
The Double Deep Learning approach advocates integrating data-centric machine self-learning techniques with machineteaching techniques to leverage the power of both and overcome their corresponding limitations. For illustration, an outline of a
$15M project is described to produce ReKo knowledge modules for medical diagnosis of about 1,000 disorders.
AI applications that are based solely on data-centric machine learning algorithms are typically point solutions for transactional
tasks that do not lend themselves to automatic generalization beyond the scope of the data sets they are based on. Today’s AI
industry is fragmented, and we are not establishing broad and deep enough foundations that will enable us to build higher level
‘generic’, ‘universal’ intelligence, let alone ‘super-intelligence’. We must find ways to create synergies between these fragments
and connect them with external knowledge sources, if we wish to scale faster the AI industry.
Examples in the article are based on- or inspired by- real-life non-transactional AI systems I deployed over decades of AI career
that benefit hundreds of millions of people around the globe. We are now in the second AI ‘spring’ after a long AI ‘winter’.
To avoid sliding again into an AI winter, it is essential that we rebalance the roles of data and knowledge. Data is
important but knowledge- deep and commonsense- are equally important.
1. Introduction
In recent years, Deep Learning (DL) algorithms achieved very important breakthroughs and outstanding results [10], [1];
primarily in image, speech and natural language understanding; leading to impressive performance in ‘transactional’ tasks such
as bank-check verification, detecting anger, alerts from a Face Recognition system, or EKG interpretation. DL algorithms use
data to automatically train neural networks to make intelligent inferences for tasks covered by the training data. They are based
on mathematical/statistical models which means that their behavior is predictable. Once trained to a certain level, they are likely
to perform consistently- within their statistical error boundaries- for cases within the scope of their training data. Another key
advantage of DL algorithms is the ‘no human touch feature engineering’ for pattern recognition tasks, meaning no need for
lengthy research projects requiring domain experts, e.g. fingerprints experts, or EKG experts, to find/extract good differentiating
features for classification decisions. Again, impressive achievements which are universal and domain-agnostic. During the early
part of my AI career, I could have certainly benefited from DL in many pattern recognition projects, including: object
recognition for a Ballistic Missile Defense system (cold war time), ultrasound wave recognition for an autonomous machine
digging coal on the moon, EKG, EEG/ERP waves, radar signals, and handwritten character recognition. I view DL’s
remarkable progress and achievements as key development in the future of AI; and as a mathematician, it is music to my ears.
However, with all due appreciation, DL algorithms have their limitations, as discussed below, specifically with ‘nontransactional’ applications that require broader and deeper reasoning; possibly involving multiple deep knowledge sources, e.g.
optimizing manufacturing and service operations, medical diagnosis and equipment troubleshooting, or military situation
assessment and mission planning. Early successes with DL led to overstating its applicability, reaching extreme claims such as
‘with DL and a sufficient amount of data you can solve all AI problems’. The examples bellow, inspired by real life AI
challenges I was presented with during my long AI career, illustrate why being radically religious about such claims limits the
progress of AI.
This article is NOT against Deep Learning’s algorithms and architecture; their track record speaks for itself. It is in favor of
enriching and amplifying DL, and AI in general, by using the wealth of knowledge that mankind developed in many fields over
thousands of years.
Following a discussion of Deep Learning’s current limitations for certain AI needs, I proceed to presenting the Double Deep
Learning approach and the idea of Wikipedia for Smart Machines as directions to overcome these limitations. The double
use of the word ‘deep’ refers to: first, data-centric deep learning as we all know it, and second, teaching computers deep
knowledge, like the difference between teaching physicians versus paramedics, or teaching engineers versus technicians. The
computer teaching process could potentially involve some automatic or semi-automatic learning from publications and other
documented sources.
“Soon We Won't Program Computers. We'll Train Them Like Dogs” (https://www.wired.com/2016/05/the-end-of-code/)
was the title of a June 2016 WIRED Magazine’s article focusing on the key difference between classic software programming
that provides explicit step by step instructions how to solve a problem, and Machine Learning software by which you provide
sample cases and a generic training algorithm that keeps iterating until the software learns how to solve the problem at a desired
performance level. The “dog” analogy was inspired by classic behaviorism studies: “…Pavlov triggered his dog's salivation not
through a deep understanding of hunger but simply by repeating a sequence of events over and over. He provided data, again
and again, until the code rewrote itself”. Paraphrasing on this title to describe the Double Deep approach, I would say: Soon We
Won't Program Computers. We'll Train Them Like We Train Human University Students with Universal Deep
Knowledge and with Sample Cases, e.g. medical students. Let us teach computers fundamental knowledge and equip them
with generic inference algorithms that will leverage these knowledge modules as they solve a specific case. In medical schools
we teach students (a) anatomy and physiology, (b) characteristics of specific diseases/disorders by means of
signs/symptoms/test results, along with sample cases, (c) how to execute a good diagnostic process, generically. Note that (a)
and (b) are knowledge modules, that are independent of (c) which is a generic inference process. In practice, when a patient
arrives he/she does not announce “I have Appendicitis”. He presents initial complaints/signs, and then a physician uses a
generic inference engine (c), while accessing knowledge modules (a) and (b) to drive a cost-effective diagnostic process. That’s
how I built with my team medical diagnostic systems for endocrinology, for emergency and critical care, for arthritis, for
space medicine and for toxicology [2], [3]. The same inference engine- based on Bayesian Inference Networks- was applied to
knowledge modules of different medical fields. (The anatomy and physiology part did not exist in the way I would do it today).
See more in the last sections of this article.
The Double Deep Learning approach could serve to build DL’s next generation solutions that integrate knowledge modules
from the wealth of humanity’s knowledge repository, and thus amplify and expand the applicability of DL. It is likely to
improve DL on both sides of the spectrum: increase quality and scope (breadth) of “good” decisions, and, not less important,
reduce the amount of glaring mistakes, such as those we see by contemporary personal assistants like Siri, Cortana, and Alexa.
(It is OK for an algorithm to make a mistake in a ‘twilight zone’ where a human professional may also err. It is totally
unacceptable for an algorithm to make glaring mistakes that even a human beginner would not make. More research effort
should be devoted to eliminating glaring mistakes by DL-based solutions as they raise very fundamental questions about the
intelligence of the algorithm and risk its credibility). Also, a lot has been said about the risks of DL machines taking over the
world. One way to start dealing with that, is to integrate at the output points of DL algorithms- (or within their neural net?)sanity checks that are based on external knowledge, to monitor DL’s output and, when needed, interfere with the ensuing
actions before they are executed, e.g. mistakenly shutting down a nuclear station, or a production line. Several examples are
given in the next two sections.
The article is based on lessons I learned over decades of AI-focused career. All examples are based on- or inspired by- real-life
non-transactional AI systems I deployed with my teams that benefit hundreds of millions of people around the globe. For
example, as of mid-2017, ClickSoftware's products schedule daily close to 750,000 Field Engineers (FE) for many of the
world’s largest service providers. Assuming that each engineer delivers on average 3 to 3.5 jobs per day, and works roughly 210
field days per year, this means that, over a year period, these products touch the life of about 500 Million people, which are
roughly 6% to 7% of the 7 Billion+ world population.
2. DL Limitations: Application-Oriented Overview
(a) When Sufficient Volume of Data does not Exist. DL requires data, massive amounts of data, e.g. thousands of
speech recording hours are needed to build a speech understanding system. For many business scenarios such volumes
of data simply do not exist (notwithstanding Big Data, see below).
Example 1: Troubleshooting New Equipment: Consider building an AI solution to support field service technicians
of a new complex medical imaging equipment that just came out. It will take several years before large and rich enough
fault data become available for training a DL-based solution that can guide service technicians with efficient fault
isolation and subsequent repair actions. Do we do nothing until sufficient data is available for a DL-based solution?
In fact, by the time sufficient data is accumulated, the current equipment model is about to be replaced by a newer
model. How would a DL algorithm know which data is only partially applicable, or no longer applicable?
Big Data. Internet of Things (IoT), connected vehicles, and other intense IT technologies are all around us producing
vast amounts of data serving as an excellent source for building AI solutions, e.g. context-aware real-time factory floor
optimization, or traffic capacity management; two AI-based solutions I am actively involved with. DL and other
machine learning technologies are perfect fit for many cases. While recognizing this potential, we should also be aware
though that ‘big data’ is sometimes not big enough for DL, as the following story illustrates:
In a recent business plan presentation of a young entrepreneur, he proposed using DL to recognize patterns of certain
business situations while referencing the recent successes in face recognition and animal classification as evidence to
the power of DL. When asked about the data he plans to use, he replied proudly and confidently “5 years of good
quality, comprehensive daily data (75 variables)”, which indeed is quite nice in a business environment. But this is just
under 2,000 cases (365*5); a small number of cases for a DL algorithm to produce useful results.
(b) Explainable AI. Today’s DL solutions operate like a “black-box” and even their developers cannot fully explain their
reasoning. For non-transactional applications in business, medicine, or military, explaining the reasoning is mandatory,
or at least very highly desirable. DARPA’ recent initiative into ‘explainable AI’ is very important [7],[9].
(c) Transactional vs non-Transactional Tasks. I used earlier the terms ‘transactional, and ’non-transactional’ tasks.
Rather than going into formal definitions, let me use EKG interpretation to clarify the difference.
Example 2: EKG Classification vs Medical Diagnosis. With thousands of EKG training data signals, a DL algorithm
can do an excellent job classifying an EKG signal shape, e.g. producing as output: Class = "ST Elevation”, see [12]
based on 64,000 EKG records. If the user now asks the DL algorithm to elaborate on the meaning of its output for
patient diagnosis, he/she is unlikely to receive a meaningful answer. A physician, on the other hand, will explain that
ST Elevation represents ventricular contraction and may indicate an artery clog that may damage the heart muscle
(Myocardial Infarction). The difference between the DL algorithm and the Physician is the level of understanding of
EKG findings. The physician’s answer is based on layers on top of layers of anatomy and physiology knowledge,
including the electrical impulses and their relationship to EKG findings. For an AI-based medical decision-making
system, just EKG classification is a narrow point solution (‘transactional’) to a small part of the problem. Such system
should go beyond signal analysis and also connect a given EKG shape to the way the heart functions and fails, and
integrate these with other patient findings.
Can a DL algorithm learn from patient data ONLY the anatomy and physiology of the human heart? That is: learn the
four chambers structure, the arteries, the valves, the conduction system and the pacemaker, the walls, …, the function
of each module and the overall blood flow. I doubt it simply because patient data does not contain the information
to enable such learning. (To appreciate the complexity, check https://www.youtube.com/watch?v=RYZ4daFwMa8 for
an excellent heart simulation that connects the human heart and EKG). Even with man-made fully documented
equipment, e.g. Semiconductor equipment, or Medical Imaging, the challenge for machine self-learning of equipment’s
structure, function and process flow is enormous. For medical diagnosis the challenge is higher because we are still
looking for the engineering design documentation of the human body, ….
As a scientist, I am all in favor of research to push further the spectrum of what data-only approaches can learn, while
also understanding its boundaries. As a business executive and AI practitioner, I believe that producing today a
working AI system for non-transactional tasks like diagnosis in cardiology and emergency medicine, a more promising
approach is to teach computers explicitly the anatomy/physiology knowledge like we teach human medical students,
rather than wait until, and if, a data-only DL algorithm will learn it from zero at a comparable level.
Several articles report that DL has been successful in medical diagnosis, e.g. for cancer. A fairer description of
the situation would be: DL algorithms support narrow aspects of the medical diagnosis process by providing
point solutions such as classifying an EKG signal or detecting tumors in a medical image, or searching for past
patients that are most similar to a given patient.
(d) Thinking versus Calculating, Newton’s Physics and DL. With data about
billions of “things”, and thousands of apples, that fall down every day, DL
algorithms, with no human touch, can certainly come up with a model to
calculate the time at which any given falling object will touch the ground.
But can DL today come up with Newton’s laws? I mean produce models that
represent deeper understanding of Earth forces along with “compact”
formula such as Time to hit the ground=SQRT(2*Hight)/g), where
g=gravity=9.8 (as opposed to a gigantic black box neural net)? If you are
only interested in building calculators for smart machines, you probably
would ask back: “Why do I need a compact formula if I get the right result
with a DL-generated neural network (possibly even more accurate if the data
also includes air resistance)?” My answer is that the importance of deep
understanding/grasping of Newton’s laws (mid 1600’s) goes way beyond a calculating tool for falling objects. For
example, the beautiful law of inertia is not at all about calculation: ‘An object at rest (or in motion) remains at rest (or
in motion), unless acted upon by an external force…’ The abstraction, generalization, and analytic formula of
Newton’s laws were the basis for future physicists to refine and expand them and discover new ones on Earth forces.
Similar processes in the 1800’s led to deeper understanding of heat and energy and the discovery of Thermodynamics
laws, then Electricity, continuing all the way to Plank and Einstein (1900-1905) with quantum physics and relativity
theory, and on to the Apollo project (1960’s), up until today, a time in which we also have fairly good understanding on
gravity forces in Space far away from Earth. In most cases, science discoveries were (a) based on sparks of brilliant
human theoretical thinking with very little data to learn from, as Sir Isaac Newton put it:” No great discovery was ever
made without a bold guess”, and (b) relied on what was already known; quoting Newton again: "If I have seen further,
it is by standing on the shoulders of giants”. As opposed to learning from zero, which is what a typical DL
algorithm does.
Why is this relevant to AI? Because physics is one of the cornerstones of engineering, which, in turn, is the basis of
all equipment around us that enable life in the modern world, e.g. agriculture equipment, food production equipment,
medical equipment, cars, airplanes, and of course, computers and mobile phones. AI systems to support equipment
maintenance and troubleshooting can improve substantially equipment’s uptime thereby offering great value. Having
led the development of the AITEST troubleshooting AI software and deployed it for dozens of complex equipment
around the globe [4], [5], I am confident that equipment-specific design knowledge, as well as universal engineering
knowledge, can greatly improve the performance of any ‘data-only’ smart machines. That’s because data can tell you
how equipment fails, equipment-specific design knowledge and universal engineering knowledge can tell you
how it works. Both types of knowledge are required to reach high performance for diagnostic and repair decisions. The
AITEST system illustrates the first steps in this direction, i.e. automatic conversion of engineering diagrams (topology
and test paths) into a Bayesian Inference network (on a massive scale).
In summary, it is one thing to learn from data how to recognize visual objects or acoustic signals. It is a totally
different challenge to derive from data full understanding of Newton’s laws, how to build a bridge, or how
equipment or human organs operate with their dynamic process flows.
(e) Glaring Mistakes: Every other day we hear jokes about glaring mistakes by AI personal assistants, e.g. Siri, Cortana,
Alexa, that quickly lead you to recognize the very limited understanding, scope and depth that the software has. DL
developers typically focus on maximizing overall percentage accuracy. How do you protect a DL-based smart machine
against glaring mistakes that even human beginners would not make? The more you push a DL algorithm for an overall
higher percentage accuracy, the higher the likelihood that glaring mistakes will sneak in (one manifestation of
overfitting).
While in some applications glaring mistakes may be something to joke about, in others, e.g. military air-defense, they
could be catastrophic. Just imagine shooting down a passenger airplane mistakenly classified as a threatening object.
Similarly, for potential misclassification mistakes by autonomous land vehicles or drones. It only takes one or a few
glaring mistakes to make users question the ‘true’ intelligence of the software to a point that the solution loses
credibility and is soon rejected/shelved (before a costly or catastrophic event makes it too late). Arguing that the overall
accuracy performance is within the, say 90%, promised error boundaries, does not help much.
Business software providers should specifically note that recovering from ‘unforgivable’ glaring-mistake events could
take a long time and be very costly for the business. The client may activate liability clauses in the contract, and worse,
the competition may run a whole marketing campaign around it to destroy your product’s reputation. That’s why I
always guide my teams with the following principle:
Glaring Mistake Protection Principle: In addition to working towards high overall accuracy, you should also always
include sanity checks to protect against glaring mistakes with individual cases. Before displaying the output of your
‘ultra-intelligent’ algorithm to a human user, or take automatic action, run sanity checks for extra protection.
Developers of autonomous vehicles are investing in multiple sensors, e.g. cameras, Lidar (Light and Radar), and
ultrasound, to overcome the limitations of each individual sensor technology. One more direction to overcome these
limitations is to add intelligent algorithms based on humanity knowledge, including common sense knowledge, as well
as deeper universal “world” knowledge, e.g. environmental, physics and engineering, and human/animal behavior.
Beyond ordinary glaring mistakes, such algorithms could also contribute to overcoming mistakes due to intentional
adversarial images designed to fool DL-based systems, see https://www.theverge.com/2017/11/2/16597276/googleai-image-attacks-adversarial-turtle-rifle-3d-printed in the context of confusing a turtle for a rifle.
External knowledge is a good way to equip DL-based solutions with ‘sanity checks’/ ‘second opinion’ to fight glaring
mistakes. The example in the next paragraph also illustrates this point.
(f) External Knowledge which is not in DL’s Data Set Could be Very Helpful: When a DL model does not produce
good enough results, those who are radically religious about ‘data-only’ ‘no human touch’ philosophy adopt a closed
garden DL doctrine and limit their options to improve performance to those within the DL world, such as: add data,
change neural net architecture, or augment the data with pre-processing operators, and then run again DL. Using
external knowledge or alternative inference algorithms are taboo. The following example provide good reasons when
and why to avoid the data-only doctrine:
Example 3: Aerial Image Interpretation; a Twilight Zone Case.
David and Abi- top notch experienced aerial image interpretation analysts at an Air-Force base- are faced with one of
those challenging cases where they cannot decide whether an object O in an image is equipment type A (agriculture) or
B (military). They have tried all options to enhance the image, but uncertainty is still very high. Sam, a security
specialist who stops by just to say hello, arrives at the peak of their heated debate on A or B. As he is waiting, he looks
at the picture and calmly says: “guys, coming from a farmer’s family, and judging by the terrain and vegetation, I
strongly doubt it is equipment A, because no farmer would use A in this situation”. After a short pause, Abi says “if
this is B, we have a major development”, and David replies: “Absolutely, I am going to wake up Jim (their boss)”.
Sam was using knowledge which is not in the picture data. Adopting the ‘data only’ doctrine of some contemporary DL
practitioners is like David and Abi ignoring Sam’s input. It limits the progress of DL. Well, a DL fan would now
suggest collecting more data that cover Sam’s farming knowledge and then re-run DL. That’s a good theoretical
exercise, but it does not go far, because equipment B only shows up about twice a week for a short while, and every
time with a slightly different silhouette, meaning about 100 pictures of B over a full year. Will this be sufficient for
DL? Indeed, data augmentation can also be used to fight the low volume of data, and, yes, separate DL networks can be
built to learn contexts where objects appear by terrain, vegetation, time of the year, etc., but how about simply asking
Sam? I mean explicitly embedding farmers’ knowledge in the AI solution. Why learn from data what humanity
already know?
Table 1 summarizes some of the above key messages and DL’s limitations. All rows in Table 1 are based on- or inspired byreal life ‘non-transactional’ AI systems I deployed and provide further evidence that knowledge plus data are likely to yield
higher AI performance than data only. See also Shoham [13]
Table 1: Data Tells you…, Knowledge tells you
Object
Data tells you…
MRI (Medical equipment)
How it fails
Frigate ZX8
What it looks like
EKG signal
Arrhythmia type C
Apples falling to the ground
How to calculate time to hit ground
Knowledge tells you…
How it works
Its capabilities
Relationship to heart’s anatomy, physiology
Deeper understanding of Earth forces
Where do we go from here?
3. Double Deep Learning
The maximum knowledge a Deep Learning (DL) algorithm can learn is encapsulated in its data set for the given task, which,
typically, is substantially less than the full humanity’s knowledge for the same task. The Double Deep Learning approach
advocates integrating data-centric machine self-learning techniques with machine-teaching techniques to leverage the power
of both, and overcome their corresponding limitations.
While the first ‘deep’ is for data-centric Deep Learning, the second ‘deep’ is for ‘machine teachers’ of knowledge and it stands
for extra focus on teaching deep ‘foundations’ and ‘first principles’ for reasoning in the task domain; in addition to shallow
prescriptive knowledge or just facts whenever needed. By ‘deep’ teaching I mean going beyond teaching just ‘shallow’
‘experiential’ knowledge which DL self-learning algorithm can also achieve if data is available. This was the main drawback of
early days rule-based expert systems of the 1970-80’s. By analogy it is like the difference between teaching physicians versus
teaching paramedics, or teaching engineers versus teaching technicians. Or quoting Aristotle: “Knowledge of the fact differs
from knowledge of the reason for the fact”. Or, as I discussed above in Example 2, knowing that EKG signal has ST changes, is
different than understanding the reasons for this finding, and what actions to take. In 2015, Dietterich and Horvitz [6] also
called to attention that:”…we have made surprisingly little progress to date on building the kinds of general intelligence that
experts and lay public envision when they think about ‘artificial Intelligence’”.
DARPA’s important initiative into ‘explainable AI’ [7] is also likely to contribute to the Double Deep Learning approach.
4. Wikipedia for Smart Machines
ReKopedia: An Open-Source Shared Knowledge Repository
ReKopedia Concept
Think about a repository that contains software representation structures of humanity’s science & technology knowledge in
various disciplines; call it the ReKopedia repository. ReKo stands for reusable knowledge. The representation structures
could be whatever we agree on, e.g. state transition graphs, neural nets, Bayesian nets, logic, fuzzy logic, frames, or rules, as
long as they are reusable by smart machines. I am proposing a community-wide initiative to establish an open-source shared
knowledge repository under which people contribute knowledge structures that are compatible with some protocols and enable
others to use them. Ideas from Service Oriented Architecture (SOA) and other software initiatives can serve as a basis to learn
from, e.g. self-contained modules, no need to know the inside technical details, and, of course, the concept of Web Services e.g.
SOAP, CORBA or REST. Modularity is key, together with mechanisms to combine modules into higher level knowledge
modules, which, when applied iteratively, create layers on top of layers of humanity knowledge.
In today’s shared economy spirit, where almost every AI algorithm can be found in open-source libraries like Python or R, the
ReKo repository can be a significant complement and energize the AI industry. The same way we agreed to share open
source software code, we will agree to share ReKo knowledge structures. Note that knowledge sharing via textbooks and
school teaching has already been a hallmark of mankind for generations. Let us do the same for smart machines.
In the early 2000’s, Google set a goal to scan all humanity’s published books. Now, in the AI era, let us embark on a similar
audacious goal to create a Wikipedia for Smart Machines (target readers are not human, but rather smart machines):
Goal - Humanity Knowledge in Software Structures: develop methodologies, tools, and automatic algorithms to convert
humanity’s documented knowledge into software structures that smart machines can use in their inference algorithms.
I am not talking about a monolithic centrally managed initiative, but rather a distributed self-organized initiative managed by a
mutually agreed upon governance.
ReKopedia Content
Content contribution to the ReKopedia repository can be made in any order and can come from everywhere, subject to some
covenant rules. As people build smart machines for a variety of applications, they will contribute knowledge modules to the
repository.
I can also envision work-groups in different disciplines, e.g. Medicine, Agriculture, Environment, Military, Engineering,
Manufacturing, Field Service, Finance, Insurance, Marketing, and Sales, each coming up with a long-term plan and priorities
for the content to populate their domain in the ReKopedia knowledge repository. Reviewing the syllabuses of schools and
universities in different areas will teach us the content we teach humans, and can be a good starting point for the content we
should teach machines. The CYC project [11], initiated by D. Lenat in the 1980’s with initial focus on commonsense
knowledge, led to a commercial product that offers extensive set of reusable knowledge modules and is available through his
company.
ReKo Representations and Algorithms
We do not have today all the answers to build ReKopedia for smart machines, in terms of knowledge and problem
representations and their corresponding algorithms, but we have very good foundations in Mathematics, Computer Science,
classic AI, Simulation theory, Management Science, Operations Research and related fields. The key principle, in my opinion,
is to maximize separation between the knowledge modules and the inference algorithms that will operate on them, as Example 4
illustrates for medical diagnosis.
Example 4: ReKopedia Modules for Medical Diagnosis: Representation and Practice
As an indication, the Bayesian Network structures (Figure 1) we developed for knowledge representation in the MEDAS [2]; an
AI system for emergency and critical care disorders and their inference algorithms (Figure 2), were also used for Space
medicine, for Arthritis, for Toxicology, and for other situation assessment applications beyond medicine. From years of
experience, I learned that, on average, about 50 human expert hours are needed to put into Bayesian Network templates the
knowledge for diagnosing a single medical disorder. This means that with 50,000 hours we can complete 1,000 disorders which
are likely to involve tens of thousands of symptoms, signs, syndromes, test results, and other findings. Assuming 50 to 100
MD’s working part time with support staff over 1 to 2 years, a $15M budget would take care of the expenses to build ReKo
modules for a non-marginal part of medicine. In fact, the MEDAS approach advocates hierarchical structures of disorders,
which means that after completing a base set, the average time per disorder will come down from 50 hours per disorder.
MEDAS reached convincing performance in its early stages [2], and by 1990 it reached “90% agreement with gold standard”
[8], long after I moved on to other areas. As for probability values on the links and nodes of the Bayesian Nets, we can start
with known values from medical publications and data bases, or expert’s subjective values, and then, as more patient data is
accumulated, apply machine learning algorithms to update them. The Bayesian Nets should be complemented and connected
with ReKo modules that capture knowledge about anatomy, bio-engineering, DNA, and other knowledge sources that can
improve the inferences made by the Bayesian inference algorithms.
Looking not too far in the future, where in large parts of the world patient data is automated starting at birth date, including
genome map data for every individual, ReKopedia-based AI systems can take intelligent healthcare automation to new
heights in terms of early warning, prevention, diagnosis and treatment, reducing cost while improving quality.
Automatic Learning of ReKo Structures from Humanity Documented Knowledge
Developing tools to automate the conversion process of natural language material; including diagrams and pictures, into ReKo
structures, can accelerate it considerably, but it requires taking Natural Language Understanding (NLU) to a higher level.
We are not there yet. To get an appreciation of the challenge, consider the task of automatic summary generation of
documents. The state of the art of this field tells us that AI NLU software is still far away from “truly understanding” what it
reads, let alone extracting knowledge from it, as compared with a college student reading a chapter in an Economy 101 textbook
and being able to solve homework exercises. This should not stop us, however, from starting to manually build the ReKo
repository for whichever field of science and technology we desire.
Figure 1: MEDAS- Hierarchical Structure of Medical Disorders [2]
Figure 2: Cycle of Diagnostic Assessment, BenBassat 1980’s [2], [3]
5. Summary
Today’s DL-based AI applications are typically point solutions for transactional tasks that do not lend themselves to automatic
generalization beyond the scope of the data sets they are based on. The AI industry is fragmented, and we are not establishing
broad and deep enough foundations that will enable us to build higher level ‘generic’, ‘universal’ intelligence, let alone ‘superintelligence’. We must find ways to create synergies between these fragments and connect them with external knowledge
sources, if we wish to scale faster the AI industry.
We are now in the second AI ‘spring’ after a long ‘winter’. To avoid sliding again into an AI winter, it is essential that
we rebalance the roles of data and knowledge. Data is important but knowledge- deep and commonsense- are equally
important.
If indeed AI is the driver of our next economic and social revolution (like electricity was), we'd better establish solid
foundations and infrastructure to develop and disseminate it; preferably with standards and fair economics.
Personal Note & Acknowledgments
This article is based on decades of my AI-focused career that are a blend of being a mathematician/statistician/computer
scientist (USC, Tel-Aviv University, UCLA), and being a business entrepreneur and 15 years CEO of ClickSoftware, a
NASDAQ (CKSW) public company. I have been researching, practicing and educating Artificial Intelligence from the first AI
“Spring” of the 1980’s, during the AI “Winter” of the 1990th and early 2000 years, and now in the AI renaissance of the 21th
century. My academic research was supported by NIH, NSF, DARPA, NASA, BMD (Ballistic Missile Defense Agency), ARI
(U.S. Army Research Institute), Israel Defense Forces, and others.
As the founder and CEO of ClickSoftware (inventor of service chain optimization- patent awarded, acquired in 2015 by a
private equity firm), and Plataine (a leader in AI and IoT-based solutions for manufacturing optimization), we leverage AI
technologies to solve large scale real-life business problems. We developed innovative AI products that benefit hundreds of
millions of people around the globe.
For more details and Publication List see http://www.moshebenbassat.com/
I am very grateful to Israel Beniaminy, my son Avner Ben-Bassat, and other colleagues for deep and useful discussions,
as well as comments on early drafts of this article.
References
[1] Brynjolfsson E. and Mcafee A. The Business of Artificial Intelligence, Harvard Business Review, July 2017
[2] Ben-Bassat M., Carlson R.W., Puri V.K., Davenport M. D., Schriver J. A, Latif M., Smith R., Portigal L. D., Lipnick E. and WeiI
M. H. Pattern-Based Interactive Diagnosis of Multiple Disorders: The MEDAS System. IEEE Transactions on Pattern Analysis
and Machine Intelligence, Vol. PAMI-2, No.2:148-160, March 1980
[3] M. Ben-Bassat, Expert Systems for Clinical Diagnosis, In Approximate Reasoning in Expert Systems. M.M. Gupta, A. Kandel,
W. Bandler, Y.s. Kiszka (eds.), North Holland, 1985, pp.671- 687.
[4] Ben-Bassat M. Use of Diagnostic Expert Systems in Aircarft Maintenance (9 real life examples), Proceedings of Aircraft
Maintenance and Engineering Conference, 1996, Singapore
[5] Ben-Bassat M., I Beniaminy, I., Joseph, D. Combining model-based and case-based expert systems, Research Perspectives and
Case Studies in System Test and Diagnosis, 179-205, 1998
[6] Dietterich and Horvitz, E. J. Rise of Concerns about AI: Reflections and Directions, Communications of the ACM, Volume 58,
Oct 2015, 38-40
[7] Gunning, D. Explainable Artificial Intelligence (XAI), http://www.cc.gatech.edu/~alanwags/DLAI2016/(Gunning)%20IJCAI16%20DLAI%20WS.pdf
[8] Georgakis, D. C. Trace, D.A. Naeymi-Rad, F. and Evens. A Statistical Evaluation of the Diagnostic Performance of MEDASThe Medical Emergency Decision Assistance System Proc Annu Symp Comput Appl Med Care. 1990 Nov 7 : 815–819.
[9] Knight, W. The Dark Secret at the Heart of AI, MIT Technology Review, April 11, 2017
[10] LeCun Y., Bengio Y, & Hinton G. Deep Learning. Nature 521, 436–444, May 2015
[11] Lenat, D.B. CYC: a large-scale investment in knowledge infrastructure, Communications of the ACM, Volume 38, Nov.
1995, 33-38
[12] Rajpurkar, P. Hannun, A. Y. Haghpanahi, M. Bourn, . Ng, A. Y. Cardiologist-Level Arrhythmia Detection with
Convolutional Neural Networks, https://arxiv.org/abs/1707.01836, July 2017
[13] Shoham, Y. Why Knowledge Representation Matters, Comm. ACM, 47-49, Jan 2016
------------------------------------------------------------------------------------------------------------------
Appendix: Key Messages
1. Machines can learn many things from data, but
data is not the only source machines can learn
from.
2. The maximum knowledge a Deep Learning (DL)
algorithm can learn is what is encapsulated in its
data set for the given task, which, typically, is
substantially less than the full humanity’s
knowledge for the same task. For instance, using
field service data, a DL algorithm can learn how
equipment fails, but DL cannot learn from such
data how it works.
3. Developing and deploying AI solutions when data
is not (yet) available is possible, with substantial
business value, by directly embedding explicit
humanity knowledge.
4. It is one thing to train a DL-neural-net to calculate
the time a falling apple will hit the ground using
data about falling apples. It is a totally different
challenge to train an algorithm from data only to
come up with Newton’s earth gravity laws. If
humans can learn from explicit teaching, why
can’t machines?
5. Machine learning technologies still have some
limitations, e.g. with non-supervised learning.
Why wait until- and if- someone trains a DL
algorithm to learn from data what humanity
already learned and documented in textbooks and
other publications? E.g. anatomy, physiology,
electrical conduction of human heart. Science-wise
it has merit. Business-wise it makes no sense.
6. In today’s AI world, data is over-rated,
knowledge is under-rated. By re-balancing the
two, AI solutions will benefit considerably. On the
other hand, by adopting ‘data-only’ doctrine, you
give up many options to improve DL’s
performance and expand its applicability.
7. The Double Deep Learning approach advocates
integrating ‘machine self-learning’ with ‘machine
teachers’. The second ‘deep’ is for ‘machine
teachers’ with extra focus on teaching deep
‘foundational’ ‘first principles’ knowledge aiming
at higher level intelligence, like the difference
between teaching physicians versus paramedics,
or teaching engineers versus technicians.
8. Being radically religious about a specific algorithm
(e.g. DL) aiming to push its applicability envelope
as far as possible, is good, even desirable, from a
science perspective. From a business perspective,
however, it may not be cost-effective, and may
even prevent you from deploying today business
solutions that deliver tremendous value.
9. Wikipedia for smart machines. AI can grow
faster by establishing an open-source shared
repository of reusable knowledge modules
(coined ReKopedia here) covering humanity’s
science & technology in various disciplines. For
illustration, a $15M project is proposed to produce
ReKopedia modules for medical diagnosis of
1,000 disorders.
| 2cs.AI
|
A Logarithmic Integrality Gap Bound for Directed Steiner Tree in
Quasi-bipartite Graphs ∗
Zachary Friggstad1 , Jochen Könemann2 , and Mohammad Shadravan3
arXiv:1604.08132v1 [cs.DS] 27 Apr 2016
1
Department of Computing Science, University of Alberta, Edmonton, AB, Canada, T6G 2E8,
[email protected]
2
Department of Combinatorics and Optimization, University of Waterloo, Waterloo, ON, Canada,
N2L 3G1, [email protected]
3
Department of Industrial Engineering and Operations Research, Columbia University, New
York, NY, USA, 10027, [email protected]
April 28, 2016
Abstract
We demonstrate that the integrality gap of the natural cut-based LP relaxation for the directed Steiner
tree problem is O(log k) in quasi-bipartite graphs with k terminals. Such instances can be seen to
generalize set cover, so the integrality gap analysis is tight up to a constant factor. A novel aspect
of our approach is that we use the primal-dual method; a technique that is rarely used in designing
approximation algorithms for network design problems in directed graphs.
1
Introduction
In an instance of the directed Steiner tree (DST) problem, we are given a directed graph G = (V, E),
non-negative costs ce for all e ∈ E, terminal nodes X ⊆ V , and a root r ∈ V . The remaining nodes in
V − (X ∪ {r}) are the Steiner nodes. The goal is to find the cheapest collection of edges F ⊆ E such that
for every terminal t ∈ X there is an r, t-path using only edges in F . Throughout, we let n denote |V | and k
denote |X|.
If X ∪ {r} = V , then the problem is simply the minimum-cost arborescence problem which can be
solved efficiently [4]. However, the general case is well-known to be NP-hard. In fact, the problem can be
seen to generalize the set-cover and group Steiner tree problems. The latter cannot be approximated within
O(log2− (n)) for any constant > 0 unless NP ⊆ DTIME(npolylog(n) ) [10].
For a DST instance G, let OP TG denote the value of the optimum solution for this instance Say that an
instance G = (V, E) of DST with terminals X is `-layered if V can be partitioned as V0 , V1 , . . . , V` where
V0 = {r}, V` = X and every edge uv ∈ E has u ∈ Vi and v ∈ Vi+1 for some 0 ≤ i < `. Zelikovsky
showed for any DST instance G and integer ` ≥ 1 that we can compute an `-layered DST instance H in
poly(n, `) time such that OP TG ≤ OP TH ≤ `·k 1/` ·OP TG and that a DST solution in H can be efficiently
mapped to a DST solution in G with the same cost [2, 17].
∗
This work was in part supported by NSERC’s Discovery grant program. The second author greatfully acknowledges the support
of the Hausdorff Institute and the Institute for Discrete Mathematics in Bonn, Germany.
1
Charikar et al. [3] exploited this fact and presented an O(`2 · k 1/` · log k)-approximation with running
time poly(n, k ` ) for any integer ` ≥ 1. In particular, this can be used to obtain an O(log3 k)-approximation
in quasi-polynomial time and a polynomial-time O(k )-approximation for any constant > 0. Finding a
polynomial-time polylogarithmic approximation remains an important open problem.
For a set of nodes S, we let δ in (S) = {uv ∈ V : u 6∈ S and v ∈ S} be the set of edges entering S. The
following is a natural linear programming (LP) relaxation for directed Steiner tree.
X
min
ce xe
(DST-Primal)
e∈E
in
s.t. x(δ (S)) ≥ 1 ∀ S ⊆ V − r, S ∩ X 6= ∅
(1)
xe ≥ 0 ∀e ∈ E
This LP is called a relaxation because of the natural correspondence between feasible solutions to a DST
instance G and feasible {0, 1}-integer solutions to the corresponding LP (DST-Primal). Thus, if we let
OP TLP denote the value of an optimum (possibly fractional) solution to LP (DST-Primal) then we have
OP TLP ≤ OP TG . For a particular instance G we say the integrality gap is OP TG /OP TLP ; we are
interested in placing the smallest possible upper bound on this quantity.
Interestingly, if |X| = 1 (the shortest path problem) or X ∪ {r} = V (the minimum-cost arborescence problem), the extreme points of (DST-Primal) are integral so the integrality gap is 1 ([12] and [4],
respectively). However, in the general case Zosin and Khuller showed that (DST-Primal) is not useful for
finding polylog(k)-approximation algorithms for DST [18]. √
The authors showed that the integrality gap of
(DST-Primal) relaxation can, unfortunately, be as bad as Ω( k), even in instances where G is a 4-layered
graph. In their examples, the number of nodes n is exponential in k so the integrality gap may still be
O(logc n) for some constant c.
On the other hand, Rothvoss recently showed that applying O(l) rounds of the semidefinite programming
Lasserre hierarchy to (the flow-based extended formulation of) (DST-Primal) yields an SDP with integrality
gap O(` · log k) for `-layered instances [14]. Subsequently, Friggstad et al. [6] showed similar results for
the weaker Sherali-Adams and Lovász-Schrijver linear programming hierarchies.
In this paper we consider the class of quasi-bipartite DST instances. An instance of DST is quasibipartite if the Steiner nodes V \(X ∪{r}) form an independent set (i.e., no directed edge has both endpoints
in V \ (X ∪ {r})). Such instances still capture the set cover problem, and thus do not admit an (1 − ) ln kapproximation for any constant > 0 unless P = NP [15, 5]. Furthermore, it is straightforward to adapt
known integrality gap constructions for set cover (e.g. [16]) to show that the integrality gap of (DST-Primal)
can be as bad as (1 − o(1)) · ln k in some instances. Hibi and Fujito [11] give an O(log k)-approximation
for quasi-bipartite instances of DST, but do not provide any integrality gap bounds.
Quasi-bipartite instances have been well-studied in the context of undirected Steiner trees. The class of
graphs was first introduced by Rajagopalan and Vazirani [13] who studied the integrality gap of (DST-Primal)
for the bidirected map of the given undirected Steiner tree instances. Currently, the best approximation for
73
quasi-bipartite instances of undirected Steiner tree is 60
by Goemans et al. [8] who also bound the integrality
gap of the bidirected cut relaxation by the same quantity. This is the same LP relaxation as (DST-Primal),
applied to the directed graph obtained by replacing each undirected edge {u, v} with the two directed edges
uv and vu. This is a slight improvement over a prior ( 73
60 + )-approximation for any constant > 0 by
Byrka et al. [1].
The best approximation for general instances of undirected Steiner tree is ln(4) + for any constant
> 0 [1]. However, the best known upper bound on the integrality gap of the bidirected cut relaxation
for non-quasi-bipartite instances is only 2; it is an open problem to determine if this integrality gap is a
constant-factor better than 2.
2
1.1
Our contributions
Our main result is the following. Let Hn =
Pn
= O(log n) be the nth harmonic number.
i=1 1/i
Theorem 1. The integrality gap of LP (DST-Primal) is at most 2 Hk = O(log k) in quasi-bipartite graphs
with k terminals. Furthermore, a Steiner tree with cost at most 2 Hk · OP TLP can be constructed in
polynomial time.
As noted above, Theorem 1 is asymptotically tight since any of the well-known Ω(log k) integrality gap
constructions for set cover instances with k items translate directly to an integrality gap lower bound for
(DST-Primal), using the usual reduction from set cover to 2-layered quasi-bipartite instances of directed
Steiner tree.
This integrality gap bound asymptotically matches the approximation guarantee proven by Hibi and
Fujito for quasi-bipartite DST instaces [11]. We remark that their approach is unlikely to give any integrality
gap bounds for (DST-Primal) because they iteratively choose low-density full Steiner trees in the same spirit
as [3] and give an O(` · log k)-approximation for finding the optimum DST solution T that does not contain
a path with ≥ ` Steiner nodes V \ (X ∪ {r}). In particular, their approach will also find an O(log k)approximation to the optimum
DST solution in 4-layered graphs and we know the integrality gap in some
√
4-layered instances is Ω( k) [18].
We prove Theorem 1 by constructing a directed Steiner tree in an iterative manner. An iteration starts
with a partial Steiner tree (see Definition 1 below), which consists of multiple directed components containing the terminals in X. Then a set of arcs are purchased to augment this partial solution to one with fewer
directed components. These arcs are discovered through a primal-dual moat growing procedure; a feasible
solution for the dual (DST-Primal) is constructed and the cost of the purchased arcs can be bounded using
this dual solution.
While the primal-dual technique has been very successful for undirected network design problems (e.g.,
see [9]), far fewer success stories are known in directed domains. Examples include a primal-dual interpretation of Dijkstra’s shortest path algorithm (e.g., see Chapter 5.4 of [12]), and Edmonds’ [4] algorithm
for minimum-cost arborescences. In both cases, the special structure of the problem is instrumental in the
primal-dual construction. One issue arising in the implementation of primal-dual approaches for directed
network design problems appears to be a certain overlap in the moat structure maintained by these algorithms. We are able to handle this difficulty here by exploiting the quasi-bipartite nature of our instances.
2
2.1
The integrality gap bound
Preliminaries and definitions
We now present an algorithmic proof of Theorem 1. As we will follow a primal-dual strategy, we first
present the LP dual of (DST-Primal).
X
max
yS
(DST-Dual)
S
s.t.
X
yS ≤ ce ∀e ∈ E
S:e∈δ in (S)
y ≥0
In (DST-Dual), the sums range only over sets of nodes S such that S ⊆ V − r and S ∩ X 6= ∅.
Our algorithm builds up partial solutions, which are defined as follows.
3
(2)
Definition 1. A partial Steiner tree is a tuple T = ({Bi , hi , Fi }`i=0 , B̄) where, for each 0 ≤ i ≤ `, Bi is a
subset of nodes, hi ∈ Bi , and Fi is a subset of edges with endpoints only in Bi such that the following hold.
• The sets B0 , B1 , . . . , B` , B̄ form a partition V .
• B̄ ⊆ V − X − r (i.e. B̄ is a subset of Steiner nodes).
• h0 = r and hi ∈ X for each 1 ≤ i ≤ `.
• For every 0 ≤ i ≤ ` and every v ∈ Bi , Fi contains an hi , v-path.
We say that B̄ is the set of free Steiner nodes in T and that hi is the head of Bi for each 0 ≤ i ≤ `. The
edges of T , denoted by E(T ), are simply ∪`i=0 Fi . We say that B0 , . . . , B` are the components of T where
B0 is the root component and B1 , . . . , B` are the non-root components.
Figure 1 illustrates a partial Steiner tree. Note that if T is a partial Steiner tree with ` = 0 non-root
components, then E(T ) is in fact a feasible DST solution.
P
Finally, for a subset of edges F we let cost(F ) = e∈F ce .
r
Figure 1: A partial Steiner tree with ` = 3 non-root components (the root is pictured at the top). The only
edges shown are those in some Fi . The white circles are the heads of the various sets Bi and the black
circles are terminals that are not heads of any components. The squares outside of the components are the
free Steiner nodes B̄. Note, in particular, that each head can reach every node in its respective component.
We do not require each Fi to be a minimal set of edges with this property.
4
2.2
High-level approach
Our algorithm builds up partial Steiner trees in an iterative manner while ensuring that the cost does not
increase by a significant amount between iterations. Specifically, we prove the following lemma in Section
3. Recall that OP TLP refers to the optimum solution value for (DST-Primal).
Lemma 1. Given a partial Steiner tree T with ` ≥ 1 non-root components, there is a polynomial-time
algorithm that finds a partial Steiner tree T 0 with `0 < ` non-root components such that
cost(E(T 0 )) ≤ cost(E(T )) + 2 · OP TLP ·
` − `0
.
`
Theorem 1 follows from Lemma 1 in a standard way.
Proof of Theorem 1. Initialize a partial Steiner tree Tk with k non-root components as follows. Let B̄ be the
set of all Steiner nodes, B0 = {r}, and F0 = ∅. Furthermore, label the terminals as t1 , . . . , tk ∈ X and for
each 1 ≤ i ≤ k let Bi = {ti }, hi = ti and Fi = ∅. Note that cost(E(Tk )) = 0.
Iterate Lemma 1 to obtain a sequence of partial Steiner trees T`0 , T`1 , T`2 , . . . , T`a where T`i has `i
non-root components such that k = `0 > `1 > . . . > `a = 0 and
cost(E(Ti+1 )) ≤ cost(E(Ti )) + 2 · OP TLP ·
`i − `i+1
`i
for each 0 ≤ i < a. Return E(T`a ) as the final Steiner tree.
That E(Ta ) can be found efficiently follows simply because we are iterating the efficient algorithm from
Lemma 1 at most k times. The cost of this Steiner tree can be bounded as follows.
cost(E(Ta )) ≤ 2 · OP TLP ·
≤ 2 · OP TLP ·
a−1
X
`i − `i+1
i=0
`i
a−1
X
`i
X
i=0 j=`i+1 +1
= 2 · OP TLP ·
a−1
X
`i
X
i=0 j=`i+1 +1
1
`i
k
X
1
1
= 2 · OP TLP ·
j
k
j=1
= 2 · OP TLP · Hk .
The idea presented above resembles one proposed by Guha et al. [7] for bounding the integrality gap
of a natural relaxation for undirected node-weighted Steiner tree by O(log k) [7]. Like our approach, Guha
et al. also build a solution incrementally. In each phase of the algorithm, the authors reduce the number of
connected components of a partial solution by adding vertices whose cost is charged carefully to the value
of a dual LP solution that the algorithm constructs simultaneously.
3
A primal-dual proof of Lemma 1
Consider a given partial Steiner tree T = ({Bi , hi , Fi }`i=0 , B̄) with ` ≥ 1 non-root components. Lemma
1 promises a partial Steiner tree T 0 with `0 < ` non-root components with cost(E(T 0 )) ≤ cost(E(T )) +
0
2 · OP TLP · `−`
` . In this section we will present an algorithm that augments forest T in the sense that it
computes a set of edges to add to T . The proof presented here is constructive: we will design a primal-dual
algorithm that maintains a feasible dual solution for (DST-Dual), and uses the structure of this solution to
guide the process of adding edges to T .
5
3.1
The algorithm
For any two nodes u, v ∈ V , let d(u, v) be the cost of the cheapest u, v-path in G. More generally, for a
subset ∅ ( S ⊆ V and a node v ∈ V we let d(S, v) = minu∈S d(u, v). We will assume that for every
0 ≤ i ≤ ` and 1 ≤ j ≤ `, j 6= i that d(Bi , hj ) > 0 as otherwise, we could merge Bi and Bj by adding the
0-cost Bi , hj -path to T .
The usual conventions of primal-dual algorithms will be adopted. We think of such an algorithm as
a continuous process that increases the value of some dual variables over time. At time t = 0, all dual
variables are initialized to a value of 0. At any point in time, exactly ` dual variables will be raised at a rate
of one unit per time unit. We will use ∆ for the time at which the algorithm terminates. As is customary, we
will say that an edge e goes tight if the dual constraint for e becomes tight as the dual variables are being
increased. When an edge goes tight, we will perform some updates to the various sets being maintained by
the algorithm. Again, the standard convention applies that if multiple edges go tight at the same time, then
we process them in any order.
Algorithm 1 describes the main subroutine that augments the partial Steiner tree T to one with fewer
components. It maintains a collection of moats Mi ⊆ V − {r} and edges Fi0 for each 1 ≤ i ≤ `, while
ensuring that the dual solution y it grows remains feasible. Mainly to aid notation, our algorithm will
maintain a so called virtual body βi for all 0 ≤ i ≤ ` such that Bi ⊆ βi ⊆ Bi ∪ B̄. We will ensure that each
v ∈ B̄ ∩ βi has a mate u ∈ Bi such that the edge uv has cost no more than ∆. For notational convenience,
we will let β0 = B0 be the virtual body of the root component. The algorithm will not grow a moat around
the root since dual variables do not exist for sets containing the root.
Our algorithm will ensure that moats are pairwise terminal-disjoint. In fact, we ensure that any two
moats may only intersect in B̄. Terminal-disjointness together with the quasi-bipartite structure of the input
graph will allow us to charge the cost of arcs added in the augmentation process to the duals grown.
An intuitive overview of our process is the following. At any time t ≥ 0, the moats Mi will consist of all
nodes v with d(v, hi ) ≤ t. The moats Mi will be grown until, at some time ∆, for at least one pair i, j with
i 6= j, there is a tight path connecting βj to hi . At this point the algorithm stops, and adds a carefully chosen
collection of tight arcs to the partial Steiner tree that merges Bj and Bi (and potentially other components).
Crucially, the cost of the added arcs will be charged to the value of the dual solution grown around the
merged components.
Due the structure of quasi-bipartite graphs, we are able to ensure that in each step of the algorithm the
active moats pay for at most one arc that is ultimately bought to form T 0 . Also, if T 0 has `0 < ` non-root
components then each arc was paid for by moats around at most ` − `0 + 1 ≤ 2(` − `0 ) different heads. So,
the total cost of all purchased arcs is at most 2(` − `0 ) · ∆. Finally, the total dual grown is ` · ∆, which is
0
≤ OP TLP due to feasibility, so the cost of the edges bought can be bounded by 2 `−`
` · OP TLP .
3.2
Algorithm and invariants
Now we will be more precise. The primal-dual procedure is presented in Algorithm 1. The following
invariants will be maintained at any time 0 ≤ t ≤ ∆ during the execution of Algorithm 1.
1. For each 1 ≤ i ≤ `, hi ∈ Mi and Mi ⊆ V − {r} (so there is a variable yMi in the dual).
2. For each 1 ≤ i ≤ `, Mi = {v ∈ V : d(v, hi ) < t} ∪ S where S ⊆ {v ∈ V : d(v, hi ) = t}.
3. Mi ∩ Mj ⊆ B̄ and both βi ∩ βj = Mi ∩ βj = ∅ for distinct 0 ≤ i, j ≤ `.
4. For each 1 ≤ i ≤ ` we have Bi ⊆ βi ⊆ Bi ∪ B̄. Furthermore, for each each v ∈ βi − Bi there is a
mate u ∈ Bi such that uv ∈ E and cuv ≤ t.
6
5. y is feasible for LP (DST-Dual) with value exactly ` · t.
These concepts are illustrated in Figure 2.
Figure 2: The moats around the two partial Steiner trees are depicted by the gray circles. The dashed edges
are those bought by the first moat and the solid edges are those bought by the second moat. Note the moats
only intersect in B (in particular, v is the only lying in both moats). Also, u lies in the virtual body for the
left partial Steiner tree and the dashed arc entering u is coming from its mate. The edges Fi from the original
partial Steiner trees are not shown. Observe that if any edge entering v goes tight then it must be from either
r or some terminal (because G is quasi-bipartite). This would allow us to merge at least one partial Steiner
tree into the body of another.
Algorithm 1 Dual Growing Procedure
1: Mi ← {v ∈ V : d(v, hi ) = 0}, 1 ≤ i ≤ `
2: βi ← Bi for 0 ≤ i ≤ `
3: y ← 0
4: Raise yMi0 uniformly for each moat Mi0 until some edge uv goes tight
5: if u ∈ βj for some 0 ≤ j ≤ ` and v ∈ Mi0 for some i0 6= j then
6:
return the partial Steiner tree T 0 described in Lemma 3.
7: else
8:
Let Mi be the unique moat with uv ∈ δ in (Mi )
9:
Mi ← Mi ∪ {u}
10:
if u ∈ βi then
11:
βi ← βi ∪ {v}
12:
go to Step (4)
3.3
. c.f. Proposition 1
Invariant analysis
Lemma 2. Invariants 1-5 are maintained by Algorithm 1 until the condition in the if statement in Step (5)
is true. Furthermore, the algorithm terminates in O(n · k) iterations.
Proof. Clearly the invariants are true after the initialization steps (at time t = 0), given that d(Bi , hj ) > 0
for any i 6= j. To see why Algorithm 1 terminates in a polynomial number of iterations, note that each
7
iteration increases the size some moat by 1 and does not decrease the size of any moats. So after at most
kn iterations some moat will grow to include the virtual body of another moat, at which point the algorithm
stops.
Assume now that the invariants are true at some point just before Step (4) is executed and that the
condition in Step (5) is false after Step (4) finishes. We will show that the invariants continue to hold just
before the next iteration starts. We let uv denote the edge that went tight that is considered in Step (4). We
also let t denote the total time the algorithm has executed (i.e. grown moats) up to this point.
0
Before proceeding with our proof, we exhibit the following useful fact. In what follows, let Mjt be the
moat around hj at any time t0 ≤ t during the algorithm. This proposition demonstrates how we control the
overlap of the moats by exploiting the quasi-bipartite structure.
0
Proposition 1. If uv ∈ δ in (Mi ), then uv 6∈ δ in (Mjt ) for any j 6= i, and for any t0 ≤ t.
0
0
Proof. Suppose, for the sake of contradiction, that uv ∈ δ in (Mjt ) for some j 6= i and t0 ≤ t. Since Mjt is
a subset of Mj , the moat containing hj at time t, we must have v ∈ Mj ∩ Mi . Invariant 3 now implies that
v ∈ B̄. Since G is quasi-bipartite, then u ∈ X. Therefore u ∈ Bj 0 for some j 0 . Since j 0 6= i or j 0 6= j, then
the terminating condition in Step (5) would have been satisfied as u ∈ βj 0 . A contradiction.
Following Proposition 1, we let i be the unique index such that uv ∈ δ in (Mi ) as in Step (8).
Invariant 1
First note that Mi never loses vertices during the algorithm’s execution, and it therefore always contains
head vertex hi . Also, vertex u is not part of B0 as otherwise the algorithm would have terminated in Step
(5). Hence Mi ∪ {u} also does not contain the root node r.
Invariant 2
This is just a reinterpretation of Dijkstra’s algorithm in the primal-dual framework (e.g. Chapter 5.4 of [12]),
coupled with the fact that no edge considered in Step (4) in some iteration crosses more than one moat at
any given time (Proposition 1).
Invariant 3
Suppose (Mi ∪ {u}) ∩ Mj 6⊆ B̄ for some i 6= j. This implies u ∈ Mj \ B̄, and hence u ∈ Bj ⊆ βj . Thus,
the termination condition in Step (5) was satisfied and the algorithm should have terminated; contradiction.
If v is not added to βi , and thus βi remains unchanged, βi ∩ βj = Mj ∩ βi = ∅ continues to hold for
j 6= i. We also must have that (Mi ∪ {u}) ∩ βj = ∅ for i 6= j, as otherwise u ∈ βj and this would violate
the termination condition in Step (5).
Now suppose that v is added to βi . Then for j 6= i we still have (βi ∪ {v}) ∩ βj = ∅ as otherwise v ∈ βj
which contradicts v ∈ Mi and the fact that Invariant 3 holds at the start of this iteration. We also have that
Mj ∩ (βi ∪ {v}) = ∅ as otherwise v ∈ Mj . But this would mean that u ∈ Mj as well by Proposition 1. We
established above that (Mi ∪ {u}) ∩ Mj ⊆ B. However, {u, v} ⊆ (Mi ∪ {u}) ∩ Mj ⊆ B contradicts the
fact that G is quasi-bipartite.
Invariant 4
That Bi ⊆ βi is clear simply because we only add nodes to the sets βi . Suppose now that v is added to
βi . In this case, v 6∈ Bi as Bi ⊆ βi from the start. We claim that v can also not be part of Bj for some
j 6= i, since otherwise ∅ 6= Bj ∩ Mi ⊆ βj ∩ Mi , contradicting Invariant 3. Hence v ∈ B̄. Note that the
quasi-bipartiteness of G implies that u ∈ X, and hence u ∈ Bi . Proposition 1 finally implies that only the
moats crossed by uv are moats around i, so since the algorithm only grows one moat around i at any time
we have cuv ≤ t, and this completes the proof of Invariant 4.
8
Invariant 5
The Step (4) stops the first time a constraint becomes tight, so feasibility is maintained. In each step, the
algorithm grows precisely ` moats simultaneously. Because the objective function of (DST-Dual) is simply
the sum of the dual variables, then the value of the dual is just ` times the total time spent growing dual
variables.
3.4
Augmenting T
To complete the final detail in the description of the algorithm, we now show how to construct the partial
Steiner tree after Step (5) has been reached. Lemma 2 shows that Invariants 1 through 5 hold just before
Step (4) in the final iteration. Say the final iteration executes for δ time units and that uv is the edge that
goes tight and was considered in Step (5).
Lemma 3. When Step (6) is reached in Algorithm 1, we can efficiently find a partial Steiner tree T 0 with
0
`0 < ` non-root components such that cost(E(T 0 )) ≤ cost(E(pt)) + 2 `−`
` · OP TLP .
Proof. Let j be the unique index such that u ∈ βj at time ∆. There is exactly one such j because βi ∩βj = ∅
for i 6= j is ensured by the invariants. Next, let J = {i0 6= j : v ∈ Mi0 } and note that J consists of all
indices i0 (except, perhaps, j) such that uv ∈ δ in (Mi0 ). By the termination condition, J 6= ∅. Vertex u lies
in βj by definition. If u 6∈ Bj then we let w be the mate of u as defined in Invariant 4. Otherwise, if w ∈ Bj ,
we let w = u.
For notational convenience, we let Pj be the path consisting of the single edge wu (or just the trivial
path with no edges if w = u). In either case, say cost of Pj is ∆ − j where j ≥ 0 (c.f. Invariant 4). For
each i0 ∈ J, let Pi0 be a shortest v, hi0 -path. Invariant 2 implies that
c(Pi0 ) = ∆ − i0 ,
for some i0 ≥ 0. Observe also that the tightness of uv at time ∆ and the definition of J imply that
X
i0 ≥ cuv .
(3)
(4)
i0 ∈J∪{j}
In fact, precisely a i0 -value of the dual variables for i0 6= j contribute to cuv ; the contribution of j’s variables
to cuv is at most i0 .
Construct a partial Steiner tree T 0 obtained from T and Algorithm 1 as follows.
• The sets Bj 0 , Fj 0 and head hj 0 are unchanged for all j 0 6∈ J ∪ {j}.
S
• Replace the components {Bi0 }i0 ∈J∪{j} with a component B := i0 ∈J∪{j} (Bi0 ∪ V (Pi0 )) having
S
head h := hj . The edges of this component in T 0 are F := i0 ∈J∪{j} (Fi0 ∪ E(Pi0 )) ∪ {uv}.
• The free Steiner nodes B̄ 0 of T 0 are the Steiner nodes not contained in any of these components.
Namely, B̄ 0 consists of those nodes in B̄ that are not contained on any path Pi0 , i0 ∈ J ∪ {j}.
We show that Steiner tree T 0 as constructed above satisfies the conditions stated in Lemma 1. We first
verify that T 0 as constructed above is indeed a valid partial Steiner tree. Clearly the new sets B̄ 0 , {Bi }i6∈J+j
and B partition V and B̄ 0 is a subset of Steiner nodes.
Note that if 0 ∈ J ∪ {j} in the above construction, then j = 0 because no moat contains r. Thus, if B0
is replaced when B is constructed, then r is the head of this new component.
9
Next, consider any b ∈ B. If b ∈ Bj then there is an hj , b-path in Fj ⊆ F . If b ∈ Bi0 , i0 6= j then it can
be reached from hj in (B, F ) as follows. Follow the hj , w-path in Fj , then the w, u path Pj , cross the edge
uv, follow Pi0 to reach hi0 , and finally follow the hi0 , b-path in Fi0 . Finally, if b 6∈ Bi0 for any i0 ∈ J + j
then b lies on some path Pi0 , in which case it can be reached in a similar way.
It is also clear that E(T ) ⊆ E(T 0 ) and that the number of non-root components in T 0 is ` − |J| < `.
Also, cost(E(T 0 )) − cost(E(T )) is at most the cost of the the paths {Pi0 }i0 ∈J+i plus cuv .
It now easily follows from (3) and (4) that
X
cost(E(Pi0 )) + cuv ≤
i0 ∈J∪{j}
X
(∆ − i0 ) + cuv ≤ (|J| + 1)∆ ≤
i0 ∈J∪{j}
|J| + 1
· OP TLP .
`
The last bound follows because the feasible dual we have grown has value ` · ∆ ≤ OP TLP . Let `0 = ` − |J|
be the number of nonroot components in T 0 . Conclude by observing |J| + 1 = ` − `0 + 1 ≤ 2(` − `0 ).
To wrap things up, executing Algorithm 1 and constructing the partial Steiner tree as in Lemma 3 yields
the partial Steiner tree that is promised by Lemma 1.
4
Conclusion
We have shown that the integrality gap of LP relaxation (DST-Primal)
is O(log k) in quasi-bipartite in√
stances of directed Steiner tree. The gap is known to be Ω( k) in 4-layered instances [18] and O(log k) in
3-layered instances [6]. Since quasi-bipartite graphs are a generalization 2-layered instances,
it is natural to
√
ask if there is a generalization of 3-layered instances which has an O(log k) or even o( k) integrality gap.
One possible generalization of 3-layered graphs would be when the subgraph of G induced by the Steiner
nodes does not have a node with both positive indegree and positive outdegree. None of the known results
on directed Steiner tree suggest such instances have a bad gap.
Even when restricted to 3-layered graphs, a straightforward adaptation of our algorithm that grow moats
around the partial Steiner tree heads until some partial Steiner trees absorbs another fails to grow a sufficiently large dual to pay for the augmentation within any reasonable factor. A new idea is needed.
References
[1] J. Byrka, F. Grandoni, T. Rothvoss, and L. Sanita. Steiner tree approximation via iterative randomized
rounding. Journal of the ACM, 60(1):6, 2013.
[2] G. Calinescu and G. Zelikovsky. The polymatroid Steiner problems. J. Combinatorial Optimization,
9(3):281–294, 2005.
[3] M. Charikar, C. Chekuri, T. Cheung, Z. Dai, A. Goel, S. Guha, and M. Li. Approximation algorithms
for directed Steiner problems. J. Algorithms, 33(1):73–91, 1999.
[4] J. Edmonds. Optimum branchings. J. Res. Natl. Bur. Stand., 71:233–240, 1967.
[5] U. Feige. A threshold of ln n for approximating Set-Cover. Journal of the ACM, 45(4):634–652, 1998.
[6] Z. Friggstad, A. Louis, Y. K. Ko, J. Könemann, M. Shadravan, and M. Tulsiani. Linear programming
hierarchies suffice for directed Steiner tree. In proceedings of IPCO, 2014.
[7] S. Guha, A. Moss, J. Naor, and B. Schieber. Efficient recovery from power outage (extended abstract).
In proceedings of STOC, 1999.
10
[8] M. X. Goemans, N. Olver, T. Rothvoss, and R. Zenklusen. Matroids and integrality gaps for hypergraphic Steiner tree relaxations. In proceedings of STOC, 2012.
[9] M. X. Goemans and D. P. Williamson. A general approximation technique for constrained forest
problems. SIAM Journal on Computing, 24(2):296–317, 1995.
[10] E. Halperin and R. Krauthgamer. Polylogarithmic inapproximability. In proceedings of STOC, 2003.
[11] T. Hibi, and T. Fujito. Multi-rooted Greedy Approximation of directed Steiner trees with applications.
In proceedings of WG, 2012.
[12] C. H. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algorithms and Complexity.
Prentice-Hall, Inc., 1982.
[13] S. Rajagopalan, and V. V. Vazirani. On the bidirected cut relaxation for the metric Steiner tree problem.
In proceedings of SODA, 742–751, 1999.
[14] T. Rothvoss. Directed Steiner tree and the Lasserre hierarchy. CoRR abs/1111.5473, 2011.
[15] D. Steurer and I. Dinur. Analytical approach to parallel repetition. CoRR abs/1305.1979, 2013.
[16] V. V. Vazirani. Approximation Algorithms. Springer-Verlag, 2003.
[17] A. Zelikovsky. A series of approximation algorithms for the acyclic directed Steiner tree problem.
Algorithmica, 18:99–110, 1997.
[18] L. Zosin and S. Khuller. On directed Steiner trees. In proceedings of SODA, 2002.
11
| 8cs.DS
|
Transfer Learning for Named-Entity Recognition with Neural Networks
Ji Young Lee∗
MIT
[email protected]
Franck Dernoncourt∗
MIT
[email protected]
arXiv:1705.06273v1 [cs.CL] 17 May 2017
Abstract
Recent approaches based on artificial neural networks (ANNs) have shown promising results for named-entity recognition
(NER). In order to achieve high performances, ANNs need to be trained on a
large labeled dataset. However, labels
might be difficult to obtain for the dataset
on which the user wants to perform NER:
label scarcity is particularly pronounced
for patient note de-identification, which
is an instance of NER. In this work, we
analyze to what extent transfer learning
may address this issue. In particular,
we demonstrate that transferring an ANN
model trained on a large labeled dataset to
another dataset with a limited number of
labels improves upon the state-of-the-art
results on two different datasets for patient
note de-identification.
1
Introduction
Electronic health records (EHRs) have been
widely adopted in some countries such as the
United States and represent gold mines of information for medical research. The majority of EHR
data exist in unstructured form such as patient
notes (Murdoch and Detsky, 2013). Applying natural language processing on patient notes can improve the phenotyping of patients (Ananthakrishnan et al., 2013; Pivovarov and Elhadad, 2015;
Halpern et al., 2016), which has many downstream applications such as the understanding of
diseases (Liao et al., 2015).
However, before patient notes can be shared
with medical investigators, some types of information, referred to as protected health information (PHI), must be removed in order to preserve
∗
These authors contributed equally to this work.
Peter Szolovits
MIT
[email protected]
patient confidentiality. In the United States, the
Health Insurance Portability and Accountability
Act (HIPAA) (Office for Civil Rights, 2002) defines 18 different types of PHI, ranging from patient names and ID numbers to addresses and
phone numbers. The task of removing PHI from a
patient note is referred to as de-identification. The
essence of de-identification is recognizing PHI in
patient notes, which is a form of named-entity
recognition (NER).
Existing de-identification systems are often
rule-based approaches or feature-based machine
learning approaches. However, these techniques
require additional lead time for developing and
fine-tuning the rules or features specific to each
new dataset. Meanwhile, recent work using ANNs
have yielded state-of-the-art performances without using any manual features (Dernoncourt et al.,
2016). Compared to the previous systems, ANNs
have a competitive advantage that the model can
be fine-tuned on a new dataset without the overhead of manual feature development, as long as
some labels for the dataset are available.
However, it may still be inefficient to mass deploy ANN-based de-identification system in practical settings, since creating annotations for patient notes is especially difficult. This is due to
the fact that only a restricted set of individuals
is authorized to access original patient notes; the
annotation task cannot be crowd-sourced, making it slow and expensive to obtain a large annotated corpus. Medical professionals are therefore
wary to explore patient notes because of this deidentification barrier, which considerably hampers
medical research.
In this paper, we analyze to what extent transfer learning may improve de-identification performances on datasets with a limited number of labels. By training an ANN model on a large dataset
(MIMIC) and transferring it to smaller datasets
(i2b2 2014 and i2b2 2016), we demonstrate that
transfer learning allows to outperform the state-ofthe-art results.
2
Related Work
Transfer learning has been studied for a long time.
There is no standard definition of transfer learning
in the literature (Li, 2012). We follow the definition from (Pan and Yang, 2010): transfer learning
aims at performing a task on a target dataset using
some knowledge learned from a source dataset.
The idea has been applied to many fields such as
speech recognition (Wang and Zheng, 2015) and
finance (Stamate et al., 2015).
The successes of ANNs for many applications
over the last few years have escalated the interest
in studying transfer learning for ANNs. In particular, much work has been done for computer
vision (Yosinski et al., 2014; Oquab et al., 2014;
Zeiler and Fergus, 2014). In these studies, some
of the parameters learned on the source dataset are
used to initialize the corresponding parameters of
the ANNs for the target dataset.
Fewer studies have been performed on transfer
learning for ANN-based models in the field of natural language processing. For example, Mou et
al. (2016) focused on transfer learning with convolutional neural networks for sentence classification. To the best of our knowledge, no study has
analyzed transfer learning for ANN-based models
in the context of NER.
3
nating the outputs of the token embedding layer
and the character LSTM layer, and outputs a sequence of vectors.
5. Fully connected layer takes the output of the
token LSTM layer as input, and outputs vectors containing the scores of each label for the
corresponding tokens.
6. Sequence optimization layer takes the sequence of vectors from the output of the fully
connected layer and outputs the most likely sequence of predicted labels, by optimizing the
sum of unigram label scores as well as bigram
label transition scores.
Figure 1 shows how these six components are interconnected to form the model. All layers are
learned jointly using stochastic gradient descent.
For regularization, dropout is applied before the
token LSTM layer, and early stopping is used on
the development set with a patience of 10 epochs.
labels of each token in the sentence
yn-1
yn
Sequence optimization
···
FC
FC
···
Fully connected
FC
FC
Token LSTM
Model
The model we use for transfer learning experiments is based on a type of recurrent neural networks called long short-term memory
(LSTM) (Hochreiter and Schmidhuber, 1997), and
utilizes both token embeddings and character embeddings. It comprises six major components:
yj
y2
y1
···
···
···
···
concatenate
Character LSTM
concatanate
···
Token
embeddings
···
1. Token embedding layer maps each token to a
token embedding.
Character embeddings
2. Character embedding layer maps each character to a character embedding.
3. Character LSTM layer takes as input character embeddings and outputs a single vector that
summarizes the information from the sequence
of characters in the corresponding token.
4. Token LSTM layer takes as input a sequence
of token vectors, which are formed by concate-
xj
jth token in the sentence
c1j
c2j
c(l-1)j clj
characters in the jth token
Figure 1: ANN model for NER. For transfer learning experiments, we train the parameters of the
model on a source dataset, and transfer all or some
of the parameters to initialize the model for training on a target dataset.
4
4.1
Experiments
Datasets
We use three de-identification datasets for the
transfer learning experiments: MIMIC, i2b2 2014,
and i2b2 2016. The MIMIC de-identification
dataset was introduced in (Dernoncourt et al.,
2016), and is a subset of the MIMIC-III
dataset (Johnson et al., 2016; Goldberger et al.,
2000; Saeed et al., 2011). The i2b2 2014 and
2016 datasets were released as part of the 2014
i2b2/UTHealth shared task Track 1 (Stubbs et al.,
2015) and the 2016 i2b2 CEGS N-GRID shared
task, respectively. Table 1 presents the datasets’
sizes.
MIMIC i2b2 2014 i2b2 2016
Vocabulary size
69,525
46,803
61,503
Number of notes
1,635
1,304
1,000
Number of tokens
2,945,228 984,723 2,689,196
Number of PHI instances 60,725
28,867
41,142
Number of PHI tokens
78,633
41,355
54,420
Table 1: Overview of the MIMIC and i2b2
datasets. PHI stands for protected health information.
4.2
Transfer learning
The goal of transfer learning is to leverage the
information present in a source dataset to improve the performance of an algorithm on a target
dataset. In our setting, we apply transfer learning
by training the parameters of the ANN model on
the source dataset (MIMIC), and using the same
ANN to retrain on the target dataset (i2b2 2014
or 2016) for fine-tuning. We use MIMIC as the
source dataset since it is the dataset with the most
labels. We perform two sets of experiments to gain
insights on how effective transfer learning is and
which parameters of the ANN are the most important to transfer.1
Experiment 1 Quantifying the impact of transfer learning for various train set sizes of the target
dataset. The primary purpose of this experiment is
to assess to what extent transfer learning improves
the performances on the target dataset. We experiment with different train set sizes to understand
how many labels are needed for the target dataset
1
Our code is an extension of the NER library
NeuroNER (Dernoncourt et al., 2017), which we committed to NeuroNER’s repository https://github.com/
Franck-Dernoncourt/NeuroNER
to achieve reasonable performances with and without transfer learning.
Experiment 2 Analyzing the importance of
each parameter of the ANN in the transfer learning. Instead of transferring all the parameters,
we experiment with transferring different combinations of parameters. The goal is to understand which components of the ANN are the most
important to transfer. The lowest layers of the
ANN tend to represent task-independent features,
whereas the topmost layers are more task-specific.
As a result, we try transferring the parameters
starting from the bottommost layer up to the topmost layer, adding one layer at a time.
5
Results
Experiment 1 Figure 2 compares the F1-scores
of the ANN trained only on the target dataset
against the ANN trained on the source dataset followed by the target dataset. Transfer learning improves the F1-scores over training only with the
target dataset, though the improvement diminishes
as the number of training samples used for the target dataset increases. This implies that the representations learned from the source dataset are
efficiently transferred and exploited for the target
dataset.
Therefore, when transfer learning is adopted,
fewer annotations are needed to achieve the same
level of performance as when the source dataset is
unused. For example, on the i2b2 2014 dataset,
performing transfer learning and using 16% of the
i2b2 train set leads to similar performance as not
using transfer learning and using 34% of the i2b2
train set. Transfer learning thus allows to cut by
half the number of labels needed on the target
dataset in this case.
For both the i2b2 2014 and 2016 datasets,
the performance gains from transfer learning are
greater when the train set size of the target dataset
is small. The largest improvement can be observed
for i2b2 2014 when using 5% of the dataset as the
train set (consisting of around 2k PHI tokens out of
50k tokens), where transfer learning increases the
F1-score by around 3.1 percent point, from 90.12
to 93.21. Even when all of the train set is used,
the F1-score improves when using transfer learning, albeit by just 0.17 percent point, from 97.80
to 97.97.
(b) i2b2 2016
98
96
96
94
F1-score (%)
F1-score (%)
(a) i2b2 2014
94
92
90
5%
10% 20% 40% 60%
Target train set size (%)
Baseline
Transfer
92
90
88 5%
10% 20% 40% 60%
Target train set size (%)
Figure 2: Impact of transfer learning on the F1-scores. Baseline corresponds to training the ANN model
only with the target dataset, and transfer learning corresponds to training on the source dataset followed
by training on the target dataset. The target train set size is the percentage of train set in the whole
dataset, and 60% corresponds to the full official train set.
(b) i2b2 2016
98
96
96
94
94
92
F1-score (%)
F1-score (%)
(a) i2b2 2014
92
90
88
No
tra
To nsfe
ke r
ne
Ch mb
ar
Ch emb
ar
To LST
Fu ken M
lly LS
co TM
nn
ec
ted
CR
F
No
tra
To nsfe
ke r
ne
Ch mb
ar
Ch emb
ar
To LST
Fu ken M
lly LS
co TM
nn
ec
ted
CR
F
90
5%
10%
20%
40%
60%
Figure 3: Impact of transferring the parameters up to each layer of the ANN model using various train
set sizes on the target dataset: 5%, 10%, 20%, 40%, and 60% (official train set).
Experiment 2 Figure 3 shows the importance of
each layer of the ANN in transfer learning. We
observe that transferring a few lower layers is almost as efficient as transferring all layers. For i2b2
2014, transferring up to the token LSTM shows
great improvements for each layer, but there is less
improvement for each added layer beyond that.
For i2b2 2016, larger improvements can be observed up to the character LSTM and less so beyond that layer.
The parameters in the lower layers therefore
seems to contain most information that are relevant to the de-identification task in general, which
supports the common hypothesis that higher layers
of ANN architectures contain the parameters that
are more specific to the task as well as the dataset
used for training.
Despite the observation that transferring a few
lower layers may be sufficient for efficient transfer learning, it is interesting to see that adding the
topmost layers to the transfer learning does not
hurt the performance. When retraining the model
on the target dataset, the ANN is able to adapt
to the target dataset quite well despite some the
higher layers being initialized to parameters that
are likely to be more specific to the source dataset.
6
Conclusion
In this work, we have studied transfer learning
with ANNs for NER, specifically patient note
de-identification, by transferring ANN parameters
trained on a large labeled dataset to another dataset
with limited human annotations. We demonstrated
that transfer learning improves the performance
over the state-of-the-art results on two datasets.
Transfer learning may be especially beneficial for
a target dataset with small number of labels.
References
Ashwin N Ananthakrishnan, Tianxi Cai, Guergana
Savova, Su-Chun Cheng, Pei Chen, Raul Guzman
Perez, Vivian S Gainer, Shawn N Murphy, Peter
Szolovits, Zongqi Xia, et al. 2013. Improving case
definition of Crohn’s disease and ulcerative colitis
in electronic medical records using natural language
processing: a novel informatics approach. Inflammatory bowel diseases 19(7):1411.
Franck Dernoncourt, Ji Young Lee, and Peter
Szolovits. 2017. NeuroNER: an easy-to-use program for named-entity recognition based on neural
networks. arXiv:1705.05487 .
Lili Mou, Zhao Meng, Rui Yan, Ge Li, Yan Xu,
Lu Zhang, and Zhi Jin. 2016. How transferable
are neural networks in NLP applications? arXiv
preprint arXiv:1603.06111 .
Travis B Murdoch and Allan S Detsky. 2013. The inevitable application of big data to health care. Jama
309(13):1351–1352.
HHS Office for Civil Rights. 2002. Standards for privacy of individually identifiable health information.
final rule. Federal Register 67(157):53181.
Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef
Sivic. 2014. Learning and transferring mid-level image representations using convolutional neural networks. In Proceedings of the IEEE conference
on computer vision and pattern recognition. pages
1717–1724.
Sinno Jialin Pan and Qiang Yang. 2010. A survey on
transfer learning. IEEE Transactions on knowledge
and data engineering 22(10):1345–1359.
Franck Dernoncourt, Ji Young Lee, Ozlem Uzuner, and
Peter Szolovits. 2016. De-identification of patient
notes with recurrent neural networks. Journal of
the American Medical Informatics Association page
ocw156.
Rimma Pivovarov and Noémie Elhadad. 2015. Automated methods for the summarization of electronic
health records. Journal of the American Medical Informatics Association 22(5):938–947.
Ary L Goldberger, Luis AN Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G
Mark, Joseph E Mietus, George B Moody, ChungKang Peng, and H Eugene Stanley. 2000. Physiobank, physiotoolkit, and physionet components of
a new research resource for complex physiologic
signals. Circulation 101(23):e215–e220.
Mohammed Saeed, Mauricio Villarroel, Andrew T
Reisner, Gari Clifford, Li-Wei Lehman, George
Moody, Thomas Heldt, Tin H Kyaw, Benjamin
Moody, and Roger G Mark. 2011. Multiparameter
intelligent monitoring in intensive care II (MIMICII): a public-access intensive care unit database.
Critical care medicine 39(5):952.
Yoni Halpern, Steven Horng, Youngduck Choi, and
David Sontag. 2016. Electronic medical record
phenotyping using the anchor and learn framework.
Journal of the American Medical Informatics Association page ocw011.
Cosmin Stamate, George D Magoulas, and
Michael SC Thomas. 2015.
Transfer learning
approach for financial applications. arXiv preprint
arXiv:1509.02807 .
Sepp Hochreiter and Jürgen Schmidhuber. 1997.
Long short-term memory. Neural computation
9(8):1735–1780.
Alistair E. W. Johnson, Tom J. Pollard, Lu Shen,
Li wei Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G. Mark. 2016. MIMIC-III,
a freely accessible critical care database. Scientific
Data .
Amber Stubbs, Christopher Kotfila, and Özlem Uzuner.
2015. Automated systems for the de-identification
of longitudinal clinical narratives: Overview of 2014
i2b2/UTHealth shared task track 1. Journal of
biomedical informatics 58:S11–S19.
Dong Wang and Thomas Fang Zheng. 2015. Transfer learning for speech and language processing. In
Signal and Information Processing Association Annual Summit and Conference (APSIPA), 2015 AsiaPacific. IEEE, pages 1225–1237.
Qi Li. 2012. Literature survey: domain adaptation algorithms for natural language processing. Department of Computer Science The Graduate Center, The
City University of New York pages 8–10.
Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod
Lipson. 2014. How transferable are features in deep
neural networks? In Advances in neural information
processing systems. pages 3320–3328.
Katherine P Liao, Tianxi Cai, Guergana K Savova,
Shawn N Murphy, Elizabeth W Karlson, Ashwin N
Ananthakrishnan, Vivian S Gainer, Stanley Y Shaw,
Zongqi Xia, Peter Szolovits, et al. 2015. Development of phenotype algorithms using electronic medical records and incorporating natural language processing. bmj 350:h1885.
Matthew D Zeiler and Rob Fergus. 2014. Visualizing
and understanding convolutional networks. In European conference on computer vision. Springer, pages
818–833.
| 2cs.AI
|
arXiv:1604.03640v1 [cs.LG] 13 Apr 2016
CBMM Memo No. 047
April 12, 2016
Bridging the Gaps Between Residual Learning,
Recurrent Neural Networks and Visual Cortex
by
Qianli Liao and Tomaso Poggio
Center for Brains, Minds and Machines, McGovern Institute, MIT
Abstract: We discuss relations between Residual Networks (ResNet), Recurrent Neural Networks (RNNs) and
the primate visual cortex. We begin with the observation that a shallow RNN is exactly equivalent to a very deep
ResNet with weight sharing among the layers. A direct implementation of such a RNN, although having orders
of magnitude fewer parameters, leads to a performance similar to the corresponding ResNet. We propose 1) a
generalization of both RNN and ResNet architectures and 2) the conjecture that a class of moderately deep RNNs
is a biologically-plausible model of the ventral stream in visual cortex. We demonstrate the effectiveness of the
architectures by testing them on the CIFAR-10 dataset.
This work was supported by the Center for Brains, Minds and Machines
(CBMM), funded by NSF STC award CCF - 1231216.
1
1
Introduction
Residual learning [8], a novel deep learning scheme characterized by ultra-deep architectures has recently achieved
state-of-the-art performance on several popular vision benchmarks. The most recent incarnation of this idea [10]
with hundreds of layers demonstrate consistent performance improvement over shallower networks. The 3.57% top-5
error achieved by residual networks on the ImageNet test set arguably rivals human performance.
Because of recent claims [33] that networks of the AlexNet[16] type successfully predict properties of neurons in
visual cortex, one natural question arises: how similar is an ultra-deep residual network to the primate cortex? A
notable difference is the depth. While a residual network has as many as 1202 layers[8], biological systems seem
to have two orders of magnitude less, if we make the customary assumption that a layer in the NN architecture
corresponds to a cortical area. In fact, there are about half a dozen areas in the ventral stream of visual cortex from
the retina to the Inferior Temporal cortex. Notice that it takes in the order of 10ms for neural activity to propagate
from one area to another one (remember that spiking activity of cortical neurons is usually well below 100 Hz).
The evolutionary advantage of having fewer layers is apparent: it supports rapid (100msec from image onset to
meaningful information in IT neural population) visual recognition, which is a key ability of human and non-human
primates [31, 28].
It is intriguingly possible to account for this discrepancy by taking into account recurrent connections within each
visual area. Areas in visual cortex comprise six different layers with lateral and feedback connections [17], which are
believed to mediate some attentional effects [2, 17, 14, 25, 12] and even learning (such as backpropagation [20]).
“Unrolling” in time the recurrent computations carried out by the visual cortex provides an equivalent “ultra-deep”
feedforward network, which might represent a more appropriate comparison with the state-of-the-art computer
vision models.
In addition, we conjecture that the effectiveness of recent “ultra-deep” neural networks primarily come from the
fact they can efficiently model the recurrent computations that are required by the recognition task. We show
compelling evidences for this conjecture by demonstrating that 1. a deep residual network is formally equivalent to
a shallow RNN; 2. such a RNN with weight sharing, thus with orders of magnitude less parameters (depending on
the unrolling depth), can retain most of the performance of the corresponding deep residual network.
Furthermore, we generalize such a RNN into a class of models that are more biologically-plausible models of cortex
and show their effectiveness on CIFAR-10.
2
2.1
Equivalence of ResNet and RNN
Intuition
We discuss here a very simple observation: a Residual Network (ResNet) approximates a specific, standard Recurrent
Neural Network (RNN) implementing the discrete dynamical system described by
ht = K ◦ (ht−1 ) + ht−1
(1)
where ht is the activity of the neural layer at time t and K is a nonlinear operator. Such a dynamical systems
corresponds to the feedback system of Figure 1 (B). Figure 1 (A) shows that unrolling in (discrete) time the feedback
system gives a deep residual network with the same (that is, shared) weights among the layers. The number of
layers in the unrolled network corresponds to the discrete time iterations of the dynamical system. The identity
shortcut mapping that characterizes residual learning appears in the figure.
Thus, ResNets with shared weights can be reformulated into the form of a recurrent system. In section 5.2) we show
experimentally that a ResNet with shared weights retains most of its performance (on CIFAR-10).
A comparison of a plain RNN and ResNet with shared weights is in the Appendix Figure 11.
2
8
h
...
h
+
I
K
h2
Unfold
+
I
I
K
Fold
h1
+
I
h0
K+I
+
K
x
x0
(A) ResNet with shared weights
xt=x0δt
(B) ResNet in recurrent form
Figure 1: A formal equivalence of a ResNet (A) with weight sharing and a RNN (B). I is the identity operator. K is
an operator denoting the nonlinear transformation called f in the main text. xt is the value of the input at time t.
δt is a Kronecker delta function.
2.2
Formulation in terms of Dynamical Systems (and Feedback)
We frame recurrent and residual neural networks in the language of dynamical systems. We consider here dynamical
systems in discrete time, though most of the definitions carry over to continuous time. A neural network (that we
assume for simplicity to have a single layer with n neurons) can be a dynamical system with a dynamics defined as
ht+1 = f (ht ; wt ) + xt
(2)
where ht ∈ Rn is the activity of the n neurons in the layer at time t and f : Rn → Rn is a continuous, bounded
function parametrized by the vector of weights wt . In a typical neural network, f is synthesized by the following
relation between the activity yt of a single neuron and its inputs xt−1 :
yt = σ(hw, xt−1 i + b),
(3)
where σ is a nonlinear function such as the linear rectifier σ(·) = | · |+ .
A standard classification of dynamical systems defines the system as
1. homogeneous if xt = 0, ∀t > 0 (alternatively the equation reads as ht+1 = f (ht ; wt ) with the inital condition
h0 = x0 )
2. time invariant if wt = w.
Residual networks with weight sharing thus correspond to homogeneous, time-invariant systems which in turn
correspond to a feedback system (see Figure 1) with an input which is non-zero only at time t = 0 (xt=0 = x0 , xt =
0 ∀t > 0) and with f (z) = (K + I) ◦ z:
3
hn = f (ht ; wt ) = (K + I)n ◦ x0
(4)
“Normal” residual networks correspond to homogeneous, time-variant systems. An analysis of the corresponding
inhomogeneous, time-invariant system is provided in the Appendix.
3
A Generalized RNN for Multi-stage Fully Recurrent Processing
As shown in the previous section, the recurrent form of a ResNet is actually shallow (if we ignore the possible depth
of the operator K). In this section, we generalize it into a moderately deep RNN that reflects the multi-stage
processing in the primate visual cortex.
3.1
Multi-state Graph
We propose a general formulation that can capture the computations performed by a multi-stage processing hierarchy
with full recurrent connections. Such a hierarchy can be characterized by a directed (cyclic) graph G with vertices
V and edges E.
G = {V, E}
(5)
where vertices V is a set contains all the processing stages (i.e., we also call them states). Take the ventral stream
of visual cortex for example, V = {LGN, V 1, V 2, V 4, IT }. Note that retina is not listed since there is no known
feedback from primate cortex to the retina. The edges E are a set that contains all the connections (i.e., transition
functions) between all vertices/states, e.g., V1-V2, V1-V4, V2-IT, etc. One example of such a graph is in Figure 2
(A).
3.2
Pre-net and Post-net
The multi-state fully recurrent system does not have to receive raw inputs. Rather, a (deep) neural network can
serve as a preprocesser. We call the preprocesser a “pre-net” as shown in Figure 2 (B). On the other hand, one also
needs a “post-net” as a postprocessor and provide supervisory signals to the recurrent system and the pre-net. The
pre-net, recurrent system and post-net are trained in an end-to-end fashion with backpropagation.
For most models in this paper, unless stated otherwise, the pre-net is a simple 3x3 convolutional layer and the
post-net is a pipeline of a batch normalization, a ReLU, a global average pooling and a fully connected layer (or a
1x1 convolution, we use these terms interchangeably).
Take primate visual system for instance, the retina is a part of the “pre-net”. It does not receive any feedback from
the cortex and thus can be separated from the recurrent system for simplicity. In Section 5.3.3, we also tried 3 layers
of 3x3 convolutions as an pre-net, which might be more similar to a retina, and observed slightly better performance.
3.3
Transition Matrix
The set of edges E can be represented as a 2-D matrix where each element (i,j) represents the transition function
from state i to state j.
One can also extend the representation of E to a 3-D matrix, where the third dimension is time and each element
(i,j,t) represents the transition function from state i to state j at time t. In this formulation, the transition functions
4
Pre-net
Post-net
...
Retina
V1
V2
V4
IT
V1
V2
V4
Loss
IT
...
Input
(A) Multi-state (Fully) Recurrent Neural Network
(B) Full model
...
t=T
V1
V2
V4
IT
Pool
Loss
Loss
h
...
Conv
...
K+I
t=2
V1
V2
V4
IT
h
K+I
t=1
V1
V2
V4
IT
h
...
Input
Input
Optional
(C) Simulating our model in time by unrolling
Conv
(D) An example ResNet: for comparison
Figure 2: Modeling the ventral stream of visual cortex using a multi-state fully recurrent neural network
can vary over time (e.g., being blocked from time t1 to time t2 , etc.). The increased expressive power of this
formulation allows us to design a system where multiple locally recurrent systems are connected sequentially: a
downstream recurrent system only receives inputs when its upstream recurrent system finishes, similar to recurrent
convolutional neural networks (e.g., [19]). This system with non-shared weights can also represent exactly the
state-of-the-art ResNet (see Figure 3).
Nevertheless, time-dependent dynamical systems, that is recurrent networks of real neurons and synapses, offer
interesting questions about the flexibility in controlling their time dependent parameters.
Example transition matrices used in this paper are shown in Figure 4.
When there are multiple transition functions to a state, their outputs are summed together.
3.4
Shared vs. Non-shared Weights
Weight sharing is described at the level of an unrolled network. Thus, it is possible to have unshared weights with a
2D transition matrix — even if the transitions are stable over time, their weights could be time-variant.
Given an unrolled network, a weight sharing configuration can be described as a set S, whose element is a set of tied
5
Pool
Conv
Pool
e.g., size: 32x32x16
...
e.g., size: 8x8x64
Loss
h
...
...
e.g., size: 16x16x32
K+I
...
e.g., size: 32x32x16
K+I
Input
h
K3+I
h3
h2
K2+I
Input
(A) ResNet without changing spacial&feature sizes
h
h1
K1+I
h1
Conv
Pre-net
Loss
h3
h2
h
e.g., size: 32x32x16
Conv
Connection available only
at a specified time t
Conv
(B) ResNet with changes of spacial&feature sizes (He et. al.)
Post-net
Pre-net
(C) Recurrent form of A
h1
h2
Subsample
&increase
features
h3
Post-net
Subsample
&increase
features
(D) Recurrent form of B
Figure 3: We show two types of ResNet and corresponding RNNs. (A) and (C): Single state ResNet. The spacial
and featural sizes are fixed. (B) and (D) 3-state ResNet. The spacial size is reduced by a factor of 2 and the featural
size is doubled at time n and 2n, where n is a meta parameter. This type corresponds to the ones proposed by He
et. al. [8].
weights s = {Wi1 ,j1 ,t1 , ..., Wim ,jm ,tm }, where Wim ,jm ,tm denotes the weight of the transition functions from state im
to jm at time tm . This requires: 1. all weights Wim ,jm ,tm ∈ s have the same initial values. 2. the actual gradients
used for updating each element of s is the sum of the gradients of all elements in s:
∀W ∈ s, (
X ∂E
∂E
)used =
(
)original
∂W
∂W 0
0
(6)
W ∈s
where E is the training objective.
For RNNs, weights are usually shared across time, but one could unshare the weights, share across states or perform
more complicated sharing using this framework.
3.5
Notations: Unrolling Depth vs. Readout Time
The meaning of “unrolling depth” may vary in different RNN models since “unrolling” a cyclic graph is not well
defined. In this paper, we adopt a biologically-plausible definition: we simulate the time after the onset of the visual
stimuli assuming each transition function takes constant time 1. We use the term “readout time” to refer to the
time the post-net reads the data from the last state.
This definition in principle allows one to have quantitive comparisons with biological systems. e.g., for a model with
readout time t in this paper, the wall clock time can be estimated to be 20t to 50t ms, considering the latency of a
single layer of biological neurons.
Regarding the initial values, at t = 0 all states are empty except that the first state has some data received from the
6
BN-ReLU-Conv x 3 (BRCx3)
BRCx2+I
BRCx2
BRCx2
BRCx2+I
V1
V2
V4
BRCx2
IT
h1
BRCx2+I
BN-ReLU-Deconv x 3 (BRDx3)
t=1...Inf.
V1
V2
V4
V1
BRCx2+I
BRCx2
BRCx2
BRCx3
V2
BRDx2
BRCx2+I
BRCx2
BRCx2
V4
BRDx2
h1
BRDx2
BRDx2
h1
BRCx2+I
BRDx2
h2
BRCx2
BRCx2+I
BRDx2
BRCx1
h3
h1
Connection available only
at a specified time t
BRCx2+I
h3
t=1...Inf.
h1
h2
h1
BRCx2+I
BRCx2
h2
BRDx2
BRCx2+I
BRCx2
BRCx2
BRDx2 BRCx2+I
t=
1...Inf.
h1
h2
h1
BRCx2+I
Conv
(only at t=n)
h2
h3
(C) an example 2-D
transition matrix
for a 2-state fully
recurrent NN
(B) an example 2-D transition
matrix for a 3-state fully
recurrent NN
BRCx2+I
h2
BRCx2+I
(A) an example 2-D transition matrix
for a 4-state fully recurrent NN for
modeling the visual cortex
BRCx2+I
h2
BRCx2
BRCx2+I
BRCx2
h3
BRDx3
h1
IT
h2
IT
h3
BRCx2
BRCx2+I
t=1...Inf.
BRDx2 BRCx2+I
h2
BRCx2
BRCx1
h1
BRDx2
BRCx2+I
BRCx2+I
h3
t=
1...Inf.
Conv
(only at t=2n)
h1
h1 BRCx2+I
BRCx2+I
(D) 3-D transition matrix of
a 3-state ResNet
(E) transition
matrix of a
1-state ResNet
Figure 4: The transition matrices used in the paper. “BN” denotes Batch Normalization and “Conv” denotes
convolution. Deconvolution layer (denoted by “Deconv”) is [34] used as a transition function from a spacially small
state to a spacially large one. BRCx2/BRDx2 denotes a BN-ReLU-Conv/Deconv-BN-ReLU-Conv/Deconv pipeline
(similar to a residual module [10]). There is always a 2x2 subsampling/upsampling between nearby states (e.g.,
V1/h1: 32x32, V2/h2: 16x16, V4/h3:8x8, IT:4x4). Stride 2 (convolution) or upsampling 2 (deconvolution) is used
in transition functions to match the spacial sizes of input and output states. The intermediate feature sizes of
transition function BRCx2/BRDx2 or BRCx3/BRDx3 are chosen to be the average feature size of input and output
states. “+I” denotes a identity shortcut mapping. The design of transition functions could be an interesting topic
for future research.
pre-net. We only start simulate a transition function when its input state is populated.
3.6
Sequential vs. Static Inputs/Outputs
As an RNN, our model supports sequential data processing and in principle all other tasks supported by traditional
RNNs. See Figure 5 for illustrations. However, if there is a batch normalization in the model, we have to use
“time-specific normalization” described in Section 3.7, which might not be feasible for some tasks.
3.7
Batch Normalizations for RNNs
As an additional observation, we found that it generally hurts performance when the normalization statistics (e.g.,
average, standard deviation, learnable scaling and shifting parameters) in batch normalization are shared across
time. This may be consistent with the observations from [18].
However, good performance is restored if we apply a procedure we call a “time-specific normalization”: mean and
standard deviation are calculated independently for every t (using training set). The learnable scaling and shifting
parameters should be time-specific. But in most models we do not use the learnable parameters of BN since they
tend not to affect the performance much.
We expect this procedure to benefit other RNNs trained with batch normalization. However, to use this procedure,
one needs to have a initial t = 0 and enumerate all possible ts. This is feasible for visual processing but needs
modifications for other tasks.
7
Input
Recurrence
Output
...
t=T
V1
V2
...
V4
YT
IT
...
XT
t=1
V1
V2
V4
Y2
IT
...
X2
V1
V2
V4
IT
...
X1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
One-to-one
...
t=2
V1
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
One-to-many
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
t=T
.
.
.
t=2
t=1
Many-to-one
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
V1
V2
V4
IT
t=T
.
.
.
t=2
t=1
Many-to-many
An example of sequential data processing
Figure 5: Our model supports sequential inputs/outputs, which includes one-to-one, many-to-one, one-to-many and
many-to-many input-output mappings.
4
Related Work
Deep Recurrent Neural Networks: Our final model is deep and similar to a stacked RNN [27, 5, 7] with several
main differences: 1. our model has feedback transitions between hidden layers and self-transition from each hidden
layer to itself. 2. our model has identity shortcut mappings inspired by residual learning. 3. our transition functions
are deep and convolutional.
As suggested by [23], the term depth in RNN could also refer to input-to-hidden, hidden-to-hidden or hidden-to-output
connections. Our model is deep in all of these senses. See Section 3.2.
Recursive Neural Networks and Convolutional Recurrent Neural Networks: When unfolding RNN
into a feedforward network, the weights of many layers are tied. This is reminiscent of Recursive Neural Networks
(Recursive NN), first proposed by [29]. Recursive NN are characterized by applying same operations recursively on
a structure. The convolutional version was first studied by [4]. Subsequent related work includes [24] and [19]. One
characteristic distinguishes our model and residual learning from Recursive NN and convolutional recurrent NN is
whether there are identity shortcut mappings. This discrepancy seems to account for the superior performance of
residual learning and of our model over the latters.
A recent report [3] we became aware of after we finished this work discusses the idea of imitating cortical feedback
by introducing loops into neural networks.
A Highway Network [30] is a feedforward network inspired by Long Short Term Memory [11] featuring more general
shortcut mappings (instead of hardwired identity mappings used by ResNet).
5
5.1
Experiments
Dataset and training details
We test all models on the standard CIFAR-10 [15] dataset. All images are 32x32 pixels with color. Data augmentation
is performed in the same way as [8].
8
Momentum was used with hyperparameter 0.9. Experiments were run for 60 epochs with batchsize 64 unless stated
otherwise. The learning rates are 0.01 for the first 40 epochs, 0.001 for epoch 41 to 50 and 0.0001 for the last 10
epochs. All experiments used the cross-entropy loss function and softmax for classification. Batch Normalization
(BN) [13] is used for all experiments. But the learnable scaling and shifting parameters are not used (except for the
last BN layer in the post-net). Network weights were initialized with the method described in [9]. Although we do
not expect the initialization to matter as long as batch normalization is used. The implementations are based on
MatConvNet[32].
5.2
5.2.1
Experiment A: ResNet with shared weights
Sharing Weights Across Time
We conjecture that the effectiveness of ResNet mainly comes from the fact that it efficiently models the recurrent
computations required by the recognition task. If this is the case, one should be able to reinterpret ResNet as a
RNN with weight sharing and achieve comparable performance to the original version. We demonstrate various
incarnations of this idea and show it is indeed the case.
We tested the 1-state and 3-state ResNets described in Figure 3 and 4. The results are shown in Figure 6.
0.2
0.1
0
0.3
0.2
0.1
0
20
40
Epoch
60
0
0.4
0.3
0.2
0.1
0
20
40
Epoch
60
0
0.6
Non−Shared
#Param 667274
Shared
#Param 76426
0.5
0.4
0.3
0.2
0.1
0
20
40
Epoch
60
0
3-state
ResNet
0.5
0.4
0.3
0.2
0.1
0
20
40
Epoch
60
0
2-state
Fully Recurrent
0.6
Non−Shared
#Param 315338
Shared
#Param 121354
Non−Shared
#Param 2365962
Shared
#Param 297994
0.5
Validation Error on CIFAR−10
0.3
0.4
0.6
Non−Shared
#Param 2365962
Shared
#Param 297994
0.5
1-state
ResNet
Validation Error on CIFAR−10
0.4
2-state
Fully Recurrent
0.6
Non−Shared
#Param 315338
Shared
#Param 121354
0.5
Training Error on CIFAR−10
0.5
Training Error on CIFAR−10
0.6
Non−Shared
#Param 667274
Shared
#Param 76426
Training Error on CIFAR−10
0.6
3-state
ResNet
Validation Error on CIFAR−10
1-state
ResNet
0.4
0.3
0.2
0.1
0
20
40
Epoch
60
0
0
20
40
Epoch
60
Figure 6: All models are robust to sharing weights across time. This supports our conjecture that deep networks
can be well approximated by shallow/moderately-deep RNNs. The transition matrices of all models are shown in
Figure 4. “#Param” denotes the number of parameters. The 1-state ResNet has a single state of size 32x32x64
(height × width × #features). It was trained and tested with readout time t=10. The 3-state ResNet has 3 states
of size 32x32x16, 16x16x32 and 8x8x64 – there is a transition (via a simple convolution) at time 4 and 8 — each
state has a self-transition unrolled 3 times. The 2-state fully recurrent NN has 2 states of the same size: 32x32x64.
It was trained and tested with readout time t=10 (same as 1-state ResNet). It is a generalization of and directly
comparable with 1-state ResNet, showing the benefit of having more states. The 2-state fully recurrent NN with
shared weights and fewer parameters outperforms 1-state and 3-state ResNet with non-shared weights.
5.2.2
Sharing Weights Across All Convolutional Layers (Less Biologically-plausible)
Out of pure engineering interests, one could further push the limit of weight sharing by not only sharing across time
but also across states. Here we show two 3-state ResNets that use a single set of convolutional weights across all
convolutional layers and achieve reasonable performance with very few parameters (Figure 7).
9
CIFAR−10 with Very Few Parameters
Feat. 30
#Param. 9460
Feat. 64
#Param 39754
0.5
Validation Error on CIFAR−10
Training Error on CIFAR−10
0.5
0.4
0.3
0.2
0.1
0
Feat. 30
#Param. 9460
Feat. 64
#Param 39754
0.4
0.3
0.2
0.1
0
50
100
150
Epoch
0
0
50
100
150
Epoch
Figure 7: A single set of convolutional weights is shared across all convolutional layers in a 3-state ResNet. The
transition (at time t= 4 and 8) between nearby states is a 2x2 max pooling with stride 2. This means that each
state has a self-transition unrolled 3 times. “Feat.” denotes the number of feature maps, which is the same across all
3 states. The learning rates were the same as other experiments except that more epochs are used (i.e., 100, 20 and
20).
5.3
5.3.1
Experiment B: Multi-state Fully/Densely Recurrent Neural Networks
Shared vs. Non-shared Weights
Although an RNN is usually implemented with shared weights across time, it is however possible to unshare the
weights and use an independent set of weights at every time t. For practical applications, whenever one can have
a initial t = 0 and enumerate all possible ts, an RNN with non-shared weights should be feasible, similar to the
time-specific batch normalization described in 3.7. The results of 2-state fully recurrent neural networks with shared
and non-shared weights are shown in Figure 6.
5.3.2
The Effect of Readout Time
In visual cortex, useful information increases as time proceeds from the onset of the visual stimuli. This suggests
that recurrent system might have better representational power as more time is allowed. We tried training and
testing the 2-state fully recurrent network with various readout time (i.e., unrolling depth, see Section 3.5) and
observe similar effects. See Figure 8.
5.3.3
Larger Models With More States
We have shown the effectiveness of 2-state fully recurrent network above by comparing it with 1-State ResNet. Now
we discuss several observations regarding 3-state and 4-state networks.
First, 3-state models seem to generally outperform 2-state ones. This is expected since more parameters are
introduced. With a 3-state models with minimum engineering, we were able to get 7.47% validation error on
CIFAR-10.
10
2−State Fully Recurrent NN With Different Readout Time t
t=2
#Param. 76426
t=3
#Param. 224138
t=5
#Param. 297994
t=10
#Param. 297994
0.4
0.5
Validation Error on CIFAR−10
Training Error on CIFAR−10
0.5
0.3
0.2
0.1
0
t=2
#Param. 76426
t=3
#Param. 224138
t=5
#Param. 297994
t=10
#Param. 297994
0.4
0.3
0.2
0.1
0
20
40
0
60
Epoch
0
20
40
60
Epoch
Figure 8: A 2-state fully recurrent network with readout time t=2, 3, 5 or 10 (See Section 3.5 for the definition of
readout time). There is consistent performance improvement as t increases. The number of parameters changes
since at some t, some recurrent connections have not been contributing to the output and thus their number of
parameters are subtracted from the total.
Next, for computational efficiency, we tried only allowing each state to have transitions to adjacent states and to
itself by disabling bypass connections (e.g., V1-V3, V2-IT, etc.). In this case, the number of transitions scales linearly
as the number of states increases, instead of quadratically. This setting performs well with 3-state networks and
slightly less well with 4-state networks (perhaps as a result of small feature/parameter sizes). With only adjacent
connections, the models are no longer fully recurrent.
Finally, for 4-state fully recurrent networks, the models tend to become overly computationally heavy if we train
it with large t or large number of feature maps. With small t and feature maps, we have not achieved better
performance than 3-state networks. Reducing the computational cost of training multi-state densely recurrent
networks would be an important future work.
For experiments in this subsection, we choose a moderately deep pre-net of three 3x3 convolutional layers to model
the layers between retina and V1: Conv-BN-ReLU-Conv-BN-ReLU-Conv. This is not essential but outperforms
shallow pre-net slightly (within 1% validation error).
The results are shown in Figure 9.
5.3.4
Generalization Across Readout Time
As an RNN, our model supports training and testing with different readout time. Based on our theoretical analyses
in Section 2.2, the representation is usually not guaranteed to converge when running a model with time t → ∞.
Nevertheless, the model exhibits good generalization over time. Results are shown in Figure 10. As a minor detail,
the model in this experiment has only adjacent connections and does not have any self-transition, but we do not
expect this to affect the conclusion.
11
3−State Fully/Densely Recurrent Neural Networks
IT
Full
V1
V2
V4
IT
0.3
0.25
0.2
0.15
0.1
Adjacent
0.05
0.3
0.25
0.2
0.15
0.1
0
20
Epoch
40
60
0.05
0.018
0.016
Training Error on CIFAR−10
V4
Full
#Param 351970
Adjacent
#Param 154958
0.35
Validation Error on CIFAR−10
V2
Training Error on CIFAR−10
V1
Full
#Param 351970
Adjacent
#Param 154958
0.35
0.02
0.4
0
20
Epoch
40
Full
#Param 4210186
Adjacent
#Param 3287946
0.088
0.086
Validation Error on CIFAR−10
4−State Fully/Densely Recurrent Neural Networks (Small Model)
0.4
0.014
0.012
0.01
0.008
0.006
0.084
0.082
0.08
0.078
0.076
0.004
0.074
0.002
0.072
60
0
40
50
Epoch
60
Full
#Param 4210186
Adjacent
#Param 3287946
0.07
40
50
Epoch
60
Figure 9: The performance of 4-state and 3-state models. The state sizes of the 4-state model are: 32x32x8, 16x16x16,
8x8x32, 4x4x64. The state sizes of the 3-state model are: 32x32x64, 16x16x128, 8x8x256. 4-state models are small
since they are computationally heavy. The readout time is t=5 for both models. All models are time-invariant
systems (i.e., weights are shared across time).
6
Discussion
The dark secret of Deep Networks: trying to imitate Recurrent Shallow Networks?
A radical conjecture would be: the effectiveness of most of the deep feedforward neural networks, including but not
limited to ResNet, can be attributed to their ability to approximate recurrent computations that are prevalent in
most tasks with larger t than shallow feedforward networks. This may offer a new perspective on the theoretical
pursuit of the long-standing question “why is deep better than shallow” [22, 21].
Equivalence between Recurrent Networks and Turing Machines
Dynamical systems (in particular discrete time systems, that is difference equations) are Turing universal (the game
“Life" is a cellular automata that has been demonstrated to be Turing universal). Thus dynamical systems such as
the feedback systems we discussed can be equivalent to Turing machine. This offers the possibility of representing
a computation more complex than a single (for instance boolean) function with the same number of learnable
parameters. Consider for instance the powercase of learning a mapping F between an input vector x and an output
vector y = F (x) that belong to the same n-dimensional space. The output can be thought as the asymptotic states
of the discrete dynamical system obtained iterating some map f . We expect that in many cases the dynamical
system that asymptotically performs the mapping may have a much simpler structure than the direct mapping F .
In other words, we expect that the mapping f such that f (n) (x) = F (x) for appropriate, possibly very large n can
be much simpler than the mapping F (here f (n) means the n-th iterate of the map f ).
Empirical Finding: Recurrent Network or Residual Networks with weight sharing work well
Our key finding is that multi-state time-invariant recurrent networks seem to perform as well as very deep residual
networks (each state corresponds to a cortical area) without shared weights. On one hand this is surprising because
the number of parameters is much reduced. On the other hand a recurrent network with fixed parameters can be
equivalent to a Turing machine and maximally powerful.
Conjecture about Cortex and Recurrent Computations in Cortical Areas
Most of the models of cortex that led to the Deep Convolutional architectures and followed them – such as the
Neocognitron [6], HMAX [26] and more recent models [1] – have neglected the layering in each cortical area and the
feedforward and recurrent connections within each area and between them. They also neglected the time evolution
of selectivity and invariance in each of the areas. The conjecture we propose in this paper changes this picture quite
drastically and makes several interesting predictions. Each area corresponds to a recurrent network and thus to a
12
The model is trained with readout time 10
0.2
Error On Training Set
Error On Test Set
0.18
0.16
Error on CIFAR−10
0.14
0.12
0.1
0.08
0.06
0.04
Train
0.02
0
6
7
8
9
10
11
12
Test Readout Time t
13
14
15
Figure 10: Training and testing with different readout time. A 3-state recurrent network with adjacent connections
is trained with readout time t=10 and test with t=6,7,8,9,10,11,12 and 15.
system with a temporal dynamics even for flashed inputs; with increasing time one expects asymptotically better
performance; masking with a mask an input image flashed briefly should disrupt recurrent computations in each
area; performance should increase with time even without a mask for briefly flashed images. Finally, we remark that
our proposal, unlike relatively shallow feedforward models, implies that cortex, and in fact its component areas are
computationally as powerful a universal Turing machines.
Acknowledgments
This work was supported by the Center for Brains, Minds and Machines (CBMM), funded by NSF STC award CCF
– 1231216.
References
References
[1] Using goal-driven deep learning models to understand sensory cortex. Nature Neuroscience, 19,3:356–365, 2016.
[2] Christian Büchel and KJ Friston. Modulation of connectivity in visual pathways by attention: cortical
interactions evaluated with structural equation modelling and fmri. Cerebral cortex, 7(8):768–778, 1997.
[3] Isaac Caswell, Chuanqi Shen, and Lisa Wang. Loopy neural nets: Imitating feedback loops in the human brain.
CS231n Report, Stanford, http://cs231n.stanford.edu/reports2016/110_Report.pdf. Google Scholar time
stamp: March 25th, 2016.
[4] David Eigen, Jason Rolfe, Rob Fergus, and Yann LeCun. Understanding deep architectures using a recursive
convolutional network. arXiv preprint arXiv:1312.1847, 2013.
13
[5] Salah El Hihi and Yoshua Bengio. Hierarchical recurrent neural networks for long-term dependencies. Citeseer.
[6] Kunihiko Fukushima. Neocognitron: A self-organizing neural network model for a mechanism of pattern
recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–202, April 1980.
[7] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.
[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv
preprint arXiv:1512.03385, 2015.
[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level
performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer
Vision, pages 1026–1034, 2015.
[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. arXiv
preprint arXiv:1603.05027, 2016.
[11] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780,
1997.
[12] JM Hupe, AC James, BR Payne, SG Lomber, P Girard, and J Bullier. Cortical feedback improves discrimination
between figure and background by v1, v2 and v3 neurons. Nature, 394(6695):784–787, 1998.
[13] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[14] Minami Ito and Charles D Gilbert. Attention modulates contextual influences in the primary visual cortex of
alert monkeys. Neuron, 22(3):593–604, 1999.
[15] Alex Krizhevsky. Learning multiple layers of features from tiny images, 2009.
[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural
networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
[17] Victor AF Lamme, Hans Super, and Henk Spekreijse. Feedforward, horizontal, and feedback processing in the
visual cortex. Current opinion in neurobiology, 8(4):529–535, 1998.
[18] César Laurent, Gabriel Pereyra, Philémon Brakel, Ying Zhang, and Yoshua Bengio. Batch normalized recurrent
neural networks. arXiv preprint arXiv:1510.01378, 2015.
[19] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, pages 3367–3375, 2015.
[20] Qianli Liao, Joel Z Leibo, and Tomaso Poggio. How important is weight symmetry in backpropagation? arXiv
preprint arXiv:1510.05067, 2015.
[21] Hrushikesh Mhaskar, Qianli Liao, and Tomaso Poggio. Learning real and boolean functions: When is deep
better than shallow. arXiv preprint arXiv:1603.00988, 2016.
[22] Guido F Montufar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. On the number of linear regions of
deep neural networks. In Advances in neural information processing systems, pages 2924–2932, 2014.
[23] Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. How to construct deep recurrent
neural networks. arXiv preprint arXiv:1312.6026, 2013.
[24] Pedro HO Pinheiro and Ronan Collobert. Recurrent convolutional neural networks for scene parsing. arXiv
preprint arXiv:1306.2795, 2013.
[25] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of
some extra-classical receptive-field effects. Nature neuroscience, 2(1):79–87, 1999.
[26] M Riesenhuber and T Poggio. Hierarchical models of object recognition in cortex. Nature Neuroscience,
2(11):1019–1025, November 1999.
[27] Jürgen Schmidhuber. Learning complex, extended sequences using the principle of history compression. Neural
Computation, 4(2):234–242, 1992.
14
[28] Thomas Serre, Aude Oliva, and Tomaso Poggio. A feedforward architecture accounts for rapid categorization.
Proceedings of the National Academy of Sciences of the United States of America, 104(15):6424–6429, 2007.
[29] Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language
with recursive neural networks. In Proceedings of the 28th international conference on machine learning
(ICML-11), pages 129–136, 2011.
[30] Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks. arXiv preprint
arXiv:1505.00387, 2015.
[31] Simon Thorpe, Denis Fize, Catherine Marlot, et al. Speed of processing in the human visual system. nature,
381(6582):520–522, 1996.
[32] Andrea Vedaldi and Karel Lenc. Matconvnet: Convolutional neural networks for matlab. In Proceedings of the
23rd Annual ACM Conference on Multimedia Conference, pages 689–692. ACM, 2015.
[33] D.L.K. Yamins and J.D. Dicarlo. Using goal-driven deep learning models to understand sensory cortex, 2016.
[34] Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks. In Computer
Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, pages 2528–2535. IEEE, 2010.
A
An Illustrative Comparison of a Plain RNN and a ResNet
h
Plain
RNN
h0
K
h1
K
X0
X
ResNet
Recurrent
Form
+
Unfold
h
K
h2
...
h2
...
X1
K
K
h0
Unrolled
RNN
h1
+
+
K
ResNet With
Weight Sharing
X0
Figure 11: A ResNet can be reformulated into a recurrent form that is almost identical to a conventional RNN.
B
Inhomogeneous, Time-invariant ResNet
The inhomogeneous, time-invariant version of ResNet is shown in Figure 12.
Let K 0 = K + I, asymptotically we have:
h = Ix + K 0 h
(7)
(I − K 0 )h = x
0 −1
h = (I − K )
x
(8)
(9)
The power series expansion of above equation is:
h = (I − K 0 )−1 x = (I + K 0 + K 0 ◦ K 0 + K 0 ◦ K 0 ◦ K 0 + ...)x
(10)
Inhomogeneous, time-invariant version of ResNet corresponds to the standard ResNet with shared weights and
shortcut connections from input to every layer. If the model has only one state, it is experimentally observed that
these shortcuts undesirably add raw inputs to the final representations and degrade the performance. However, if
15
8
h
...
h
+
I
K'
Unfold
h2
I
+
I
K'
Fold
h1
+
+
I
h0
K'
x0
K'
x
xt=x0
K'=K+I
(A) ResNet with shared weights
and shortcuts from input to all layers
(B) Folded
Figure 12: Inhomogeneous, Time-invariant ResNet
the model has multiple states (like the visual cortex), it might be biologically-plausible for the first state (V1) to
receive constant inputs from the pre-net (retina and LGN). Figure 13 shows the performance of an inhomogeneous
3-state recurrent network in comparison with homogeneous ones.
3−State Fully/Densely Recurrent Neural Networks
0.02
Full
#Param 4210186
Adjacent
#Param 3287946
Adjacent, inhomogeneous
#Param 3287946
V2
V4
IT
V1
V2
V4
IT
t=4
V1
V2
V4
IT
V1
V2
V4
IT
t=3
V1
V1
Input
V2
V2
V4
V4
IT
V1
IT
Homogeneous
V1
Input
V2
V2
V4
V4
IT
IT
t=2
t=1
Inhomogeneous
0.016
Training Error on CIFAR−10
V1
0.014
0.086
0.012
0.01
0.008
0.006
0.084
0.082
0.08
0.078
0.076
0.004
0.074
0.002
0.072
0
40
45
50
Epoch
55
Full
#Param 4210186
Adjacent
#Param 3287946
Adjacent, inhomogeneous
#Param 3287946
0.088
Validation Error on CIFAR−10
0.018
60
0.07
40
45
50
Epoch
55
Figure 13: Inhomogeneous 3-state models. All settings are the same as Figure 9. All models are time-invariant.
16
60
| 9cs.NE
|
EXISTENCE OF SLICES ON A TAME CONTEXT
SOPHIE MARQUES
arXiv:1401.1409v1 [math.AG] 7 Jan 2014
Contents
Keywords
Acknowledgment
Introduction
Hypotheses, basic concepts and notation
1. Definition of slices
2. Slice theorem for actions by finite étale (smooth) group scheme
3. Unramified case
3.1. Definitions
3.2. Global slice theorem
3.3. Local slice theorem
4. Linearly reductive group scheme
4.1. Definition
4.2. Cohomological properties
4.3. Liftings of linearly reductive group schemes
5. Tame quotient stack
5.1. Coarse moduli spaces
5.2. On the exactness of the functor of invariants
5.3. Definition of tame quotient stack
5.4. Local definition of tameness
5.5. Existence of a torsor
5.6. Slice theorem for tame quotient stacks
6. Tame action by affine group scheme vs Tame quotient stack
6.1. Algebraic interlude
6.2. Tame actions
6.3. Exactness of the functor of the invariants for tame actions
6.4. Relationship between the two notions of tameness
References
Abstract. We study the ramification theory for actions involving group schemes, focusing on the tame ramification. We consider the notion of tame quotient stack introduced
in [AOV] and the one of tame action introduced in [CEPT]. We establish a local Slice
theorem for unramified actions and after proving some interesting lifting properties for
linearly reductive group schemes, we establish a Slice theorem for actions by commutative group schemes inducing tame quotient stacks. Roughly speaking, we show that
these actions are induced from an action of an extension of the inertia group on a finitely
presented flat neighborhood. We finally consider the notion of tame action and determine
how this notion is related to the one of tame quotient stack previously considered.
Date: 28 février 2018.
1. SUBMITTED IN CEJM
1
2
2
2
3
6
7
8
8
9
9
10
10
10
11
12
12
13
14
15
17
18
18
18
19
20
22
23
Keywords. Group schemes, Ramification, Unramified, Freeness, Tameness, Slice theorem, Linearly reductive group schemes, Lifting, Actions, Quotient stack, Trivial cohomology.
Acknowledgment. I would like to thanks my supervisors Boas Erez and Marco Garuti but also Angelo
Vistoli and Yuri Tschinkel for their help and advise.
Introduction
It is well-known that locally, for the étale topology, actions of constant group
schemes (and slightly more generally etale group schemes) are induced from actions
of their inertia groups: one can reconstruct the original action from the action of an
inertia group at a point (Theorem 2.1) (this is a direct extension of the classic result
of decomposition of finite extensions of valuation field after moving to the completion
[Ser68, II, §3, Théorème 1, Corollaire 4 and Proposition 4]). This statement is an
instance of a slice theorem (see Theorem 2.2).
In this paper, we establish a more general slice theorem, under a tameness hypothesis. One motivation for studying slices in this generality is the theory of tame covers,
in the sense of Grothendieck and Murre: these admit étales slices (see Abhyankar’s
lemma in [GM71], [CE92]). Another motivation is the fundamental theorem of Luna,
which states that actions of linearly reductive algebraic groups on an affine varieties
are tame (see [Lun73]) and admit slices.
In the simple case when the action has trivial inertia groups our slice theorem
is simply the statement that freeness is local (Proposition 3.3). More general tame
quotient stacks were introduced by Abramovich, Olsson and Vistoli in [AOV08,
§3] in their study of certain ramification issues arising in moduli theory. Here, we
characterize tameness for quotient stacks for actions of finite commutative group
schemes via the existence of finitely presented flat slices with linearly reductive
slice groups (Theorem 5.16). Roughly speaking, we show that these actions are
induced from an action of an extension of the inertia group on a finitely presented
flat neighborhood. Furthermore, this lifting of the inertia group can be constructed
also in the non-commutative case, as a subgroup of the initial group and as a flat
and linearly reductive group scheme (Theorem 4.4). Moreover, [AOV08, Theorem
3.2] shows that tameness is characterized by the property that all inertia groups
at topological points are linearly reductive (it suffices to require this only at the
geometric points).
This is what one could have expected as a definition of tameness. In fact, let us
consider a Γ-extension of Dedekind rings B{BΓ , where Γ is an abstract finite group
and BΓ is the invariant ring for the action of Γ on B. For any prime ideal p of B we
write kppq; then, it is well known, that B{BΓ is tame if and only if for all prime p
of B, the inertia group Γ0 ppq has order prime to the characteristic of kppq. This last
condition is equivalent to requiring that the group algebra kppqrΓ0 ppqs be semisimple,
i.e., that the constant group scheme Γ0 ppq attached to Γ0 ppq is linearly reductive.
2
But as kppq-group, Γ0 ppq is exactly the inertia group of the action of the constant
group scheme attached to Γ on X “ SpecpBq (cf [DG70, III, §2, n˚2, Example 2.4]).
Independently, in [CEPT96, §2], Chinburg, Pappas, Erez and Taylor introduced
the notion of tame actions. Thus it is natural to anticipate a relationship between
these two notions of tameness. We prove that under quite general hypotheses, tame
actions define tame quotient stacks (Theorem 6.8). In fact, under additional hypotheses such as finiteness, the two notions are equivalent (Theorem 6.9). Thus, the
previous results apply to this notion of tameness and we can answer more precisely
[CEPT96, §4, Question 2 and 3].
Hypotheses, basic concepts and notation
We write fppf for faithfully flat and finitely presented. Throughout, we fix the
following notation.
Let R be a commutative, noetherian, unitary base ring: all modules and algebras
are over R and all algebras are commutative. Let S :“ SpecpRq be the corresponding
affine base scheme: all schemes are S-schemes. For V a scheme over S and S1 an Sscheme, let VS1 (resp. S1 V) be the base change V ˆS S1 (resp. S1 ˆS V). In particular,
for V a scheme over S and R1 an R-algebra, we write VR1 (resp. R1 V) instead of
VSpecpR1 q (resp. SpecpR1 q V) for the base change V ˆS SpecpR1 q (resp. SpecpR1 q ˆS V).
Given S-morphisms of schemes V Ñ W and Z Ñ W, we have the fiber product
V ˆW Z
pr2
/
Z
pr1
/
V
W
together with the natural projections.
Let A “ pA, δ, ǫ, sq be a flat, finitely presented commutative Hopf algebra over R
where δ denotes the comultiplicaton, ǫ the unit map and s the antipode, G the affine
flat group scheme associated to A over S and X :“ SpecpBq an affine scheme over S.
Remark. For simplicity, we only consider actions involving affine schemes even
though most of the following results are true for actions involving general schemes by
glueing. Only the definition of tameness by [CEPT96] cannot be always generalized
for actions involving the non-affine schemes (see [CEPT96, §7]).
An action of G on X over S is denoted by pX, Gq; we write µX : X ˆS G Ñ X
for its structure map, (we write x ¨ g :“ µX px, gq, for any x P X and g P G) and
3
ρB : B Ñ B bR A for the structure map giving the A-comodule structure on B (see
[CEPT96, §1.b]).
We will use the sigma notation, well-established in the Hopf algebra literature.
ř
Specifically, for a P A, we write δpaq “ paq a1 b a2 . This presentation itself is
purely symbolic; the terms a1 and a2 do not stand for particular elements of A. The
comultiplication δ takes values in A bR A, and we know that:
δpaq “ pa1,1 b a1,2 q ` pa2,1 b a2,2 q ` pa3,1 b a3,2 q ` . . . ` pan,1 b an,2 q,
for some elements ai, j of A and some integer n. The Sigma notation is a way to
separate the ai,1 from the a j,2 . In other words, the notation a1 stands for the generic
ai,1 and the notation a2 stands for the generic a j,2 . Similarly, for any right comodule
ř
(respectively left comodule) M, any m P M, we write ρM pmq “ mp0q b mp1q (reř
spectively M ρpmq “ mp´1q b mp0q ).
For p P X, we denote IG ppq the inertia group at p to be the fiber product:
IG ppq
pr2
/
Specpkppqq
pr1
∆˝ξ
X ˆS G
pµX ,pr1 q
/
X ˆS X
where kppq is the residue field at p, ξ : Specpkppqq Ñ X is the morphism induced by
the canonical morphism B Ñ kppq and pµX , pr1q : X ˆS G Ñ X ˆS X is the Galois
map sending px, gq to px ¨ g, xq.
Let
C :“ BA “ tb P B|ρB pbq “ b b 1u
be the ring of invariants for this action, Y :“ SpecpCq and π : X Ñ Y the morphism
induced by the inclusion C Ď B.
Remark. We can also then consider X as a scheme over Y. Then the data of the
S-scheme morphism µX : X ˆS G Ñ X defining an action pX, Gq over S is equivalent
to the data of the Y-scheme morphism µX : X ˆY GY Ñ X defining an action pX, GY q
1
over Y. For Y1 a Y-scheme, we denote XYY the base change to Y1 for X considered
as a Y-scheme.
In the following, we will say that a U-scheme Z together with a G-action over S
where U is a S-scheme is a G-torsor over U if Z Ñ U is fppf and the Galois map
pµZ , pr1q : Z ˆS G » Z ˆU Z is an isomorphism. (Here, we choose to work with the
4
fppf topology which is the more reasonable topology to work under our assumptions).
We write rX{Gs for the quotient stack associated to this action. Recall, that it is
defined on the T-points, where T is a S-scheme as:
rX{GspTq “ tP is G-torsor over T together with a G-equivariant morphism P Ñ Xu
We get a canonical S-morphism p : X Ñ rX{Gs by sending a T-point T Ñ X to trivial G-torsor T ˆS G over T together with the G-equivariant morphism T ˆS G Ñ X.
Under the present hypothesis, p is fppf and by the Artin criterium (see [LMB00,
Théorème 10.1]), rX{Gs is an Artin stack. We denote BS G the classifying stack
that is the quotient stack associated to the trivial action of G on S.
Moreover, X{G stands for the categorical quotient in the category of algebraic
spaces, that is X{G is an algebraic space, there is a morphism φ : X Ñ X{G such
that φ ˝ µX “ φ ˝ pr1 and for any algebraic space Z and any morphism ψ : X Ñ Z
such that ψ ˝ µX “ ψ ˝ pr1, ψ factorizes via φ, e.g. there is a morphism X{G Ñ Z
making the following diagram commute:
X
φ
ψ
/
= Z
⑤
⑤⑤
⑤⑤
⑤
⑤⑤
X{G
If G is finite, flat over S then X{G “ Y, by [Con05, Theorem 3.1].
Remark. The quotient stack rX{Gs can always be defined, as soon as G is a flat
group scheme, instead X{G does not exist necessarily. Even if X{G does exist, the
quotient stack rX{Gs gives more information about the action than the categorical
quotient X{G. More precisely, the only difference between the two holds in the existence of automorphisms of points for the stack rX{Gs, but these automorphisms
correspond to the inertia groups for the action pX, Gq. A slice theorem reducing, fppf
locally, the action pX, Gq to an action by a lifting of an inertia group would imply
that the data of the quotient stack associated to X and G is enough to rebuild locally
the action pX, Gq. This is the case for action of finite and étale group scheme (see
Theorem 2.1).
We denote by MA (resp. A M) the category of the right (resp. left) A-comodules
and by (MR (resp. R M) the category of right (resp. left) R-modules). Let pM, ρM q
and pN, ρN q be two right A-comodules. Denote by ComA pM, Nq the R-module of the
A-comodule morphisms from M to N. We recall that a R-linear map g : M Ñ N is
called A-comodule morphism if ρN ˝ g “ pg b IdA q ˝ ρM .
5
We denote by B MA the category of (B,A)-modules. The objects are R-modules
N which are also right A-comodules and left B-modules such that the structural map
of comodules is B-linear, i.e.,
ÿ
ρN pbnq “ ρB pbqρN pnq “
bp0q np0q b bp1q np1q , for all n P N, b P B.
The morphisms of B MA are morphisms that are simultaneously B-linear and Acomodule morphisms. For any M, N P B MA , we write B BimA pM, Nq for the set of such
morphisms from M to N. For a pB, Aq-module M, its submodule of invariants
pMqA is defined by the exact sequence
0 Ñ pMqA Ñ M
ρM /
Mb1
/
MbA
This defines the functor of invariants
p´qA : B MA Ñ C M
which is left exact, by definition. We denote QcohG pXq the G-equivariant quasicoherent sheaves over X (see [AOV08, 2.1]). We have that QcohG pXq » B MA . For
X an Artin stack, we denote by QcohpXq the quasi-coherent sheaves over X.
We have QcohprX{Gsq » QcohG pXq (see [AOV08, 2.1] or [Mar, Chapter IV, §9]).
1. Definition of slices
Here we recall the notion of slices introduced in [CEPT96, Definition 3.1]:
Definition 1.1. We say that the action pX, Gq over S admits étale (respectively
finitely presented flat) slices if:
(1) There is a categorical quotient Q :“ X{G, in the category of algebraic spaces.
(2) For any q P Q, there exist:
(a) a S-scheme Q1 and an étale (resp. finitely presented flat) S-morphism
Q1 Ñ Q such that there is q1 P Q1 which maps to q via this morphism.
(b) a closed subgroup Gq of GQ1 over Q1 which stabilizes some point p1 of
X ˆQ Q1 above q1 , i.e., Gq kpp1 q » IGQ1 pp1 q,
(c) a Q1 -scheme Z with a Gq -action such that Q1 “ Z{Gq and the action
pX ˆQ Q1 , GQ1 q is induced by pZ, Gq q.
The subgroup Gq is called a slice group.
Remark 1.2.
(1) Roughly speaking, for the étale (respectively fppf ) topology, an
action which admits slices can be described by the action of a lifting of an
inertia group of a point.
(2) The action pXQ1 , GQ1 q should be thought of as a G-stable neighborhood of an
orbit. Such a neighborhood induced from an action pZ, Hq, where H is a lifting
of an inertia group of a point, is called a “tubular neighborhood”.
6
2. Slice theorem for actions by finite étale (smooth) group scheme
Since a finite and smooth group scheme is étale, and an étale group scheme is locally constant for the étale topology, it is enough to consider an action by a constant
group scheme.
Let Γ be an abstract finite group. Denote by Γ the constant group scheme associated
to Γ over S. The associated Hopf algebra is then of the form MappΓ, Rq. We recall
first some important facts:
(1) The data of an action of G on X is equivalent to the data of an action of Γ on
B, C “ BΓ “ tγ P Γ|γb “ bu (see [CEPT96, §1.d]) and X{Γ “ Y (see [Con05,
Theorem 3.1]). This particular case also permits to make the transition from
algebra and classic number theory to the algebraic geometry context and to
understand later definitions.
(2) For q P Y, Γ operates transitively on the prime ideals above q (see [Bou81,
Chap. V, §2, Theorem 2]).
(3) The inertia group scheme at a point p P X is the constant group scheme
associated to the algebraic inertia group
Γ0 ppq “ tγ P Γ|γp “ p and γ.x “ x, @x P kppqu
for the action Γ on B induced by pX, Gq at the ideal p (see [DG70, III, §2, n˚
2, Example 2.4]).
In order to alleviate writing, we use that the morphism B Ñ C is integral (see [Bou81,
Chap. V, §1, n˚9, Proposition 22]).Let q be a prime ideal on C, if we write Csh
q for
sh
sh
the strict henselization of C at q, then the base change BbC Cq Ñ Cq is also integral
and since we want a result locally for the étale topology, without loss of generality,
we suppose in the following that the base is local strictly henselian with maximal
ideal q and that B is semi-local over C that is, the product of its local component
pBm q where m are runs though the set of the prime (maximal) ideals of B over the
maximal ideal q of C. Let p be a prime ideal over q. We define MapΓ0 ppq pΓ, Bp q to be
the set of maps
u : Γ Ñ Bp
such that upγiq “ i´1upγq, for any γ P Γ and i P Γ0 pPq; it carries a Γ-action via
λ.upγq “ upλ´1 γq,
for any
λ, γ P Γ.
With these notations, we obtain easily the following Γ-equivariant isomorphism:
pMappΓ, Cq bC Bp qΓ0 ppq » MapΓ0 ppq pΓ, Bp q
With all the previous notations, one obtains that the action of Γ on B can be rebuilt
locally at q thanks to the action of an inertia group at p on Bp :
7
Lemma 2.1. [Ray70, Chapitre X] We have a canonical isomorphism φ, compatible
with the actions of Γ defined by
φ : B Ñ MapΓ0 ppq pΓ, Bp q
b ÞÑ u : γ ÞÑ pγ´1 bqp
Γ ppq
Moreover, C » Bp0 .
Proof. Since Bp » iBp , for any i P Γ0 ppq, φ is well-defined. Let p1 , ....,ps be the
prime ideals over q, we put p1 :“ p. Since the action of Γ is transitive on the
pi ’s, for any i ‰ 1, we have a γi P Γ, such that pi “ γi p. Moreover, Bp j » γ j Bp .
Since we suppose C strictly henselian and B is integral over C, then we know that
B “ ‘si“1 Bpi » ‘si“1 γi Bp . As a consequence, any b P B can be written uniquely as
pγ1 b1 , γ2 b2 , ..., γsbs q in ‘si“1 γi Bp with bi P Bp . Each γi acts on this representation, by
left translation and permuting the term on the direct sum. In particular, we obtain
pγ´1
bqP “ γ´1
γ j b j “ b j , and this proves that φ is an isomorphism. Finally, the
j
j
composite map
B »φ MapΓ0 ppq pΓ, Bp q Ñ Bp
u
ÞÑ up1Γ q
Γ ppq
induces an isomorphism between C and Bp0 .
Rewriting the previous theorem in algebro-geometric terms gives us exactly étales
slices for any action by a finite étale group scheme.
Theorem 2.2. An action of a finite étale group scheme G on X over S admits étale
slices.
Proof. First, Y “ X{G is a categorical quotient in the category of algebraic space (see
[Con05, Theorem 3.1]). The rest of the proof is a direct consequence of the previous
theorem since as we already mentioned we can assume without loss of generality that
G is constant of the form Γ using the previous notation: take q P Y and p P X over
q. Write Chs
q for the strict Henselization of C at q. In Definition 1.1, since X Ñ Y
is finite, we can take Y1 “ SpecpC1 q as an étale subextension of SpecpChs
q q over Y
1
containing q in its image with C a local ring, the slice group Gq to be the constant
group scheme associated to Γ0 ppq over Y1 and Z :“ SpecppB bC C1 qm q, where m is a
prime ideal above the maximal ideal of the local ring C1 corresponding to q.
3. Unramified case
3.1. Definitions.
Definition 3.1. We say that the action pX, Gq is unramified at p P X if the
inertia group IG ppq is trivial. We say that the action pX, Gq is unramified if pX, Gq
is unramified at any p. We say the action is free if the Galois morphism pµX , pr1q :
X ˆS G Ñ X ˆS X is a closed immersion.
8
Lemma 3.2. [DG70, III, §2, n˚ 2, 2.2] The following assertions are equivalent:
(1) pX, Gq is free.
(2) pX, Gq is unramified.
3.2. Global slice theorem. We already have the following theorem.
Theorem 3.3. [DG70, III, §2, n˚2 & n˚3] Suppose that G is finite, flat over S. The
following assertions are equivalent:
(1) pX, Gq is unramified.
(2) X is a G-torsor over Y.
3.3. Local slice theorem.
Lemma 3.4. Suppose that S “ Y is local, with maximal ideal q. The following
assertions are equivalent:
(1) pX, Gq is a free action.
(2) IG ppq is trivial for some p over q.
Proof. The first implication follows from the definition. Let us prove that p2q ñ p1q.
Let kpqq “ R{q be the residue field of R at q and f : X Ñ S be the quotient morphism:
since G is finite, f is finite. Set S0 :“ Specpkpqqq and X0 :“ f ´1 pqq “ tp1 , .., pr u “
X ˆS S0 where for any i P 1, ..., r, pi are the primes above q. Since the inertia is trivial
at some prime over q, it is trivial at any prime over q since all the prime ideals over
q are conjugate up to a finitely presented flat finite base change. Then, up to this
finitely presented flat finite base change, the action pXS0 , GS0 q is free, in other words
X0 ˆS0 GS0 Ñ X0 ˆS0 X0 is a closed immersion (see [DG70, III, §2, n˚2, Proposition
2.2]). That is, we have the surjection
pB bR Bq bR R{q Ñ pB bR Aq bR R{q.
Since BbR B and BbR A are finite R-algebras, by Nakayama’s lemma, BbR B Ñ BbR A
is a surjection, hence the action is free.
From this lemma, we can deduce easily the following theorem which is a local slice
theorem for free action:
Theorem 3.5. Suppose that G is finite and flat over S. Let p P X and q P Y its
image via the morphism π : X Ñ Y. The following assertions are equivalent:
(1) The inertia group scheme is trivial at p, up to a finite base change.
(2) There is an finite, flat morphism Y1 Ñ Y containing q in its image such that
the action pXYY1 , GY1 q over Y1 is free. In other words, pXYY1 ˆY XYY1 , GXY1 q is
induced by the action pXYY1 ˆY1 XYY1 , eXY1 q, where eXY1 denotes the trivial group
Y
scheme over XYY1 .
9
Y
Proof. p1q ñ p2q From the previous lemma, there is an finite, flat morphism Y1 Ñ Y
containing q in its image such that pXYY1 , GY1 q is free, and since GY1 is finite over Y1 ,
XY1 is a GY1 -torsor over Y1 (see Theorem 3.3) hence XYY1 is an fppf morphism over Y1 ,
there is a GY1 -equivariant isomorphism
pµXY1 , pr1q : pXYY1 ˆY1 GY1 q{eXY1 » XYY1 ˆY1 XYY1
Y
Y
where on the left hand side GY1 acts only on the factor GY1 by translation and on
the right hand side GY1 acts on the first factor only.
p2q ñ p1q Trivial.
4. Linearly reductive group scheme
4.1. Definition. From the notion of tameness, we can define a class of group
schemes with nice properties, which will be very important for the following.
Definition 4.1. We say that G is linearly reductive group scheme, if p´qA :
A
R M Ñ R M is exact.
4.2. Cohomological properties. The cohomology of linearly reductive group schemes
has some very interesting cohomological vanishing properties that are useful for deformation involving such group schemes. We can even characterize linearly reductivity by the vanishing of the cohomology:
Lemma 4.2. (see [Kem00, Proposition 1]) Suppose that G is flat, finite over S,
that X finitely presented over S and S “ Specpkq where k is a field. The following
assertion are equivalent:
(1) G is linearly reductive over S.
(2) H1 pG, Mq “ 0 for any M P MA .
The following lemma is part of the proof of [AOV08, Lemma 2.13]:
Lemma 4.3. If G is linearly reductive over S, then
(1)
Exti pLBk G{k , F q “ 0, f or i ‰ ´1, 0.
for any coherent sheaf F on Bk G.
(2) If we suppose that G is also smooth,
Exti pLBk G{k , F q “ 0, f or i ‰ ´1
for any coherent sheaf F on Bk G.
Proof. By [AOV08, Lemma 2.15], the cotangent complex LBk G{k of the structural
r´1,0s
morphism Bk G Ñ k (where k is a field) belongs to Dcoh pOBk G q. Since k is a field,
any coherent sheaf on Bk G is locally free, and therefore for any coherent sheaf F on
Bk G, we have:
r´1,0s
RHompLBk G{k , F q P Dcoh pOBk G q.
10
(1) Since the global section functor is exact on the category CohpOBk G q (since G
is linearly reductive), we obtain that then
Exti pLBk G{k , F q “ 0, f or i ‰ ´1, 0.
for any F is a coherent sheaf on Bk G.
r0s
(2) When G is smooth, then LBk G{k P Dcoh pOBk G q and
Exti pLBk G{k , F q “ 0, f or i ‰ ´1
for any F is a coherent sheaf on Bk G.
4.3. Liftings of linearly reductive group schemes. By [AOV08, Proposition
2.18], we know that a linearly reductive group scheme over some point can be lifted
as linearly reductive group to an étale neighborhood. Thanks to this result, we were
able to prove that we can lift this group fppf locally as a subgroup in the following
sense.
Theorem 4.4. Let p be a point of S, G be a finite, flat group scheme over S and H0
be a finite, linearly reductive closed subgroup scheme of Gkppq over Specpkppqq. Then,
there exists a flat, finitely presented morphism U Ñ S, a point q P U mapping to p
and a flat linearly reductive closed subgroup scheme H of GU over U whose pullback
Hkpqq is isomorphic to the pullback of H0 kpqq .
Proof. Let p P S. By [AOV08, Proposition 2.18], there exists an étale morphism
U Ñ S, a point q P U mapping to p and a linearly reductive group scheme H over
U such that Hkpqq » H0 kpqq . Set Un :“ SpecpR{qn`1 q. One has that H0 is a subgroup
scheme of Gkppq , and this defines a representable morphism of algebraic stacks x0 :
Bkpqq H0 Ñ Bkpqq Gkpqq over kpqq. We want to prove the existence a representable
morphism of algebraic stacks x : BU H Ñ BU GU filling in the following 2-commutative
diagram
/
Bkpqq H0
✷✷ ❑❑❑
✷✷ ❑❑x❑
❑❑❑
✷✷
❑%
✷✷
✷✷
Bkpqq Gkpqq
✷✷
✷✷
✷✷
✷✷
g
✷✷
✷✷
✷✷
U0
BU H
✴✴
x1
✴✴
✴✴
#
✴✴
✴✴ / BU GU
✴✴
✴✴
✴✴
✴✴
✴✴
✴✴
✴
/ U
We can prove, thanks to Grothendieck ’s Existence theorem for algebraic stacks (see
[Ols07, Theorem 11.1]) and Artin’s approximation theorem (see [Art69, Théorème
11
1.12]), that the existence of such x only depends on the existence of a formal deformation, that is morphisms xn : BUn HUn Ñ BUn GUn filling in the following 2-commutative
diagram:
Bkpqq H0
in
✸✸ ▲▲▲ x
✸✸ ▲▲▲1
▲▲▲
✸✸
▲&
✸✸
✸✸
B
kpqq Gkpqq
✸✸
✸✸
✸✸
✸✸
✸✸
✸✸
✸✸
Specpkpqqq
/
/
BUn´1 HUn´1
jn
✻✻ ◆◆◆ xn´1
◆◆◆
✻✻
◆◆◆
✻✻
◆◆'
✻✻
/ B
✻✻
Un´1 GUn´1
✻✻
✻✻
✻✻
✻✻
✻✻
✻✻
✻✻
✻
/ U
n´1
BU n H U n
✶✶
xn
✶✶
✶✶
%
✶✶
/ B G
✶✶
Un Un
✶✶
✶✶
✶✶
✶✶
✶✶
✶✶
✶✶
/ U
n
By [Ols06, Theorem 1.5], the obstruction to extend the morphism xn´1 to xn lies in
Ext1 pLx˚ LBkpqq Gkpqq {kpqq , mn {mn`1 bkpqq OBkpqq Gkpqq q,
which is trivial for any n P N by Lemma 4.3. It follows that there exists an arrow xn
filling the previous diagram. This leads to the existence of a representable morphism
of stacks F : BU HU Ñ BU GU . Let Q Ñ U be the GU -torsor which is the image via F of
the trivial HU -torsor HU Ñ U. Furthermore, the functor F induces a homomorphism
from AutS pHQ Ñ Qq “ HQ to the automorphism group scheme of the image of the
HQ -torsor HQ Ñ Q in BQ GQ . Since this image is the pullback of Q to Q over BQ GQ ,
which is canonically a trivial torsor, its automorphism group is GQ . This defines a
group morphism HQ Ñ GQ , with Q Ñ U an fppf morphism as GU -torsor. Since
F is representable, this morphism is injective. Finally, since H is proper and G is
separated, HV is closed in GV .
5. Tame quotient stack
5.1. Coarse moduli spaces. Abramovich, Olsson and Vistoli introduced in [AOV08]
the notion of tame stack. Before recalling their definition, we need some additional
terminology.
Definition 5.1. A coarse moduli space for the quotient stack rX{Gs is a
couple pM, φq where M is a algebraic space and φ : rX{Gs Ñ M a morphism of stacks
such that any morphism from rX{Gs to an algebraic space factor through φ and such
that for any algebraically closed field Ω, |rX{GspΩq| » MpΩq where |rX{GspΩq| is
the set of isomorphism classes of the geometric points of rX{Gs taking value on Ω.
Remark 5.2.
(1) The notion of coarse moduli spaces can be generalized for general Artin stack. Coarse moduli space for quotient stack are in particular a
categorical quotient in the category of algebraic spaces for the action pX, Gq.
More precisely,
12
(a) Let M be an algebraic space. The datum of a morphism φ : rX{Gs Ñ M
is equivalent to the datum of a morphism of S-algebraic spaces f : X Ñ M
such that f ˝ pr1 “ f ˝ µX . In particular, the canonical map π : X Ñ Y
induces a map ρ : rX{Gs Ñ Y.
(b) If f : X Ñ M is a geometric quotient which is categorical in the category
of algebraic spaces, then rX{Gs Ñ M is a coarse moduli space.
For more details, see [Mar, 10.4].
(2) If G is finite, flat over S then ρ : rX{Gs Ñ Y is a coarse moduli space for
rX{Gs (see [Con05, §3]).
Theorem 5.3. Suppose that the group scheme G and the scheme X are finitely
presented over S and that all the inertia group schemes are finite. The quotient
stack rX{Gs admits a coarse moduli space φ : rX{Gs Ñ M such that φ is proper.
Moreover, φ induces a functor φ˚ : QcohprX{Gsq Ñ QcohpMq.
Proof. Since G is finitely presented over S, so is X Ñ rX{Gs (X is a G-torsor over
rX{Gs). Since X Ñ rX{Gs is surjective, flat of finite presentation and X Ñ S is also
of finite presentation, rX{Gs Ñ S is of finite presentation. By [Con05, Theorem 1.1],
the hypotheses insure that the quotient stack admits a coarse moduli space that we
denote φ : rX{Gs Ñ M such that φ is proper. In particular, φ is quasi-compact
and quasi-separated, thus by [Ols07, Lem. 6.5(i)], the induced morphism on the
quasi-coherent sheaves φ˚ : QcohprX{Gsq Ñ QcohpMq is well defined.
5.2. On the exactness of the functor of invariants.
Lemma 5.4. The map ρ : rX{Gs Ñ Y (defined in Remark 5.2) induces a functor
ρ˚ : QcohprX{Gsq Ñ QcohpYq. Then, the functor ρ˚ is exact if and only if the
functor of invariants p´qA : B MA Ñ C M is exact.
Proof. By [Mar, Lemma IV.10], we know that the map ρ is quasi-compact and quasiseparated, thus by [Ols07, Lem. 6.5(i)], the induced morphism on the quasi-coherent
sheaves ρ˚ : QcohprX{Gsq Ñ QcohpMq is well defined. ρ is defined as making the
diagram bellow commute:
ρ
rX{Gs
b❊❊
❊❊
❊
p ❊❊❊
/
@
Y
π
X
Thus, it induces the following commutating diagram of functors:
QcohprX{Gsq
ρ˚
/
QcohpYq
q8
qqq
q
q
q
qqq π˚
g❖❖❖
❖❖❖
❖❖
p˚ ❖❖❖
QcohG pXq
13
The functor p˚ is an equivalence of categories (see [Mar, Proposition IV.24]). Moreover, by [Har77, §5], we have the following commutative diagram:
QcohG pXq
π˚
/
QcohpYq
ΓpX, q
ΓpY, q
A
BM
p qA
/
CM
where the functors of global section ΓpX, q and ΓpY, q are equivalences of categories.
This proves the lemma.
Theorem 5.5. Suppose that X is noetherian, finitely presented over S, that G is
finitely presented over S and that all the inertia groups are finite. If the functor of
invariants p´qA is exact, then the map ρ : rX{Gs Ñ Y is a coarse moduli space of
rX{Gs and ρ is proper.
Proof. , By [Alp08, Theorem 6.6], the quotient π : X Ñ Y is categorical in the
category of algebraic spaces. [Con05, Theorem 1.1] insures that the quotient stack
admits a coarse moduli space that we denote ρ : rX{Gs Ñ M such that ρ is proper.
But, since ρ ˝ p : X Ñ rX{Gs Ñ M and π : X Ñ Y are categorical quotients
in the category of algebraic spaces, we obtain by unicity that M » Y and that
ρ : rX{Gs Ñ Y is a coarse moduli space.
5.3. Definition of tame quotient stack. Finally, we can define a tame quotient
stack. Since we need the existence of a coarse moduli space. Until the end of section
5, we suppose that G is flat, finitely presented over S, that X is finitely presented
over S and that all the inertia groups at the geometric points for the action pX, Gq
are finite. Moreover, we denote M the coarse moduli space and φ : rX{Gs Ñ M the
proper map (see Theorem 5.3).
Definition 5.6. We say that the quotient stack rX{Gs is tame if the functor φ˚ :
QcohrX{Gs Ñ QcohpMq is exact.
Remark 5.7.
(1) The notion of tameness can be defined similarly for general
stacks (see [AOV08, Definition 3.1]).
(2) When G is finite flat over S, rX{Gs is tame if and only if the functor of
invariants p´qA is exact. (It is a consequence of Lemma 5.4 since Y is a
coarse moduli space for G finite flat over S, by [Con05, §3]).
(3) When G is finite flat over S, G is linearly reductive if and only if the classifying stack BS G is tame.
Tameness is local. In fact,
14
Lemma 5.8. [Alp08, Proposition 3.10] For any morphism of S-schemes g : M1 Ñ
M, we consider the following 2-cartesian diagram:
g1
rX{GsM1
/
rX{Gs
φ1
φ
M1
g
/
M
Suppose that φ (resp. φ1 ) is the coarse moduli space for rX{Gs (resp. rX{GsM1 ).
(1) If g is faithfully flat and the quotient stack rX{GsM1 is tame then the quotient
stack rX{Gs is tame.
(2) If the quotient stack rX{Gs is tame then the quotient stack rX{GsM1 is also
tame.
Proof.
(1) From the 2-cartesian diagram we deduce that the functors g˚ φ˚ and
φ1˚ g1˚ are isomorphic. Since g is flat, g1 is flat as well and g1˚ is exact; also
φ1˚ is exact by assumption, so the composite φ1˚ g1˚ is exact, hence so is g˚ φ˚ .
But, since g is faithfully flat, we have that φ˚ is also exact as required.
(2) First suppose that g is an open, quasi-compact immersion. Let
0 Ñ F11 Ñ F21 Ñ F31 Ñ 0
be an exact sequence of OrX{GsM1 -modules. Set F3 :“ g1˚ F2 {g1˚ F1 . Then
0 Ñ g1˚ F11 Ñ g1˚ F21 Ñ F3 Ñ 0
is exact. Moreover, g1˚ F3 » F31 since the adjunction morphism g1˚ g1˚ Ñ id is
an isomorphism. Since φ˚ is exact by assumption, φ˚ g1˚ F21 Ñ φ˚ F3 is also
surjective, but then g˚ φ1˚ F21 Ñ φ˚ F3 is surjective as well. Since g is an open
immersion, φ1˚ F21 Ñ g˚ φ˚ F3 is surjective. Finally, since g˚ φ˚ and φ1˚ g1˚ are
isomorphic functors, φ1˚ F21 Ñ φ1˚ F31 is surjective.
We consider now any morphism of schemes g : M1 Ñ M. Since the tameness
property on a stack is Zariski local, we can assume M1 and M affine. Then g1
is also affine, so the functor g1˚ is exact. By assumption φ˚ is exact, therefore
φ˚ g1˚ “ g˚ φ1˚ is exact. But the functor g˚ has the property that a sequence
F1 Ñ F2 Ñ F3 is exact if and only if g˚ F1 Ñ g˚ F2 Ñ g˚ F3 is exact. It
follows that φ1˚ is exact as required.
5.4. Local definition of tameness.
Theorem 5.9. [AOV08, Theorem 3.2] The following assertions are equivalent:
(1) The quotient stack rX{Gs is tame.
(2) The inertia groups IG pξq Ñ Specpkq are linearly reductive groups, for any
ξ : Specpkq Ñ X, where k is a field.
15
(3) The inertia groups IG pξq Ñ Specpkq are linearly reductive groups, for any
geometric point ξ : Specpkq Ñ X, where k is an algebraically closed field.
(4) The inertia groups IG ppq Ñ Specpkppqq are linearly reductive groups, for any
p P X.
(5) For any point p P X, denote q its image via the morphism X Ñ M there
exist an fppf (also can be chosen étale surjective) morphism M1 Ñ M containing q in its image, a linearly reductive group scheme H Ñ M1 , such that
Hkpqq » IG ppq, acting on a finite and finitely presented scheme P Ñ M1 and
an isomorphism of algebraic stacks over M1
rX{Gs ˆM M1 » rP{Hs.
Proof. p1q ñ p2q Let ξ : Specpkq Ñ X be a k-point, where k is a field, and IG pξq is the
inertia group in ξ. The quotient stack rGk {IG pξqs is a scheme, denote it Gk {IG pξq.
Since the square
Gk {IG pξq
/
Bk IG pξq
/
X ˆS Specpkq
rX{Gs ˆS Specpkq
is 2-cartesian, Bk IG pξq Ñ rX{Gs ˆS Specpkq is affine since Gk {IG pξq Ñ X ˆS Specpkq
is affine.
Now, let us consider the following commutative diagram:
Bk IG pξq
g
/
rX{Gs p‹q
p´qIG pξq
φ
Specpkpqqq
f
/
M
Since we have seen that g is affine, g˚ : QcohpBkIG pξqq Ñ QcohprX{Gsq is an exact
functor and φ˚ : QcohprX{Gsq Ñ QcohpYq is exact by definition of tameness. Since
f˚ φ1˚ “ φ˚ g˚ , if
0 Ñ V1 Ñ V2 Ñ V3 Ñ 0
is an exact sequence of G-representations, considered as exact sequence of quasi
coherent sheaves over Bk IG pξq, we have the following exact sequence:
0 Ñ f˚ pV1 qIG pξq Ñ f˚ pV2 qIG pξq Ñ f˚pV3 qIG pξq Ñ 0.
Moreover, p´qIG pξq is left exact and this implies that
0 Ñ pV1 qIG pξq Ñ pV2 qIG pξq Ñ pV3 qIG pξq Ñ 0
is exact. So, IG pξq is linearly reductive.
p2q ñ p3q Immediate.
p3q ñ p4q By [AOV08, Theorem 2.16], it is enough to prove that IG ppqΩ is trivial,
16
for any Ω algebraically closed field. Denoting by pΩ the composite SpecpΩq Ñ
Specpkppqq Ñ X. But, IG ppqΩ is a subgroup of IG ppΩ q which is trivial by assumption.
Thus, IG ppqΩ is trivial.
p4q ñ p5q See [AOV08, Theorem 3.2]
p5q ñ p1q See Lemma 5.8.
We obtain the following corollary.
Corollary 5.10. [AOV08, Corollary 3.5] The stack rX{Gs Ñ S is tame if and only if
for any morphism Specpkq Ñ S, where k is an algebraically closed field, the geometric
fiber rX{Gs ˆS Specpkq Ñ Specpkq is tame.
Definition 5.11. We say that rX{Gs is tame at p if the inertia group IG ppq is
linearly reductive.
Remark 5.12. By the previous theorem, rX{Gs is tame if and only if rX{Gs is tame
at p, for any p P X.
5.5. Existence of a torsor. We state now an interesting consequence of the previous theorem which permits to define a torsor for a tame quotient stack for actions
of finite commutative group schemes.
Proposition 5.13. Suppose that G is finite, commutative flat over S. If the quotient
stack rX{Gs is tame, then for any point p P X, denoting by q P Y its image by the
morphism X Ñ Y, there exist a finitely presented flat morphism Y1 Ñ Y containing
q in its image and a subgroup H of GY1 over Y1 lifting the inertia group at p such
that XYY1 {H is a GY1 {H-torsor over Y1 .
Proof. Notice first that under the assumption of the theorem, GY1 {H is a finite affine
group scheme. By the previous theorem, we know that the inertia at p is linearly
reductive. So, by Lemma 4.4, there are a finitely presented flat morphism Y1 Ñ Y
containing q in its image and a linearly reductive group H Ñ Y1 lifting this inertia
group as a subgroup of GY1 . Moreover, the inertia group at p, image of p by the
quotient morphism XYY1 Ñ XYY1 {H for the action pXYY1 {H, GY1 {Hq, is equal to
IGY1 {H ppq “ IG ppq{pIG ppq X Hkppq q “ teu
By Theorem 3.3, up to passing to a finitely presented flat neighborhood of q, the
action pXYY1 {H, GY1 {Hq is free, thus XYY1 {H is a GY1 {H-torsor over Y1 since GY1 {H is
finite (see Theorem 3.3).
Remark 5.14. It would be useful to establish the previous proposition in general
if G is not commutative. But, in this case G{H is not necessarily a group scheme
and we cannot even define the notion of action or of torsor. But, taking the normal
closure of H in GY1 instead of H, we can establish the same result.
17
5.6. Slice theorem for tame quotient stacks. We manage to prove a slice theorem for actions by a finite commutative group scheme, using the following lemma.
Lemma 5.15. [CEPT96, Proposition 6.5] Let H be a subgroup of G such that the
quotient for the natural translation action exists and is universal. Let ψ : X Ñ G{H
be a morphism of schemes preserving the G-actions. Let
Z “ X ˆG{H e
be defined as the fibered product of the two maps ψ and the inclusion e “ H{H Ñ
G{H. Assume that the balanced product Z ˆH G exists and is a universal quotient.
Then we have an isomorphism of G-actions
pX, Gq » pZ ˆH G, Gq.
Finally, we get the following slice theorem which extends [CEPT96, Theorem 6.4]
as we will see above:
Theorem 5.16. Suppose that G is commutative and finite over S. The quotient
stack rX{Gs is tame if and only if the action pX, Gq admits finitely presented flat
slices such that the slice group at p P X are linearly reductive.
Proof. When G is commutative, since by proposition 5.13, up to passing to a finite
finitely presented flat neighborhood of q, there is a subgroup H of G such that
pX{H, G{Hq defines a torsor over Y, then X{H ˆY X{H » X{H ˆY G{H. This gives
us a G-equivariant morphism ψ : X Ñ G{H after the fppf base change X{H Ñ Y. So,
pX, Gq is induced by the action pZ, Hq using the notation of the previous lemma. The
converse is due to Theorem 5.9 since if the action pX, Gq admits finitely presented
flat slices such that the slice group at p P X are linearly reductive, then for any
p P X, IG ppq is linearly reductive.
6. Tame action by affine group scheme vs Tame quotient stack
6.1. Algebraic interlude. The following lemma is quite important in the sense
that it relates the functor of invariants to the functor ComA p´, Bq. As a consequence,
it relates in particular the exactness of this functors which will be important in order
to compare the two notions of tameness (see Lemma 6.7).
Lemma 6.1. Suppose that the Hopf algebra A is flat over R. Let B P MA and
M P R M be finitely presented as R-modules. There is a natural (functorial on M)
isomorphism:
pB bR M˚ qA » ComA pM, Bq
18
Proof. The commutativity of the right square of the diagram below insures the
existence of λ which is an isomorphism and so α and γ as well.
/
0
pB bR M˚ qA
/
B bR M˚
0
/
ComA pM, Bq
/
pB bR M˚ q bR A
γ
α
λ
φ
/
HomR pM, Bq
ψ
/
HomR pM, B bR Aq
where φ :“ ρBbR M˚ ´ IdB b IdM˚ b 1, ψ :“ pρB ˝ ´q ´ ρM˚ and for any b b f b a P
ř
B bR M˚ bR A and m P M, γpb b f b aqpmq “ b. f pmp0q q b a.mp´1q
Indeed, the map γ is an isomorphism as composite of the isomorphism γ1 : B bR
M˚ bR A Ñ B bR A bR M˚ defined for any b b f b a P B bR M˚ bR A by γ1 pb b f b aq “
b b a.mp´1q b f pmp0q q (Its inverse map is defined for any b b a b f P B bR A bR M˚
ř
by γ´1
pb b a b f q “ b b f pmp0q q b a.Spmp´1q q) with the canonical isomorphism
1
γ2 : B bR A bR M˚ Ñ HomR pM, B bR Aq. For any b b f P B bR M˚ , we have
ř
ψ ˝ αpb b f q “ ψprm ÞÑ b f pmqsq “ rm ÞÑ ρB pbq f pmq ´ b f pmp0q q b mp´1q s
and
γ ˝ φpb b f q “
“
“
“
ř
γprm ÞÑ bp0q b f pmp0q q b bp1q Spmp´1q q ´ b b f pmq b 1sq
ř
rm ÞÑ pbp0q b bp1q Spmp1 q qpmp0q qp´1q q f ppmp0q qp0q ´ b f pmp0q b mp´1q s
ř
ř
rm ÞÑ pbp0q b bp1q ǫpmp´1q qq f pmp0q q ´ b f pmp0q q b mp´1q s
ř
rm ÞÑ ρB pbq f pmq ´ b f pmp0q q b mp´1q s.
6.2. Tame actions. Chinburg, Erez, Pappas and Taylor defined in their article
[CEPT96] the notion of tame action. We recall here this definition and some useful
properties.
Definition 6.2. We say that an action pX, Gq is tame if there is a unitary (that is
αp1A q “ 1B ) morphism of A-comodules α : A Ñ B, which means that α is a R-linear
map such that ρB ˝ α “ pα b IdA q ˝ δ, such a morphism is called a total integral.
Tame actions are stable under base change.
Lemma 6.3. If the action pX, Gq is tame, then after an affine base change R1 Ñ R
the action pXR1 , GR1 q is also tame.
Proof. Since the action pX, Gq is tame, there is a AC -comodule map A Ñ B, which
induces naturally a comodule map AR1 Ñ BR1 .
The next lemma will allow us to assume that the base is equal to the quotient,
if the structural map X Ñ S has the same properties as the quotient morphism
X Ñ Y.
Lemma 6.4. The following assertions are equivalent:
19
(1) The action pX, CGq over C is tame .
(2) The action pX, Gq over S is tame.
(3) The action pC X, C Gq over C is tame.
Proof. p1q ñ p2q Let α : C A Ñ B be a total integral for the tame action pX, CGq.
Since B is a C A-comodule via ρB : B Ñ B bR A » B bC C A, the composite α1 :“
α ˝ iA : A Ñ B, where iA : A Ñ C bR A maps a to 1 b a, is a unitary A-comodule
map so a total integral for the action pX, Gq.
p2q ñ p3q Follows from base change, by Lemma 6.3.
p3q ñ p1q Denote by β : C A Ñ C B the total integral for the action pC X, CGq. Recall
that IdC bR A is a C A-comodule via IdC bδ and C B is a C A-comodule map via IdC bρB .
Consider the composite β1 :“ µC ˝ β : C bR A Ñ B where µC : C bR B Ñ B comes
from the algebra multiplication of B. For any b P B and c P C,
ř
pµC b IdA qppIdC b ρB qpc b bqq “ pµC b IdA qp c b bp0q b bp1q q
ř
“
cbp0q b bp1q “ ρB pcqρB pbq (since C “ BA )
“ ρB pµC pb b cqq (since B is an A-comodule algebra).
Thus, µC is an AC -comodule map and β1 also being compositions of AC -comodule
maps.
6.3. Exactness of the functor of the invariants for tame actions. In the case
of a constant group scheme, by [CEPT96, Lemma 2.2], the tameness of an action is
equivalent to the surjectivity of the trace map, which generalizes a characterization
of tameness in number theory (see [Cas67, Chapter I, §5, Theorem 2]) and justifies
the choice of the terminology. In the general case, we can define a projector which
plays the role of the trace map in the case of an action by a constant group scheme
(see [Doi90, section 1]).
Lemma 6.5. [Doi90, §1] From a total integral map α : A Ñ B and M P B MA , we
can define a R-linear projector called Reynold operator
prM : M Ñ MA
ř
m ÞÑ prM pmq :“ prM,αpmq “ pmq αpSpmp1q qqmp0q .
Proof. For any m P M and a P A, we have
ρM pαpaqq “ pα b Aqδpaq “ pα b Aqp
ÿ
a1 b a2 q “
ÿ
Thus,
ρM pm.αpaqq “
ÿ
mp0q αpa1 q b mp1q a2 p˝q.
20
αpa1 q b a2 .
We obtain:
ř
ρM p mp0q αpSpmp1q qq
ř
pm q αpSpmp1q q1 b pmp0q q1 Spmp1q q1 q pby p˝qq
ř p0q p0q
pm q αpSppmp1q q2 q b pmp0q q1 Sppmp1q q1 q (since S is an antimorphism)
ř p0q p0q
ř
mp0q αpSpmp3q qq b mp1q Spmp2q q “ mp0q αpSpmp2q qq b pmp1q q1 Spmp1q q2
ř
m Spmp2q q b ǫpmp1q q1 (by the definition of an antipodeq
ř p0q
ř
mp0q Spmp1q ǫpmp1q qq b 1 “ mp0q Sppmp1q q1 ǫpmp1q q1 q b 1
ř
mp0q Spmp1q q b 1(by the properties of the counity).
prM pmq b 1
ρM pprM pmqq “
“
“
“
“
“
“
“
So, prM pmq P pMqA and prM 2 “ prM . Moreover, for any m P pMqA , ρM pmq “ m b 1,
Hence, prM pmq “ mαpsp1qq “ mαp1q “ m since αp1q “ 1. This proves that prM is a
Reynold operator.
The existence of this projector insures that the functor of invariants in exact, for
tame actions. The following lemma will permit to relate the previous two notions of
tameness.
Lemma 6.6. [CEPT96, Lemma 2.3] If the action pX, Gq is tame then the functor
of invariants p´qA : B MA Ñ C M is exact.
Proof. Let 0 Ñ M1 Ñ M2 Ñ M3 Ñ 0 be an exact sequence of BMA . Using the
notation of the previous lemma, we have the following commutative diagram:
0
prM1
|
pM1 qA “ prM1 pM1 q
/
0
φ1 |pM
1q
A
A
M1
φ1
pM2 q “ prM2 pM2 q
/
0
φ2 |pM
2q
A
0
/
prM2
|
/
|
/
M2
φ2
prM3
pM3 qA “ prM3 pM3 q
ρM1
/
M3
M1 b1
//
M1 b A
//
M2 b A
//
M3 b A
ρM2
M2 b1
ρM3
M3 b1
0
Left exactness is automatic, right exactness follows from the previous diagram.
Adding finite hypothesis, we are able to prove that the exactness of the functor
of invariant is equivalent to tameness of the action.
Proposition 6.7. Suppose that C is locally noetherian, that B is flat over C and
that A is finite, locally free over R. Then, the following assertions are equivalent:
(1) The action pX, Gq is tame.
21
(2) The functor p´qA : B MA Ñ C M, N ÞÑ pNqA is exact.
(1’) The action pX, CGq is tame.
(2’) The functor p´qC A : B MC A Ñ C M, N ÞÑ pNqCA is exact.
Proof. p1q ô p11 q follows from lemma 6.4.
p2q ô p21 q follows from B MA “ B MAC .
p1q ñ p2q follows from lemma 6.5
p21 q ñ p11 q Suppose that p´qA is exact. Since C A is finite over C (base change of A
which is finite over R) and B is finite over C, B bC C A is finite over C. Moreover, as
C is locally noetherian, B and B bR A are also of finite presentation as algebras over
C so, in particular, as C-modules. By lemma 6.1 since we suppose B flat over C, we
have the following isomorphism
pB bR B˚ qC A » ComC A pB, Bq and pB bR pB bR Aq˚ qC A » ComC A pB bC C A, Bq
Since pB b ǫq is a C-linear section of ρB : B Ñ B bR A, pB bC C Aq˚ Ñ B˚ is surjective.
Thus, from the exactness of p´qC A , we obtain the surjectivity of
pB bR pB bR Aq˚ qC A Ñ pB bR B˚ qC A
Finally, the isomorphisms above imply the surjectivity of the natural map
ComC A pB bC C A, Bq Ñ ComC A pB, Bq
This insures the existence of a C A-comodule map λB : B bC C A Ñ B such that
λB ˝ ρB “ B. The previous lemma permits to conclude the proof.
6.4. Relationship between the two notions of tameness. First, we also can
prove easily that the tame actions defines always tame quotient stacks.
Theorem 6.8. Suppose that X is noetherian, finitely presented over S, that G is
finitely presented over S and and that all the inertia groups are finite. If the action
pX, Gq is tame then the quotient stack rX{Gs is tame.
Proof. This follows directly from Lemma 6.6 and Corollary 5.5.
We manage to prove the equivalence of these two notions of tameness defined in
this paper for actions involving finite group schemes such that the B is flat over C:
Theorem 6.9. Suppose that G is finite and locally free over S. The action pX, Gq is
tame if and only if the quotient stack rX{Gs is tame. If moreover, C is noetherian
and B flat over C, the converse is true.
Proof. By [Con05, §3], we know that ρ : rX{Gs Ñ Y is a coarse moduli space.
Moreover, ρ˚ : QcohprX{Gsq Ñ QcohpYq is exact by Lemma 6.6, thus rX{Gs is tame.
The converse follows from Proposition 6.7..
22
Remark 6.10. We can replace the hypothesis Y noetherian by X of finite type over
S. In fact, by [Con05, Theorem 3.1, (2)], Y is of finite type over S so also noetherian
since S is supposed noetherian.
We obtain easily the following corollary.
Corollary 6.11. Suppose that S is noetherian and G is finite and flat over S. Then,
G is linearly reductive over S if and only if the trivial action pS, Gq is tame over S.
Remark 6.12. We say that a Hopf algebra A is relatively cosemisimple if for all
M P MA the submodules which are direct summands in MR are also direct summands
in MA . The Hopf algebra A is relatively cosemisimple if and only if the trivial action
pSpecpRq, Gq is tame (see [BW03, 16.10]). In particular, when G is finite, flat over
S, this is also equivalent to G being linearly reductive.
The following result can be seen as an analogue of the trace surjectivity that we
have proved in the constant case.
Corollary 6.13. Suppose that G is finite, locally free over S, C is locally noetherian
and B flat over C. The following assertions are equivalent:
(1) The action pX, Gq is tame.
(2) The quotient stack rX{Gs is tame.
(3) The functor p´qA : B MA Ñ C M, N ÞÑ pNqA is exact.
(4) There is a Reynold operator prM : M Ñ MA for any M P B MA .
Proof. p1q ô p2q ô p3q Follows from the previous theorem.
p1q ñ p4q follows from Lemma 6.5.
p4q ñ p2q As the functor p´qA is left exact, it is enough to prove exactness on the
right. So, let ξ : M Ñ N P B BimA pM, Nq be an epimorphism. It induces a morphism
ξA : MA Ñ N A . For n P N A , by the surjectivity of ξ, there is m P M such that
ξpmq “ n. Moreover, prM pmq P MA , so ξpprM pmqq “ prN pξpmqq “ n. Therefore ξ is
surjective.
References
[Alp08]
Jarod Alper : Good moduli spaces for Artin stacks. ProQuest LLC, Ann Arbor, MI, 2008.
Thesis (Ph.D.)–Stanford University.
[AOV08]
D. Abramovich, M. Olsson et A. Vistoli : Tame stacks in positive characteristic. Ann. Inst.
Fourier (Grenoble), 58(4):1057–1091, 2008.
[Art69]
M. Artin : Algebraic approximation of structures over complete local rings. Inst. Hautes Études
Sci. Publ. Math., (36):23–58, 1969.
[Bou81]
N. Bourbaki : Éléments de mathématique. Lecture Notes in Mathematics, Vol. 305. Masson,
Paris, 1981. Algèbre. Chapitres 4 à 7. [Algebra. Chapters 4–7].
[BW03]
T. Brzezinski et R. Wisbauer : Corings and comodules, volume 309 de London Mathematical
Society Lecture Note Series. Cambridge University Press, Cambridge, 2003.
23
[Cas67]
Algebraic number theory. Proceedings of an instructional conference organized by the London Mathematical Society (a NATO Advanced Study Institute) with the support of the Inter
national Mathematical Union. Edited by J. W. S. Cassels and A. Fröhlich. Academic Press,
London, 1967.
[CE92]
T. Chinburg et B. Erez : Equivariant Euler-Poincaré characteristics and tameness. Astérisque,
(209):13, 179–194, 1992. Journées Arithmétiques, 1991 (Geneva).
[CEPT96] T. Chinburg, B. Erez, G. Pappas et M. J. Taylor : Tame actions of group schemes: integrals
and slices. Duke Math. J., 82(2):269–308, 1996.
[Con05]
B. Conrad : The Keel-Mori theorem via stacks. Preprint from the University of Stanford
available at http://math.stanford.edu/˜conrad/papers/coarsespace.pdf, 2005.
[DG70]
M. Demazure et P. Gabriel : Groupes algébriques. Tome I: Géométrie algébrique, généralités,
groupes commutatifs. Masson & Cie, Éditeur, Paris, 1970. Avec un appendice Corps de classes
local par Michiel Hazewinkel.
[Doi90]
Y. Doi : Hopf extensions of algebras and Maschke type theorems. Israel J. Math., 72(1-2):99–
108, 1990. Hopf algebras.
[GM71]
A. Grothendieck et J. P. Murre : The tame fundamental group of a formal neighbourhood of
a divisor with normal crossings on a scheme. Lecture Notes in Mathematics, Vol. 208. SpringerVerlag, Berlin, 1971.
[Har77]
R. Hartshorne : Algebraic geometry. Springer-Verlag, New York, 1977. Graduate Texts in
Mathematics, No. 52.
[Kem00]
G. Kemper : A characterization of linearly reductive groups by their invariants. Transform.
Groups, 5:85–92, 2000.
[LMB00]
G. Laumon et L. Moret-Bailly : Champs algébriques, volume 39 de Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results
in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics].
Springer-Verlag, Berlin, 2000.
[Lun73]
D. Luna : Slices étales. In Sur les groupes algébriques, pages 81–105. Bull. Soc. Math. France,
Paris, Mémoire 33. Soc. Math. France, Paris, 1973.
[Mar]
Sophie Marques : Tameness for actions of affine group schemes and quotient stacks. Preprint
available at https://files.nyu.edu/sm5439/public/memoire/these.pdf.
[Ols06]
M. C. Olsson : Deformation theory of representable morphisms of algebraic stacks. Math. Z.,
253(1):25–62, 2006.
[Ols07]
M. C. Olsson : Sheaves on Artin stacks. J. Reine Angew. Math., 603:55–112, 2007.
[Ray70]
M. Raynaud : Anneaux locaux henséliens. Lecture Notes in Mathematics, Vol. 169. SpringerVerlag, Berlin, 1970.
[Ser68]
J.-P. Serre : Corps locaux. Hermann, Paris, 1968. Deuxième édition, Publications de
l’Université de Nancago, No. VIII.
(MARQUES Sophie) Visiting assistant professor at New York University, Courant institute of mathematical sciences, 251 Mercer St, New York, NY 10012, USA
E-mail address: [email protected]
24
| 0math.AC
|
arXiv:1611.07237v2 [math.ST] 24 Nov 2016
Multivariate Intensity Estimation via Hyperbolic Wavelet
Selection
NATHALIE AKAKPO
Laboratoire de Probabilités et Modèles Aléatoires (LPMA), UMR 7599
Université Pierre et Marie Curie (UPMC), Paris
Centre de Recherches Mathématiques (CRM), UMI 3457
Université de Montréal (UdeM)
Abstract. We propose a new statistical procedure able in some way to overcome the curse
of dimensionality without structural assumptions on the function to estimate. It relies on
a least-squares type penalized criterion and a new collection of models built from hyperbolic
biorthogonal wavelet bases. We study its properties in a unifying intensity estimation framework,
where an oracle-type inequality and adaptation to mixed smoothness are shown to hold. Besides,
we describe an algorithm for implementing the estimator with a quite reasonable complexity.
Keywords: Hyperbolic wavelets; Biorthogonal wavelets; Mixed smoothness; Model selection; Density;
Copula; Poisson process; Lévy process.
Contents
1. Introduction
2. Framework and examples
2.1. General framework
2.2. Examples
3. Estimation on a given pyramidal wavelet model
3.1. Wavelets on L2 ([0, 1])
3.2. Hyperbolic wavelet basis on L2 ([0, 1]d )
3.3. Pyramidal models
3.4. Least-squares type estimator on a pyramidal model
3.5. Quadratic risk on a pyramidal model
4. Wavelet pyramid model selection
4.1. Penalized pyramid selection
4.2. Combinatorial complexity and choice of the penalty function
4.3. Back to the examples
5. Adaptivity to mixed smoothness
5.1. Function spaces with dominating mixed smoothness
5.2. Link with structural assumptions
5.3. Approximation qualities and minimax rate
6. Implementing wavelet pyramid selection
6.1. Algorithm and computational complexity
6.2. Illustrative examples
7. Proofs
7.1. Proof of Proposition 2
7.2. Proof of Proposition 3
E-mail address: [email protected].
Date: November 28, 2016.
1
2
3
3
3
5
5
6
7
7
8
10
10
10
12
14
14
15
16
18
18
18
20
20
21
2
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
7.3. Proof of Proposition 4
7.4. Proof of Theorem 1
7.5. Proofs of Corollaries 1 to 5
7.6. Proof of Proposition 5
7.7. Proof of Theorem 2
References
22
23
26
27
28
29
1. Introduction
Over the last decades, many wavelet procedures have been developed in various statistical
frameworks. Yet, in multivariate settings, most of them are based on isotropic wavelet bases.
These indeed have the advantage of being as easily tractable as their univariate counterparts since
each isotropic wavelet is a tensor product of univariate wavelets coming from the same resolution
level. Notable counterexamples are [Don97], [Neu00] and [NvS97], or [ACF15] and [ACF14]. They
underline the usefulness of hyperbolic wavelet bases, where coordinatewise varying resolution levels
are allowed, so as to recover a wider range of functions, and in particular functions with anisotropic
smoothness.
Much attention has also been paid to the so-called curse of dimensionality. A common way
to overcome this problem in Statistics is to impose structural assumptions on the function to
estimate. In a regression framework, beyond the well-known additive and single-index models,
we may cite the work of [HM07] who propose a spline-based method in an additive model with
unknown link function, or the use of ANOVA-like decompositions in [IS07] or [DIT14]. Besides,
two landmark papers consider a general framework of composite functions, encompassing several
classical structural assumptions: [JLT09] propose a kernel-based procedure in the white noise
framework, whereas [BB14] propose a general model selection procedure with a wide scope of
applications. Finally, Lepski [Lep13] (see also [Reb15b, Reb15a]) consider density estimation
with adaptation to a possibly multiplicative structure of the density. In the meanwhile, in the
field of Approximation Theory and Numerical Analysis, a renewed interest in function spaces
with dominating mixed smoothness has been growing (see for instance [DTU16]), due to their
tractability for multivariate integration for instance. Such spaces do not impose any structure,
but only that the highest order derivative is a mixed derivative. Surprisingly, in the statistical
literature, it seems that only the thresholding-type procedures of [Neu00] and [BPP13] deal with
such spaces, either in the white noise framework or in a functional deconvolution model.
In order to fill this gap, this paper is devoted to a new statistical procedure based on wavelet
selection from hyperbolic biorthogonal bases. We underline its universality by studying it in a
general intensity estimation framework, encompassing many examples of interest such as density,
copula density, Poisson intensity or Lévy jump intensity estimation. We first define a whole
collection of linear subspaces, called models, generated by subsets of the dual hyperbolic basis,
and a least-squares type criterion adapted to the norm induced by the primal hyperbolic basis.
Then we describe a procedure to choose the best model from the data by using a penalized
approach similar to [BBM99]. Our procedure satisfies an oracle-type inequality provided the
intensity to estimate is bounded. Besides, it reaches the minimax rate up to a constant factor,
or up to a logarithmic factor, over a wide range of spaces with dominating mixed smoothness,
and this rate is akin to the one we would obtain in a univariate framework. Notice that, contrary
to [Neu00] or [BPP13], we allow for a greater variety of such spaces (of Sobolev, Hölder or Besov
type smoothness) and also for spatially nonhomogeneous smoothness. For that purpose, we prove
a key result from nonlinear approximation theory, in the spirit of [BM00], that may be of interest
for other types of model selection procedures (see for instance [Bir06, Bar11, BB16]). Depending
on the kind of intensity to estimate, different structural assumptions might make sense, some of
which have been considered in [JLT09], [BB14], [Lep13], [Reb15b, Reb15a], but not all. We explain
in what respect these structural assumptions fall within the scope of estimation under dominating
mixed smoothness. Yet, we emphasize that we do not need to impose any structural assumptions
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
3
on the target function. Thus in some way our method is adaptive at the same time to many
structures. Besides, it can be implemented with a computational complexity linear in the sample
size, up to logarithmic factors.
The plan of the paper is as follows. In Section 2, we describe the general intensity estimation
framework and several examples of interest. In Section 3, we define the so-called pyramidal wavelet
models and a least-squares type criterion, and provide a detailed account of estimation on a given
model. Section 4 is devoted to the choice of an adequate penalty so as to perform data-driven
model selection. The optimality of the resulting procedure from the minimax point of view is then
discussed in Section 5, under mixed smoothness assumptions. The algorithm for implementing our
wavelet procedure and an illustrative example are given in Section 6. All proofs are postponed to
Section 7. Let us end with some remark about the notation. Throughout the paper, C, C1 , . . . will
stand for numerical constants, and C(θ), C1 (θ), . . . for positive reals that only depend on some θ.
Their values are allowed to change from line to line.
2. Framework and examples
Qd
2.1. General framework. Let d ∈ N, d ≥ 2, and Q = k=1 [ak , bk ] be a given hyperrectangle in
Rd equipped with its Borel σ−algebra B(Q) and the Lebesgue measure. We denote by L2 (Q) the
space of square integrable functions on Q, equipped with its usual norm
sZ
t2 (x)dx
(1)
ktk =
Q
and scalar product h., .i. In this article, we are interested in a nonnegative measure on B(Q) that
admits a bounded density s with respect to the Lebesgue measure, and our aim is to estimate that
function s over Q. Given a probability space (Ω, E, P), we assume that there exists some random
measure M defined on (Ω, E, P), with values in the set of Borel measures on Q such that, for all
A ∈ B(Q),
E [M (A)] = h1IA , si.
(2)
By classical convergence theorems, this condition implies that, for all nonnegative or bounded
measurable functions t,
Z
(3)
E
tdM = ht, si.
Q
c, which is close enough to M in a sense to
We assume that we observe some random measure M
c = M.
be made precise later. When M can be observed, we set of course M
2.2. Examples. Our general framework encompasses several special frameworks of interest, as
we shall now show.
2.2.1. Example 1: density estimation. Given n ∈ N? , we observe identically distributed random variables Y1 , . . . , Yn with common density s with respect to the Lebesgue measure on Q =
Qd
k=1 [ak , bk ] . The observed empirical measure is then given by
n
X
c(A) = M (A) = 1
M
1IA (Yi ), for A ∈ B(Q),
n i=1
and obviously satisfies (2).
2.2.2. Example 2: copula density estimation. Given n ∈ N? , we observe independent and identically distributed random variables X1 , . . . , Xn with values in Rd . For i = 1, . . . , n and j = 1, . . . , d,
the j-th coordinate Xij of Xi has continuous distribution function Fj . We recall that, from Sklar’s
Theorem [Skl59] (see also [Nel06], for instance), there exists a unique distribution function C on
[0, 1]d with uniform marginals such that, for all (x1 , . . . , xd ) ∈ Rd ,
P(Xi1 ≤ x1 , . . . , Xid ≤ xd ) = C(F1 (x1 ), . . . , Fd (xd )).
4
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
This function C is called the copula of Xi1 , . . . , Xid . We assume that it admits a density s with
respect to the Lebesgue measure on Q = [0, 1]d . Since C is the joint distribution function of the
Fj (X1j ), j = 1, . . . , d, a random measure satisfying (2) is given by
n
M (A) =
1X
1IA (F1 (Xi1 ), . . . , Fd (Xid )) , for A ∈ B([0, 1]d ).
n i=1
As the marginal distributions Fj are usually unknown, we replace them by the empirical distribution functions F̂nj , where
n
1X
1IX ≤t ,
F̂nj (t) =
n i=1 ij
and define
n
X
c(A) = 1
M
1IA F̂n1 (Xi1 ), . . . , F̂nd (Xid ) , for A ∈ B([0, 1]d ).
n i=1
2.2.3. Example 3: Poisson intensity estimation. Let us denote by Vold (Q) the Lebesgue meaQd
sure of Q = k=1 [ak , bk ] . We observe a Poisson process N on Q whose mean measure has
intensity Vold (Q)s. Otherwise said, for all finite family (Ak )1≤k≤K of disjoint measurable subsets
of Q, N (A
R independent Poisson random variables with respective parameters
R 1 ), . . . , N (AK ) are
Vold (Q) A1 s, . . . , Vold (Q) AK s. Therefore the empirical measure
c(A) = M (A) = N (A) , for A ∈ B(Q),
M
Vold (Q)
does satisfy (2). We do not assume s to be constant throughout Q so that the Poisson process
may be nonhomogeneous.
2.2.4. Example 4: Lévy jump intensity estimation (continuous time). Let T be a fixed positive
real, we observe on [0, T ] a Lévy process X = (Xt )t≥0 with values in Rd . Otherwise said, X is
a process starting at 0, with stationary and independent increments, and which is continuous in
probability with càdlàg trajectories (see for instance [Ber96, Sat99, CT04]). This process may
have jumps, whose sizes are ruled by the so-called jump intensity measure or Lévy measure. An
important example of such process is the compound Poisson process
Xt =
Nt
X
ξi , t ≥ 0,
i=1
where (Nt )t≥0 is a univariate homogeneous Poisson process, (ξi )i≥i are i.i.d. with values in Rd
and distribution ρ with no mass at 0, and (Nt )t≥0 and (ξi )i≥i are independent. In this case, ρ is
also the Lévy measure of X.
Here, we assume that the Lévy measure admits a density f with respect to the Lebesgue
Qd
measure on Rd \{0}. Given some compact hyperrectangle Q = k=1 [ak , bk ] ⊂ Rd \{0}, our aim is
to estimate the restriction s of f to Q. For that purpose, we use the observed empirical measure
ZZ
c(A) = M (A) = 1
M
N (dt, dx), for A ∈ B(Q).
T
[0,T ]×A
A well-known property
of Lévy processes states that the random measure N defined for B ∈
B [0, +∞) × Rd \{0} by
N (B) = ]{t > 0/(t, Xt − Xt− ) ∈ B}
is a Poisson process with mean measure
Z Z
µ(B) =
f (x)dtdx,
B
so that M satisfies (2).
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
5
2.2.5. Example 5: Lévy jump intensity estimation (discrete time). The framework is the same as
in Example 4, except that (Xt )t≥0 is not observed. Given some time step ∆ > 0 and n ∈ N? , we
only have at our disposal the random variables
Yi = Xi∆ − X(i−1)∆ , i = 1, . . . n.
In order to estimate s on Q, we consider the random measure
ZZ
1
N (dt, dx), for A ∈ B(Q),
M (A) =
n∆
[0,n∆]×A
which is unobserved, and replaced for estimation purpose with
n
X
c(A) = 1
M
1IA (Yi ), for A ∈ B(Q).
n∆ i=1
3. Estimation on a given pyramidal wavelet model
The first step of our estimation procedure relies on the definition of finite dimensional linear
subspaces of L2 (Q), called models, generated by some finite families of biorthogonal wavelets. We
only describe here models for Q = [0, 1]d . For a general hyperrectangle Q, the adequate models
can be deduced by translation and scaling. We then introduce a least-squares type contrast that
allows to define an estimator of s within a given wavelet model.
3.1. Wavelets on L2 ([0, 1]). We shall first introduce a multiresolution analysis and a wavelet
basis for L2 ([0, 1]) satisfying the same general assumptions as in [Hoc02b] and [Hoc02a]. Concrete
examples of wavelet bases satisfying those assumptions may be found in [CDV93] and [DKU99] for
instance. In the sequel, we denote by κ some positive constant, that only depends on the choice
of the bases. We fix the coarsest resolution level at j0 ∈ N. On the one hand, we assume that the
scaling spaces
Vj = Vect{φλ ; λ ∈ ∆j } and Vj? = Vect{φ?λ ; λ ∈ ∆j }, j ≥ j0 ,
satisfy the following hypotheses:
S.i) (Riesz bases) For all j ≥ j0 , {φλ ; λ ∈ ∆j } are linearly independent functions from L2 ([0, 1]),
P
∼
so are {φ?λ ; λ ∈ ∆j }, and they form Riesz bases of Vj and Vj? , i.e.
λ∈∆j aλ φλ
P
1/2
P
2
?
∼
λ∈∆j aλ
λ∈∆j aλ φλ .
S.ii) (Dimension) There exists some nonnegative integer B such that, for all j ≥ j0 , dim(Vj ) =
dim(Vj? ) = ]∆j = 2j + B.
?
S.iii) (Nesting) For all j ≥ j0 , Vj ⊂ Vj+1 and Vj? ⊂ Vj+1
.
?
S.iv) (Density) ∪j≥j0 Vj = ∪j≥j0 Vj = L2 ([0, 1]).
S.v) (Biorthogonality) Let j ≥ j0 , for all λ, µ ∈ ∆j , hφλ , φ?µ i = δλ,µ .
S.vi) (Localization) Let j ≥ j0 , for all λ ∈ ∆j , |Supp(φλ )| ∼ |Supp(φ?λ )| ∼ 2−j .
S.vii) (Almost disjoint supports) For all j ≥ j0 and all λ ∈ ∆j ,
max ]{µ ∈ ∆j s.t. Supp(φλ ) ∩ Supp(φµ ) 6= ∅}, ]{µ ∈ ∆j s.t. Supp(φ?λ ) ∩ Supp(φ?µ ) 6= ∅} ≤ κ.
S.viii) (Norms) For all j ≥ j0 and all λ ∈ ∆j , kφλ k = kφ?λ k = 1 and max(kφλ k∞ , kφ?λ k∞ ) ≤ κ2j/2 .
S.ix) (Polynomial reproducibility) The primal scaling spaces are exact of order N , i.e. for all
j ≥ j0 , ΠN −1 ⊂ Vj , where ΠN −1 is the set of all polynomial functions with degree ≤ N − 1
over [0, 1].
On the other hand, the wavelet spaces
Wj = Vect{ψλ ; λ ∈ ∇j } and Wj? = Vect{ψλ? ; λ ∈ ∇j }, j ≥ j0 + 1,
fulfill the following conditions:
W.i) (Riesz bases) The functions {ψλ ; λ ∈ ∪j≥j0 +1 ∇j } are linearly independent. Together with
the {φλ ; λ ∈ ∆j0 }, they form a Riesz basis for L2 ([0, 1]). The same holds for the ψ ? and
the φ? .
6
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
?
?
?
, with Vj ⊥ Wj+1
= Vj? ⊕ Wj+1
W.ii) (Orthogonality) For all j ≥ j0 , Vj+1 = Vj ⊕ Wj+1 and Vj+1
and Vj? ⊥ Wj+1 .
W.iii) (Biorthogonality) Let j ≥ j0 , for all λ, µ ∈ ∇j+1 , hψλ , ψµ? i = δλ,µ .
W.iv) (Localization) Let j ≥ j0 , for all λ ∈ ∇j+1 , |Supp(ψλ )| ∼ |Supp(ψλ? )| ∼ 2−j .
W.v) (Almost disjoint supports) For all j ≥ j0 and all λ ∈ ∇j+1 ,
max ]{µ ∈ ∇j+1 s.t. Supp(ψλ ) ∩ Supp(ψµ ) 6= ∅}, ]{µ ∈ ∇j+1 s.t. Supp(ψλ? ) ∩ Supp(ψµ? ) 6= ∅} ≤ κ.
W.vi) (Norms) For all j ≥ j0 and all λ ∈ ∇j+1 , kψλ k = kψλ? k = 1 and max(kψλ k∞ , kψλ? k∞ ) ≤
κ2j/2 .
W.vii) (Fast Wavelet Transform) Let j ≥ j0 , for all λ ∈ ∇j+1 ,
]{µ ∈ ∆j+1 |hψλ , φµ i =
6 0} ≤ κ
and for all µ ∈ ∆j+1
The same holds for the
ψλ?
|hψλ , φµ i| ≤ κ.
and the φ?λ .
Remarks:
• These properties imply that any function f ∈ L2 ([0, 1]) may be decomposed as
X
X X
(4)
f=
hf, φλ iφ?λ +
hf, ψλ iψλ? .
λ∈∆j0
j≥j0 +1 λ∈∇j
• Properties S.ii) and W.ii) imply that dim(Wj+1 ) = 2j .
• Property W.vii) means in particular that, for each resolution level j, any wavelet can be
represented as a linear combination of scaling functions from the same resolution level
with a number of components bounded independently of the level as well as the amplitude
of the coefficients.
As is well known, contrary to orthogonal bases, biorthogonal bases allow for both symmetric
and smooth wavelets. Besides, properties of dual biorthogonal bases are usually not the same.
Usually, in decomposition (4), the analysis wavelets φλ and ψλ are the one with most null moments,
whereas the synthesis wavelets φ?λ and ψλ? are the one with greatest smoothness. Yet, we may
sometimes need the following smoothness assumptions on the analysis wavelets (not very restrictive
c.
in practice), only to bound residual terms due to the replacement of M with M
Assumption (L). For all λ ∈ ∆j0 , for all j ≥ j0 and all µ ∈ ∇j+1 , φλ and ψµ are Lipschitz
functions with Lipschitz norms satisfying kφλ kL ≤ κ23j0 /2 and kψµ kL ≤ κ23j/2 .
We still refer to [CDV93] and [DKU99] for examples of wavelet bases satisfying this additional
assumption.
3.2. Hyperbolic wavelet basis on L2 ([0, 1]d ). In the sequel, for ease of notation, we set Nj0 =
{j ∈ N, j ≥ j0 }, ∇j0 = ∆j0 , Wj0 = Vj0 and Wj?0 = Vj?0 , and for λ ∈ ∇j0 , ψλ = φλ and
ψλ? = φ?λ . Given a biorthogonal basis of L2 ([0, 1]) chosen according to 3.1, we deduce biorthogonal
wavelets of L2 ([0, 1]d ) by tensor product. More precisely, for j = (j1 , . . . , jd ) ∈ Ndj0 , we set ∇j =
∇j1 × . . . × ∇jd and for all λ = (λ1 , . . . , λd ) ∈ ∇j , we define Ψλ (x1 , . . . , xd ) = ψλ1 (x1 ) . . . ψλd (xd )
and Ψ?λ (x1 , . . . , xd ) = ψλ?1 (x1 ) . . . ψλ?d (xd ). Contrary to most statistical works based on wavelets,
we thus allow for tensor products of univariate wavelets coming from different resolution levels
j1 , . . . , jd . Writing Λ = ∪j∈Ndj ∇j , the families {Ψλ ; λ ∈ Λ} and {Ψ?λ ; λ ∈ Λ} define biorthogonal
0
bases of L2 ([0, 1]d ) called biorthogonal hyperbolic bases. Indeed,
L2 ([0, 1]d ) = ∪j≥j0 Vj ⊗ . . . ⊗ Vj ,
and for all j ≥ j0 ,
Vj ⊗ . . . ⊗ Vj = (Wj0 ⊕ Wj0 +1 ⊕ . . . ⊕ Wj ) ⊗ . . . ⊗ (Wj0 ⊕ Wj0 +1 ⊕ . . . ⊕ Wj )
=
⊕
j0 ≤k1 ,...,kd ≤j
Wk1 ⊗ . . . ⊗ Wkd .
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
7
In the same way,
L2 ([0, 1]d ) = ∪j≥j0
⊕
j0 ≤k1 ,...,kd ≤j
Wk?1 ⊗ . . . ⊗ Wk?d .
Besides, they induce on L2 ([0, 1]d ) the norms
sX
sX
2
(5)
ktkΨ =
ht, Ψλ i and ktkΨ? =
ht, Ψ?λ i2 ,
λ∈Λ
λ∈Λ
which are both equivalent to k.k, with equality when the wavelet basis is orthogonal. It should be
noticed that the scalar product derived from k.kΨ , for instance, is
X
(6)
ht, uiΨ =
ht, Ψλ ihu, Ψλ i.
λ∈Λ
3.3. Pyramidal models. A wavelet basis in dimension 1 has a natural pyramidal structure when
the wavelets are grouped according to their resolution level. A hyperbolic basis too, provided we
define a proper notion of resolution level that takes into account anisotropy: for a wavelet Ψλ or
Ψ?λ with λ ∈ ∇j , we define the global resolution level as |j| := j1 + . . . + jd . Thus, the supports
of all wavelets corresponding to a given global resolution level ` ∈ Ndj0 have a volume of roughly
2−` but exhibit very different shapes. For all ` ∈ Ndj0 , we define J` = {j ∈ Ndj0 /|j| = `}, and
U ∇(`) = ∪j∈J` ∇j the index set for d-variate wavelets at resolution level `.
Given some maximal resolution level L• ∈ Ndj0 , we define, for all `1 ∈ {dj0 + 1, . . . , L• + 1}, the
family MP
`1 of all sets m of the form
!
`[
L•[
−`1
1 −1
m=
m(`1 + k) ,
U ∇(`) ∪
k=0
`=dj0
where, for all 0 ≤ k ≤ L• − `1 , m(`1 + k) may be any subset of U ∇(`1 + k) with N (`1 , k) elements.
Typically, N (`1 , k) will be chosen so as to impose some sparsity: it is expected to be smaller than
the total number of wavelets at level `1 + k and to decrease when the resolution level increases.
An adequate choice of N (`1 , k) will be proposed in Proposition 4. Thus, choosing a set in MP
`1
amounts to keep all hyperbolic wavelets at level at most `1 − 1, but only a few at deeper levels.
P
• +1
We set MP = ∪L
`1 =dj0 +1 M`1 and define a pyramidal model as any finite dimensional subspace of
the form
?
Sm
= Vect{Ψ?λ ; λ ∈ m}, for m ∈ MP .
SL •
?
We denote by Dm the dimension of Sm
. Setting m• = `=dj
U ∇(`), we can see that all pyramidal
0
?
models are included in Sm
.
•
3.4. Least-squares type estimator on a pyramidal model. Let us fix some model m ∈ MP .
If the random measure M is observed, then we can build a least-squares type estimator š?m for s
?
with values in Sm
and associated with the norm k.kΨ defined by (5). Indeed, setting
X
γ(t) = ktk2Ψ − 2
ht, Ψλ iβ̌λ ,
λ∈Λ
where
Z
β̌λ =
Ψλ dM,
Q
we deduce from (3) that s minimizes over t ∈ L2 (Q)
X
ks − tk2Ψ − ksk2Ψ = ktk2Ψ − 2
ht, Ψλ ihs, Ψλ i = E[γ(t)],
λ∈Λ
so we introduce
š?m = argmin γ(t).
?
t∈Sm
8
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
For all sequences of reals (αλ )λ∈m ,
!
(7)
γ
X
αλ Ψ?λ
=
λ∈m
X
αλ − β̌λ
2
−
λ∈m
X
β̌λ2 ,
λ∈m
hence
š?m =
X
β̌λ Ψ?λ .
λ∈m
c, we consider the pseudo-least-squares contrast
Since we only observe the random measure M
X
γ
b(t) = ktk2 − 2
ht, Ψλ iβbλ ,
Ψ
λ∈Λ
where
Z
c,
Ψλ dM
βbλ =
Q
?
as
and we define the best estimator of s within Sm
sb?m = argmin γ
b(t) =
?
t∈Sm
X
βbλ Ψ?λ .
λ∈m
3.5. Quadratic risk on a pyramidal model. Let us introduce the orthogonal projection of s
?
for the norm k.kΨ , that is
on Sm
X
s?m =
βλ Ψ?λ ,
λ∈m
where
βλ = hΨλ , si.
It follows from (3) that β̌λ is an unbiased estimator for βλ , so that š?m is an unbiased estimator
for s?m . Thanks to Pythagoras’ equality, we recover for š?m the usual decomposition
X
(8)
E ks − š?m k2Ψ = ks − s?m k2Ψ +
Var(β̌λ ),
λ∈m
where the first term is a bias term or approximation error and the second term is a variance
c is observed, combining the triangle inequality, the basic
term or estimation error. When only M
inequality (14) and (8) easily provides at least an upper-bound akin to (8), up to a residual term.
Proposition 1. For all θ > 0,
!
X
Var(β̌λ ) + (1 + 1/θ)E kš?m − ŝ?m k2Ψ .
E ks − ŝ?m k2Ψ ≤ (1 + θ) ks − s?m k2Ψ +
λ∈m
c = M, θ can be taken equal to 0 and equality holds.
When M
In all the examples introduced in Section 2.2, we shall verify that the quadratic risks satisfies, for
all θ > 0,
(9)
ksk∞ Dm
E ks − ŝ?m k2Ψ ≤ c1 ks − s?m k2Ψ + c2
+ r1 (n̄),
n̄
where n̄ describes the amount of available data, and the residual term r1 (n̄) does not weigh too
much upon the estimation rate.
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
9
3.5.1. Example 1: density estimation (continued). In this framework, the empirical coefficients
are of the form
n
1X
β̌λ =
Ψλ (Yi ).
n i=1
As the wavelets are normalized and s is bounded,
Z
1
ksk∞
Ψ2 (x)s(x)dx ≤
,
Var(β̌λ ) ≤
n Q λ
n
so
ksk∞ Dm
E ks − ŝ?m k2Ψ ≤ ks − s?m k2Ψ +
.
n
Hence (9) is satisfied for instance with n̄ = n, c1 = c2 = 1, r1 (n) = 0.
3.5.2. Example 2: copula density estimation (continued). In this case,
n
β̌λ =
while
1X
Ψλ (F1 (Xi1 ), . . . , Fd (Xid )) ,
n i=1
n
1X
Ψλ F̂n1 (Xi1 ), . . . , F̂nd (Xid ) .
βbλ =
n i=1
As in Example 1, Var(β̌λ ) ≤ ksk∞ /n. Besides we prove in Section 7.1 the following upper-bound
for the residual terms.
Proposition 2. Under Assumption (L), for all m ∈ MP ,
E[kš?m − ŝ?m k2 ] ≤ C(κ, d)Ld−1
24L• log(n)/n.
•
√
√
Hence choosing n̄ = n, 24L• = n/ log(n), c1 = c2 = 2, and r1 (n) = C(κ, d) log(n)d−1 / n
yields (9).
3.5.3. Example 3: Poisson intensity estimation (continued). In this case,
Z
1
Ψλ (x)N (dx).
β̌λ =
Vold (Q) Q
From Campbell’s formula,
Var(β̌λ ) =
1
Vold (Q)
Z
Ψ2λ (x)s(x)dx
Q
so
ksk∞ Dm
E ks − ŝ?m k2Ψ ≤ ks − s?m k2Ψ +
,
n̄
with n̄ = Vold (Q).
3.5.4. Example 4: Lévy jump intensity estimation with continuous time observations (continued).
In this case,
ZZ
1
β̌λ =
Ψλ (x)N (dt, dx).
T
[0,T ]×Q
From Campbell’s formula again,
Var(β̌λ ) =
1
T2
ZZ
Ψ2λ (x)s(x)dtdx
[0,T ]×Q
so
ksk∞ Dm
E ks − ŝ?m k2Ψ ≤ ks − s?m k2Ψ +
,
n̄
with n̄ = T.
10
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
3.5.5. Example 5: Lévy jump intensity estimation with discrete time observations (continued). In
this case, the empirical coefficients and their approximate counterparts are of the form
ZZ
n
1 X
1
b
Ψλ (x)N (dt, dx) and βλ =
β̌λ =
Ψλ (Xi∆ − X(i−1)∆ ).
n∆
n∆ i=1
[0,n∆]×Q
We deduce as previously that Var(β̌λ ) ≤ ksk∞ /n̄ with n̄ = n∆. Besides we can bound the residual
term thanks to the following proposition, proved in Section 7.2.
Proposition 3. Under Assumption (L), for all m ∈ MP ,
E[kš?m − ŝ?m k2 ] ≤ 8
24L• n∆3 + 23L• ∆
ksk∞ Dm
+ C(κ, d, f, Q)Ld−1
.
•
n∆
n∆
provided ∆ is small enough.
Assuming n∆2 stays bounded while n∆ → ∞ as n → ∞, and choosing 24L• = n∆, we deduce that (9) is satisfied under Assumption (L) with n̄ = n∆, c1 = 2, c2 = 18, and r1 (n̄) =
C(κ, d, f, Q) log(n̄)d−1 /n̄. Notice that these assumptions on n and ∆ are classical in the so-called
framework of high-frequency observations.
Remark: Proposition 3 extends [FL09] to a multivariate model with a complex structure due
to the use of hyperbolic wavelets, instead of isotropic ones. Yet, the extension is not so straightforward, so we give a detailed proof in Section 7.
4. Wavelet pyramid model selection
The upper-bound (9) for the risk on one pyramidal model suggests that a good model should
be large enough so that the approximation error is small, and small enough so that the estimation
error is small. Without prior knowledge on the function s to estimate, choosing the best pyramidal
model is thus impossible. In this section, we describe a data-driven procedure that selects the best
pyramidal model from the data, without using any smoothness assumption on s. We provide
theoretical results that guarantee the performance of such a procedure. We underline how these
properties are linked with the structure of the collection of models.
4.1. Penalized pyramid selection. When M is observed, we deduce from (8) that
X
E ks − š?m k2Ψ − ksk2Ψ = −ks?m k2Ψ +
Var(β̌λ )
λ∈m
that γ(š?m )
P
+
−kš?m k2Ψ .
=
Following the work of [BBM99], we introduce a penalty
and from (7)
function pen : M → R and choose a best pyramidal model from the data defined as
m̂P = argmin (γ̂(ŝ?m ) + pen(m)).
m∈MP
In order to choose the pyramidal model with smallest quadratic risk, the penalty pen(m) is expected to behave roughly as the estimation error within model m. We provide such a penalty in
the following Section. Our final estimator for s is then
s̃P = ŝ?m̂P .
4.2. Combinatorial complexity and choice of the penalty function. As widely examplified
in [Mas07, BGH09] for instance, the choice of an adequate penalty depends on the combinatorial
complexity of the collection of models, which is measured through the index
log ]MP
`1
,
(10)
max
dj0 +1≤`1 ≤L• +1
D(`1 )
where D(`1 ) is the common dimension of all pyramidal models in MP
`1 . Ideally, this index should
be upper-bounded independently of the sample size for the resulting model selection procedure
to reach the optimal estimation rate. The following proposition describes the combinatorial complexity of the collection of pyramidal models.
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
11
Proposition 4. Let M = 2+B/2j0 −1 . For all `1 ∈ {dj0 +1, . . . , L• +1} and all k ∈ {0, . . . , L• −`1 },
let
N (`1 , k) = b2]U ∇(`1 + k)(k + 2)−(d+2) 2−k M −d c
(11)
and D(`1 ) be the common dimension of all models in MP
`1 . There exists positive reals κ1 (d),
κ2 (j0 , B, d) and κ3 (j0 , B, d) such that
κ1 (d)(`1 − dj0 + d − 2)d−1 2`1 ≤ D(`1 ) ≤ κ2 (j0 , B, d)(`1 − dj0 + d − 2)d−1 2`1
and
log ]MP
`1 ≤ κ3 (j0 , B, d)D(`1 ).
We remind that B is defined in Section 3.1(Assumption S.ii)). Possible values for κ1 , κ2 and κ3
are given in the proof, which is postponed
to Section 7.3. In the same way, we could prove a
matching lower-bound for log ]MP
for
large
enough `1 , so that the whole family MP contains
`1
d−1 L•
of order of L• 2 models. Typically, we will choose L• such that 2L• is a power of the sample
size n̄. So while MP contains at least an exponential number of models, the number of models
per dimension is moderate enough so that the combinatorial index (10) bounded.
From now on, we assume that (11) is satisfied, as well as the following hypotheses. For all
?
, let
subfamily T of Sm
•
!
Z X
Z(T ) = sup
ht, Ψλ iΨλ dM − hs, tiΨ .
t∈T
Q λ∈m
•
Assumption
(Conc). There exist positive reals n̄, κ01 , κ02 , κ03 such that, for all countable subfamily
?
T of t ∈ Sm• |ktkΨ = 1 satisfying
sup
t∈T
X
ht, Ψλ iΨλ
λ∈m•
≤ B(T )
∞
for some positive constant B(T ), we have, for all x > 0,
r
x
x
P Z(T ) ≥ κ01 E [Z(T )] + κ02 ksk∞ + κ03 B(T )
≤ exp(−x).
n̄
n̄
Assumption (Var). There exist a nonnegative constant κ04 and a collection of estimators
(σ̂λ2 )λ∈m• such that, for all λ ∈ m• ,
E σ̂λ2 ≤ κ04 max(ksk∞ , 1).
Besides there exist a nonnegative constant κ05 , a nonnegative function w such that w(n̄)/n̄ −−−−→ 0,
n̄→∞
and a measurable event Ωσ on which, for all λ ∈ m• ,
Var(β̌λ ) ≤ κ05
max{σ̂λ2 , 1}
n̄
and such that
pσ := P(Ωcσ ) ≤
w(n̄)
.
n̄
Assumption (Rem). For the same function w as in Assumption (Var) and some nonnegative
constant κ06 ,
ksk∞ Dm
w(n̄)
E kš?m − ŝ?m k2Ψ ≤ κ06
+
, for all m ⊂ m• ,
n̄
n̄
and
q
q
1
w(n̄)
? − ŝ? k4 ,
? − ŝ? k4
max
E
kš
p
E
kš
≤
.
σ
m•
m• Ψ
m•
m• Ψ
(d+1)/2
n̄
n̄(log(n̄)/d)
12
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
Assumption (Conc) describes how the random measure M concentrates around the
measure to
estimate. Assumption (Var) ensures that we can estimate the variance terms E ks?m − š?m k2Ψ
c is to M.
over each m ∈ MP . Last, Assumption (Rem) describes how close M
Theorem 1. Assume that (11), Assumptions (Conc), (Var),(Rem) are satisfied, and that
max(ksk∞ , 1) ≤ R̄. Choose L• such that
n̄
2L• ≤
2d
((log n̄)/d)
and a penalty of the form
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
, m ∈ MP .
n̄
λ∈m
If c1 , c2 are positive and large enough, then
max ksk2Ψ , ksk∞ , 1
R̄Dm
−3(d+1)/2
P 2
? 2
1 + (log(n̄)/d)
+ w(n̄)
+C2
E ks − s̃ kΨ ≤ C1 min ks − sm kΨ +
n̄
n̄
m∈MP
where C1 may depend on κ01 , κ02 , κ04 , κ05 , κ06 , c1 , c2 and C2 may depend κ01 , κ02 , κ03 , κ07 , j0 , d.
In practice, the penalty constants c1 and c2 are calibrated by simulation study. We may also
replace R̄ in the penalty by max{kŝ?m• k∞ , 1}, and extend Theorem 1 to a random R̄ by using
arguments similar to [AL11].
?
4.3. Back to the examples. First, two general remarks are in order. For t ∈ Sm
, let ft =
•
P
?
λ∈m• ht, Ψλ iΨλ , then kft k = ktkΨ and by (3), for all countable subfamily T of Sm• ,
Z
Z
Z(T ) = sup
ft dM − E
ft dM
.
t∈T
Q
Q
So Assumption (Conc) usually proceeds from a Talagrand type concentration inequality. Besides,
we have seen in Section 3.5 that in general
ksk∞
.
λ∈m•
n̄
Thus, whenever some upper-bound R∞ for ksk∞ is known, Assumption (Var) is satisfied with
σ̂λ2 = R∞ for all λ ∈ m• , Ωσ = Ω, w(n̄) = 0, κ04 = κ05 = 1. One may also estimate each variance
term: this is what we propose in the following results, proved in Section 7.5.
max Var(β̌λ ) ≤
Corollary 1. In the density estimation framework (see 2.2.1), let R̄ ≥ max(ksk∞ , 1), 2L• =
−2d
n((log n)/d) ,
n
σ̂λ2 =
i−1
XX
1
2
(Ψλ (Yi ) − Ψλ (Yj )) for all λ ∈ m• ,
n(n − 1) i=2 j=1
and
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
, for all m ∈ MP .
n
λ∈m
If c1 , c2 are positive and large enough, then
max ksk2Ψ , R̄
R̄Dm
E ks − s̃P k2Ψ ≤ C1 min ks − s?m k2Ψ +
+ C2
n
n
m∈MP
where C1 may depend on κ, d, c1 , c2 and C2 may depend κ, j0 , d.
Corollary 2. In the copula density estimation framework (see 2.2.2), let R̄ ≥ max(ksk∞ , 1) and
−2d
2L• = min{n1/8 (log n)−1/4 , n((log n)/d) }. For all λ ∈ m• , define
σ̂λ2 =
n X
i−1
2
X
1
Ψλ F̂n1 (Xi1 ), . . . , F̂nd (Xid ) − Ψλ F̂n1 (Xj1 ), . . . , F̂nd (Xjd )
,
n(n − 1) i=2 j=1
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
13
and for all m ∈ MP , let
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
.
n
λ∈m
Under Assumption (L), and if c1 , c2 are positive and large enough, then
(log n)d−1
R̄Dm
P 2
? 2
√
+ C2 max ksk2Ψ , R̄
E ks − s̃ kΨ ≤ C1 min ks − sm kΨ +
n
n
m∈MP
where C1 may depend on κ, d, c1 , c2 and C2 may depend κ, j0 , d.
Corollary 3. In the Poisson intensity estimation framework (see 2.2.3), let R̄ ≥ max(ksk∞ , 1),
−2d
2L• = Vold (Q)((log Vold (Q))/d) ,
Z
1
2
Ψ2 dN, for all λ ∈ m• ,
σ̂λ =
Vold (Q) Q λ
and
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
, for all m ∈ MP .
Vold (Q)
λ∈m
If c1 , c2 are positive and large enough, then
max ksk2Ψ , R̄
R̄Dm
P 2
? 2
E ks − s̃ kΨ ≤ C1 min ks − sm kΨ +
+ C2
Vold (Q)
Vold (Q)
m∈MP
where C1 may depend on κ, d, c1 , c2 and C2 may depend κ, j0 , d.
Corollary 4. In the Lévy jump intensity estimation framework with continuous time observations
−2d
(see 2.2.4), let R̄ ≥ max(ksk∞ , 1), 2L• = T ((log T )/d) ,
ZZ
1
σ̂λ2 =
Ψ2λ (x)N (dt, dx), for all λ ∈ m• ,
T
[0,T ]×Q
and
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
, for all m ∈ MP .
T
λ∈m
If c1 , c2 are positive and large enough, then
max ksk2Ψ , R̄
R̄Dm
P 2
? 2
E ks − s̃ kΨ ≤ C1 min ks − sm kΨ +
+ C2
T
T
m∈MP
where C1 may depend on κ, d, c1 , c2 and C2 may depend κ, j0 , d.
Corollary 5. In the Lévy jump intensity estimation
framework with discrete
o time observations
n
−2d
(see 2.2.5), let R̄ ≥ max(ksk∞ , 1), 2L• = min (n∆)1/4 , n∆(log(n∆)/d)
,
n
σ̂λ2 =
1 X 2
Ψλ Xi∆ − X(i−1)∆ , for all λ ∈ m• ,
n∆ i=1
and
pen(m) =
X c1 σ̂ 2 + c2 R̄
λ
, for all m ∈ MP .
n∆
λ∈m
2
If Assumption (L) is satisfied, if n∆ stays bounded while n∆ → ∞ as n → ∞, and if c1 , c2 are
positive and large enough, then
R̄Dm
logd−1 (n∆)
E ks − s̃P k2Ψ ≤ C1 min ks − s?m k2Ψ +
+ C2 max ksk2Ψ , R̄
T
n∆
m∈MP
where C1 may depend on κ, d, c1 , c2 and C2 may depend κ, j0 , d, Q, f.
Corollaries 3, 4, 5 extend respectively the works of [RB03, FLH09, UK11] to a multivariate framework, with a complex family of models allowing for nonhomogeneous smoothness, and a more
refined penalty.
14
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
5. Adaptivity to mixed smoothness
There remains to compare the performance of our procedure s̃P to that of other estimators.
For that purpose, we derive the estimation rate of s̃P under smoothness assumptions that induce
sparsity on the hyperbolic wavelet coefficients of s. We then compare it to the minimax rate.
5.1. Function spaces with dominating mixed smoothness. For α ∈ N? and 1 ≤ p ≤ ∞, the
mixed Sobolev space with smoothness α measured in the Lp −norm is defined as
r1 +...+rd
X
∂
f
α
α
<
∞
,
= f ∈ Lp ([0, 1]d ) kf kSWp,(d)
:=
SWp,(d)
∂ r1 x 1 . . . ∂ rd x d p
0≤r1 ,...,rd ≤α
while the classical Sobolev space is
α
α
:=
Wp,(d)
= f ∈ Lp ([0, 1]d ) kf kWp,(d)
X
0≤r1 +...+rd ≤α
r1 +...+rd
∂
f
∂ r1 x 1 . . . ∂ rd x d
<∞
p
.
The former contains functions whose highest order derivative is the mixed derivative ∂ dα f /∂ α x1 . . . ∂ α xd ,
while the latter contains all derivatives up to global order dα. Both spaces coincide in dimension
d = 1, and otherwise we have the obvious continuous embeddings
(12)
dα
α
α
Wp,(d)
,→ SWp,(d)
,→ Wp,(d)
.
Hölder and Besov spaces with mixed dominating smoothness may be defined thanks to mixed
differences. For f : [0, 1] → R, x ∈ [0, 1] and h > 0,
∆0h (f, x) = f (x),
∆1h (f, x) = f (x + h) − f (x)
and more generally, for r ∈ N? , the r-th order univariate difference operator is
∆rh = ∆1h ◦ ∆hr−1 ,
so that
(13)
∆rh (f, x)
=
r
X
r
k=0
k
(−1)r−k f (x + kh).
Then for t > 0 the univariate modulus of continuity of order r in Lp is defined as
wr (f, t)p = sup k∆rh (f, .)kp .
0<h<t
For f : [0, 1]d → R, x = (x1 , . . . , xd ) ∈ [0, 1]d , r ∈ N? and h` > 0, we denote by ∆rh` ,` the univariate
difference operator applied to the `-th coordinate while keeping the other ones fixed, so that
r
X
r
r
∆h` ,` (f, x) =
(−1)r−k f (x1 , . . . , x` + kh` , . . . , xd ).
k
k=0
For any subset e of {1, . . . , d} and h = (h1 , . . . , hd ) ∈ (0, +∞)d , the r-th order mixed difference
operator is given by
Y
∆r,e
∆rh` ,` .
h :=
`∈e
For t = (t1 , . . . , td ) ∈ (0, +∞)d , we set te = (t` )`∈e , and define the mixed modulus of continuity
wre (f, te )p =
sup
0<h` <t` ,`∈e
k∆r,e
h (f, .)kp .
α
For α > 0 and 0 < p ≤ ∞, the mixed Hölder space SHp,(d)
is the space of all functions f : [0, 1]d →
R such that
X
Y
e
α
kf kSHp,(d)
:=
t−α
sup
` wbαc+1 (f, te )p
e⊂{1,...,d}
t>0
`∈e
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
15
is finite, where by convention the term associated with e = ∅ is kf kp . More generally, for α > 0
α
and 0 < p, q ≤ ∞, the mixed Besov space SBp,q,(d)
is the space of all functions f : [0, 1]d → R
such that
!1/q
!q
Z
Z
X
Y
Y dt`
−α
e
α
kf kSBp,q,(d)
:=
,
...
t` wbαc+1 (f, te )p
t`
(0,1)
(0,1)
`∈e
e⊂{1,...,d}
`∈e
α
α
where the Lq -norm is replaced by a sup-norm in case q = ∞, so that SBp,∞,(d)
= SHp,(d)
. By
α
comparison, the usual Besov space Bp,q,(d) may be defined as the space of all functions f ∈
Lp ([0, 1]d ) such that
1/q
q
Pd R
{`}
`
if 0 < q < ∞
kf kp + `=1 (0,1) t−α
wbαc+1 (f, t` )p dt
`
t`
α
kf kBp,q,(d)
:=
kf k + Pd sup
−α {`}
t w
(f, t )
if q = ∞
p
`=1
t` >0 `
bαc+1
` p
is finite. Extending (12), the recent results of [NS16a] confirm that the continuous embeddings
dα
α
α
Bp,q,(d)
,→ SBp,q,(d)
,→ Bp,q,(d)
,
hold under fairly general assumptions on α, p, q, d..
On the other hand, given α > 0, 0 < p < ∞, 0 < q ≤ ∞, we define
P
q/p 1/q
P
P
q`(α+1/2−1/p)
p
2
|hf,
Ψ
i|
λ
j∈J`
`≥dj0
λ∈∇j
NΨ,α,p,q (f ) =
P
1/p
p
sup`≥dj 2`(α+1/2−1/p) supj∈J
|hf,
Ψ
i|
λ
λ∈∇j
0
`
if 0 < q < ∞
if q = ∞
and NΨ,α,∞,q in the same way by replacing the `p -norm with a sup-norm. Then for α > 0, 0 <
p, q ≤ ∞, R > 0, we denote by SB(α, p, q, R) the set of all functions f ∈ Lp ([0, 1]d ) such that
NΨ,α,p,q (f ) ≤ R.
Under appropriate conditions on the smoothness of Ψ? , that we will assume to be satisfied in
the sequel, the sets SB(α, p, q, R) may be interpreted as balls with radius R in Besov spaces with
α
dominating mixed smoothness SBp,q,(d)
(see for instance [ST87, Hoc02a, Hep04, DTU16]). Mixed
Sobolev spaces are not easily characterized in terms of wavelet coefficients, but they satisfy the
compact embeddings
α
α
α
SBp,min(p,2),(d)
,→ SWp,(d)
,→ SBp,max(p,2),(d)
, for 1 < p < ∞
and
α
α
α
SB1,1,(d)
,→ SW1,(d)
,→ SB1,∞,(d)
(see [DTU16], Section 3.3). So, without loss of generality, we shall mostly turn our attention to
Besov-Hölder spaces in the sequel.
5.2. Link with structural assumptions. The following property collects examples of composite
functions with mixed dominating smoothness built from lower dimensional functions with classical
Sobolev or Besov smoothness. The proof and upper-bounds for the norms of the composite
functions are given in Section 7.6. An analogous property for (mixed) Sobolev smoothness instead
of (mixed) Besov smoothness can be proved straightforwardly.
Proposition 5. Let α > 0 and 0 < p, q ≤ ∞.
Pd
α
α
, then f (x) = `=1 u` (x` ) ∈ SBp,q,(d)
.
(i) If u1 , . . . , ud ∈ Bp,q,(1)
αI
(ii) Let P be some partition of {1, . . . , d}. If, for all I ∈ P, uI ∈ Bp,q,(|I|)
, then f (x) =
Q
ᾱ
u
(x
)
∈
SB
where
ᾱ
=
min
(α
/|I|).
I
I
I∈P
I
I∈P
p,q,(d)
dα
α
(iii) Let α ∈ N? and p > 1, if g ∈ W∞,(1)
and u` ∈ Wp,1
for ` = 1, . . . , d, then f (x) =
P
d
α
g
`=1 u` (x` ) ∈ SWp,(d) .
α
(iv) If f ∈ SBp,q,(d)
with α > 1 and ∂ d f /∂x1 . . . ∂xd ∈ Lp ([0, 1]d ), then ∂ d f /∂x1 . . . ∂xd ∈
α−1
SBp,q,(d)
.
16
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
α
(v) If f1 and f2 ∈ SBp,p,(d)
where either 1 < p ≤ ∞ and α > 1/p, or p = 1 and α ≥ 1, then the
α
product function x 7→ f1 (x)f2 (x) ∈ SBp,p,(d)
.
Notice that in (i) (resp. (ii), (iii)), the assumptions on the component functions u` , uI or g are
dα
dᾱ
dα
not enough to ensure that f ∈ Bp,q,(d)
(resp. Bp,q,(d)
, Wp,(d)
).
Remark: We believe that a generalization of (iii) to Besov or fractional Sobolev smoothness holds.
Yet such a generalization would require refined arguments from Approximation Theory in the spirit
of [BS11, Mou11] which are beyond the scope of that paper.
The structural assumption (ii) may be satisfied in the multivariate density estimation framework 2.2.1 whenever Y1 = (Y11 , . . . , Y1d ) can be split into independent sub-groups of coordinates,
and has recently been considered in [Lep13, Reb15b, Reb15a]. Case (i) and its generalization
(iii) may not be directly of use in our multivariate intensity framework, but they will allow to
draw a comparison with [HM07, BB14]. Combining (iii) and (iv) is of interest for copula density
estimation 2.2.2, having in mind that a wide nonparametric family of copulas are Archimedean
copulas (see [Nel06], Chapter 4), which have densities of the form
s(x1 , . . . , xd ) = (φ−1 )0 (x1 ) . . . (φ−1 )0 (xd )φ(d) φ−1 (x1 ) + . . . + φ−1 (xd )
provided the generator φ is smooth enough (see for instance [MN09]). Combining (iii), (iv), (v)
may be of interest for Lévy intensity estimation in 2.2.4 or 2.2.5. Indeed, a popular way to build
multivariate Lévy intensities is based on Lévy copulas studied in [KT06] (see also [CT04], Chapter
5). The resulting Lévy intensities then have the form
f (x1 , . . . , xd ) = f1 (x1 ) . . . fd (xd )F (1,...,1) (U1 (x1 ) + . . . + Ud (xd ))
R∞
where F is a so-called Lévy copula, F (1,...,1) = ∂ d F/∂t1 . . . ∂td and U` (x` ) = x` f` (t)dt. Besides,
a common form for F is
F (x) = φ φ−1 (x1 ) + . . . + φ−1 (xd )
under appropriate smoothness assumptions on φ. Last, let us emphasize that any linear combinaα
tion (mixtures for instance) of functions in SBp,q,(d)
inherits the same smoothness. Consequently,
mixed dominating smoothness may be thought as a fully nonparametric surrogate for a wide range
of structural assumptions.
5.3. Approximation qualities and minimax rate. We provide in Section 7.7 a constructive
proof for the following nonlinear approximation result, in the spirit of [BBM99].
Theorem 2. Let R > 0, 0 < p < ∞, 0 < q ≤ ∞, α > max(1/p − 1/2, 0), and f ∈ L2 ([0, 1]d ) ∩
SB(α, p, q, R). Under (11), for all `1 ∈ {dj0 +1, . . . , L• +1}, there exists some model m`1 (f ) ∈ MP
`1
?
and some approximation A(f, `1 ) ∈ Sm
for f such that
` (f )
1
kf − A(f, `1 )k2Ψ
2(d−1)(1/2−1/q)+ −2L• (α−(1/p−1/2)+ )
2(d−1)(1/2−1/ max(p,q)) −2α`1
≤ C(B, j0 , α, p, d)R2 L•
2
+ `1
2
.
Remark: When p ≥ 2, the same kind of result still holds with all N (`1 , k) = 0. But Assumption (11)
is really useful when p < 2, the so-called non-homogeneous smoothness case.
The first term in the upper-bound is a linear approximation error by the highest dimensional model
?
Sm
in the collection. As Dm• is of order Ld−1
2L• , we deduce from [DTU16] (Section 4.3) that this
•
•
α
first term is optimal over SBp,q , at least for 1 < p < ∞, 1 ≤ q ≤ ∞ and α > max(1/p − 1/2, 0), for
instance. The second term in the upper-bound is a nonlinear approximation error of f within the
?
model Sm
, with dimension Dm`1 (f ) of order `1d−1 2`1 . So we deduce from [DTU16] (Theorem
` (f )
1
−2α
7.6) that this second term, which is of order Dm
(log Dm`1 (f ) )2(d−1)(α+1/2−1/q) , is also optimal
`1 (f )
α
up to a constant factor over SBp,q , at least for 1 < p < ∞, p ≤ q ≤ ∞ and α > max(1/p − 1/2, 0).
α
Notice that, under the classical Besov smoothness assumption f ∈ L2 ([0, 1]d ) ∩ Bp,q,(d)
, the best
possible approximation rate for f by D-dimensional linear subspaces in the L2 -norm would be
of order D−2α/d . Thus with a mixed smoothness of order α in dimension d, we recover the same
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
17
approximation rate as with a classical smoothness of order dα in dimension d, up to a logarithmic
factor.
Let us define, for α, p, q, R, R0 > 0,
SB(α, p, q, R, R0 ) = {f ∈ SB(α, p, q, R)/kf k∞ ≤ R0 }.
In the sequel, we use the notation a C(θ)b when there exist positive reals C1 (θ), C2 (θ) such that
C1 (θ)b ≤ a ≤ C2 (θ)b.
Corollary 6. Assume L• is large enough, then for all 0 < p < ∞, 0 < q ≤ ∞, α > (1/p −
1/2)+ , R ≥ n̄−1 , R0 > 0,
2/(1+2α)
(d−1)(α+1/2−1/ max(p,q))
Es ks − s̃P k2Ψ ≤ C(B, d, α, p, R0 ) log(n̄R2 )
sup
Rn̄−α
.
s∈SB(α,p,q,R,R0 )
Proof. In order to minimize approximately the upper-bound, we choose `1 such that
2(d−1)(1/2−1/ max(p,q)) −2α`1
`1
2
R2 C(α, p, q, d)`1d−1 2`1 /n̄,
that is for instance
2`1 C(α, p, q, d)
1/(1+2α)
−2(d−1)/ max(p,q)
,
log(n̄R2 )
(n̄R2 )
which yields the announced upper-bound.
Remember that a similar result holds when replacing the Ψ-norm by the equivalent L2 -norm.
Though unusual, the upper-bound in Corollary 6 is indeed related to the minimax rate.
Proposition 6. In the density estimation framework, assume R2 ≥ n−1 , R0 > 0, p > 0, 0 < q ≤
∞, and either α > (1/p − 1/2)+ and q ≥ 2 or α > (1/p − 1/2)+ + 1/ min(p, q, 2) − 1/ min(p, 2),
then
2/(1+2α)
(d−1)(α+1/2−1/q)
inf
sup
Rn−α
Es ks − ŝk2 C(α, p, q, d) log(nR2 )
.
ŝ estimator of s
s∈SB(α,p,q,R,R0 )
Proof. One may derive from [DTU16] (Theorem 6.20), [Dun01] (proof of Theorem 1) and the link
between entropy number and Kolmogorov entropy that the Kolmogorov -entropy of SB(α, p, q, R)
is
(d−1)(α+1/2−1/q)/α
H (α, p, q, R) = (R/)1/α (log(R/))
.
According to [YB99] (Proposition 1), in the density estimation framework, the minimax risk over
SB(α, p, q, R, R0 ) is of order ρ2n where ρ2n = Hρn (α, p, q, R)/n, which yields the announced rate.
Consequently, in the density estimation framework, the penalized pyramid selection procedure is
minimax over SB(α, p, q, R) up to a constant factor if p ≤ q ≤ ∞, and only up to a logarithmic
factor otherwise.
Let us end with some comments about these estimation rates. First, we remind that the minα
imax rate under the assumption s ∈ Bp,q,(d)
is of order n−2α/d/(1+2α/d) . Thus, under a mixed
smoothness assumption of order α, we recover, up to a logarithmic factor, the same rate as with
smoothness of order α in dimension 1, which can only be obtained with smoothness of order dα
under a classical smoothness assumption in dimension d. Besides, under the multiplicative constraint (ii) of Proposition 5, we recover the same rate as [Reb15a], up to a logarithmic factor. And
under the generalized additive constraint (iii) of Proposition 5, we recover the same rate as [BB14]
(Section 4.3), up to a logarithmic factor. Regarding Neumann seminal work on estimation under
mixed smoothness [Neu00] (see his Section 3), a first adaptive wavelet thresholding is proved to
r
r
be optimal up to a logarithmic factor over SW2,(d)
= SB2,2,(d)
, and another, nonadaptive one, is
r
proved to be optimal up to a constant over SB1,∞,(d) , where r is a positive integer. Our procedure thus outperforms [Neu00] by being at the same time adaptive and minimax optimal up to a
constant over these two classes, and many other ones.
18
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
6. Implementing wavelet pyramid selection
We end this paper with a quick overview of practical issues related to wavelet pyramid selection.
As we perform selection within a large collection of models, where typically the number of models
is exponential in the sample size, we must guarantee that the estimator can still be computed in
a reasonable time. Besides, we provide simulation based examples illustrating the interest of this
new method.
6.1. Algorithm and computational complexity. Theorem 1 supports the choice of an additive
penalty of the form
X
pen(m) =
v̂λ2 ,
λ∈m
where detailed expressions for v̂λ2 in several statistical frameworks have been given in Section 4.3.
P
As γ̂(ŝ?m ) = − λ∈m β̂λ2 , the penalized selection procedure amounts to choose
m̂P = argmax crit(m)
m∈MP
where
X
crit(m) =
(β̂λ2 − v̂λ2 ).
λ∈m
v̂λ2
Since each
is roughly an (over)estimate for the variance of β̂λ2 , our method, though different
from a thresholding procedure, will mainly retain empirical wavelet coefficients β̂λ2 which are
significantly larger than their variance.
A remarkable thing is that, due to both the structure of the collection of models and of the
penalty function, the penalized estimator can be determined without computing all the preliminary
estimators (ŝ?m )m∈MP , which makes the computation of s̃P feasible in practice. Indeed, we can
proceed as follows.
Step 1. For each `1 ∈ {dj0 + 1, . . . , L• + 1}, determine
X
m̂`1 = argmax
(β̂λ2 − v̂λ2 ).
m∈MP
`1 λ∈m
For that purpose, it is enough, for each k ∈ {0, . . . , L• − `1 }, to
• compute and sort in decreasing order all the coefficients (β̂λ2 − v̂λ2 )λ∈U ∇(`1 +k) ;
• keep the N (`1 , k) indices in U ∇(`1 + k) that yield the N (`1 , k) greatest such coefficients.
Step 2. Determine the integer `ˆ ∈ {dj0 + 1, . . . , L• + 1} such that
m̂`ˆ =
argmax
crit(m̂`1 ).
dj0 +1≤`1 ≤L• +1
The global computational complexity of s̃P is thus O(log(L• )Ld• 2L• ). Typically, we will choose
L• at most of order log2 (n̄) so the resulting computational complexity will be at most of order
O(log(log(n̄)) logd (n̄)n̄).
6.2. Illustrative examples. In this section, we study two examples in dimension d = 2 by using
Haar wavelets.
First, in the density estimation framework, we consider an example where the coordinates of
Yi = (Yi1 , Yi2 ) are independent conditionally on a K-way categorical variable Z, so that the density
of Yi may be written as
K
X
s(x1 , x2 ) =
πk s1,k (x1 )s2,k (x2 ),
k=1
where π = (π1 , . . . , πK ) is the probability vector characterizing the distribution of Z. For a compact
interval I, and a, b > 0, let us denote by β(I; a, b) the Beta density with parameters a, b shifted
and rescaled to have support I, and by U(I) the uniform density on I. In our example, we take
• K = 4 and π = (3/5, 1/10, 1/40, 11/40) ;
• s1,1 = β ([0, 3/5]; 4, 4) and s2,1 = β ([0, 2/5]; 4, 4) ;
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
19
• s1,2 = β ([2/5, 1]; 100, 100) and s2,2 = β ([2/5, 1]; 20, 20) ;
• s1,3 = U ([0, 1]) and s2,3 = U ([0, 1]) ;
• s1,4 = β ([3/5, 1]; 8, 4) and s2,4 = U ([2/5, 1]) .
The resulting mixture density s of Yi is shown in Figure 1 (b). We choose L• = n/((log n)/2)2
and first compute the least-squares estimator ŝ?• of s on the model VL?• /2 ⊗ . . . ⊗ VL?• /2 , which
provides the estimator R̂ = max{kŝ?• k, 1} for R̄. We then use the penalty
pen(m) =
X 1.5σ̂ 2 + 0.5R̂
λ
.
n
λ∈m
For a sample with size n = 2000, Figure 1 illustrates how the procedure first selects a rough model
m̂`ˆ (Figure 1 (c)) and then add some details wherever needed (Figure 1 (d)). Summing up the two
yields the pyramid selection estimator s̃P (Figure 1 (e)). By way of comparison, we also represent
in Figure 1 (f) a widely used estimator: the bivariate Gaussian kernel estimator, with the "known
support" option, implemented in MATLAB ksdensity function. We observe that, contrary to the
kernel density estimator, the pyramid selection estimator recovers indeed the main three modes,
and in particular the sharp peak.
Figure 1. Pyramid selection and standard kernel for an example of mixture of
multiplicative densities.
In the copula density estimation framework, we consider an example where the copula of Xi =
(Xi1 , Xi2 ) is either a Frank copula or a Clayton copula conditionally to a binary variable Z. More
precisely, we consider the mixture copula
s(x1 , x2 ) = 0.5sF (x1 , x2 ) + 0.5sC (x1 , x2 )
where sF is the density of a Frank copula with parameter 4 and sC is the density of a Clayton copula
with parameter 2. These two examples of Archimedean copula densities are shown in Figure 2
and the resulting mixture in Figure 3 (b). We use the same penalty as in the previous example,
adapted of course to the copula density estimation framework. We illustrate in Figure 3 the
pyramid selection procedure on a sample with size n = 2000. Though not all theoretical conditions
are fully satisfied here, the pyramid selection procedure still provides a reliable estimator.
As a conclusion, those examples suggest that the Haar pyramid selection already provides a
useful new estimation procedure. This is most encouraging for pyramid selection based on higher
order wavelets, whose full calibration based on an extensive simulation study in each framework
will be the subject of another work.
20
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
Figure 2. Left: Frank copula density with parameter 4; Right: Clayton copula
density with parameter 2.
Figure 3. Pyramid selection for an example of mixture copula density.
7. Proofs
We shall use repeatedly the classical inequality
(14)
1
2ab ≤ θa2 + b2
θ
for all positive θ, a, b.
7.1. Proof of Proposition 2. We only have to prove 2 for m• . Indeed, as any pyramidal model
m is a subset of m• , a common upper-bound for the residual terms is
kš?m − ŝ?m k2 ≤ kš?m• − ŝ?m• k2 .
Under Assumption (L), and thanks to assumptions S.iii) and W.v), we have that for all λ ∈ m• ,
x = (x1 , . . . , xd ) and y = (y1 , . . . , yd ) ∈ Q,
|Ψλ (x) − Ψλ (y)| ≤ κd 23L• /2
d
X
|xk − yk |.
k=1
According to Massart’s version of Dworetzky-Kiefer-Wolfowitz inequality (see [Mas90]), for any
√
positive z, and 1 ≤ k ≤ d, there exists some event Ωk (z) on which kF̂nk − Fk k∞ ≤ z/ n and such
c
2
d
that P(Ωk (z)) ≤ 2 exp(−2z ). Setting Ω(z) = ∩k=1 Ωk (z), we thus have for all λ ∈ m•
√
|β̌λ − βbλ | ≤ κd 23L• /2 d(z/ n)1IΩ(z) + κd 23L• /2 d1IΩc (z) ,
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
21
hence
E[kš?m• − ŝ?m• k2 ] ≤ κ2d 23L• d2 (z 2 /n + 2d exp(−2z 2 ))Dm• .
Finally, Dm• is of order Ld−1
2L• (see Proposition 4), so by choosing 2z 2 = log(n),
•
E[kš?m• − ŝ?m• k2 ] ≤ C(κ, d)L•d−1 24L• log(n)/n.
7.2. Proof of RProposition 3. For all bounded measurable function g, let us denote D∆ (g) =
E[g(X∆ )]/∆ − Q gs. For all λ ∈ Λ,
2
D∆ (Ψ2λ )
ksk∞
2
2
(15) E β̌λ − βbλ
≤ 4 Var(β̌λ ) + Var(βbλ ) + D∆
+4
+ 4D∆
(Ψλ ) ≤ 8
(Ψλ ).
n∆
n∆
We shall bound D∆ (Ψλ ) by using the decomposition of a Lévy process into a big jump compound
Poisson process and an independent small jump Lévy process. Let us fix ε > 0 small enough so
Qd
that Q = k=1 [ak , bk ] ⊂ {kxk > ε} and denote by (Σ, µ, ν) the characteristic Lévy triplet of
X = (Xt )t≥0 , where µ stands for the drift and ν is the Lévy measure, with density f with respect
to the Lebesgue measure on Rd (see Section 2.2). Then X is distributed as Xε + X̃ε , where Xε
and X̃ε are independent Lévy processes with following characteristics. First, Xε is a Lévy process
with characteristic Lévy triplet (Σ, µε , ν ), where the drift is
Z
µε = µ −
x f (x)dx.
ε<kxk≤1
and the Lévy measure is
νε (dx) = 1Ikxk≤ε f (x)dx.
ε
The process X̃ is the compound Poisson process
X̃tε
=
Ñt
X
ξi ,
i=1
where Ñ is a homogeneous Poisson process with intensity λε = ν({kxk > ε}), (ξi )i≥1 are i.i.d.
with density λ−1
ε 1Ikxk>ε f (x), and Ñ and (ξi )i≥1 are independent.
Conditioning by Ñ and using the aforementioned independence properties yields
"
!#
j+2
∞
ε
X
X
(λε ∆)j
E[Ψλ (X∆ )]
−λε ∆ E[Ψλ (X∆ )]
−λε ∆
ε
2
−λε ∆
ε
=e
+λε e
E[Ψλ (X∆ +ξ1 )]+λε ∆e
E Ψλ X∆ +
ξi
.
∆
∆
(j + 2)!
j=0
i=1
ε
and ξ1 then yields
Conditioning by ξ1 and using independence between X∆
Z
ε
ε
λε E[Ψλ (X∆
+ ξ1 )] =
E[Ψλ (X∆
+ x)]f (x)dx.
kxk>ε
Writing hΨλ , si = e
−λε ∆
hΨλ , si + (1 − e
(1)
−λε ∆
)hΨλ , si and using (1 − e−λε ∆ ) ≤ λε ∆ leads to
(2)
(3)
(4)
|D∆ (Ψλ )| ≤ R∆ (Ψλ ) + R∆ (Ψλ ) + R∆ (Ψλ ) + R∆ (Ψλ ),
where
(1)
R∆ (Ψλ ) = e−λε ∆
(16)
ε
E[Ψλ (X∆
)]
,
∆
(3)
(2)
R∆ (Ψλ ) = e−λε ∆
R∆ (Ψλ ) = λε ∆kΨλ k1 ksk∞ ,
Z
ε
|E [Ψλ (X∆
+ x) − Ψλ (x)]| f (x)dx,
kxk>ε
(4)
R∆ (Ψλ ) = λ2ε ∆kΨλ k∞ .
As Ψλ has compact support Q,
ε
P(X∆
∈ Q)
.
∆
ε
ε
Let us denote by X∆,k
the k-th coordinate of X∆
and by dQ the maximal distance from [ak , bk ] to
0, for k = 1, . . . , d, reached for instance at k = k0 . We deduce from the proof of Lemma 2 in [RW02]
(see also [FLH09], equation (3.3)) that there exists z0 = z0 (ε) such that if ∆ < dQ /z0 (ε),
(1)
R∆ (Ψλ ) ≤ e−λε ∆ kΨλ k∞
ε
ε
P(X∆
∈ Q) ≤ P(|X∆,k
| ≥ dQ ) ≤ exp ((z0 log(z0 ) + u − u log(u))/(2ε)) ∆dQ /(2ε)
0
22
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
so that
(1)
R∆ (Ψλ ) ≤ C(dQ , ε)e−λε ∆ kΨλ k∞ ∆dQ /(2ε)−1 .
(17)
Under Assumption (L), Ψλ is Lipschitz on Q, so
ε
ε
.
+kΨλ k∞ 1I{X∆ε +x∈Q}∩{x∈Q}
|Ψλ (X∆
+ x) − Ψλ (x)| ≤ kΨλ kL kX∆
k1 1I{X∆ε +x∈Q}∩{x∈Q} +|Ψλ (x)|1I{X∆ε +x∈Q}∩{x∈Q}
/
/
Besides, as Q is compact and bounded away from the origin, there exists δQ and ρQ > 0 such that
ε
ε
{X∆
+ x ∈ Q} ∩ {x ∈ Q} ⊂ {kX∆
k ≥ δQ }
ε
ε
ε
({X∆
+ x ∈ Q} ∩ {x ∈
/ Q}) ∪ ({X∆
+x∈
/ Q} ∩ {x ∈ Q}) ⊂ {kX∆
k ≥ ρQ }
ε
The Lévy measure of X is compactly supported and satisfies
Z
Z
kxk2 νε (dx) =
kxk2 ν(dx)
kxk≤ε
which is finite since ν is a Lévy measure (see for instance [Sat99], Theorem 8.1). So we deduce
from [Mil71], Theorem 2.1, that
ε 2
E kX∆
k ≤ C(d, f )∆,
hence
ε
ε 2
−1
−1
E kX∆
k1 1IkX∆ε k≥δQ ≤ C(d)δQ
E kX∆
k ≤ C(d, f )δQ
∆
and from Markov inequality
ε
P(kX∆
k ≥ ρQ ) ≤ C(d, f )ρ−2
Q ∆.
Finally, fixing 0 < ε < min(dQ /4, inf x∈Q kxk), we have for all 0 < ∆ < min(dQ /z0 (), 1)
(18)
(2)
−1
−2
R∆ (Ψλ ) ≤ C(d, f )e−λε ∆ ∆(δQ
λε kΨλ kL + ρ−2
Q ksk∞ kΨλ k1 + ρQ λε kΨλ k∞ ).
For all λ ∈ m• ,
max(kΨλ kL , kΨλ k∞ , kΨλ k1 ) ≤ C(κ)23L• /2 , max(kΨ2λ kL , kΨ2λ k∞ , kΨ2λ k1 ) ≤ C(κ)22L• ,
so that combining (15), (17), (18) and (16) yields
E[kš?m − ŝ?m k2 ] ≤ 8
24L• n∆3 + 23L• ∆
ksk∞ Dm
+ C(κ, d, f, Q, ε)L•d−1
.
n∆
n∆
7.3. Proof of Proposition 4. Due to hypotheses S.ii) and W.ii), we have for all j ≥ j0 ,
2j−1 ≤ ]∇j ≤ M 2j−1 ,
hence, for all j ∈ Ndj0 ,
(1/2)d 2|j| ≤ ]∇j ≤ (M/2)d 2|j| .
Let us fix ` ∈ {dj0 , . . . , L• }. The number of d-uples j ∈ Ndj0 such that |j| = ` is equal to the
number of partititions of the integer ` − dj0 into d nonnegative integers, hence
d−1
Y
` − dj0 + d − 1
` − dj0
]J` =
=
1+
.
d−1
k
k=1
The last two displays and the classical upper-bound for binomial coefficient (see for instance [Mas07],
Proposition 2.5) yield
(19)
c0 (d)(` − dj0 + d − 1)d−1 2` ≤ ]U ∇(`) ≤ c1 (M, d)(` − dj0 + d − 1)d−1 2` ,
where c0 (d) = 2−d (d − 1)−(d−1) and c1 (M, d) = (M/2)d (e/(d − 1))d−1 .
Let us now fix `1 ∈ {dj0 + 1, . . . , L• + 1}. Any model m ∈ MP
`1 satisfies
Dm =
`X
1 −1
`=dj0
]U ∇(`) +
LX
• −`1
N (`1 , k).
k=0
So we obviously have
Dm ≥ ]U ∇(`1 − 1) ≥ κ1 (d)(`1 − dj0 + d − 2)d−1 2`1 ,
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
23
with κ1 (d) = c0 (d)/2 = 2−(d+1) (d − 1)−(d−1) . Besides, with our choice of N (`1 , k),
`X
1 −1
Dm ≤ c1 (M, d)(`1 − dj0 + d − 2)d−1
2` + 2M −d c1 (M, d)s1 (d)(`1 − dj0 + d − 2)d−1 2`1 ,
`=dj0
so that Proposition 4 holds with κ2 (d, j0 , B) = c1 (M, d)(1 + 2M −d c1 (M, d)s1 (d)), where
∞
X
(1 + k/(d − 1))d−1
s1 (d) =
.
(2 + k)d+2
k=0
MP
`1
The number of subsets of Λ in
satisfies
LY
N (`1 ,k)
LY
−`
• −`1
•
1
e ]U ∇(`1 + k)
]U ∇(`1 + k)
≤
]MP
=
.
`1
N (`1 , k)
N (`1 , k)
k=0
k=0
d+2 k
d
For k ∈ {0, . . . , L• − `1 }, let f (k) = (k + 2) 2 M /2, then N (`1 , k) ≤ ]U ∇(`1 + k)/f (k). As the
function x ∈ [0, U ] 7→ x log(eU/x) is increasing, we deduce
log(]MP
`1 ) ≤ D(`1 )
LX
• −`1
k=0
Setting
s2 =
∞
X
k=0
1
,
(k + 2)3
s3 =
]U ∇(`1 + k) 1 + log(f (k))
.
]U ∇(`1 − 1)
f (k)
∞
X
log(k + 2)
k=0
(k + 2)3
, s4 =
∞
X
k=0
1
,
(k + 2)2
one may take for instance κ3 (j0 , B, d) = (log(e/2) + d log(M ))s2 + (d + 2)s3 + log(2)s4 in Proposition 4.
7.4. Proof of Theorem 1.
7.4.1. Notation and preliminary results. Hyperbolic wavelet bases inherit from the underlying
univariate wavelet bases a localization property which can be stated as follows.
d−1
Lemma 1. Let D(L• ) = (e(L• − dj0 + d − 1)/(d − 1))
2L• /2 , then for all real-valued sequence
(aλ )λ∈m• ,
X
X
aλ Ψ?λ
≤ κ07 max |aλ |D(L• ),
aλ Ψλ
,
max
λ∈m•
λ∈m•
λ∈m•
∞
∞
√
where κ07 = κ2d (2 + 2) for instance.
Proof. For all x = (x1 , . . . , xd ) ∈ [0, 1]d , using assumptions S.vi), S.vii), S.viii), W.iv), W.v), W.vi)
in Section 3.1, we get
L• X Y
d
X
X
X
aλ Ψλ ≤ max |aλ |
|ψλk (xk )|
λ∈m•
λ∈m•
`=dj0 j∈J` k=1
≤ κ2d max |aλ |
λ∈m•
L• X
X
λk ∈∇jk
2`/2 .
`=dj0 j∈J`
d−1
We deduce from the proof of Proposition 4 the upper-bound ]J` ≤ (e(L• − dj0 + d − 1)/(d − 1))
which allows to conclude.
For all t ∈ L2 ([0, 1]d ), we define
X
ν(t) =
ht, Ψλ i(β̌λ − hs, Ψλ i),
νR (t) =
λ∈Λ
X
ht, Ψλ i(β̂λ − β̌λ ),
ν̂(t) = ν(t) + νR (t),
λ∈Λ
and for all m ∈ MP , we set
χ(m) =
sup
? |ktk =1
t∈Sm
Ψ
ν(t),
χR (m) =
sup
? |ktk =1
t∈Sm
Ψ
νR (t).
24
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
P
Lemma 2. For all m ∈ MP , let t?m = λ∈m (ν(Ψ?λ )/χ(m))Ψ?λ , then
sX
χ(m) =
ν 2 (Ψ?λ ) = ks?m − š?m kΨ = ν(t?m ),
λ∈m
χR (m) =
sX
2 (Ψ? ) = kš? − ŝ? k .
νR
m
m Ψ
λ
λ∈m
Proof. The proof follows from the linearity of ν and νR and Cauchy-Schwarz inequality.
Lemma 3. Let = κ02 ksk∞ /(κ03 κ07 D(L• )) and
ΩT = ∩λ∈m• {|ν(Ψ?λ )| ≤ } .
For all x > 0, there exists a measurable event Ωm (x) on which
X max{σ̂ 2 , 1}
x
0
λ
χ2 (m)1IΩT ∩Ωσ ≤ 2κ02
+ 8κ02 ksk∞ .
1 κ5
n̄
n̄
λ∈m
and such that
P(Ωcm (x))
≤ exp(−x).
?
|ktkΨ = 1}. Let us set z =
Proof.
We observe that χ(m) = Z(Tm ) where Tm = {t ∈ Sm
p
?
0
0
|ktkΨ = 1, maxλ∈m |ht, Ψλ i| ≤
κ2 ksk∞ x/n̄ and consider a countable and dense subset Tm of {t ∈ Sm
/z}. Thanks to the localization property in Lemma 1,
p
X
κ0 ksk∞
sup
ht, Ψλ iΨλ
≤ κ03 p2
.
0
x/n̄
t∈Tm
λ∈m•
∞
So Assumption (Conc) ensures that there exists Ωm (x) such that P(Ωcm (x)) ≤ exp(−x) and on
which
r
x
0
0
0
Z(Tm ) ≤ κ1 E [Z(Tm )] + 2 κ02 ksk∞ ,
n̄
hence
x
2
0
0
Z 2 (Tm0 ) ≤ 2κ02
1 E [Z(Tm )] + 8κ2 ksk∞ .
n̄
As Z(Tm0 ) ≤ χ(m), we obtain by convexity and Lemma 2
X
E2 [Z(Tm0 )] ≤ E χ2 (m) =
Var(β̌λ ).
λ∈m
t?m
On ΩT ∩ {χ(m) ≥ z},
given by Lemma 2 satisfies supλ∈m |ht?m , Ψλ i| ≤ /z, so that χ2 (m) =
2
0
Z (Tm ), while on ΩT ∩{χ(m) < z}, χ2 (m) < κ02 ksk∞ x/n̄. The proof then follows from Assumption
(Var).
7.4.2. Proof of Theorem 1. Let us fix m ∈ MP . From the definition of m̂P and of ŝ?m , we get
γ̂(s̃P ) + pen(m̂P ) ≤ γ̂(s?m ) + pen(m).
For all t, u ∈ L2 ([0, 1]d ),
γ̂(t) − γ̂(u) = kt − sk2Ψ − ku − sk2Ψ − 2ν̂(t − u),
so
ks − s̃P k2Ψ ≤ ks − s?m k2Ψ + 2ν̂(s̃P − s?m ) + pen(m) − pen(m̂P ).
Using the triangle inequality and Inequality (14) with θ = 1/4 and θ = 1, we get
2ν̂(s̃P − s?m ) ≤ 2ks̃P − s?m kΨ (χ(m ∪ m̂P ) + χR (m ∪ m̂P ))
1
1
≤ ks − s̃P k2Ψ + ks − s?m k2Ψ + 8χ2 (m ∪ m̂P ) + 8χ2R (m ∪ m̂P ),
2
2
hence
(20)
ks − s̃P k2Ψ ≤ 3ks − s?m k2Ψ + 16χ2 (m ∪ m̂P ) + 2(pen(m) − pen(m̂P )) + 16χ2R (m ∪ m̂P ).
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
25
Let us fix ζ > 0 and set ω = κ3 (j0 , B, d) + log(2) and Ω? (ζ) = ∩m0 ∈MP Ωm∪m0 (ζ + ωDm ). We
deduce from Lemma 3 that on Ω? (ζ)
X max{σ̂ 2 , 1}
ω(Dm + Dm̂P )
ζ
0
λ
+ 8κ02 ksk∞
+ 8κ02 ksk∞ .
(21) χ2 (m ∪ m̂P )1IΩT ∩Ωσ ≤ 2κ02
1 κ5
n̄
n̄
n̄
P
λ∈m∪m̂
Besides, given Proposition 4, our choice of ω leads to
LX
• +1
log(]MP
` )
P(Ωc? (ζ)) ≤ e−ζ
≤ e−ζ .
exp −D(`) ω −
D(`)
`=dj0 +1
Choosing for instance
pen(m) = c1
X σ̂ 2
R̄Dm
λ
+ c2
,
n̄
n̄
λ∈m
0
0
0
with c1 ≥ 16κ02
1 κ5 and c2 ≥ 64κ2 ω + 8κ6 and integrating with respect to ζ > 0, we deduce
from (20), (21), Assumption (Var) and Assumption (Conc) that
w(n̄)
R̄Dm
ksk∞
+ 64κ02
+8
,
E ks − s̃P k2Ψ 1IΩT ∩Ωσ ≤ 3ks − s?m k2Ψ + C
n̄
n̄
n̄
where C may depend on κ01 , κ02 , κ04 , κ05 , κ06 , c1 , c2 .
In order to bound E ks − s̃P k2Ψ 1IΩcT ∪Ωcσ , we first notice that from the triangle inequality and
Lemma 2
(22)
ks − s̃P kΨ ≤ ks − s?m̂P kΨ + ks?m̂P − ŝ?m̂P kΨ
≤ kskΨ + χ(m̂) + χR (m̂),
hence
ks − s̃P k2Ψ ≤ ksk2Ψ + 4χ2 (m• ) + 4χ2R (m• ).
Then setting pT = P(ΩcT ) and pσ = P(Ωcσ ), Cauchy-Schwarz inequality entails
q
p
√
E [χ4 (m• )] + E [χ4R (m• )] .
E ks − s̃P k2Ψ 1IΩcT ∪Ωcσ ≤ 2(pT + pσ )ksk2Ψ + 4 pT + pσ
Let λ ∈ m• , kΨ?λ k∞ ≤ κd 2L• /2 , so applying Assumption (Conc) with T = {Ψ?λ } and T = {−Ψ?λ },
we get
n̄2
n̄
?
P (|ν(Ψλ )| ≥ ) ≤ 2 exp − min
.
,
4κ02 ksk∞ 2κ03 κd 2L• /2
d−1
Then setting ι = (e(L• − dj0 + d − 1)/(d − 1))
pT ≤ 2ι2L• exp −Cksk∞
, Proposition 4 yields
C
n̄
,
≤ 2
ι2 2L•
n̄ (log(n̄)/d)d+1
where C may depend on κ02 , κ03 , κ07 , j0 , d. Besides, we deduce from Assumption (Conc) and Lemma 1
that, for all x > 0,
!
r
r
ksk∞ Dm•
x
x
0
0 0
0
P χ(m• ) ≥ κ1
+ κ2 ksk∞ + κ3 κ7 D(L• )
≤ exp(−x).
n̄
n̄
n̄
For a nonnegative random variable U, Fubini’s inequality implies
Z ∞
4
E[U ] =
4xp−1 P(U ≥ x)dx
0
so
4
E[χ (m• )] ≤ C max
where C may depend on
ι4 22L• ι2 22L•
,
n̄4
n̄2
κ01 , κ02 , κ03 , κ07 , j0 , d.
≤
C
2(d+1)
(log(n̄)/d)
Remembering (22) , we conclude that
R̄Dm
ksk∞
E ks − s̃P k2Ψ ≤ 3ks−s?m k2Ψ +C1
+C2
+C3 max{ksk2Ψ , 1}
n̄
n̄
1
3(d+1)/2
n̄ (log(n̄)/d)
w(n̄)
+
n̄
!
,
26
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
where C1 may depend on κ01 , κ02 , κ04 , κ05 , κ06 , c1 , c2 , C2 may depend on κ02 , C3 may depend κ01 , κ02 , κ03 , κ07 , j0 , d.
7.5. Proofs of Corollaries 1 to 5.
7.5.1. Proof of Corollary 1. Assumption (Conc) is a straightforward consequence of Talagrand’s
inequality, as stated for instance in [Mas07] (Inequality (5.50), and is satisfied, whatever θ > 0,
for
(23)
n̄ = n, κ01 = 1 + θ, κ02 = 2, κ03 = (1/3 + 1/θ)/2.
For all λ ∈ m• , σ̂λ2 is an unbiased estimator for Var(Ψλ (Y1 )). Besides, the existence of Ωσ follows
from Lemma 1 in [RBRTM11] with γ = 2. Thus Assumptions (Var) and (Rem) are satisfied by
taking κ04 = 1, κ05 that only depends on κ and d, κ06 = 0, w(n) = C(κ, j0 , d)/ logd+1 (n).
7.5.2. Proof of Corollary 2. Setting Yi = (F1 (Xi1 ), . . . , Fd (Xid )), i = 1, . . . , n, we recover the
previous density estimation framework, so Assumption (Conc) is still satisfied with (24). Setting
Ŷi = (Fn1 (Xi1 ), . . . , F̂nd (Xid )), i = 1, . . . , n, and
n
σ̌λ2
i−1
XX
1
2
=
(Ψλ (Yi ) − Ψλ (Yj )) ,
n(n − 1) i=2 j=1
we observe that, for all λ ∈ m•
max σ̂λ2 − 4σ̌λ2 , σ̌λ2 − 4σ̂λ2 ≤ 8Rλ (n)
where
Rλ (n) =
n
2
X
1
(i − 1) Ψλ (Ŷi ) − Ψλ (Yi ) .
n(n − 1) i=2
Using the same arguments as in the proof of Proposition 2, we get for all λ ∈ m• and all m ⊂ m•
E [Rλ (n)] ≤ C(κ, d)23L• log(n)/n,
Rλ (n) ≤ C(κ, d)23L• log(n)/n
except on a set with probability smaller than 2d/n,
E kš?m − ŝ?m k2 ≤ C(κ, d, j0 )L•d−1 24L• log(n)/n,
and
q
√
E kš?m• − ŝ?m• k4 ≤ C(κ, d, j0 )L•d−1 24L• / n.
Building on the proof of Corollary 1, we conclude that Assumptions (Var) and (Rem) are satisfied
√
with κ04 , κ05 that only depend on κ, j0 , d, κ06 = 0, and w(n) = n logd−1 (n).
7.5.3. Proof of Corollary 3. Assumption (Conc) is a straightforward consequence of Talagrand’s
inequality for Poisson processes proved by [RB03] (Corollary 2), and is satisfied, whatever θ > 0,
by
n̄ = Vold (Q), κ01 = 1 + θ, κ02 = 12, κ03 = (1.25 + 32/θ).
R
For all λ ∈ m• , σ̂λ2 is an unbiased estimator for Q Ψ2λ s = Vold (Q)Var(β̌λ ). Besides, the existence
of Ωσ follows from Lemma 6.1 in [RBR10]. Thus Assumptions (Var) and (Rem) are satisfied by
taking κ04 = 1, κ05 that only depends on κ and d, κ06 = 0, w(n̄) = C(κ, j0 , d)/ logd+1 (n̄).
(24)
7.5.4. Proof of Corollary 4. The proof is similar to that of Corollary 3 with n̄ = T.
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
27
7.5.5. Proof of Corollary 5. Regarding Assumption (Conc), the proof is similar to that of Corollary 3 with n̄ = n∆. For all λ ∈ m• , let
ZZ
1
2
σ̌λ =
Ψ2λ (x)N (dt, dx).
n∆
[0,n∆]×Q
For any bounded measurable function g on Q, let
Z
Z
Z
c
R(g) =
g(dM − dM ), I(g) =
gdM − E
gdM ,
Q
Q
ˆ =
I(g)
Q
Z
Z
c−E
gdM
Q
c
gdM ,
Q
then
where D∆
ˆ − I(g) + D∆ (g)
R(g) = I(g)
has been defined in the proof of Proposition 3. Notice that
σ̂λ2 − σ̌λ2 = R(Ψ2λ )
and
kŝ?m• − š?m• k2Ψ =
X
R2 (Ψλ ).
λ∈m•
In the course of the proof of Proposition 3, we have shown that, for bounded and Lipschitz
functions g on Q,
|D∆ (g)| ≤ C(λε , ε, f, Q) max {kgk1 , kgk∞ , kgkL } ∆
ˆ
provided ∆ and ε are small enough. Besides, both I(g)
and I(g) satisfy Bernstein inequalities
(Bernstein inequality as stated in [Mas07], Proposition 2.9, for the former, and Bernstein inequality as stated in [RB03], Proposition 7, for the latter). Combining all these arguments yields
Corollary 5.
7.6. Proof of Proposition 5. For α > 0, we set r = bαc + 1.
{`}
(i). From (13), it is easy to see that ∆rh` ,` (f, x) = ∆rh` (u` , x` ). Thus wr (f, t` )p = wr (u` , t` )p
and wre (f, te )p = 0 as soon as e ⊂ {1, . . . , d} contains at least two elements. Therefore,
kf k
α
SBp,q,(d)
≤ C(p)
d
X
α
ku` kBp,q,(1)
.
`=1
(ii). For the sake of readability, we shall detail only two special cases. Let us first deal with the
Qd
α
case f (x) = `=1 u` (x` ) where each u` ∈ Bp,q,(1)
. From (13),
Y
Y
∆r,e
∆rh` (u` , x` )
u` (x` ),
h (f, x) =
`∈e
/
`∈e
so
α
kf kSBp,q,(d)
≤ 2d
d
Y
α
ku` kBp,q,(1)
.
`=1
α1
Let us now assume that d = 3 and that f (x) = u1 (x1 )u2,3 (x2 , x3 ) where u1 ∈ Bp,q,(1)
and
α2
u2,3 ∈ Bp,q,(2) . We set r` = bα` c + 1 for ` = 1, 2, and r̄ = bᾱc + 1, where ᾱ = min(α1 , α2 /2). For
0 < t1 , t2 , t3 < 1, we easily have
kf kp = ku1 kp ku2,3 kp
{1}
t1−ᾱ wr̄ (f, t1 )p ≤ t1−α1 wr1 (u1 , t1 )p ku2,3 kp
{`}
t`−ᾱ wr̄ (f, t` )p ≤ ku1 kp t`−α` wr{`}
(u2,3 , t` )p , for ` = 2, 3
`
{1,`}
ᾱ
t1−ᾱ t−
(f, t1 , t` )p ≤ t1−α1 wr1 (u1 , t1 )p t`−α` wr{`}
(u2,3 , t` )p , for ` = 2, 3.
` wr̄
`
Besides, we deduce from (13) that
k∆r̄h (g, .)kp ≤ C(r̄, p)kgkp ,
and as operators ∆r̄h` ,` commute, we have
{2,3}
ᾱ −ᾱ
t−
2 t3 wr̄
n
o
{2}
{3}
ᾱ −ᾱ
(f, t2 , t3 )p ≤ C(p, r̄)ku1 kp t−
t
min
w
(u
,
t
)
,
w
(u
,
t
)
2,3 2 p
2,3 3 p .
r̄
r̄
2
3
28
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
ᾱ −ᾱ
ᾱ
ᾱ
The inequality of arithmetic and geometric means entails that 2t−
≤ t−2
+ t−2
, so
2 t3
2
3
{2,3}
t2−ᾱ t3−ᾱ wr̄
(f, t2 , t3 )p ≤ C(p, r̄)ku1 kp t2−α2 wr{2}
(u2,3 , t2 )p + t3−α3 wr{3}
(u2,3 , t3 )p .
2
2
In the same way,
{1,2,3}
ᾱ −ᾱ −ᾱ
t−
1 t2 t3 wr̄
−α3 {3}
2
(u
,
t
)
.
w
(u
,
t
)
+
t
wr{2}
(f, t1 , t2 , t3 )p ≤ C(p, r̄)t1−α1 wr1 (u1 , t1 )p t−α
2,3
3
p
2,3
2
p
r
3
2
2
2
Consequently,
ᾱ
≤ C(p, r̄)ku1 kB α1
kf kSBp,q,(d)
p,q,(1)
ku2,3 kB α2
p,q,(2)
.
(iii). The proof follows from the chain rule for higher order derivatives of a composite function.
(r)
(r)
α−r
Notice that for all 1 ≤ ` ≤ d and 1 ≤ r ≤ α − 1, u` ∈ Wp,(1)
, with α − r > 1/p, so u` is bounded.
(iv). The proof follows from a d-variate extension of Theorem 4.1, Inequality (10) in [PST13]
(see also [DL93] Chapter 6, Theorem 3.1).
(v). See Theorem 3.10 in [NS16b].
7.7. Proof of Theorem 2. We recall that for any finite sequence (ai )i∈I , and 0 < p1 , p2 < ∞,
!1/p2
!1/p1
X
X
p2
(1/p2 −1/p1 )+
p1
|ai |
≤ |I|
|ai |
.
i∈I
i∈I
Besides, we have proved in the course of the proof of Proposition 4 that
J` ≤ c1 (M, d)(` − dj0 + d − 1)d−1 .
In the hyperbolic basis, f admits a unique decomposition of the form
∞
X
X
hf, Ψλ iΨ?λ .
f=
`=dj0 λ∈U ∇(`)
Defining
f• =
L•
X
X
hf, Ψλ iΨ?λ ,
`=dj0 λ∈U ∇(`)
we have for finite q > 0, using the aforementioned reminders,
∞
X
X X
kf − f• k2Ψ =
hf, Ψλ i2
`=L• +1 j∈J` λ∈∇j
≤
∞
X
2/p
X
X
2(1/2−1/p)+
(]∇j )
λ∈∇j
`=L• +1 j∈J`
22`(1/2−1/p)+
`=L• +1
X
j∈J`
∞
X
|hf, Ψλ i|p
λ∈∇j
q/p 2/q
X
X
2(1/2−1/q)+
22`(1/2−1/p)+ ]J`
|hf, Ψλ i|p
`=L• +1
≤ C(B, d, p)
X
∞
X
≤ C(B, d, p)
2/p
∞
X
≤ C(B, d, p)
|hf, Ψλ i|p
j∈J`
λ∈∇j
22`(1/2−1/p)+ (` − dj0 + d − 1)2(d−1)(1/2−1/q)+ R2 2−2`(α+1/2−1/p)
`=L• +1
∞
X
2
≤ C(B, d, p)R
(` − dj0 + d − 1)2(d−1)(1/2−1/q)+ 2−2`(α−(1/p−1/2)+ )
`=L• +1
2(d−1)(1/2−1/q)+ −2L• (α−(1/p−1/2)+ )
≤ C(B, α, p, d)R2 L•
2
The case q = ∞ can be treated in the same way.
.
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
29
Let us fix k ∈ {0, . . . , L• − `1 } and define m̄(`1 + k, f ) as the subset of U ∇(`1 + k) such that
{|hf, Ψλ i|; λ ∈ m̄(`1 + k, f )} are the N (`1 , k) largest elements among {|hf, Ψλ i|; λ ∈ U ∇(`1 + k)}.
We then consider the approximation for f given by
`X
1 −1
A(`1 , f ) =
X
LX
• −`1
hf, Ψλ iΨ?λ +
`=dj0 λ∈U ∇(`)
and the set
m`1 (f ) =
`[
1 −1
X
hf, Ψλ i2
k=0 λ∈m̄(`1 +k,f )
U ∇(`) ∪
L•[
−`1
!
m̄(`1 + k, f ) .
k=0
`=dj0
Let us first assume that 0 < p ≤ 2. Using Lemma 4.16 in [Mas07] and (7.7), we get
kf• − A(`1 , f ))k2Ψ =
LX
• −`1
X
hf, Ψλ i2
k=0 λ∈U ∇(`1 +k)\m̄(`1 +k,f )
≤
LX
• −`1
X
|hf, Ψλ i|p
k=0
≤
2/p
LX
• −`1
/(N (`1 , k) + 1)2(1/p−1/2)
λ∈U ∇(`1 +k)
q/p 2/q
X
X
2(1/p−1/q)+
|hf, Ψλ i|p /(N (`1 , k) + 1)2(1/p−1/2) .
]J`1 +k
j∈J`1 +k
k=0
λ∈∇j
Besides, it follows from (11) that
N (`1 , k) + 1 ≥ 2M −d 2−d (d − 1)−(d−1) (`1 + k − dj0 + d − 1)d−1 2`1 (k + 2)−(d+2) .
Therefore
kf• − A(`1 , f ))k2Ψ ≤ C(α, p, d)R2 (`1 − dj0 + d − 1)2(d−1)(1/2−1/ max(p,q)) 2−2α`1 .
In case p ≥ 2, the same kind of upper-bound follows from
kf• − A(`1 , f ))k2Ψ ≤
LX
• −`1
2/p
]U ∇(`1 + k)2(1/2−1/p)
X
|hf, Ψλ i|p
.
λ∈U ∇(`1 +k)
k=0
Last,
kf − A(`1 , f ))k2Ψ = kf − f• k2Ψ + kf• − A(`1 , f ))k2Ψ
which completes the proof.
References
[ACF14] F. Autin, G. Claeskens, and J.-M. Freyermuth. Hyperbolic wavelet thresholding
methods and the curse of dimensionality through the maxiset approach. Applied and
Computational Harmonic Analysis, 36(2):239 – 255, 2014.
[ACF15] Florent Autin, Gerda Claeskens, and Jean-Marc Freyermuth. Asymptotic performance of projection estimators in standard and hyperbolic wavelet bases. Electron.
J. Statist., 9(2):1852–1883, 2015.
[AL11] Nathalie Akakpo and Claire Lacour. Inhomogeneous and anisotropic conditional
density estimation from dependent data. Electronic journal of statistics, 5:1618–
1653, 2011.
[Bar11] Yannick Baraud. Estimator selection with respect to hellinger-type risks. Probability
Theory and Related Fields, 151(1):353–401, 2011.
[BB14] Yannick Baraud and Lucien Birgé. Estimating composite functions by model selection. Ann. Inst. H. Poincaré Probab. Statist., 50(1):285–314, 02 2014.
[BB16] Yannick Baraud and Lucien Birgé. Rho-estimators revisited: general theory and
applications. Working paper or preprint, June 2016.
30
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
[BBM99] Andrew Barron, Lucien Birgé, and Pascal Massart. Risk bounds for model selection
via penalization. Probab. Theory Related Fields, 113(3):301–413, 1999.
[Ber96] Jean Bertoin. Lévy processes, volume 121 of Cambridge Tracts in Mathematics.
Cambridge University Press, Cambridge, 1996.
[BGH09] Yannick Baraud, Christophe Giraud, and Sylvie Huet. Gaussian model selection with
an unknown variance. Ann. Statist., 37(2):630–672, 2009.
[Bir06] Lucien Birgé. Model selection via testing : an alternative to (penalized) maximum
likelihood estimators. Annales de l’I.H.P. Probabilités et statistiques, 42(3):273–325,
2006.
[BM00] L. Birgé and P. Massart. An adaptive compression algorithm in besov spaces. Constructive Approximation, 16(1):1–36, 2000.
[BPP13] Rida Benhaddou, Marianna Pensky, and Dominique Picard. Anisotropic de-noising
in functional deconvolution model with dimension-free convergence rates. Electron.
J. Statist., 7:1686–1715, 2013.
[BS11] Gérard Bourdaud and Winfried Sickel. Composition operators on function spaces
with fractional order of smoothness. RIMS Kokyuroku Bessatsu B, 26:93–132, 2011.
[CDV93] Albert Cohen, Ingrid Daubechies, and Pierre Vial. Wavelets on the interval and fast
wavelet transforms. Appl. Comput. Harmon. Anal., 1(1):54–81, 1993.
[CT04] Rama Cont and Peter Tankov. Financial modelling with jump processes. Chapman &
Hall/CRC Financial Mathematics Series. Chapman & Hall/CRC, Boca Raton, FL,
2004.
[DIT14] Arnak Dalalyan, Yuri Ingster, and Alexandre B. Tsybakov. Statistical inference in
compound functional models. Probability Theory and Related Fields, 158(3):513–532,
2014.
[DKU99] Wolfgang Dahmen, Angela Kunoth, and Karsten Urban. Biorthogonal spline wavelets
on the interval—stability and moment conditions. Appl. Comput. Harmon. Anal.,
6(2):132–196, 1999.
[DL93] Ronald A. DeVore and George G. Lorentz. Constructive approximation, volume
303 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of
Mathematical Sciences]. Springer-Verlag, Berlin, 1993.
[Don97] D. L. Donoho. CART and best-ortho-basis: a connection. Ann. Statist., 25(5):1870–
1911, 1997.
[DTU16] Dinh Dung, Vladimir N. Temlyakov, and Tino Ullrich. Hyperbolic cross approximation. arXiv preprint arXiv:1601.03978v1, 2016.
[Dun01] Dinh Dung. Non-linear approximations using sets of finite cardinality or finite pseudodimension. Journal of Complexity, 17(2):467 – 492, 2001.
[FL09] José E Figueroa-López. Nonparametric estimation for lévy models based on discretesampling. Lecture notes-monograph series, pages 117–146, 2009.
[FLH09] José E. Figueroa-López and Christian Houdré. Small-time expansions for the transition distributions of Lévy processes. Stochastic Process. Appl., 119(11):3862–3889,
2009.
[Hep04] Wang Heping. Representation and approximation of multivariate functions with
mixed smoothness by hyperbolic wavelets. J. Math. Anal. Appl., 291(2):698–715,
2004.
[HM07] Joel L. Horowitz and Enno Mammen. Rate-optimal estimation for a general class
of nonparametric regression models with unknown link functions. Ann. Statist.,
35(6):2589–2619, 12 2007.
[Hoc02a] Reinhard Hochmuth. N -term approximation in anisotropic function spaces. Math.
Nachr., 244:131–149, 2002.
[Hoc02b] Reinhard Hochmuth. Wavelet characterizations for anisotropic Besov spaces. Appl.
Comput. Harmon. Anal., 12(2):179–208, 2002.
[IS07] Yu. Ingster and I. Suslina. Estimation and detection of high-variable functions from
Sloan–Woźniakowski space. Mathematical Methods of Statistics, 16(4):318–353, 2007.
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
31
[JLT09] Anatoli B. Juditsky, Oleg V. Lepski, and Alexandre B. Tsybakov. Nonparametric
estimation of composite functions. Ann. Statist., 37(3):1360–1404, 06 2009.
[KT06] Jan Kallsen and Peter Tankov. Characterization of dependence of multidimensional
Lévy processes using Lévy copulas. Journal of Multivariate Analysis, 97(7):1551 –
1572, 2006.
[Lep13] Oleg Lepski. Multivariate density estimation under sup-norm loss: Oracle approach,
adaptation and independence structure. Ann. Statist., 41(2):1005–1034, 04 2013.
[Mas90] P. Massart. The tight constant in the Dvoretzky-Kiefer-Wolfowitz inequality. Ann.
Probab., 18(3):1269–1283, 1990.
[Mas07] P. Massart. Concentration inequalities and model selection, volume 1896 of Lecture
Notes in Mathematics. Springer, Berlin, 2007. Lectures from the 33rd Summer School
on Probability Theory held in Saint-Flour, July 6–23, 2003, With a foreword by Jean
Picard.
[Mil71] P. W. Millar. Path behavior of processes with stationary independent increments.
Z. Wahrscheinlichkeitstheorie und Verw. Gebiete, 17:53–73, 1971.
[MN09] Alexander J. McNeil and Johanna Nešlehová. Multivariate Archimedean copulas, d-monotone functions and `1 -norm symmetric distributions. Ann. Statist.,
37(5B):3059–3097, 10 2009.
[Mou11] Madani Moussai. The composition in multidimensional Triebel–Lizorkin spaces.
Mathematische Nachrichten, 284(2-3):317–331, 2011.
[Nel06] Roger B. Nelsen. An introduction to copulas. Springer Series in Statistics. Springer,
New York, second edition, 2006.
[Neu00] Michael H. Neumann. Multivariate wavelet thresholding in anisotropic function
spaces. Statist. Sinica, 10(2):399–431, 2000.
[NS16a] Van Kien Nguyen and Winfried Sickel. Isotropic and dominating mixed Besov spaces:
a comparison. arXiv preprint arXiv:1601.04000, 2016.
[NS16b] Van Kien Nguyen and Winfried Sickel. Pointwise multipliers for Sobolev and Besov
spaces of dominating mixed smoothness. arXiv preprint arXiv:1608.03414, 2016.
[NvS97] Michael H. Neumann and Rainer von Sachs. Wavelet thresholding in anisotropic
function classes and application to adaptive estimation of evolutionary spectra. Ann.
Statist., 25(1):38–76, 1997.
[PST13] MK Potapov, BV Simonov, and S Yu Tikhonov. Mixed moduli of smoothness in
Lp , 1 < p < ∞: a survey. Surveys in Approximation Theory, 8(18), 2013.
[RB03] Patricia Reynaud-Bouret. Adaptive estimation of the intensity of inhomogeneous
Poisson processes via concentration inequalities. Probab. Theory Related Fields,
126(1):103–153, 2003.
[RBR10] Patricia Reynaud-Bouret and Vincent Rivoirard. Near optimal thresholding estimation of a Poisson intensity on the real line. Electron. J. Stat., 4:172–238, 2010.
[RBRTM11] Patricia Reynaud-Bouret, Vincent Rivoirard, and Christine Tuleau-Malot. Adaptive
density estimation: a curse of support? J. Statist. Plann. Inference, 141(1):115–139,
2011.
[Reb15a] Gilles Rebelles. Lp - adaptive estimation of an anisotropic density under independence
hypothesis. Electron. J. Statist., 9(1):106–134, 2015.
[Reb15b] Gilles Rebelles. Pointwise adaptive estimation of a multivariate density under independence hypothesis. Bernoulli, 21(4):1984–2023, 11 2015.
[RW02] Ludger Rüschendorf and Jeannette H. C. Woerner. Expansion of transition distributions of Lévy processes in small time. Bernoulli, 8(1):81–96, 2002.
[Sat99] Ken-iti Sato. Lévy processes and infinitely divisible distributions, volume 68 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge,
1999. Translated from the 1990 Japanese original, Revised by the author.
[Skl59] A. Sklar. Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist.
Univ. Paris, 8:229–231, 1959.
[ST87] Hans-Jürgen Schmeisser and Hans Triebel. Topics in Fourier analysis and function
spaces. A Wiley-Interscience Publication. John Wiley & Sons, Ltd., Chichester, 1987.
32
MULTIVARIATE INTENSITY ESTIMATION VIA HYPERBOLIC WAVELET SELECTION
[UK11] Florian AJ Ueltzhöfer and Claudia Klüppelberg. An oracle inequality for penalised
projection estimation of lévy densities from high-frequency observations. Journal of
Nonparametric Statistics, 23(4):967–989, 2011.
[YB99] Yuhong Yang and Andrew Barron. Information-theoretic determination of minimax
rates of convergence. Ann. Statist., 27(5):1564–1599, 1999.
| 10math.ST
|
arXiv:1611.00297v4 [cs.IT] 3 Dec 2017
Generalized Entropy Concentration for Counts
Kostas N. Oikonomou
AT&T Labs Research
Middletown, NJ 07748, U.S.A.
Email: [email protected]
November 2017
Abstract
The phenomenon of entropy concentration provides strong support for the maximum entropy method, MaxEnt, for inferring a probability vector from information
in the form of constraints. Here we extend this phenomenon, in a discrete setting, to
non-negative integral vectors not necessarily summing to 1. We show that linear constraints that simply bound the allowable sums suffice for concentration to occur even
in this setting. This requires a new, ‘generalized’ entropy measure in which the sum of
the vector plays a role. We measure the concentration in terms of deviation from the
maximum generalized entropy value, or in terms of the distance from the maximum
generalized entropy vector. We provide non-asymptotic bounds on the concentration
in terms of various parameters, including a tolerance on the constraints which ensures
that they are always satisfied by an integral vector. Generalized entropy maximization
is not only compatible with ordinary MaxEnt, but can also be considered an extension of it, as it allows us to address problems that cannot be formulated as MaxEnt
problems.
Keywords: maximum generalized entropy, counts, concentration, linear constraints,
inequalities, norms, tolerances
Contents
1 Introduction
2 The
2.1
2.2
2.3
2.4
2.5
2
generalized entropy G
Basic properties . . . . . . . . . . . . .
Monotonicity and concavity properties
Lower bounds . . . . . . . . . . . . . .
Maximization . . . . . . . . . . . . . .
A connection with I-divergence . . . .
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
8
9
10
11
12
3 Constraints, scaling, sensitivity, and the optimal count vector
3.1 Constraints with tolerances . . . . . . . . . . . . . . . . . . . . .
3.2 Effect of tolerances on the optimality of x∗ . . . . . . . . . . . .
3.3 Scaling of the data and bounds on the allowable sums . . . . . .
3.4 The optimal count vector ν ∗ . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
14
15
16
17
4 Concentration with respect to entropy difference
4.1 Realizations of the optimal count vector . . . . . .
4.2 Realizations of the sets with smaller entropy . . . .
4.3 The scaling factor needed for concentration . . . .
4.3.1 Bounds on the concentration threshold . . .
4.4 Examples . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
18
19
21
23
25
26
5 Concentration with respect to distance from the MaxGEnt vector
5.1 Realizations of the sets far from the MaxGEnt vector . . . . . . . . . . . .
5.2 Scaling and concentration around the MaxGEnt count vector . . . . . . .
5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
29
31
34
6 Conclusion
37
A Proofs
37
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introduction
The maximum entropy method or principle, originally proposed by E.T. Jaynes in 1957,
now appears in standard textbooks on engineering probability and information theory,
[PP02], [CT06]. Commonly referred to as MaxEnt, the principle essentially states that if
the only information available about a probability vector is in the form of linear constraints
on its elements, then, among all others, the preferred probability vector is the one that
maximizes the Shannon entropy under these constraints. Besides the great wealth and
diversity of its applications, MaxEnt can be justified on a variety of theoretical grounds:
axiomatic formulations ([SJ80], [Ski89], [Csi91], [Cat12]), the concentration phenomenon
([Jay83], [Gr8], [Cat12], [OG16]), decision- and game-theoretic interpretations ([Gr8] and
references therein), and its unification with Bayesian inference ([GC07], [Cat12]).
Among these justifications, in a discrete setting, the appeal of concentration lies in its
conceptual simplicity. It is essentially a combinatorial argument, first presented by E.T.
Jaynes [Jay83], who called it “concentration of distributions at entropy maxima”. The
concentration viewpoint was further developed in [Gr8] and [OG16], which presented generalizations, improved results, eliminated the asymptotics, and studied additional aspects.
In this paper we adopt a discrete, finite, non-probabilistic, combinatorial approach,
and show that the concentration phenomenon arises in a new setting, that of non-negative
2
vectors which are not necessarily density vectors1.1 . Among other things, this requires introducing a new, ‘generalized’ entropy measure. This new concentration phenomenon lends
support to an extension of the MaxEnt method to what we call “maximum generalized
entropy”, or MaxGEnt.
The basics of entropy concentration are easiest to explain in terms of the abstract
“balls and bins” paradigm ([Jay03]). There are m labelled, distinguishable bins, to which
n indistinguishable balls are to be allocated one-by-one. The final content of the bins
is described by a count vector ν = (ν1 , . . . , νm ) which sums to n, and a corresponding
frequency vector f = ν/n, summing to 1.P
Suppose that theP
frequency vector must satisfy
a set of linear equalities and inequalities, i aij fi = bj and i aij fi 6 bj , with aij , bj ∈ R.
The concentration phenomenon is that as n becomes large, the overwhelming majority of
the allocations which accord with the constraints have frequency vectors that are close to
the m-vector which maximizes the Shannon entropy subject to the constraints.
In our extension there is no longer a given number of balls. Therefore we cannot define
a unique frequeny vector, but must deal directly with count vectors ν whose sums are
unknown (Example 1.1 below makes this clear). The linear constraints are now placed on
the counts νi , again with coefficients in R. Our only assumption about the constraints is
that they limit the sums of the count vectors to lie in a finite range [s1 , s2 ]. With just
this assumption, we show that as the counts are allowed to become larger and larger (by
a process of scaling the problem, explained in §3), the vast majority of allocations that
satisfy the constraints in fact have count vectors close to the non-negative m-vector x∗
that maximizes the generalized entropy G(x). A precise statement of this concentration
phenomenon needs some additional preliminaries, and is given at the end of this section.
Our main results are, in §2, a new generalized entropy function G, defined on arbitrary
non-negative vectors, which reduces to the Shannon entropy H on vectors summing to 1;
its properties are studied in §2 and §3, where the scaling process is also introduced. In §4
we demonstrate the new concentration pheonomenon with respect to deviations from the
maximum generalized entropy value G∗ . Theorem 4.1 gives a lower bound on the ratio of
the number of realizations of the MaxGEnt vector to that of the set of count vectors ν
whose generalized entropies G(ν) are far from the maximum value G∗ . Then Theorem 4.2
completes the picture by deriving how large the problem must be for the above ratio to
be suitably large. In §5 we establish concentration with respect to the `1 norm distance
of the count vectors from the MaxGEnt vector ; we present Theorems 5.1 and 5.2, which
are analogous to those of §4, and also Theorem 5.3, an optimized version of Theorem 5.2.
In all the theorems, ‘far’, ‘large’, etc. are defined in terms of parameters, introduced in
Table 1.2 below. None of our results involve any asymptotic considerations, and we give a
number of numerical illustrations.
The following example demonstrates the basic issues referred to above in a very simple
1.1
What we call here ‘density’ or ‘frequency’ vectors would be called “discrete probability distributions”,
possibly ‘empirical’, if we were operating in a probabilistic setting.
3
setting, which highlights the differences with the usual frequency vector case. After this,
we proceed to the precise statement of generalized entropy concentration.
Example 1.1 A number of indistinguishable balls1.2 are to be placed one-by-one in three
bins, red, green, and blue. The final content (νr , νg , νb ) of the bins must satisfy νr + νg = 4
and νg + νb 6 6. Thus the total number of balls that may be put in the bins cannot be too
small, e.g. 3, or too large, e.g. 20. Each assignment of balls to the bins is described by a
sequence made from the letters r, g, b, with a corresponding count vector ν = (νr , νg , νb );
the sequence can be of any length n consistent with the constraints. Table 1.1 lists all the
count vectors that satisfy the constraints, their sums n, and their number of realizations
#ν, i.e. the number of sequences that result in these counts, given by a multinomial
7
coefficient, e.g. 3,1,3
= 140. [In the terminology of the theory of types, #ν is the size
of the type class T (ν/n).] What can be said about the “most likely” final content of the
bins?
νr
0
1
2
3
4
0
1
2
3
4
νg
4
3
2
1
0
4
3
2
1
0
νb
0
0
0
0
1
1
1
1
1
1
n
4
5
#ν
1
4
6
4
1
5
20
30
20
5
νr
0
1
2
3
4
1
2
3
4
νg
4
3
2
1
0
3
2
1
0
νb
2
2
2
2
2
3
3
3
3
n
6
7
#ν
15
60
90
60
15
140
210
140
35
νr
2
3
4
3
4
4
νg
2
1
0
1
0
0
νb
4
4
4
5
5
6
n
8
9
10
#ν
420
280
70
504
126
210
Table 1.1: The count vectors ν = (νr , νg , νb ) satisfying νr + νg = 4, νg + νb 6 6, their sum n, and
their number of realizations #ν. If we had the additional constraint νr + νg + νb = 7, only the
n = 7 section of the table would apply, and we would reduce to a MaxEnt problem.
This example makes two points. First, it does not seem possible to find a single frequency vector that can be naturally associated with the problem; without that, one cannot
think about maximizing the usual entropy1.3 . Second, one may think that starting with
the largest possible number of balls, 10 in this case, would lead to the greatest number of
realizations. But this is not so: the count vector with the most realizations sums to 9, and
even vectors summing to 8 have more realizations than the one summing to 10.
1.2
The balls don’t have to be indistinguishable, we just ignore distinguishing characteristics, if they have
any. However, in modelling some situations, such as in Example 3.1, indistinguishability is essential.
1.3
In the ordinary entropy problem where we have a single n, the distinction between count and frequency
vectors doesn’t really matter, there is a 1-1 correspondence; but this is not true here.
4
Next we give a precise statement, GC below, of generalized entropy concentration. To
do that we need to (a) define the generalized entropy and describe how to find the vector
that maximizes it, (b) specify how to derive the bounds s1 , s2 from the constraints, (c)
describe how to ensure the existence of integral solutions (count vectors) to the constraints,
and (d) introduce parameters that define the concentration.
To find the vector with the largest number of realizations in a problem like that of
Example 1.1, we first assume that the problem does not admit arbitrarily large solutions.
This is made precise in (1.2) below, but a necessary condition is that each element of ν
appears in some constraint1.4 . Next we relax the integrality requirement on the counts,
and set up a continuous maximization problem
P
P
P
maxx∈C G(x), where G(x) = − i xi ln xi + ( i xi ) ln ( i xi )
,
(1.1)
and
C = {x ∈ Rm | AE x = bE , AI x 6 bI , x > 0}.
Here G(x) is the generalized entropy of the real vector x > 0, and the constraints on x are
expressed via the real matrices AE , AI and vectors bE , bI . We assume that the constraints
(a) are satisfiable, and (b) they bound the possible sums of the x ∈ C; this is equivalent to
assuming that all xi are bounded. Thus C is a non-empty polytope in Rm and (1.1) is a
concave maximization problem (see e.g. [BV04]) with a solution x∗ . We will refer to (1.1)
as the “MaxGEnt problem” and to x∗ as “the MaxGEnt vector” or as “the optimal
relaxed count vector”. Since the function G is concave but not strictly concave, see Fig.
2.1 in §2, it is not immediate that the solution x∗ is unique; however, we show that this is
the case in §2.4.
P
The boundedness assumption is that i xi lies between (finite) numbers s1 and s2 ;
these are determined by solving the linear programs
s1 , min(x1 + · · · + xm ),
s2 , max(x1 + · · · + xm ).
x∈C
x∈C
(1.2)
(A technicality is that the constraints may force some elements of x∗ to be 0; for reasons
explained in §3 it is convenient to eliminate such elements, so that in the end all elements
of x∗ can be assumed to be positive reals.) Finally, from x∗ we derive an integral vector ν ∗ ,
to which we refer as the optimal, or MaxGEnt count vector, by a procedure explained in
§3.
Because in the end we are interested only in integral/count vectors in the set C of (1.1),
we will introduce, as explained in §3, tolerances on the satisfaction on the constraints,
governed by a parameter δ. This will turn C into C(δ). To describe the concentration
we need two more parameters, ε specifying the strength of the concentration, and η or ϑ
describing the size of the region in which it occurs. The parameters are summarized in
Table 1.2.
Lastly, when we have ordinary entropy and frequency vectors, concentration occurs by
increasing the number of balls n. With count vectors, this is replaced by increasing bE , bI ,
1.4
But this is not sufficient: consider, e.g. m = 2 and ν1 , ν2 > 0, ν1 − ν2 = 10.
5
δ:
ε:
η:
ϑ:
relative tolerance in satisfying the constraints
concentration tolerance, on number of realizations
relative tolerance in deviation from the maximum generalized entropy value G∗
absolute tolerance in deviation (distance) from the optimal relaxed count vector x∗
Table 1.2: Parameters for the concentration results.
the values of the constraints. The increase we consider here consists in multiplying these
vectors by a scalar c > 1, a process which we call scaling. This scaling results in larger and
larger count vectors being admissible and is described in detail in §3.
Now we can give the precise statement of the concentration phenomenon for count
vectors:
GC: Theorems 4.2 and 5.2 compute a number ĉ(δ, ε, η) and ĉ(δ, ε, ϑ), respectively, called the “concentration threshold”, such that if the problem data bE , bI
is scaled by any factor c > ĉ, the number of assignments/sequences that result
in the optimal count vector ν ∗ is at least 1/ε times greater than the number of
all assignments that result in count vectors with entropy less than (1 − η)G∗
or farther than ϑ from x∗ by `1 norm.
Significance
In a problem where the only available information is embodied in the constraints and
which otherwise admits a large number of probability vectors as solutions, the concentration phenomenon provides a powerful argument for the MaxEnt method, which selects a
particular solution, the one with maximum entropy, in preference to all others1.5 . Likewise,
the concentration results in this paper support the maximization of generalized entropy
for problems involving general non-negative vectors. We believe that MaxGEnt can be
considered to be a compatible extension of MaxEnt. The compatibility is that any MaxEnt problem over the reals with constraints AE x = bE , AI x 6 bI can be formulated as
aPMaxGEnt problem of the form (1.1) with the same constraints, plus the constraint
∗
m
i xi = 1; both problems will have the same solution x ∈ R , and the maximum en∗
∗
tropy H(x ) will equal the maximum generalized entropy G(x ). Also, if P
the constraints
of the MaxGEnt problem either explicitly or implicitly fix the value of i xi , then the
problem can be reduced to a MaxEnt problem over the reals. The extension consists in
the fact that MaxGEnt addresses problems involving un-normalized vectors that cannot
be formulated as MaxEnt problems, as we saw in Example 1.1; more examples of such
problems are given in §3, §4, and §5.
1.5
MaxEnt solves the inference problem, not the decision problem. It does not claim that the maximum
entropy object is the one to use no matter what use one has in mind.
6
Related work
Our term “generalized entropy” for G is neither imaginative nor distinctive, and there
are many other generalized entropy measures. The most general of these are Csiszár’s
f -entropies and f -divergences [Csi96], and the related Φ-entropies of [BLM13]. Any relationship of G to Φ-entropies remains to be investigated. The function G, in the form
of the log of a multinomial coefficient with “variable numerator”, appeared in [OS06] and
[Oik12].
The problem of inferring a non-negative real vector from information in the form of
linear equalities was considered by Skilling [Ski89], where such vectors were termed “positive additive distributions”, and by Csizsár, [Csi91], [Csi96]. Both authors gave axiomatic
justifications, which do not involve probabilities, for minimizing the I-divergence, a generalization of relative entropy to un-normalized vectors. A further generalization is the α, β
divergences of [CCA11]. We discuss a connection between I-divergence and our generalized
entropy in §2.5.
With respect to concentration, recent developments for the discrete, normalized case
were given in [OG16]. The continuous normalized case, for relative entropy, is examined
in [Cat12] from the viewpoint of information geometry. Countable spaces are also treated
in [Gr8]. But these references do not provide explicit bounds such as the ones here and
in [OG16]. To our knowledge, concentration for non-density vectors has not been studied
before.
The structure and some of the presentation of this paper are similar to [OG16] because
of the similar subject matter, entropy concentration from a combinatorial viewpoint. Many
of the results here that appear similar to those of section III of [OG16] are generalizations
of those results, insofar as G is a generalization of H. However the main theorems here
do not actually subsume corresponding theorems in [OG16], because in both cases the
theorems include optimizations specific to count or frequency vectors, respectively.
2
The generalized entropy G
In this section we introduce the generalized entropy function G, and study its properties,
relationships with other functions, and its maximization under linear constraints.
Given a real vector x > 0, its generalized entropy is
X X
X
G(x) , H(x) +
xi ln
xi =
xi H(χ), x > 0.
(2.1)
i
i
i
Here H(x) is the form − i xi ln xi extended to vectors in Rm
+ that are not necessarily
density vectors, and χ is the density, or normalized, or probability, vector corresponding
to x. (2.1) gives two ways to look at G(x): it is the (extended) entropy of x plus the sum
of x times its log, or the sum of x times the ordinary entropy of the normalized x. If x is
already normalized G(x) coincides with H(x). Fig. 2.1 is a plot of G(x) for m = 2.
P
7
5
4
4
3
3
2
2
4.0
3.5
3.0
x2
2.5
2.0
1.5
1.0
1.0
1.5
2.0
2.5
3.0
3.5
4.0
1.0
x1
1.5
2.0
2.5
x1
3.0
3.5
4.0
1.0
1.5
2.0
2.5
3.0
3.5
G(x1 , x2 )
G( x 1 , x 2 )
5
4.0
x2
Figure 2.1: G(x1 , x2 ). Note that G(x, x) = (2 ln 2)x, which destroys the strict concavity of
G.
2.1
Basic properties
We list some important properties of the function G:
m
to “second Stirling
P1 G(x1 , . . . , xm ) is the log of the multinomial coefficient xx1 +···+x
1 ,...,xm
√
1
ϑ
,ϑ ∈
order”: by using the first two terms of ln x! = x ln x − x + 2 ln x + ln 2π + 12x
(0, 1),we find that
x1 + · · · + xm
ln
≈ G(x1 , . . . , xm ).
x1 , . . . , x m
This interpretation was given in [Oik12], where it was used to derive “most likely”
matrices, i.e. those with the largest number of realizations, from incomplete information.
P2 G is related to the ordinary entropy (of density vectors) and the extended entropy (of
arbitrary non-negative vectors) H in the two ways specified in (2.1).
P3 Unlike the entropy of normalized vectors which is bounded by ln m, the generalized
entropy G(x) increases without bound as the elements of x become larger: for any
x, y, if y > x then G(y) > G(x). This is shown in Proposition 2.3. One consequence
is that if x, y are close in norm, i.e. kx − yk 6 ζ, |G(x) − G(y)| cannot be bounded
by an expression involving only m and ζ.
P4 G(x) is positive, unless x has just one non-0 element, in which case G(x) = 0. This
follows from the second form in (2.1).
P5 Given any p.d. p = (p1 , . . . , pm ) and any n-sequence σ with count vector ν, the
probability of σ under p can be written as
Prp (σ) = e−(G(ν)+nD(f kp))
8
where D(· | ·) is the divergence, or relative entropy, between two probability vectors
and f = ν/n is the frequency vector corresponding to ν. By substituting G(ν) =
nH(f ) we obtain the well-known expression for the same probability in terms of the
ordinary entropy of a frequency vector.
P6 Like the ordinary or the extended H, G(x1 , . . . , xm ) is concave over the domain x1 >
0, . . . , xm > 0, but unlike H, it is not strictly concave. See Proposition 2.2 in §2.2.
P
P7 The maximum of G(x1 , . . . , xm ) subject just to the constraint
i xi = s is s ln m.
When s = 1, x is a density vector and this reduces to the maximum of H.
P8 What is the relationship between maximizing G and maximizing the extended H?
Consider maximizing the first formPin (2.1), subject to AE x = bE , AI x 6 bI , by
imposing the additional constraint i xi = s and treating s as a parameter taking
values in [s1 , s2 ]. For a given s, there will be a unique maximum since H is strictly
∗
concave2.1 . Further, some
subject to
P s = s will achieve maxs maxx (s ln s + H(x))
E
E
I
I
A x = b , A x 6 b , i xi = s; this maximum value will equal G(x∗ ). Using the
second form in (2.1) we see that there is a similar relationship between maximizing
G(x) and maximizing the function sH(x/s).
P9 G has a scaling (or homogeneity) property, which H does not: for any c > 0 and any
x ∈ Rm
+ , G(cx) = cG(x). This is most easily seen from the second form in (2.1).
P10 G has a further important scaling property: if x∗ maximizes G(x) under Ax 6 b, then
for any c > 0, cx∗ maximizes G(x) under Ax 6 cb. We show this in §3.3, Proposition
3.2.
2.2
Monotonicity and concavity properties
As we noted in property 2.1 in §2.1, G is an increasing function in the sense that
Proposition 2.1 For any x, y, if y > x then G(y) > G(x), and if the inequality is strict
in some places, then G(y) > G(x).
We will use this property in §2.3. Now we turn to concavity.
The extended ordinary entropy H is strictly concave, and in addition, strongly concave
for any modulus γ 6 1/a when defined over [0, a]m . The generalized entropy G is also
concave, but neither strictly concave, nor strongly concave for any modulus. However −G
is sublinear, whereas −H is not. These properties are collected in the following proposition:
Proposition 2.2
1. The function G(x1 , .., xm ) is concave over Rm
+.
2.1
One may also maximize H without the constraint
9
P
xi = s, but what would the result mean?
2. G is not strictly concave over Rm
+.
3. G is not strongly concave over Rm
+ for any modulus γ > 0.
4. If the definition of G is extended over all of Rm by setting G(x) = −∞ if any xi is < 0,
then for all α, β > 0 and for all x, y ∈ Rm , G(αx + βy) > αG(x) + βG(y).
The last property is stronger than (implies) concavity since α, β are not required to sum
to 1. The absence of strict concavity means that more care is needed with maximization,
we address this in §2.4.
2.3
Lower bounds
Given a point x, if some other point y is close to it in the distance/norm sense, how much
smaller than G(x) can G(y) be? We will need the answer in §4. Proposition 2.1 implies that
if we have a hypercube centered at x, say kx − yk∞ 6 ζ, then G(·) attains its maximum at
the “upper right-hand” corner of the hypercube and its minimum at the “lower left-hand”
corner. Specifically, for any ζ > 0, let ζ denote the m-vector (ζ, . . . , ζ), and let x > ζ.
Then it can be seen from Proposition 2.1 that for any y > 0
kx − yk∞ 6 ζ
⇒
G(x − ζ) 6 G(y) 6 G(x + ζ).
(2.2)
Using this observation we can show that
Lemma 2.1 Given ζ > 0 and x, y ∈ Rm
+ , if x > ζ and ky − xk∞ 6 ζ, then
G(y) > G(x) −
X
i
ln
kxk1
1 X 1
m
ζ−
−
ζ 2.
xi
2
xi − ζ
kxk1 /m − ζ
i
The coefficient of ζ 2 is positive unless all xi are equal, in which case it becomes 0.
The lower bound above does not depend on y, only on x and ζ. The restriction x > ζ
applies to the ‘reference’ point x, not to the ‘variable’ y; see also Remark 4.2. Lastly, since
kx − yk1 6 ζ ⇒ kx − yk∞ 6 ζ, the lemma holds also when the `∞ norm is replaced by the
`1 norm.
We will use Lemma 2.1 in §4.1 to bound how far from the maximum G(x∗ ) the value
G(x) can be if x is close to x∗ . We also comment there, Remark 4.3, on how the above
bound compares to bounds obtainable from the relationship between (ordinary) entropy
difference and `1 norm.
10
2.4
Maximization
Let C(0) denote the subset of Rm defined by the constraints in (1.1)2.2 . Here we point out
that the point x∗ solving (1.1) occupies a special location in this set, and consequently is
the unique optimal solution of our maximization problem, despite the fact that G is not a
strictly concave function (recall Proposition 2.2, part 2).
Proposition 2.3
1. The set C(0) does not contain any x s.t. x > x∗ with at least one strict inequality.
2. The point x∗ is the unique optimal solution of problem (1.1).
Figure 2.2 illustrates the first statement of the proposition.
x2
x1
0
Figure 2.2: A 2-dimensional polytope C(0). By Proposition 2.3, x∗ can lie only on the heavy black
line.
Finally we look at the form of the solution x∗ in terms of Lagrange multipliers. The
Lagrangean for problem (1.1) is
L(x, λE , λI ) = G(x) − λE · (AE x − bE ) − λI · (AI x − bI ),
(2.3)
where λE , λI are the vectors of the Lagrange multipliers corresponding to the equality and
inequality constraints. The solution x∗ will satisfy some of the inequality constraints with
equality (and these are called binding or active at x∗ ), and some with strict inequality.
It is known that multipliers λIj corresponding to inequalities non-binding at x∗ will be 0,
while the rest of them will be > 0 (see, e.g., [HUL96], Ch. VII, §2.4). Thus, denoting
the sub-vector of λI corresponding to binding inequalities by λBI and the corresponding
sub-matrix of AI by ABI , it follows from (2.3) that x∗ can be written as
x∗j = (x∗1 + · · · + x∗m )e−(λ
E ·AE +λBI ·ABI )
.j
.j
.
This expression determines the elements of the density vector χ∗ = x∗ /
the multipliers, but it does not determine the vector x∗ itself.
2.2
(2.4)
∗
i xi
P
The reason for the “0” will be seen in §3.1, where we discuss tolerances on constraints.
11
in terms of
Remark 2.1 It is clear that the form (2.4) cannot express any elements of x∗ that are 0, if
the multipliers λ are to be finite. To avoid introducing special cases in the sequel to handle
the zeros, we will assume as a convenience that any elements of the solution to problem
(1.1) that are forced to be exactly 0 by the constraints are eliminated from consideration
either before or after the solution is found. We have already alluded to this after (1.2).
Thus, whenever we speak of x∗ in what follows we will assume that all of its elements are
positive. See Example 5.3 in §5. A more detailed discussion of the issue of 0s is in [OG16],
§II.A.
Example 2.1 Returning to Example 1.1, it is possible to maximize G analytically under
the given constraints. Introducing real variables x1 , x2 , x3 corresponding to νr , νg , νb and
letting the constraints be x1 + x2 = a and x2 + x3 6 b, the solution turns out to be
√
a + b + a2 + b2
∗
∗
∗
∗
∗
∗
∗
x1 = s − b, x2 = a + b − s , x3 = s − a,
s =
.
2
Further, the bounds s1 , s2 of (1.2) on the possible sums are s1 = a and s2 = a + b. We
see that the MaxGEnt solution to the problem is never trivial,
in the sense that for all
√
13
1
∗
∗
a, b, we have s1 < s < s2 ; when a = b we have s = 2 1 + 5 s2 ≈ 0.861s2 . With
a = 4, b = 6 we find s∗ = 8.61 and x∗ = (2.61, 1.39, 4.61); compare with Table 1.1.
2.5
A connection with I-divergence
For density vectors, the relationship between ordinary entropy H(x) and divergence D(xky)
is well known: with uniform y, D(xky) reduces to H(x) to within a constant, and its
minimization is equivalent to the maximization of H(x). Here we look at whether G(x)
has any analogous properties.
P
First, if in D(xky) we take y to have all of its elements equal to
i xi , we obtain
2.3 . For example, minimizing D(xky)
−G(x). However, this is merely
a
formal
relationship
P
P
with respect to x when y = ( i xi , . . . , i xi ) cannot be given the same interpretation as
minimizing D(xky) with respect to x given a fixed ‘prior’ y. So even if x, y summed to 1,
neither the axiomatic nor the concentration justifications for cross-entropy minimization
would apply.
Second, the concentration properties we establish in §4 and §5 support the maximization
of G(x) as a method of inference of non-negative vectors from limited information. Another
method for doing this, suggested in [Ski89], [Csi96], is based on minimizing the I-divergence
(information divergence) between non-negative vectors
D(ukv) ,
X
i
2.3
ui ln
X
ui X
−
ui +
vi ,
vi
i
This is pointed out in [BV04], Ch. 3, Example 3.19.
12
i
u, v ∈ Rm
+.
(2.5)
This reduces to D(ukv) when u, v sum to 1. The inference problem is “problem (iii)” in
[Csi96]: infer a non-negative function p(z), not necessarily summing or integrating to 1,
given that (a) it belongs to a certain feasible set F of functions defined by linear equality
constraints, and (b) a default model q(z)2.4 . It is shown that the solution of this problem is
the p∗ ∈ F that minimizes the I-divergence D(pkq). (Recently, minimization of I-divergence
and generalizations to “α, β divergences” has found many applications in the area known
as “non-negative matrix factorization”, see [CCA11].)
There is a relationship between minimizing I-divergence and maximizing generalized
entropy:
Proposition 2.4 Let (AE , bE ), (AI , bI ) be linear equality and inequality constraints on a
∗
vector in Rm
+ , and let x be the solution of the MaxGEnt problem with these constraints on
∗
x. Given a prior v ∈ Rm
+ , let u (v) be the solution to the minimum I-divergence problem
with the same constraints on u. Then there is a prior ṽ which makes the two solutions
coincide, i.e. u∗ (ṽ) = x∗ . That prior is ṽ = (s∗ , . . . , s∗ ).
This follows from the fact that the minimum I-divergence solution to a problem with
prior v and constraints AE u = bE and AI u 6 bI on u is
u∗j = vj e−(λ
E ·AE +λBI ·ABI )
.j
.j
.
(2.6)
If we set vj = s∗ , it can be seen from expression (2.4) that u∗j = x∗j satisfies (2.6).
Inference by minimizing I-divergence under equality constraints has an axiomatic basis,
but as pointed out in §3 and §7 of [Csi96], the combinatorial, concentration rationale that we
are advocating here does not seem to apply to it. Proposition 2.4 shows that the adoption of
a particular prior furnishes this rationale, except that this prior cannot be properly viewed
as independent of the solution (posterior) u∗ . This dependence may shed some light on the
difficulty of finding the concentration rationale in general. [As an illustration, Example 2.1
can be solved by I-divergence minimization assuming a constant prior v = (α, α, α). An
analytical solution u∗ is possible, and it has the same form as the MaxGEnt solution, but
it is a function of α ∈ (0, ∞); the question then becomes what value to adopt for α.]
3
Constraints, scaling, sensitivity, and the optimal count
vector
In §3.1 we discuss the necessity of introducing tolerances into the constraints defining the
MaxGEnt problem, and in §3.2 the effect of these tolerances on the maximization of G.
In §3.3 we turn to the scaling of the problem, i.e. multiplying the data vector b by some
c > 0, and the important properties of this scaling. Lastly, in §3.4 we discuss the optimal ,
or MaxGEnt count vector ν ∗ , constructed from the real vector x∗ solving problem (1.1).
2.4
The sense of ‘default’ is that if q is in F, then, in the absence of any constraints, the method should
infer p∗ = q.
13
3.1
Constraints with tolerances
We pointed out the necessity of introducing tolerances into linear constraints when establishing concentration of ordinary entropy in [OG16]. The constraints involved real
coefficients, and the solutions had to be rational (frequency) vectors with a particular
denominator. Here the solutions need to be integral (count) vectors, but the equality constraints may not have any integral solution; e.g. x1 − x2 = 1, x1 + x2 = 4 are satisfied only
for (x1 , x2 ) = (2.5, 1.5), and likewise with inequalities, e.g. 1.3 6 x1 6 1.99. We therefore define the set of real m-vectors x that satisfy the constraints in (1.1) with a relative
accuracy or tolerance δ > 0:
C(δ) , {x ∈ Rm : bE − δ|β E | 6 AE x 6 bE + δ|β E |, AI x 6 bI + δ|β I |},
(3.1)
where β E , β I are identical to bE , bI , except that any elements that are 0 are replaced
by appropriate small positive constants. The tolerances are only on the values b of the
constraints, not on their structure A. Recall that the generalized entropy is maximized
over C(0), which we have assumed to be non-empty, problem (1.1).
There are three main points concerning the introduction of δ. First, the existence of
integral solutions, which is elaborated in Proposition 3.1 below. Second, and related to
the first, δ ensures that the concentration statement GC in §1 holds for all scalings of the
problem larger than a threshold ĉ. This is analogous to having concentration for frequency
(rational) vectors hold for all denominators n larger than some N , as in [OG16]. Third, δ
has an effect on the maximization of G; this the subject of §3.2.
Proposition 3.1 below gives the fundamental facts about the existence of count vectors
in C(δ). Given an x in C(0), any other vector y close enough to it is in C(δ), and, if δ is
not too small, the count vector obtained by rounding x element-wise is in C(δ); in other
words, for every real vector in C(0) there is an integral vector in C(δ). The “close enough”
and the “not too small” depend on a number ϑ∞ :
Proposition 3.1 With β E , β I as in (3.1), define
ϑ∞ , min(|β E |min / 9 AE 9∞ , |β I |min / 9 AI 9∞ ),
or ∞ if there are no constraints3.1 . Then if x is any point in C(0),
1. Given any δ > 0, any y ∈ Rm
+ such that ky − xk∞ 6 δ ϑ∞ is in C(δ).
2. In particular, if δ > 1/(2ϑ∞ ), the integral/count vector [x] is in C(δ).
As we add constraints to a problem, ϑ∞ can only decrease, or at best stay the same.
This proposition is used in §4.3, eq. (4.23), and in §5.2, after (5.9).
3.1
Recall that the infinity norm 9 · 9∞ of a matrix is the maximum of the `1 norms of the rows.
14
Example 3.1 Fig. 3.1 shows a network consisting of 6 nodes and 6 links. The links
are subject to a certain impairment x and xi is the quantity associated with link i. The
impairment is additive, e.g. its value over the path AB consisting of links 4, 1, 6 is x4 +
x1 + x6 .
B
x4 + x1 + x6 = b1
x6 + x3 + x5 = b2
x4 + x2 + x5 = b3
x4, x5, x6 6 b4
6
3
1
A
4
5
2
C
Figure 3.1: Data b on the impairment x in a 6-node, 6-link network.
Suppose that x is measured over the 3 paths AB, BC, CA, and it is also known that
the access links 4, 5, 6 contribute no more than a certain amount, as shown in Fig. 3.1.
The structure matrices AE , AI and data vectors bE , bI then are
#
"
"
#
"
#
"
#
AE =
1
0
0
0
0
1
0
1
0
1
0
1
0
1
1
1
1
0
,
bE =
b1
b2
b3
,
AI =
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
1
,
bI =
b4
b4
b4
.
The problem is to infer the impairment vector x from the measurement vector b. Clearly,
the values of the bi depend on the chosen units and can change under various conditions,
whereas the elements of AE , AI are constants defining the structure of the network, and
independent of any units.
Suppose we take (b1 , . . . , b4 ) = (10.5, 18.3, 8.7, 4). Then with β E , β I = bE , bI we have
in Proposition 3.1 |bE |min / 9 AE 9∞ = 8.7/3 and |bI |min / 9 AI 9∞ = 4/1, so ϑ∞ = 2.9. The
vector x = (6.591, 5.326, 13.26, 1.120, 2.253, 2.789) satisfies the constraints exactly. The
rounded vector [x] = (7, 5, 13, 1, 2, 3) is in the set C(δ) defined by (3.1) for any δ > 0.172.
3.2
Effect of tolerances on the optimality of x∗
With the constraints x1 − x2 = 1, x1 + x2 = 4,and x1 , x2 > 0, C(0) is a 0-dimensional
polytope in R2 , the point (2.5, 1.5). However, introducing the tolerance δ = 0.05 turns
the equalities into inequalities and C(0.05) = {0.95 6 x1 − x2 6 1.05, 3.8 6 x1 + x2 6 4.2}
becomes 2-dimensional. Apart from the change in dimension, C(0.05) also contains the
point (2.55, 1.55) at which G assumes a value greater than G∗ = G(2.5, 1.5), its maximum
over C(0). This must be taken into account, since concentration refers to the vectors in
C(δ), not those in C(0). The following lemma shows that the amount by which the value
of G can exceed G∗ due to the widening of the domain C(0) to C(δ) is bounded by a linear
function of δ; it generalizes Prop. II.2 of [OG16] for the ordinary entropy H:
15
Lemma 3.1 Let (λE , λBI ) be the vector of Lagrange multipliers in (2.4) corresponding to
the solution G∗ = G∗ (0), x∗ = x∗ (0) of the maximization problem (2.1). Define
Λ∗ , |λE | · |bE | + λBI · |bBI |,
Λ∗ > G ∗ .
Then with δ > 0, for any ν ∈ C(δ)
G(ν) 6 G∗ + Λ∗ δ − nD(f kχ∗ ),
P
where n = i νi , f is the frequency vector corresponding to ν, and χ∗ is the density vector
corresponding to x∗ .
The upper bound on G(ν) is at least (1 + δ)G∗ − nD(f kχ∗ ). When δ = 0 the lemma
says simply that G∗ is the maximum of G over C(0). The D(·k·) term is positive, and
equals 0 iff ν = αx∗ for some α > 03.2 . Leaving aside that this is possible only for special
x∗ and α, Lemma 3.1 says that if the resulting ν is in C(δ), then G(ν) = αG∗ 6 G∗ + Λ∗ δ,
i.e. the allowable α is limited by δ. Also, if we have even one equality constraint, δ limits
the size of the allowable α even further.
3.3
Scaling of the data and bounds on the allowable sums
We establish a fundamental property, 2.1 in §2.1, of maximizing the generalized entropy
G: if the problem data b is scaled by the factor c > 0, all aspects of the solution scale by
the same factor.
∗ maximizes G(x) under the linear
Proposition 3.2 Suppose that the relaxed count vector xP
constraints AE x = bE , AI x 6 bI , which also imply that i xi is between the bounds s1 , s2 .
Let c > 0 be any constant. Then the vector cx∗ maximizes G(x) under the scaled constraints
P
AE x = cbE , AI x 6 cbI , the maximum value of G is cG(x∗ ), and the new bounds on i xi
are cs1 , cs2 .
How do s1 and s2 , defined in (1.2), depend on the structure matrices AE , AI and the
data bE , bI ? In general, the problem of bounding s1 or s2 doesn’t have a simple answer:
by scaling the variables, any linear program whose objective function is a positive linear
combination of the variables can be converted to one where the objective function is simply
the sum of the variables. But in some special cases we can derive simple bounds on s1 and
s2 :
Proposition 3.3 Bounds on the sums s1 and s2 .
1. If there are some equality constraints, then s1 > kbE k1 / 9 (AE )T 9∞ . (This bound
can only increase if there are also inequalities.)
3.2
The only way the density vectors can be equal is if the un-normalized vectors are proportional.
16
I
2. Suppose allP
of AE , AI , bE
P, b I areI > 0, and EeachI xi occurs in at least one constraint.
E
E
Then s2 6 i bi /αi + i bi /αi , where αi , αi , is the smallest non-zero element of
row i of AE , respectively AI , if that element is < 1, and 1 otherwise.
Recall from §1 that “each xi occurs in at least one constraint” is a necessary condition
for the problem to be bounded. The proposition applies to Example 3.1: we find that
s1 > (b1 + b2 + b3 )/2 and s2 6 b1 + b2 + b3 + b4 .
3.4
The optimal count vector ν ∗
Given the relaxed optimal count vector x∗ , we construct from it a count vector ν ∗ which
is a reasonable approximation to the integral vector that solves problem (1.1), in the sense
that (a) its sum is close to that of x∗ , and (b) its distance from x∗ is small in `1 norm.
These properties will be needed in §4 and §5. We will require ν ∗ to sum to n∗ , where
X
n∗ , ds∗ e, s∗ ,
x∗i .
(3.2)
i
For any x > 0, let [x] be the vector obtained by rounding each of the elements of x up or
down to the nearest integer. ν ∗ is obtained from x∗ by a process of rounding and adjusting:
Definition 3.1 ([OG16], Defn. III.1) Given x∗ , form the density P
vector χ∗ = x∗ /s∗
and set ν̃ = [n∗ χ∗ ]. Construct ν ∗ by adjusting ν̃ as follows. Let d = i ν̃i − n∗ ∈ Z. If
d = 0, set ν ∗ = ν̃. Otherwise, if d < 0, add 1 to |d| elements of ν̃ that were rounded down,
and if d > 0, subtract 1 from |d| elements that were rounded up. The resulting vector is ν ∗ .
We will refer to ν ∗ as “the optimal count vector” or “the MaxGEnt count vector” (even
though it is not unique). It sums to n∗ , and does not differ too much from x∗ in norm:
Proposition 3.4 The optimal count vector ν ∗ of Definition 3.1 is such that
X
3m
3m
νi∗ = n∗ , kν ∗ − x∗ k1 6
+ 1, kν ∗ − x∗ k∞ 6 1, kf ∗ − χ∗ k1 6 ∗ .
4
4n
16i6m
There are other approximations to the integral solution of problem (1.1); for example,
simply [x∗ ] achieves smaller norms than ν ∗ : k[x∗ ] − x∗ k1 6 m/2, k[x∗ ] − x∗ k∞ 6 1/2. ([x∗ ]
is the point of Nm that minimizes the Euclidean distance kν − x∗ k2 from x∗ .) But [x∗ ] does
not have the required sum n∗ .
Another, more sophisticated
definition for ν ∗ ,Pwould use the solution of the integer
Pm
m
∗
linear programPminν∈Nm i=1 |νi − x∗i | subject to
P i=1 νi = n . [This is a linear program
because minz i |zi − ci | is equivalent to mina,z i ai subject to ai ≥ zi − ci , ai ≤ zi − ci .]
A ν ∗ better than that of Definition 3.1 would improve the bound in (4.9) below3.3 .
No integral vector can achieve `1 norm smaller than kx∗ − [x∗ ]k1 ; this solution to the linear program
ignores the constraint, and minimizes each term of the objective function individually.
3.3
17
4
Concentration with respect to entropy difference
It is not clear that concentration should occur at all in a situation like the one of Example
1.1. The fact that G has a global maximum G∗ over C(0) is not enough. In this section we
demonstrate that concentration around G∗ does indeed occur, in the sense of the statement
GC of §1, pertaining to entropies η-far from G∗ . This is done in two stages, by Theorem
4.1 in §4.2 and Theorem 4.2 in §4.3.
Consider the count vectors that sum to n and satisfy the constraints. We divide them
into two sets, An , Bn , according to the deviation of their generalized entropy from G∗ : given
δ, η > 0,
An (δ, η) , {ν ∈ Nn ∩ C(δ), G(ν) > (1 − η)G∗ },
(4.1)
Bn (δ, η) , {ν ∈ Nn ∩ C(δ), G(ν) < (1 − η)G∗ }.
Irrespective of the values of δ and η, An (δ, η) ] Bn (δ, η) = Nn ∩ C(δ). Now we discuss the
possible range of n.
E
E
I
I
P We have assumed that the problem constraints A x = b , A x 6 b imply that s1 6
i xi 6 s2 , where the bounds s1 , s2 on the sum of x are found by solving the linear
programs (1.2). So any integral vector that satisfies the constraints exactly, i.e. is in C(0),
must have a sum n between n1 = ds1 e and n2 = bs2 c. We will use a slight modification of
this definition
n1 , ds1 e, n2 , ds2 e.
(4.2)
With n∗ defined by (3.2), we have n1 6 n∗ 6 n2 . We may assume without loss of generality
that n1 6 n2 + 1; otherwise all count vectors sum to a known n, and we reduce to the case
of frequency vectors which was studied in [OG16].
Remark 4.1 There is a certain degree of arbitrariness (or flexibility) in the definitions
of n1 , n2 . Setting n1 = ds1 e, n2 = bs2 c says that the allowable sums are those of count
vectors which belong to C(0); it does not say that the only allowable vectors are those in
C(0). Now it could be argued that after introducing the tolerance δ, the numbers n1 , n2
should be allowed to become functions of δ. However, this would introduce significant
extra complexity. Our definition makes concessions to simplicity by restricting somewhat
the allowable sums, and by slightly adjusting the value of n2 to handle the ‘boundary’ case
bs2 c < s∗ 6 s2 more easily.
Having defined the range of allowable sums n as n1 6 n 6 n2 , we will use the (disjoint)
unions of the sets (4.1) over n ∈ {n1 , . . . , n2 }
P
An1 :n2 (δ, η) , {ν| i νi = n, n1 6 n 6 n2 , ν ∈ C(δ), G(ν) > (1 − η)G∗ } ,
P
(4.3)
Bn1 :n2 (δ, η) , {ν| i νi = n, n1 6 n 6 n2 , ν ∈ C(δ), G(ν) < (1 − η)G∗ } .
Irrespective of δ and η we have
An1 :n2 (δ, η) ] Bn1 :n2 (δ, η) = Nn1 :n2 ∩ C(δ).
18
(4.4)
We note the following relationship among the numbers of realizations of the optimal count
vector ν ∗ and those of the sets An1 :n2 (δ, η) and Bn1 :n2 (δ, η): if ν ∗ ∈ An1 :n2 (δ, η), then
#ν ∗
#Bn1 :n2 (δ, η)
>
1
ε
⇒
#An1 :n2 (δ, η) + # Bn1 :n2 (δ, η)
#An1 :n2 (δ, η)
#An1 :n2 (δ, η)
#(Nn1 :n2 ∩ C(δ))
ν∗
6
1+ε ⇒
>
1
1+ε
> 1 − ε.
(4.5)
w.r.t. realizations, then the
In other words, if the single vector
dominates the set Bn1 :n2
set An1 :n2 dominates the set Nn1 :n2 ∩ C(δ) likewise.
The concentration statement GC in §1 says that given δ, ε, η > 0, there is a number
ĉ = ĉ(δ, ε, η) > 1 such that when the data bE , bI are scaled by any factor c > ĉ, then
ν ∗ ∈ An1 :n2 (δ, η)
and
#ν ∗
1
> .
#Bn1 :n2 (δ, η)
ε
(4.6)
We establish the inequality in (4.6) by finding a lower bound on #ν ∗ in §4.1 and an upper
bound on #Bn1 :n2 (δ, η) in §4.2. Theorem 4.1 presents the ratio of these bounds. Then in
§4.3 we find the concentration threshold ĉ that ensures (4.6); that is given by Theorem 4.2.
Table 4.1 describes our notation for the process of scaling the problem data.
Basic quantities
x∗ 7→ cx∗
s1 , s2 , s∗ 7→ cs1 , cs2 , cs∗
G∗ 7→ cG∗
ϑ∞ 7→ cϑ∞
Derived quantities
ν∗
n1 , n2 , n∗
Table 4.1: The data scaling process b 7→ cb. The symbols x∗ , s1 , s2 , . . . on the left denote quantities before scaling. The symbols ν ∗ , . . . on the right are quantities derived from the scaled basic
quantities.
4.1
Realizations of the optimal count vector
∗
∗
m
In this section we find a lower bound on #ν ∗ = ν1ν ∗+···+ν
where ν ∗ is the m-vector of
∗
1 ,...,νm
Definition 3.1, in terms of quantities related to the generalized entropy.
Like the number of realizations of a frequency vector and its entropy, the number of
realizations #ν of a count vector ν is related to its generalized entropy. Given ν ∈ Nm ,
w.l.o.g. let ν1 , . . . , νk , k > 1 be its non-zero elements; then
√
k
n
1
− 12
G(ν)
G(ν)
e
S(ν)e
6 #ν 6 S(ν)e
,
S(ν) ,
.
(4.7)
√
(k−1)/2
ν1 · · · νk
(2π)
19
This follows immediately from eq. (III.6) in [OG16], or Problem 2.2 in [CK11]; the bounds
hold even when k = 1 and #ν = 1. Since ν ∗ has no 0 elements (Remark 2.1) we can take
k = m in (4.7), so
∗
#ν ∗ > e−m/12 S(ν ∗ )eG(ν ) .
(4.8)
Next we want to bound G(ν ∗ ) in terms of G∗ = G(x∗ ). By Proposition 3.4, kν ∗ −x∗ k∞ 6 1.
If we assume that x∗ > 1, Lemma 2.1 applies to ν ∗ and x∗ and we get
x∗ > 1
G(ν ∗ ) > G∗ −
⇒
X
ln
i
1
1 X 1
m
−
.
−
χ∗i
2
x∗i − 1 s∗ /m − 1
(4.9)
i
Returning to (4.8), it remains to find a convenient lower bound for S(ν ∗ ). Since kν ∗ −
x∗ k∞ 6 1, we can use νi∗ 6 x∗i + 1 in (4.7) to obtain
√
Y
s∗
1
∗
p ∗
.
(4.10)
S(ν ) >
(m−1)/2
(2π)
xi + 1
16i6m
[Another, simpler bound, is obtained by noting that ν1 ν2 · · · νm is maximum when all νi
are equal to n/m. The bound (4.10) is generally better, but can become slightly worse in
some exceptional situations.] Putting (4.10) and (4.9) in (4.8),
√
m
m
1
−m/12 s∗ − 1 Pm
Y
−
e
χ∗i
∗
∗
∗
i=1 x −1
2
s /m−1
i
p
#ν ∗ >
eG
e
∗
(m−1)/2
(2π)
xi + 1
, if x∗ > 1
(4.11)
i=1
∗
, C0 (x∗ )eG .
The form of C0 (x∗ ) is convenient for scaling according to Table 4.1.
Remark 4.2 On the condition x∗ > 1. It is certainly possible to formulate MaxGEnt
problems whose solutions have some elements that are smaller than 1, in fact arbitrarily
close to 0, and thus invalidate (4.9) and (4.11). Here however we are dealing with ‘large’
problems, where x∗ is scaled by c > 1 for concentration to arise; see Theorem 4.2 below. So
one way to deal with such problem formulations is to take as “the problem” a certain prescaling of the original, one might say pathological, problem. Nevertheless, if one wanted
to avoid the x∗ > 1 issue entirely, one could use a weaker bound than (4.9) not subject to
this restriction; see, for example, Remark 4.3.
Remark 4.3 We compare the bound (4.9), derived for count vectors, to one adapted from
a bound for density vectors. In [OG16], proof of Proposition III.1, we derived the bound
3m
3m
∗
∗
H(f ) > H(χ ) − ∗ ln(m − 1) − h
(4.12)
8s
8s∗
20
where h(·) is the binary entropy function; there we had n in place of s∗ . [This is based on
the bound |H(χ) − H(ψ)| 6 21 kχ − ψk1 ln(m − 1) + h 21 kχ − ψk1 ; see [CK11] problem
3.10, or [Zha07]. An improved version, using both the `1 and `∞ norms is in [Sas13].]
By multiplying both sides of (4.12) by n∗ and then using the fact that n∗ H(f ∗ ) = G(ν ∗ ),
n∗ > s∗ , and s∗ G(χ∗ ) = G(x∗ ) , G∗ , we obtain
3m
3m
∗
∗
∗
ln(m − 1) − s h
.
(4.13)
G(ν ) > G −
8
8s∗
One way to compare the bounds (4.9) and (4.13) is to ask how the right-hand sides, apart
from the G∗ term, behave under scaling
P of the problem by c (§3.3): we see that as c
increases, the r.h.s. of (4.9) tends to − i ln(1/χ∗i ) while the r.h.s. of (4.13) goes to −∞.
4.2
Realizations of the sets with smaller entropy
Here we derive upper bounds on the number of realizations of the sets Bn (δ, η) and
Bn1 :n2 (δ, η). By combining them with the lower bound on #ν ∗ of §4.1, we establish our
first main result, Theorem 4.1.
From (4.1) and (4.7),
X
X
∗
#Bn (δ, η) 6
S(ν),
S(ν)eG(ν) 6 e(1−η)G
ν∈Nn ∩C(δ), G(ν)<(1−η)G∗
ν∈Nn
where in going from the 1st to the 2nd inequality we ignored all the constraints. Using
(4.7) and proceeding as in [OG16], proof of Lemma III.1,
√
m
X
X
m
n
1
(1−η)G∗
#Bn (δ, η) 6 e
√
(k−1)/2
k (2π)
ν1 · · · νk
ν1 +···+νk =n
k=1
√
Z
m
X
m
n
dx1 · · · dxk
∗
6 e(1−η)G
√
(k−1)/2
x
+
·
·
·
+
x
=
n
1
k
k (2π)
x1 · · · xk
k=1
x1 , . . . , xk > 0
(4.14)
√
m
k/2
X
m
n
π
∗
= e(1−η)G
nk/2−1
k (2π)(k−1)/2 Γ(k/2)
k=1
m
X
p
m
n k/2
1
(1−η)G∗
= e
.
2π/n
2
Γ(k/2)
k
k=1
m/2
We show in the Appendix §A that the sum in the last line above is bounded by (n/2)
Γ(m/2) 1+
p
p
m
m
m/n . This is better than the 4 1 + n/4
bound for the same sum obtained in
[OG16], proof of Lemma III.1, as it is asymptotically tight (m fixed, n → ∞). Using this
improved bound in (4.14),
m
p
(n/2)m/2
∗p
#Bn (δ, η) < e(1−η)G 2π/n
(4.15)
1 + m/n .
Γ(m/2)
21
We now turn to the set Bn1 :n2 (δ, η) defined in (4.3). By (4.15),
X
#Bn (δ, η)
#Bn1 :n2 (δ, η) =
n1 6n6n2
(1−η)G∗
<e
√
m
X
p
1 n m2
2π
√
1 + m/n .
Γ(m/2) n 6n6n n 2
1
2
Bounding the sum in the 2nd line by an integral,
Z s2 +2
X
√ m
√ m
1 √
1 √
1
1
√
n + m 6 m/2
y + m dy
√
m/2
y
n
2
2
s1
n1 6n6n2
√
√ m+1
√ m+1
√
1
1
s2 + 2 + m
s1 + m
= m/2−1
−
m+1
2
where in the first line we have widened the interval of integration from [n1 , n2 + 1] to
[s1 , s2 + 2]; recall the definition (4.2) of n1 , n2 . Therefore
∗
#Bn1 :n2 (δ, η) < C1 (s1 , s2 )e(1−η)cG ,
√
√
√ m+1
√ m+1
√
π
C1 (s1 , s2 ) ,
s
+
2
+
m
−
s
+
m
,
2
1
(m + 1)2(m−3)/2 Γ(m/2)
(4.16)
where the sums s1 , s2 have been defined in (1.2).
By combining (4.11) and (4.16) we arrive at our first main result, a lower bound on
the ratio of the number of realizations of the optimal count vector ν ∗ to those of the set
Bn1 :n2 (δ, η), of count vectors with generalized entropy η-far from G∗ = G(x∗ ):
Theorem 4.1 Given structure matrices AE , AI and data vectors bE , bI , let (x∗ , s1 , s2 ) be
the optimal solution to problem (1.1), (1.2). Assume that x∗ > 1; recall Remark 4.2. Then
for any δ, η > 0,
#ν ∗
(m + 1)e−m/12 Γ(m/2) C2 (x∗ )C4 (x∗ ) ηG∗
e
>
m
#Bn1 :n2 (δ, η)
C3 (s1 , s2 )
2π 2
where the constants are
√
χ∗
p ∗i
,
xi + 1
16i6m
1 X
1
m
C4 (x∗ ) = exp −
−
,
2
x∗i − 1 s∗ /m − 1
16i6m
m+1
√
√
√
√ m+1
C3 (s1 , s2 ) =
m + s2 + 2
−
m + s1
.
C2 (x∗ ) =
s∗
Y
One use of the theorem is when the problem is already ‘large’ enough and doesn’t
require further scaling. Then one may substitute appropriate values for δ and η and see
what kind of concentration is achieved. Note that the concentration tolerance ε does not
appear in the theorem.
22
4.3
The scaling factor needed for concentration
What happens to the lower bound of Theorem 4.1 as the size of the problem increases? In
this section we establish Theorem 4.2, our first concentration result, which shows that the
bound can exceed 1/ε for any given ε > 0.
Introducing into the bound of Theorem 4.1 a scaling factor c > 1,
(m + 1)e−m/12 Γ(m/2) C2 (cx∗ )C4 (cx∗ ) cηG∗
#ν ∗
>
e
.
m
#Bn1 :n2 (δ, η)
C3 (cs1 , cs2 )
2π 2
(4.17)
To facilitate scaling, we develop bounds on the functions C2 , C3 , C4 of c appearing above.
First,
√
√
Y
Y
s∗
s∗
χ∗i
χ∗
∗
p ∗
p ∗i
C2 (cx ) = (m−1)/2
> (m−1)/2
, c > 1,
(4.18)
c
c
xi + 1
xi + 1/c
16i6m
16i6m
since χ∗ is invariant under scaling, and the first product above increases as c %. Next,
writing C3 as
p
p
p
m+1
√ m+1
m/c + s2 + 2/c
−
m/c + s1
,
C3 (cs1 , cs2 ) = c(m+1)/2
it can be shown that the function of c multiplying c(m+1)/2 above decreases as c %4.1 , so
its maximum occurs at c = 1. Thus
√
m+1
√
√
√ m+1
C3 (cs1 , cs2 ) 6 c(m+1)/2
m + s2 + 2
−
m + s1
, c > 1.
(4.19)
Finally, for C4 (cx∗ ),
1
m
1 X
1
1 X
1
1 X
−
>
−
>
−
,
−
∗
∗
∗
∗
2
cxi − 1 cs /m − 1
2
cxi − 1
2
xi − 1
16i6m
16i6m
16i6m
since x∗ > 1 and c > 1, and so
− 12
C4 (cx∗ ) > e
Pm
1
i=1 x∗ −1
i
.
(4.20)
Putting (4.18), (4.19), and (4.20) into (4.17), if x∗ > 1,
#ν ∗
∗
> Bc−m ecηG ,
#Bn1 :n2 (δ, η)
(4.21)
where the constant
√ Q
χ∗
P
s∗ 16i6m √ ∗i
1
xi +1
− 12 m
(m + 1)Γ(m/2)e−m/12
i=1 x∗ −1
i
B ,
e
m
√
√
√
√ m+1
m+1
2π 2
m + s2 + 2
−
m + s1
4.1
After some algebra, its derivative can be shown to be negative if s2 > s1 .
23
is 1. By (4.6), the scaling factor c to be applied to the original problem must be such
that the r.h.s. of (4.21) is > 1/ε, and also such that ν ∗ belongs to An1 :n2 (δ, η). The first
of these requirements translates into
cηG∗ − m ln c > − ln(εB),
(4.22)
If c1 is the largest of the two solutions of the equality version of (4.22)4.2 , the inequality
(4.22) will hold for all c > c1 .
The second requirement on c, that ν ∗ ∈ An1 :n2 (δ, η), which is really ν ∗ ∈ An∗ (δ, η),
has two parts. For the first part we need ν ∗ ∈ C(δ); by Proposition 3.1 this is ensured by
kν ∗ − cx∗ k∞ 6 δcϑ∞ , and since the l.h.s. is 6 1 by Proposition 3.4, this will hold if c > c2
where
1
c2 ,
.
(4.23)
δϑ∞
For the second part we need c to be s.t. G(ν ∗ ) > (1 − η)cG∗ . By Proposition 3.2 and (4.9),
this is ensured by
X
1
1 X
m
1
cG∗ −
ln ∗ −
−
> (1 − η)cG∗
χi
2
cx∗i − 1 cs∗ /m − 1
16i6m
16i6m
X
1 X
1
1
⇐
+
ln ∗ < cηG∗ ,
∗
2
cxi − 1
χi
16i6m
16i6m
X
1 X
1
1
⇐
(4.24)
+
ln ∗ 6 cηG∗ ,
∗
2c
xi − 1
χi
16i6m
16i6m
where the last implication follows from c > 1 and x∗ > 1. So we need c > c3 , the largest
solution of the (quadratic) equation version of (4.24).
Given tolerances δ, ε, η, we have now established how to compute a lower bound ĉ, the
concentration threshold , on the scaling factor required for concentration to occur around
the point ν ∗ or in the set An1 :n2 , to the extent specified by δ, ε, η. This is our second main
result, which establishes the statement GC in §1 concerning deviation from the value G∗ :
Theorem 4.2 With the conditions of Theorem 4.1, for any δ, ε, η > 0, define the concentration threshold
ĉ , max(c1 , c2 , c3 ),
where c1 , c2 , c3 have been defined in (4.22)-(4.24). Then when the data bE , bI is scaled by
a factor c > ĉ, the count vector ν ∗ of Definition 3.1 belongs to the set An1 :n2 (δ, η) and we
have
1
#An1 :n2 (δ, η)
#ν ∗
>
and
> 1 − ε,
#Bn1 :n2 (δ, η)
ε
#(Nn1 :n2 ∩ C(δ))
where n1 = dcs1 e, n2 = dcs2 e, and the sets An1 :n2 , Bn1 :n2 have been defined in (4.3).
4.2
An equation of this type generally has two roots, one small and one large. For example ex /x = 10 has
roots 0.1118 and 3.577.
24
Note that the constraint information AE , bE , AI , bI appears implicitly, via s1 , s2 , and ϑ∞ .
The various sets figuring in the theorem are depicted in Figure 4.1.
kν − x∗ k∞ 6 δϑ∞
Nn1 :n2 ∩ C(δ)
Bn1 :n2 (δ, η)
x∗
ν∗
G(ν) < (1 − η)G∗
An∗ (δ, η)
An1 :n2 (δ, η)
G(ν) > (1 − η)G∗
Figure 4.1: The outer ellipse, the set Nn1 :n2 ∩ C(δ) of count vectors that satisfy the constraints
to within tolerance δ, is partitioned into Bn1 :n2 , shown in gray, and An1 :n2 , the inner white ellipse.
The relationship shown between kν − x∗ k∞ 6 δϑ∞ and An1 :n2 (δ, η) is not the only one possible.
Likewise for x∗ and An∗ (δ, η).
4.3.1
Bounds on the concentration threshold
It is useful to know something about how the threshold ĉ depends on the solution x∗ , G∗ to
the MaxGEnt problem and on the parameters δ, ε, η, without having to solve equations.
We derive some bounds on ĉ with regard to convenience, not tightness4.3 .
If ci > Li , then ĉ = maxi ci > maxi Li . Hence we have the lower bound
ĉ > max
− ln(εB)
ηG∗
,
1
,
δϑ∞
(4.25)
since c1 must be bigger than the first term on the r.h.s., and c2 equals the second. As
intuitively expected, the bound says that the smaller δ, ε, or η are, the more scaling we
need. By looking at the expression for B after (4.21), we see that the same holds the
farther apart the bounds s1 , s2 on the possible sums are from each other; this accords with
intuition, and we discuss it further in Example 4.2.
Next, if ci 6 Ui , then maxi ci 6 maxi Ui . So
sP
m
∗
2m
m − ln(εB) ln(εB) 1
i=1 1/(xi − 1)
ĉ 6 max
ln
−
,
,
,
(4.26)
ηG∗
ηG∗
ηG∗ δϑ∞
2ηG∗
where the expressions on the r.h.s. are upper bounds on c1 , c2 , c3 , respectively, as shown
in the Appendix4.4 . The upper bound says that the larger G∗ is, the less scaling we need;
likewise for the elements of x∗ . Both of these implications agree with intuition. Further
illustrations of the bounds (4.25) and (4.26) are in Example 4.1.
4.3
4.4
The bounds still require knowing the solution x∗ to the MaxGEnt problem.
Concerning the last expression, recall our assumption x∗ > 1 and Remark 4.2.
25
4.4
Examples
We give two examples. The first continues Example 3.1, illustrates the bounds on the concentration threshold, and points out a, at first sight, surprising behavior of the threshold.
The second example illustrates an intuitively-expected relationship between concentration
and the bounds s1 , s2 .
Example 4.1 Returning to Example 3.1, we find
s1 = 21.5, s2 = 37.5, x∗ = (6.591, 5.326, 13.26, 1.120, 2.253, 2.789), s∗ = 31.34, G∗ = 47.53.
Thus ν ∗ = (7, 5, 14, 1, 2, 3) and n∗ = 32. Also, ϑ∞ = 2.9. Table 4.2 shows what happens
when the problem data b is scaled by the factor ĉ dictated by the given δ, ε, η. [We don’t
use a special notation for the quantities appearing in the unscaled vs. the scaled problem,
so whenever we write x∗ , ν ∗ , bE , etc. a scaling factor, which could be 1, is implied.]
η
0.05
0.02
0.01
ĉ
34.48
91.27
191.9
0.05
0.02
0.01
40.25
106.8
222.9
δ ∈ [0.01, 1], ε = 10−9
b
n1
n∗
(362.1,631.0,300.0,137.9)
880
1081
(958.3,1670,794.0,365.1)
2328 2861
(2015,3512,1670,767.7)
4894 6015
δ ∈ [0.01, 1], ε = 10−15
(422.7,736.6,350.2,161.0)
1027 1262
(1121.3,1954.3,929.1,427.2)
2724 3347
(2340.2,4078.6,1939.0,891.5) 5684 6985
n2
1294
3423
7197
ν∗
(227,184,457,39,78,96)
(602,486,1210,102,206,255)
(1265,1022,2545,215,433,535)
1510
4005
8358
(266,214,534,45,91,112)
(703,569,1416,120,241,298)
(1469,1187,2955,250,502,622)
Table 4.2: Scaling of the problem of Example 3.1 for the given δ, ε, η.
With respect to the discrete solution, in the first row of Table 4.2 for example, we
have kx∗ − ν ∗ k∞ = 0.370. Further, ν ∗ satisfies the equality constraints with tolerance
kAE ν ∗ − bE k∞ / min |bE | = 0.0033 and the inequality constraints with tolerance 0. We
see that the scaling factor ĉ is quite sensitive to η and rather insensitive to ε; this can
be surmised from (4.25). One way to interpret the scaling is as a change in the scale of
measurement of the data b, e.g. a change in the units. Then scaling by a larger factor
means choosing more refined units, and the above results show that the concentration
increases, as intuitively expected.
With respect to the bounds (4.25) and (4.26) on the threshold ĉ, for the first row of
the table with δ = 0.01, they yield ĉ ∈ [34.48, 41.87]. For δ ∈ [0.02, 0.05] they yield ĉ ∈
[25.1, 41.87]. For the second row, the bounds give ĉ ∈ [62.8, 116.2] for any δ ∈ [0.01, 0.05].
Now suppose that the problem data is pre-scaled by 34.5. Then for the first row the
bounds say that ĉ ∈ [1.0, 1.0], i.e. no further scaling is needed. For the second row,
Theorem 4.2 gives ĉ = 2.39 and the bounds give ĉ ∈ [2.23, 2.55]. So the original problem
had a threshold ĉ = 91.27, but when scaled by 34.5, the threshold becomes only ĉ = 2.39 <
91.27
34.5 = 2.64. Apparently, unlike the rest of the problem (Proposition 3.2), the concentration
26
threshold does not behave linearly with scaling: ĉ (34.5 × problem) < 34.5ĉ (problem). The
explanation for this at first sight disconcerting behavior is two-fold: first, Theorem 4.2 does
not say that ĉ is the minimum required scaling factor for a given problem; second, there
are many approximations involved in the derivation of ĉ, and many get better as the size
of the problem increases.
Example 4.2 Intuition says that the bounds s1 , s2 on the possible sums of the admissible
count vectors have something to do with concentration: if they are wide, concentration
should be more difficult to achieve. Suppose that, somehow, the MaxGEnt vector x∗
from which ν ∗ is derived remains fixed; then the wider the range s1 , s2 allowed by the
constraints, the larger should be the scaling factor required for ν ∗ to dominate. The bound
(4.25) agrees with this, due to the expression for B after (4.21). We now give a simple
situation in which the difference between s1 and s2 can increase while x∗ remains fixed.
Consider a 2-dimensional problem with box constraints b1 6 x1 6 b2 , b3 6 x2 6 b4 ,
depicted in Fig. 4.2. Then s1 = b1 + b3 , s2 = b2 + b4 and G is maximum at the upper right
corner of the box (Proposition 2.3). If we reduce b1 , b3 to b01 , b03 , the lower left corner of
the box moves down and to the left while the upper right corner remains fixed, as shown
in the figure. Thus we widen the bounds s1 , s2 while leaving s∗ , G∗ unchanged, and the
x2
kx − x∗ k 6 δϑ∞
kx − x∗ k 6 δϑ0∞
(b2 , b4 ) = x∗
(b1 , b3 )
(b01 , b03 )
x1
Figure 4.2: Reducing s1 while leaving s2 and x∗ unchanged.
problem with the new box constraints requires more scaling than the original problem.
The construction generalizes immediately to m dimensions, see §2.3.
5
Concentration with respect to distance from the MaxGEnt vector
In this section we provide results analogous to those of §4, but with the sets A, B formulated
in terms of the distance of their elements from the optimal vector x∗ , as measured by the
`1 norm. This is a more intuitive measure than difference in entropy. There are three main
results: Theorems 5.1 and 5.2, analogues of Theorems 4.1 and 4.2, and Theorem 5.3, an
27
optimized version of Theorem 5.2 that does not require specifying a δ. In various places
we reuse results and methods from §4, so the presentation here is more succinct.
For given n and δ > 0, we want to consider the count vectors in Nn that lie in C(δ)
and whose distance from x∗ is no more than ϑ > 0 in `1 norm, and those that lie in C(δ)
but are farther from x∗ than ϑ in `1 norm. The situation is less straightforward than with
frequency/density vectors. First, given two real m-vectors, the norm of their difference can
never be smaller than the difference of their norms, so it does not make sense to require
that this norm be too small5.1 . Second, we will be considering norms that can be large
numbers, especially after scaling of the problem, so it will not do to consider a fixed-size
region around x∗ . For these reasons, we define for ϑ > 0
An (δ, ϑ) , {ν ∈ Nn ∩ C(δ), kν − x∗ k1 6 |n − s∗ | + min(n, s∗ )ϑ},
Bn (δ, ϑ) , {ν ∈ Nn ∩ C(δ), kν − x∗ k1 > |n − s∗ | + min(n, s∗ )ϑ}.
(5.1)
This is more complicated that the definition for frequency vectors in [OG16], but here ϑ is
again a small number < 1. If n were equal to s∗ , (5.1) would say that the density vectors
f and χ∗ are such that kf − χ∗ k1 is 6 ϑ in An and > ϑ in Bn . In general, (5.1) says that
the norm of ν − x∗ is close to |n − s∗ |: if n 6 s∗ , the bound is s∗ − (1 − ϑ)n, and if n > s∗
it is n − (1 − ϑ)s∗ .
We will consider the (disjoint) unions of the sets (5.1) over n ∈ {n1 , . . . , n2 }, with n1 , n2
given by (4.2):
X
An1 :n2 (δ, ϑ) , ν |
νi = n, n1 6 n 6 n2 , ν ∈ C(δ), kν − x∗ k1 6 |n − s∗ | + min(n, s∗ )ϑ ,
i
X
Bn1 :n2 (δ, ϑ) , ν |
νi = n, n1 6 n 6 n2 , ν ∈ C(δ), kν − x∗ k1 > |n − s∗ | + min(n, s∗ )ϑ .
i
(5.2)
∩ C(δ), the set of count vectors that sum to a
For any δ, ϑ, these two sets partition Nn1 :n2
number between n1 and n2 and lie in C(δ).
With these definitions, we will establish an analogue of (4.6) in §4: given δ, ε, ϑ > 0,
there is a concentration threshold ĉ = ĉ(δ, ε, ϑ) s.t. if the problem data bE , bI is scaled by
any factor c > ĉ, then the MaxGEnt count vector ν ∗ is in the set An1 :n2 (δ, ϑ) and has at
least 1/ε times the realizations of all vectors in the set Bn1 :n2 (δ, ϑ):
ν ∗ ∈ An1 :n2 (δ, ϑ)
and
#ν ∗
1
> .
#Bn1 :n2 (δ, ϑ)
ε
(5.3)
There is one important difference with §4, that here the tolerances δ and ϑ cannot be
chosen independently of one another, they must obey a certain restriction.
5.1
In the case of frequency vectors, this lower bound is 0. See Proposition 5.1 for more details.
28
Remark 5.1 G∗ is the maximum of G over the domain C(0), with no tolerances on the
constraints. As we said in §3.1, a tolerance δ > 0 widens this domain to C(δ), may move
the vector that maximizes G from x∗ (0) to x∗ (δ), and may change the maximum value
from G∗ (0) to G∗ (δ). Here we are looking for concentration in a region of size ϑ around the
point x∗ . If δ is too large, we cannot expect such a region to dominate the count vectors
in C(δ) w.r.t. the number of realizations, since x∗ (δ) may even lie inside the set B(δ, ϑ);
by Proposition 2.3, it already lies on the ‘boundary’ of C(0). If ϑ is given, concentration
in A(δ, ϑ) requires an upper bound on the allowable δ; see (5.8) below.
In the setting of §4 there is no limitation on the magnitude of δ with respect to that of
η. It is perfectly fine if the set An (δ, η) contains ν with G(ν) > G∗ (0), but not if Bn (δ, η)
does. But Bn (δ, η) can’t contain any such ν by its definition (4.1): if there are any such ν,
all of them have to be in An (δ, η).
5.1
Realizations of the sets far from the MaxGEnt vector
To bound the number of realizations of Bn1 :n2 (δ, ϑ) we need to show that if ν is far from
x∗ , in the kν − x∗ k1 sense, then G(ν) is far from G(x∗ ). To simplify the notation, in this
section we denote x∗ (0), χ∗ (0), G∗ (0) simply by x∗ , χ∗ , G∗ .
We first need an auxiliary relationship between the norm of the difference of two real
vectors and the norm of the difference of their normalized versions:
Proposition 5.1 Let k · k be any vector norm, such as k · k1 , k · k2 , k · k∞ etc. Then for
any x, y ∈ Rm and ϑ > 0,
kx − yk > |kxk − kyk| + ϑ
x
ϑ
y
>
−
.
kxk kyk
min(kxk, kyk)
⇒
What we want to show about G(ν) and G∗ follows by taking Lemma 3.1, bounding the
divergence term D(·k·) in terms of the `1 norm, and then using Proposition 5.1 with the
`1 norm and min(kνk, kx∗ k)ϑ in place of ϑ:
Lemma 5.1 Given δ > 0 and ϑ > 0, with the notation of Lemma 3.1, for any count vector
ν ∈ C(δ) with sum n,
kν − x∗ k1 > |n − s∗ | + min(n, s∗ )ϑ
where
γ∗ ,
In general, γ ∗ >
1
1 − β∗
ln
,
4(1 − 2β ∗ )
β∗
1
2
and
1−χ∗max
2
β∗ ,
⇒
max
G(ν) 6 G∗ + Λ∗ δ − γ ∗ ϑ2 n
I⊂{1,...,m}
X
X
min
χ∗i , 1 −
χ∗i .
i∈I
6 β ∗ 6 12 . If β ∗ = 1/2, γ ∗ , 1/2.
29
i∈I
The bound on the divergence that we used above, D(pkq) > γ(q)kp − qk21 , is due to
[OW05]. The closeness of the number β(q) to 1/2 can be thought of as measuring how far
away the density vector q is from having a partition5.2 . [BHK14] is also relevant here, as
the authors study inf p D(pkq) subject to kp − qk1 > `. They refer to 1 − β > 1/2, where
β is as in Lemma 5.1, as the “balance coefficient”. Their Theorem 1b provides an exact
value for inf p D(pkq) as a function of 1 − β, q, and `, valid for ` 6 4(1/2 − β); this could
be used in Lemma 5.1, at the expense of an additional condition between, in our notation,
ϑ and β ∗ . They also show that β > 1/2 − qmax /2, where qmax is the largest element of q,
a result which we have incorporated into Lemma 5.1.
We can now proceed to find an upper bound on #Bn1 :n2 (δ, ϑ). Beginning with #Bn (δ, ϑ),
by (5.1) and (4.7)
X
#Bn (δ, ϑ) 6
S(ν)eG(ν) .
ν ∈ Nn ∩ C(δ)
kν − x∗ k1 > |n − s∗ | + min(n, s∗ )ϑ
Applying Lemma 5.1 to G(ν) and, similarly to what we did in §4.2, ignoring the condition
involving the norm in the sum as well as the intersection with C(δ),
X
∗
∗
∗ 2
S(ν).
#Bn (δ, ϑ) 6 eG +Λ δ−γ ϑ n
ν∈Nn
The sum above is identical to that in the expression for #Bn (δ, η) given at the beginning
of §4.2, so following the development that led to (4.15),
#Bn (δ, ϑ) 6
p
p
m ∗ ∗ ∗ 2
(n/2)m/2
2π/n
1 + m/n eG +Λ δ−γ ϑ n .
Γ(m/2)
Compare with (4.15). Consequently,
X
#Bn1 :n2 (δ, ϑ) =
#Bn (δ, ϑ)
√
n1 6n6n2
X
√ m
1 √
2π
∗ 2
G∗ +Λ∗ δ
√
e
m + n e−γ ϑ n
m/2
n
2
Γ(m/2)
n1 6n6n2
√
2π
∗
∗
6 m/2
eG +Λ δ
2
Γ(m/2)
√
√
√ m+1 −γ ∗ ϑ2 (s∗ +1)
√ m+1 −γ ∗ ϑ2 s1
2
e
+
e
s2 + 2 + m
s∗ + 2 + m
(m + 1)
(5.4)
where the inequality implied in the last line is derived in the Appendix. This bound on
#Bn1 :n2 (δ, ϑ) is to be compared with the bound (4.16) on #Bn1 :n2 (δ, η).
6
5.2
In the sense of the NP-complete problem Partition.
30
Combining (4.11) with (5.4) we obtain a lower bound on the ratio of numbers of realizations analogous to that of Theorem 4.1:
Theorem 5.1 Given structure matrices AE , AI and data vectors bE , bI , let (x∗ , s1 , s2 ) be
the optimal solution of problem (1.1). Assume that x∗ > 1; recall Remark 4.2. Then for
any δ, ε, ϑ > 0,
(m + 1)Γ(m/2)e−m/12 C2 (x∗ )C4 (x∗ ) γ ∗ ϑ2 s1 −Λ∗ δ
#ν ∗
>
e
,
#Bn1 :n2 (δ, ϑ)
C30 (s∗ , s1 , s2 )
2π m/2
where the constants C2 (x∗ ), C4 (x∗ ) are the same as in (4.17),
√
√
√ m+1 −γ ∗ ϑ2 (s∗ +1)
√ m+1 −γ ∗ ϑ2 s1
C30 (s∗ , s1 , s2 ) =
s2 + 2 + m
e
+
s∗ + 2 + m
e
,
and Λ∗ , γ ∗ have been defined in Lemmas 3.1 and 5.1 respectively.
The lower bound will not be useful if the exponent γ ∗ ϑ2 s1 − Λ∗ δ is not positive. We
elaborate on this in §5.2. Also, like Theorem 4.1, the theorem says nothing about how
large the bound is for a given problem. This is the job of Theorem 5.2.
5.2
Scaling and concentration around the MaxGEnt count vector
As we did in §4.3, we now investigate what happens to the lower bound of Theorem 5.1
when the problem data b is scaled by a factor c > 1. The end results are the concentration
Theorems 5.2 and 5.3 below.
Table 4.1 described how scaling the data affects the quantities appearing in the bound,
except for Λ∗ , which is new to §5. Scaling b has the effect x∗ 7→ cx∗ , and from (2.4) in
§2.4 we see that the Lagrange multipliers remain unchanged5.3 . Then the definition of Λ∗
in Lemma 3.1 shows that the end result of scaling is Λ∗ 7→ cΛ∗ . This and s1 7→ cs1 imply
that scaling by c multiplies the exponent γ ∗ ϑ2 s1 − Λ∗ δ in Theorem 5.1 by c. The effect of
scaling on C2 and C4 is given by (4.18) and (4.20), and finally, since c > 1,
C30 (cs∗ , cs1 , cs2 ) <
√
√
√
√
m+1
∗ 2
∗ 2 ∗
∗ 2
c 2 ( s2 + 2 + m)m+1 e−γ ϑ e−cγ ϑ s + ( s∗ + 2 + m)m+1 e−cγ ϑ s1
√
√
√
√
m+1
∗ 2
∗ 2 ∗
∗ 2
6 c 2 e−cγ ϑ s1 ( s2 + 2 + m)m+1 e−γ ϑ e−γ ϑ (s −s1 ) + ( s∗ + 2 + m)m+1
m+1
∗ 2
, c 2 e−cγ ϑ s1 C300 .
(5.5)
In conclusion, when the data b is scaled by the factor c > 1, Theorem 5.1 says that if
x∗ > 1, then
#ν ∗
B0
∗ 2
∗
> 00 c−m e(2γ ϑ s1 −Λ δ)c ,
(5.6)
#Bn1 :n2 (δ, ϑ)
C3
This also follows from expression (A.8) in the proof of Lemma 3.1, for G∗ in terms of the multipliers
and G∗ 7→ cG∗ .
5.3
31
where C300 is defined in (5.5) and
B0 ,
(m + 1)Γ(m/2)e−m/12 √ ∗ − 12
s e
2π m/2
Pm
1
i=1 x∗ −1
i
Y
p
16i6m
χ∗i
.
x∗i + 1
(5.7)
(5.6) and (5.7) are to be compared with (4.21).
Recalling Remark 5.1, an important consequence of (5.6) is that if concentration is to
occur the tolerances δ and ϑ must satisfy
ϑ2 >
Λ∗
δ.
2γ ∗ s1
(5.8)
This can be ensured by choosing small enough δ for the given ϑ, or large enough ϑ for the
given δ. (The results of this paper do not immediately translate to the frequency vector
case, but (5.8) can be compared with the similar condition in Theorem IV.2 of [OG16].)
By (5.6), the concentration statement (5.3) will hold if the scaling factor c is such that
(2γ ∗ ϑ2 s1 − Λ∗ δ)c − m ln c > ln
C300
.
εB 0
(5.9)
This inequality is of the same form as (4.22), and will hold for all c greater than the larger
of the two roots of the equality version of it.
As in §4.3, we also need ν ∗ to be in the set An1 :n2 (δ, ϑ) of (5.2), more specifically in
An∗ (δ, ϑ). For this, we must first have ν ∗ ∈ C(δ); this is ensured by c > c2 , with c2 as
in (4.23). Second, by the definition (5.1) of An (δ, ϑ), we need kν ∗ − x∗ k1 6 |n∗ − s∗ | +
min(n∗ , s∗ )ϑ; by Proposition 3.4 this will hold if ϑ > (3m/4 + 1)/(cs∗ ).
We have now established the desired analogue of Theorem 4.2, and proved the statement
GC of §1, in terms of distance from the MaxGEnt vector x∗ :
Theorem 5.2 With the same conditions as in Theorem 5.1, suppose that the tolerances
δ, ϑ satisfy (5.8), where Λ∗ , γ ∗ have been defined in Lemmas 3.1 and 5.1. Let
c1 ,
3m/4 + 1
,
ϑs∗
c2 ,
1
,
δϑ∞
and given ε > 0, let c3 be the largest root c of the equality version of (5.9). Finally, define
the concentration threshold
ĉ , max(c1 , c2 , c3 ).
Then when the data bE , bI is scaled by any c > ĉ, the MaxGEnt count vector ν ∗ of
Definition 3.1 belongs to the set An1 :n2 (δ, ϑ) of (5.2), specifically to An∗ (δ, ϑ), and is such
that
1
#An1 :n2 (δ, ϑ)
#ν ∗
>
and
> 1 − ε.
#Bn1 :n2 (δ, ϑ)
ε
#(Nn1 :n2 ∩ C(δ))
32
The second inequality in the claim of the theorem follows from the first by (4.5) in §4,
which holds whether the sets An1 :n2 and Bn1 :n2 are defined as they were in §4 or as they
were defined here. As in Theorem 4.2, the constraint information AE , bE , AI , bI appears
implicitly in Theorem 5.2, via s1 , s2 , and ϑ∞ . Bounds on the concentration threshold can
be derived similarly to §4.3.1.
Finally, Fig. 5.1 depicts the various sets involved in the definition of the threshold ĉ
appearing in the theorem.
Nn1 :n2 ∩ C(δ)
kν − x∗ k∞ 6 δϑ∞
Bn1 :n2 (δ, ϑ)
kν − x k1 > |n − s∗ | + min(n, s∗ )ϑ
∗
ν∗
x∗
An∗ (δ, ϑ)
An1 :n2 (δ, ϑ)
kν − x∗ k1 6 |n − s∗ | + min(n, s∗ )ϑ
Figure 5.1: Concentration around ν ∗ w.r.t. `1 norm. The MaxGEnt vector ν ∗ has 1/ε times
more realizations than the entire set Bn1 :n2 (δ, ϑ), shown in gray. The relationship we show between
kν − x∗ k∞ 6 δϑ∞ and An1 :n2 (δ, ϑ) is not the only one possible; likewise for x∗ and An∗ (δ, ϑ).
From the definition of ĉ in Theorem 5.2 we see that as δ increases, the constants c2
and c3 behave in opposite ways: c2 decreases but c3 increases. If one cares only about the
tolerances ε and ϑ, and does not care to specify a particular δ, this opens the possibility
of reducing ĉ by choosing δ so as to minimize the largest of c2 , c3 :
Theorem 5.3 Given ε, ϑ, suppose that
Λ > mϑ∞
∗
2
ϑ <
and
Λ∗
p
s∗ /m + 1 1
,
γ ∗ ϑ∞ s1
ε1/m
where the various quantities are as in Theorem 5.2. If so, the equation for δ
2γ ∗ ϑ2 s1 1
C 00
Λ∗
+ m ln δ = ln 3 0 +
− m ln ϑ∞
ϑ∞ δ
εB
ϑ∞
has a root δ0 ∈ 0, 2γ ∗ ϑ2 s1 /Λ∗ , and we define
ĉ , max
1
3m/4 + 1
,
.
δ 0 ϑ∞
ϑs∗
Then when the data bE , bI is scaled by any c > ĉ, the MaxGEnt count vector ν ∗ of
Definition 3.1 belongs to the set An∗ (δ0 , ϑ) of (5.1), and is such that
#ν ∗
1
>
#Bn1 :n2 (δ0 , ϑ)
ε
and
#An1 :n2 (δ0 , ϑ)
> 1 − ε.
#(Nn1 :n2 ∩ C(δ0 ))
33
In this situation a simple lower bound on the concentration threshold ĉ is
3m
H(χ∗ ) 1
ĉ > max
,
2γ ∗ ϑ∞ ϑ2 4s∗ ϑ
(5.10)
where for the first expression we used the upper bound on δ0 and Λ∗ > s∗ H(χ∗ ). The ratio
H(χ∗ )/γ ∗ is small for imbalanced distributions χ∗ , e.g. with a single dominant element,
in which case γ ∗ is large, and approaches 2 ln m for perfectly balanced ones. The bound
(5.10) says that ĉ increases with ϑ−2 , and this can be seen in Example 5.2 below.
5.3
Examples
The first two examples illustrate Theorems 5.2 and 5.3, while the third illustrates the
removal of 0s from the solution mentioned in §2.4 and the ‘boundary’ case in which the
MaxGEnt vector x∗ sums to the maximum allowable s∗ = s2 .
Example 5.1 We return to Example 4.1. Recall that
s1 = 21.5, s2 = 37.5, x∗ = (6.591, 5.326, 13.26, 1.120, 2.253, 2.789), s∗ = 31.34, G∗ = 47.53.
We have ϑ∞ = 2.9, γ ∗ = 0.5, and Λ∗ = G∗ = 47.53. The constraint (5.8) on δ and ϑ is
ϑ2 > 1.864δ. This means that if we want small ϑ, we must have a correspondingly small
δ, as we commented after (5.8). Table 5.1 lists various values of ĉ(δ, ε, ϑ) obtained from
Theorem 5.2.
ĉ
δ
0.001
10−4
10−5
0.001
10−4
10−5
0.001
10−4
10−5
10−4
10−5
10−5
ϑ
0.08
0.07
0.06
0.05
0.03
0.01
0.008
ε = 10−9
862.7
3448
34483
1322
3448
34483
2392
3448
34483
3448
34483
34483
60376
111472
ε = 10−15
989.2
3448
34483
1511
3448
34483
2722
3448
34483
3448
34483
34483
67351
123967
Table 5.1: Scaling of the problem of Example 4.1 for the given δ, ε, ϑ. The threshold ĉ does not
behave smoothly because of the max() in Theorem 5.2.
34
Example 5.2 Consider the same data as in Table 5.1, but with only ε, ϑ specified; we
don’t care about a particular δ, as long as it ensures that ν ∗ ∈ An∗ (δ, ϑ). With δ chosen
automatically by Theorem 5.3, Table 5.2 below shows that the concentration threshold ĉ
is significantly reduced.
ĉ
ϑ
0.08
0.07
0.06
0.05
0.04
0.03
0.01
0.008
ε = 10−9
704.4
933.5
1292
1896
3032
5548
55345
88189
ε = 10−15
793.4
1050
1450
2124
3387
6178
60991
97004
Table 5.2: The threshold ĉ for given ε, ϑ with optimal selection of δ = δ0 . Compare with Table
5.1.
The variation of ĉ with ϑ−2 implied by the lower bound (5.10) is evident.
Example 5.3 Fig. 5.2 shows four cities connected by road segments. We assume that
vehicles travelling from one city to another follow the most direct route, and that there is
no traffic from a city to itself.
4
1
3
2
Figure 5.2: Four cities connected by (bidirectional) road segments. Arrows indicate the constrained
directions.
The number of vehicles in city i is known, ui , which puts upper bounds on the number
that leaves each city; also, from observations we have lower bounds lij on the number of
vehicles on the road segments 2 → 3, 3 → 1, and 3 → 4. From this information we want
to infer how many vehicles travel from city i to city j, i.e. infer the 4 × 4 matrix of counts
0 v12 v13 v14
v21 0 v23 v24
v=
v31 v32 0 v34 .
v41 v42 v43 0
35
So suppose the constraints on v are
X
vii = 0,
vij 6 100, 120, 80, 90, v23 +v24 > 80,
v31 +v41 > 59,
v14 +v24 +v34 > 70,
j
where the last three reflect the “direct route” assumption. Then we have s1 = 139, s2 = 390.
We define the 12-element vector x for the MaxGEnt method as (v12 , v13 , v14 , v21 , v23 , v24 ,
. . . , v43 ). [Note that if we knew that all vehicles in a city leave the city, then we could
define a frequency matrix by dividing the matrix v by u1 + · · · + u4 and thus formulate a
MaxEnt problem.]
The MaxGEnt solution in matrix form is
0
33.333 33.333 33.333
40.0
0
40.0
40.0
v∗ =
27.765 26.118
0
26.118
31.235 29.382 29.382
0
with sum s∗ = s2 = 390 and maximum generalized entropy G∗ = 964.62 = Λ∗ , γ ∗ = 0.5.
So here we have the boundary case in which the sum of x∗ is the maximum possible.
(Problems involving matrices subject to constraints of the above type, for which analytical
solutions are possible, were studied in [Oik12].)
Applying Theorem 5.3 with ϑ = 0.04, ε = 10−15 the ‘optimal’ δ is δ0 ≈ 4.36 · 10−5 and
yields the threshold ĉ = 1166.45. Using a scaling factor c = 1167.0 results in the integral
matrix
0
38900 38900 38900
46681
0
46680 46680
ν∗ =
32401 30479
0
30479
36452 34289 34289
0
with sum n∗ = 455130 ∈ [162213, 455130]. This matrix has at least 1015 times the number
of realizations of the entire set B166213:455130 (4.36·10−5 , 0.04) defined in (5.2). To gain some
appreciation of what this means, it is not easy to determine the size of this set, but just
the particular subset of it
B455129 (0, 0.04) = {ν ∈ C(0), kν − x∗ k1 > 18206}
contains at least 2.768 · 1038 elements5.4 . For comparison, the whole of C(0) contains 1.274 ·
1056 count vectors. (We compute these numbers with the barvinok software, [VWBC05].
For B455129 we get a lower bound by using the stronger constraint |ν1 − x∗1 | > 18205 in
place of kν − x∗ k1 > 18206, which is hard to express.)
5.4
We have |455129 − 1167 · 390| + min(455129, 1167390) · 0.04 = 18206.16.
36
6
Conclusion
We demonstrated an extension of the phenomenon of entropy concentration, hitherto
known to apply to probability or frequency vectors, to the realm of count vectors, whose
elements are natural numbers. This required introducing a new entropy function in which
the sum of the count vector plays a role. Still, like the Shannon entropy, this generalized
entropy can be viewed combinatorially as an approximation to the log of a multinomial
coefficient. Our derivations are carried out in a fully discrete, finite, non-asymptotic framework, do not involve any probabilities, and all of the objects about which we make any
claims are fully constructible. This discrete, combinatorial setting is an attempt to reduce
the phenomenon of entropy concentration to its essence. We believe that this concentration
phenomenon supports viewing the maximization of our generalized entropy as a compatible
extension of the well-known MaxEnt method of inference.
Acknowledgments
Thanks to Peter Grünwald for his comments on a previous version of the manuscript, and
for many useful discussions on the subject.
A
Proofs
Proof of Proposition 2.1
Given a y > x, y can be reached from x by a sequence of steps each of which increases a
single coordinate, and the value of G increases at each step because all its partial derivatives
are positive. (The derivatives are 0 only at points x that consist of a single non-zero element;
a direct proof can be given for that case.)
0
For a more formal proof, we note that the directional derivative
u) of G at any
P G (ξ;
0
m
point ξ is > 0 in any direction u > 0: G (ξ; u) = ∇G(ξ) · u = i ui ln ξ1 +···+ξ
. So any
ξi
move away from ξ in a direction u > 0 will increase G. More precisely, by the mean value
theorem, for any y that can be written as x + u for some u > 0, there is a ξ on the line
segment from x to x + u s.t. G(x + u) − G(x) = ∇G(ξ) · u > 0. Finally, if some element of
u is strictly positive, then ∇G(ξ) · u > 0.
Proof of Proposition 2.2
1. To establish concavity it suffices to show that ∇2 G(x), the Hessian of G, is negative
semi-definite. We find
1
1
1
∇2 G(x) =
Um − diag
,...,
,
(A.1)
x1 + · · · + xm
x1
xm
37
where Um is a m × m matrix all of whose entries are 1. Given x, for an arbitrary
y = (y1 , . . . , ym ) we must have y T ∇2 G(x)y 6 0. To show this, first write ∇2 G(x) as
∇2 G(x) =
x + · · · + x
1
x1 + · · · + xm
1
m
,...,
Um − diag
.
x1 + · · · + xm
x1
xm
Now define ξi = xi /(x1 + · · · + xm ). Then y T ∇2 G(x)y 6 0 is equivalent to
2
(y1 + · · · + ym )2 6 y12 /ξ1 + · · · + ym
/ξm ,
(A.2)
2 /ξ is a convex
where the ξi are > 0 and sum to 1. But for fixed y, y12 /ξ1 + · · · + ym
m
function of ξ = (ξ1 , . . . , ξm ) over the domain ξ 0, and its minimum under the
constraint ξ1 + · · · + ξm = 1 occurs at ξi = yi /(y1 + · · · + ym ). So the least value of
the r.h.s. of (A.2) as a function of ξ1 , . . . , ξm is (y1 + · · · + ym )2 , and this establishes
(A.2).
T 2
For P
a given x we
P see that y ∇ G(x)y is 0 exactly at points y such that for all i,
xi / j xj = yi / j yj , i.e. iff y = cx for some c ∈ R.
2. The fact that the Hessian fails to be negative definite does not imply that G is not
strictly concave; negative definiteness is a sufficient, but not a necessary condition
for strict concavity.
It can be seen that G is not strictly concave because of the scaling or homogeneity
property 2.1 in §2.1: consider the distinct points x and y = 2x; strict concavity would
require G((x + y)/2) > G(x)/2 + G(y)/2, which is not true.
3. Proposition 1.1.2 in Chapter IV of [HUL96] says that a function F (x) is strongly
convex on a convex set C with modulus γ > 0 iff the modified function F (x)− 21 γkxk22
is convex on C. Applying this to our function G, by the proof carried out in part 1,
m
we would have to show that given any x ∈ Rm
+ , for all y ∈ R+
2
2
(y1 + · · · + ym )2 − (y12 /ξ1 + · · · + ym
/ξm ) + γ(y12 + · · · + ym
) 6 0
for the chosen modulus γ > 0. But for any x and any γ > 0, this condition is false
at the point y = γx.
4. By Definition 1.1.1 in [HUL96] Ch. V, §1.1, a convex and positively homogeneous
function F defined over the extended real numbers R ∪ {±∞} is sublinear. If we
define G(·) over all of Rm by setting G(x1 , . . . , xm ) = −∞ if any xi is negative, the
above statement applies to F = −G. Finally, a sublinear function has the property
F (αx + βy) 6 αF (x) + βF (y).
38
Proof of Lemma 2.1
By (2.2), if kx − yk∞ 6 ζ we have G(y) > G(x − ζ). Now we expand G(x − ζ) in a Taylor
series around x. Since G(·) is a twice-differentiable function on the open set x > ζ, if x, x0
are two points in this set, then there is a x̃ = (1 − α)x + αx0 with α ∈ [0, 1], such that
1
G(x0 ) = G(x) + ∇G(x) · (x0 − x) + (x0 − x)T · ∇2 G(x̃) · (x0 − x)
2
(Theorem 12.14 of [Apo74]). Set x0 = x − ζ, so x̃ = x − αζ. Noting that
x + ··· + x
x1 + · · · + xm
1
m
, . . . , ln
,
∇G(x) = ln
x1
xm
1
1
1
∇2 G(x̃) =
Um − diag
,...,
,
x̃1 + · · · + x̃m
x̃1
x̃m
X
m
1
(x0 − x)T · ∇2 G(x̃) · (x0 − x) = ζ 2
,
−
x̃1 + · · · + x̃m x̃i
i
where the second equality is (A.1) in the proof of Proposition 2.2, we find that for any
x>ζ
G(x − ζ) = G(x) − ζ
X
ln
1
x 1 + · · · + x m 1 2 X
m
− ζ
−
, (A.3)
xi
2
xi − αζ
kxk1 /m − αζ
i
i
where we know that the sum of the ζ and the ζ 2 terms on the right is negative. [We
chose to expand around the point x − ζ because then the sign of the terms ∇G(x) · ζ and
ζ T · ∇2 G(x̃) · ζ is known.] Now for fixed x define the function
g(α, ζ) ,
X
16i6m
m
1
−
,
xi − αζ
kxk1 /m − αζ
α ∈ [0, 1], ζ > 0, x > ζ.
(A.4)
This function is > 0 and increasing with α. To see that g(α, ζ) > 0, set ui = xi − αζ
so that kxk1 /m − αζ becomes the arithmetic mean ū of the ui ; then use a fundamental
property of the power means: for any u > 0, and any weights wi summing to 1,
X
wi u−k
i
−1/k
6
i
X
wi ui ,
k>1
(A.5)
i
(see [HLP97], Theorem 16). The desired result follows by choosing all wi = 1/m. To show
that g(α, ζ) increases with α,
X
ζ
∂g
mζ
=
−
2
∂α
(xi − αζ)
(kxk1 /m − αζ)2
i
39
and this is always > 0 by the same power means technique (A.5). [Similarly, ∂ 2 g/∂α2 > 0,
so g(α, ζ) is a convex function of α.] We therefore see that for any ζ > 0
min g(α, ζ) = g(0, ζ)
α∈[0,1]
and
max g(α, ζ) = g(1, ζ).
α∈[0,1]
(A.6)
It now follows from G(y) > G(x − ζ), (A.3), and (A.6) that for any x > ζ and any y s.t.
ky − xk∞ 6 ζ
G(y) > G(x) − ζ
X
i
ln
kxk1 1 2 X 1
m
− ζ
−
.
xi
2
xi − ζ
kxk1 /m − ζ
i
This establishes the lemma. The coefficient of ζ 2 above is > 0 and equals 0 iff all elements
of x are equal ([HLP97], Theorem 16).
Proof of Proposition 2.3
1. Suppose there is a ball around x∗ contained in C(0); then this ball will contain
points x∗ + u for sufficiently small u > 0, and by Proposition 2.1 we will have
G(x∗ + u) > G(x∗ ). But x∗ maximizes G over C(0), so there can be no such x∗ + u;
in other words, C(0) cannot contain any x > x∗ with any strict inequalities.
2. Let z be another point in C(0) such that G(z) = G(x∗ ); by part 1 above, at least one
element of z must be strictly less than the corresponding element of x∗ . But then by
Proposition 2.1 we must have G(z) < G(x∗ ).
Proof of Proposition 3.1
Consider the equality constraints first. Writing them as |AE y −bE | 6 δ|bE |, we see that this
E
E
E
will be satisfied if maxi |AE y − bE |i 6 δ mini |bE
i |, or kA y − b k∞ 6 δ|b |min . Now for any
m
E
E
E
E
E
y ∈ R , A y − b = A (y − x), since x ∈ C(0). Thus kA y − b k∞ = kAE (y − x)k∞ . But
kAE (y − x)k∞ 6 9AE 9∞ ky − xk∞ , where the (rectangular) matrix norm 9 · 9∞ is defined
as the largest of the `1 norms of the rowsA.1 . Therefore, to ensure kAE y − bE k∞ 6 δ|bE |min
it suffices to require that ky − xk∞ 6 δ|bE |min / 9 AE 9∞ , as claimed.
Turning to the inequality constraints, write them as AI (x + y − x) 6 bI + δ|bI |, or
I
A x−bI 6 AI (x−y)+δ|bI |. Since AI x−bI 6 0, this inequality will be satisfied if AI (y−x) 6
δ|bI |. This will certainly hold if maxi (AI (y − x))i 6 δ mini |bIi |, which is equivalent to
kAI (y − x)k∞ 6 δ|bI |min . In turn, this will hold if we require 9AI 9∞ ky − xk∞ 6 |bI |min .
For both types of constraints the final condition is stronger than necessary, but more
so in the case of inequalities. Finally, part 2 of the proposition follows from part 1 since
k[x] − xk∞ 6 1/2.
For any rectangular matrixPA and compatible vector x, kAxk∞ 6 9A 9∞ kxk∞ holds because the l.h.s.
is maxi |Ai. x|. This is 6 maxi j |aij xj | 6 maxi kxk∞ kAi. k1 = kxk∞ 9 A9∞ .
A.1
40
Proof of Proposition 3.2
P
From (2.4) we can write the elements of x∗ in the form x∗j = ( i x∗i ) Ej ,where Ej is an
expression involving the vectors λE , λBI and the matrices AE , ABI . The elements of λE , λBI
are determined by substituting the x∗j into the constraints. Thus the kth equality constraint
leads to an equation of the form
X
x∗i (expression involving the Ej ) = bE
(A.7)
k
i
and similarly for each binding inequality constraint. But the solution of a system of equations of the form (A.7) is unchanged if the x∗i on the l.h.s. and the bE and bBI on the r.h.s.
are both multiplied by the same constant c > 0. This establishes the first claim. The claim
about the maximum of G follows from property 2.1 of G in the list of §2.
Coming to the bounds on x1 + · · · + xm , the fact that they scale with b is just a
property P
of general linear programs. That is, if y is the solution to the
P linear program
m
m
minx∈R
i αi xi subject to Ax 6 b, then cy is the solution to minx∈R
i αi xi subject to
Ax 6 cb. Similarly for the maximum.
Proof of Proposition 3.3
For part 1, given AE x = bE , we have kAE xk1 = kbE k1 . Now, omitting the superscript to
simplify the notation,
kAxk1 = |a11 x1 + · · · + a1m xm | + |a21 x1 + · · · + a2m xm | + · · · + |a`1 x1 + · · · + a`m xm |
6 |a11 + a21 + · · · + a`1 ||x1 | + |a12 + a22 + · · · + a`2 ||x2 | + · · ·
6 9AT 9∞ kxk1 .
Hence 9(AE )T 9∞ kxk1 > kbE k1 , and since x > 0, kxk1 is simply the sum of the xi .
For part 2, any x ∈ Rm satisfying AE x = bE , AI x 6 bI will satisfy AE x 6 bE , AI x 6 bI
as well. Divide each inequality in this system by the smallest non-0 element of the l.h.s.,
if that element is < 1, otherwise leave the inequality as is. Since each xi appears in
some constraint, if we add all the above
the resulting l.h.s. will be
P inequalities
PbyI sides
E
E
I
> x1 + · · · + xm , and the r.h.s. will be i bi /αi + i bi /αi , where the αi are defined as
in the Proposition.
Proof of Proposition 3.4
First, the adjustment performed on ν̃ is always possible: if d < 0 there must be at least |d|
elements of n∗ χ∗ that were rounded to their floors, and if d > 0 to their ceilings. It is clear
that the adjustment makes ν ∗ sum to n∗ . Now suppose that k ∈ N and χ is an m-element
density vector; then kχ sums to k, and the sum of the rounded version [kχ] differs by no
more than m/2 from k. Thus d 6 m/2.
41
For the bound on kν ∗ − x∗ k1 , we first show that kν ∗ − n∗ χ∗ k1 6 3m/4. The adjustment
of ν̃ causes d of the elements of ν ∗ to differ from the corresponding elements of n∗ χ∗ by
< 1, and the rest to differ by 6 1/2, so kν ∗ − n∗ χ∗ k1 6 maxd (d + (m − d)/2) 6 3m/4.
Next,
kν ∗ − x∗ k1 = kν ∗ − s∗ χ∗ k1 6 kν ∗ − n∗ χ∗ k1 + kn∗ χ∗ − s∗ χ∗ k1 6 3m/4 + |n∗ − s∗ |,
since χ∗ sums to 1, and lastly |n∗ − s∗ | < 1 by (3.2).
That kν ∗ − x∗ k∞ 6 1 follows from this last statement and the fact that ν ∗ sums to n∗ .
Finally, the bound on kf ∗ − χ∗ k1 follows from that on kν ∗ − n∗ χ∗ k1 .
Proof of Lemma 3.1
For brevity, in this proof we denote G∗ (0), x∗ (0), χ∗ (0) simply by G∗ , x∗ , χ∗ .
E
E
BI
BI
Given the vector x∗ , set s∗ = x∗1 + · · · + x∗m . Then from (2.4), x∗j = s∗ e−(λ ·A.j +λ ·A.j ) .
Therefore
X
X
BI
x∗i ln x∗i =
x∗i ln s∗ − (λE · AE
· ABI
.i + λ
.i )
i
i
∗
= s ln s∗ −
X
BI
x∗i (λE · AE
· ABI
.i + λ
.i )
i
∗
∗
= s ln s − (λE · bE + λBI · bBI ),
since x∗ satisfies the equalities and the binding inequalities. Substituting the above in (2.1),
the maximum generalized entropy can be expressed in terms of the Lagrange multipliers
and the data as
G∗ = λE · bE + λBI · bBI .
(A.8)
This implies that the quantity Λ∗ is at least as large as G∗ , as claimed.
Now if σ is an arbitrary sequence with count vector ν, its probability under χ∗ is
Prχ∗ (σ) = (χ∗1 )ν1 · · · (χ∗m )νm
where χ∗j = e−(λ
E ·AE +λBI ·ABI )
.j
.j
Prχ∗ (σ) = e−ξ(ν) ,
. Therefore
where
ξ(ν) =
X
i
E
λE
i (Ai. · ν) +
X
BI
λBI
i (Ai. · ν).
(A.9)
i
The rest of the proof is analogous to that of Proposition II.2 in [OG16]. If ν is in C(δ),
then
E
E
E
E
BI
BI
bE
ABI
i − δ|bi | 6 Ai. · ν 6 bi + δ|bi |,
i. · ν 6 bi + δ|bi |.
Therefore from (A.9), noting that λBI > 0 but the λE
i can be positive or negative,
maxν∈C(δ) ξ(ν) 6 λE · bE + (|λE | · |bE |)δ + λBI · (bBI + δ|bBI |),
P
BI
minν∈C(δ) ξ(ν) > λE · bE − (|λE | · |bE |)δ + minν∈C(δ) i λBI
i (Ai. · ν).
42
(The | · | around λE cannot be removed.) Using (A.8) in the above,
maxν∈C(δ) ξ(ν) 6 G∗ + (|λE | · |bE | + λBI · |bBI |)δ,
P
BI
BI
minν∈C(δ) ξ(ν) > G∗ − (|λE | · |bE |)δ + minν∈C(δ) i λBI
i (Ai. · ν − bi )
= G∗ − (|λE | · |bE |)δ − ∆(C(δ)),
where
∆(C(δ)) , max
ν∈C(δ)
X
(A.10)
BI
BI
λBI
i (bi − Ai. · ν).
i
Finally, for any p.d. p and any n-sequence σ with count vector ν, Prp (σ) is given by the
expression in property 2.1 in §2.1. Comparing that with (A.9), ξ(ν) = G(ν) + nD(f kχ∗ ),
so by using (A.10)
G∗ − (|λE | · |bE |)δ − ∆(C(δ)) 6 G(ν) + nD(f | χ∗ ) 6 G∗ + (|λE | · |bE | + λBI · |bBI |)δ,
where f = ν/n, and the claim of the lemma follows.
Proof of inequality (4.15)
p
P
m k
Let y = n/2. The sum m
k=1 k y /Γ(k/2) can be found in closed form by noticing
that if it is split over even and odd k, each of the sums is hypergeometric. However, the
resulting expression is too complicated for our purposes. We will obtain a tractable bound
that matches the highest power of y in the sum, i.e. y m /Γ(m/2).
We need an auxiliary fact, relating Γ(k/2) for k < m to Γ(m/2). From Gautschi’s
inequality
p for the gamma function (see [OLBC10], 5.6.4) it follows that Γ((µ − 1)/2) >
Γ(µ/2)/ µ/2, for any µ > 1. Applying this recursively we find that for k > 1
m − k
2k/2 Γ(m/2)
Γ
>
1/2
2
m(m − 1) · · · (m − k + 1)
(A.11)
2k/2 Γ(m/2)
> k/2 −k(k−1)/(4m)
m e
where the 2nd line follows by using 1 − z < e−z , for z < 1, in the denominator of the first
line.
Now pulling out the last term of our sum, reversing the order of the other terms, and
43
applying (A.11) to each term, we get
m
m−1
X
X m m k/2
k(k−1)
m k 1
ym
1
y
y m−k e− 4m
<
+
k
Γ(k/2)
Γ(m/2) Γ(m/2)
m−k
2
k=1
k=1
m−1
m − k−1 k/2
ym X m
ym
e 2m
+
=
Γ(m/2) Γ(m/2)
k
2y 2
k=1
p
m
ym
ym
2
<
+
1 + m/(2y ) − 1
Γ(m/2) Γ(m/2)
m
p
m
(n/2) 2
1 + m/n ,
=
Γ(m/2)
where in going from the 2nd to
pthe 3d line
p last
m we ignored the exponential factor and the
term in the expansion of 1 + m/(2y 2 ) , and in the last line we substituted y = n/2.
The ratio of the sum and this last expression tends to 1 as n → ∞.
Proofs of inequality (4.26)
The first term is an upper bound on c1 . (4.22) is an inequality of the type x > α ln x + β,
with α, β > 0. We will show that if α + β > 1, this inequality is satisfied by x = 2α ln(α +
β) + β. [This expression is motivated by the method of successive substitutions: with
x0 = β, we get x2 = α ln(a ln β + β) + β; but this satisfies the inequality only if β < 1.]
Substituting into the inequality we get
(α + β)2 > 2α ln(α + β) + β
⇔
α+β >
α
β
2 ln(α + β) +
.
α+β
α+β
Therefore this will hold if α+β > max(2 ln(α+β), 1). Now we have assumed that α+β > 1,
and x > 2 ln x is always true for x > 0, so our claim is established. Turning to the case
α + β < 1, we can suppose that α < 1, otherwise we fall into the case α + β > 1. Then it
suffices to find a x that satisfies x > ln x + β, and that is so for x = 1.5β + ln β.
The third term in (4.26) is an upper bound on c3 . Write (4.24) as 21 Σ1 − cΣ2 6 c2 ηG∗ .
p
This will hold if c >
Σ22 + 2Σ1 ηG∗ − Σ2 /(2ηG∗ ), so the r.h.s. can be taken to be c3 . If
√
√
√
a, b > 0, which is guaranteed
by our assumption that x∗ > 1, then a + b < a + b, so
q
Pm
∗
√
i=1 1/(xi −1)
2Σ1 ηG∗ /(2ηG∗ ) =
is an upper bound on c3 .
2ηG∗
Proof of Proposition 5.1
To ease the notation, let kxk = s, kyk = t. First we show that
x y
−
6ϑ
s
t
⇒
|kx − yk − |s − t|| 6 min(s, t)ϑ.
44
(A.12)
We have
x y y y
x y
−
6ϑ⇔
−
−
6ϑ
+
s
t
s
s
t
s
1
1 1
⇔ kx − yk −
y 6ϑ
−
s
s
t
⇔ |kx − yk − |s − t|| 6 sϑ.
x y
y y
−
−
−
6ϑ
s
s
s
t
1
1 1
⇔ kx − yk − − t 6 ϑ
s
s
t
⇒
Exchanging x with y and s with t in this derivation, it also follows that
x y
6 ϑ ⇒ |kx − yk − |s − t|| 6 tϑ,
−
s
t
and this establishes (A.12). Now (A.12) implies that
x
y
6ϑ
−
kxk kyk
⇒
kx − yk 6 min(kxk, kyk)ϑ + |kxk − kyk|
and taking the contrapositive of this
kx − yk > min(kxk, kyk)ϑ + |kxk − kyk|
x
y
−
> ϑ,
kxk kyk
⇒
from which the claim of the proposition follows.
Proof of the inequality in (5.4)
This is an improvement over bounding the sum in the second line of (5.4) by simply pulling
∗ 2
out e−γ ϑ n1 and then bounding the rest by an integral. Splitting the sum around the point
n∗ ,
√ m
√
n2
X
( n + m) −γ ∗ ϑ2 n
√
e
6
n
n=n1
√ m
√ m
√
√
n2
n∗
X
X
( n + m)
( n + m)
−γ ∗ ϑ2 (n∗ +1)
−γ ∗ ϑ2 n1
√
√
+e
6
e
n
n
n=n1
n=n∗ +1
Z √n∗ +1
Z √s2 +2
√ m
√ m
−γ ∗ ϑ2 s1
−γ ∗ ϑ2 (s∗ +1)
2e
u
+
m
du
+
2e
u
+
m du,
√
√
n∗ +1
s1
since the summand is an increasing function of n. The last line can be written as
√
√ m+1 −γ ∗ ϑ2 s1
2
∗ 2 ∗
(e
− e−γ ϑ (s +1) )+
n∗ + 1 + m
m+1
√ m+1 −γ ∗ ϑ2 (s∗ +1)
√ m+1 −γ ∗ ϑ2 s1
√
2 √
s2 + 2 + m
e
−
s1 + m
e
m+1
and the desired result follows by neglecting the second exponential from each of the two
summands.
45
Proof of Theorem 5.3
We minimize the max of c2 (δ), c3 (δ) by setting them equal to each other. Substituting c2
for c into (5.9) which defines c3 , we get the equation for δ in the theorem:
2γ ∗ ϑ2 s1 1
Λ∗
C 00
− m ln ϑ∞ .
+ m ln δ = ln 3 0 +
ϑ∞ δ
εB
ϑ∞
(A.13)
Let f (δ) stand for the function of δ on the l.h.s. This function decreases for δ < 2γ ∗ ϑ2 s1 /
(mϑ∞ ). From the condition between ϑ and δ of Theorem 5.2, we must have δ < δmax =
2γ ∗ ϑ2 s1 /Λ∗ . So if 2γ ∗ ϑ2 s1 /(mϑ∞ ) > δmax , which will hold if Λ∗ > mϑ∞ , then f (δ) will
decrease with δ ∈ (0, δmax ). If f (δmax ) is less than the r.h.s. of (A.13) then (A.13) will
have a root δ0 ∈ (0, δmax ]. This condition on f (δmax ) boils down to
2γ ∗ ϑ∞ ϑ2 s1
Λ∗
m
C300
<
,
εB 0
or
ε
Λ∗
ϑ <
2γ ∗ ϑ∞ s1
1/m 2
C300
B0
1/m
.
(A.14)
To arrive at the condition of the theorem we find a simple lower bound on (C300 /B 0 )1/m .
From (5.5),
√
√
m+1
√ m+1
s∗ + 2 + m
>
s∗ + m + 2
.
C300 >
Therefore from (5.7)
m+1
√ ∗
2π m/2 em/12
s +m+2
C300
√
>
B0
(m + 1)Γ(m/2) s∗
√
m
2π m/2 em/12
>
s∗ + m + 2
(m + 1)Γ(m/2)
where in the first line we used the fact that the product of the last two factors in the
expression (5.7) for B 0 is < 1. It follows that
C300
B0
1/m
√
√
21/m πe1/12
s∗ + m + 2
>
((m + 1)Γ(m/2))1/m
√
√
21/m πe1/12 m p ∗
=
s /m + 1 + 2/m
((m + 1)Γ(m/2))1/m
p
> 2 s∗ /m + 1.
[To go from the 2nd to the 3d line, it can be shown that the first factor on the 2nd line is
an increasing function of m; its minimum occurs at m = 2 and is ≈ 2.22.] It follows that
condition (A.14) for the existence of the root δ0 will be satisfied if
ε1/m ϑ2 6
Λ∗ p ∗
s /m + 1,
γ ∗ ϑ∞ s1
46
as stated in the
theorem. Now since we have ensured c2 (δ0 ) = c3 (δ0 ), we can take ĉ =
max c2 (δ0 ), c1 , where c1 is as in Theorem 5.2.
Finally, it is quite likely that c2 (δ0 ) > c1 so that ĉ = c2 (δ0 ). Given that δ0 < δmax =
2γ ∗ ϑ2 s1 /Λ∗ and Λ∗ > mϑ∞ , it can be seen that this will be so if if ϑ < s∗ /(2γ ∗ s1 ).
References
[Apo74]
T.M. Apostol. Mathematical Analysis, 2nd Ed. Addison-Wesley, 1974.
[BHK14]
D. Berend, P. Harremoës, and A. Kontorovich. Minimum KL-divergence
on complements of L1 balls. IEEE Transactions on Information Theory,
60(6):3172–3177, 2014. Also http://arxiv.org/abs/1206.6544v8.
[BLM13]
S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A
Nonasymptotic Theory of Independence. Oxford University Press, 2013.
[BV04]
S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge, 2004.
[Cat12]
A. Caticha. Entropic Inference and the Foundations of Physics. In EBEB2012, the 11th Brazilian Meeting on Bayesian Statistics, 2012. Also http:
//arxiv.org/abs/1212.6967.
[CCA11]
A. Cichocki, S. Cruces, and S-I. Amari. Generalized Alpha-Beta Divergences
and Their Application to Robust Nonnegative Matrix Factorization. Entropy,
13:134–170, 2011.
[CK11]
I. Csiszár and J. Körner. Information Theory: Coding Theorems for Discrete
Memoryless Systems. Cambridge, 2nd edition, 2011.
[Csi91]
I. Csiszár. Why Least Squares and Maximum Entropy? An Axiomatic Approach
to Inference for Linear Inverse Problems. The Annals of Statistics, 19(4):2032–
2066, 1991.
[Csi96]
I. Csiszár. Maxent, Mathematics, and Information Theory. In K.M. Hanson
and R.N. Silver, editors, Maximum Entropy and Bayesian Methods, 15th Int’l
Workshop, Santa Fe, New Mexico, U.S.A., 1996. Kluwer Academic.
[CT06]
T.M. Cover and J.A. Thomas. Elements of Information Theory. J. Wiley, 2nd
edition, 2006.
[GC07]
A. Giffin and A. Caticha. Updating probabilities with data and moments. In
K.H. Knuth et al, editor, Bayesian Inference and Maximum Entropy Methods
in Science and Engineering, 28. AIP Conf. Proc. 954, 2007.
47
[Gr8]
P.D. Grünwald. Entropy Concentration and the Empirical Coding Game. Statistica Neerlandica, 62(3):374–392, 2008. Also http://arxiv.org/abs/0809.
1017.
[HLP97]
G.H. Hardy, J.E. Littlewood, and G. Pólya. Inequalities, 2nd Ed. Cambridge
University Press, 1997.
[HUL96]
J.B. Hiriart-Urruty and C. Lemaréchal. Convex Analysis and Minimization
Algorithms I. Springer-Verlag, 1996.
[Jay83]
E.T. Jaynes. Concentration of Distributions at Entropy Maxima. In R.D.
Rosenkrantz, editor, E.T. Jaynes: Papers on Probability, Statistics, and Statistical Physics. D. Reidel, 1983.
[Jay03]
E.T. Jaynes. Probability Theory: The Logic of Science. Cambridge University
Press, 2003.
[OG16]
K.N. Oikonomou and P.D. Grünwald. Explicit Bounds for Entropy Concentration under Linear Constraints. IEEE Transactions on Information Theory,
62:1206–1230, March 2016. Also http://arxiv.org/abs/1107.6004.
[Oik12]
K.N. Oikonomou. Analytical Forms for Most Likely Matrices Derived from
Incomplete Information. International Journal of Systems Science, 43:443–458,
March 2012. Also http://arxiv.org/abs/1110.0819.
[OLBC10] F.W. Olver, D.W. Lozier, R.F. Boisvert, and C.W. Clark, editors. NIST Handbook of Mathematical Functions. Cambridge University Press, 2010.
[OS06]
K.N. Oikonomou and R.K. Sinha. Network Design and Cost Analysis of Optical
VPNs. In Proceedings of the OFC, Anaheim, CA, U.S.A., March 2006. Optical
Society of America.
[OW05]
E. Ordentlich and M.J. Weinberger. A Distribution-Dependent Refinement of
Pinsker’s Inequality. IEEE Transactions on Information Theory, 51(5):1836–
1840, May 2005.
[PP02]
A. Papoulis and S.U Pillai. Probability, Random Variables, and Stochastic Processes, 4th Ed. Mc. Graw Hill, 2002.
[Sas13]
I. Sason. Entropy Bounds for Discrete Random Variables via Maximal Coupling.
IEEE Transactions on Information Theory, 59(11):7118–7131, 2013.
[SJ80]
J.E Shore and R.W Johnson. Axiomatic derivation of the principle of maximum
entropy and the principle of minimum cross-entropy. IEEE Transactions on
Information Theory, 26:26–37, 1980.
48
[Ski89]
J. Skilling. Classic Maximum Entropy. In J. Skilling, editor, Maximum Entropy
and Bayesian Methods. Kluwer Academic, 1989.
[VWBC05] S. Verdoolaege, K. Woods, M. Bruynooghe, and R. Cools. Computation and
Manipulation of Enumerators of Integer Projections of Parametric Polytopes.
Technical Report Report CW 392, K.U. Leuven, March 2005.
[Zha07]
Z. Zhang. Estimating mutual information via Kolmogorov distance. IEEE
Transactions on Information Theory, 53:3280–3282, 2007.
49
| 10math.ST
|
An integral-transform approach to the bioheat transfer problems in
magnetic hyperthermia
Kenya Murase
Department of Medical Physics and Engineering, Division of Medical
Technology and Science, Faculty of Health Science, Graduate School of
Medicine, Osaka University
1-7 Yamadaoka, Suita, Osaka 565-0871, Japan
Short title:
An integral-transform approach to magnetic hyperthermia
Address correspondence to:
Kenya Murase, Dr. Med. Sci., Dr. Eng.
Department of Medical Physics and Engineering, Division of Medical
Technology and Science, Faculty of Health Science, Graduate School of
Medicine, Osaka University
1-7 Yamadaoka, Suita, Osaka 565-0871, Japan
Tel & Fax: (81)-6-6879-2571,
E-mail: [email protected]
1
Abstract
Our purpose in this study was to present an integral-transform approach
to the analytical solutions of the Pennes' bioheat transfer equation and to apply it
to the calculation of temperature distribution in tissues in hyperthermia with
magnetic nanoparticles (magnetic hyperthermia).
The validity of our method was investigated by comparison with the
analytical solutions obtained by the Green's function method for point and shell
heat sources and the numerical solutions obtained by the finite-difference
method for Gaussian-distributed and step-function sources.
There was good agreement between the radial profiles of temperature
calculated by our method and those obtained by the Green's function method.
There was also good agreement between our method and the finite-difference
method except for the central temperature for a step-function source that had
approximately a 0.3% difference. We also found that the equations describing
the steady-state solutions for point and shell sources obtained by our method
agreed with those obtained by the Green’s function method. These results
appear to indicate the validity of our method.
In conclusion, we presented an integral-transform approach to the
bioheat transfer problems in magnetic hyperthermia, and this study
demonstrated the validity of our method. The analytical solutions presented in
this study will be useful for gaining some insight into the heat diffusion process
during magnetic hyperthermia, for testing numerical codes and/or more
2
complicated approaches, and for performing sensitivity analysis and
optimization of the parameters that affect the thermal diffusion process in
magnetic hyperthermia.
Keywords: Magnetic hyperthermia, magnetic nanoparticle, Pennes' bioheat
transfer equation, integral-transform method, Green’s function
method
3
1 Introduction
Hyperthermia is one of the promising approaches to cancer therapy. The
most commonly used heating method in the clinical setting is capacitive heating
by use of a radiofrequency (RF) electric field [1]. However, a major technical
problem with hyperthermia is the difficulty of heating the targeted tumor to the
desired temperature without damaging the surrounding tissues, as the
electromagnetic energy must be directed from an external source and penetrate
normal tissue. Other hyperthermia modalities including RF ablation and
ultrasound hyperthermia have been reported [2, 3], but the efficacies of these
modalities depend on the size and depth of the tumor, and disadvantages include
a limited ability to target the tumor and control the exposure.
Hyperthermia
using
magnetic
nanoparticles
(MNPs)
(magnetic
hyperthermia) was developed in the 1950s [4] and is still under development for
overcoming the above disadvantages [5, 6]. MNPs generate heat in an
alternating magnetic field as a result of hysteresis and relaxational losses,
resulting in heating of the tissue in which MNPs accumulate [7]. With the
development of precise methods for synthesizing functionalized MNPs [8],
MNPs with functionalized surfaces, which have high specificity for a tumor
tissue, have been developed as heating elements for magnetic hyperthermia [9].
Furthermore, there is renewed interest in magnetic hyperthermia as a treatment
modality for cancer, especially when it is combined with other, more traditional
therapeutic approaches such as the co-delivery of anticancer drugs or
4
photodynamic therapy [10]. From these aspects, magnetic hyperthermia has
received much recent attention.
The bioheat transfer equation proposed by Pennes [11] is the basis for
understanding the kinetics of the tumor and tissue heating. The solution of this
equation is important both for treatment planning and for the design of new
clinical heating systems [12].
Various investigations have attempted to obtain analytical solutions to
the Pennes’ bioheat transfer equation. Durkee and Antich [13] solved it in
one-dimensional multi-region Cartesian and spherical geometry, based on the
method of separation of variables and Green’s function method. Vyas and
Rustgi [14] obtained an analytical solution by using the Green’s function
method to describe the temperature distribution due to a laser beam with a
Gaussian profile. Andra et al. [15] solved it for a constant heat source embedded
in an infinite medium without blood perfusion by using the Laplace transform.
Deng and Liu [16] derived analytical solutions to the bioheat transfer problems
with generalized spatial or transient heating both on the skin surface and inside
biological bodies by using the Green’s function method. Bagaria and Johnson
[17] modeled diseased and healthy tissues as two finite concentric spherical
regions and included the blood perfusion effect in both regions. They obtained
analytical solutions to the model by separation of variables. Recently, Giordano
et al. [18] derived fundamental solutions of the Pennes’ bioheat transfer
equation in rectangular, cylindrical, and spherical coordinates.
5
Although the Green's function method is a convenient way to describe
thermal problems [19, 20] and has often been applied to solving the bioheat
transfer equation as described above [12, 14], it is not rare for its handling to
become complicated. Besides the Green’s function method, analytical solutions
to the Pennes’ bioheat transfer equation can be obtained by use of the
integral-transform method [19, 20], which is considered to be easier to
implement than the Green's function method. To the best of our knowledge,
however, there are few studies that have used this approach.
Our purpose in this study was to present an integral-transform approach
to the analytical solutions of the Pennes' bioheat transfer equation for the
calculation of the temperature distribution in tissues in magnetic hyperthermia
and to investigate its validity by comparison with the Green's function method
and the finite-difference method for several heat source models.
2 Materials and methods
2.1 Pennes' bioheat transfer equation
To estimate the temperature distribution in vivo, we solved the Pennes' bioheat
transfer equation [11] given by
c p
T
b c pbb Ta T Qmet P ,
t
(1)
where T is the temperature of tissue, the thermal conductivity of tissue, b the
density of blood, cpb the specific heat of blood, b the blood perfusion rate, Ta
6
the temperature of arterial blood, Qmet the rate of metabolic heat generation, and
P the energy dissipation. and cp are the density and specific heat of tissue,
respectively. In this study, it was assumed that the volume flow of blood per
unit volume is constant and uniform throughout tissue, which means that b is
constant. Furthermore, the above thermo-physical properties such as and cp
and Qmet were assumed to be constant. Therefore, Eq. (1) is reduced to
c p
T
b c pbb Tc T P ,
t
(2)
where
Tc Ta
Qmet
b c pbb .
(3)
Tc is considered to be the temperature of tissue in the steady state prior to
heating or the core body temperature maintained by the balance between
metabolic heat generation and blood perfusion. It should be noted that, when
Qmet
Ta , Tc can be assumed to be equal to Ta, as is often seen in the
b c pbb
literature [12, 18]. When we describe Eq. (1) in spherical coordinates, Eq. (1)
becomes
c p
T 2 T
r
b c pbb Tc T P .
t r 2 r r
2.2 Integral-transform method
7
(4)
Applying the integral transform (Fourier sine transform) to Eq. (4) yields (see
Appendix A)
2 1
T Tc
r
where K
0
F ( )
1 e
2 2
2
2
t
K
sin( r )d ,
(5)
b c pbb
c p
2
and
, and F ( ) is given by Eq. (A7).
When r=0, using the formula lim sin x 1 , Eq. (5) is reduced to
x
x 0
T Tc
2 1
0
F ( )
1 e
2 2
In the steady state, i.e., when
T Tc
When r=0 and
2
2
t
K
d .
(6)
t , Eq. (5) is reduced to
2 1 F ( )
sin( r )d .
r 0 2 2
(7)
t , Eq. (6) is reduced to
T Tc
2 1
0
F ( )
d .
2 2
(8)
As illustrative examples, we considered four heat sources (point, shell,
Gaussian-distributed, and step-function sources).
2.2.1 Point source
In this case, P is given by
P
P0
(r ) ,
4r 2
(9)
8
where (r ) is a Dirac's delta function and P0 is the point heating energy. For
this source, F ( ) given by Eq. (A7) becomes
F ( )
2 P0
.
4
(10)
Substituting Eq. (10) into Eq. (5) and using the formula
0
x sin(ax)
dx e ab ,
2
2
2
b x
we obtain
r
T Tc
e
e 0 2
2 2r 2
2
P0
In the steady state, i.e., when
T Tc
P0
4r
2
2
t
K
sin( r )d .
(11)
t , Eq. (11) is reduced to
e r .
(12)
It should be noted that Eq. (12) is also obtained from Eq. (7), and that Eqs. (11)
and (12) have a singularity at r=0, which is represented by the factor 1/r in the
equations and reveals the highly localized effect of a point source.
2.2.2 Shell source
In this case, P is given by
P
P0
(r r0 ) ,
4r 2
(13)
and F ( ) given by Eq. (A7) becomes
9
F ( )
Substituting
2 P0
sin r0 .
4r0
Eq.
(14)
into
(14)
Eq.
(5)
1
sin(ax) sin(bx) [cos(a b) cos(a b)] and
2
and
0
using
formulae
cos(ax)
ab
, we obtain
dx
e
2
2
2b
b x
sin( r )
r r0
r r0
0
T Tc
e
e
2
e
0 2
2 2rr0
4
P0
the
2
2
t
K
sin( r )d .
(15)
When r=0, substituting Eq. (14 ) into Eq. (6) yields
r
sin( r )
0
0
T Tc
e
e 0
2 2r0 2
2 2
P0
2
2
t
K
d .
(16)
In the steady state, Eqs. (15) and (16) become
T Tc
P0
r r0
r r0
e
e
8rr0
(17)
and
T Tc
P0
4r0
e r0 ,
(18)
respectively. Note that Eqs. (17) and (18) are also obtained from Eqs. (7) and (8),
respectively.
2.2.3 Gaussian-distributed source
In this case, P is given by
10
P P0 e
r2
r02
,
(19)
where P0 is the maximum value of the energy dissipation at the center and r0 is a
radius that is associated with how far from the center the heating is affecting the
tissue. For this source, F ( ) given by Eq. (A7) becomes
2
F ( )
P0 r03 e
4
2r02
4
.
(20)
Substituting Eq. (20) into Eq. (5) yields
T Tc
P0 r03
2 r 0
e r 4
2 2
2 2
0
In the steady state, i.e., when
T Tc
P0 r03
2 r 0
1 e
2
2
t
K
sin( r )d .
(21)
t , Eq. (21) is reduced to
e r 4
sin( r )d ,
2 2
2 2
0
(22)
When r=0, Eqs. (21) and (22) become
P0 r03
T Tc
2
2 e r 4
2 2
2 2
0
0
1 e
2
2
t
K
d
(23)
and
T Tc
P0 r03
2
0
2 e r 4
d ,
2 2
2 2
0
respectively.
2.2.4 Step-function source
11
(24)
In this case, P is given by
for 0 r r0
P
P 0
0
for r0 r
,
(25)
and F ( ) given by Eq. (A7) becomes
F ( )
1
r
P0 2 sin( r0 ) 0 cos(r0 ) .
2
(26)
Substituting Eq. (26) into Eq. (5) yields
2P
T Tc 0
r
0
sin( r0 ) r0 cos( r0 )
1 e
2 ( 2 2 )
2
2
t
K
sin( r )d . (27)
When r=0, Eq. (27) becomes
T Tc
2 P0
0
sin( r0 ) r0 cos(r0 )
1 e
( 2 2 )
2
2
t
K
d .
(28)
2.3 Green’s function method
2.3.1 Point source
The Green’s function of the Pennes’ bioheat transfer equation for radial flow in
an infinite domain in spherical coordinates has been given by Giordano et al.
[18]. When using this function, we obtain the temperature for a point source as
(see Appendix B)
b ( t )
T Tc
aP0
r2
4 a ( t )
e
d ,
8 3 2 0 [a(t )]3 2
t
12
(29)
where a
b c pbb
and b
. It should be noted that, when r=0, the
c p
c p
integral in Eq. (29) diverges to infinity.
In the steady state, i.e., when
2
e
analytical solution:
ar
b
r
a
t , the integral in Eq. (29) has an
. Thus, the steady-state solution obtained by the
Greens' function method for a point source becomes
T Tc
P0
4r
b
r
a
e
.
(30)
2.3.2 Shell source
When using the Green's function given by Giordano et al. [18], we obtain the
temperature for a shell source as (see Appendix B)
( r r0 )
( r r0 )
e b (t ) 4 a (t ) 4 a (t )
e
d .
e
a (t )
2
aP0
T Tc
8rr0
t
0
2
(31)
When r=0, we obtain from Eq. (B8)
b ( t )
T Tc
aP0
r02
4 a ( t )
e
d .
8 3 2 0 [a(t )]3 2
t
13
(32)
In the steady state, i.e., when
1
[e
ab
solution:
b
r r0
a
e
t , the integral in Eq. (31) has an analytical
b
( r r0 )
a
] . Thus, the steady-state solution obtained by
the Greens' function method for a shell source becomes
a
[e
b
P0
T Tc
8rr0
Similarly, the integral at
2
e
ar0
b
r0
a
b
r r0
a
e
b
( r r0 )
a
].
(33)
t in Eq. (32) has an analytical solution:
. Thus, the steady-state solution for a shell source at r=0 becomes
T Tc
P0
4r0
e
b
r0
a
.
(34)
2.4 Finite-difference method
We
also
solved
Eq.
(4)
by
using
the
finite-difference
method
(forward-difference scheme) (see Appendix C) for Gaussian-distributed and
step-function sources for comparison. When we used the finite-difference
method (Appendix C), the outer radius of the domain for analysis was taken as
15 cm, and the spatial and time intervals ( r and t ) were taken as 0.3 mm
c p r 2
and
, respectively.
2
2.5 Energy dissipation of magnetic nanoparticles
14
Rosensweig [7] developed analytical relationships and computations of the
energy dissipation of MNPs subjected to an alternating magnetic field (AMF).
From this theory, P in Eq. (4) can be given by [6, 7]
P 0 0 H 02 f
2f
,
1 (2f ) 2
(35)
where 0 is the permeability of free space, 0 the equilibrium susceptibility, and
H0 and f the amplitude and frequency of the AMF, respectively. is the
effective relaxation time given by
1
1
N
1
B ,
(36)
where N and B are the Neel relaxation and Brownian relaxation time,
respectively [6, 7]. N and B are given by the following relationships [6, 7]:
N 0
3VH
and B
,
k BT
2
e
(37)
where 0 is the average relaxation time in response to a thermal fluctuation,
the viscosity of medium, kB the Boltzmann constant, T the temperature, and
KVM / k BT , with K being the anisotropy constant of MNP. VH is taken as
the hydrodynamic volume of MNP that is larger than the magnetic volume
VM D 3 / 6 for MNP of diameter D. As a model for VH, it is assumed that
VH (1 2 / D) 3VM , where is the thickness of a sorbed surfactant layer.
Because the actual equilibrium susceptibility 0 is dependent on the magnetic
15
field, 0 is assumed to be the chord susceptibility corresponding to the Langevin
equation, given by
0 i
3
1
(coth ) ,
(38)
where i 0M d2VM 3k BT , 0 M d HVM k BT , H H 0 cos(2ft ) , Md is the
domain magnetization of a suspended particle, and is the volume fraction of
MNPs.
In this study, we considered magnetite (Fe3O4) as MNPs. The above
parameters for magnetite were taken to be as follows: M d 446 kA/m ,
K 9 kJ/m 3 , c p 670 J/kg/K , and 5180 kg/m 3 [21]. was taken as 0.003,
which is close to the typical magnetite dosage of ~10 mg Fe per gram of tumor
that has been reported in clinical studies [22].
Figure 1(a) shows the relationship between P and D for magnetite, in
which H0 was fixed at 5 mT and f was varied from 100 kHz to 1000 kHz with an
interval of 100 kHz, whereas Fig. 1(b) shows the case when f was fixed at 500
kHz and H0 was varied from 1 mT to 10 mT with an interval of 1 mT. It should
be noted that the unit of mT can be converted to kA/m by use of the relationship
1 mT 0.796 kA/m . As shown in Fig. 1, P largely depends on D and its
maximum value increases with increasing f and H0.
As an illustrative example, we considered the case with D 19 nm ,
f 500 kHz , and H 0 5 mT . In this case, P 2.28 106 W/m 3 . For a point
16
source, we assumed that MNPs were located within a sphere with a radius of 1
mm. From the relationship P P0 /V , where V is the volume of the region
where MNPs are located, P0 in Eq. (9) was taken as 0.0096 W. For a shell
source, r0 and the width of the shell were assumed to be 5 mm and 1 mm,
respectively, resulting in V 315.2 mm 3 . Thus, P0 in Eq. (13) was assumed to
be
0.72
W.
For
Gaussian-distributed
and
step-function
sources,
P0 P 2.28 106 W/m 3 was used in Eqs. (19) and (25).
2.6 Numerical studies
Numerical studies were performed under the following conditions: The values
for the thermo-physical properties of blood and tissue were assumed to be as
follows [21]: = 0.502 W/m/K, = 1060 kg/m3, cp = 3600 J/kg/K, b = 1000
kg/m3, cpb = 4180 J/kg/K, b = 6.4×10-3 s-1, and Tc = 310 K. In this study, the r0
values in Eqs. (13), (19), and (25) were all taken as 5 mm.
3 Results
Figure 2 shows a comparison of the radial profiles of temperature calculated by
our method and those calculated by the Green's function method for a point
source at three time points (5, 10, and 100 s). As shown in Fig. 2, there was
good agreement between them.
17
Figure 3 shows a comparison of the radial profiles of temperature
calculated by our method and those calculated by the Green's function method
for a shell source at four time points (10, 50, 100, and 1000 s). As shown in Fig.
3, there was good agreement between them.
Figure 4 shows a comparison of the radial profiles of temperature
calculated by our method and those calculated by the finite-difference method
for a Gaussian-distributed source at four time points (10, 50, 100, and 500 s). As
shown in Fig. 4, there was good agreement between them.
Figure 5 shows a comparison of the radial profiles of temperature
calculated by our method and those calculated by the finite-difference method
for a step-function source at four time points (10, 50, 100, and 500 s). As shown
in Fig. 5, although some difference (approximately 0.3%) was observed at r=0,
there was good agreement between them except for the central temperature.
4 Discussion
In this study, we presented an integral-transform approach to the bioheat
transfer problems in magnetic hyperthermia and derived the transient and
steady-state analytical solutions to the Pennes' bioheat transfer equation for
several heat source models by using this approach. Furthermore, we investigated
the validity of this approach by comparison with the analytical solutions
obtained by the Green’s function method for point and shell sources and the
numerical
solutions
obtained
by
the
18
finite-difference
method
for
Gaussian-distributed and step-function sources. To the best of our knowledge,
these analytical solutions obtained by the integral-transform approach have not
been reported previously. The largest difference was observed between the
central temperature obtained by our method and that obtained by the
finite-difference method for a step-function source (Fig. 5), but the difference
was approximately 0.3% at most. Excluding this case, there was good
agreement between our method and the Green’s function method or the
finite-difference method (Figs. 2-5), indicating the validity of our method.
As previously described, the steady-state solutions obtained by the
Greens' function method for point and shell sources are given by Eqs. (30) and
(33), respectively. Because
b
, Eqs. (30) and (33) agree with Eqs. (12)
a
and (17) derived from our method, respectively. Furthermore, the steady-state
solution for a shell source at r=0 obtained by the Greens' function method [Eq.
(34)] also agrees with that obtained by our method [Eq. (18)]. These results also
appear to indicate the validity of our method.
The shell source used in this study is a model consisting of a thin shell of
MNPs in the outer surface of a spherical solid tumor whose outer region extends
to infinity and represents the normal tissue. As pointed out by Giordano et al.
[18], this model is a realistic model distribution that provides an approximately
constant therapeutic temperature inside the tumor. For this model, there was
also good agreement between our method and the Green's function method.
19
The Green's function method is a convenient way of solving differential
equations such as the Pennes’ bioheat transfer equation [12, 14]. Mathematically,
the Green’s function is the solution to a differential equation with an
instantaneous point source. When the temperature distribution for various heat
sources is calculated by use of the Green’s function method, it is necessary to
compute the integral of the product of the Green’s function and the function
describing the heat source as shown in Eqs. (B4) and (B7). In general, this
integral becomes a double integral with respect to the temporal and spatial
variables. For point and shell sources that can be described by a Dirac’s delta
function as shown in Eqs. (9) and (13), it is relatively easy to compute the
double integral. However, it is not always easy to compute the double integral
for heat sources whose function cannot be described by a Dirac’s delta function,
such as Guassian-distributed and step-function sources. On the other hand, the
integral-transform method presented in this study appears much easier to
implement than the Green's function method.
In the integral-transform method presented in this study, the kernel for
the integral transform was taken to be
2 sin( r ) . In general, the kernel
should be chosen depending on the boundary conditions at r 0 [19]. When
the boundary condition at r 0 is of the first kind, the kernel should be
2 sin( r ) , whereas it should be
2 cos(r ) for the boundary condition
of the second kind [19]. In this study, the parameter [ r (see Appendix
20
A)] is always zero at r 0 , that is, the boundary condition at r 0 is of the
first kind. Thus, we used
2 sin( r ) as the kernel for the integral transform
in this study.
The analytical solutions presented in this study were based on several
assumptions. First, the domain for analysis was assumed to be infinite. Although
this assumption is considered to be valid for deep tumors surrounded by normal
tissue, our method cannot be applied to the case of relatively superficial tumors.
Second, the thermo-physical properties of blood and tissue were assumed to be
the same in both the tumor and normal tissue. Third, the shape of tumors and the
distribution of MNPs were assumed to be spherically symmetric. Although the
analytical solutions derived in this study cannot be applied to cases with
complex geometries and/or a heterogeneous medium, they will provide useful
tools for testing of numerical codes and/or more complicated approaches, and
for performing sensitivity analysis of the parameters involved in a problem [18].
In conclusion, we presented an integral-transform approach to the
bioheat transfer problems in magnetic hyperthermia, and this study
demonstrated the validity of our method. The analytical solutions presented in
this study will be useful for gaining some insight into the heat-diffusion process
during magnetic hyperthermia, for testing of numerical codes and/or more
complicated approaches, and for performing sensitivity analysis and
optimization of the parameters that affect the thermal diffusion process in
21
magnetic hyperthermia.
22
Appendix A
When the following parameter is introduced:
T Tc ,
(A1)
Eq. (4) is reduced to
c p
2
r
b c pbb P .
t r 2 r r
(A2)
Furthermore, if we perform the following variable transformation: r , Eq.
(A2) becomes
2
rP
K
2 2
t r
,
(A3)
where
K
b c pbb
c p
2
and
.
(A4)
If we apply the integral transform (Fourier sine transform) [19, 20] to Eq. (A3),
we obtain
K
F ( )
( 2 2 )
t
,
(A5)
where is defined by [20]
2
0
sin( r )dr ,
(A6)
and
23
2
F ( )
0
rP(r ) sin( r )dr .
(A7)
denotes the Fourier-transform variable, which is assumed to take all values
from 0 to infinity continuously. It should be noted that and
at
r
r
were taken as zero to obtain Eq. (A5).
Solving Eq. (A5) with respect to t yields
e
2 2
K
t
F ( )
(0)
1 e
( 2 2 )
2
2
t
K
,
(A8)
where (0) is the value of at t 0 . If we assume that the temperature (T)
at t 0 is equal to Tc, we obtain (0) 0 . Using the following inverse
Fourier transformation [20]:
2
0
( ) sin( r)d ,
(A9)
we obtain
2 1
0
F ( )
2 2
1 e
2
2
t
K
sin( r )d .
(A10)
Finally, by use of and Eqs. (A1) and (A4), the temperature (T) can be
r
obtained by Eq. (9).
24
Appendix B
The Green's function of Eq. (A2) in an infinite domain is given by [18]
( r r ')
( r r ')
e
4 a ( t )
4 a ( t )
e
,
G ( r , t ; r ' , )
e
2rr' a (t )
2
b ( t )
2
(B1)
where
a
1
c p K
(B2)
and
b c pbb 2
b
c p
K .
(B3)
Point source
For a point source model, P is given by Eq. (12). In this case, the
solution to Eq. (A2) is given by
P0 (r ' )
dr ' d
0 r '0
4r '2
( r r ') 2
( r r ') 2
aP0 t e b (t ) (r ' ) 4 a (t ) 4 a (t )
. (B4)
e
e
dr
'
d
8 0 r '0 rr ' a (t )
a
t
r '2 G ( r , t ; r ' , )
Using the following relationship:
e
( r r ')2
4 a ( t )
e
( r r ')2
4 a ( t )
4 a(t )
0
e a (t ) sin(r ) sin(r ' )d
2
25
(B5)
and the formula
0
e
a 2 x 2
cos(bx)dx
b ( t )
aP0
2a
e
b2
4a2
, we obtain
r2
4 a ( t )
e
d .
8 3 2 0 [a(t )]3 2
t
(B6)
Thus, we obtain Eq. (29).
Shell source
In this case, P is given by Eq. (16). In this case, the solution to Eq. (A2)
is given by
a
t
0 r '0
aP0
8rr0
P0 (r 'r0 )
dr ' d
2
4r '
( r r0 ) 2
( r r0 ) 2
b ( t )
e
.
e 4 a (t ) e 4 a (t ) d
a (t )
r '2 G ( r , t ; r ' , )
t
0
(B7)
Thus, we obtain Eq. (31).
When r=0, using Eq. (B5) and the formula lim sin x 1 yields
x
x 0
b ( t )
aP0
r02
4 a ( t )
e
d .
8 0 [a(t )]3 2
t
32
Thus, we obtain Eq. (32).
26
(B8)
Appendix C
To solve Eq. (4), we used the following finite-difference method
(forward-difference scheme). First, we divide the spatial and time domains into
r and t such that
small intervals
r (i 1) r (i 1,2,, M ) and
t ( j 1) t ( j 1,2,, N ) , and we denote the temperature at the nodal point
i r at the time j t by Ti , j . For r 0 , i.e., i 1 , Eq. (4) is reduced to
c p
Ti , j 1 Ti , j
t
2 Ti 1, j Ti , j Ti 1, j 2Ti , j Ti 1, j
2
i
1
r
r 2
b c pbb (Tc Ti , j ) Pi , j
,
(C1)
where Pi , j denotes the energy dissipation at the nodal point i r at the time
j t . Thus, Ti , j 1 can be computed from
Ti , j 1 Ti , j
t
2
(1
)(Ti 1, j Ti , j ) (Ti , j Ti 1, j )
i 1
c p r
2
b c pbb t
t
(Tc Ti , j )
P
c p
c p i , j
(C2)
for i=2, 3, …, M and j=1, 2, …, N.
For r 0 , i.e., i 1 , we used the following L'Hopital's rule [23] to
avoid dividing by zero:
1 T
2T
lim
lim 2 .
r 0 r r
r 0 r
(C3)
Then, we obtain
c p
T1, j 1 T1, j
t
6
T2, j T1, j
r 2
b c pbb (Tc T1, j ) P1, j
27
(C4)
or
T1, j 1 T1, j
b c pbb t
6t
t
(T2, j T1, j )
(Tc T1, j )
P1, j
2
c p
c p
c p r
(C5)
for j=1, 2, …, N. For numerical stability, the following condition should be
satisfied [24]:
t
c p r 2
0.5 .
As boundary conditions,
(C6)
T
was taken as zero at the center and outer
r
boundary, i.e., T2, j T1, j and TM , j TM 1, j for j=1, 2, ..., N. As initial conditions,
the temperature at t=0 was assumed to be Tc, i.e., Ti ,1 Tc for i=1, 2, ..., M.
28
References
1.
Abe M, Hiraoka M, Takahashi M, Egawa S, Matsuda C, Onoyama Y,
Morita K, Kakehi M, Sugahara T. Multi-institutional studies on
hyperthermia using an 8-MHz radiofrequency capacitive heating device
(Thermotron RF-8) in combination with radiation for cancer therapy.
Cancer. 1986; 58: 1589-95.
2.
Oura S, Tamaki T, Hirai I, Yoshimasu T, Ohta F, Nakamura R,
Okamura Y. Radiofrequency ablation therapy in patients with breast
cancers two centimeters or less in size. Breast Cancer. 2007; 14: 48-54.
3.
Seip R, Ebbini ES. Noninvasive estimation of tissue temperature
response to heating fields using diagnostic ultrasound. IEEE Trans
Biomed Eng. 1995; 42: 828-39.
4.
Gilchrist RK, Medal R, Shorey WD, Hanselman RC, Parrott JC, Taylor
CB. Selective inductive heating of lymph nodes. Ann Surg. 1957; 146:
596-606.
5.
Jordan A, Scholz R, Maier-Hauff K, Johannsen M, Wust P, Nodobny J,
Schirra H, Schmidt H, Deger S, Loening S, Lanksch W, Felix R.
Presentation of a new magnetic field therapy system for the treatment of
human solid tumors with magnetic fluid hyperthermia. J Magn Magn
Mat. 2001; 225: 118-26.
6.
Murase K, Oonoki J, Takata H, Song R, Angraini A, Ausanai P,
Matsushita T. Simulation and experimental studies on magnetic
29
hyperthermia with use of superparamagnetic iron oxide nanoparticles.
Radiol Phys Technol. 2011; 4: 194-202.
7.
Rosensweig RE. Heating magnetic fluid with alternating magnetic field.
J Magn Magn Mat. 2002; 252: 370-4.
8.
Neuberger T, Schopf B, Hofmann H, Hofmann M, von Rechenberga B.
Superparamagnetic
nanoparticles
for
biomedical
applications:
possibilities and limitations of a new drugdelivery system. J Magn Magn
Mat. 2005; 293: 483-96.
9.
Ito A, Shinkai M, Honda H, Kobayashi T. Medical applications of
functionalized magnetic nanoparticles. J Biosci Bioeng. 2005; 100: 1-11.
10.
Balivada S, Rachakatla RS, Wang H, Samarakoon TN, Dani RK, Pyle M,
Kroh FO, Walker B, Leaym X, Koper OB, Tamura M, Chikan V,
Bossmann SH, Troyer DL. A/C magnetic hyperthermia of melanoma
mediated by iron(0)/iron oxide core/shell magnetic nanoparticles: a
mouse study. BMC Cancer. 2010; 10: 119-27.
11.
Pennes HH. Analysis of tissue and arterial blood temperatures in the
resting human forearm. J Appl Physiol. 1948; 1:93-122.
12.
Gao B, Langer S, Corry PM. Application of the time-dependent Green's
function and Fourier transforms to the solution of the bioheat equation.
Int J Hyperthermia. 1995; 11: 267-85.
13.
Durkee JW, Antich PP, Lee CE. Exact solutions to the multiregion
time-dependent bioheat equation. I: solution development. Phys Med
30
Bio. 1990; 35: 847-67.
14.
Vyas R, Rustgi ML. Green's function solution to the tissue bioheat
equation. Med Phys. 1992; 19: 1319-24.
15.
Andra W, d’Ambly CG, Hergt R, Hilger I, Kaiser WA. Temperature
distribution as function of time around a small spherical heat source of
local magnetic hyperthermia. J Magn Magn Mat. 1999; 194: 197-203.
16.
Deng Z-S, Liu J. Analytical study on bioheat transfer problems with
spatial or transient heating on skin surface or inside biological bodies. J
Biomech Eng. 2002; 124: 638-49.
17.
Bagaria HG, Johnson DT. Transient solution to the bioheat equation and
optimization for magnetic fluid hyperthermia treatment. Int J
Hyperthermia. 2005; 21: 57-75.
18.
Giordano MA, Gutierrez G, Rinaldi C. Fundamental solutions to the
bioheat equation and their application to magnetic fluid hyperthermia.
Int J Hyperthermia. 2010; 26: 475-84.
19.
Ozisik MN. 'Boundary Value Problems of Heat Conduction',
International Textbook Company, Scranton, Pennsylvania, 1968, pp.
43-124.
20.
Carslaw HS, Jaeger JC. 'Conduction of Heat in Solids', Oxford at the
Clarendon Press, Oxford, 1959, pp. 455-465.
21.
Maenosono S, Saita S. Theoretical assessment of FePt nanoparticles as
heating elements for magnetic hyperthermia. IEEE Trans Magn. 2006;
31
42: 1638-42.
22.
Jordan A, Scholz R, Wust P, Fahling H, Krause J, Wlodarczyk W,
Sander B, Vogl T, Felix R. Effects of magnetic fluid hyperthermia
(MFH) on C3H mammary carcinoma in vivo. Int J Hyperthermia. 1997;
13: 587-605.
23.
Spivak M. 'Calculus', Publish or Perish, Houston, 1994, pp. 201-202.
24.
Smith GD. 'Numerical Solution of Partial Differential Equations with
Exercises and Worked Solutions', Oxford University Press, London,
1965, p. 93.
32
Figure Legends
Fig. 1
(a) Relationship between the energy dissipation (P) and the diameter
of magnetic nanoparticles (MNPs) (D) for magnetite, in which the
amplitude of the alternating magnetic field (H0) was fixed at 5 mT and
the frequency (f) was varied from 100 kHz to 1000 kHz with an
interval of 100 kHz.
(b) Relationship between P and D for magnetite, in which f was fixed
at 500 kHz and H0 was varied from 1 mT to 10 mT with an interval of
1 mT. The unit of mT can be converted to kA/m by use of the
relationship 1 mT 0.796 kA/m .
Fig. 2
Comparison of the radial profiles of temperature calculated by our
method and those obtained by the Green's function method for a point
source. The solid, dashed, and dotted lines show the results calculated
by our method at t=5 s, 10 s, and 100 s, respectively, whereas the
closed circles, squares, and triangles show the results obtained by the
Green's function method at t=5 s, 10 s, and 100 s, respectively.
Fig. 3
Comparison of the radial profiles of temperature calculated by our
method and those obtained by the Green's function method for a shell
source. The solid, long dashed, dashed, and dotted lines show the
results calculated by our method at t=10 s, 50 s, 100 s, and 1000 s,
respectively, whereas the closed circles, squares, triangles, and
33
diamonds show the results obtained by the Green's function method at
t=10 s, 50 s, 100 s, and 1000 s, respectively.
Fig. 4
Comparison of the radial profiles of temperature calculated by our
method and those obtained by the finite-difference method for a
Gaussian-distributed source. The solid, long dashed, dashed, and
dotted lines show the results calculated by our method at t=10 s, 50 s,
100 s, and 500 s, respectively, whereas the closed circles, squares,
triangles, and diamonds show the results obtained by the
finite-difference method at t=10 s, 50 s, 100 s, and 500 s, respectively.
Fig. 5
Comparison of the radial profiles of temperature calculated by our
method and those obtained by the finite-difference method for a
step-function source. The solid, long dashed, dashed, and dotted lines
show the results calculated by our method at t=10 s, 50 s, 100 s, and
500 s, respectively, whereas the closed circles, squares, triangles, and
diamonds show the results obtained by the finite-difference method at
t=10 s, 50 s, 100 s, and 500 s, respectively.
34
Fig. 1(a)
Fig. 1(b)
35
Fig. 2
Fig. 3
36
Fig. 4
Fig. 5
37
| 5cs.CE
|
LaSalle Invariance Principle for Discrete-time Dynamical Systems:
A Concise and Self-contained Tutorial
arXiv:1710.03710v2 [math.DS] 24 Oct 2017
Wenjun Mei and Francesco Bullo
Center for Control, Dynamical-system, and Computation
University of California at Santa Barbara
Summary. As a method to establish the Lyapunov stability of differential/difference equations systems, the LaSalle invariance
principle was originally proposed in the 1950’s and has become a fundamental mathematical tool in the area of dynamical systems
and control. In both theoretical research and engineering practice, discrete-time dynamical systems have been at least as extensively
studied as continuous-time systems. For example, model predictive control is typically studied in discrete-time via Lyapunov
methods. However, there is a peculiar absence in the standard literature of standard treatments of Lyapunov functions and LaSalle
invariance principle for discrete-time nonlinear systems. Most of the textbooks on nonlinear dynamical systems focus only on
continuous-time systems. For example, the classic textbook by Khalil on nonlinear systems [10] relegates discrete-time systems
to a few exercises at the end of Chapter 4. The textbook by Vidyasagar [14] does not present the discrete-time LaSalle invariance
principle. In Chapter 1 of the book by LaSalle [11], the author establishes the LaSalle invariance principle for difference equation
systems. However, all the useful lemmas in [11] are given in the form of exercises with no proof provided. In this document,
we provide the proofs of all the lemmas proposed in [11] that are needed to derive the main theorem on the LaSalle invariance
principle for discrete-time dynamical systems. We organize all the materials in a self-contained manner. We first introduce some
basic concepts and definitions in Section 1, such as dynamical systems, invariant sets, and limit sets. In Section 2 we present and
prove some useful lemmas on the properties of invariant sets and limit sets. Finally, we establish the original LaSalle invariance
principle for discrete-time dynamical systems and a simple extension in Section 3. In Section 4, we provide some references on
extensions of LaSalle invariance principles for further reading. This document is intended for educational and tutorial purposes and
contains lemmas that might be useful as a reference for researchers.
Key words: LaSalle invariance principle, difference equations, discrete-time nonlinear dynamics, limit set, Lyapunov function
1 Basic Concepts: Dynamical System, Motion, Limit Set, and Invariant Set
Before reviewing the basic concepts in discrete-time dynamical systems, we first introduce some frequently used notations. Let N
be the set of natural numbers, i.e, {0, 1, 2, . . . }. Denote by Z and Z+ the set of integers and positive integers respectively. The set of
real numbers is denoted by R and the m-dimension Euclidean space is denoted by Rm . Let 0m be the all-zeros m × 1 column vector
and let 1m be the all-ones m × 1 column vector. We use φ to denote the empty set. For any sequence {xk }k∈N , by {xk } → y we mean
xk → y as k → ∞. Without causing any confusion, sometimes we omit the subscript k ∈ N when we refer to a sequence {xk }k∈N .
1.1 Discrete dynamical system
Given a map T : Rm → Rm , the following equations system:
x(n + 1) = T x(n) ,
for any n ∈ N,
(1)
is referred to as a m-dimension first-order difference equations system. Equation (1) together with an additional condition x(0) =
x0 ∈ Rm defines an initial-value problem for the m-dimension first-order difference equations system. A sequence {x(n)}n∈N is the
solution to the initial-value problem if x(n) = T n (x0 ) for any n ∈ N. Similarly, for any k ∈ Z+ , a map g : Rm × · · · × Rm → Rm
{z
}
|
k
defines a m-dimension k-th order difference equations system, and the corresponding initial-value problem is of the following form:
2
x(n + 1)
x(k − 1)
x(0)
= g x(n), . . . , x(n − k + 1) ,
= xk−1 ,
...
= x0 .
Lemma 1 (Equivalence to first-order system). Any one-dimension m-th order difference equations system
u(n + 1) = g u(n), . . . , u(n − m + 1)
is equivalent to an m-dimension first-order difference equations system.
Proof: For any n ≥ m, let xm (n) = u(n − m + 1), xm−1 (n) = u(n − m + 2), . . . , x2 (n) = u(n − 1), and x1 (n) = u(n). Define x(n) ∈ Rm
>
as x(n) = x1 (n), . . . , xm (n) . We thereby have:
x1 (n + 1) = g x1 (n), . . . , xm (n) ,
x2 (n + 1) = x1 (n),
...
xm (n + 1) = xm−1 (n).
Define T : Rm → Rm as
g x1 (n), . . . , xm (n)
x1 (n)
.
T x(n) =
...
xm−1 (n)
We obtain an m-dimension first order difference equations system x(n + 1) = T x(n) . This concludes the proof.
t
u
For simplicity, in the rest of this document, whenever we refer to a difference equations system, we assume it is m-dimension
and first-order, unless specified. In addition, no matter whether it is emphasized, the map T : Rm → Rm is always assumed to be
continuous in this document.
Definition 1 (discrete semi-dynamical system and dynamical system). A map π : N × Rm → Rm is a discrete semi-dynamical
system on Rm if, for any n, k ∈ N and any x ∈ Rm ,
1) π(0, x) = x;
2) π n, π(k, x) = π(n + k, x) (semi-group property);
3) π is continuous.
The map π is a discrete dynamical system if 2) above holds for any n, k ∈ Z (group property).
Remark 1. The semi-group property implies the uniqueness of the solution to equations x(n + 1) = π x(n) , with the initial condition x(0) = x0 , in the forward direction, i.e., for any n ∈ N. The group property leads to the uniqueness of the solution in both
direction, i.e., for any n ∈ Z.
1.2 Limit set and invariant set
Invariant set and limit set are two important concepts in dynamical systems. Limit set characterizes the asymptotic behavior of a
difference equations system, i.e., the limit behavior of T n (x0 ), as n → ∞. A compact invariant set implies the existence of limit sets.
In this subsection we present the definitions of invariant set and limit set.
Definition 2 (Invariant sets). Given a map T : Rm → Rm and a set H ⊂ Rm , define the set T (H) = {y ∈ Rm | y = T (x) for some x ∈
H}. The set H is positively invariant if T (H) ⊂ H, negatively invariant if T (H) ⊃ H, and invariant if T (H) = H. A set H ⊂ Rm is
invariantly connected if
1) H is closed and invariant;
3
2) H is not a union of two non-empty disjoint closed invariant sets.
For any E ⊂ Rm , a set M is the largest invariant set in E, if M ⊂ E, T (M) = M, and T (A ∪ M) 6= A ∪ M for any set A such that
A ⊂ E and A ∩ M = φ .
Remark 2. The following statements can be inferred from Definition 2: 1) Any set of countably many (more than one) isolated fixed
points is not invariantly connected; 2) If M is the largest invariant set in E, then, for any A ⊂ E, T (A ∪ M) = A ∪ M implies A ⊂ M.
Before presenting the definition of limit set, we first introduce some preliminary notions.
Definition 3 (Motion, periodicity, fixed point, and extension of motion). Given a map T : Rm → Rm and a vector x ∈ Rm , the
motion from x refers to the sequence {T n (x)}n∈N and is denoted by T n x. A motion T n x is periodic if there exists some k ∈ Z+ such
that T k (x) = x. The least such k is referred to as the period of the motion T n . A point x is called a fixed point of T if the motion
T n x has period k = 1, i.e., T (x) = x. A set Tn x = {Tn (x)}n∈Z ⊂ Rm is referred to as an extension of the motion T n x if T0 x = x and
T (Tn x) = Tn+1 x for any n ∈ Z.
Remark 3. The extension of a motion T n x is not unique if the map T is non-invertible.
Definition 4 (Distance and convergence). For any point x ∈ Rm and set S ⊂ Rm , define the distance between
x and S as ρ(x, S) =
inf kx − yk, where k·k is some norm defined in Rm . A motion T n x converges to the set S if lim ρ T n (x), S = 0.
y∈S
n→∞
Definition5 (Interior, closure, and boundary). For any x ∈ Rm and r > 0, define the open ball around x with radius r as the set
B(x, r) = y ∈ Rm ky − xk < r . For any set S ⊂ Rm , x ∈ S is an interior point of S is there exists
ε > 0 such that B(x, ε) ⊂ S.
Denote by int S the set of all the interior points of S. Define the closure of S as the set S = x ∈ Rm ρ(x, S) = 0 . Define the
boundary of S as the set ∂ S = S \ int S.
Now we present the definitions of limit point and limit set.
Definition 6 (Limit point and limit set). Given a motion T n x, y ∈ Rm is a limit point of the motion T n x if there exists a subsequence
{nk }k∈N such that nk → ∞ and T nk (x) → y as k → ∞. When there is no ambiguity about the map T , we also refer to y as a limit
point of x. The set of all the limit points of x is referred to as the limit set of x, denoted by Ω (x). That is,
n
o
Ω (x) = y ∈ Rm There exists sequence {nk } ⊂ N such that nk → ∞ and T nk (x) → y as k → ∞ .
Given a set H ⊂ Rm , the limit set of the set H is denoted by Ω (H) and defined as
n
o
Ω (H) = y ∈ Rm There exists sequence {nk } ⊂ N and {yk } ⊂ H such that nk → ∞ and T nk (yk ) → y as k → ∞ .
2 Properties of Invariant Sets and Limit Sets
In this section we present and prove some important properties of the limit sets and the invariant sets of difference equations
systems. These properties will be used in the proof of the LaSalle invariant principle.
2.1 Properties of invariant sets
We first present some lemmas on the properties of invariant sets.
Lemma 2 (Invariantly connected set and periodic motion). For any given continuous map T : Rm → Rm , suppose H is an
invariant set with finite elements. The set H is invariantly connected if and only if H is a periodic motion T n x.
4
Proof: Suppose H is a periodic motion with period K. By definition, H = {x, T (x), . . . , T K−1 (x)} is invariant and closed. It is also
straightforward to check that H = {x, T (x), . . . , T K−1 (x)} cannot be a union of any two non-empty disjoint closed invariant set.
Therefore, H is invariantly connected.
Now suppose H is invariantly connected. As a consequence, any x ∈ H cannot be a fixed point of the map T , otherwise either
T (H) ⊂ H or H is a union of two disjoint closed invariant sets: {x} and H \ {x}. Since x is not a fixed point and H is invariant,
there must exists at least one positive integer k > 1 such that T k(x) (x) = x. Denote the least such integer as k(x). Since H is a set of
finite elements, the least common multiple of all the k(x) such that x ∈ H is a finite positive integer, denoted by K. Therefore, for
any x ∈ H, T K (x) = x, which implies that H is a periodic motion. This conclude the proof.
t
u
Lemma 3 (Closures of invariant sets). For any continuous map T : Rm → Rm , the following statements hold:
1) The closure of any positively invariant set of T is positively invariant;
2) The closure of any bounded invariant set of T is invariant.
Proof: Suppose H is a positively invariant set of map T . Due to the continuity of T , if T (x) ∈
/ H for some x ∈ ∂ H, then there exists
y ∈ int(H) such that T (y) ∈
/ H. This implies that H is not positively invariant, which leads to a contradiction. Therefore, T (∂ H) ⊂ H
and thus T (H) ⊂ H. This concludes the proof for statement 1).
According to statement 1), the closure of a bounded invariant set is positively invariant. Suppose H is bounded and invariant. We
have T (H) ⊃ H, which leads to the following result: For any x ∈ H, there exists y ∈ H such that T (y) = x. Since H is bounded,
∂ H is well-defined and bounded. For any x∗ ∈ ∂ H, there exists a sequence {xn } ⊂ H such that {xn } → x∗ . For any xn ∈ H, there
exists yn ∈ H such that T (yn ) = xn . Now we obtain a sequence {yn } ⊂ H ⊂ H. Since H is a compact set, there exists a sequence
{nk } → ∞ such that the subsequence {ynk } converges to some y∗ ∈ H. Moreover, since map T is continuous,
T (y∗ ) = T lim ynk = lim T (ynk ) = lim xnk = x∗ ∈ H.
k→∞
Therefore, for any x∗
∈ H, there exists y∗
k→∞
k→∞
such that T (y∗ ) = x∗
∈H
∈ T (H). Now we have obtained both T (H) ⊂ H and T (H) ⊃ H.
Therefore, T (H) = H, which concludes the proof of statement 2).
t
u
Lemma 4 (Invariant set and extension of motion). For any continuous map T : Rm → Rm , a set H is an invariant set of T if and
only if each motion T n x starting in H has an extension in H.
Proof: Suppose H is an invariant set. Since H is positively invariant, for any x ∈ H, we have T n (x) ∈ H for any n ∈ N, which leads
to T n ⊂ H. On the other hand, since H is also negatively invariant, for any x ∈ H, there exists x−1 ∈ H such that T (x−1 ) = x. Let
T−1 (x) = x−1 . Since x−1 ∈ H, there exists x−2 ∈ H such that T (x−2 ) ∈ H. Following this argument and let Tn (x) = T n (x) for any
n ∈ N, we construct an extension Tn x = {Tn (x)}n∈Z of the motion T n x and the extension Tn x is in H.
Now suppose that each motion T n x starting in H has an extension Tn x in H. Since x ∈ H leads to Tn x ⊂ H for any x ∈ H, we have
T (x) ∈ H, which implies T (H) ⊂ H. In addition, for any x ∈ H, since there exists y = T−1 (x) ∈ H such that T (y) = x, we have
T (H) ⊃ H. Therefore, H is an invariant set of T . This concludes the proof.
t
u
Lemma 5 (Properties of the largest invariant set). For any continuous map T ∈ Rm → Rm and any set E ⊂ Rm , if M is the largest
invariant set in E, then the following statements hold:
1) M is the union all the extensions of motions that remain in E for all n ∈ Z;
2) x ∈ M if and only if there exists an extension of motion Tn x such that Tn x ⊂ E;
3) If E is compact, then M is compact.
Proof: For any extension of motion staring at x, denoted by Tn x, one can easily check that the set Tn x = {Tn (x)}n∈Z is invariant,
i.e., T (Tn x) = Tn x. If Tn x ∈ E, then, according to Remark 2, T (M ∪ Tn x) = T (M) ∪ T (Tn x) = M ∪ Tn x implies that Tn x ⊂ M. On the
other hand, for any x ∈ M, since M is invariant, we have T (x) ∈ M and there exists x−1 ∈ M such that T (x−1 ) = x. Then we further
have T 2 (x) ∈ M and there exists x−2 ∈ M such that T (x−2 ) = x−1 . Following this argument, we conclude that any x ∈ M is in some
extension of motion in M. This concludes the proof of statement 1).
Statement 2) is a straightforward result of statement 1).
5
Now we proceed to prove statement 3). For any sequence {xn } ⊂ M such that {xn } → x∗ , since E is compact, we have x∗ ∈ E.
Moreover, since xn ∈ M leads to T k (xn ) ∈ M for any k ∈ N, due to the continuity of T , we have
lim T k (xn ) = T k lim xn = T k (x∗ ).
n→∞
n→∞
Therefore, the motion T n x∗ satisfies T n x∗ ⊂ E. Moreover, since M is invariant and {xn } ⊂ M, for any n ∈ N, there exists yn ∈ M
such that T (yn ) = xn . Now we obtain a sequence {yn } ⊂ M ⊂ E. Since E is compact, there exists a subsequence {ynk } such that
{nk } → ∞ and {ynk } → y∗ ∈ E as k → ∞. Due to the continuity of map T , we have
T (y∗ ) = T lim ynk = lim T (ynk ) = lim xnk = x∗ .
k→∞
(x∗ )
y∗
k→∞
x∗
k→∞
y∗ ,
Let T−1
= ∈ E. Applying the same argument above for to
we obtain T−2 (x∗ ) ∈ E such that T (T−2 (x∗ )) = T−1 (x∗ ).
∗
n
∗
Continue this argument, we get an extension Tn x of the motion T x such that Tn x∗ ⊂ E.According to statement 1), we have
x∗ ∈ M. Now we have proved that any {xn } ⊂ M such that {xn } → x∗ leads to x∗ ∈ M. Therefore, M is a compact set.
t
u
2.2 Properties of limit sets
Lemma 6 (Closed forms of limit sets). Given a continuous map T : Rm → Rm , for any x ∈ Rm , the limit set of x given by Definition 6 satisfies
n
(2)
Ω (x) = ∩∞j=0 ∪∞
n= j {T (x)};
Similarly, for any set H ⊂ Rm , the limit set of H, given by Definition 6, satisfies
n
Ω (H) = ∩∞j=0 ∪∞
n= j T (H).
(3)
Proof: We first prove equation (2). Suppose y ∈ Ω (x). For any j ∈ N, since
(i there exists {nk }k∈N such that nk → ∞ and T nk (x) → y as k → ∞;
n
nk
(ii ρ(y, ∪∞
n= j {T (x)}) ≤ ρ(y, T (y)) for any k such that nk > j,
n
∞
n
by letting k → ∞, we have ρ(y, ∪∞
n= j {T (x)}) = 0, which implies that y ∈ ∪n= j {T (x)}. Since the argument above holds for any
∞
∞
∞
n
n
n
j ∈ N, we have y ∈ ∩∞j=1 ∪∞
n= j {T (x)}. On the other hand, for any y ∈ ∩ j=1 ∪n= j {T (x)}, since y ∈ ∪n= j {T (x)} for any j ∈ N,
∞
n
we have ρ(y, ∪n= j {T (x)}) = 0, that is,
inf
ρ(y,
z)
=
0,
for
any
j
∈
N.
Let
0
<
ε
<
1
and
j
=
1.
There
must exist some
∞
z∈∪n= j {T n (x)}
n1 ≥ 1 such that
Then let j = n1 + 1. There must exist some n2 ≥ j > n1 such that ρ(y, T n2 (x)) < ε 2 . Following
this argument, we construct a subsequence {nk }k∈Z+ such that {nk } → ∞ and ρ(y, T nk (x)) < ε k for any k ∈ Z+ , which implies that
T nk → y as k → ∞. Therefore, according to Definition 6, y ∈ Ω (x). This concludes the proof for equation (2).
ρ(y, T n1 (x)) < ε.
Similarly, equation (3) is proved by the following argument:
y ∈ Ω (H) ⇐⇒ ∃ {nk } → ∞ and {yk } ⊂ H, s.t. T nk (yk ) → y as k → ∞,
⇐⇒ ∀ j ∈ N, ∀ k ∈ N, ∃ nk ≥ j and yk ∈ H s.t. T nk (yk ) ∈ B(y, ε k ), for some 0 < ε < 1,
⇐⇒ ∀ j ∈ N,
⇐⇒ ∀ j ∈ N,
inf ky − T n (x)k = 0,
n≥ j, x∈H
n
ρ(y, ∪∞
n= j T (H)) = 0.
t
u
Lemma 7 (Invariance and asymptotic properties of limit set Ω (x)). For any continuous map T : Rm → Rm and any x ∈ Rm , the
following statements hold:
1) The limit set Ω (x) is closed and positively invariant;
2) If the motion T n x is a bounded set, then Ω (x) is
a) nonempty;
b) compact;
c) invariant;
d) invariantly connected;
e) the smallest set that T n (x) approaches as n → ∞.
6
n
Proof: The sketch of tis proof can be found on Page 4 of [11]. Let A j (x) = ∪∞
n= j {T (x)}. According to Lemma 6, we have
∞
Ω (x) = ∩ j=0 A j (x). Each A j (x) is a closed set since it is the closure of a set of countably many points in Rm . Therefore, as the
intersection of countably many closed sets A j (x)’s, Ω (x) is closed. Moreover, for any y ∈ Ω (x), by definition, there exists asequence
{nk } → ∞ such that T nk (x) → y as k → ∞. Let `k = nk + 1. Due to the continuity of map T , we haveT `k (x) = T T nk (x) → T (y),
which implies that T (y) is also a limit point of x, i.e., T (y) ∈ Ω (x). Therefore, we obtain T Ω (x) ⊂ Ω (x). This concludes the
proof of statement 1).
Now we prove statement 2). Since T n x is bounded, there exists a sequence {nk } → ∞ such that T nk (x) converges. By definition,
limk→∞ T nk (x) ∈ Ω (x). Therefore, Ω (x) is non-empty. This concludes the proof of statement 2)a).
n
∞
n
n
By definition, Ω (x) ⊂ ∪∞
n=0 {T (x)}. If T x is bounded, then ∪n=0 {T (x)} is bounded, which in turn implied that Ω (x) is bounded.
According to statement 1), Ω 9x) is closed. Therefore, Ω (x) ⊂ Rm is compact. This concludes the proof of statement 2)b).
For any y ∈ Ω (x), there exists {nk } → ∞ such that T nk (x) → y. Let `k = nk − 1, then T T `k (x) → y as k → ∞. Since {T `k (x)}k∈N ⊂
T n x is bounded, there exists a sequence {kr } → ∞ such that T `kr (x) converges to some z ∈ Ω (x) as r → ∞. Due to the continuity of
T , we have
T (z) = T lim T `kr (x) = lim T `kr +1 (x) = lim T nkr (x) = y.
r→∞
r→∞
r→∞
Now wehave shown that, for any y ∈ Ω (x), there exists z ∈ Ω (x) such that T (z) = y. Therefore, T Ω (x) ⊃ Ω (x). We already have
T Ω (x) ⊂ Ω (x) according to statement 1). Therefore, Ω (x) is invariant. This concludes the proof of statement 2)c).
We prove statement 2)d) by contradiction. Suppose Ω (x) is a union
of two disjoint sets Ω1 (x) and Ω2 (x), which
are both closed,
non-empty, and invariant. As a consequence, d = ρ Ω1 (x), Ω2 (x) > 0. Due to the
continuity of T and T Ω1 (x) = Ω1 (x), for any
0 < δ1 < d/2, there exists ε1 satisfying 0 < ε1 < δ1 such that T B(Ω1 (x), ε1 ) ⊂ B(Ω1 (x), δ1 ). Here we adopt the generalized
definition of an open ball around a set. That is, for any set S, B(S, ε) = {x ∈ Rm | ρ(x, S) < ε}. Similar to the argument above, we
have that, for any 0 < δ2 < d/2, there exists 0 < ε2 < δ2 such that T B(Ω2 (x), ε2 ) ⊂ B(Ω2 (x), δ2 ). Moreover, since Ω1 (x) ⊂
Ω (x), there must exist a sequence {nk } → ∞ such that T nk (x) enters B Ω1 (x), ε1 for infinite times. For any k, there must exist a
positive integer `k > nk such that T `k (x) ∈ B Ω2 (x), ε2 , otherwise Ω2 (x) cannot be a subset of the limit set of x. However, since
T nk (x) ∈ B Ω1 (x), ε1 , T nk +1 (x) cannot be in B Ω2 (x), ε2 . Therefore, for each k ∈ N, there exists `˜k satisfying nk < `˜k < `k such
˜
˜
that T `k (x) ∈
/ B Ω1 (x), ε1 and T `k (x) ∈
/ B Ω2 (x), ε2 . That is, for any k ∈ N,
˜
T `k (x) ∈ T n x \ B Ω1 (x), ε1 ∪ B Ω2 (x), ε2 ,
˜
which is a bounded and closed set. Since there are infinitely many `˜k ’s such that T `k (x) is in the compact set T n x \ B Ω1 (x), ε1 ∪
B Ω2 (x), ε2 , this compact set must contain at least one limit point z ∈ Ω (x), which contradicts the assumption that Ω (x) =
Ω1 (x) ∪ Ω2 (x) since z ∈
/ Ω1 (x) and z ∈
/ Ω2 (x). This concludes the proof of statement 2)d).
n
To prove statement 2)e), we first
Ω (x). We point out that T n (x) approaches Ω (x)
prove by contradiction that nT (x) approaches
n
if and only if ρ T (x), Ω (x) → 0 as n → ∞. Suppose ρ T (x),
Ω (x) does not converges to 0. Then there exists ε > 0 such
that, for any k ∈ N, there exists n ≥ k such that ρ T n (x), Ω (x) ≥ ε. By letting k = 1, 2, . . . , we obtain a subsequence {nk } → ∞
such that ρ T nk (x), Ω (x) ≥ ε for any k. However, since {T nk (x)}k∈N is bounded, there exists a sequence {kr } → ∞ such that
T nkr → y ∈ Ω (x), which implies
ρ T nkr (x), y ≥ ρ T nkr (x), Ω (x) → 0
and contradicts with ρ T nkr (x), Ω (x) ≥ ε for any r. Therefore ρ T n (x), Ω (x) → 0.
Now we proceed to prove that Ω (x) is the smallest set that T n (x) approaches. For any set A, if T n (x) approaches A, then T n (x) also
n
approaches A. Therefore, we only need to discuss
the case when A is a closed set. Suppose T (x) → A as n → ∞. Then we have
n
k
that, for any sequence {nk } → ∞, ρ T (x), A → 0. In addition, for any y ∈ Ω (x), there exists a sequence {nk } → ∞ such that
nk
T nk → y as k → ∞. Suppose y ∈
/ A. SinceA is closed, we have ρ(A, y) > 0. For
such sequence
{nk } that T (x) → y, we already
obtain ρ T nk (x), A → 0 and ρ T nk (x), y → 0. Since ρ(A, y) ≤ ρ T nk (x), A + ρ T nk (x), y for any k ∈ N, we have ρ(A, y) = 0,
which contradicts ρ(A, y) > 0. Therefore, y ∈ A and thus Ω (x) ⊂ A. This concludes the proof of statement 2)e).
t
u
The following lemma presents some important properties of the limit set of any set H, i.e., Ω (H). The proof follows the same line
of argument in the proof of Lemma 7.
Lemma 8 (Invariance and asymptotic properties of limit set Ω (H)). For any continuous map T : Rm → Rm and any set H ⊂ Rm ,
the following statements for the limit set Ω (H) hold:
7
1) For any x ∈ H, Ω (x) ⊂ Ω (H);
2) Ω (H) is closed and positively invariant;
n
3) If ∪∞
n=0 T (H) is bounded, then
a) Ω (H) is non-empty, compact, and invariant;
b) for any x ∈ H, T n (x) approaches Ω (H);
c) Ω (H) is the smallest set that T n (H) approaches as n → ∞.
n
Remark 4. Unlike the limit set of a point, Ω (H) is not necessarily invariantly connected, even if ∪∞
n=0 T (H) is bounded. For
example, consider the following one-dimension first-order difference equation system:
x(n + 1) = T x(n) = εx(n)2 + (1 − ε)x(n),
where x(n) ∈ R for any n ∈ N and 0 < ε < 1. One can easily check that the set of fixed points is {0, 1} and Ω ([0, 1]) = {0, 1}, which
is a union of two disjoint, non-empty, and closed invariant sets: {0} and {1}.
Lemma 9 (Limit set of compact & positively invariant set). For any continuous map T : Rm → Rm , suppose E ⊂ Rm is a compact
and positively invariant set. The following statements hold:
n
1) Ω (E) = ∩∞
n=0 T (E);
2) Ω (E) is non-empty, compact, and invariant;
3) Ω (E) is the largest invariant set in E.
Proof: Since E is compact and T is continuous, T n (E) is compact for any n ∈ N. Since E is positively invariant, T (E) ⊂ E and
T n+1 (E) ⊂ T n (E) for any n ∈ N. Therefore, by definition,
∞
∞
j
n
j
Ω (E) = ∩∞j=0 ∪∞
n= j T (E) = ∩ j=0 T (E) = ∩ j=0 T (E).
This concludes the proof for statement 1).
n
0
Since ∪∞
n=0 T (E) = T (E) = E is bounded, according to statement 3)a) in Lemma 8, Ω (E) is non-empty, compact, and invariant.
This proves statement 2).
Now we prove by contradiction that Ω (E) is the largest invariant set in E. Suppose there exists a non-empty set A ⊂ E such that
A ∩ Ω (E) = φ , and T A ∪ Ω (E) = A ∪ Ω (E).
Since T A ∪ Ω (E) = T (A) ∪ T Ω (E) = T (A) ∪ Ω (E), we have T (E) ⊃ T (A) ⊃ A. In addition,
n
T A ∪ Ω (E) = A ∪ Ω (E) ⇒ T n A ∪ Ω (E) = A ∪ Ω (E), for any n ∈ N. ⇒ ∩∞
n=0 T A ∪ Ω (E) = A ∪ Ω (E).
In the meanwhile,
T A ∪ Ω (E) = T (A) ∪ Ω (E)
⇒
T n A ∪ Ω (E) = T n (A) ∪ Ω (E) ⊂ T n (E) ∪ Ω (E)
∞
n
∞
n
n
∩∞
= ∩n=0 T n (E) ∪ Ω (E) = Ω (E)
n=0 T A ∪ Ω (E) ⊂ ∩n=0 T (E) ∪ T Ω (E)
⇒
A ∪ Ω (E) = Ω (E)
⇒
A ⊂ Ω (E),
⇒
which contradicts A ∩ Ω (E) = φ . This concludes the proof for statement 3).
t
u
3 LaSalle Invariance Principle and Its Extension
With all the preparation work in Section 1 and Section 2, now we are ready to present the main theorem on the original LaSalle
invariance principle for discrete-time dynamical systems. The proof can be found on Page 6 of [11].
Theorem 1 (Discrete-time LaSalle invariance principle). Let G be any set in Rm . Consider a difference equations system defined
by a map T : Rm → Rm that is well-defined for any x ∈ G and continuous at any x ∈ G. Suppose there exists a scalar map V : Rm → R
satisfying
8
i) V (x) is continuous at any x ∈ G;
ii) V T (x) −V (x) ≤ 0 for any x ∈ G.
For any x0 ∈ G, if the solution to the following initial-value problem
(
x(n + 1) = T x(n) ,
x(0)
= x0
−1
satisfies that ∪∞
n=0 {x(n)} is bounded and x(n) ∈ G for any n ∈ N, then there exists
x ∈ R such that x(n) → M ∩V (c) as n → ∞,
−1
m
where V (c) = {x ∈ R |V (x) = c} and M is the largest invariant set in E = x ∈ G V T (x) −V (x) = 0 .
Proof: Let X = ∪∞
n=0 {x(n)}. We have that X ⊂ G is compact. Since V (x) is continuous on G,
V (x) is lower bounded on X. Moreover,
since V x(n) is non-increasing with n for any n ∈ N, there exists c ∈ R such that V x(n) → c as n → ∞.
nk
For any
y ∈ Ω (x0 ), since there exists {nk } → ∞ such that x(nk ) = T (x0 ) → y as k → ∞ and due to the continuity of V , we have
V xnk → V (y). Therefore,
V (y) = lim V x(nk ) = lim V x(n) = c, for any y ∈ Ω (x0 ),
k→∞
which leads to Ω (x0
n→∞
) ∈ V −1 (c).
Moreover, for any y∈ Ω (x0 ), since X is bounded,
according to Lemma 7, Ω (x0 ) is invariant. Therefore, T (y) ∈ Ω (x0 ), which
implies that V T (y) = c and thus V T (y) −V (y) = 0 for any y ∈ Ω (x0 ). Now we obtain Ω (x0 ) ⊂ E. Since Ω (x0 ) ⊂ E and M is
the largest invariant set of E, we have Ω (x0 ) ⊂ M ∩V −1 (c). Finally, since T n (x) approaches Ω (x0 ) as n → ∞, x(n) → M ∩V −1 (c)
as n → ∞.
t
u
The classic LaSalle invariance principle stated in Theorem 1 requires that both T and V are well-defined on G. Below we present a
simple extension of the classic LaSalle invariance principle. This extension establish the converge of the solution x(n) when T and
V are not defined on ∂ G but x(n) is uniformly bounded from ∂ G after some finite time N.
Theorem 2 (Extension of LaSalle invariance principle). Consider the following difference equations system:
x(n + 1) = T x(n) ,
(4)
where T : Rm → Rm is continuous
on some set G ⊂ Rm . Suppose there exists a map V : Rm → R satisfying: i) V (x) is continuous
at any x ∈ G; ii) V T (x) − V (x) ≤ 0 for any x ∈ G. For any x0 ∈ G, if there exists a compact set Gc and N ∈ N such that the
−1
solution x(n) to equation (4) with x(0) = x0 satisfies x(n) ∈ Gc for any n ≥ N, then there exists
c ∈ R such that
x(n) → M ∩V (c)
as n → ∞, where V −1 (c) = {x ∈ Rm |V (x) = c} and M is the largest invariant set in E = x ∈ Gc |V T (x) −V (x) = 0 .
Proof: For any such x0 ∈ G, since x(n) is the solution, we have x(n) = T n (x0 ) for any n ∈ N. According to Lemma 6,
∞
n
n
Ω (x0 ) = ∩∞j=0 ∪∞
n= j {T (x0 )} ⊂ ∪n=N {T (x0 )} ⊂ Gc .
For any n ≥ N, since x(n) ∈ ∪∞j=N {x( j)} ⊂ Gc and V is continuous on Gc , V x(n) is uniformly lower bounded for all n ≥ N. In
addition, since V x(n) is non-increasing, there exists c ∈ R such that limn→∞ V x(t) = c.
For any y ∈ Ω (x0 ), there exists a sequence {nk } → ∞ such that {x(nk )} → y as k → ∞. Since V is continuous on G, we have
V x(nk ) → V (y). Moreover, limn→∞ V (n) = c leads to V (y) = c. Therefore, Ω (x0 ) ⊂ V −1 (c).
According to Lemma 7, Ω (x0 ) is invariant. Therefore, T (y) ∈ Ω (x0 ) for any y ∈ Ω (x0 ), which in turn implies that V T (y) −V (y) =
0 for any y ∈ Ω (x0 ) and thereby Ω (x0 ) ⊂ E. Moreover, since Ω (x0 ) is invariant, Ω (x0 ) ⊂ M. Therefore, Ω (x0 ) ⊂ M ∩ V −1 (c).
Since x(n) = T n (x) approaches Ω (x0 ), x(n) → M ∩V −1 (c) as n → ∞. This concludes the proof.
t
u
4 Advanced Versions of LaSalle Invariance Principle
In this section we provide an incomplete list of references on the extensions and more advanced versions of LaSalle invariance
principle for the interest of further reading. In Section 8, Chapter 1 of [11], the author discusses the vector Lyapunov functions;
9
Cortés et al. [5] propose a sufficient condition for the convergence of discrete-time systems to the fixed points. Hale [7] extends the
LaSalle invariance principles to autonomous systems with infinite dimensions; Shevitz and Paden [13] discusses LaSalle invariance
principle in non-smooth systems; Extensions of LaSalle’s results to switched systems are provided by Hespanha et al. [8], Bacciotti
et al. [2], and Mancilla et al. [12]; Alberto et al. [1] consider the invariance principle for discrete-time dynamical systems with
generalized Lyapunov functions of which the first difference are positive in some bounded regions; Results on Lyapunov functions
and invariance principles for difference inclusions systems can be found in the research articles by Kellett and Teel [9] and Bullo et
al. [3] (see Lemma 4.1), as well as in the book by Bullo et al. [4] (see Theorem 1.21); We refer to the book by Goebel et al. [6] for
a systematic treatment of hybrid dynamical systems.
References
1. L. F. C. Alberto, T. R. Calliero, and A. C. P. Martins. An invariance principle for nonlinear discrete autonomous dynamical
systems. IEEE Transactions on Automatic Control, 52:692–697, 2007. doi:10.1109/TAC.2007.894532.
2. A. Bacciotti and L. Mazzi. An invariance principle for nonlinear switched systems. Systems & Control Letters, 54(11):1109–
1119, 2005. doi:10.1016/j.sysconle.2005.04.003.
3. F. Bullo, R. Carli, and P. Frasca. Gossip coverage control for robotic networks: Dynamical systems on the space of partitions.
SIAM Journal on Control and Optimization, 50(1):419–447, 2012. doi:10.1137/100806370.
4. F. Bullo, J. Cortés, and S. Martı́nez. Distributed Control of Robotic Networks. Princeton University Press, 2009. URL:
http://www.coordinationbook.info.
5. J. Cortés, S. Martı́nez, T. Karatas, and F. Bullo. Coverage control for mobile sensing networks. IEEE Transactions on Robotics
and Automation, 20(2):243–255, 2004. doi:10.1109/TRA.2004.824698.
6. R. Goebel, R. G. Sanfelice, and A. R. Teel. Hybrid Dynamical Systems: Modeling, Stability, and Robustness. Princeton
University Press, 2012.
7. J. K. Hale. Dynamical systems and stability. Journal of Mathematical Analysis and Applications, 26:39–59, 1969. doi:
10.1016/0022-247X(69)90175-9.
8. J. Hespanha, D. Liberzon, D. Angeli, and E. D. Sontag. Nonlinear norm-observability notions and stability of switched systems.
IEEE Transactions on Automatic Control, 50(2):154–168, 2005. doi:10.1109/TAC.2004.841937.
9. C. M. Kellett and A. R. Teel. Smooth Lyapunov functions and robustness of stability for difference inclusions. Systems &
Control Letters, 52:395–405, 2004. doi:10.1016/j.sysconle.2004.02.015.
10. H. K. Khalil. Nonlinear Systems. Prentice Hall, 3 edition, 2002.
11. J. P. LaSalle. The Stability of Dynamical Systems. SIAM, 1976. doi:10.1137/1.9781611970432.
12. J. L. Mancilla-Aguilar and R. A. Garcı́a. An extension of LaSalle’s invariance principle for switched systems. Systems &
Control Letters, 55:376–384, 2006. doi:10.1016/j.sysconle.2005.07.009.
13. D. Shevitz and B. Paden. Lyapunov stability theory of nonsmooth systems. IEEE Transactions on Automatic Control,
39(9):1910–1914, 1994.
14. M. Vidyasagar. Nonlinear Systems Analysis. SIAM, 2002. doi:10.1137/1.9780898719185.
| 3cs.SY
|
1
Distributed Fusion of Labeled Multi-Object
Densities Via Label Spaces Matching
arXiv:1603.08336v1 [cs.SY] 28 Mar 2016
Bailu Wang, Wei Yi, Suqi Li, Lingjiang Kong and Xiaobo Yang
University of Electronic Science and Technology of China, School of Electronic Engineering, Chengdu City,
China
Email: [email protected]
Abstract—In this paper, we address the problem of the
distributed multi-target tracking with labeled set filters in the
framework of Generalized Covariance Intersection (GCI). Our
analyses show that the label space mismatching (LS-DM) phenomenon, which means the same realization drawn from label
spaces of different sensors does not have the same implication,
is quite common in practical scenarios and may bring serious
problems. Our contributions are two-fold. Firstly, we provide a
principled mathematical definition of “label spaces matching (LSDM)” based on information divergence, which is also referred
to as LS-M criterion. Then, to handle the LS-DM, we propose a
novel two-step distributed fusion algorithm, named as GCI fusion
via label spaces matching (GCI-LSM). The first step is to match
the label spaces from different sensors. To this end, we build a
ranked assignment problem and design a cost function consistent
with LS-M criterion to seek the optimal solution of matching
correspondence between label spaces of different sensors. The
second step is to perform the GCI fusion on the matched label
space. We also derive the GCI fusion with generic labeled multiobject (LMO) densities based on LS-M, which is the foundation
of labeled distributed fusion algorithms. Simulation results for
Gaussian mixture implementation highlight the performance of
the proposed GCI-LSM algorithm in two different tracking
scenarios.
I. I NTRODUCTION
Compared with centralized multi-object tracking methods,
distributed multi-sensor multi-object tracking (DMMT) methods generally benefit from lower communication cost and
higher fault tolerance. As such, they have increasingly attracted interest from tracing community. When the correlations
between the estimates from different sensors are not known,
devising DMMT solutions becomes particularly challenging.
The optimal fusion to this problem was developed in [1], but
the computational cost of calculating the common information
can make the solution intractable in practical applications.
An alternative is to use suboptimal fusion technique, namely,
Generalized Covariance Intersection (GCI) or exponential
mixture Densities (EMDs) [2] pioneered by Mahler [3]. The
highlight of GCI is that it is capable to fuse both Gaussian and
non-Gaussian formed multi-object distributions from different
sensors with completely unknown correlation.
Based on GCI fusion rule, distributed fusion with the
probability hypothesis density [4], [5] (PHD)/cardinalized
PHD [6], [7] and multi-Bernoulli [8]–[14] filters has been
explored in [3], [15]–[20]. However, the aforementioned filters
on one hand are not multi-object trackers as target states
are indistinguishable, and on the other hand are almost not
the closed-form solution to the optimal Bayssian filter even
though a special observation model, i.e., standard observation
model [19] is assumed. Recently, the notion of labeled random
finite set (RFS) is introduced to address target trajectories and
their uniqueness in [21]–[27]. Vo et al. proposed a class of
generalized labeled multi-Bernoulli (GLMB) 1 densities which
is a conjugate prior and also closed under the ChapmanKolmogorov equation for the standard observation model in
Bayesian inference. Moreover, the relevant stronger results,
δ-GLMB filter, which can be directly used to multi-target
tracking, can not only produce trajectories formally but also
outperform the aforementioned filters. Except for the standard
observation model, the labeled set filter also has achieved
some good results for generic observation model. In [27],
Papi et al. proposed a δ-GLMB density approximation of the
labeled multi-object (LMO) density and developed an efficient
δ-GLMB filter for the generic observation model. [27] also
provides a further detailed expression for the universal LMO
density, which is the product of the joint existence probability
of the label set and the joint probability density of states
conditional on their corresponding labels.
Due to the advantages of labeled set filters, it is meaningful
to investigate their generalization to the distributed environment. In [28], Fantacci et al. derived the closed-form solutions
of GCI fusion with marginalized δ-GLMB (Mδ-GLMB) and
labeled multi-Bernoulli (LMB) posteriors, and highlight the
performance of the relevant DMMT algorithms based on the
assumption that different sensors share the same label space.
However, our analyses show that this assumption is hard to
be satisfied in many real world applications. In other word,
the label spaces of each sensors always mismatch in the sense
that the same realization drawn from label spaces of different sensors does not have the same implication in practical
scenarios, which is referred to as “label space mismatching
(LS-DM)”2. When LS-DM happens, the direct fusion with the
labeled posteriors from different sensors will exhibit a counterintuitive behavior: the fusion will be performed between
objects with different labels, making the fusion performance
poor. Therefore, there is a lack of robustness in practice if one
perform fusion with labeled posteriors from different sensors
directly.
1 GLMB distribution is also simply named as Vo-Vo distribution by Malher
in his book [30] first time.
2 The letter “D” in the abbreviation of label space mismatching means
double “M”, i.e., “miss” and “matching”.
2
To get rid of the bad influences of LS-DM, two promising
thoughts can be employed: one is to perform GCI fusion with
unlabeled version of posteriors from different sensors, which
is firstly proposed in [29], and the other is to match the label
spaces of different sensors and then perform GCI fusion on
the matched label space. This paper focuses on the latter and
our contributions are two-fold:
i) We provide a principled mathematical definition for “label space matching” based on information divergence.
This definition also provides a criterion to judge whether
the label spaces are matching or not. Moreover to make
this criterion have practicality, we derive the specified
expression of set marginal density for single-object case.
ii) We proposed a two-step distributed fusion algorithm,
namely, GCI fusion with LMO densities via label spaces
matching (GCI-LSM for short). First step is to match the
label spaces from different sensors. To this end, the ranked
assignment problem is built to seek the optimal solution of
the matching correspondence with the cost function based
on LS-M criterion. Then, perform GCI fusion with LMO
densities on matched label space. In addition, we derive
the GCI fusion with the generic LMO density based on
the assumption that label spaces are matching, which is
the foundation of many labeled DMMT..
In numerical results, the performance of the proposed fusion
algorithm with Gaussian mixture (GM) implementation is
verified.
II. BACKGROUND
A. Notation
In this paper, we inhere the convention that single-target
states are denoted by the small letter “x”, e.g., x, x and
the multi-target states are denoted by capital letter “X”, e.g.,
X, X. To distinguish labeled states and distributions from the
unlabeled ones, bold face letters are adopted for the labeled
ones, e.g., x, X, π. Observations generated by single-target
states are denoted by the small letter “z”, i.e., z, and the multitarget observations are denoted by capital letter “Z”, i.e., Z.
Moreover, blackboard bold letters represent spaces, e.g., the
state space is represented by X, the label space by L, and the
observation space by Z. The collection of all finite sets of X
is denoted by F (X) and Fn (X) denotes all finite subsets with
n elements.
The labeled single target state x is constructed by augmenting a state x ∈ X with an label ℓ ∈ L. The labels are usually
drawn form a discrete label space, L = {αi : i ∈ N}, where
all αi are distinct and the index space N is the set of positive
integers.
The multi-target state X, the labeled multi-target state X
and the multi-target observation Z are modelled by the finite
set of single-target states, the finite set of labeled single-target
states, and the finite set of observations generated by singletarget states, respectively, i.e.,
X ={x1 , · · · , xn } ⊂ X
X ={x1 , · · · , xn } ⊂ X × L
Z ={z1 , · · · , zm } ⊂ Z
(1)
We use the multi-object exponential notation
Y
hX ,
h(x)
(2)
x∈X
for real-valued function h, with h∅ = 1 by convention. To
admit arbitrary arguments like sets, vectors and integers, the
generalized Kronecker delta function is given by
1, if X = Y
δY (X) ,
(3)
0, otherwise
and the inclusion function is given by
1, ifX ⊆ Y
1Y (X) ,
0, otherwise
(4)
If X is a singleton, i.e., X = {x}, the notation 1Y (X) is used
instead of 1Y ({x}).
Also notice that the labeled multi-target state is an RFS on
X × L with distinct labels. The set of labels of an labeled RFS
X is given by L(X) = {L(x) : x ∈ X}, where L : X×L → L
is the projection defined by L((x, ℓ)) = ℓ. The distinct label
indicator
△ (X) = δ|X| (|L(X)|)
(5)
B. Labeled Multi-Object Density
For an arbitrary labeled RFS, its LMO density can be
represented as the expression given in Lemma 1 [27].
Lemma 1. Given an labeled multi-object density π on F (X×
L), and for any positive integer n, we define the joint existence
probability of the label set {ℓ1 , ℓ2 , · · · , ℓn } by
Z
w({ℓ1 , · · · , ℓn }) = π({(x1 , ℓ1 ),· · · ,(xn , ℓn )})d(x1 ,· · ·, xn )
(6)
and the joint probability density on Xn of the states x1 , · · · , xn
conditional on their corresponding labels ℓ1 , · · · , ℓn by
π({(x1 , ℓ1 ), · · · , (xn , ℓn )})
w({ℓn , · · · , ℓn })
(7)
Thus, the LMO density can be expressed as
P ({(x1 , ℓ1 ), · · · , (xn , ℓn )}) =
π(X) = w(L(X))P (X).
(8)
C. Multi-target Bayesian filter
Finite Set Statistics (FISST) proposed by Mahler, has provided a rigorous and elegant mathematical framework for the
multi-target detection, tracking and classification problem in
an unified Bayesian paradigm.
In the FISST framework, the optimal multi-target Bayesian
filter 1 propagates RFS based posterior density πk (Xk |Z 1:k )
conditioned on the sets of observations up to time k, Z 1:k , in
time with the following recursion [19]:
πk|k−1 (Xk |Z 1:k−1 )
Z
= fk|k−1 (Xk |Xk−1 )πk−1 (Xk−1 |Z 1:k−1 )δXk−1 ,
(9)
1 Note that the multi-object Bayesian filter in (9) and (10) is also appropriate
for the labeled set posterior, and the labeled set integrals defined as [21] are
involving.
3
πk (Xk |Z 1:k ) = R
gk (Zk |Xk )πk|k−1 (Xk |Z 1:k−1 )
(10)
gk (Zk |Xk )πk|k−1 (Xk |Z 1:k−1 )δXk
where fk|k−1 (Xk |Xk−1 ) is the multi-target Markov transition
function andR gk (Zk |Xk ) is the multi-target likelihood function
of Zk , and ·δX denotes the set integral [19] defined by
Z
Z
∞
X
1
f ({x1 , · · · , xn })dx1 · · · dxn . (11)
f (X)δX =
n!
n=0
D. GCI Fusion Rule
The GCI was proposed by Mahler specifically to extend
FISST to distributed environments [9]. Consider two nodes 1
and 2 in the sensor network. At time k, each nodes maintain
its own local posteriors π1 (X|Z1 ) and π2 (X|Z2 ) which are
both the RFS based densities. Under the GCI 2 proposed by
Mahler, the fused distribution is the geometric mean, or the
exponential mixture of the local posteriors [3],
πω (X|Z1 , Z2 ) = R
π1 (X|Z1 )ω1 π2 (X|Z2 )ω2
π1 (X|Z1 )ω1 π2 (X|Z2 )ω2 δX
(12)
where ω1 , ω2 (ω1 + ω2 = 1) are the parameters determining
the relative fusion weight of each distributions.
Eq. (12) is derived by following that the distribution that
minimizes the weighted sum of its Kullback-Leibler divergence (KLD) with respect to a given set of distributions is
an EMD [3],
πω = arg min(ω1 DKL (π k π1 ) + ω2 DKL (π k π2 ))
π
where DKL denotes the KLD with
Z
f (X)
DKL (f ||g) , f (X) log
dX.
g(X)
(13)
(14)
Note the integral in (14) must be interpreted as a set integral.
III. L ABEL S PACE M ISMATCHING P HENOMENON AND
P ROMISING S OLUTIONS
The GCI formula in (12) is generally computationally
intractable for the set integrals need to integrate over all
joint state spaces, considering each cardinality (number of
objects). Fortunately, it is tractable to derive the closedform solutions for GCI fusion with many simplistic labeled
densities including LMB and Mδ-GLMB densities, which can
simplify the G-CI formula largely. However, these closedform solutions are derived based on the assumption that the
label spaces of different local labeled set filters are matching,
and this assumption is really harsh in practice making these
solutions restrictive in realworld DMMT.
In this section, we firstly analyze the causes of label spaces
mismatching (LS-DM) phenomenon in terms of two popular
birth procedures, and then provide two novel methods to solve
this challenge problem.
2 Note that GCI fusion rule in (12) is also appropriate for the labeled set
posterior, and the labeled set integrals defined as [21] are involving.
A. LS-DM Phenomenon
The essential meaning of LS-DM is that the same realization
drawn from label spaces of different sensors does not have the
same implication. The underlying implication LS-DM is that
the posterior spatial distributions of the same object in different
sensors only have a tiny discrepancy. This phenomenon is quite
common in labeled DMMT. It may originate from any time
steps during the recursion of multi-object filtering and fusing,
and will last during the subsequent time steps in many cases.
Naturally, the birth procedure has a decisive influence on the
matching of label spaces of different sensors. Hence, in the
following, we analyze the causes of LS-DM in terms of two
popular birth procedures.
(1) Adaptive birth procedure (ABP) [22]. This birth procedure
is widely used, in which new-born targets are based on the
observations not associated to the persisting targets. Due
to the randomness of the observations, it is really difficult
to guarantee that the same births of different local set
filters are labeled using the same object label. In addition,
the observation sets provided by different sensors incorporating noisy observations of objects, stochastic missdetections, and stochastic clutters are also contributed to
the LS-DM of persisting objects. For instance, if sensor 1
loses a object due to miss-detection and re-initiate it later,
while sensor 2 keep locking on this object always, then
the mismatching of the label of this object will arise.
(2) Priori knowledge based on birth procedure (PBP) [8].
This birth procedure is often used in some well-known
scenarios with the priori of positions of object births, e.g.,
entrance of marketplace, airport etc. Generally, the object
label has two dimension, ℓ = (k, i), where k is the time of
birth, and i is a unique index to distinguish objects. The
priori of the born positions for PBP can provide reference
for the object index ℓ, but contribute little to the birth
time k, hence there still exists a chance of mismatching
for the births since it is easily effected by the uncertain of
measurement noise, clutter and the variance of the prior of
the born position. In addition, due to that persisting objects
may be wrongly dominated by clutters, or be truncated due
to miss-detection in the following time steps, the LS-DM
for persisting objects also happen sometimes.
The above analyses suggest that for both ABP and PBP,
it is difficult to ensure different sensors share the same label
space. Note that PBP suffer less from LS-MM than ABP for
it can use the prior information as a reference. In a word,
to ensure the matching of label spaces of each sensors, an
ideal detecting environment, in which each sensor dose not
have miss-detections and clutters, and the estimate accuracy
of each sensor is enough high, is required.
B. Promising Solutions for LS-DM
To break away the bad influence of LS-MM, we propose
two solutions:
• The first method [29] is that the GCI fusion is performed
on unlabeled state space via transforming the labeled RFS
densities to their unlabeled versions. Therefore, this fusion
method has robustness. For the GLMB family, we had
4
proved that their unlabeled versions is the generalized multiBernoulli (GMB) distributions [29], and the GCI fusion with
GMB distributions (GCI-GMB) is also proposed in [29].
• The second method is that firstly match the label spaces
from different sensors, then perform GCI fusion with labeled densities on macthed labeled state space as shown in
Fig 1. This approach is referred to as GCI fusion with label
space matching (GCI-LSM).
This paper mainly focuses on the GCI-LSM fusion method.
Labeled RFS
Density 1
Labeled RFS
Density 2
Label Space
Matching
share the same
label space
GCI
Fusion
Fused
Density
Fig. 1. GCI-LSM: the label spaces of different sensors are matched through
some means each time firstly, then perform GCI fusion on labeled state space.
IV. GCI F USION
VIA
L ABEL S PACE M ATCHING
Based on the second solution of LS-DM, label spaces
matching and the closed-form solution of GCI fusion with
LMOs are the two key points need to be addressed. To clearly
describe the concept of label space matching (LS-M), we
firstly give the mathematical definition of label spaces matching based on information divergence, also referred to as LS-M
criterion, which is the foundation of GCI-LMS fusion method.
Then by solving the built ranked assignment problem about
the matching relationship of objects between different sensors,
we then get matched label spaces. Finally with this condition
different label spaces are matched, the closed-form solution of
GCI fusion with universal LMO densities is derived.
A. The Mathematical Description of Label Space Matching
In order to clearly describe “label space matching”, we
formulate it in a rigorous mathematical model shown in
Definition 1. Definition 1 also provides a criterion to judge
whether the label spaces from different sensors are matching
or not, and we call this criterion as LS-M criterion.
Definition 1. Consider a scenario that sensor 1 and sensor
2 observe the same spatial region. Suppose that π1 (·) and
π 2 (·) with state space X and label space Li are the multiobject posteriors of sensor 1 and sensor 2 respectively. The
RFS Ψi , i = 1, 2 with the probability density πi (·) can be
represented as the union
] (ℓ)
Ψi =
(15)
ψi
ℓ∈L
(ℓ)
ψi
with the state space X × {ℓ} is the random finite
where
subset of Ψ.
Then L1 and L2 are said to be matching only if
(ℓ)
(ℓ)
L1 = L2 and ∀ ℓ ∈ L1 , D( π 1 || π 2 ) ≤ Γm
(ℓ)
(16)
where π s (·) is the probability density of ψ (ℓ)
s , ℓ ∈ Ls , s =
(ℓ)
(ℓ)
1, 2, Γm is a given threshold, and D( π 1 (·)|| π2 (·)) describes the “distance” between two distributions.
The condition that L1 = L2 demands that both the cardinalities and each elements of the label spaces of sensor 1
and sensor 2 have the same values. Note ψ (ℓ)
s is the random
(ℓ)
finite subset related the object with label ℓ, and π s is
the probability density for object ℓ in sensor s. Hence, the
(ℓ)
(ℓ)
condition D( π 1 || π2 ) < Γm demands that the densities
of object ℓ in sensors 1 and 2 have a slight difference,
which ensures that the object label ℓ of sensors 1 and 2 are
(ℓ)
matching. As π s incorporates all the statistical information
about object ℓ, and thus it is reasonable to judge the matching
relationship of label ℓ based on its probability densities. In
a word, to match the label spaces L1 and L2 , one-to-one
object matching constrain should be satisfied between different
sensors.
Remark 1. The parameter Γm in (16) is a given threshold,
and the slighter its values is, the harsher the LS-M criterion is.
In ideal case, Γ0 which means the objects of different sensors
share the same density. The “distance” D(f ||g) in (16) usually
chooses information-based divergences including KLD, Rényi,
Csiszár-Morimoto (AliSilvey), and Cauchy-Schwarz, etc., to
measure the similarlies/differences of different multi-object
density.
A key point of using the LS-M criterion is to compute
(ℓ)
the probability density πs (·) from the global multi-object
(ℓ)
density π s (·). π s (·) is also called as set marginal density
(ℓ)
of π s (·) with respective to π s (·). In [31], we preliminarily
give the concept of the set marginal density as shown in
Definition 3, its generalized computing method as shown
in Lemma 1, and its specified computing method for joint
multi-Bernoulli RFS. In [32], the set marginal density is
extened to labeled multi-object density. In this section, we
derive the specified expression for the set marginal density of
(ℓ)
labeled random finite subset π s with respective to π s in
Proposition 1. Proposition 1 guarantees the practicability of
the LS-M criterion.
Definition 2. Let Ψ be an RFS. Then for any random finite
subset of Ψ, denoted by ψ, its set density function fψ (X), is
called set marginal density of ψ with respect to Ψ.
Lemma 2. Let Ψ be an RFS. Then for any random finite
subset of Ψ, denoted by ψ, its set marginal density of ψ with
respect to Ψ, denoted by fψ (X) can be derived as
fψ (X) =
δ Pr(ψ ⊆ S, Ψ/ψ ⊆ X)
δX
(17)
S=∅
where Ψ/ψ , {x|x ∈ Ψ and x ∈
/ ψ} and “δ/δX” denotes a
set derivative [19].
Remark 2. Eq. (17) (Lemma 2) makes us convenient get a
set marginal density of a random finite subset of an labeled
RFS Ψ. Indeed, the local statistical properties of an labeled
RFS can be learned by the set marginal density. Also, the
relations of label spaces or correlations among different RFSs
densities can be known via analyzing the relevances of their
corresponding set marginal densities.
According Proposition 2 in [32], the set marginal density
5
with single object space is derived in the following,
Proposition 1. Given π s (·) = ws (L(·))Ps (·) be the multiobject posterior of sensor s. For any ℓ ∈ Ls , the set marginal
density of its corresponding subset ψ (ℓ)
on space X × {ℓ}
s
(ℓ)
is an labeled Bernoulli distribution with parameters π s =
(ℓ) (ℓ)
{(rs , ps )} are shown as
X
1I (ℓ)ws (∪I),
(18)
rs(ℓ) =
I∈F (Ls )
p(ℓ)
s
=
1
r(ℓ)
X
1I (ℓ)ws (∪I)pI−{ℓ} ({x, ℓ})
(19)
I∈F (Ls )
where
p{ℓ1 ,··· ,ℓ} ({x, ℓ})
(20)
Z
= Ps ({(x, ℓ), (x1 , ℓ1 ), · · · , (xn , ℓn )})dx1 , · · · , dxn .
(21)
Remark 3. Proposition 1 indicates that the set marginal density of each ψ (ℓ)
s , ℓ ∈ Ls is an labeled Bernoulli distribution.
The class of Bernoulli densities own a congenital advantage
that it can get tractable results for information divergence
(ℓ)
(ℓ)
generally making the computation of D( π1 || π 2 ) < Γm
simplistic, thus enhance the practicability of the LS-M criterion largely.
B. Label Space Matching via Ranked Assignment Problem
Section III-B showed that LS-DM phenomenon is quite
common in practical scenarios. Actually, for different sensors
observing the same spatial region, the tracks of a sensor have
only one definite correspondence with the tracks of another
sensor in ideal case, consistent with Γm = 0 in (16). However, due to the influence of the stochastic noise, there exist
great uncertainty for the matching correspondence. Then the
problem of seeking the solution of matching correspondence
is essentially a optimization problem.
In this section, we firstly provide the mathematical representation of the matching correspondence between different
sensors using a mapping function. Then based on the LS-M
criterion given in Definition 1, we build a ranked assignment
problem and design a principle cost function to seek the
solution of optimal matching correspondence, where the information divergence employs the Rényi Divergence (RD) which
is the generalized form of the KLD with the free parameter
α → 1.
Definition 3. A fusion map is a function τ : L1 → {0} ∪ L2
′
′
such that τ (i) = τ (i ) > 0 implies i = i . The set of all such
fusion maps is called the fusion map space denoted by T .
label space. This can be accomplished by solving the following
ranked assignment problem.
Enumerating L1 and L2 , each fusion map τ ∈ T (L1 ) can be
represented by an |L1 | × |L2 | assignment matrix S consisting
of 0 or 1 entries with every row and column summing to either
1 or 0. For i ∈ {1, . . . , |L1 |}, j ∈ {1, . . . , |L2 |}, Si,j = 1 if
(i)
(j)
and only if track ℓ1 of sensor 1 is assigned to track ℓ2
(i)
(j)
of sensor 2, i.e. τ (ℓ1 ) = ℓ2 . An all-zero row i means that
(i)
track ℓ1 of sensor 1 is a false track or the corresponding track
of sensor 2 is misdetected while all-zero column j means that
(j)
track ℓ2 of sensor 2 is a false track or the corresponding track
of sensor 1 is misdetedted. Conversion from the assignment
P|L1 | (j)
(i)
ℓ2 δ1 (Si,j ).
(matrix) S to τ is given by τ (ℓ1 ) = j=1
The cost matrix of an optimal assignment problem is the
|L1 | × |L2 | matrix:
C1,1
· · · C1,|L1 |
..
..
..
CL1 ,L2 =
.
.
.
C|L1 |,1
· · · C|L1 |,|L2 |
where for i ∈ {1, . . . , |L1 |}, j ∈ {1, . . . , |L2 |}, Ci,j is the cost
(j)
(i)
of the assigning the ℓ2 th track of sensor 2 to ℓ1 th track of
sensor 1.
According the Definition 1, if two label spaces L1 and
L2 are matched, the distance between arbitrary two single
object densities (Bernoulli density) indicating the same true
object from the two sensors respectively is tiny enough. RD,
specifically, for α = 0.5, equals the Hellinger affinity, and thus
the cost selection criterion becomes the equality of Hellinger
distance, which can be used to describe the distance between
two densities, which is also consistent with the LS-M criterion.
The Proposition 1 shows that single object density follows
labeled Bernoulli distribution, thus using the formula for Renyi
divergence between two Bernoulli distributions, it can easily
be shown that
(ℓ′ )
(ℓ)
Ci,j =
Rα ( π 1 (X)|| π2 (X))
Z
1
(ℓ)
(ℓ′ )
π 1 (Xn )(1−α) π 2 (Xn )α dX
log
=
α−1
X
Z
∞
X
1
1
(ℓ)
=
log
π 1 ({x1 , · · · , x2 })α
α−1
n!
n=0
(ℓ′ )
π 2 ({x1 , · · · , x2 })1−α dx1 · · · dxn
=
1
(ℓ)
(ℓ′ )
log((1 − r1 )α (1 − r2 )1−α +
α−1
Z
(ℓ) α (ℓ′ ) 1−α
r2
r1
(ℓ)
(ℓ′ )
p1 (x)α p2 (x)1−α dx).
(23)
(ℓ)
Each fusion map τ ∈ T describes one possible (hypothesis)
matching relationship of different label spaces,
where π s (X), s = 1, 2, is the set marginal density provided
in Proposition 1.
{(ℓ, τ (ℓ))}ℓ∈L1
The cost of S is the combined costs of every true track
of sensor 1 to the track of sensor 2, which can be succinctly
written as the Frobenius inner product
(22)
and the number of fusion maps grows exponentially with the
number of objects.
Due to the uncertainty of τ ∗ , we need to seek the optimal
estimation of τ ∗ in order to perform GCI fusion on matching
T
J(S) = tr(S CL1 ,L2 ) =
|L1 | |L2 |
X
X
i=1 j=1
Ci,j Si,j .
6
The optimal assignment problem seeks an assignment matrix S∗ (τ ∗ ) that minimizes the cost function J(S):
S∗ = arg min J(S)
S
(24)
where S∗ (τ ∗ ) denotes the assignment matrix for the best
matching hypothesis or mapping case.
By solving the equation (24) using Murty’s algorithm [33],
the true matching hypothesis τ ∗ is specified, then the consensual label space is given by
Lc = {(ℓ1 , τ ∗ (ℓ1 ))|τ ∗ (ℓ1 ) > 0}ℓ1 ∈L1 .
The tracks in one sensor with no corresponding matched tracks
in another sensor are leaved out considering the uncertainty of
them.
Remark 4. The optimal τ ∗ establishes the optimal solution
of one-to-one matching correspondence between two label
spaces, hence the consensual label space Lc is obtained, which
makes the assumption that different sensors share the same
label space come true.
C. GCI Fusion with Labeled Multi-Object Density
When fusing LMO densities via the GCI rule (12), the
main challenge is that the GCI formula is computationally
intractable due to the set-integral that integrates over all joint
target-spaces. However, when the condition of different label
spaces matching is hold, the problem of GCI fusion with LMO
densities (GCI-LMO) is great simplified, for it doesn’t need
to consider all possible matching correspondence between
different label spaces [29].
In Proposition 2, we derived the GCI-fusion for generic
LMO density based on the matching of label spaces.
Proposition 2. Let πs (X) = ws (L(·))Ps (·) be the labeled
multi-object posterior of sensor s, s = 1, 2, and their label
spaces L1 and L2 are matching. Then the distributed fusion
with π 1 (X) and π2 (X) via GCI rule in (12) is given by
πω (X) = wω (L(X))pω (X)
(25)
where
w1ω1 (I)w2ω2 (I)ηω (I)
,
ω1
ω2
I∈F (L1 ) w1 (I)w2 (I)ηω (I)
wω (I) = P
pω (X) =
ω2
1
pω
1 (X)p2 (X)
ηω (I)
(26)
(27)
LMB and Mδ-GLMB respectively (GCI-LMB and GCI-MδGLMB).
D. Pseudo-Code
A pseudo-code of the proposed GCI-LSM fusion algorithm
is given in Algorithm 1.
Algorithm 1: The proposed GCI-LSM fusion.
Inputs: Receive posteriors π s from nodes s = 1 : Ns ;
(ℓ)
Step 1: Calculate the set marginal density of πs with
respect to πs according to Proposition 1;
Step 2: Perform GCI-LSM fusion by adopting iteration
method:
Initial: π ω = π1 ;
for s=2:NS do
1) Obtain the consensual label space Lc of π s and
πw according to (24);
2) Perform GCI-LSM fusion with πω and πs
according to (25), then output the fused posterior
πω ;
end
Return: the fused posterior π ω ({x1 , . . . , xn }) in the
form of (8).
V. G AUSSIAN M IXTURE I MPLEMENT
We now detail the computation of the cost function Ci,j
in (23) for the ranked assignment problem of the GCI-LSM
fusion for special formed LMO densities (e.g., LMB and
Mδ-GLMB). In the present work, each single-object density
(ℓ)
conditional on its existence ps (x) of sensor s is represented
by a GM of the form
(ℓ)
p(ℓ)
s (x) =
j=1
ws(ℓ,j) N x; ms(ℓ,j) , Ps(ℓ,j) .
where
(i)
K=
(j)
J1 J2
X
X
wω ,
(31)
m=1 n=1
(28)
2
pω
2 ({(x1 , ℓ1 ), · · · , (xn , ℓn )})d(x1 , · · · , xn ).
Especially, for special formed LMO densities (e.g., LMB
[23] and Mδ-GLMB [26] densities), their closed-form solutions had been shown in [28] under the assumption that
different label spaces share the same birth space, and their
corresponding GCI fusion is referred as GCI fusion with
(29)
Since the calculation of cost function C involves exponentiation of GMs which, in general, do not provide a GM. To
preserve the GM form, a suitable approximation of the GM
exponentiation proposed in [17] is adopted. Thus, Eq. (23)
turns out to be
1
(i)
(j)
(i)
(j)
C i,j = − log (q1 )α (q2 )β + (r1 )α (r2 )β K
(30)
β
with
ηω ({ℓ1 , · · · , ℓn })
Z
1
= pω
1 ({(x1 , ℓ1 ), · · · , (xn , ℓn )})
Js
X
wω = w
eω N
(i,m)
P
(i,m)
(j,n)
m1
−m2 ; 0, 1
ω1
(i,m) ω1
w
eω = (w1
(j,n) ω2
) (w2
(i)
,
(32)
,ω2 )
(33)
(i,m)
(j,n)
(j)
(j)
) ρ(P1
(i)
!
(j,n)
P
+ 2
ω2
,ω1 )ρ(P2
and β = 1 − α, q1 = 1 − r1 , q2 = 1 − r2 , ρ(P, ω) =
p
det[2πP ω −1 ](det[2πP ])−ω .
Moreover, the GM implementation of the GCI fusion for
special formed LMO densities (e.g., LMB and Mδ-GLMB)
7
under the assumption that different sensors share the same
label space can refer to (55) in [17].
600
400
where In and 0n denote the n × n identity and zero matrices,
∆ = 1 second (s) is the sampling period, and σν = 5m/s2
is the standard deviation of the process noise. The probability
of target survival is PS,k = 0.99; The probability of target
detection in each sensor is independent of the probability of
detection at all sensors and is PD = 0.99. The single-target
observation model is also a linear Gaussian with
Hk = I2 02 , Rk = σε2 I2 ,
where σε = 1.4m, is the standard deviation of the measurement noise. The number of clutter reports in each scan is
Poisson distributed with λ = 10. Each clutter report is sampled
uniformly over the whole surveillance region.
The parameters of GM implementation have chosen as
follows: the truncation threshold is γt = 10−4 ; the prune
threshold is γp = 10−5 ; the merging threshold is γm = 4; the
maximum number of Gaussian components is Nmax = 10.
All performance metrics are given in term of the optimal subpattern assignment (OSPA) error [34].
200
T2: born at 20s, dies at 60s
0
T1: born at 0s, dies at 40s
−200
−400
Communication line
−600
Sensor 1
−800
−800 −600
Sensor 2
−400 −200
0
200
400
600
800
X coordinate (m)
Fig. 2. The scenario of simple distributed sensor network with two sensors
tracking two targets.
of the Bernoulli birth distribution at time k + 1 depending on
the measurement zk is proportional to the probability that zk
is not assigned to any target during the updated at time k:
!
1 − rU,k (z)
rB,k+1 (z) = min rB,max , P
· λB,k+1
ξ∈Zk 1 − rU,k (ξ)
(34)
where
X
rU,k =
I ,θ
1θ (z)wk+
(35)
(I+ ,θ)∈F (L+)×ΘI+
I ,θ
with wk+ is given by (59) in [23], and λB,k+1 is the expected
number of target birth at time k + 1 and rB,max ∈ [0, 1] is the
maximum existence probability of a new born target.
100
GCI−LSM
GCI−LMB
80
OSPA [m]
The performance of the proposed GCI-LSM fusion is evaluated in two 2-dimensional multi-object tracking scenarios. The
GCI-LSM is implemented using the GM approach proposed
in Section V. Since this paper does not focus on the problem
of weight selection, we choose the Metropolis weights for
convenience (notice that this may have an impact on the fusion
performance). The LMB filter is adopted by local filters. The
efficiency of LMB filter has been demonstrated in [23].
All targets travel in straight paths and with different but
constant velocities. The number of targets is time varying due
to births and deaths. The following target and observation
models are used. The target state variable is a vector of plannar
position and velocity xk = [px,k , ṗx,k , py,k , ṗy,k ]⊤ , where
“⊤ ” denotes the matrix transpose. The single-target transition
model is linear Gaussian specified by
1
I2 12 ∆I2
I2 ∆I2
2
4
, Qk = σv 1
Fk =
02 I2
I2
3 02
Y coordinate (m)
VI. P ERFORMANCE A SSESSMENT
60
40
20
A. Scenario 1
To demonstrate the effectiveness of GCI-LSM fusion, the
performance of GCI-LSM fusion is compared with GCI-LMB
[28] (under the assumption that different sensors share the
same label space) in two experiments with ABP and PBP used
respectively For this purpose, an simple scenario involving two
sensors and two objects is considered as shown in Fig 2. The
duration of this scenario is Ts = 60s.
1) Experiment 1: The preceding analyses show that the LSDM phenomenon arises frequently when the ABP is adopted
by local filters. To prove this and the effectiveness of the
proposed GCI-LSM fusion, the GCI-LSM fusion is compared
with GCI-LMB fusion under the ABP situation.
The adaptive birth procedure proposed in [23] is employed
for this scenario. More specifically, the existence probability
0
10
20
30
Time k
40
50
60
Fig. 3. ABP: OSPA errors of GCI-LSM and GCI-LMB fusion algorithms
with order p = 1 and cut-off c = 100 with adaptive birth (200 MC runs).
Fig. 3 illustrates that the performance of GCI-LSM fusion
is significant better than GCI-LMB fusion. Since the method
of the ABP depends on the observations with randomness, the
labels of the birth targets each time are also randomness with
their corresponding observations. This result of GCI-LMB
fusion shows that the ABP leads to the LS-DM frequently, and
it is necessity to match the label spaces from different sensors
to ensure the consensual between them, or the performance
of the GCI fusion will collapse. The result of GCI-LSM also
evidences this viewpoint. It can seen that once the LS-DM is
8
removed, the GCI-fusion perform really excellent, exactly as
the GCI-LSM fusion. The outstanding performance of GCILSM also gets benefit from the well-designed cost function in
ranked assignment problems.
600
Y coordinate (m)
400
2) Experiment 2: This experiment analyzes the problem of
the LS-DM under PBP situation. The preceding analyses show
that the PBP also suffers from the LS-DM even though it can
obtain some priors about the births. In this experiment, the
performance of GCI-LSM and GCI-LMB fusion is compared
using PBP [23].
T2
T1
0
−200
T3
−400
Communication line
Sensor 2
Sensor 3
−600
Sensor 1
600
600
True
Estimated Tracks of Sensor 1
400
Y coordinate (m)
200
400
600
800
Birth with label (1,1)
200
Fig. 5. The scenario of distributed sensor network with three sensors tracking
five targets. T1 born at 0s, dies at 55s; T2 born at 0s, dies at 55s; T3 born at
10s, dies at 65s, T4 born at 25s, dies at 65s; T5 born at 40s, dies at 65s.
0
−200
Birth with label (1,1)
−400
−600
−800 −600 −400 −200
0
200 400
X coordinate (m)
600
800
−600
−800 −600 −400 −200
0
200 400
X coordinate (m)
600
B. Scenario 2
800
To further test the performance of the proposed GCI-LSM
fusion in challenging scenarios, a sensor network scenario
involving five targets is considered as shown in Fig. 5. In
the experiment, the proposed GCI-LSM fusion is compared to
the GCI-GMB fusion mentioned in Section III-B [29] and the
GCI fusion with PHD filter (GCI-PHD) [16]. Both GCI-LSM
and GCI-GMB fusion use ABP introduced in scenario 1 and
the adaptive birth distribution of GCI-PHD is introduced in
[5]. The duration of this scenario is Ts = 65s.
(b)
(a)
100
True
Estimated Tracks of GCI−LMB
400
GCI−LSM
GCI−LMB
80
OSPA [m]
200
0
60
40
−200
20
−400
−600
−800 −600 −400 −200
0
200 400
X coordinate (m)
600
800
0
10
20
30
Time k
40
50
60
6
GCI−LSM
GCI−GMB
GCI−PHD
50
(d)
40
Fig. 4. PBP procedure: (a) multi-object state estimation of sensor 1 (single
MC run), (b) multi-object state estimation of sensor 2 (single MC run), (c)
multi-object state estimation of GCI-LMB fusion (single MC run),(d) OSPA
errors of GCI-LSM and GCI-LMB fusion algorithms with order p = 1 and
cut-off c = 100 (200 MC runs).
30
20
10
0
Figs. 4 (a)-(c) show the multi-object estimations of local
filters and GCI-LMB fusion respectively for single MC run.
It can be seen that in this run, the GCI-LMB fusion fails
to perform fusing for the 2th track while both local filters
accurately estimate this track. Due to that the prior information
for births only provide the initial positions, but fail to provide
the initial time, 2th object is initialized at different time
step in different local filters. Hence, the labels of the 2th
object of different local filters are mismatching obviously,
leading to that the GCI-LMB fusion algorithm completely
lose the 2th object. The performance comparison between
GCI-LSM fusion and GCI-LMB is also shown in Fig. 4
(d). As expected, the performance of the GCI-LSM fusion
has remarkable advantages towards GCI-LMB fusion, and the
GCI-LMB fusion is getting worse with target births and deaths.
This result is consistent with the single Monte Carlo (MC)
run’s. The above results confirm that the GCI-LSM fusion is
able to handle the LS-DM, while the GCI-LMB fusion cannot.
5
Estimated Cardinality
(c)
OSPA [m]
Y coordinate (m)
0
−400
Y coordinate (m)
0
X coordinate (m)
Birth with label (20,2)
200
600
−400 −200
400
Birth with label (21,2)
−200
−800
−800 −600
True Tracks
Estimated Tracks of Sensor 2
T5
T4
200
True Cardinality
GCI−GMB
GCI−LSM
GCI−PHD
4
3
2
1
10
20
30
40
Time k
(a)
50
60
0
10
20
30
40
Time k
50
60
(b)
Fig. 6. (a) OSPA distance of order p = 1 and cut-off c = 100 for the
GM implementation with adaptive birth, (b) cardinality estimation, (200 MC
runs).
Both the OSPA distance and the cardinality estimation in
Fig. 6 illustrate the performance differences among the three
fusion methods. It can be seen that the performance of GCILSM is almost the same as GCI-GMB after the performances
converge. Also the GCI-LSM performs slightly worse than
GCI-GMB when objects are born, and the explanation here
is that GCI-GMB fusion considered all possible matching
correspondences between label spaces of different sensors
jointly, while GCI-LSM fusion utilizes an optimal estimation
of the matching correspondence. Moreover the tiny performance loss of GCI-LMS fusion toward GCI-GMB fusion
also demonstrates the superiority of the optimal estimation
9
of matching correspondence. In other words, the ranked assignment problem built in Section IV-B can match the label
spaces from different sensors accurately and consistently. In
addition, Fig. 6 also reveals that both the GCI-LSM and GCIGMB fusion outperform the GCI-PHD fusion for both OSPA
error and the cardinality. This result also demonstrates the
effectiveness GCI-LSM fusion.
VII. C ONCLUSION
This paper investigates the problem of distributed multitarget tracking (DMMT) with labeled multi-object density
based on generalized covariance intersection. Firstly, we provided a principled mathematical definition of label spaces
matching (LS-M) based on information divergence, referred
to as LS-M criterion. Then we proposed a novel two-step
distributed fusion algorithm. Firstly, to match the label spaces
from different sensors, we build a ranked assignment problem
to seek the optimal solution of matching correspondence
between objects of different sensors based on LS-M criterion. Then, GCI fusion is performed on the matched label
space. Moreover, we derive the GCI fusion with generic
labeled multi-object (LMO) densities. A Gaussian mixture
implementation of the proposed GCI-LSM is also given, and
its effectiveness and better performance are demonstrated in
numerical results. At the present stage, the impact of objects
closely spaced on the fusion is not very clearly, thus further
work will study the GCI-LSM fusion considering objects in
proximity.
R EFERENCES
[1] C. Y. Chong, S. Mori, and K. C. Chang, “Distributed multitarget multisensor tracking,” Multitarget-Multisensor Tracking: Advanced Applications;
Y. Bar-Shalom (ed.); Artech House Chapter 8, 1990.
[2] S. J. Julier, T. Bailey, and J. K. Uhlmann, “Using exponential mixture
models for suboptimal distributed data fusion,” in Proc. IEEE Nonlinear
Stat. Signal Proc. Workshop, pp. 160-163, Sep. 2006.
[3] R. Mahler, “Optimal/Robust distributed data fusion: a unified approach,”
in Proc. SPIE Defense Sec. Symp., 2000.
[4] B. N. Vo and W. K. Ma, “The Gaussian mixture probability hypothesis
density filter,” IEEE Trans. on Signal Process., vol. 54, no. 11, pp. 40914104, Nov. 2006.
[5] B. Ristic, D. Clark, B. N. Vo, and B. T. Vo, “Adaptive target birth intensity
for PHD and CPHD filters,” IEEE Trans. Aerosp. Electron. Syst., vol. 48,
no. 2, pp. 1656-1668, 2012.
[6] B. T. Vo, B. N. Vo, and A. Cantoni, “Analytic implementations of the
cardinalized probability hypothesis density filter,” IEEE Trans. on Signal
Process., vol. 55, no. 7, pp. 3553-3567, Jul. 2007.
[7] D. Fränken, M. Schmidt, and M.Ulmke, ““Spooky action at a distanc” in
the cardinalized probability hypothesis density filter,” IEEE Trans. Aerosp.
Electron. Syst., vol. 45, no. 4, pp. 1657-1664, 2009.
[8] B. T. Vo, B. N. Vo, and A. Cantoni, “The cardinality balanced multitarget multi-Bernoulli filter and its implementations,” IEEE Trans. on
Signal Process., vol. 57, no. 2, pp. 409-423, Oct. 2009.
[9] B. T. Vo, B. N. Vo, N. T. Pham and D. Suter, “Joint detection and
estimation of multiple Objects from image observation,” IEEE Trans. on
Signal Process., vol. 58, no. 10, pp. 5129-5141, Oct. 2010.
[10] B. T. Vo, B. N. Vo and R. Hoseinnezhad, “Multi-Bernoulli based trackbefore-detect with road constraints,” in Proc. IEEE Int. Fusion Conf., pp.
840-846, Jul. 2012.
[11] K.G. Amirali, R. Hoseinnezhad and B.H. Alireza, “Robust multiBernoulli sensor selection for multi-target tracking in sensor networks,”
IEEE Signal Process. Lett., vol. 20, no. 12, pp.1167-1170, Dec. 2013
[12] K.G. Amirali, R. Hoseinnezhad and B.H. Alireza, “Multi-bernoulli
sensor control via minimization of expected estimation errors,” IEEE
Trans. Aerosp. Electron. Syst., vol. 51, no. 3, pp. 1762-1773, Jul. 2015.
[13] R. Hoseinnezhad, B. N. Vo, B. T. Vo, and D. Suter, “Bayesian integration
of audio and visual information for multi-target tracking using a CBMEMBER filter,” in Proc. Int. Conf. Acoust., Speech, Signal Process.
(ICASSP), Prague, Czech Republic, pp. 2300-2303, May 2011.
[14] R. Hoseinnezhad, B. N. Vo and B. T. Vo, “Visual tracking in background
subtracted image sequences via multi-Bernoulli filtering,” IEEE Trans. on
Signal Process., pp: 392-397,vol. 61, no. 2, Jan. 2013.
[15] D. Clark, S. Julier, R. Mahler, and B. Ristić, “Robust multi-object sensor
fusion with unknown correlations” in Proc. Sens. Signal Process. Defence
(SSPD 10), Sep. 2010.
[16] M. Üney, D. Clark, and S. Julier, “Distributed fusion of PHD filters via
exponential mixture densities,” IEEE J. Sel. Topics Signal Process., vol.
7, no. 3, pp. 521-531, Apr. 2013.
[17] G. Battistelli, L. Chisci, C. Fantacci, A. Farina, and A. Graziano,
“Consensus CPHD filter for distributed multitarget tracking,” IEEE J.
Sel. Topics Signal Process., vol. 7, no. 3, pp. 508-520, Mar. 2013.
[18] M. B. Guldogan, “Consensus Bernoulli filter for distributed detection
and tracking using multi-static doppler shifts,” IEEE Signal Process. Lett.,
vol. 21, no. 6, pp. 672-676, Jun. 2014.
[19] R. P. S. Mahler, Statistical Multisource-Multitarget Information Fusion.
Norwell, MA, USA: Artech House, 2007.
[20] B. L. Wang, W. Yi, R. Hoseinnezhad, S. Q. Li, L. J. Kong and
X. B. Yang, “Distributed fusion with multi-Bernoulli filter based on
generalized Covariance Intersection,” under review for IEEE Trans. on
Signal Process., Jun. 2016.
[21] B. N. Vo, B. T. Vo, “Labeled random finite sets and multi-object
conjugate priors.” IEEE Trans. on Signal Process., vol. 61, no. 10, pp.
3460-3475, Jul. 2013.
[22] B. N. Vo, B. T. Vo, and D. Phung, “Labeled random finite sets and
the Bayes multi-target tracking filter,” IEEE Trans. on Signal Process.,
vol.PP, no.99, pp.1, Oct. 2014.
[23] S. Reuter, B. T. Vo, B. N. Vo, and K. Dietmayer, “The labeled multiBernoulli filter,” IEEE Trans. on Signal Process., vol. 62, no. 12, pp.32463260, Jun. 2014.
[24] M. Beard, B. T. Vo, and B. N. Vo, “Bayesian multi-target tracking with
merged measurements using labelled random finite sets,” IEEE Trans. on
Signal Process., 2015.
[25] F. Papi and D. Y. Kim, “A particle multi-target tracker for superpositional measurements using labeled random finite sets,” arXiv preprint
arXiv:1501.02248, 2014.
[26] C. Fantacci, B. T. Vo, F. Papi and B. N. Vo, “The marginalized δ-GLMB
filter,” http://arxiv.org/abs/1501.01579. Accessed Jan. 2015.
[27] F. Papi, B. N. Vo, B. T. Vo, C. Fantacci, and M. Beard, “Generalized
labeled multi-Bernoulli approximation of multi-object densities,” arXiv
preprint, 2014, arXiv:1412.5294.
[28] C. Fantacci, B. T. Vo and B. N. Vo, “Consensus labeled
random finite set filtering for distributed multi-object tracking,”
http://arxiv.org/abs/1501.00926. Accessed Jan. 2015.
[29] B. L. Wang, W. Yi, S. Q. Li, L. J. Kong and X. B. Yang, “Distributed
multi-target tracking via generalized multi-Bernoulli random finite sets,”
in Proc. 18th Int. Conf. Inf. Fusion, pp. 253-361, Jul. 2015.
[30] R. P. Mahler, Advances in Statistical Multisource-Multitarget Information Fusion. Artech House, 2014.
[31] S. Q. Li, W. Yi, B. L. Wang, and L. J. Kong, “Joint multi-Bernoulli
random finite set for two-target scenario,”, under review in IEEE Trans.
on Signal Process..
[32] S. Q. Li, W. Yi and L. J. Kong, “Enhanced approximation of labeled
multi-object density based on correlation analysiss,” submmitted in Proc.
19th Int. Conf. Inf. Fusion, 2016.
[33] K. G. Murty,“An Algorithm for Ranking all the Assignments in Order of
Increasing Cost” Operations Research, vol. 16, no. 3, pp. 682-687,1968.
[34] D. Schumacher, B. T. Vo, B. N. Vo, “A consistent metric for performance
evaluation of multi-object filters,” IEEE Trans. on Signal Process., vol.56,
no. 8, pp. 3447-3457, Aug. 2008.
| 3cs.SY
|
arXiv:1504.08014v3 [math.AC] 29 Dec 2015
TESTING FOR THE GORENSTEIN PROPERTY
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
Abstract. We answer a question of Celikbas, Dao, and Takahashi by establishing the following characterization of Gorenstein rings: a commutative
noetherian local ring (R, m) is Gorenstein if and only if it admits an integrally
closed m-primary ideal of finite Gorenstein dimension. This is accomplished
through a detailed study of certain test complexes. Along the way we construct such a test complex that detect finiteness of Gorenstein dimension, but
not that of projective dimension.
1. Introduction
Throughout this paper R denotes a commutative noetherian local ring with
unique maximal ideal m and residue field k.
A celebrated theorem of Auslander, Buchsbaum, and Serre [3, 33] tells us that
R is regular if and only if k has finite projective dimension. Burch [8, p. 947,
Corollary 3] extended this by proving that R is regular if and only if pd(R/I) < ∞
for some integrally closed m-primary ideal I of R.
Auslander and Bridger [1, 2] introduced the G-dimension as a generalization of
projective dimension. (See 2.3 for the definition.) Analogous to the regular setting,
the finiteness of G-dimR (k) characterizes the Gorensteinness of R. In our local
setting, Goto and Hayasaka [21] studied Gorenstein dimension of integrally closed
m-primary ideals and, analogous to Burch’s result, established the following; see
the question of K. Yoshida stated in the discussion following [21, (1.1)].
1.1 ([21, (1.1)]). Let I be an integrally closed m-primary ideal of R. Assume I
contains a non-zerodivisor of R, or R satisfies Serre’s condition (S1 ). Then R is
Gorenstein if and only if G-dimR (R/I) < ∞.
Our aim in this paper is to remove the hypothesis “I contains a non-zerodivisor
of R, or R satisfies Serre’s condition (S1 )” from 1.1. We accomplish this in the following result and hence obtain a complete generalization of Burch’s aforementioned
result; see also Corollary 4.7 for a further generalization.
Theorem 4.8. Let I ⊆ R be an integrally closed ideal with depth(R/I) = 0, e.g.,
such that I is m-primary. Then R is Gorenstein if and only if G-dimR (R/I) < ∞.
Date: January 5, 2018.
2010 Mathematics Subject Classification. 13B22, 13D02,13D07, 13D09, 13H10.
Key words and phrases. Integrally closed ideals, G-dimension, projective dimension, semidualizing complexes, test complexes.
Sean Sather-Wagstaff was supported in part by a grant from the NSA.
1
2
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
Our argument is quite different from that of Goto and Hayasaka [21] since it
uses G-dim-test complexes. For this part of the introduction, we focus on the case
of H-dim-test modules, defined next. Note that the pd-test modules (i.e., the case
where H-dim = pd) are from [9].
Definition 1.2. Let H-dim denote either projective dimension pd or G-dimension
G-dim. Let M be a finitely generated R-module. Then M is an H-dim-test module
over R if the following condition holds for all finitely generated R-modules N : If
TorR
i (M, N ) = 0 for all i ≫ 0, then H-dimR (N ) < ∞.
It is straightforward to show that if M is a pd-test R-module, then it is also
G-dim-test. Example 3.14 shows that the converse of this statement fails in general.
As part of our proof of Theorem 4.8, we also answer the following questions; see
Corollaries 3.11(c) and 3.12.
Question 1.3. Let M be a pd-test module over R.
c be a pd-test module over R?
b
(a) Must M
(b) ([9, (3.5)]) If G-dimR (M ) < ∞, must R be Gorenstein?
Affirmative answers to Question 1.3(b) under additional hypotheses are in [9,
(1.3)] and [10, (2.15)]. Also, Majadas [28] gives an affirmative answer to a version
of Question 1.3(a) that uses a more restrictive version of test modules.
Theorem 4.8 follows from the next, significantly stronger result.
Theorem 4.4. Let M be a G-dim-test R-module such that ExtiR (M, R) = 0 for
i ≫ 0. Then R is Gorenstein.
In turn, this follows from the much more general Theorem 4.1 and Corollary 4.2,
which are results for detecting dualizing complexes.
We conclude this introduction by summarizing the contents of this paper. Section 2 consists of background material for use throughout the paper, and contains
some technical lemmas for later use. In Section 3, we develop foundational properties of various H-dim-test objects, and answer Question 1.3. And Section 4 contains
the theorems highlighted above.
2. Derived Categories and Semidualizing Complexes
2.1. Throughout this paper we work in the derived category D(R) whose objects
are the chain complexes of R-modules, i.e., the R-complexes X with homological
differential ∂iX : Xi → Xi−1 . References for this include [13, 23, 36, 37]. Our
notation is consistent with [12]. In particular, RHomR (X, Y ) and X ⊗L
R Y are the
derived Hom-complex and derived tensor product of two R-complexes X and Y .
Isomorphisms in D(R) are identified by the symbol ≃. The projective dimension
and flat dimension of an R-complex X ∈ Db (R) are denoted pdR (X) and fdR (X).
The subcategory of D(R) consisting of homologically bounded R-complexes (i.e.,
complexes X such that Hi (X) = 0 for |i| ≫ 0) is Db (R). The subcategory of
D(R) consisting of homologically finite R-complexes (i.e., complexes X such that
H(X) := ⊕i∈Z Hi (X) is finitely generated) is denoted Dbf (R).
2.2. A homologically finite R-complex C is semidualzing if the natural morphism
R → RHomR (C, C) in D(R) is an isomorphism. For example, an R-module is
semidualizing if and only if HomR (C, C) ∼
= R and ExtiR (C, C) = 0 for i > 1. In
TESTING FOR THE GORENSTEIN PROPERTY
3
particular, R is a semidualizing R-module. A dualizing R-complex is a semidualizing R-complex of finite injective dimension.
2.2.1. If R is a homomorphic image of a local Gorenstein ring Q, then R has a
dualizing complex, by [23, V.10]. (The converse holds by work of Kawasaki [27].)
b has a
In particular, the Cohen Structure Theorem shows that the completion R
dualizing complex. When R has a dualizing complex D, and C is a semidualizing
R-complex, the dual RHomR (C, D) is also semidualizing over R, by [12, (2.12)].
2.2.2. Let ϕ : R → S be a flat local ring homomorphism, and let C be a semidualizing R-complex. Then the S-complex S ⊗L
R C is semidualizing, by [12, (5.6)].
If the closed fibre S/mS is Gorenstein and R has a dualizing complex DR , then
R
DS := S ⊗L
R D is dualizing for S by [5, (5.1)].
Dualizing complexes were introduced by Grothendieck and Harshorne [23]. The
more general semidualizing complexes originated in special cases, e.g., in [6, 16,
19, 35], with the general version premiering in [12]. The notion of G-dimension,
summarized next, started with the work of Auslander and Bridger [1, 2] for modules.
Foxby and Yassemi [38] recognized the connection with derived reflexivity, with the
general situation given in [12].
2.3. Let C be a semidualizing R-complex and X ∈ Dbf (R). Write GC -dimR (X) < ∞
when X is “derived C-reflexive”, i.e., when RHomR (X, C) ∈ Db (R) and the natural
morphism X → RHomR (RHomR (X, C), C) in D(R) is an isomorphism. In the case
C = R, we write G-dimR (X) < ∞ instead of GR -dimR (X) < ∞.
2.3.1. The complex C is dualizing if and only if every R-complex in Dbf (R) is
derived C-reflexive, by [12, (8.4)]. In particular, R is Gorenstein if and only if
every R-complex X ∈ Dbf (R) has G-dimR (X) < ∞.
2.3.2. Let R → S be a flat local ring homormorphism. Given an R-complex
X ∈ Dbf (R), one has GC -dimR (X) < ∞ if and only if GS⊗LR C -dimS (S ⊗L
R X) < ∞,
by [12, (5.10)]; see 2.2.2.
Auslander and Bass classes, defined next, arrived in special cases in [6, 16], again
with the general case described in [12].
2.4. Let C be a semidualizing R-complex. The Auslander class AC (R) consists of
the R-complexes X ∈ Db (R) such that C ⊗L
R X ∈ Db (R) and the natural morphism
C
γX
: X → RHomR (C, C ⊗L
X)
in
D(R)
is
an
isomorphism. The Bass class BC (R)
R
consists of all the R-complexes X ∈ Db (R) such that RHomR (C, X) ∈ Db (R) and
C
such that the natural morphism ξX
: C ⊗L
R RHomR (C, X) → X in D(R) is an
isomorphism.
2.4.1. When R has a dualizing complex D, given an R-complex X ∈ Dbf (R),
one has GC -dimR (X) < ∞ if and only if X ∈ ARHomR (C,D) (R), by [12, (4.7)];
this uses 2.2.1 and 2.3.1, which imply that RHomR (C, D) is semidualizing and
C ≃ RHomR (RHomR (C, D), D).
2.4.2. Let R → S be a flat local ring homormorphism. Given an S-complex X,
one has X ∈ AC (R) if and only if X ∈ AS⊗LR C (S), by [12, (5.3.a)].
The following two lemmas are proved like [17, (4.4)] and [25, (7.3)], respectively.
4
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
Lemma 2.5. Let X, P ∈ Dbf (R) such that P 6≃ 0 and pdR (P ) < ∞. Let C be a
semidualizing R-complex.
(a) The following conditions are equivalent:
(i) X ∈ AC (R),
(ii) P ⊗L
R X ∈ AC (R), and
(iii) RHomR (P, X) ∈ AC (R).
(b) The following conditions are equivalent:
(i) X ∈ BC (R),
(ii) P ⊗L
R X ∈ BC (R), and
(iii) RHomR (P, X) ∈ BC (R).
Lemma 2.6. Let R → S be a flat local ring homomorphism such that S/mS is
Gorenstein. Let X ∈ Dbf (S) such that each homology module Hi (X) is finitely
generated over R.
(a) One has GC -dimR (X) < ∞ if and only if GS⊗LR C -dimS (X) < ∞.
(b) One has G-dimR (X) < ∞ if and only if G-dimS (X) < ∞.
The next result, essentially from [8, Theorem 5(ii)], is key for Theorem 4.8.
Lemma 2.7. Let I be an integrally closed ideal such that depth(R/I) = 0, and let
R
M be a finitely generated R-module. If TorR
i (R/I, M ) = 0 = Tori+1 (R/I, M ) for
some i > 1, then pdR (M ) 6 i. In particular, R/I is a pd-test module over R.
R
Proof. Assume that TorR
i (R/I, M ) = 0 = Tori+1 (R/I, M ), and suppose that
pdR (M ) > i. The Auslander-Buchsbaum-Serre Theorem states that k is a pdtest R-module, so we assume without loss of generality that I ( m. Hence, we have
I : m 6= R so I : m ⊆ m. From [8, Theorem 5(ii)], we conclude that mI = m(I : m).
Claim: I : m = I. One containment (⊇) is standard. For the reverse containment, let r ∈ I : m ⊆ m. To show that r is in I, it suffices to show that r is integral
over I, since I is integrally closed. To this end, we use the “determinantal trick”
from [24, (1.1.8)]: it suffices to show that (1) we have rm ⊆ Im, and (2) whenever
am = 0 for some a ∈ R, we have ar = 0. For (1), since r is in I : m, we have
rm ⊆ (I : m)m = Im, as desired. For (2), if am = 0, then the fact that r is in m
implies that ar = 0, as desired. This completes the proof of the claim.
Now, the fact that depth(R/I) = 0 implies that there is an element x ∈ R r I
such that xm ⊆ I. In other words, x ∈ (I : m)rI, contradicting the above claim.
2.8. Let I be an integrally closed ideal of R. If one assumes that I is m-primary
(stronger than the assumption depth(R/I) = 0 from Lemma 2.7) then one gets
the following very strong conclusion. Given a finitely generated R-module M , if
TorR
i (R/I, M ) = 0 for some i > 1, then pdR (M ) < i, by [14, (3.3)].
3. H-dim-Test Complexes
In this section, let C be a semidualizing R-complex.
We now introduce the main object of study for this paper.
Definition 3.1. Let M ∈ Dbf (R), and let H-dim denote either projective dimension
pd or GC -dimension GC -dim. Then M is an H-dim-test complex over R if the
following condition holds for all N ∈ Dbf (R): If TorR
i (M, N ) = 0 for all i ≫ 0, i.e.,
if M ⊗L
N
∈
D
(R),
then
H-dim
(N
)
<
∞.
b
R
R
TESTING FOR THE GORENSTEIN PROPERTY
5
3.2. Let M be an R-module. A standard truncation argument shows that M is a
H-dim-test module if and only if it is a H-dim-test complex; see [9, Proof of (3.2)].
3.3. Examples of pd-test modules are given in 2.8 and Lemma 2.7. Note that this
includes the standard example k = R/m. See also [10, Appendix A].
3.3.1. Given an R-complex X ∈ Dbf (R), if pdR (X) is finite, then so is GC -dimR (X),
by [12, (2.9)]. Thus, if M is a pd-test complex, then it is also a GC -dim-test, in
particular, M is a G-dim-test complex.
3.3.2. If R is G-regular (i.e., if every R-module of finite G-dimension has finite
projective dimension), then the pd-test complexes and G-dim-test complexes over
R are the same. Examples of G-regular rings include regular rings and CohenMacaulay rings of minimal multiplicity; see [34].
3.3.3. Examples of GC -dim test modules, e.g., of G-dim test modules, that are not
pd test modules are more mysterious. See Example 3.14 for a non-trivial example.
3.3.4. Assume that R has a dualizing complex D. A natural candidate for a
GC -dim-test complex is C † = RHomR (C, D). Indeed, if GC -dimR (X) < ∞, then
X ∈ AC † (R) by 2.4.1, so by definition we have C † ⊗L
R X ∈ Db (R). In particular, a
natural candidate for a G-dim-test complex is D; see, e.g., Corollary 4.5.
However, D can fail to be a G-dim-test complex. Indeed, Jorgensen and Şega [26,
(1.7)] construct an artinian local ring R with a finitely generated module L that
satisfies ExtiR (L, R) = 0 for all i > 1 and G-dimR (L) = ∞. Since R is local and
artinian, it has a dualizing complex, namely D = ER (k) the injective hull of the
residue field k. We claim that TorR
i (D, L) = 0 for i > 1. (This shows that D is not
G-dim-test over R.) To this end, recall the following for i > 1:
HomR (TorR (D, L), D) ∼
= Exti (L, R) = 0.
i
R
The fact that D is faithfully injective implies that TorR
i (D, L) = 0 for i > 1.
3.4. The ring R is regular if and only if every X ∈ Dbf (R) has pdR (X) < ∞. Hence,
the trivial complex 0 is a pd-test complex if and only if R is regular, equivalently,
if and only if R has a pd-test complex of finite projective dimension. Similarly, if
C is a semidualizing R-complex, then 0 is a GC -dim-test complex if and only if
C is dualizing, equivalently, if and only if R has a GC -dim-test complex of finite
projective dimension; see 2.3.1. In particular, 0 is a G-dim-test complex if and only
if R is Gorenstein, equivalently, if and only if R has a G-dim-test complex of finite
projective dimension.
We continue with a discussion of ascent and descent of test complexes.
ϕ
Theorem 3.5. Let R −
→ S be a flat local ring homomorphism, and let M ∈ Dbf (R).
L
(a) If S ⊗R M is GS⊗LR C -dim-test over S, then M is GC -dim-test over R.
(b) If S ⊗L
R M is G-dim-test over S, then M is G-dim-test over R.
(c) If S ⊗L
R M is pd-test over S, then M is pd-test over R.
-dim-test complex over S. To show that
Proof. (a) Assume that S ⊗L
R M is a GS⊗L
RC
M is a GC -dim-test complex over R, let X ∈ Dbf (R) such that M ⊗L
R X ∈ Db (R).
L
L
L
X)
are
all in Dbf (S).
(M
⊗
M
,
and
S
⊗
X,
S
⊗
By flatness, the complexes S ⊗L
R
R
R
R
Moreover, we have the following isomorphisms in D(S):
L
L
L
L
L
L
(S ⊗L
R M ) ⊗S (S ⊗R X) ≃ (S ⊗R M ) ⊗R X ≃ S ⊗R (M ⊗R X).
6
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
As S ⊗L
-dim-test complex over S, we have GS⊗LR C -dimS (S ⊗L
R M is a GS⊗L
R X) <
RC
∞. Using 2.3.2, we conclude that GC -dimR (X) < ∞, as desired.
(b) This is the special case C = R of part (a).
(c) Argue as in part (a), using [6, (1.5.3)] in place of 2.3.2.
Note that the conditions on ϕ in the following three items hold for the natural
b or to its henselization Rh .
maps from R to its completion R
Remark 3.6. Let ϕ : R → S be a flat local ring homomorphism, and assume that
the closed fibre S/mS is module-finite over k. Let M ∈ Dbf (R) and X ∈ Dbf (S)
L
such that (S ⊗L
R M ) ⊗S X ∈ Db (S). Let x = x1 , . . . , xn ∈ m be a generating
sequence for m, and set K := K S (x), the Koszul complex on x over S. It follows
L
L
that K ⊗L
S ((S ⊗R M ) ⊗S X) ∈ Db (S). From the following isomorphisms
L
L
L
L
L
L
L
K ⊗L
S ((S ⊗R M ) ⊗S X) ≃ (K ⊗S X) ⊗S (S ⊗R M ) ≃ (K ⊗S X) ⊗R M
L
we conclude that (K ⊗L
S X) ⊗R M ∈ Db (R).
L
f
Note that K ⊗S X ∈ Db (R). Indeed, we already have K ⊗L
S X ∈ Db (R), so
it suffices to show that every homology module Hi (K ⊗L
X)
is
finitely generated
S
over R. We know that Hi (K ⊗L
X)
is
finitely
generated
over
S.
Moreover, it is
S
annihilated by (x)S = mS. Thus, it is a finitely generated S/mS-module; since
S/mS is module finite over k, each Hi (K ⊗L
S X) is finitely generated over k, so it is
finitely generated over R.
Theorem 3.7. Let ϕ : R → S be a flat local ring homomorphism, and let M ∈
Dbf (R). Assume that the closed fibre S/mS is Gorenstein and module-finite over k.
(a) The R-complex M is GC -dim-test if and only if S ⊗L
-dim-test
R M is GS⊗L
RC
over S.
(b) The R-complex M is G-dim-test if and only if S ⊗L
R M is G-dim-test over S.
Proof. (a) One implication is covered by Theorem 3.5(a). For the reverse implication, assume that M is a GC -dim-test complex over R. To show that S ⊗L
R M is a
L
GS⊗LR C -dim-test complex over S, let X ∈ Dbf (S) such that (S ⊗L
R M )⊗S X ∈ Db (S).
Let x = x1 , . . . , xn ∈ m be a generating sequence for m, and set K := K S (x). By
L
L
f
Remark 3.6 we have (K ⊗L
S X) ⊗R M ∈ Db (R) and K ⊗S X ∈ Db (R). Since M
L
is an GC -dim-test complex over R, one has GC -dimR (K ⊗S X) < ∞. It follows
from Lemma 2.6 that GS⊗LR C -dimS (K ⊗L
S X) < ∞. We deduce from [17, (4.4)] that
GS⊗LR C -dimS (X) < ∞, as desired.
(b) This is the special case C = R of part (a).
Here is one of our main results; see also Corollary 3.12 and Theorem 4.4.
Theorem 3.8. Let ϕ : (R, m) → (S, n) be a flat local ring homomorphism, and let
M ∈ Dbf (R). Assume the induced map R/m → S/mS is a finite field extension,
i.e., we have mS = n and the induced map R/m → S/mS is finite. Then M is a
pd-test complex over R if and only if S ⊗L
R M is a pd-test complex over S.
Proof. One implication is covered by Theorem 3.5(c). For the reverse implication,
assume that M is a pd-test complex over R.
Case 1: S is complete. To show that S ⊗L
R M is a pd-test complex over S, let X ∈
f
L
Db (S) such that (S ⊗L
M
)
⊗
X
∈
D
(S).
Let x = x1 , . . . , xn ∈ m be a generating
b
R
S
L
L
sequence for m, and set K := K S (x). It follows that K⊗L
S ((S⊗R M )⊗S X) ∈ Db (S).
TESTING FOR THE GORENSTEIN PROPERTY
7
L
L
f
By Remark 3.6 we have (K ⊗L
S X) ⊗R M ∈ Db (R) and K ⊗S X ∈ Db (R). As M is
L
an pd-test complex over R, we have pdR (K ⊗S X) < ∞. It follows from [7, (2.5)]
that pdS (K ⊗L
S X) < ∞, and [6, (1.5.3)] implies that pdS (X) < ∞.
Case 2: the general case. Case 1 implies that Sb ⊗L
R M is a pd-test complex over
L
L
b
b
b
S, i.e., S ⊗S (S ⊗R M ) is a pd-test complex over S. So, S ⊗L
R M is a pd-test complex
over S, by Theorem 3.5(c), as desired.
The next example, from discussions with Ryo Takahashi, shows that the hypothesis mS = n is necessary for the conclusion of Theorem 3.8.
Example 3.9. Let k be a field, R = k and S = k[[y]]/(y 2 ). Then the natural map
R → S is a finite free map since S is free over R with R-basis {1, y}. Let M = R.
Then, since R is regular, M is a pd-test module over R. However N = M ⊗R S = S
is not a pd-test module over S since S is not regular; see 3.4.
On the other hand, we do not know whether or not having a regular closed fibre
in Theorem 3.8 is sufficient, as we note next.
Question 3.10. Let ϕ : (R, m) → (S, n) be a flat local ring homomorphism, and
let M ∈ Dbf (R). Assume that S/mS is regular. If M is a pd-test complex over R,
then must S ⊗L
R M be a pd-test complex over S?
The next corollary answers Question 1.3(a).
b := R
b ⊗L C.
Corollary 3.11. Let M be an R-module, and set C
R
c
b
(a) The module M is GC -dim-test over R if and only if M is GCb -dim-test over R.
c is G-dim-test over R.
b
(b) The module M is G-dim-test over R if and only if M
c
b
(c) The module M is pd-test over R if and only if M is pd-test over R.
b ⊗L M ≃ M
c in Df (R),
b the desired conclusions follow from TheoProof. Since R
R
b
rems 3.7 and 3.8.
The next corollary answers Question 1.3(b). We are able to improve this result
significantly in the next section; see Theorem 4.4 and the subsequent paragraph.
Corollary 3.12. Let M be a pd-test module over R. If G-dimR (M ) < ∞, then R
is Gorenstein.
c is a pd-test module for R,
b with G-dim b (M
c) <
Proof. Corollary 3.11(c) says that M
R
b
∞ by 2.3.2. Using [9, (1.3)], we conclude that R is Gorenstein, hence so is R.
We end this section by building a module that is G-dim-test but not pd-test; see
Example 3.14.
Proposition 3.13. Let ϕ : (A, n, F ) → R be a flat local ring homomorphism, and
set N := R/nR ≃ R ⊗L
A F.
(a) The R-module N is Tor-rigid, i.e., for any finitely generated R-module M , if
R
TorR
i (M, N ) = 0 for some i > 1, then Torj (M, N ) = 0 for all j > i.
(b) Let B be a semidualizing A-complex, and set C := R ⊗L
A B. If R/nR is Gorenstein, then N is a GC -dim-test complex over R.
(c) If R/nR is Gorenstein, then N is a G-dim-test complex over R.
8
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
Proof. First, note that we have N := R/nR ∼
= R ⊗A (A/n) = R ⊗A F ≃ R ⊗L
A F
since R is flat over A. Furthermore, for every R-complex M , there are isomorphisms
L
L
L
M ⊗L
R N ≃ M ⊗R (R ⊗A F ) ≃ M ⊗A F.
(3.13.1)
A
∼
In particular, one has TorR
i (M, N ) = Tori (M, F ) for all i.
(a) Let M be a finitely generated R-module. From [29, (22.3)], if TorA
1 (M, F ) =
0, then M is flat over A, so we have TorA
(M,
F
)
=
0
for
all
j
>
1.
More
generally,
by
j
A
A
dimension-shifting, if Tori (M, F ) = 0 for some i > 1, then we have Torj (M, F ) = 0
for all j > i. Thus, the isomorphism from the previous paragraph implies the desired
Tor-rigidity.
b is a G b L -dim-test
(b) Corollary 3.11(a) shows that it suffices to show that N
R⊗R C
b Note that the induced map ϕ
b → R
b is flat and local with
complex over R.
b: A
\ Also, there are isomorphisms in D(R):
b
Gorenstein closed fibre R/nR.
L
b ⊗Lb (A
b ⊗L
b L
b L
b L
R
A B) ≃ R ⊗A B ≃ R ⊗R (R ⊗A B) ≃ R ⊗R C.
A
Thus, we may replace ϕ with the induced map ϕ
b to assume for the rest of the proof
that A and R are complete. Let DA be a dualizing A-complex; see 2.2.1. Then the
A
†
A
R-complex DR := R ⊗L
A D is dualizing for R, by 2.2.2. Set B := RHomA (B, D )
†
R
†
L
†
and C := RHomR (C, D ), noting that C ≃ R ⊗A B by [12, proof of (5.10)(*)].
Let M ∈ Dbf (R) such that M ⊗L
R N ∈ Db (R). We need to show GC -dimR (M ) <
∞ i.e., that M ∈ AC † (R); see 2.4.1. By (3.13.1), the complex M ⊗L
A F is in Db (R).
Since M is in Dbf (R), we conclude that fdA (M ) < ∞ by [4, (5.5.F)]. It follows
from [12, (4.4)] that M ∈ AB † (A), so M ∈ AR⊗LA B † (R) = AC † (R) by [12, (5.3.a)].
(c) This is the special case B = A, hence C = R, of part (b).
Example 3.14. Let k be a field. Consider the finite-dimensional local k-algebras
A := k[y, z]/(y, z)2 and
R := k[x, y, z]/(x2 , y 2 , z 2 , yz) ∼
= A[x]/(x2 ).
Notice that R is free over A, hence flat. Also, the natural map A → R is local
with Gorenstein closed fibre R/nR ∼
= k[x]/(x2 ); here, as in Proposition 3.13, we let
n denote the maximal ideal of A. Since the assumptions of Proposition 3.13 are
satisfied, the R-module
N = R ⊗A k = R ⊗A (A/(y, z)A) ∼
= R/(y, z)R
is Tor-rigid and G-dim-test over R. Furthermore, from [9, (4.1)], we know that N
is not pd-test.
Since A is artinian and local, the injective hull DA = ER (k) is a dualizing AA
module. Thus, the R-module DR := R ⊗A DA ≃ R ⊗L
is dualizing for R,
A D
R
by 2.2.2. We conclude by showing that D is also a G-dim-test module that is not
pd-test and, moreover, is not Tor-rigid.
Note that A has length 3 and type 2. From this, we construct an exact sequence
over A of the following form:
0 → k 3 → A2 → DA → 0.
(3.14.1)
TESTING FOR THE GORENSTEIN PROPERTY
9
Indeed, the condition type(A) = 2 says that DA is minimally generated by 2 elep
ments. Let A2 −
→ DA → 0 be a minimal presentation, and consider the corresponding short exact sequence
p
0 → Ker(p) → A2 −
→ DA → 0.
From the minimality of the presentation, it follows that Ker(p) ⊆ nA2 . Since
n2 = 0, we conclude that Ker(p) is a k-vector space, so we need only verify that
lenA (Ker(p)) = 3. This equality follows from the additivity of length, via the
condition lenA (DA ) = len(A) = 3.
Since R is flat over A, we apply the base-change functor R ⊗A − to the sequence (3.14.1) to obtain the next exact sequence over R:
0 → N 3 → R2 → DR → 0.
For any R-module M , the associate long exact in TorR
i (M, −) shows that we have
R
R
3
R ∼
Tori (M, D ) = Tori−1 (M, N ) for all i > 2. In particular, we have
R
R
TorR
i (M, D ) = 0 if and only if Tori−1 (M, N ) = 0, for i > 2.
(3.14.2)
Claim: DR is G-dim-test over R. To show this, let M be a finitely generated
R
R-module such that TorR
i (M, D ) = 0 for i ≫ 0. The display (3.14.2) implies that
R
Tori (M, N ) = 0 for i ≫ 0. Since N is G-dim-test over R, we have G-dimR (M ) <
∞, as desired.
Claim: DR is not pd-test over R. To show this, suppose by way of contradiction
that DR is pd-test over R. We show that N is pd-test, contradicting [9, (4.1)]. Let
M be a finitely generated R-module such that TorR
i (M, N ) = 0 for i ≫ 0. The
R
display (3.14.2) implies that TorR
(M,
D
)
=
0
for
i
≫ 0. Since DR is pd-test, we
i
have pdR (M ) < ∞. Thus, N is pd-test, giving the advertised contradiction, and
establishing the claim.
R
Claim: If M is a finitely generated R-module such that TorR
i (M, D ) = 0 for
R
R
R
some i > 2, then Torj (M, D ) = 0 for all j > i. (This shows that D is almost
Tor-rigid.) Since N is Tor-rigid by Proposition 3.13(a), this follows from (3.14.2).
Claim: DR is not Tor-rigid over R. To this end, we follow a construction
R
of [15, Chapter 3] and build an R-module L such that TorR
1 (L, D ) = 0 and
R
R
Tori (L, D ) 6= 0 for all i > 2. Let f1 , f2 be a minimal generating sequence for
HomA (k, A) ∼
= k 2 . For instance, f1 (1) = y and f2 (1) = z will work here. Define
2
u : k → A by the formula a 7→ (f1 (a), f2 (a)). Since k is simple, the non-zero map
u is a monomorphism. Let M := Coker(u). The long exact sequence in ExtA (−, A)
associated to the sequence
u
0→k−
→ A2 → M → 0
(3.14.3)
shows that Ext1A (M, A) = 0.
Set L := R ⊗A M . To make things concrete, if one uses the specific functions
f1 , f2 suggested in the previous paragraph, then M has the following minimal free
presentation over A
( yz )
A −−−→ A2 → M → 0
so L has the following minimal free presentation over R
( yz )
R −−−→ R2 → L → 0.
10
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
Now, flat base-change implies that
Ext1 (L, R) ∼
= R ⊗A Ext1 (M, A) = 0.
A
R
Thus, we have
1
R
R ∼
HomR (TorR
1 (D , L), D ) = ExtR (L, R) = 0.
R
The fact that DR is faithfully injective over R implies that TorR
1 (D , L) = 0.
A
Since A is not regular, we have pdA (k) = ∞ and Tori (k, k) 6= 0 for all i > 0.
Therefore, by (3.14.3), we conclude that pdA (M ) = ∞. In particular, this implies
that TorA
1 (k, M ) 6= 0. So it follows from (3.14.1) and (3.14.3) that
A
A
Tor (D , M ) ∼
= TorA (k 3 , M ) ∼
= TorA (k 3 , k) ∼
= (TorA (k, k))3 6= 0 for all i > 2
i
i−1
i−2
i−2
and
A 3
A
A
3
∼
∼
TorA
2 (D , M ) = Tor1 (k , M ) = (Tor1 (k, M )) 6= 0.
Thus flat base-change implies that
TorR (DR , L) ∼
= R ⊗A TorA (DA , M ) 6= 0 for all i > 2.
i
i
This completes the claim and the example.
4. Detecting the Dualizing and Gorenstein Properties
Our next result yields both Theorems 4.4 and 4.8 highlighted in the introduction. Note that condition (2) in this result does not assume a priori that R has
a dualizing complex; however, the result shows that this condition implies that R
has a dualizing complex.
Theorem 4.1. Let B, C be semidualizing R-complexes. Let M 6≃ 0 be a GB -dimtest R-complex such that RHomR (M, C) ∈ Db (R). Assume that one of the following
conditions holds:
(1) The ring R has a dualizing complex D, and GB † -dimR (C) < ∞ where B † :=
RHomR (B, D), or
(2) One has C ∈ AB (R).
Then B ⊗L
R C is dualizing for R.
Proof. (1) Assume that R has a dualizing complex D, and GB † -dimR (C) < ∞. Set
C † := RHomR (C, D) which is semidualizing by 2.2.1. Let x = x1 , . . . , xn ∈ m be
a generating sequence for m, and consider the Koszul complex K := K R (x). Set
X := K ⊗L
R RHomR (C, E) where E = ER (k) is the injective hull of k.
We note that X is in Dbf (R). Indeed, the complex RHomR (K, C) is homologically
finite since C is. The total homology module H(RHomR (K, C)) is annihilated by
(x)R = m, so it is a finite dimensional vector space over k. By Matlis duality, the
total homology module of RHomR (RHomR (K, C), E) is also a finite dimensional
vector space over k, so we have
f
X = K ⊗L
R RHomR (C, E) ≃ RHomR (RHomR (K, C), E) ∈ Db (R)
(4.1.1)
by Hom-evaluation [4, (4.4)].
The assumption RHomR (M, C) ∈ Db (R) implies that
M ⊗L
R RHomR (C, E) ≃ RHomR (RHomR (M, C), E) ∈ Db (R)
again by Hom-evaluation [4, (4.4)]. From this, we conclude that
L
L
L
L
M ⊗L
R X = M ⊗R (K ⊗R RHomR (C, E)) ≃ K ⊗R (M ⊗R RHomR (C, E)) ∈ Db (R).
TESTING FOR THE GORENSTEIN PROPERTY
11
Since M is a GB -dim-test complex, this implies that GB -dimR (X) < ∞. From 2.4.1
we conclude that X ∈ AB † (R), i.e., we have RHomR (RHomR (K, C), E) ∈ AB † (R)
by (4.1.1). Since E is faithfully injective, argue as in the proof of [11, (3.2.9)] to
conclude that RHomR (K, C) ∈ BB † (R); see also [32, 4.14]. By assumption, we
have C ∈ Dbf (R), so Lemma 2.5(b) shows that C ∈ BB † (R).
By assumption, we have GB † -dimR (C) < ∞, so [18, (1.3)] implies that B † ∈
BC (R). We conclude from [18, (1.4) and (4.10.4)] that C and B † are isomorphic
up to a shift in D(R). Apply a shift to C to assume that C ≃ B † . From [12, (4.4)]
†
L
we have D ≃ B ⊗L
R B ≃ B ⊗R C, as desired.
b has a dualizing complex
(2) Assume now that C ∈ AB (R). The completion R
b
L
R
L
b
b
b
D , by 2.2.1, and the complexes R ⊗R B and R ⊗R C are semidualizing over R,
L
L
b ⊗ M 6≃ 0 by faithful flatness of R,
b and the complex R
b⊗ M
by 2.2.2. We have R
R
R
is GR⊗
-dim-test
by
Theorem
3.7(a).
Also,
by
faithful
flatness,
the
condition
L
b
B
R
RHomR (M, C) ∈ Db (R) implies that
b ⊗L
b L
b L
b
RHomRb (R
R M, R ⊗R C) ≃ R ⊗R RHomR (M, C) ∈ Db (R)
by [17, (1.9.a)].
b
b ⊗L B ≃ (R
b ⊗L B)††
With (−)† := RHomRb (−, DR ), we have an isomorphism R
R
R
by 2.3.1. In addition, from [12, (5.8)], the assumption C ∈ AB (R) implies that
b ⊗L C ∈ A b L (R)
b = A b L †† (R).
b
we have R
We conclude from 2.4.1 that
R
R⊗R B
(R⊗R B)
b L
G(R⊗
b L B)† -dimR
b (R ⊗R C) < ∞.
R
b Thus, the R-complex
b
It follows that condition (1) is satisfied over R.
b ⊗L B) ⊗L (R
b ⊗L C) ≃ (R
b ⊗L B) ⊗L C ≃ R
b ⊗L (B ⊗L C)
(R
R
b
R
R
R
R
R
R
b Note that the condition C ∈ AB (R) implies by definition that
is dualizing for R.
f
B ⊗L
C
∈
D
(R).
Thus, the condition B, C ∈ Dbf (R) implies that B ⊗L
b
R
R C ∈ Db (R).
b ⊗L (B ⊗L C) is dualizing for R
b implies that B ⊗L C is
From this, the fact that R
R
R
R
dualizing for R, by [5, (5.1)], as desired.
We now give several consequences of Theorem 4.1. Compare the next result
to [9, (3.4)].
Corollary 4.2. Let M be a G-dim-test R-complex. Let C be a semidualizing Rcomplex such that that RHomR (M, C) ∈ Db (R). Then C is dualizing for R.
Proof. By 3.4, we assume that M 6≃ 0. The desired conclusion follows from Theorem 4.1 with B = R, once we note that C ∈ Db (R) = AR (R).
Corollary 4.3. Let M be a G-dim-test R-complex such that RHomR (M, R) ∈
Db (R). Then R is Gorenstein.
Proof. Use C = R in Corollary 4.2.
Theorem 4.4. Let M be a G-dim-test R-module such that
i ≫ 0, e.g., with G-dimR (M ) < ∞. Then R is Gorenstein.
Proof. This is immediate from Corollary 4.3.
ExtiR (M, R)
= 0 for
We note that the hypotheses of Theorem 4.4 are weaker than those in Corollary 3.12. Indeed, Example 3.14 above exhibits a G-dim-test module that is not a
12
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
pd-test module. Furthermore, as we noted in 3.3.4, there exist examples of finitely
generated modules L such that ExtiR (L, R) = 0 for all i > 1 and G-dimR (L) = ∞.
Corollary 4.5. Let C be a semidualizing R-complex. If C is G-dim-test over R,
then C is dualizing for R.
Proof. This follows from Corollary 4.2 since RHomR (C, C) ≃ R ∈ Db (R).
Remark 4.6. In light of Corollary 4.5, it is worth noting that there are rings with
semidualizing complexes that are not dualizing and that have infinite projective dimension. (In particular, these complexes are neither pd-test nor G-dim-test by 3.3.1
and Corollary 4.5.) The first examples were constructed (though not published) by
Foxby. See also [12, (7.8)] and [30, 31].
It is also worth noting that the converse of Corollary 4.5 fails in general by 3.3.4.
Corollary 4.7. Let I ⊆ R be an integrally closed ideal such that depth(R/I) = 0,
e.g., such that I is m-primary, and let C be a semidualizing R-complex. Then C is
dualizing for R if and only if GC -dimR (R/I) < ∞.
Proof. Note that I is a pd-test by Lemma 2.7, and apply 2.3.1 and Corollary 4.2.
Recall that the next result has been initially obtained by Goto and Hayasaka [21,
(1.1)] under some extra conditions; see also [22, (2.2)].
Theorem 4.8. Let I ⊆ R be an integrally closed ideal with depth(R/I) = 0, e.g.,
such that I is m-primary. Then R is Gorenstein if and only if G-dimR (R/I) < ∞.
Proof. Apply Corollary 4.7 with C = R, or use Theorem 4.4 with Lemma 2.7.
We finish this section by giving two examples that show the integrally closed and
depth hypotheses of Theorem 4.8 are necessary:
Example 4.9. Let k be a field and let R = k[[x, y, z]]/(x2 , y 2 , z 2 , yz), as in Example
3.14. Then, since R is Artinian, each proper ideal of R is m-primary but is not integrally closed; see [24, (1.1.3)(3)]. In particular the principal ideal I of R generated
by x is m-primary but not integrally closed. Note that R is not Gorenstein. Also,
we have G-dim(I) = 0 because the fact that R is of the form A[x]/(x2 ) implies that
x
x
x
I has a complete resolution · · · −
→R−
→R−
→ ···.
Example 4.10. Let k be a field, set S := k[[x1 , x2 , x3 , y1 , y2 , y3 ]], and let I be
the ideal of S generated by x1 y2 − x2 y1 , x1 y3 − x3 y1 , and x2 y3 − y2 x3 . Set R :=
S/I. Then R is a four-dimensional normal Cohen-Macaulay domain that is not
Gorenstein; see [20, Theorem (a)]. Let 0 6= f ∈ m. Then the ideal f R of R
generated by f is integrally closed; see [24, (1.5.2)]. Furthermore pdR (R/f R) =
G-dimR (R/f R) = 1 and depth(R/f R) = 3.
Acknowledgments
Parts of this work were completed when Celikbas visited North Dakota State
University in November 2013, and when Sather-Wagstaff visited the University of
Connecticut in April 2015. We are grateful for the kind hospitality and the generous
support of the NDSU and UConn Mathematics Departments. We are also grateful
to Jerzy Weyman for supporting Sather-Wagstaff’s visit, and to Irena Swanson and
Ryo Takahashi for helpful feedback on this work. We thank Naoki Taniguchi and
Shiro Goto for pointing out Example 4.10 to us. We also thank the referee for
his/her valuable corrections and suggestions on the manuscript.
TESTING FOR THE GORENSTEIN PROPERTY
13
References
1. M. Auslander, Anneaux de Gorenstein, et torsion en algèbre commutative, Séminaire
d’Algèbre Commutative dirigé par Pierre Samuel, vol. 1966/67, Secrétariat mathématique,
Paris, 1967. MR 37 #1435
2. M. Auslander and M. Bridger, Stable module theory, Memoirs of the American Mathematical
Society, No. 94, American Mathematical Society, Providence, R.I., 1969. MR MR0269685 (42
#4580)
3. M. Auslander and D. A. Buchsbaum, Homological dimension in local rings, Trans. Amer.
Math. Soc. 85 (1957), 390–405. MR 0086822 (19,249d)
4. L. L. Avramov and H.-B. Foxby, Homological dimensions of unbounded complexes, J. Pure
Appl. Algebra 71 (1991), 129–155. MR 93g:18017
5.
, Locally Gorenstein homomorphisms, Amer. J. Math. 114 (1992), no. 5, 1007–1047.
MR 1183530 (93i:13019)
, Ring homomorphisms and finite Gorenstein dimension, Proc. London Math. Soc. (3)
6.
75 (1997), no. 2, 241–270. MR 98d:13014
7. L. L. Avramov, S. Iyengar, and C. Miller, Homology over local homomorphisms, Amer. J.
Math. 128 (2006), no. 1, 23–90. MR 2197067
8. L. Burch, On ideals of finite homological dimension in local rings, Proc. Cambridge Philos.
Soc. 64 (1968), 941–948. MR 0229634 (37 #5208)
9. O. Celikbas, H. Dao, and R. Takahashi, Modules that detect finite homological dimensions,
Kyoto J. Math. 54 (2014), no. 2, 295–310. MR 3215569
10. O. Celikbas, M. Gheibi, A. Sadeghi, and M. R. Zargar, Homological dimensions of rigid
modules, preprint (2014), arXiv:1405.5188.
11. L. W. Christensen, Gorenstein dimensions, Lecture Notes in Mathematics, vol. 1747, SpringerVerlag, Berlin, 2000. MR MR1799866 (2002e:13032)
12.
, Semi-dualizing complexes and their Auslander categories, Trans. Amer. Math. Soc.
353 (2001), no. 5, 1839–1883 (electronic). MR 1813596 (2002a:13017)
13. L. W. Christensen, H.-B. Foxby, and H. Holm, Derived category methods in commutative
algebra, Preprint, 2012.
14. A. Corso, C. Huneke, D. Katz, and W. V. Vasconcelos, Integral closure of ideals and annihilators of homology, Commutative algebra, Lect. Notes Pure Appl. Math., vol. 244, Chapman
& Hall/CRC, Boca Raton, FL, 2006, pp. 33–48. MR 2184788 (2006h:13015)
15. E. G. Evans and P. Griffith, Syzygies, London Mathematical Society Lecture Note Series, vol.
106, Cambridge University Press, Cambridge, 1985. MR MR811636 (87b:13001)
16. H.-B. Foxby, Gorenstein modules and related modules, Math. Scand. 31 (1972), 267–284
(1973). MR 48 #6094
17. A. Frankild and S. Sather-Wagstaff, Reflexivity and ring homomorphisms of finite flat dimension, Comm. Algebra 35 (2007), no. 2, 461–500. MR 2294611
18. A. J. Frankild, S. Sather-Wagstaff, and A. Taylor, Relations between semidualizing complexes,
J. Commut. Algebra 1 (2009), no. 3, 393–436. MR 2524860
19. E. S. Golod, G-dimension and generalized perfect ideals, Trudy Mat. Inst. Steklov. 165 (1984),
62–66, Algebraic geometry and its applications. MR 752933 (85m:13011)
20. S. Goto, When do the determinantal ideals define Gorenstein rings?, Sci. Rep. Tokyo Kyoiku
Daigaku Sect. A, 12, (1974), 129–145. MR 0360567 (50 #13015)
21. S. Goto and F. Hayasaka, Finite homological dimension and primes associated to integrally
closed ideals, Proc. Amer. Math. Soc. 130 (2002), no. 11, 3159–3164 (electronic). MR 1912992
(2003g:13014)
22. S. Goto and F. Hayasaka, Finite homological dimension and primes associated to integrally
closed ideals II, J. Math. Kyoto Univ. 42 (2002), no. 4, 631–639. MR 1967051 (2004e:13035)
23. R. Hartshorne, Residues and duality, Lecture notes of a seminar on the work of A.
Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne. Lecture Notes
in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. MR 0222093 (36 #5145)
24. C. Huneke and I. Swanson, Integral closure of ideals, rings, and modules, London Mathematical Society Lecture Note Series, vol. 336, Cambridge University Press, Cambridge, 2006.
MR 2266432 (2008m:13013)
14
OLGUR CELIKBAS AND SEAN SATHER-WAGSTAFF
25. S. Iyengar and S. Sather-Wagstaff, G-dimension over local homomorphisms. Applications
to the Frobenius endomorphism, Illinois J. Math. 48 (2004), no. 1, 241–272. MR 2048224
(2005c:13016)
26. David A. Jorgensen and Liana M. Şega, Independence of the total reflexivity conditions for
modules, Algebr. Represent. Theory 9 (2006), no. 2, 217–226. MR 2238367 (2007c:13022)
27. Takesi Kawasaki, On Macaulayfication of Noetherian schemes, Trans. Amer. Math. Soc. 352
(2000), no. 6, 2517–2552. MR 1707481 (2000j:14077)
28. J. Majadas, On test modules for flat dimension, J. Algebra Appl. 13 (2014), no. 3, 1350107,
6. MR 3125875
29. H. Matsumura, Commutative ring theory, second ed., Cambridge Studies in Advanced Mathematics, vol. 8, Cambridge University Press, Cambridge, 1989, Translated from the Japanese
by M. Reid. MR MR1011461 (90i:13001)
30. S. Sather-Wagstaff, Semidualizing modules and the divisor class group, Illinois J. Math. 51
(2007), no. 1, 255–285. MR 2346197
31.
, Lower bounds for the number of semidualizing complexes over a local ring, Math.
Scand. 110 (2012), no. 1, 5–17. MR 2900066
32. S. Sather-Wagstaff and R. Wicklein, Adic Foxby classes, in preparation.
33. J.-P. Serre, Sur la dimension homologique des anneaux et des modules noethériens, Proceedings of the international symposium on algebraic number theory, Tokyo & Nikko, 1955
(Tokyo), Science Council of Japan, 1956, pp. 175–189. MR 19,119a
34. R. Takahashi, On G-regular local rings, Comm. Algebra 36 (2008), no. 12, 4472–4491.
MR 2473342 (2009i:13040)
35. W. V. Vasconcelos, Divisor theory in module categories, North-Holland Publishing Co., Amsterdam, 1974, North-Holland Mathematics Studies, No. 14, Notas de Matemática No. 53.
[Notes on Mathematics, No. 53]. MR 0498530 (58 #16637)
36. J.-L. Verdier, Catégories dérivées, SGA 4 12 , Springer-Verlag, Berlin, 1977, Lecture Notes in
Mathematics, Vol. 569, pp. 262–311. MR 57 #3132
, Des catégories dérivées des catégories abéliennes, Astérisque (1996), no. 239, xii+253
37.
pp. (1997), With a preface by Luc Illusie, Edited and with a note by Georges Maltsiniotis.
MR 98c:18007
38. S. Yassemi, G-dimension, Math. Scand. 77 (1995), no. 2, 161–174. MR 97d:13017
University of Connecticut, Department of Mathematics, Storrs, CT 06269–3009 USA
E-mail address: [email protected]
NDSU Department of Mathematics # 2750, PO Box 6050, Fargo, ND 58108-6050 USA
Current address: Department of Mathematical Sciences, Clemson University, O-110 Martin
Hall, Box 340975, Clemson, S.C. 29634 USA
E-mail address: [email protected]
URL: http://people.clemson.edu/~ssather/
| 0math.AC
|
1
arXiv:1709.08385v1 [cs.CR] 25 Sep 2017
Deep Learning Based Cryptographic Primitive Classification
Gregory D. Hill
Xavier J. A. Bellekens
Division of Computing and Mathematics
Abertay University
Dundee, Scotland
Email: [email protected]
Division of Computing and Mathematics
Abertay University
Dundee, Scotland
Email: [email protected]
Cryptovirological augmentations present an immediate, incomparable threat. Over the last decade, the substantial proliferation
of crypto-ransomware has had widespread consequences for consumers and organisations alike. Established preventive measures
perform well, however, the problem has not ceased. Reverse
engineering potentially malicious software is a cumbersome task
due to platform eccentricities and obfuscated transmutation
mechanisms, hence requiring smarter, more efficient detection
strategies. The following manuscript presents a novel approach
for the classification of cryptographic primitives in compiled
binary executables using deep learning. The model blueprint,
a Dynamic Convolutional Neural Network (DCNN), is fittingly
configured to learn from variable-length control flow diagnostics
output from a dynamic trace. To rival the size and variability
of contemporary data compendiums, hence feeding the model
cognition, a methodology for the procedural generation of synthetic cryptographic binaries is defined, utilising core primitives
from OpenSSL with multivariate obfuscation, to draw a vastly
scalable distribution. The library, CryptoKnight, rendered an
algorithmic pool of AES, RC4, Blowfish, MD5 and RSA to
synthesis combinable variants which are automatically fed in
its core model. Converging at 91% accuracy, CryptoKnight is
successfully able to classify the sample algorithms with minimal
loss.
Index Terms—Deep Learning, Convolutional Neural Network,
Cryptovirology, Ransomware, Binary Analysis
I. I NTRODUCTION
T
HE idea of cryptovirology was first introduced by [1]
to describe the offensive nature of cryptography for
extortion-based security threats. It comprises a set of revolutionary attacks that combine strong (symmetric and asymmetric) cryptographic techniques with unique viral technology.
The fierce proliferation of ‘crypto-ransomware’ is rather troubling for a number of reasons. Designed to infect, encrypt
and lock-down available hosts, this category of malware has
had disastrous consequences for many [2] [3]. For those who
can afford to reclaim their private data, the financial loss is
typically quite substantial, despite the fact that there is no
guaranteed recovery. Ultimately, without a backup, there is
little that can be done. Preventative frameworks have been
proven to effectively halt unusual activity [4] [5] by closely
monitoring the file system’s input / output, but administrators
are not always likely to follow best practices [6]. In any
case, unknown weaknesses can still be exploited to further
an attacker’s goal.
The cryptovirological landscape has evolved in recent years.
A distinct growth has been noted in the overall number
of targeted attacks and variants [7]. A long-term study of
1,359 ransomware samples [4] observed between 2006 and
2014 found a distinct number of variants with cryptographic
capabilities. During analysis, these instances were found to
utilise both standard and customised cryptography with generational enhancements, specifically in terms of key generation and management. Tailored cryptosystems particularly
limit the scale of effective analysis [8]. Infamous variants
are known to have employed well-established documented
algorithms [8] [9], but a number of deviations are significant. In some instances, malware variants have employed
custom ‘cryptography’, lesser-known algorithms (such as the
Soviet/Russian symmetric block cipher GOST [10]), common
substitution-ciphers, or exclusively, encoding mechanisms.
The field of malware analysis seeks to determine the
potential impact of malicious software by examining it in
a controlled environment. Investigators find flaws otherwise
unknown to current identification technologies - sourcing keys
and blocking further infection [2]. When reverse engineering
a potentially malicious executable, several issues should be
addressed. Possible problems include the accuracy of analysis,
quality of the application’s obfuscation, and the lifetime of
findings. The analysis of a binary can typically be considered
from two viewpoints: static or dynamic. Static analysis is performed in a non-runtime environment, therefore examination
is relatively safe; however potential morphism restricts the
accuracy of results [11] [12]. Alternatively, dynamic analysis [13] sequentially assesses a binary throughout its execution, which can provide significantly more accurate results
and contend with obfuscatory measures [14] [15], but if not
properly handled, samples could prove somewhat hazardous.
This manuscript focuses on the latter methodology.
Cryptographic algorithm identification facilitates malware
analysis in a number of ways, but in this case, when assessing
ransomware strains, it yields a starting point for investigation.
This is essential when analytical time is restricted. With the uncertainty surrounding an application’s custom, undocumented
or established cryptosystem, analysts struggle to maintain
complete awareness of the field - which makes this task ideal
for automation. To effectively model cryptographic execution,
previous research has relied on a number of assumptions and
observations. These features do not necessarily always depict
cryptographic code, but provide a baseline for analysis. For
instance, cryptographic algorithms naturally involve the use
of bitwise integer arithmetic and logical operations. These
activities frequently reside in loops, for example, block ciphers
2
typically loop over an input buffer to decrypt it block-by-block.
[8] also postulated that any encrypted data is likely to have a
higher information entropy than decrypted data.
Deep learning studies intricate Artificial Neural Networks
(ANNs) with multiple hidden computational layers [16] that
effectively model representations of data with multiple layers
of abstraction, in which the high-level representations can
amplify aspects of the input that are important for discrimination. These techniques have been used amongst others
to identify network threats [17] or encrypted traffic on a
network [18] [19]. A Convolutional Neural Network (CNN), is
a specialised architecture of ANN that employs a convolution
operation in at least one of its layers [20] [21]. A variety of
substantiated CNN architectures have been used to great effect
in computer vision [22] and even Natural Language Processing
(NLP), with empirically distinguished superiority in semantic
matching [23], compared to other models.
CryptoKnight is developed in coordination with this
methodology. We introduce a scalable learning system that
can easily incorporate new samples through the scalable synthesis of customisable cryptographic algorithms. Its entirely
automated core architecture is aimed to minimise human
interaction, thus allowing the composition of an effective
model. We tested the framework on a number of externally
sourced applications utilising non-library linked functionality.
Our experimental analysis indicates that CryptoKnight is a
flexible solution that can quickly learn from new cryptographic execution patterns to classify unknown software. This
manuscript presents the following contributions:
• Our unique convolutional neural network architecture
fits variable-length data to map an application’s timeinvariant cryptographic execution.
• Complimented by procedural synthesis, we address the
issue of this task’s disproportionate latent feature space.
• The realised framework, CryptoKnight, has demonstrably
faster results compared to that of previous methodologies,
and is extensively re-trainable.
II. R ELATED W ORK
The cryptovirological threat model has rapidly evolved over
the last decade. A number of notable individuals and research
groups have attempted to address the problem of cryptographic
primitive identification. We will discuss the consequences of
their findings here and address intrinsic problems.
A. Heuristics
Heuristical methods [24] are often utilised to locate an
optimal strategy for capturing the most appropriate solution.
These measures have previously shown great success in cryptographic primitive identification. A joint project from ETH
Zürich and Google, Inc. [8] detailed the automated decryption
of encrypted network communication in memory, to identify
the location and time a subject binary interacted with decrypted input. From an execution trace which dynamically
extracted memory access patterns and control flow data [8],
was able to identify the necessary factors required to retrieve
the relevant data in a new process. His implementation was
successfully able to identify the location of several decrypted
webpages in memory fetched using cURL and OpenSSL, and
successfully extracted the decrypted output from a Kraken
malware binary. The entropy metric was found to negatively
affect the recognition of simple substitution ciphers as they do
not typically effect the information entropy, which was also
found to affect the analysis of GnuPG.
Gröbert at al. [25] [9] utilised fine-grained dynamic binary
analysis to generate a high-level control flow graph for evaluation using three heuristics. The chains heuristic measured
the ordered concatenation of all mnemonics in a basic block,
comparing known signatures; the mnemonic-const heuristic
extent the former method by assessing the combination of
instructions and constants; and the verifier heuristic confirmed
a relationship between the input and output (I/O) of a permutation block. Trialled on cURL, the tool detected both RivestShamir-Adleman (RSA) and Advanced Encryption Standard
(AES) in a traced Secure Sockets Layer (SSL) session. When
tested on a real-world malware sample,GpCode, the operation
successfully detected the cryptosystem and extracted its keys.
However, the trace took fourteen hours with an extra eight
hours for analysis.
The Crypto Intelligence System [26] conglomerates a number of these heuristical measures to counter the problem of
situational dependencies. As cryptography can be reduced
to Rice’s Theorem, it is suggested that different heuristics
can provide more accurate readings. Evaluating the detection
methods used by [8] and [25], Matenaar et al. found that [27]
presented the least false positives in their tests. However, in
presenting these comparisons, the problem with heuristics is
shown as in that they do not always generalise suitably.
B. Machine Learning
Attempting to address a difficulty with past methodologies,
one thesis [28] studied the suitability of machine learning.
While automated and highly efficient, thresholds [9] [11] [14]
often require manual adjustment to manage the identification
of new algorithmic samples. Hosfelt sought to emphasise the
ease of model retraining by analysing the performance of:
Support Vector Machine, Kernels, Naive Bayes, Decision Tree
and K-means Clustering. The study was met with varying
success, but ultimately suffered from a limited sampling of
the latent feature space, preventing adequate scaling for more
complex data - i.e. multi-purpose applications that may use
cryptography in addition to other functions that may unintentionally obfuscate the control flow. EldeRan [29] similarly
assessed the suitability of automated dynamic analysis for
ransomware. A regularized logistic regression classifier was
utilised in conjuncture with a dataset of 582 ransomware and
942 ‘good’ applications - thus producing an Area Under Curve
(AUC) of 0.995. While the results in this case were positive,
the methodology required a large number of malicious samples
to have been run to generate a sufficient distribution.
C. Obfuscation
A tool by the name of CryptoHunt [14] was recently developed to identify cryptographic implementations in binary code
3
Fig. 1: Framework Architecture
despite advanced obfuscation. The implementation tracked the
dynamic execution of a reference binary at instruction level,
to further identify and transform loop bodies into boolean
formulas. Each formula was designed to successfully abstract
the particular primitive, but remain compact to describe the
most emblematic features. Unlike sole I/O verification, performed by [9], this semantic depth more prominently revealed
distinguishable features regardless of obfuscation. Aligot [30]
was also designed for obfuscatory resilience, but instead chose
to focus on I/O. Both tools performed well on a variety of
samples, but required pre-existing reference implementations
and manual integration.
D. Data Flow Analysis
Two papers studied representational patterns of cryptographic data through dynamic analysis [31] [32]. By closely
monitoring an application’s I/O each methodology aimed to
pinpoint a cryptographic algorithm that matched a similar
pattern. [15] alternatively assessed the ‘avalanche effect’ as
a unique discriminatory feature, where a small change in the
input would dramatically alter the output. Although effective,
none of the methods would likely adapt to unique obfuscations.
A unique and fairly effective approach for the identification
of symmetric algorithms in binary code was based on subgraph isomorphism and static analysis [11]. Lestringant et al.
resolved each cryptographic algorithm to a Data Flow Graph
(DFG), normalising their structure without breaking semantics.
The proposed subgraph isomorphism step then assessed which
signatures were contained within the normalised DFG. For the
targeted sample pool (XTEA, Message Digest 5 (MD5), AES),
their implementation reached 100% accuracy. Unfortunately,
the formula relied on the manual selection of appropriate
signatures which distinguish the applicable algorithms. For
these three instances, generation was elementary, but would
not realistically scale to the dimensions sought in this paper.
III. OVERVIEW
With any given subject application, this methodology aims
to automatically verify the existence of pre-determined cryptographic signatures in unknown binary code. The intention is
to provide a solution that can easily generalise when presented
with new conditions, without manually adjusting a number of
thresholds. The full system, as in Figure 1, is comprised of
three higher-level stages.
1) Procedural generation guides the synthesis of unique
cryptographic binaries with variable obfuscation and
alternate compilation.
2) Assumptions of cryptographic code aid the discrimination of diagnostics from the dynamic analysis of
synthetic or reference binaries, to build an ‘image’ of
execution.
3) A DCNN fits variable-length matrices for ease of training and the immediate classification of new samples.
IV. S YNTHESIS
To construct a reasonably sized dataset with enough variation to satisfy the abstraction of cryptographic primitives,
it is not enough to simply hand-write a small number of
applications with little diversity in terms of operational outliers. For example, extracting features from the execution
of a manually implemented single-purpose binary may give
an appropriate feature vector, but re-running the extraction
process will not provide any variation for repeating labels,
outside of environmental setup. This methodology leverages
procedural generation to include elements that provide some
obfuscation without directly altering the intended control flow.
For the three main algorithmic categories - symmetric, asymmetric and hashing - interpretation should correlate the related
components to dynamically construct a unique executable.
A. Artefacts
OpenSSL is an open source cryptographic library that provides an Application Programming Interface (API) for accessing its algorithmic definitions. Review of its documentation revealed a number of similarities in the intended implementation
of the each function. These specificities are either: variable,
differ for each primitive; or constant, true for each category.
This approach exclusively examined C in experimentation, but
C++ can also be integrated.
4
Via appropriate headers, an application first imports the
libraries which provide the expectant functionality when later
compiled. In this case, from either OpenSSL or the C Standard Library. Each primitive naturally requires contrastive
functionality, so this is variable. Within the scope of an
application’s main body however, each symmetric algorithm
requires the specification of a key and Initialization Vector
(IV), asymmetric algorithms require a certificate declaration
whereas hashing definitions do not expect either. These rules
are categorically constant, therefore, their definitions can be
specified by type. Next, a plaintext sequence is loaded into
memory - directly or from a file - and ciphertext memory
is allocated, also constant. Each sample will then employ its
unique algorithm through differing declarations, reading in the
plaintext, key or IV.
code will dramatically fluctuate. Real-world instances are
rarely compiled identically so multivariate output can provide
further assurance for generalisation.
V. F EATURE E XTRACTION
Cryptographic execution is time-invariant. Therefore, a reference binary may employ its associated functions at any point
within a trace. Unintentional obfuscation of the control flow
will negatively affect discriminatory performance, so granularity needs to be high. An underlying problem of previous
work [28]. This following approach opts to draw appropriate
features from a reference binary using dynamic instrumentation via Intel’s Pin API. Through the disassembly of runtime instruction data, this section’s outlined measurements
principally assess the activity’s importance in relation to the
assumptions of cryptographic code.
B. Obfuscation
Two primary transformation mechanisms were highlighted
by [14]. The first technique discusses the abstraction of
relevant data groups to decrease their perceptible mapping. For
example, a multidimensional array may be concatenated into
a single column and either expanded or accessed as necessary.
The second technique concerns itself with the splitting of
variables [33] to disguise their representation. Colloquially
known as ‘data aggregation’ and ‘data splitting’, these methods
partially obfuscate the data flow without subtracting from a
application’s distinct activity.
Outside of such distinct obfuscation, the inclusion of structured loops, arithmetical or bitwise operations create discriminatory irregularity in an otherwise translucent process.
Analogously, many ‘training’ images for computer vision will
contain noise that suitably detract from the trivial classification
of its subject. This process aims to replicate such uncertainty.
C. Interpretation
Formatting each respective variable artefact to allow ease
of parsing in a similarly structured markup tree will allow
the interpretation of unique cryptographic applications with
alternate obfuscation. Stochastically generated keys, IVs and
plaintext will add additional variation into each image. Algorithm 1 outlines the pseudo-code for this procedure.
Algorithm 1 Cryptographic Synthesis
Input: cryptographic constants & variables
Output: application code
New Sample :
1: select obfuscation - (aggregation, split, normal)
2: write to file:
import statements
abstracted keys
encryption routines
3: inject randomised arithmetic
4: return relative location
When compiling the resultant collection of cryptographic
applications, data variability can be further increased. With
alternate compilers or optimisation options, the resultant object
A. Basic Blocks & Loops
A Basic Block (BBL) is a sequential series of instructions
executed in a particular order, exclusively defined when there
is one branch in (entry) and one branch out (exit). A BBL
ends with one of the following conditions.
• unconditional or conditional branch - direct / indirect.
• return to caller.
Each instruction is evaluated in a linear trace, if any criteria
are met, it is marked as a tail. The following instruction is
delimited as the head of the subsequent sequence, but can
similarly be identified as a tail. The ‘stack’ of execution
stores relevant data from each instruction, and two boolean
expressions indicate the predetermined blocks. As each BBL
is dynamically revealed, non-executed instructions will unfortunately not be observed [9], but we can monitor indirect
branches.
Many high level languages share a distinctly strict definition
of a loop, contrarily, common interpretations of amorphous
code are loose. Extending previous definitions [34] [35] we
hence delineate a loop upon the immediate re-iteration of any
BBL, as output from Algorithm 2.
Algorithm 2 Instruction Sequencing, BBL Detection
Input: run-time hooks
Output: path of execution
Callback :
1: head = false
2: if (last instruction is tail) then
3:
head = true
4: end if
5: if (instruction is branch, call or return) then
6:
tail = true
7: else
8:
tail = false
9: end if
10: if (write) then
11:
get entropy (memory write)
12: end if
13: return stack
5
8,000
B. Instructions
Conventional architectures use a common instruction format, interpretable as: opcode operand (destination/source). In
x86, there are zero to three operands (separated by commas),
two of which specify the destination and source. For example, when AES performs a single round of an encryption
flow it calls the instruction 66 0f 38 dc d1 which can be
disassembled as aesenc xmm2, xmm1. Directly operating on
the first operand, in this case xmm2, it performs a round of
AES encryption using the 128-bit round key from xmm1. Although this is a interesting example, the Advanced Encryption
Standard New Instructions (AES-NI) architecture presents a
problem for later generalisation as cryptographic acceleration
prevents detailed analysis.
Alternate object code is typically quite distinctive, especially in cryptographic code. A primitive may employ a number of operations, in any order, and it is important not to dwell
on specificities - i.e. exact semantics. For each instruction
in the ‘carved’ linear trace we weight its ratio of bitwise
operations upon the cross-correlation of prominent operators
from a pool of cryptographic routines for discriminatory
emphasis.
8,000
Blowfish
6,000
6,000
4,000
4,000
2,000
2,000
0
RC4
0
0
8,000
1,000
2,000
0
1,000
8,000
MD5
6,000
6,000
4,000
4,000
2,000
2,000
0
2,000
RSA
0
0
1,000
2,000
0
1,000 2,000 3,000 4,000 5,000
·104
6
AES
4
2
0
0
1,000 2,000 3,000 4,000 5,000
C. Entropy
As characterized by [36], the associated uncertainty of a
finite discrete probability distribution p = (p1 , p2 , . . . , pn ) can
be measured using
PnShannon’s Entropy. Suppose pk > 0(k =
1, 2, . . . , n) and k=1 pk = 1, distribution p is measured by
quantity H[p] = H(p1 , p2 , . . . , pn ) hence defined as:
H(p1 , p2 , . . . , pn ) =
n
X
k=1
pk log2
1
pk
(1)
Upon detecting a memory write, the respective location’s
contents can be replicated. Casting its value to distribution p
will allow the immediate calculation of H[p]. Related memory
can then be deleted to prevent unnecessary exhaustion. Each
BBLs absolute entropy increase / decrease can then be scored
by its relation to prior activations over opposing registers and
then summated as in Figure 2 where each BBL is ∈ W.
VI. M ODEL
Founded in earlier research [37], this proposed definition
of DCNN treats the input in a manner similar to that of a
sentence. For each word, embeddings are defined as d, where
di corresponds to the total weight of a particular operation,
multiplied by its entropic score. Feature vector wi ∈ Wd is a
therefore a column in sentence matrix s such that s ∈ Wd×s .
Let’s say w = 128, exclusively assessing pure arithmetic
impact would make d = 12, so s would equal 12 × 128.
The model itself combines a number of one wide convolutions
with (dynamic) k-max pooling and folding to map variablelength input. Topologically presented in Figure 3, the model
combines a number of one wide convolutions with (dynamic)
k-max pooling and folding to map variable-length input.
Mathematically, convolution is an operation on two functions of a real-valued argument, in this case a vector of weights
Fig. 2: Entropy Scoring
The relative entropy scoring (y) of each BBL (x) within a trace
from a sample set of algorithms.
m ∈ Wm and a vector of inputs s ∈ Ws , to yield a new
sequence c. With kernel m, the convolved sequence in a onewide convolution takes the form c ∈ Rs+m−1 , thus preserving
the number of defined embeddings.
Equation 2 describes the selection of k, a distinct subset of s
that most relevantly depicts an l-th orders progression. Based
on the total number of convolutional layers L, the current
layer l, the projected sentence length s and the predefined
final pooling parameter ktop , any particular layer’s selection
is delimited as:
kl = max(ktop , [
L−l
s])
L
(2)
With value k and sequence p ∈ Rp of length p ≥ k the base
k-max pooling operation selects the subsequence pkmax of the
k most active features. Completely unreactive to positional
variation, it not only preserves the original perspective, but
can also distinguish repetitious features. Since identifiable
cryptographic routines may execute at any point in a sequential
trace, this operation fits perfectly.
Another significant phase in this procedure simplifies
the way the model perceives complex dependencies over
rows. Veritable feature independence is removed through
component-wise summation of every two rows, shrinking d/2.
Nested between a convolutional layer and (dynamic) k-max
pooling, a folding layer halves the respective matrix.
While the model itself can automatically scale with new
variants, a number of hyper-parameters may need to be ad-
6
80
Accuracy
justed to better suit the sample pool. Manual tuning is an
inefficient, costly process that frequently offers no advantage.
[38] empirically show that randomly chosen trials are more
efficient for hyper-parameter optimization than manual or
grid search over the same domain - calculating the most
viable constants in a fraction of the time. Taking the shuffled
Cartesian product of all hyper-parameter subsets allows the
ambiguated selection of distinct constants for trial on a small
number of epochs.
60
40
20
0
50
100
150
200
Epoch
Fig. 4: Accuracy
with more tanh non-linear activations and k-max pooling further reduced the feature space. Due to the number of embeddings the model utilised two folding operations simultaneously
prior to the penultimate pooling layer, after its convolution.
The final convolution built two additional feature maps and
then pooled on the stipulated topmost magnitude of 56. A
linear transformation, of output size 8, was then applied
with softmax to fully connect the model. Filter widths were
specified for the first and last convolutional layers of 20 and 12
respectfully, the remaining hidden layers shared a width of 10.
In total, the model employed 14 convolutions.
Category
Symmetric
Asymmetric
Hashing
Convolution
k-Max Pooling
Folding
Fully Connected
Fig. 3: Dynamic Convolutional Neural Network (DCNN)
The architecture of a DCNN as illustrated by [37]. In this case, the
model is intended for a seven word input sentence of embedding
d = 4 with two convolutional layers (m = 3 and m = 2), two
(dynamic) k-max pooling layers (k = 5 and k = 3) within two
feature maps.
VII. E XPERIMENTATION
Table I presents a range of popular algorithms which were
selected for experimental analysis. These primitives are widely
utilised for both legitimate and fraudulent purposes.
CryptoKnight was configured to build two feature maps
from its first convolution, with a dropout probability of 0.6,
through a round of k-max pooling and non-linear activation.
Eleven hidden one-dimensional wide convolutions interspersed
Algorithm
AES
RC4
Blowfish
RSA
MD5
TABLE I: Cryptographic Algorithms
Based on a frequency analysis of simplified opcodes within
the sample pool, the mnemonics add, sub, inc, dec, shr, shl,
and, or, xor, pxor, test & lea were selected for weighting.
Therefore the final design matrix, of embedding 12, contained
a variable number of vectors corresponding to the numeration and associated weightings of each basic block in a
subject binary. A distribution of size n = 750 was drawn
in which 75% was used for training and 25% remained for
testing. After 200 epochs, the model successfully converged
at 91.3% accuracy with a minimal loss of 0.29. Figure 4 shows
the test accuracy over 200 epochs, and Figure 5 displays its
simultaneous loss. An additional collection was then drawn
for validation. Table II diagnoses the pre-trained model’s
associated confusion on an additional n = 300 samples.
We collected five open source implementations, presented
in Table III, from GitHub 1 for Rivest Cipher 4 (RC4) and
1 https://github.com
7
AES
RC4
BLF
MD5
RSA
R/A
10
Loss
8
6
AES
37
0
0
0
19
0
RC4
0
50
8
2
0
0
BLF
0
0
42
7
0
0
MD5
0
0
0
41
0
0
RSA
13
0
0
0
30
0
R/A
0
0
0
0
1
50
TABLE II: Validation Results
4
x = predicted, y = actual
2
0
0
50
100
150
200
Epoch
No representational RSA or MD5 instances were identified at time of testing. AES was overlooked due to the
model having been trained on cryptographically accelerated
binaries. The classification rate varied with hyper-parameter
optimisations and distribution sizes, but was able to correctly
classify 4/5 RC4 and 4/5 Blowfish samples in our tests.
Algorithm
Fig. 5: Loss
RC4
80
Accuracy
Blowfish
70
60
Source
github.com/ogay/rc4
github.com/tomaslu/rc4
github.com/Maihj/RC4
github.com/anthonywei/rc4/tree/master/c
github.com/shiffthq/rc4
github.com/jdiez17/blowfish
github.com/Rupan/blowfish
github.com/rkhullar/cipher-c
github.com/genterist/simpleBlowFishX
github.com/JaseP88/Blowfish-Algorithm
TABLE III: Open Source Implementations
50
0
50
100
150
200
Epoch
Fig. 6: No Entropy
Blowfish. The resultant collection of binaries leveraged pure
(non-library linked) cryptographic functionality to assess our
method.
Analysing GnuPG 1.4.20, the software was directed to
encrypt an empty text document using 256-bit AES. With a
trace time of 1 minute 40 seconds, CryptoKnight predicted the
utilisation of RSA with AES.
VIII. D ISCUSSION
Traditional cryptographic identification techniques are inherently expensive [8] [9] [11] [14] and heavily rely on human
intuition. CryptoKnight was built to reduce this associated
error-prone interaction. With refined sampling of the latent
feature space, a procedurally synthesised distribution allowed
our DCNN to map proportional linear sequences with a finer
granularity than that of conventional architectures without
overfitting, CryptoKnight converged at 91% accuracy without
extensive hyper-parameter optimisation. The model ultimately
fit the synthetic distribution with veritable ease, with performance on par to that of [37].
The impediment of dynamic binary instrumentation was
made clear by [9] who highlighted an extensive twenty two
hour trace and analysis time. CryptoKnight’s analysis time also
varied, but not quite to this extent. Hence, for the sample
binaries, analysis took up to a maximum of around one
minute. Consequently, large collections saw exponential draw
times of indefinable length. Since manual analysis often takes
invariably longer than an adequate draw time, this footprint
is arguably marginal. Once trained, the proposed framework
would be most beneficial as part of an analyst’s toolkit - to
quickly verify any cryptographic instances.
8
The DCNN was intended to map time-invariant cryptographic execution despite control or data flow obfuscation.
An intrinsic problem of initial work and the approach by
used [28]. While successful, this formulation still has a
few preliminary issues to address. The predefined operator
embeddings explicitly define the entire feature set, therefore
new samples which perhaps deviate from traditional operation
may be unidentifiable. Should the framework not immediately
generalise to an additional algorithm, correlating its most predominant operators by interchanging embeddings or enlarging
the scope should enhance cognition.
A fundamental part of CryptoKnight’s supervised design is
that it can only classify known samples. New cryptographic
algorithms must be added to the generation pool - a process
this framework has strived to simplify, but an unavoidable
limitation of the proposed architecture. This also makes custom cryptography more difficult to classify, as no high-level
reference implementations would feasibly exist to import. Integrating an unsupervised component into the core model could
facilitate the detection of non-cryptographic signatures. This,
or alternatively advanced synthesis could negate the need for
procedural generation entirely, to further reduce the presently
expensive time requirement and/or aid in the classification of
customised cryptography. A multi-class element that learns
application invariant primitives would also be prove beneficial
as CryptoKnight was exclusively trained on pre-combined
functions. A similar model could be re-purposed to decompile
binaries with more accuracy than traditional methods - which
typically only manage simplistic control flow.
The entropy metric assumes that a cryptographic function’s
associated uncertainty is higher than that of conventional interaction. In the case of [8], this negatively affected the recognition of simple substitution ciphers, however it is unlikely
to affect CryptoKnight in the same way due to its scoring
mechanics and demonstrably high accuracy in subjective tests
without the metric. Figure 6 proves that the entropy metric
does not impact classification rate, converging at 83.3% - 8̃%
difference.
Problems with cryptographic acceleration played an important role in the detection of native AES implementations.
Intel’s AES-NI extension was proposed in 2008 for boosting
the relative speed of encryption and decryption on microprocessors from Intel and AMD. [39] describe the instruction set
with regard to its breakthrough performance increase. The six
instructions, prefixed by AES, directly perform each of the
cipher’s operations on the Streaming SIMD Extensions (SSE)
XMM registers, however, the natural progression of the cipher
could not be fully observed.
IX. C ONCLUSION
Despite advanced countermeasures, the cryptovirological
threat has significantly increased over the last decade, incentivising the aforementioned research. Our research demonstrated that cryptographic primitive classification in compiled
binary executables could be achieved successfully using a dynamic convolutional neural network. We also demonstrated our
implementation and achieved 91% accuracy without extensive
hyper-parameter optimisation. Moreover, our implementation
is fundamentally more flexible than that of previous work,
marginalising the error prone human element. The framework
successfully detected every implementation of RC4 and Blowfish (including externally sourced, native (non-library) written
compositions) and maintained a distinctively high accuracy on
synthetic implementations. Future work includes the detection
of cryptographic functions in parallel with the execution of a
subject binary.
R EFERENCES
[1] A. Young and M. Yung, “Cryptovirology: extortion-based security
threats and countermeasures,” in Proceedings 1996 IEEE Symposium
on Security and Privacy, May 1996, pp. 129–140.
[2] J. Snow, “Cryptxxx ransomware,” Apr 2016. [Online]. Available:
https://blog.kaspersky.com/cryptxxx-ransomware/11939/
[3] A.
Chiu,
“Player
3
has
entered
the
game:
Say
hello
to
’wannacry’,”
May
2017.
[Online].
Available:
http://blog.talosintelligence.com/2017/05/wannacry.html
[4] A. Kharraz, W. Robertson, D. Balzarotti, L. Bilge, and E. Kirda, “Cutting
the gordian knot: A look under the hood of ransomware attacks,” in
International Conference on Detection of Intrusions and Malware, and
Vulnerability Assessment. Springer, 2015, pp. 3–24.
[5] N. Scaife, H. Carter, P. Traynor, and K. R. Butler, “Cryptolock (and drop
it): stopping ransomware attacks on user data,” in Distributed Computing
Systems (ICDCS), 2016 IEEE 36th International Conference on. IEEE,
2016, pp. 303–312.
[6] C. Deane-McKenna, “Nhs ransomware cyber-attack was preventable,”
May 2017. [Online]. Available: https://theconversation.com/nhsransomware-cyber-attack-was-preventable-77674
[7] C. Beek, “Mcafee labs threats report,” Intel Security, Sep 2016.
[Online]. Available: https://www.mcafee.com/us/resources/reports/rpquarterly-threats-sep-2016.pdf
[8] N. Lutz, “Towards revealing attackers’ intent by automatically decrypting network traffic,” Master’s thesis, ETH Zürich, Switzerland, 7 2008,
a joint project between the ETH Zurich and Google, Inc.
[9] F. Gröbert, C. Willems, and T. Holz, Automated Identification of
Cryptographic Primitives in Binary Programs. Berlin, Heidelberg:
Springer Berlin Heidelberg, 2011, pp. 41–60. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-23644-0 3
[10] IBM.
(2016)
Bucbi
ransomware.
[Online].
Available:
https://exchange.xforce.ibmcloud.com/collection/Bucbi-Ransomware16eef23d3b7ea484ed69ecd78b6c1232
[11] P. Lestringant, F. Guihéry, and P.-A. Fouque, “Automated identification
of cryptographic primitives in binary code with data flow graph
isomorphism,” in Proceedings of the 10th ACM Symposium on
Information, Computer and Communications Security, ser. ASIA CCS
’15. New York, NY, USA: ACM, 2015, pp. 203–214. [Online].
Available: http://doi.acm.org/10.1145/2714576.2714639
[12] A. Moser, C. Kruegel, and E. Kirda, “Limits of static analysis for
malware detection,” in Computer security applications conference, 2007.
ACSAC 2007. Twenty-third annual. IEEE, 2007, pp. 421–430.
[13] C.-K. Luk, R. Cohn, R. Muth, H. Patil, A. Klauser, G. Lowney,
S. Wallace, V. J. Reddi, and K. Hazelwood, “Pin: building customized
program analysis tools with dynamic instrumentation,” in Acm sigplan
notices, vol. 40. ACM, 2005, pp. 190–200.
[14] D. Xu, J. Ming, and D. Wu, “Cryptographic function detection in obfuscated binaries via bit-precise symbolic loop mapping,” in In Proceedings
of the 38th IEEE Symposium on Security and Privacy, 2017.
[15] X. Li, X. Wang, and W. Chang, “Cipherxray: Exposing cryptographic
operations and transient secrets from monitored binary execution,” IEEE
transactions on dependable and secure computing, vol. 11, no. 2, pp.
101–114, 2014.
[16] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
[17] E. Hodo, X. Bellekens, A. Hamilton, P. L. Dubouilh, E. Iorkyase,
C. Tachtatzis, and R. Atkinson, “Threat analysis of iot networks using artificial neural network intrusion detection system,” in 2016 International
Symposium on Networks, Computers and Communications (ISNCC),
May 2016, pp. 1–6.
9
[18] E. Hodo, X. Bellekens, E. Iorkyase, A. Hamilton, C. Tachtatzis,
and R. Atkinson, “Machine learning approach for detection of
nontor traffic,” in Proceedings of the 12th International Conference
on Availability, Reliability and Security, ser. ARES ’17. New
York, NY, USA: ACM, 2017, pp. 85:1–85:6. [Online]. Available:
http://doi.acm.org/10.1145/3098954.3106068
[19] P. Fruehwirt, S. Schrittwieser, and E. Weippl, “Using machine learning
techniques for traffic classification and preliminary surveying of an
attackers profile,” in Proc. of Int. Conf. on Privacy, Security, Risk and
Trust, 2014.
[20] Y. LeCun et al., “Generalization and network design strategies,” Connectionism in perspective, pp. 143–155, 1989.
[21] E. Hodo, X. J. A. Bellekens, A. Hamilton, C. Tachtatzis, and R. C.
Atkinson, “Shallow and deep networks intrusion detection system: A
taxonomy and survey,” CoRR, vol. abs/1701.02145, 2017. [Online].
Available: http://arxiv.org/abs/1701.02145
[22] Y. LeCun, K. Kavukcuoglu, and C. Farabet, “Convolutional networks
and applications in vision,” in Circuits and Systems (ISCAS), Proceedings of 2010 IEEE International Symposium on. IEEE, 2010, pp. 253–
256.
[23] B. Hu, Z. Lu, H. Li, and Q. Chen, “Convolutional neural network
architectures for matching natural language sentences,” in Advances in
neural information processing systems, 2014, pp. 2042–2050.
[24] J. Pearl, “Heuristics: intelligent search strategies for computer problem
solving,” 1984.
[25] F. Grbert, “Automatic identification of cryptographic primitives in software,” 2010.
[26] F. Matenaar, A. Wichmann, F. Leder, and E. Gerhards-Padilla, “Cis:
The crypto intelligence system for automatic detection and localization
of cryptographic functions in current malware.”
[27] J. Caballero, P. Poosankam, C. Kreibich, and D. Song, “Dispatcher:
Enabling active botnet infiltration using automatic protocol reverseengineering,” 2009.
[28] D. D. Hosfelt, “Automated detection and classification of
cryptographic algorithms in binary programs through machine
learning,” CoRR, vol. abs/1503.01186, 2015. [Online]. Available:
http://arxiv.org/abs/1503.01186
[29] D. Sgandurra, L. Muñoz-González, R. Mohsen, and E. C. Lupu, “Automated dynamic analysis of ransomware: Benefits, limitations and use
for detection,” arXiv preprint arXiv:1609.03020, 2016.
[30] J. Calvet, J. M. Fernandez, and J.-Y. Marion, “Aligot: cryptographic
function identification in obfuscated binary programs,” in Proceedings
of the 2012 ACM conference on Computer and communications security.
ACM, 2012, pp. 169–182.
[31] Z. Puhan, W. Jianxiong, W. Xin, and W. Zehui, “Decrypted data
detection algorithm based on dynamic dataflow analysis,” in 2014 International Conference on Computer, Information and Telecommunication
Systems (CITS), July 2014, pp. 1–4.
[32] R. Zhao, D. Gu, J. Li, and R. Yu, Detection and Analysis
of Cryptographic Data Inside Software.
Berlin, Heidelberg:
Springer Berlin Heidelberg, 2011, pp. 182–196. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-24861-0 13
[33] S. Drape, “Intellectual property protection using obfuscation,” 2009.
[34] J. Tubella and A. Gonzalez, “Control speculation in multithreaded
processors through dynamic loop detection,” in Proceedings 1998 Fourth
International Symposium on High-Performance Computer Architecture,
Feb 1998, pp. 14–23.
[35] T. Moseley, D. Grunwald, D. A. Connors, R. Ramanujam, V. Tovinkere,
and R. Peri, “Loopprof: Dynamic techniques for loop detection and profiling,” in Proceedings of the 2006 Workshop on Binary Instrumentation
and Applications (WBIA), 2006.
[36] A. Rényi et al., “On measures of entropy and information,” in Proceedings of the fourth Berkeley symposium on mathematical statistics and
probability, vol. 1, 1961, pp. 547–561.
[37] N. Kalchbrenner, E. Grefenstette, and P. Blunsom, “A convolutional
neural network for modelling sentences,” CoRR, vol. abs/1404.2188,
2014. [Online]. Available: http://arxiv.org/abs/1404.2188
[38] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13, no. Feb, pp.
281–305, 2012.
[39] K. Akdemir, M. Dixon, W. Feghali, P. Fay, V. Gopal, J. Guilford,
E. Ozturk, G. Wolrich, and R. Zohar, “Breakthrough aes performance
with intel aes new instructions,” White paper, June, 2010.
| 9cs.NE
|
Heuristic Online Goal Recognition in Continuous Domains
arXiv:1709.09839v1 [cs.AI] 28 Sep 2017
Mor Vered
Bar Ilan University, Ramat-Gan, Israel
[email protected]
Abstract
Goal recognition is the problem of inferring the
goal of an agent, based on its observed actions. An
inspiring approach—plan recognition by planning
(PRP)—uses off-the-shelf planners to dynamically
generate plans for given goals, eliminating the need
for the traditional plan library. However, existing
PRP formulation is inherently inefficient in online
recognition, and cannot be used with motion planners for continuous spaces. In this paper, we utilize a different PRP formulation which allows for
online goal recognition, and for application in continuous spaces. We present an online recognition
algorithm, where two heuristic decision points may
be used to improve run-time significantly over existing work. We specify heuristics for continuous
domains, prove guarantees on their use, and empirically evaluate the algorithm over n hundreds of
experiments in both a 3D navigational environment
and a cooperative robotic team task.
1
Introduction
Goal recognition is the problem of inferring the (unobserved)
goal of an agent, based on a sequence of its observed actions [Hong, 2001; Blaylock and Allen, 2006; Baker et al.,
2007; Lesh and Etzioni, 1995]. It is a fundamental research
problem in artificial intelligence, closely related to plan, activity, and intent recognition [Sukthankar et al., 2014]. The
traditional approach to plan recognition is through the use
of a plan-library, a pre-calculated library of known plans to
achieve known goals [Sukthankar et al., 2014]. Ramirez and
Geffner [2010] introduced a seminal recognition approach
which avoids the use of a plan library completely. Given a
set of goals G, the Plan Recognition by Planning (PRP) approach uses off-the-shelf planners in a blackbox fashion, to
dynamically generate recognition hypotheses as needed.
The use of PRP in continuous domains, and in an online fashion (i.e., when observations are made incrementally)
raises new challenges. The original [2010] formulation, relies on synthesizing two optimal plans for every goal g ∈ G:
(i) a plan to reach goal g in a manner compatible with the observations O; and (ii) a plan to reach goal g while (at least
partially) deviating from O, i.e. complying with O. The likelihood of each goal is computed from the difference in costs
Gal A. Kaminka
Bar Ilan University, Ramat-Gan, Israel
[email protected]
of optimal solutions to the two plans. Overall, 2|G| planning
problems are solved, two for each goal.
However, in online recognition the set O is incrementally
revealed, and O changes with it. Thus two new planning
problems are solved with every new observation, for a total
of 2|G||O| calls to the planner instead of 2|G|. In addition,
using an off-the-shelf continuous-space planner to generate a
plan that may partially go through previous observations, but
must not go through all of them, is currently impossible given
the state of the art.
We present a general heuristic algorithm for online recognition in continuous domains that solves at most |G|(|O| + 1)
planning problems, and at best, |G|. The algorithm relies on
an alternative formulation, that does not use O. It has two
key decision points where appropriate heuristics reduce the
number of calls to the planner: for each new observation, the
first decision is whether to generate and solve a new planning problem for each g, or remain with the former calculated
plans. In the best case, this may reduce the number of overall calls to the planner to only |G| calls. A second decision
is whether to prune unlikely goal candidates, incrementally
reducing |G|, thus making fewer calls to the planner.
We describe the algorithm in detail, and examine several heuristic variants. Utilizing off-the-shelf continuousspace planners, without any modification, we evaluate the
different variants in hundreds of recognition problems, in
two continuous-environment tasks: a standard motion planning benchmark, and simulated ROS-enabled robots utilizing
goal-recognition for coordination.
2
Related Work
Sukthankar et al. [2014] provide a survey of recent work in
goal and plan recognition, most of it assuming a library of
plans for recognition of goals. Though successful in many
applications, library-based methods are limited to recognizing known plans. Alternative methods have been sought.
Geib [2015], and Sadeghipour et al. [2011] offer methods that utilize the same library for both planning and planrecognition. Hong [2001] presents an online method, with
no use of a library, but lacking the ranking of the recognized
goals. Baker et.al [2005] present a Bayesian framework to
calculate goal likelihoods, marginalizing over possible actions. Keren et al. [Keren et al., 2015] investigate ways to
ease goal recognition by modifying the domain.
Ramirez and Geffner [2010] proposed the PRP formulation (which plans for g twice: with O and with O), for offline
recognition. We build on their earlier formulation [2009], in
which they did not probabilistically rank the hypotheses, as
we do here. This allows us to more efficiently compute the
likelihood of different goals, given incrementally revealed observations. We embed this formulation in a definition of plan
recognition for continuous spaces, which also varies from the
original in that the recognizer observes effects, rather than
actions.
Other investigations of PRP exist. [Masters and Sardina,
2017] provided a simpler formula than that of [2010] achieving identical results in half the time, still in discrete environments. Sohrabi et. al [2016] also observe effects, though
in discrete environments, and have also sought to eliminate
planner calls, by using a k-best planner in an offline manner to sample the plans explaining the observations. Ramirez
and Geffner [2011] extend the model to include POMDP settings with partially observable states. Martin et. al [2015] and
Pereira et al. [2017] refrain from using a planner at all, instead
using pre-computed information (cost estimates and landmarks, resp.) to significantly speed up the recognition. These
approaches complement ours. Vered and Kaminka [2016]
present an online recognizer, which we prove is a special case
of the algorithm we present here. We go a significant step beyond by introducing heuristics to significantly improve both
run-time and accuracy.
3
Goal Recognition in Continuous Spaces
We begin by giving a general definition of the goal recognition problem in continuous spaces ( Section 3.1). We proceed
to develop an efficient online recognizer, which can utilize
heuristics to further improve the efficiency ( Section 3.2). We
then discuss such heuristics in detail ( Section 3.3).
3.1
Problem Formulation
We define R, the online goal recognition problem in continuous spaces as a quintuple R = hW, I, G, O, M i. W ⊆ Rn
is the world in which the observed motion takes place, as defined in standard motion planning [LaValle, 2006]. I ∈ W ,
the initial pose of the agent. G, a set of goals; each goal
g ∈ W , i.e., a point. O, a discrete set of observations, where
for all o ∈ O, o ⊂ W , i.e., each observation a specific subset
of the work area i.e., a point or trajectory. M , a (potentially
infinite) set of plan trajectories, each beginning in I, and ending in one of the goal positions g ∈ G. For each goal g, there
exists at least one plan mg ∈ M that has it as its end point.
Intuitively, given the problem R, a solution to the goal
recognition problem is a specific goal v ∈ G that best
matches the observations O. For each goal g, trajectories mg
(ending with g) are matched against the observations O.
Formally, we seek to determine v ≡ argmaxg∈G P r(g|O).
Ramirez and Geffner [2009, Thm. 7] have shown that necessarily, a goal g is a solution to the goal recognition problem
iff the cost of an optimal plan to achieve g (denoted here ig ,
for ideal plan) is equal to the cost of an optimal plan that
achieves g, while including all the observations (a plan we
refer to as mg ). Vered and Kaminka [2016] build on this
to establish a ranking over the goals. They define the ratio
cost(i )
score(g) ≡ cost(mgg ) , and rank goals higher as score(g) gets
closer to 1. They show experimentally that the ratio works
well in continuous domains, and thus we use it here (ignoring priors on P r(g) for simplicity): P P
r(g|O) ≡ ηscore(g),
where the normalizing constant η is 1/ g∈G score(g).
The next step is to compute the plans ig (ideal plan, from
initial pose I to goal g) and mg (an optimal plan that includes
the observations). As described in [Vered et al., 2016], computing ig is a straightforward application of a planner. The
synthesis of mg is a bit more complex, as mg candidates must
minimize the error in matching the observations.
To do this, we take advantage of the opportunity afforded
by the equal footing of observations O and plans in M in
continuous environments. Each observation is a trajectory or
point in continuous space. Each plan is likewise a trajectory
in the same space, as plans are modeled by their effects. Thus
generating a plan mg that perfectly matches the observations
is done by composing it from two parts:
• A plan prefix, (denoted m−
g ) is built by concatenating
all observations in O into a single trajectory ( [Masters
and Sardina, 2017] have shown that the same plan prefix
may be generated for all possible trajectories ).
• A plan suffix (denoted m+
g ) is generated by calling the
planner, to generate a trajectory from the last observed
point in the prefix m−
g (the ending point of the last observation in O) to the goal g.
Using ⊕ to denote trajectory concatenation, a plan mg ≡
+
m−
g ⊕ mg is a trajectory from the first observed point in O,
to g. Notice that mg necessarily perfectly matches the observations O, since it incorporates them.
Given a goal g and a sequence of observations O, the planner is called twice: to generate ig and to generate m+
g , used
to construct mg . The cost of ig and mg is contrasted using
a scoring procedure, denoted match(mg , ig ), which uses the
ratio as described above. As ig does not depend on O, it can
be generated once for every goal, while mg needs to be resynthesized from its component parts as O is incrementally
revealed. This establishes the baseline of (1 + |O|)|G| calls
to the planner [Vered et al., 2016]. We now generalize this
procedure to further improve on this baseline.
3.2
Heuristic Online Recognition Algorithm
We identify two key decision points in the baseline recognition process described above, that can be used to improve its
efficiency:
• Recompute plans only if necessary, i.e., if the new
observation may change the ranking (captured by a
RECOM P U T E function);
• Prune (eliminate) goals which are impossible or extremely unlikely (as they deviate too much from the ideal
plan ig ), (captured by the P RU N E function).
A good RECOM P U T E heuristic reduces calls to the planner by avoiding unnecessary computation of m+
g for new observations. A good P RU N E heuristic reduces calls to the
planner by eliminating goals from being considered for future observations. Using appropriate heuristics in these functions, we can reduce the number of calls made to the planner
and consequently overall recognition run-time. This section
presents the algorithm. The next section will examine candidate heuristics.
Algorithm 1 begins (lines 3–5) by computing the ideal plan
ig for all goals, once. It also sets ig as a default plan suffix
m+
g . This suffix guarantees that valid (though not necessarily
optimal plans) mg can be created from m+
g , even in the ex+
treme case where no computation of mg is ever done. Then,
the main loop begins (line 6), iterating over observations as
they are made available. We then reach the first decision:
should we recompute the suffix m+
g (line 7).
We will begin by giving a general outline.
The
RECOM P U T E function takes the current winning trajectory mυ (υ is the current top-ranked goal) and the latest observation o. It matches the observation to mυ and heuristically
determines (see next section) whether o may cause a change
in the ranking of the top goal υ. If so, then the suffixes m+
g
of all goals (lines 8–12) will be recomputed (lines 11–12),
unless pruned (lines 9–10). Otherwise (lines 13–15) the current suffix m+
g of all goals will be modified based on o, but
without calling the planner.
Algorithm 1 O NLINE G OAL R ECOGNITION (R, planner)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
∀g : mg , m−
g ←∅
υ←∅
B the top-ranked goal
for all g ∈ G do
ig ← planner(I, g)
B default value for plan suffix
m+
g ← ig
while new o ∈ O available do
if RECOM P U T E(mυ , o) then
for all g ∈ G do
if P RU N E(m+
g , o, g) then
G ← G − {g}
else
m+
g ← planner(o, g)
else
for all g ∈ G do
+
prefix(o, m+
m+
g)
g ← mg
for all g ∈ G do
−
m−
g ← mg ⊕ o
−
mg ← mg ⊕ m+
g
for all g ∈ G do
P r(g|O) ← η · score(g)
υ ← argmaxg∈G P (g|O)
Recomputing m+
g . Here a straightforward call to the planner is made per the discussion in section 3.1, to generate an
optimal trajectory. From the initial point (the last point in o,
as o might contain more than a single point), to the goal g.
Modifying m+
g . When no recomputation of the suffix is
deemed necessary, o will be added to the prefix m−
g , and the
+
existing mg must be updated so that it continues m−
g and
leads towards g. The baseline algorithm calls a planner to do
this, but the point of this step is to approximate the planner
call so as to avoid its run-time cost. This is done by removing
(denoted by ) any parts that are inconsistent with respect to
the observation from the beginning of the old suffix m+
g .
−
The old m+
begins
where
the
old
m
(without
o)
ended.
g
g
should
ideally
begin
with
the
last
point
of the
The new m+
g
−
new mg (which is the new observation o), and continue as
much as possible with the old m+
g . Thus a prefix of the old
+
m+
,
denoted
(pref
ix(o,
m
),
line
15) is made redundant by
g
g
o and needs to be removed. If o is directly on m+
g , then
pref ix(o, m+
)
is
exactly
the
trajectory
from
the
beginning
g
point of m+
g to o. But in general, we cannot expect o to be
directly on m+
g . We thus define the ending point for the prefix
to be oe, the geometrically closest point to o on m+
g.
Pruning. Intuitively, when the newest observation o leads
away from a goal g, we may want to eliminate the goal from
being considered further, by permanently removing it from G.
This is a risky decision, as a mistake will cause the algorithm
to become unsound (will not return the correct result, even
given all the observations). On the other hand, a series of
correct decisions here can incrementally reduce G down to a
singleton (|G| = 1), which will mean that the number of calls
to the planner in the best case will approximate (|O| + 1).
Finally, when the algorithm reaches line 16, a valid suffix
m+
g is available for all goals in G. For all of them, it then
concatenates the latest observation to the prefix m−
g (line 17),
and creates a new plan mg by concatenating the prefix and
suffix (line 18). This means that a new score(g) can be used
to estimate P r(g|O) (lines 19–20), and a (potentially new)
top-ranked goal υ to be selected (line 21).
3.3
Recognition Heuristics
Algorithm 1 is a generalization of the algorithm described
in [Vered et al., 2016]. By varying the heuristic functions
used, we can specialize its behavior to be exactly the same
(Thm. 1), or change its behavior in different ways.
Theorem 1. If RECOM P U T E = > and P RU N E = ⊥
then Algorithm 1 will generate exactly the same number of
planner calls as the algorithm in [Vered et al., 2016].
Proof. (Sketch) By setting RECOM P U T E to be always
true, and P RU N E to be always false, initially a single call
to the planner will be made to calculate ig , and then a new
call to generate m+
g will be made for all goals and every observation (since no calls will be skipped and no goals would
be pruned). This is in accordance with the behavior of the
algorithm reported in [Vered et al., 2016].
Let us now turn to examine how to avoid unnecessary planner calls. An ideal scenario would be for Alg. 1 to never
compute a new suffix m+
g , for any goal. In line 5 of Alg. 1
the initial suffixes m+
g are set to the ideal plans, ig . If
RECOM P U T E is always false, then no new planner calls
will be made and m+
g will be incrementally modified (Alg. 1,
line 15) to accommodate the observations.
This approach offers significant savings (Thm. 2), and in
the best case, when the observations closely match the originally calculated paths, can produce good recognition results.
However, realistically, the observations may contain a certain
amount of noise or the observed agent may not be perfectly
rational. Moreover, it could be that the observed agent is perfectly rational and there is no noise in the observations and
yet the approach will fail. This is due to cases where there
are multiple perfectly-rational (optimal) plans, which differ
from each other but have the exact same optimal cost. In
such cases, it is possible that the planner used by the recognizer will generate an ideal plan ig which differs from an
equivalent—but different—ideal plan mg carried out by the
observed agent.
Theorem 2. If RECOM P U T E = ⊥ there will be exactly
|G| calls to the planner.
Proof. Straightforward, omitted for space.
RECOMPUTE. As we cannot realistically expect the
observations to perfectly match the predictions, we need a
heuristic that evaluates to false when the new observation o
does not alter the top ranked goal υ (saving a redundant |G|
calls to the planner), and evaluates to true otherwise.
A suggestion for such a heuristic in continuous domains is
to measure the shortest distance dist(o, mg ) between o and
all plans mg . If dist(o, mυ ) is shortest we can assume the
observed agent is still heading towards the same goal and do
not need to re-call the planner, keeping the current rankings.
PRUNE. Finally, we introduce a pruning heuristic for observing rational agents in continuous domains. It is inspired
by studies of human estimates of intentionality and intended
action [Bonchek-Dokow and Kaminka, 2014]. Such studies
have shown a strong bias on part of humans to prefer hypotheses that interpret motions as continuing in straight lines,
i.e., without deviations from, or corrections to, the heading of
movements. Once a rational agent is moving away or past a
goal point g, it is considered an unlikely target.
Figure 1: Illustration of goal angles used in pruning heuristic.
To capture this, the P RU N E heuristic takes a geometric
approach. We calculate αg , the angle created between the
end-point of the old m−
g , the newly received observation o
and the previously calculated plan mg . αg is calculated using
the cosine formula, cos(α) = (~u · ~v )/(||~u||||~v ||), where ~u is
the vector created by the previous and new observation and ~v ,
the vector created by the previously calculated plan and the
new observation.
Figure 1 presents an illustration of the heuristic approach
in 2D. For the new observation, o2 , we measure the angle αi
created by the new observation o2 , the previous observation
o1 (which ends m−
g ) and the previous plans m1,2 (shown as
the dashed lines). If the angle is bigger than a given threshold
we deduce that the previous path is heading in the wrong direction and prune the goal. By defining different sized threshold angles we can relax or strengthen the pruning process as
needed.
4
Evaluation
We empirically evaluated our online recognition approach
and the suggested heuristics over hundreds of goal recognition problems while measuring both the efficiency of the approach in terms of run-time and overall number of calls to
the planner, and the performance of the approach in terms
of convergence and correct ranking of the chosen goal. We
additionally implemented our approach on simulated ROSenabled robots while measuring the efficiency of the algorithm as compared to two separate approaches, one containing full knowledge of the observed agents’ intentions and the
other containing no knowledge and no reasoning mechanism.
4.1
Online Goal Recognition In A 3D Navigation
Domain
We implemented our approach and the proposed heuristics
to recognize the goals of navigation in 3D worlds. We used
TRRT (Transition-based Rapidly-exploring Random Trees),
an off-the-shelf planner that guarantees asymptotic nearoptimality , available as part of the Open Motion Planning
Library (OMPL [Şucan et al., 2012]) along with the OMPL
cubicles environment and default robot. Each call to the planner was given a time limit of 1 sec. The cost measure being
the length of the path. For the Pruning heuristic we used a
threshold angle of 120◦ .
We set 11 points spread through the cubicles environments.
We then generated two observed paths from each point to all
others, for a total of 110 × 2 goal recognition problems. The
observations were obtained by running an RRT* planner on
each pair of points, with a time limit of 5 minutes per run.
RRT* was chosen because it is an optimized planner that
guarantees asymptotic optimality. The longer the run-time
the more optimal the path. Each problem contained between
20-76 observed points.
Performance Measures We use two measures of recognition performance : (1) the time (measured by number of observations from the end) in which the recognizer converged to
the correct hypothesis (including 0 if it failed). Higher values
indicate earlier convergence and are therefore better; and (2)
the number of times they ranked the correct hypothesis at the
top (i.e., rank 1), which indicates their general accuracy. The
more frequently the recognizer ranked the correct hypothesis
at the top, the more reliable it is, hence a larger value is better.
Efficiency Measures In order to evaluate the overall efficiency of each approach we also used two separate measures:
(1) the number of times the planner was called within the
recognition process; and (2) the overall time (in sec.) spent
planning. Though these two parameters are closely linked,
they are not wholly dependant. While a reduction in overall number of calls to the planner will also necessarily result
in a reduction in planner run-time, the total amount of time
allowed for each planner run may vary according to the difficulty of the planning problem and therefore create considerable differences.
4.2
Effects of the different heuristic approaches
We ran the TRRT based recognizer on the above-mentioned
220 problems, comparing the different approaches. The results are displayed in Table 1, columns 1–4. Baseline, refers
to the algorithm of Vered et al. [2016], making (O + 1)|G|
planner calls. The second approach, No Recomp, refers to
the method of no recomputation at all, meaning the planner is
only utilized once in the beginning of the process, to calculate
ig , the ideal path, for all of the goals. The third approach, Recompute, measures the effect of RECOMPUTE which aims
to reduce overall number of calls to the planner (Section 3.2).
The fourth approach, Prune, measures the effect of PRUNE
which aims to reduce the overall number of goals by eliminating unlikely goal candidates (Section 3.2). And the last
approach, Both, measures the effects of utilizing a combination of both the Pruning and Recompute Heuristics.
Efficiency Table 1, column 1, displays the average of the
results of each approach as the mean of total planner run-time
measured in seconds. When only calling the planner once in
the recognition process the No Recompute approach takes an
average of only 7.1 sec.
Baseline has a time average of 105 sec.
The Pruning heuristic reduces the average time to only
74.9 sec. And the Recompute heuristic further reduces the
average time to 49.9 sec. When utilizing both heuristics we
achieved a reduction to 36.5 sec. an improvement of a substantial
65.25% from the Baseline approach.
The second column displays the average of the results in
terms of number of calls made by the recognizer to the planner. The No Recompute approach had an average of an extremely efficient 9 calls, i.e. the number of goals. Baseline
had an average of 265.05 calls while the Recomputation and
Pruning heuristics had similar success with a further reduction to 148.9 and 151.4 calls each. Using both heuristics the
number of calls was reduced to an average of only 90.6 calls,
a reduction of 63.3% from the Baseline approach.
In conclusion we see that employing the heuristics makes a
big impact on run-time and successfully reduces overall number of calls to the planner. While the recomputation heuristic outperformed the pruning heuristic, both in run-time and
overall number of calls, utilizing both heuristics can reduce
both run-time and number of calls made to the planner by over
60% from the baseline approach. The most efficient method
proved to be the No Recompute approach, only calculating
|G| plans. We will now show that this improvement in efficiency costs considerably in performance.
Performance Table 1, column 3, measures the average
convergence to the correct result percent, higher values are
better. As we can see with no reuse of the planner at all,
No Recompute only produces 6.7% convergence. As this approach does not make use of the incrementally revealed observations within the recognition process, any deviation from
the initially calculated path, ig , will have considerable impact
on recognition results.
By converting to the online Baseline algorithm, we were
able to more than double the convergence percent to 21.8%.
Each incremental observation was now taken into account
during the reuse of the planner and therefore had greater
weight on the ranking of the goals. Applying both the Pruning and Recomputation heuristics further improve the overall convergence. By eliminating goals the ranking process
now proved to be easier, as there were less goals to compare to. Furthermore, the early elimination of goals in the
pruning process was able to also eliminate the further noise
these goals might introduce to the ranking process, when their
paths deviated from the optimal. The Recomputation heuristic increases it to 25.4% and the Pruning to 42.2%, an improvement of 20.4% from the Baseline approach. When utilizing both heuristics we see that the high convergence level
obtained by the Pruning heuristic is maintained.
Column 4, measures the percent of times the correct goal
was ranked first. Here too a higher value is better and will
reflect on overall reliability of the ranking procedure. The
results mostly agree with the convergence results. With no
planner reuse at all, No Recompute, performs poorly with a
low 9.5%. Baseline more than doubles the success here as
well, to 20.2%. The Recomputation heuristic achieves 33.9%
and the Pruning heuristic increases the results to 40.5%, an
improvement of 20.3% from the Baseline approach. Again,
when applying both heuristics the success level of the Pruning method is obtained.
Employing the heuristics has made a big impact on overall
performance successfully increasing convergence and overall
correct rankings. The Pruning heuristic outperformed the Recomputation heuristic in both measures and a combination of
both heuristics maintains the high success rate leading to an
improvement of over 20% in both measures.
4.3
Sensitivity to recognition difficulty
In online, continuous domains, the hardness of the recognition problem could possibly effect recognizer performance
and efficiency. We wanted to evaluate the sensitivity of the
results shown above to the hardness of the recognition problems. We therefore added another 9 goal points (e.g., 19 potential goals in each recognition problem), for a total of 380
recognition problems. These extra points were specifically
added in close proximity to some of the preexisting 10 points,
such that navigating towards any one of them appears (to human eyes) to be just as possible as any other.
Table 1, columns 5–6, examines the efficiency of the different online recognition approaches over the harder clustered goals problems. We omitted the No Recompute heuristic in these instances as the behavior of this heuristic is very
straightforward. The results are consistent with the results
from the original scenario. The Baseline approach is the least
Baseline
Recompute
Prune
Both
No Recompute
10 goals
Efficiency
Run-Time
PlannerCalls
105.02
265.05
49.98
148.94
74.90
151.46
36.49
90.68
7.10
9.00
Performance
Conv.
Rank.
21.82
20.24
25.44
33.91
42.16
40.50
42.41
40.21
6.77
9.54
19 goals
Efficiency
Run-Time
PlannerCalls
194.65
516.57
126.75
397.85
160.29
386.53
97.63
287.36
-
Performance
Conv.
Rank.
16.37
19.54
18.7
22.76
23.18
24.03
20.98
25.82
-
Table 1: Comparison of all approaches across scattered and clustered goal scenarios.
efficient, having a higher run-time and larger number of calls
to the planner, than the rest. The most efficient approach is
still the approach of utilizing both the Pruning heuristic and
the Recompute heuristic together. In run-time the Recompute
heuristic is still more efficient than the Pruning however for
the measure of number of calls made to the planner we see
that, for more clustered goals scenarios, the Pruning heuristic
slightly outperforms the Recompute heuristic.
Table 1, columns 7–8, examines the performance of the
different online recognition approaches over the harder clustered goals problems. For the harder problems the best performance achieved, in terms of convergence, was by the Pruning heuristic with a convergence of 23.18% from the end. In
terms of the amount of times the correct goal was ranked first
the Both approach, combining both Pruning and Recompute
heuristics, only slightly outperformed the Pruning approach.
The worst performance was achieved by the Baseline approach, in terms of both criteria measured; convergence and
ranked first, in congruence with the performance results of
the scattered goal scenario.
Table 2 measures the deterioration in efficiency and performance with comparison to the scattered goal scenario. The
deterioration is measured in terms of deterioration percent,
hence a 100% deterioration in run-time means the planner
took twice as long on average, on the harder problems. Therefore lower values are better. In terms of efficiency, we can
clearly see that the least deterioration, both in run-time and
number of calls to the planner, occurred for the Baseline approach proving this approach to be very reliable with a deterioration of 85.35% and 94.90% respectively. The biggest
deterioration in terms of run-time occurred for the combination of both heuristics with a deterioration of 167.58%. This
was considerably caused by the substantial deterioration of
the Recompute approach which deteriorated by 153.58%. The
Pruning heuristic deteriorated considerable less in terms of
run-time with only 114% deterioration.
In terms of number of calls made to the planner, again,
the worst deterioration occurred for the Both approach, with
a deterioration of 216.9% while the deterioration for each of
the heuristics was considerably less; 155.2% for the Pruning
heuristic and 153.6% for the Recomputation heuristic.
In terms of performance deterioration we again see that the
most resilient approach in terms of performance, as well as
efficiency, proved to be the Baseline both in terms of Convergence and Ranked first with a deterioration of 25.98% in
convergence and only 3.46% in ranked first. The biggest deterioration in convergence occurred for the Both approach, as
was in the efficiency results. However, in terms of ranked first
the biggest deterioration occurred for the Pruning heuristic.
This was, in part, due to the fact that clustered goals make the
pruning process considerably less efficient as the goals are
too close to be pruned.
4.4
Online Goal Recognition on Robots
As a final set of experiments, and to show the applicability of
our approach, we implemented Alg. 1 in a cooperative robotic
team task. We used ROS [Quigley et al., 2009] to control simulated robots in Gazebo, using the default ROS motion planner, move_base, in the recognition process. We simulated
a soccer field, with two robots operating as members of the
same team (Figure 2). The observed robot was given an initial
goal to travel to, proceeding to execute the plan in a straightforward manner, and the observing robot had to strategically
place itself in a pre-chosen position to assist the other robot
team member. If the observed robot navigated to goal 4 the
strategic place to assist it on the offense would be to navigate
to goal 3 and vice versa. Likewise also with goals 1 and 2.
The observed robot always started at the same initial point
in the middle of the field, while we experimented with 3 different starting points for the observing robot; two points behind the observed robots position and on parallel sides (Figure 2, init points 1 and 2) and one point past the observed
robot in the middle of the field (init point 3). We ran 10–20
runs from each initial position to each of the goals for a total
of 193 problems.
We compared our online goal recognizer (OGR) in its baseline form, to two different approaches: (a) giving full knowledge (FK) of the intended goal to the observing robot, ahead
of time, allowing the observing robot to navigate directly towards it; and (b) giving no (zero) knowledge (ZK) of intended
goal, thus forcing the observing robot to wait for its team
member to reach its desired goal, before it can navigate towards the complementary location.
To evaluate the different approaches we measured the overall time (in seconds) the simulated robot ran until reaching its
target goal. The lower the time the more efficient the robot.
The results are displayed in Table 3.
The results show that the goal recognition approach substantially improves on the zero knowledge approach, while
requiring no precalculations; all needed plans are generated
on-the-fly via the planner. Understandably our approach falls
short of the full knowledge approach as it generates hypotheses on the fly following observations, which leads to some
deviations from the optimal, direct route.
5
Summary
We presented an efficient, heuristic, online goal recognition
approach which utilizes a planner in the recognition process
Baseline
Recompute
Prune
Both
Deterioration
Efficiency
Performance
Run-Time
PlannerCalls
Conv.
Rank.
85.35%
94.90%
24.98%
3.46%
153.58%
167.11%
26.49%
32.88%
114.01%
155.20%
45.02%
40.67%
167.58%
216.90%
50.53%
35.79%
Table 2: Deterioration of performance and efficiency between scattered and clustered goal scenarios.
Figure 2: Experiment setup (via RVIZ)
I1
I2
I3
G1
G2
G3
G4
G1
G2
G3
G4
G1
G2
G3
G4
FK
10.00
5.80
9.10
5.80
5.80
9.10
10.00
12.35
12.41
10.10
9.24
5.72
OGR
17.35
12.31
16.19
17.10
14.09
19.56
15.89
15.32
17.36
18.65
10.62
13.40
ZK
21.50
17.18
20.43
26.03
17.67
24.45
25.45
32.62
20.88
24.26
20.66
20.40
Table 3: Online goal recognizer vs. full and zero knowledge
to generate recognition hypotheses. We identified key decision points which effect both overall run-time and the number of calls made to the planner and introduced a generic
online goal recognition algorithm along with two heuristics
to improve planner performance and efficiency in navigation
goal recognition. We evaluated the approach in a challenging
navigational goals domain over hundreds of experiments and
varying levels of problem complexity. The results demonstrate the power of our proposed heuristics and show that,
while powerful by themselves, a combination of them leads
to a reduction of a substantial 63% of the calls the recognizer
makes to the planner and planner run-time in comparison with
previous work. This, while showing an increase of over 20%
in recognition measures. We further demonstrated the algorithm in a realistic simulation of a simple robotic team task,
and showed that it is capable of recognizing goals using standard robotics motion planners.
References
[Baker et al., 2005] Chris Baker, Rebecca Saxe, and
Joshua B Tenenbaum.
Bayesian models of human
action understanding. In Advances in neural information
processing systems, pages 99–106, 2005.
[Baker et al., 2007] Chris L Baker, Joshua B Tenenbaum,
and Rebecca R Saxe. Goal inference as inverse planning.
In Proceedings of the 29th Annual Meeting of the Cognitive Science Society, 2007.
[Blaylock and Allen, 2006] Nate Blaylock and James Allen.
Fast hierarchical goal schema recognition. In Proceedings
of the Twenty-First National Conference on Artificial Intelligence (AAAI-06), pages 796–801, 2006.
[Bonchek-Dokow and Kaminka, 2014] Elisheva BonchekDokow and Gal A Kaminka. Towards computational
models of intention detection and intention prediction.
Cognitive Systems Research, 28:44–79, 2014.
[Geib, 2015] Christopher Geib. Lexicalized reasoning. In
Proceedings of the Third Annual Conference on Advances
in Cognitive Systems, 2015.
[Hong, 2001] Jun Hong. Goal recognition through goal
graph analysis. Journal of Artificial Intelligence Research,
15:1–30, 2001.
[Keren et al., 2015] Sarah Keren, Avigdor Gal, and Erez
Karpas. Goal recognition design for non-optimal agents.
pages 3298–3304, 2015.
[LaValle, 2006] Steven M. LaValle. Planning Algorithms.
Cambridge University Press, 2006.
[Lesh and Etzioni, 1995] Neal Lesh and Oren Etzioni. A
sound and fast goal recognizer. In Proceedings of the
International Joint Conference on Artificial Intelligence
(IJCAI-95), 1995.
[Martin et al., 2015] Yolanda E. Martin, Maria D. R.
Moreno, David E Smith, et al. A fast goal recognition technique based on interaction estimates. In Twenty-Fourth
International Joint Conference on Artificial Intelligence,
pages 761–768, 2015.
[Masters and Sardina, 2017] Peta Masters and Sebastian
Sardina. Cost-based goal recognition for path-planning.
In Proceedings of the 16th Conference on Autonomous
Agents and MultiAgent Systems, pages 750–758. International Foundation for Autonomous Agents and Multiagent
Systems, 2017.
[Pereira et al., 2017] Ramon Fraga Pereira, Nir Oren, and
Felipe Meneguzzi. Landmark-based heuristics for goal
recognition. 2017.
[Quigley et al., 2009] Morgan Quigley, Ken Conley, Brian
Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob
Wheeler, and Andrew Y Ng. Ros: an open-source robot
operating system. In ICRA workshop on open source software, volume 3, page 5. Kobe, Japan, 2009.
[Ramırez and Geffner, 2009] Miquel Ramırez and Hector
Geffner. Plan recognition as planning. In International
Joint Conference on Artifical Intelligence, pages 1778–
1783, 2009.
[Ramırez and Geffner, 2010] Miquel Ramırez and Hector
Geffner. Probabilistic plan recognition using off-the-shelf
classical planners. In International Joint Conference on
Artificial Intelligence, 2010.
[Ramırez and Geffner, 2011] Miquel Ramırez and Hector
Geffner. Goal recognition over POMDPs: Inferring the
intention of a pomdp agent. In Proceedings of the International Joint Conference on Artificial Intelligence, pages
2009–2014, 2011.
[Sadeghipour and Kopp, 2011] Amir Sadeghipour and Stefan Kopp. Embodied gesture processing: Motor-based
integration of perception and action in social artificial
agents. Cognitive Computation, 3(3):419–435, 2011.
[Sohrabi et al., 2016] Shirin Sohrabi, Anton V. Riabov, and
Octavian Udrea. Plan recognition as planning revisited.
The Twenty-Fifth International Joint Conference on Artificial Intelligence, pages 3258–3264, 2016.
[Şucan et al., 2012] Ioan A. Şucan, Mark Moll, and Lydia E.
Kavraki. The Open Motion Planning Library. IEEE
Robotics & Automation Magazine, 19(4):72–82, December 2012.
[Sukthankar et al., 2014] Gita Sukthankar, Robert P. Goldman, Christopher Geib, David V. Pynadath, and Hung Bui,
editors. Plan, Activity, and Intent Recognition. Morgan
Kaufmann, 2014.
[Vered et al., 2016] Mor Vered, Gal A. Kaminka, and Sivan
Biham. Online goal recognition through mirroring: Humans and agents. In Proceedings of the Annual Conference on Advances in Cognitive Systems, 2016. A slightly
modified version appears in Proceedings of the IJCAI 2016
workshop on Human-Agent Interaction Design and Models (HAIDM).
| 2cs.AI
|
Tropical Land Use Land Cover Mapping in Pará (Brazil) using
Discriminative Markov Random Fields and Multi-temporal TerraSAR-X
Data
Ron Hagensiekera,∗, Ribana Roscherb , Johannes Rosentretera , Benjamin Jakimowc , Björn Waskea
a Freie
Universität Berlin, Institute of Geographical Sciences, Malteserstr. 74-100, 12249 Berlin, Germany
Friedrich-Wilhelms-Universität Bonn, Institute of Geodesy and Geoinformation, Nussallee 15, 53115 Bonn,
Germany
c Humboldt-Universität zu Berlin, Geography Department, Unter den Linden 6, 10099 Berlin, Germany
arXiv:1709.07794v1 [cs.CV] 22 Sep 2017
b Rheinische
Abstract
Remote sensing satellite data offer the unique possibility to map land use land cover transformations by
providing spatially explicit information. However, detection of short-term processes and land use patterns
of high spatial-temporal variability is a challenging task.
We present a novel framework using multi-temporal TerraSAR-X data and machine learning techniques,
namely Discriminative Markov Random Fields with spatio-temporal priors, and Import Vector Machines, in
order to advance the mapping of land cover characterized by short-term changes. Our study region covers
a current deforestation frontier in the Brazilian state Pará with land cover dominated by primary forests,
different types of pasture land and secondary vegetation, and land use dominated by short-term processes
such as slash-and-burn activities. The data set comprises multi-temporal TerraSAR-X imagery acquired
over the course of the 2014 dry season, as well as optical data (RapidEye, Landsat) for reference. Results
show that land use land cover is reliably mapped, resulting in spatially adjusted overall accuracies of up to
79% in a five class setting, yet limitations for the differentiation of different pasture types remain.
The proposed method is applicable on multi-temporal data sets, and constitutes a feasible approach to map
land use land cover in regions that are affected by high-frequent temporal changes.
Keywords: Markov Random Fields (MRF), Import Vector Machines (IVM), Multi-temporal LULC
mapping, Deforestation, Amazon, SAR
∗ Corresponding
author. Tel.: (+49 30) 838- 458 683
Email addresses: [email protected] (Ron Hagensieker), [email protected] (Ribana Roscher),
[email protected] (Johannes Rosentreter), [email protected] (Benjamin Jakimow),
[email protected] (Björn Waske)
Preprint submitted to International Journal of Applied Earth Observation and Geoinformation
September 25, 2017
1. Introduction
The Brazilian Amazon is the largest area of tropical rain forest shared by a single country. In the
last decades it has become increasingly threatened by large scale deforestation, forest degradation, and the
expansion of agriculture (Davidson et al., 2012; Lapola et al., 2014). They affect the Earth’s ecosystems and
ecosystem services far beyond the boundaries of the original region, and can influence the climate directly
at local and even regional scales (Foley, 2005; Vitousek, 1997). Thus, detailed knowledge and information
on land use and land cover (LULC) offers valuable input for decision support and environmental monitoring
systems.
Remote sensing satellite data offers the unique possibility to generate consistent LULC maps over large
areas at a temporally high resolution. Mapping of LULC change in the Amazon is predominantly achieved
by analyzing multi-spectral remote sensing data (INPE, 2015; Wulder et al., 2012; Hansen et al., 2013).
However, a limitation of the analysis of multi-spectral remote sensing data is imposed by its dependency
on cloud-free conditions. These are rare in tropical regions and in general not met during wet season (e.g.
Rufin et al., 2015; Müller et al., 2015). Synthetic aperture radar (SAR) data can overcome these problems
and various studies demonstrate the potential for mapping LULC and their changes (Pfeifer et al., 2016; Qi
et al., 2012; Bovolo and Bruzzone, 2005), also in the context of deforestation and related processes (Sarker
et al., 2013; Reiche et al., 2015; Englhart et al., 2011; Almeida-Filho et al., 2009). Such mapping approaches
become even more attractive due to recent missions with increased repetition rates, higher spatial resolution
(e.g. TerraSAR-X and Sentinel-1), as well as better data availability, e.g., by virtue of the Copernicus data
policy (Aschbacher and Milagro-Pérez, 2012). TerraSAR-X and the Sentinel-1 constellation guarantee cloud
free coverage within 11 and 6 days respectively, while the repetition rate of the Sentinel-2 constellation (5
days) and Landsat-8 (16 days) might be affected by clouds.
Although the classification accuracy of SAR data can be limited in direct comparison to multi-spectral
data, various approaches exist to increase the mapping accuracy. These include the integration of one-pass
interferometry (Schlund et al., 2013), contextual spatial information derived from texture parameters or
segmentation (Cutler et al., 2012; Sarker et al., 2013; Schlund et al., 2013; Waske and van der Linden,
2008), or the utilization of multi-temporal or multi-sensoral data (Reiche et al., 2013; Stefanski et al.,
2014a; Waske and Braun, 2009). Although limitations of short wavelength SAR data for the classification
of dense vegetation are well documented (e.g. Kumar and Patnaik, 2013), various studies have highlighted
the potentials of this data for LULC mapping (e.g. Schlund et al., 2013; Qi et al., 2015; Uhlmann and
Kiranyaz, 2014; Qi et al., 2012; Khatami et al., 2016; Sonobe et al., 2014), e.g. by utilization of multi2
temporal data, modern classification algorithms, or spatial context. Multi-temporal data sets are generally
more adequate when classes can be characterized by clearly defined temporal signatures, e.g. caused by
differences in the phenology of crops, land use management, or seasonal cycles (Blaes et al., 2005; McNairn
et al., 2009). While the single classification of a multitemporal data set might be useful for study sites
without or long-term changes (Waske and Braun, 2009; Stefanski et al., 2014b), it might be limited for
study sites with temporally high-frequent changes in land cover, e.g. slash-and-burn activities, at arbitrary
points in time. Recent studies have shown great potentials to tackle these problems by time series analysis
of multispectral data (Zhu and Woodcock, 2014), but SAR speckle and quick succession processes still pose
difficult challenges using such methods, especially if very long time series are often not available.
In the context of multi-temporal data analysis, a main drawback is often the assumption of non changing
land cover during the investigation period. Consequently, temporally dynamic LULC, such as slash-and-burn
activities or transitions between clean and shrubby pasture, are neglected. Various studies emphasize the
usage of an adequate classification approach to ensure a high mapping accuracy (Liu et al., 2006; Waske and
Benediktsson, 2007; Waske and Braun, 2009). Especially the integration of spatial information by means of
region-based classification or spatial features such as texture lead to a gain in accuracy. In addition, Markov
Random Fields (MRFs) are a promising approach to integrate spatial context (Moser et al., 2013; Moser
and Serpico, 2013; Liu et al., 2006). MRFs are employed to model prior knowledge about neighborhood
relations within the image, called spatial relations, but can also be extended to describe relations of the
same area but at different acquisition dates (temporal relations). Since the early 1990s, approaches based
on MRFs have been utilized in remote sensing for various purposes (Bouman and Shapiro, 1994; Xie et al.,
2002; Tran et al., 2005; Solberg et al., 1996). Liu et al. (2008) use locally variant transition models to
account for spatial heterogeneity and have applied the model on subsets of two Landsat scenes from 1990
and 2001. More recently, Wehmann and Liu (2015) have adapted an integrated kernel as proposed by Moser
and Serpico (2013), and used Iterative Conditional Modes (ICM) as optimization technique with spatiallyvariant transitions for classifying Landsat data. Hoberg et al. (2015) apply multi-temporal Conditional
Random Fields to regularize annual remote sensing imagery from different high resolution scales (IKONOS,
RapidEye, Landsat) over the course of five years.
With the emergence of efficient probabilistic classifiers over the last decade, standard MRFs have been
extended to discriminative MRFs (Kumar and Hebert, 2003), and turn out to be increasingly useful to
optimize land cover classifications (Moser and Serpico, 2010; Tarabalka et al., 2010; Voisin et al., 2013). Liu
et al. (2006) highlight the advantages of utilizing non-parametric, probabilistic Support Vector Machines
3
(SVMs, Platt (1999)) over a maximum likelihood classifier. However, although many remote sensing studies
highlight the positive capabilities of MRFs, only few studies aim on using MRFs for landscape-scale mapping
with multi-temporal data sets (e.g. Cai et al., 2014; Wehmann and Liu, 2015; Olding et al., 2015), for example,
to map forest cover change (Liu et al., 2008, 2006). If multi-temporal data sets are available, MRFs can also
be used to optimize the corresponding maps by considering predefined spatial-temporal inter-dependencies
between neighboring pixels, which are stored in transition matrices.
We present a novel framework for classification of a TS-X time series using discriminative MRFs and Import
Vector Machine (IVM), a probabilistic, discriminative, non-parametric classifier. Each scene is separately
classified using IVM, afterwards MRFs are utilized in an independent step to post-regularize the classification
map. We chose IVMs over commonly used probabilistic SVMs, since they have proven to offer a more reliable
probabilistic output (Zhu and Hastie, 2005; Roscher et al., 2012a,b). For MRF optimization we choose Loopy
Belief Propagation (LBP) over ICM as this method has been shown repeatedly to yield higher accuracies
(Szeliski et al., 2006; Andres et al., 2010a). Few studies have utilized LBP in the field of remote sensing (Li
et al., 2013), and as a novelty we integrate LBP into a multi-temporal setting.
The presented framework aims on the classification of each individual acquisition, and thus enables mapping
of high frequency spatial-temporal LULC patterns. In contrast to related studies, we use a multi-temporal
MRF model on SAR data to detect short-term transitions within one season and Loopy Belief Propagation
(LBP) for inference.
The overall goal of this research is focused on two objectives: (i) to map LULC in a tropical setting with
short-term processes, by adapting recent MRF methods, and (ii) to assess the potential for LULC mapping
using time-series image data of short wavelength SAR. The specific objective is to map LULC in Pará, Brazil,
where transformations of forest to pasture are the major driver of deforestation. Pasture management in the
study region tends to fall into one of two categories: long-term processes of intensively managed pasture land
(pasto limpo), or short-term processes of episodically managed pasture land with a high degree of successive
dynamics (pasto sujo). Pasture management in general is characterized by slash-and-burn processes resulting
in sudden changes in LULC.
2. Study Area & Data
2.1. Study Area
The study area lies in the Northern part of the Novo Progresso municipality (southern Pará state,
Brazil), and is intersected by the BR-163 highway in the Southwest 1. The BR-163 is accompanied by
4
Figure 1: Composite of three TerraSAR-X acquisitions (Red: VV June 8, Green: HH September 4, Blue: VV November 9 of
2014). True color ETM (L8, 12 September 2014) in the background shows the diverse LULC properties.
5
Table 1: Scenes utilized in this study. All scenes were collected over the same area using the same incidence angle.
Date
2014-06-08
2014-06-30
2014-07-22
2014-08-24
2014-09-04
Polarization
VV-VH
HH-HV
VV-VH
VV-VH
HH-HV
fishbone structures indicative of deforestation (Ahmed et al., 2013; Coy and Klingler, 2014). A major driver
of deforestation in the study area is the transformation of forests into pasture land. The climate in the study
region is characterized by a wet and a dry season. While the dry season, between June and September, sees
abrupt land cover changes in the form of large scale burning and clear cuts, the wet season is defined by
gradual regrowth, yet deforestation rates over the wet season are on the rise.
Figure 2: Photograph illustrating the fluent transitions and interactions of different land cover types in the study region.
6
2.2. Remote Sensing Data
The data base for the study consists of five TS-X strip map scenes with 5m × 5m spatial resolution
(Table 1). All images are ordered in single-look complex format, comprising different VV-VH and HH-HV
polarization at an incidence angle of 37.75◦ , and cover a swath of roughly 50km × 15km (5663 × 11856
pixels). Data is calibrated and processed according to common procedures (see Section 2.2). Preprocessing
in the context of this study includes all necessary steps before random sampling of training and test data
is performed. After random sampling, training of IVM, and MRF regularization are taken out, land cover
maps are generated and validated, and average measures are calculated.
Preprocessing of the TS-X scenes is conducted using the Sentinel 1-Toolbox and the Geospatial Data Abstraction Library (GDAL Development Team, 2015). All scenes were processed separately in the following
order:
Multilooking: 3 range looks, 2 azimuth looks, yielding a ground resolution of ∼ 4.7m × 4.7m.
Terrain & Radiometric Correction: Range-Doppler terrain correction (SRTM 3Sec) and resampling to
5m × 5m pixel spacing. The data is projected into UTM Zone 21S. γ 0 radiometric normalization is
applied using an SRTM.
GLCM-Texture Texture measurements are widely used to increase the mapping accuracy of SAR data
(Sarker et al., 2012; Dekker, 2003; Cutler et al., 2012). Gray Level Co-occurence Matrices (GLCM) are
calculated with 11 × 11 moving window size into all symmetric directions with offset one. Probabilistic
quantization is conducted into 64 levels. Ten texture parameters are separately derived for any available
polarization and any available scene: contrast, dissimilarity, homogeneity, angular second moment
(ASM), energy, maximum probability, entropy, GLCM mean, GLCM variance, and GLCM correlation.
These will be included as additional features to improve the IVM classification. For more information
on GLCM-based texture parameters see Haralick et al. (1973); Sarker et al. (2012). In respect to
findings by Sarker et al. (2013); Nyoungui et al. (2002), and our own experiments, we abstain from
combining texture metrics with speckle filtering. Since we use 10 texture measures per layer, we have
a total of 22 features per scene for the classification process.
2.3. Reference Data
Reference data includes multispectral RapidEye and Landsat data, in situ data, as well as land cover
data from various Brazilian agencies (e.g. PRODES, TerraClass). PRODES (Programa de Cálculo do
7
Table 2: Number of sample points available for training distinguished by class, extracted from polygons.
Date
Class
Burnt Pasture
Clean Pasture
Shrubby Pasture
Water
Forest
06-08
06-30
07-22
08-24
09-04
15
487
749
28
799
55
444
783
28
701
139
375
789
28
704
469
384
438
28
616
434
420
428
28
516
Desflorestamento da Amazônia) is an effort by the Brazilian space agency (INPE) to generate annual maps
documenting deforestation of primary forests inside the Legal Amazon with a minimum mapping unit of
6.25ha (INPE, 2015). Targeting only the sites that PRODES considers deforested, TerraClass is an effort
to determine LULC classes of the affected areas (Almeida et al., 2016). The overall coverage of all available
TS-X swaths constitutes the study area (Figure 1), and is sufficiently covered by reference information.
While forests as well as clean and shrubby pasture are present in the study area, occurrence of water and
burnt pasture is overall scarce. To address this issue, polygons are manually distributed over the entire area.
Afterwards, each polygon is assigned one class label for each date covered by TS-X to address changes of
LULC. If necessary, polygons are split to avoid class ambiguity within different temporal instances. E.g., if
a coherent pasture area is only partially burnt, the polygon gets split. The generation of reference data is
supported by visual interpretation of RapidEye as well as Landsat 7 ETM+ and Landsat 8 OLI data of the
same time period. In addition to Landsat and RapidEye imagery, fire products derived from MODIS are
also considered. Moreover, photographs from a field campaign conducted in August 2014 are available.
Sampling is conducted by two of the authors in close cooperation and was harmonized with classification
schemes by INPE (Instituto Nacional de Pesquisas Espaciais). The following LULC classes are considered:
clean pasture, shrubby pasture, burnt pasture, water, and forest. Clean pasture, also called
pasto limpo, describes pasture land that is intensively worked. This includes regular tillage and burning of
land to support cattle ranching. Shrubby pasture, also called pasto sujo, is not intensively managed and
thus affected by bush encroachment. The coarser appearence of shrubby pasture generally allows a visual
separation from clean pasture in high resolution images. Burnt pasture includes clean as well as shrubby
pasture areas which were recently burned, and are characterized by open soil and vegetation residues. Such
areas can be easily identified using false color composites. Forest, beside primary forests, includes areas
of secondary vegetation and regeneration as these are usually non-separable by X-band SAR. Forests have
a very characteristic appearance in TS-X images and high resolution multispectral imagery. Table 2 gives
8
Figure 3: Classification scheme.
In-Situ
Multispectral (RE, LS)
Water
Forest
Pasto Burnt
Pasto Limpo
Pasto Sujo
Class
9
an overview of the number of available training samples for each class and date. It should be underlined
that the burning season usually starts around end of July. Hence only few burned pasture areas could be
identified before that period. Water bodies are also very scarce and only two lakes over the entire study area
are included. Figure 3 visualizes the classes considered in our classification scheme. The considered LULC
classes match comparable studies using TS-X data in Brazilian, or tropical settings, respectively (Garcia
et al., 2011; Schlund et al., 2013). As the time period of our study falls into the dry season between June
and September, corresponding multispectral remote sensing data could be interpreted sufficiently well. Yet,
some challenges remain:
• For the study region two dominating pasture types can be identified: pasto sujo, i.e., shrubby pasture,
and pasto limpo, i.e., clean pasture (Almeida and Vieira, 2008; Adami et al., 2015). While both
types are generally used for cattle ranching in this region, pasto sujo is characterized by bushes and
occasional early stages of succession. However, the transition between these types can be gradual and
consequently hard to interpret from remote sensing imagery alone; even at 5m ground resolution as
offered by RapidEye.
• Transitions from pasto sujo into early stages of secondary vegetation are hard to distinguish, due to the
gradual nature of this process. However, it is not as relevant in our study site since significantly less
areas are affected. To allow a solid separation of these classes, we consider multi-annual time series to
identify pasture management. In addition, we include information offered by TerraClass which reliably
separate different types of secondary vegetation and pasture land.
• Primary and secondary forests, as well as secondary vegetation, are combined into one class, as various
studies and own preliminary tests indicate the limitation of X-band for separating these two classes.
3. Methods
The proposed framework consists of four steps: (i) preprocessing, (ii) random sampling, (iii) classification
of each single scene using IVM, and (iv) optimization of the MRF model. Final validation is performed on
averages over 10 independent runs, using a random sampling (50:50) into spatially disjoint train and test
polygons. As pixels sampled from training polygons are solely used for IVM parameterization (grid search)
and model training, pixels sampled from test polygons enable an independent validation.
Throughout the paper we use the following notation: Let there be a training set (xn , yn ) ∈ T comprising
N feature vectors {x1 , . . . , xn , . . . , xN } and corresponding class labels yn ∈ {1, ..., K}, distributed over an
10
t+1
t
t−1
Figure 4: Temporal (green) and spatial (yellow) neighbors of a given pixel (blue).
image lattice I . We later address image samples at any given coordinate as xi , and probability estimates as
pi = [p1 , ..., pnk , ..., pnK ] with pnk = p(yn = k | xn ).
3.1. Import Vector Machines
IVM is a discriminative and probabilistic classifier based on kernel logistic regression and has first been
introduced by Zhu and Hastie (2005). Roscher et al. (2012b) have shown that IVMs provide more reliable
probabilities than probabilistic SVMs, since IVMs’ probabilities are more balanced, whereas SVMs generally
overestimate maximum probabilities. To account for complex decision boundaries between classes, IVM
generally benefit from integrating a kernel function. For this study, we utilize the radial basis function (rbf)
kernel parameterized by kernel width σ, which is a standard for remote sensing purposes. Parameterization
is achieved analogously to standard SVM practices using a grid search, to estimate the cost parameter C
and σ. For a more encompassing description of IVM see Zhu and Hastie (2005) and Roscher et al. (2012a).
3.2. Markov Random Field
In this study, we use post classification MRF with spatio-temporal neighborhood relations between
pixels, as illustrated in Figure 4. Parameterization is achieved through transition matrices, which are 5 × 5
matrices indicating spatial and temporal transition probabilities between the five classes. For our description
of MRF, we adapt a terminology similar to Moser et al. (2013); Melgani and Serpico (2003). Therefore, with
xi denoting pixel features and yi its corresponding class label, we reformulate the IVM-based probabilities
p(yi | xi ) as energy terms
UX = −
X
ln p(yi | xi ).
(1)
i∈I
As the energy UX are equivalent, minimization of UX is identical to maximization of P . Now consider
a function for the spatial neighborhood Usp , with i ∼sp j applying to any two pixels, which are direct,
11
4-connected, spatial neighbors, and a function δ(yi , yj ) to assign weights to neighboring classes:
Usp =
X
1 − δ(yi , yj )
(2)
i∼s j
In this case the function δ yields a K×K matrix which can be used to favor certain neighboring constellations.
The function δ is generally defined as Potts model to result in an identity matrix, which encourages the
generation of homogeneous areas. The standard mono-temporal MRF model is given by summation of (1)
and (2)
UX ,sp = −
X
ln p(yi | xi ) + β
X
1 − δ(yi , yj ) ,
(3)
i∼s j
i∈I
with weight parameter β to regulate importance of the spatial component. For the multi-temporal case
we consider co-registered images with the temporal neighbors being the spatially congruent cells at the
neighboring acquisition times t − 1 and t + 1. Only if pixel j is the temporal successor of i, i ∼t+1 j applies;
and only if pixel j is the temporal predecessor of i, i ∼t−1 j applies. Temporal energy is hence given by (4),
analogous to the spatial case.
Utemp =
X
1 − τ1 (yi , yj ) +
i∼t+1 j
X
1 − τ2 (yi , yj )
(4)
i∼t−1 j
Here, τ1 and τ2 are K × K matrices defining the temporal transitions as observed from land cover trajectories. In opposition to the spatial weighting δ, we require multiple, non symmetrical matrices τ to respect
trajectories with regard to the future, or the past. The overall energy function is defined by integrating the
temporal vicinity into (3), which yields
U =−
X
+βtemp
1 − δ(yi , yj )
i∼s j
i∈I
X
X
ln p(yi | xi ) + βsp
1 − τ1 (yi , yj ) +
i∼t+1 j
X
1 − τ2 (yi , yj ) .
(5)
i∼t−1 j
This function combines (3) and (4). Weight parameters β(·) can be used to adjust the importance of temporal
and spatial weights.
3.3. Passing scheme & transition matrices
LBP is an inference algorithm utilizing Message Passing (Pearl, 1982), and is shown to approximate
maximum values sufficiently well (Murphy et al., 1999). We choose LBP over graph-cut based methods for
12
Figure 5: Passing schedule as applied in this study. One pass over all layers corresponds to one iteration of LBP.
Action
Description
Step 1: Generation of a fallback copy of the
current energy layer (blue). Necessary for future
calculation of messages passed to the red layer.
Step 2: Messages are passed from the previous
fallback, the next energy layer, and the current
layer, also factoring in the unaries upon receiving. This procedure is performed using a set of
moving windows.
Step 3: Discarding of the previous fallback, and
backing up of the next energy layer (see Step 1).
Iterate over each layer.
their more general applicability, as graph-cuts are specifically defined for symmetrical binary factors (Boykov
et al., 2001), and can not be applied in non-symmetric environments (Kolmogorov and Zabin, 2004). ICM
(Iterated Conditional Modes) is another algorithm which is commonly used to achieve inference, especially
in remote sensing and using multi-temporal data sets (Liu et al., 2008, 2006; Wehmann and Liu, 2015).
While it has low computational cost, it is generally outperformed by LBP in terms of accuracy (Szeliski
et al., 2006; Andres et al., 2010b). For this reason we formulate an implementation of LBP using moving
windows, which can be applied to image stacks of arbitrarily large image stacks sufficiently well. Figure 6
illustrates the neighborhood of one pixel in a factor graph, analogous to the MRF neighborhood as described
in Section 3.2. Using the Potts function to define δ is common practice in remote sensing literature (Moser
et al., 2013), and since the focus of this study lies on the examination of MRF for a multi-temporal linking of
classifications, we follow this practice. The Potts function can be represented by an identity matrix, which
supports assignment of neighboring pixels to the same class. It is in general not sensible to formulate an
asymmetric message passing for the two spatial dimensions, as a pixel will assume the same properties of
its left as of its right neighbor. More specifically, the Potts model is a way to reflect Tobler’s assumption on
13
τ2
δ
I
τ1
δ
δ
τ1
δ
τ2
Figure 6: Factor-graph as implemented in this study. Variable nodes illustrated by yellow (spatial neighbor) and green nodes
(temporal). Circles mark the corresponding factor nodes and the unary IVM-based energy.
spatial autocorrelation, promoting the idea of close objects to be more alike than distant objects.
In contrast to the spatial transitions, utilization of the Potts model for temporal transitions can lead to
serious distortions and cause equalization of all subjected probability maps, i.e., it would prohibit any land
cover changes from occuring. While we can not assume any spatially directional patterns in the area, and
thus rely on the Potts model, a pixel will pass different messages to its temporal successor opposed to its
predecessor. Through adjustment of τ it is possible to assign probabilities for all possible types of class
transitions. We therefore express temporal transitions through two asymmetric transition matrices (τ1 and
τ2 ). While matrix τ1 illustrates the messages pixels pass from the scene at time t to its neighbor at t − 1,
τ2 defines the messages the pixels pass from t to t + 1. This differentiation is important, e.g. considering
that burnt pasture at t will prohibit primary forests at t + 1, yet it might endorse pasture at t − 1. We
integrate the user-defined transition matrices τ1 and τ2 as interface to inject a-priori expert knowledge into
our regularization model. While not empirically derived, these matrices are based on weak assumptions on
land cover trajectories.
The assumptions include that Water and Forest are regarded very consistent classes, yet that pasture
areas have some kind of interaction with each other. This especially concerns the transition of pasture land
to burnt pasture land, which is explicitly tolerated. Furthermore, forest does not explicitly prohibit
predecessing non-forest areas, which is due to forest including secondary vegetation and to offer the model
14
some tolerance with regard to misclassifications. Hence, the formalization of land cover trajectories is
relatively straight forward, and not necessarily based on elaborate a-priori knowledge. Previous tests showed
a very similar outcome concerning the modification of these parameters, yet using very strict transitions
could lead to undesired results and suppress dynamics entirely. In the scope of this study we utilize different
transition matrices due to the varying time gap between the five TS-X acqusitions. The revisit rate of TS-X
is eleven days, and the available imagery shows one gap of eleven days, two gaps of 22 days, as well as one
gap of 33 days in between neighboring acquisitions. We hence linearly modify the transitions to adjust to
the varying temporal resolution, since with increasing time, more change is expected to occur.
The following summarizes the most relevant assumptions we made for specification of the transition matrices.
• Pasture areas can potentially be burnt. After burning, likelihood is high to transition back into clean
pasture or shrubby pasture.
• Transitions from shrubby pasture to clean pasture are permitted.
• Clean pasture is considered stable, yet may transition into shrubby pasture or forest. Following
observations of TerraClass, for the study region we assume a slow shift from clean pasture into
shrubby pasture overall.
• Forest is the most consistent class. It can get removed, yet especially shrubby pasture can develop
into forest as the class also includes secondary vegetation.
• Water is used to describe bodies of water which are permanently filled within this dry season.
• For any class there is a small tolerance to evade to counteract inconsistent transitions which may be
caused by misclassifications.
3.4. Classification & Validation
Three different types of classification are compared: (i) the baseline IVM classification, (ii) the spatialonly MRF with βt = 0, from now on referred to as s-MRF, and (iii) the spatial-temporal MRF, referred to
as st-MRF. While many studies rely on supervised classification using SVM and Random Forest classifiers,
various studies show that IVM perform at least equally well in terms of accuracy ((Roscher et al., 2012b;
Braun et al., 2012)). Therefore the original IVM classification is considered as adequate baseline classification.
Reference polygons exclusively comprise either training or test samples to avoid spatial autocorrelation.
15
For training purposes, 15 samples per polygon are randomly selected, using a minimum sampling distance
of 30 meters. A systematic sampling ensures that an adequate number of training samples is selected for
all five classes; clean pasture, shrubby pasture, burnt pasture, water, and forest. Validation is
conducted considering the current terms of good practice as laid out by Olofsson et al. (2014). Samples
are clustered in polygons to improve on the spatial variability of both, training and test samples, with
pixels being the assessment unit. This sampling strategy is a necessary trade-off between ideal conditions
of independent random sampling and the difficulties of obtaining large scale, multi-temporal reference data
in a challenging environment (Olofsson et al., 2014). Error matrices are derived to serve as a basis for
the estimation of overall accuracies (OA), user accuracies (UA), producer accuracies (PA), and their corresponding confidence intervals (CI). In addition, we calculate area measures and their confidence intervals at
each acquisition date to estimate the development of burnt pasture land over the entire 2014’s dry season.
Classification and validation is conducted ten times using different training and test sets and the results are
averaged.
4. Results
We show that we can benefit from the high repetition rate and high ground resolution of TS-X, and
that the proposed framework outperforms common classification approaches in terms of area adjusted mapping accuracy (Olofsson et al., 2014). Table 3 illustrates the average area adjusted OA for the five TS-X
scenes, using the three different methods. Irrespective of the acquisition date, the accuracy was significantly
improved by the MRF, with the s-MRF consistently outperforming the IVM only results, and st-MRF consistently outperforming s-MRF to a lesser degree. The weakest classification of IVM (60%) and s-MRF
(69%) could be clearly improved by up to 19 and 10 percentage points compared to the classification results
achieved with st-MRF (79%). On average, OA could be improved by 8.6 percentage points using s-MRF,
and 12.2 percentage points using st-MRF when compared to the IVM classification. As recommended by
Olofsson et al. (2014), we additionally calculated variance measures for results, yet with confidence intervals
Table 3: Area adjusted overall accuracies for different dates. The shown values are means over 10 iterations.
Acquisition Date
2014-06-08
2014-06-30
2014-07-22
2014-08-24
2014-09-04
Polarization
IVM
s-MRF
st-MRF
VV-VH
HH-HV
VV-VH
VV-VH
HH-HV
0.65
0.60
0.66
0.69
0.68
0.75
0.69
0.76
0.74
0.77
0.77
0.79
0.78
0.76
0.78
16
2014-06-08
2014-06-30
2014-07-22
2014-08-24
2014-09-04
(%)
100
80
60
40
20
0
100
80
60
40
20
0
100
80
60
40
20
0
100
80
60
40
20
0
100
80
60
40
20
0
IVM
s-MRF
st-MRF
UA PA UA PA UA PA UA PA UA PA
Forest P. Burnt P. Clean P. Shrub Water
Figure 7: User’s and Producer’s Accuracy
for all the classes at each date.
17
generally falling well below 1 percentage point, we will not further address these measurements.
Figure 7 summarizes the average UA’s and PA’s of the three approaches (IVM, s-MRF, st-MRF). IVM
yields the lowest accuracies, while st-MRF generally shows the highest, and most balanced class-specific
accuracies. All approaches are especially reliable concerning the classification of forest areas, with st-MRF
achieving especially high PA’s for this class (≥ 90%). For any given approach, the three pasture classes
are classified with significantly lower accuracies than forest areas. Shrubby pasture and clean pasture
are overall underrepresented, with PA’s between 30% and 60% depending on the scene and method. Clean
pasture generally yields accuracies of approximately 50%, but also classification of this class is particularly
problematic concerning the PA of the last scene (around 40%). Weak classification results of the different
pasture classes are generally caused by confusion within the different pasture types, and reflects findings
of comparable studies which utilize X-band SAR data (Schlund et al., 2013). In general, st-MRF shows
higher accuracies, when compared to the classification results achieved by the other approaches, and is
capable of mapping burnt pasture starting from 2014-08-24. This is very notable, as a general concern
regarding multi-temporal MRF’s are its smoothing effects, which could cause the suppression of sporadic
events. Due to the low number of burnt pasture areas before the end of July, we are not able to reliably
calculate accuracies for burnt pasture areas at every date. Only up to two burnt fields exist for the first
two acquisitions, which do not allow for an adequate classification and validation. However, this is also
in accordance to the typical land management in the region, insofar slash-and-burn activities usually start
later in the season. Nevertheless, the class is kept as st-MRF utilizes any class for the scenes and to have
a consistent classification scheme over the entire period. Although some additional burnt pasture areas
occur in July (ten areas over the entire study area), the classification accuracy remains very low. Despite
consisting of few samples, possibly due to its temporal consistency and very distinct signature, water is
mapped especially well. As water encompasses just very few areas over the entire study site, yet yields
PAs of 85% and higher for the baseline approach. This weakness appears to get enhanced by the s-MRF
approach, which yields a remarkable drop in the PA of water at some dates, while the spatio-temporal
MRF appears to ensure its further designation. This behavior underlines capabilities of st-MRF to not only
increase mapping accuracy of temporally sporadic classes, such as burnt pasture, but, remarkably, also
proves the value of st-MRF regarding the mapping of classes which are static, yet spatially small scaled.
Contrastingly, using mono-temporal MRF such classes tend to get suppressed more frequently. Regarding
UA, water is mapped very convincingly with accuracies of over 95% using the MRF approaches, yet the
IVM classification shows much less reliable accuracies.
18
Figure 8: Comparison of different classifications inside the subsetted area of Figure 1.
Water
Pas. Clean
s-MRF
st-MRF
2014-09-04
2014-08-24
2014-07-22
2014-06-30
2014-06-08
IVM
Pas. Burnt
19
Pas. Scrub
Forest
Reference
3000
Area of Pasture Types in km2
2500
2000
1500
1000
500
Burnt
Clean
Shrubby
0
06-08
06-30
07-22
Date
08-24
09-04
Figure 9: Growth of burnt areas over the 2014 dry season. Error bars are indicative of the 95% confidence interval.
The visual assessment of the classification maps underlines the positive effect of the MRF-based approaches
(Figures 8 and 10). A large number of speckle induced misclassification can be attributed to any of the
maps classified using IVM with texture parameters. This effect is suppressed to some extent by s-MRF, yet
individual clusters of misclassification can still be located. While not entirely homogeneous, st-MRF suppresses the noise considerably, yet maintaining the general spatial patterns of LULC. Fine spatial structures
appear to not get suppressed by the st-MRF, despite conservative IVM estimates.
Using land cover maps derived from st-MRF, Figure 9 illustrates that clear trends can still be derived using
the proposed data and methods. The figure illustrates a high percentage of shrubby pasture land early in
the dry season. Over the course of the dry season, this amount is continuously shrinking, while the burning
of pasture starts growing exponentially at the end of July. At the end of the dry season, the area of clean
pasture land is comparable to that of shrubby pasture.
20
21
(a) 2014-06-08
(c) 2014-07-22
(d) 2014-08-24
Figure 10: Final classification results using the st-MRF approach.
(b) 2014-06-30
(e) 201409-04
5. Discussion
The main objective of this study, i.e., the adaptation of recent methods for the mapping of dynamic LULC
in a tropical setting, is shown to be generally positive in our study. The proposed approach of using spatialtemporal MRF with expert knowledge is generally able to capture short term LULC dynamics, which are
challenging to map using standard classification techniques. Although our validation confirms limitations of
short-wavelength SAR data (e.g., when differentiating different pasture types, especially mono-temporally),
the proposed approach enables the generation of a meaningful time series of homogeneous LULC maps using
only SAR data. Consequently, the reliable prompt mapping of LULC change can be achieved independent
of cloud cover and atmospheric inference.
The results show that the use of the proposed approach outperforms standard IVM classifications utilizing texture parameters only, as well as common spatial MRF, in terms of classification accuracy. Visual
inspection of burnt pasture areas of early dates shows bright and overall heterogeneous backscatter within
the class, and similarity to the other pasture classes at TS-X images, while Landsat and RapidEye images
unambiguously indicate burnt pasture. Possible reasons for this could be organic debris or wet conditions.
Contrary to that, many burnt areas of subsequent scenes, after the occurrence of large scale burning, can
be identified more clearly at X-band as areas of low backscatter.
Regarding X-band data the potential transfer of the approach to the wet season, which is characterized by
higher saturation of backscatter intensity, is another challenge. While separation of pasture types already
appears difficult in the dry season, the integration of temporal context via the MRF might allow for a reliable
separation of pasture and forest areas over the wet season. Additional testing showed furthermore that utilization of temporal trajectories alone, despite generally not as effective as utilization of the spatial context
(s-MRF), can be used to significantly elevate all accuracies above 70%. In particular the weak classification
of 2014-06-30 could benefit from this approach as the variance of classification outcomes is reduced between
different scenes.
Our findings are in accordance with the results of other recent studies, which were able to improve the
classification accuracy via implementation of multi-temporal MRF (Wehmann and Liu, 2015; Liu et al.,
2008). While Wehmann and Liu (2015) use regionally optimized transition matrices, and a state of the art
integrated kernel based on Moser and Serpico (2013) to achieve high classification accuracies over long time
periods, the proposed method aims on the detection of short term land cover change in SAR imagery, and
utilizes LBP for inference as well as IVM for classification. The visual assessment of the classification results
confirms the positive effect of the MRF on the classification accuracy. Although the maps provided by the
22
23
Figure 11: Difference map between 2014-06-30 classifications of st-MRF and IVM. Light colors indicate agreement between the
two maps. Dark colors indicate class ambiguities, while the class of the final st-MRF classification is presented.
conventional IVM classification show general land use patterns, the results are affected by typical SARinherent noise. Even homogeneous areas appear very noisy, despite texture parameters that were included
in the classification procedure. Boundaries between individual land cover and land use classes may appear
blurred and are hard to identify. This drawback is significantly reduced by both MRF-based methods. As
LBP tries to minimize the transition energy by homogenizing adjacent pixels, areas become overall more
concentrated and edges along different LULC classes can be more clearly identified. Benefits can also be
attributed to the classification of the interior of areas, as the application of MRF suppresses outliers. Thus,
the results confirm the edge preserving capabilities of MRF, even for challenging spatial class transitions
(e.g. forest to shrubby pasture).
With regard to class-specific accuracies, the spatio-temporal MRF offers preferable results over just IVM
and the spatial-only approach. Figure 11 illustrates the differences between a 2014-06-30 classification of
the IVM and the st-MRF approaches, underlining the potentials of solving confusion between forest and
shrubby pasture. It is colorized to highlight disagreements of the classifications, with pale colors signifying consenting classifications, and opaque colors indicating classes as assigned by st-MRF (for legend see
Figure 8). It is especially obvious that with increasing vegetation density confusion also rises, and that
clean pasture is classified congruently in both approaches. While the classification of clean pasture
and shrubby pasture remains challenging, TS-X data constitutes an adequate data source for forest /
non-forest mapping. The PA and UA for forest are higher compared to the accuracies achieved for the
other classes, and are in accordance with the accuracies of comparable studies (Schlund et al., 2013; Garcia
et al., 2011). While we did not perform specific analysis on the differences of HH-HV and VV-VH polarized
data sets, Table 3 shows the OA of the internal HH-HV polarized scenes (06-30) to especially benefit from
the multi-temporal integration, and also its neighboring scenes to benefit disproportionately. We can thus
assume synergetic effects to be transferable through a multi-temporal MRF, yielding a promising outlook
for the multi-sensoral integration of various data sources.
Regarding the low requirements concerning parameterization and the implementation through moving windows, we consider the introduced method to be transferable to other study regions. Adaptation of the
transition matrices allows the method to be fitted to more static environments, or also to address multiannual time series data. Despite the ambitious goals of this study, i.e. to perform land cover mapping in
a densely vegetated and dynamic tropical study region using TS-X data, and some documented limitations
concerning the separability of different pasture types, we were able to achieve improvements over standard
classifications. As the method incorporates adjacency information, potential shortcomings exist when the
24
ground resolution is coarse relative to the mapped land cover. In this case, fragmented structures might
get suppressed. Further adjustment would also be required if assumptions on land cover trajectories are
variant in a multi-temporal setup. For example, two scenes from the dry season carry a different transition
probability with regard to burning than two scenes from the wet season. While this can be easily solved
through different transition models, within this study we just included slight modifications of transition
matrices to account for the different intervals between TS-X acquisitions.
6. Conclusion
The results show clearly that the integration of spatial-temporal MRFs are advantageous to the baseline
classification approach and spatial MRF methods. Especially the classification of forest areas yields very
high accuracies. We were able to successfully implement an LBP optimization for the regularization of high
resolution, multi-temporal TS-X images of a tropical context. We were furthermore able to give adequate
estimates on the spatio-temporal pattern of land use dynamics such as burned pastures. Importantly, the
suggested approach is able to handle process of small spatio-temporal scale, and despite its smoothing effects does not suppress fine structures. Separation of different types of pasture (pasto sujo and pasto limpo)
remains a challenging task at the short wavelength. Classification of burnt pasture early in the season
highlights limitations of the MRF-based model, which arise when the underlying classification accuracy is
already limited. While the approach is well suited to regularize small classification errors using contextual
information, it is not able to sufficiently address misclassifications in complex, transitional environments
with weak classification accuracies. The sometimes relatively low class accuracies are not necessarily a limitation of the proposed method, but rather caused by the short-wave TS-X data as well as class-specific
characteristics.
Particularly for study sites which are characterized by land use patterns of high spatial-temporal variability,
the proposed approach (i.e., using spatial-temporal MRF with expert knowledge) appears feasible. Using
expert knowledge on land cover trajectories, we could positively influence model performance and bypass
computationally demanding techniques for the estimation of MRF parameters.
When derived from multiple classifications, change maps are generally strongly affected by weak initial classifications.
The proposed method is formalized to be transferable to large, possibly multi-sensoral, image stacks. For
future studies our aim is to integrate the regularization of short-term, intra-annual dynamics with long-term
dynamics such as deforestation and agricultural trends, using multi-sensoral imagery.
25
Acknowledgments
The study was carried out as part of the SenseCarbon research project funded by DLR/BMWi (50 EE
1255). SenseCarbon is part of GFOI-R&D (BRA-2). We thank the countless developers behind the freely
accessible software utilized in this study (i.e. Python, GDAL, S1TBX, and Q-GIS). For their supporting
efforts we direct further thanks to Niklas Potthoff, Paul Wagner, Rolf Rissiek, Bernd Melchers, Merry
Crowson, and Christian Lamparter, as well as the anonymous reviewers.
References
References
Adami, M., Gomes, A. R., Coutinho, A. C., Esquerdo, J. C. D. M., Venturieri, A., 2015. Dinâmica do uso e cobertura da terra
no estado do pará entre os anos de 2008 a 2012. XVII Simpósio Brasileiro de Sensoriamento Remoto.
Ahmed, S. E., Souza, C. M., Riberio, J., Ewers, R. M., Jan 2013. Temporal patterns of road network development in the
Brazilian amazon. Reg Environ Change 13 (5), 927–937.
URL http://dx.doi.org/10.1007/s10113-012-0397-z
Almeida, A. S. d., Vieira, I. C. G., 2008. Dinâmica da cobertura vegetal e uso da terra no municı́pio de são francisco do pará
(pará, brasil) com o uso da técnica de sensoriamento remoto. Boletim do Museu Paraense Emı́lio Goeldi Ciências Naturais
3 (1), 81–92.
Almeida, C. A. d., Coutinho, A. C., Esquerdo, J. C. D. M., Adami, M., Venturieri, A., Diniz, C. G., Dessay, N., Durieux, L.,
Gomes, A. R., Sep 2016. High spatial resolution land use and land cover mapping of the Brazilian legal amazon in 2008 using
landsat-5/tm and modis data. Acta Amaz. 46 (3), 291–302.
URL http://dx.doi.org/10.1590/1809-4392201505504
Almeida-Filho, R., Shimabukuro, Y. E., Rosenqvist, A., Sanchez, G. A., Jul 2009. Using dual-polarized alos PALSAR data for
detecting new fronts of deforestation in the Brazilian amaznia. International Journal of Remote Sensing 30 (14), 3735–3743.
URL http://dx.doi.org/10.1080/01431160902777175
Andres, B., Kappes, J. H., Köthe, U., Schnörr, C., Hamprecht, F. A., 2010a. An empirical comparison of inference algorithms
for graphical models with higher order factors using opengm. Pattern Recognition, 353–362.
URL http://dx.doi.org/10.1007/978-3-642-15986-2_36
Andres, B., Kappes, J. H., Köthe, U., Schnörr, C., Hamprecht, F. A., 2010b. An empirical comparison of inference algorithms
for graphical models with higher order factors using opengm. In: Pattern Recognition. Springer, pp. 353–362.
Aschbacher, J., Milagro-Pérez, M. P., May 2012. The European earth monitoring (GMes) programme: Status and perspectives.
Remote Sensing of Environment 120, 3–8.
URL http://dx.doi.org/10.1016/j.rse.2011.08.028
Blaes, X., Vanhalle, L., Defourny, P., 2005. Efficiency of crop identification based on optical and sar image time series. Remote
sensing of environment 96 (3), 352–365.
26
Bouman, C. A., Shapiro, M., 1994. A multiscale random field model for bayesian image segmentation. Image Processing, IEEE
Transactions on 3 (2), 162–177.
Bovolo, F., Bruzzone, L., Dec 2005. A detail-preserving scale-driven approach to change detection in multitemporal SAR
images. IEEE Transactions on Geoscience and Remote Sensing 43 (12), 2963–2972.
URL http://dx.doi.org/10.1109/TGRS.2005.857987
Boykov, Y., Veksler, O., Zabih, R., 2001. Fast approximate energy minimization via graph cuts. Pattern Analysis and Machine
Intelligence, IEEE Transactions on 23 (11), 1222–1239.
Braun, A. C., Weidner, U., Hinz, S., apr 2012. Classification in high-dimensional feature spaces assessment using SVM, IVM
and RVM with focus on simulated EnMAP data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote
Sensing 5 (2), 436–443.
Cai, S., Liu, D., Sulla-Menashe, D., Friedl, M. A., May 2014. Enhancing modis land cover product with a spatial-temporal
modeling algorithm. Remote Sensing of Environment 147, 243–255.
URL http://dx.doi.org/10.1016/j.rse.2014.03.012
Coy, M., Klingler, M., 2014. Frentes pioneiras em transformação: o eixo da br-163 e os desafios socioambientais. Territórios e
Fronteiras 7 (1), 1–26.
Cutler, M., Boyd, D., Foody, G., Vetrivel, A., 2012. Estimating tropical forest biomass with a combination of sar image texture
and landsat tm data: An assessment of predictions between regions. ISPRS Journal of Photogrammetry and Remote Sensing
70, 66–77.
Davidson, E. A., de Araújo, A. C., Artaxo, P., Balch, J. K., Brown, I. F., Bustamante, M. M., Coe, M. T., DeFries, R. S.,
Keller, M., Longo, M., et al., 2012. The amazon basin in transition. Nature 481 (7381), 321–328.
Dekker, R. J., 2003. Texture analysis and classification of ers sar images for map updating of urban areas in the netherlands.
Geoscience and Remote Sensing, IEEE Transactions on 41 (9), 1950–1958.
Englhart, S., Keuck, V., Siegert, F., May 2011. Aboveground biomass retrieval in tropical forests – the potential of combined
x- and l-band SAR data use. Remote Sensing of Environment 115 (5), 1260–1271.
URL http://dx.doi.org/10.1016/j.rse.2011.01.008
Foley, J. A., Jul 2005. Global consequences of land use. Science 309 (5734), 570–574.
URL http://dx.doi.org/10.1126/science.1111772
Garcia, C. E., dos SANTOS, J. R., Mura, J. C., Kux, H. J. H., 2011. Análise do potencial de imagem terrasar-x para
mapeamento temático no sudoeste da amazônia brasileira. Acta Amazonica 42 (2).
GDAL Development Team, 2015. GDAL - Geospatial Data Abstraction Library, Version 2.0.1. Open Source Geospatial Foundation.
URL http://www.gdal.org
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., Thau, D., Stehman, S. V., Goetz,
S. J., Loveland, T. R., Kommareddy, A., Egorov, A., Chini, L., Justice, C. O., Townshend, J. R. G., 2013. High-resolution
global maps of 21st-century forest cover change. Science 342 (6160), 850–853.
URL http://science.sciencemag.org/content/342/6160/850
Haralick, R. M., Shanmugam, K., Dinstein, I. H., 1973. Textural features for image classification. Systems, Man and Cybernetics,
IEEE Transactions on (6), 610–621.
Hoberg, T., Rottensteiner, F., Queiroz Feitosa, R., Heipke, C., 2015. Conditional random fields for multitemporal and multiscale
27
classification of optical satellite imagery. Geoscience and Remote Sensing, IEEE Transactions on 53 (2), 659–673.
INPE, 2015. Projeto prodes - monitoramento da floresta amazonica brasileira por satelite.
Khatami, R., Mountrakis, G., Stehman, S. V., May 2016. A meta-analysis of remote sensing research on supervised pixelbased land-cover image classification processes: General guidelines for practitioners and future research. Remote Sensing of
Environment 177, 89–100.
URL http://dx.doi.org/10.1016/j.rse.2016.02.028
Kolmogorov, V., Zabin, R., 2004. What energy functions can be minimized via graph cuts? Pattern Analysis and Machine
Intelligence, IEEE Transactions on 26 (2), 147–159.
Kumar, S., Hebert, 2003. Discriminative random fields: a discriminative framework for contextual interaction in classification.
Proceedings Ninth IEEE International Conference on Computer Vision.
URL http://dx.doi.org/10.1109/ICCV.2003.1238478
Kumar, T., Patnaik, C., 2013. Discrimination of mangrove forests and characterization of adjoining land cover classes using temporal c-band synthetic aperture radar data: A case study of sundarbans. International Journal of Applied Earth
Observation and Geoinformation 23, 119–131.
Lapola, D. M., Martinelli, L. A., Peres, C. A., Ometto, J. P., Ferreira, M. E., Nobre, C. A., Aguiar, A. P. D., Bustamante,
M. M., Cardoso, M. F., Costa, M. H., et al., 2014. Pervasive transition of the brazilian land-use system. Nature climate
change 4 (1), 27–35.
Li, J., Bioucas-Dias, J. M., Plaza, A., feb 2013. Spectral spatial classification of hyperspectral data using loopy belief propagation
and active learning. IEEE Transactions on Geoscience and Remote Sensing 51 (2), 844–856.
Liu, D., Kelly, M., Gong, P., 2006. A spatial–temporal approach to monitoring forest disease spread using multi-temporal high
spatial resolution imagery. Remote sensing of environment 101 (2), 167–180.
Liu, D., Song, K., Townshend, J. R., Gong, P., 2008. Using local transition probability models in markov random fields for
forest change detection. Remote Sensing of Environment 112 (5), 2222–2231.
McNairn, H., Champagne, C., Shang, J., Holmstrom, D., Reichert, G., 2009. Integration of optical and synthetic aperture radar
(sar) imagery for delivering operational annual crop inventories. ISPRS Journal of Photogrammetry and Remote Sensing
64 (5), 434–449.
Melgani, F., Serpico, S. B., 2003. A markov random field approach to spatio-temporal contextual image classification. Geoscience
and Remote Sensing, IEEE Transactions on 41 (11), 2478–2487.
Moser, G., Serpico, S. B., 2010. Contextual remote-sensing image classification by support vector machines and markov random
fields. In: Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International. IEEE, pp. 3728–3731.
Moser, G., Serpico, S. B., 2013. Combining support vector machines and markov random fields in an integrated framework for
contextual image classification. Geoscience and Remote Sensing, IEEE Transactions on 51 (5), 2734–2752.
Moser, G., Serpico, S. B., Benediktsson, J. A., 2013. Land-cover mapping by markov modeling of spatial–contextual information
in very-high-resolution remote sensing images.
Müller, H., Rufin, P., Griffiths, P., Barros Siqueira, A. J., Hostert, P., Jan 2015. Mining dense landsat time series for separating
cropland and pasture in a heterogeneous Brazilian savanna landscape. Remote Sensing of Environment 156, 490–499.
URL http://dx.doi.org/10.1016/j.rse.2014.10.014
Murphy, K. P., Weiss, Y., Jordan, M. I., 1999. Loopy belief propagation for approximate inference: An empirical study. In:
Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence. Morgan Kaufmann Publishers Inc., pp.
28
467–475.
Nyoungui, A. N., Tonye, E., Akono, A., Jan 2002. Evaluation of speckle filtering and texture analysis methods for land cover
classification from SAR images. International Journal of Remote Sensing 23 (9), 1895–1925.
URL http://dx.doi.org/10.1080/01431160110036157
Olding, W. C., Olivier, J. C., Salmon, B. P., Jul 2015. A Markov random field model for decision level fusion of multi-source
image segments. 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS).
URL http://dx.doi.org/10.1109/IGARSS.2015.7326289
Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., Wulder, M. A., May 2014. Good practices for
estimating area and assessing accuracy of land change. Remote Sensing of Environment 148, 42–57.
URL http://dx.doi.org/10.1016/j.rse.2014.02.015
Pearl, J., 1982. Reverend bayes on inference engines: A distributed hierarchical approach. In: AAAI. pp. 133–136.
Pfeifer, M., Kor, L., Nilus, R., Turner, E., Cusack, J., Lysenko, I., Khoo, M., Chey, V., Chung, A., Ewers, R., Apr 2016.
Mapping the structure of borneos tropical forests across a degradation gradient. Remote Sensing of Environment 176, 84–97.
URL http://dx.doi.org/10.1016/j.rse.2016.01.014
Platt, J. C., 1999. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In:
Advances in large margin classifiers. Citeseer.
Qi, Z., Yeh, A. G.-O., Li, X., Lin, Z., Mar 2012. A novel algorithm for land use and land cover classification using RADarsat-2
polarimetric SAR data. Remote Sensing of Environment 118, 21–39.
URL http://dx.doi.org/10.1016/j.rse.2011.11.001
Qi, Z., Yeh, A. G.-O., Li, X., Xian, S., Zhang, X., Jul 2015. Monthly short-term detection of land development using radarsat-2
polarimetric SAR imagery. Remote Sensing of Environment 164, 179–196.
URL http://dx.doi.org/10.1016/j.rse.2015.04.018
Reiche, J., Souzax, C. M., Hoekman, D. H., Verbesselt, J., Persaud, H., Herold, M., Oct 2013. Feature level fusion of multitemporal alos PALSAR and landsat data for mapping and monitoring of tropical deforestation and forest degradation. IEEE
J. Sel. Top. Appl. Earth Observations Remote Sensing 6 (5), 2159–2173.
URL http://dx.doi.org/10.1109/JSTARS.2013.2245101
Reiche, J., Verbesselt, J., Hoekman, D., Herold, M., 2015. Fusing landsat and sar time series to detect deforestation in the
tropics. Remote Sensing of Environment 156, 276–293.
Roscher, R., Förstner, W., Waske, B., 2012a. I2 vm: Incremental import vector machines. Image and Vision Computing 30 (4),
263–278.
Roscher, R., Waske, B., Förstner, W., 2012b. Incremental import vector machines for classifying hyperspectral data. Geoscience
and Remote Sensing, IEEE Transactions on 50 (9), 3463–3473.
Rufin, P., Müller, H., Pflugmacher, D., Hostert, P., Sep 2015. Land use intensity trajectories on amazonian pastures derived
from landsat time series. International Journal of Applied Earth Observation and Geoinformation 41, 1–10.
URL http://dx.doi.org/10.1016/j.jag.2015.04.010
Sarker, M. L. R., Nichol, J., Ahmad, B., Busu, I., Rahman, A. A., 2012. Potential of texture measurements of two-date dual
polarization palsar data for the improvement of forest biomass estimation. ISPRS Journal of Photogrammetry and Remote
Sensing 69, 146–166.
Sarker, M. L. R., Nichol, J., Iz, H. B., Ahmad, B., Rahman, A. A., 2013. Forest biomass estimation using texture measurements
29
of high-resolution dual-polarization c-band sar data.
Schlund, M., von Poncet, F., Hoekman, D. H., Kuntz, S., Schmullius, C., 2013. Importance of bistatic sar features from
tandem-x for forest mapping and monitoring. Remote Sensing of Environment.
Solberg, A. H. S., Taxt, T., Jain, A. K., 1996. A markov random field model for classification of multisource satellite imagery.
Geoscience and Remote Sensing, IEEE Transactions on 34 (1), 100–113.
Sonobe, R., Tani, H., Wang, X., Kobayashi, N., Shimamura, H., Feb 2014. Random forest classification of crop type using
multi-temporal terrasar-x dual-polarimetric data. Remote Sensing Letters 5 (2), 157–164.
URL http://dx.doi.org/10.1080/2150704X.2014.889863
Stefanski, J., Chaskovskyy, O., Waske, B., 2014a. Mapping and monitoring of land use changes in post-soviet western ukraine
using remote sensing data. Applied Geography 55, 155–164.
Stefanski, J., Chaskovskyy, O., Waske, B., Dec 2014b. Mapping and monitoring of land use changes in post-soviet western
Ukraine using remote sensing data. Applied Geography 55, 155–164.
URL http://dx.doi.org/10.1016/j.apgeog.2014.08.003
Szeliski, R., Zabih, R., Scharstein, D., Veksler, O., Kolmogorov, V., Agarwala, A., Tappen, M., Rother, C., 2006. A comparative
study of energy minimization methods for Markov random fields. Lecture Notes in Computer Science, 16–29.
URL http://dx.doi.org/10.1007/11744047_2
Tarabalka, Y., Fauvel, M., Chanussot, J., Benediktsson, J. A., 2010. Svm-and mrf-based method for accurate classification of
hyperspectral images. Geoscience and Remote Sensing Letters, IEEE 7 (4), 736–740.
Tran, T. N., Wehrens, R., Hoekman, D. H., Buydens, L. M., 2005. Initialization of markov random field clustering of large
remote sensing images. Geoscience and Remote Sensing, IEEE Transactions on 43 (8), 1912–1919.
Uhlmann, S., Kiranyaz, S., Apr 2014. Classification of dual- and single polarized SAR images by incorporating visual features.
ISPRS Journal of Photogrammetry and Remote Sensing 90, 10–22.
URL http://dx.doi.org/10.1016/j.isprsjprs.2014.01.005
Vitousek, P. M., Jul 1997. Human domination of earths ecosystems. Science 277 (5325), 494–499.
URL http://dx.doi.org/10.1126/science.277.5325.494
Voisin, A., Krylov, V. A., Moser, G., Serpico, S. B., Zerubia, J., 2013. classification of very high resolution sar images of urban
areas using copulas and texture in a hierarchical markov random field model. Geoscience and Remote Sensing Letters, IEEE
10 (1), 96–100.
Waske, B., Benediktsson, J. A., 2007. Fusion of support vector machines for classification of multisensor data. Geoscience and
Remote Sensing, IEEE Transactions on 45 (12), 3858–3866.
Waske, B., Braun, M., 2009. Classifier ensembles for land cover mapping using multitemporal sar imagery. ISPRS Journal of
Photogrammetry and Remote Sensing 64 (5), 450–457.
Waske, B., van der Linden, S., 2008. Classifying multilevel imagery from sar and optical sensors by decision fusion. Geoscience
and Remote Sensing, IEEE Transactions on 46 (5), 1457–1466.
Wehmann, A., Liu, D., 2015. A spatial–temporal contextual markovian kernel method for multi-temporal land cover mapping.
ISPRS Journal of Photogrammetry and Remote Sensing.
Wulder, M. A., Masek, J. G., Cohen, W. B., Loveland, T. R., Woodcock, C. E., 2012. Opening the archive: How free data has
enabled the science and monitoring promise of landsat. Remote Sensing of Environment 122, 2–10, landsat Legacy Special
Issue.
30
URL http://www.sciencedirect.com/science/article/pii/S003442571200034X
Xie, H., Pierce, L. E., Ulaby, F. T., 2002. Sar speckle reduction using wavelet denoising and markov random field modeling.
Geoscience and Remote Sensing, IEEE Transactions on 40 (10), 2196–2212.
Zhu, J., Hastie, T., 2005. Kernel logistic regression and the import vector machine. Journal of Computational and Graphical
Statistics 14 (1), 185–205.
Zhu, Z., Woodcock, C. E., 2014. Continuous change detection and classification of land cover using all available landsat data.
Remote sensing of Environment 144, 152–171.
References
31
| 1cs.CV
|
Separators in region intersection graphs
arXiv:1608.01612v3 [math.CO] 27 Jul 2017
James R. Lee∗
Abstract
For undirected graphs G (V, E) and G0 (V0 , E0 ), say that G is a region intersection
graph over G0 if there is a family of connected subsets {R u ⊆ V0 : u ∈ V } of G0 such that
{u, v} ∈ E ⇐⇒ R u ∩ R v , ∅.
We show if G0 excludes the complete graph K h as a minor for some h > 1, then every region
√
intersection graph G over G0 with m edges has a balanced separator with at most c h m nodes,
where c h is a constant depending only on h. If G additionally has uniformly bounded vertex
degrees, then such a separator is found by spectral partitioning.
A string graph is the intersection graph of continuous arcs in the plane. String graphs are
precisely region intersection graphs over planar graphs. Thus the preceding result implies that
√
every string graph with m edges has a balanced separator of size O( m). This bound is optimal,
as it generalizes the planar separator theorem. It confirms a conjecture of Fox and Pach (2010),
√
and improves over the O( m log m) bound of Matoušek (2013).
Contents
1
Introduction
1.1 Balanced separators and extremal spread
1.2 Eigenvalues and L2 -extremal spread . . .
1.3 Additional applications . . . . . . . . . . .
1.4 Preliminaries . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
3
7
7
7
2
Vertex separators and conformal graph metrics
8
2.1 Conformal graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Padded partitions and random separators . . . . . . . . . . . . . . . . . . . . . . . . . 10
3
Multi-flows, congestion, and crossings
12
3.1 Duality between conformal metrics and multi-flows . . . . . . . . . . . . . . . . . . . 13
3.2 Crossing congestion and excluded minors . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Vertex congestion in rigs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4
Careful minors and random separators
4.1 Careful minors in rigs . . . . . . . . . . . . .
4.2 Chopping trees . . . . . . . . . . . . . . . . .
4.3 The random separator construction . . . . . .
4.4 A diameter bound for well-spaced subgraphs
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
16
19
20
Applications and discussion
24
5.1 Spectral bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 Weighted separators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Bi-Lipschitz embedding problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
∗ University
of Washington. Partially supported by NSF CCF-1407779.
1
1
Introduction
Consider an undirected graph G0 (V0 , E0 ). A graph G (V, E) is said to be a region intersection
graph (rig) over G0 if the vertices of G correspond to connected subsets of G0 and there is an edge
between two vertices of G precisely when those subsets intersect. Concretely, there is a family of
connected subsets {R u ⊆ V0 : u ∈ V } such that {u, v} ∈ E ⇐⇒ R u ∩ R v , ∅. For succinctness, we
will often refer to G as a rig over G0 .
Let rig(G0 ) denote the family of all finite rigs over G0 . Prominent examples of such graphs
include the intersection graphs of pathwise-connected regions on a surface (which are intersection
graphs over graphs that can be drawn on that surface).
For instance, string graphs are the intersection graphs of continuous arcs in the plane. It is easy
to see that every finite string graph G is a rig over some planar graph: By a simple compactness
argument, we may assume that every two strings intersect a finite number of times. Now consider
the planar graph G0 whose vertices lie at the intersection points of strings and with edges between
two vertices that are adjacent on a string (see Figure 1). Then G ∈ rig(G0 ). It is not too difficult to
see that the converse is also true; see Lemma 1.4.
To illustrate the non-trivial nature of such objects, we recall that there are string graphs on n
strings that require 2Ω(n) intersections in any such representation [KM91]. The recognition problem
for string graphs is NP-hard [Kra91]. Decidability of the recognition problem was established in
[SŠ04], and membership in NP was proved in [SSŠ03]. We refer to the recent survey [Mat15] for
more of the background and history behind string graphs.
Even when G0 is planar, the rigs over G0 can be dense: Every complete graph is a rig over some
planar graph (in particular, every complete graph is a string graph). It has been conjectured
by Fox
√
and Pach [FP10] that every m-edge string graph has a balanced separator with O( m) nodes. Fox
√
and Pach proved that such graphs have separators of size O(m 3/4 log m) and presented a number
of applications
of their separator theorem. Matoušek [Mat14] obtained a near-optimal bound of
√
O( m log m). In the present work, we confirm the conjecture of Fox and Pach, and generalize the
result to include all rigs over graphs that exclude a fixed minor.
Theorem 1.1. If G ∈ rig(G0 ) and G0 excludes K h as a minor, then G has a 23 -balanced separator of size at
√
√
most c h m where m is the number of edges in G. Moreover, one has the estimate c h 6 O(h 3 log h).
In the preceding statement, an ε-balanced separator of G (V, E) is a subset S ⊆ V such that in
the induced graph G[V \ S], every connected component contains at most ε|V | vertices.
The proof of Theorem 1.1 is constructive, as it is based on solving and rounding a linear program;
it yields a polynomial-time algorithm for constructing the claimed separator. In the case when there
is a bound on the maximum degree of G, one can use the well-known spectral bisection algorithm
(see Section 1.2).
√Since planar graphs exclude K5 as a minor, Theorem 1.1 implies that m-edge string graphs have
O( m)-node balanced separators. Since
p the graphs that can be drawn on any compact surface of
genus 1 exclude a K h minor for h 6 O( 1 + 1), Theorem 1.1 also applies to string graphs over any
fixed compact surface.
In addition, it implies the Alon-Seymour-Thomas [AST90] separator theorem1 for graphs
excluding a fixed minor, for the following reason. Let us define the subdivision of a graph G to be the
graph GÛ obtained by subdividing every edge of G into a path of length two. Then every graph G is
Û and it is not hard to see that for h > 1, G has a K h minor if and only if GÛ has a K h minor.
a rig over G,
1Note that Theorem 1.1 is quantitatively weaker in the sense that [AST90] shows the existence of separators with
√
√
O(h 3/2 n) vertices. Since every K h -minor-free graph has at most O(nh log h) edges [Kos82, Tho84], our bound is
√
O(h 7/2 (log h)3/4 n).
2
Figure 1: A string graph as a rig over a planar graph.
Applications in topological graph theory. We mention two applications of Theorem 1.1 in graph
theory. In [FP14], the authors present some applications of separator theorems for string graphs. In
two cases, the tight bound for separators leads to tight bounds for other problems. The next two
theorems confirm conjectures of Fox and Pach; as proved in [FP14], they follow from Theorem 1.1.
Both results are tight up to a constant factor.
Theorem 1.2. There is a constant c > 0 such that for every t > 1, it holds that every K t,t -free string graph
on n vertices has at most cnt(log t) edges.
A topological graph is a graph drawn in the plane so that its vertices are represented by points
and its edges by curves connecting the corresponding pairs of points.
Theorem 1.3. In every topological graph with n vertices and m > 4n edges, there are two disjoint sets, each
of cardinality
m2
Ω 2
(1.1)
n log mn
so that every edge in one set crosses all edges in the other.
This improves over the bound of Ω
m2
n c
n 2 (log m
)
for some c > 0 proved in [FPT10], where the
authors also show that the bound (1.1) is tight. Before we conclude this section, let us justify the
observation made earlier.
Lemma 1.4. Finite string graphs are precisely finite region intersection graphs over planar graphs.
Proof. We have already argued that string graphs are planar rigs. Consider now a planar graph
G0 (V0 , E0 ) and a finite graph G (V, E) such that G ∈ rig(G0 ). Let {R u ⊆ V0 : u ∈ V } be a
representation of G as a rig over G0 .
Since G is finite, we may assume that each region R u is finite. To see this, for v ∈ V0 , let its type
be the set T(v) {u ∈ V : v ∈ R u }. Then since G is finite, there are only finitely many types. For
any region R u ⊆ V0 , let R̃ u be a finite set of vertices that exhausts every type in R u , and let R̂ u be a
finite spanning tree of R̃ u in the induced graph G0 [R u ]. Then the regions { R̂ u : u ∈ V } are finite
and connected, and also form a representation of G as a rig over G0 .
When each region R u is finite, we may assume also that G0 is finite. Now take a planar drawing
of G0 in 2 where the edges of G0 are drawn as continuous arcs, and for every u ∈ V, let Tu ⊆ 2
be the drawing of the spanning tree of R u . Each Tu can be represented by a string (simply trace the
tree using an in-order traversal that begins and ends at some fixed node), and thus G is a string
graph.
1.1
Balanced separators and extremal spread
Since complete graphs are string graphs, we do not have access to topological methods based on
the exclusion of minors. Instead, we highlight a more delicate structural theory. The following fact
is an exercise.
3
Fact. If GÛ is a string graph, then G is planar.
More generally, we recall that H is a minor of G if H can be obtained from G by a sequence
of edge contractions, edge deletions, and vertex deletions. If H can be obtained using only edge
contractions and vertex deletions, we say that H is a strict minor of G. The following lemma appears
in Section 4.
Lemma 1.5. If G ∈ rig(G0 ) and HÛ is a strict minor of G, then H is a minor of G0 .
This topological structure of (forbidden) strict minors in G interacts nicely with “conformal
geometry” on G, as we now explain. Consider the family of all pseudo-metric spaces that arise from
a finite graph G by assigning non-negative lengths to its edges and taking the induced shortest path
distance. Certainly if we add an edge to G, the family of such spaces can only grow (since by giving
the edge length equal to the diameter of the space, we effectively remove it from consideration). In
particular, if G K n is the complete graph on n vertices, then every n-point metric space is a path
metric on G.
The same phenomenon does not arise when one instead considers vertex-weighted path metrics
on G. A conformal graph is a pair (G, ω) such that G (V, E) is a graph, and ω : V → + . This
ω(u)+ω(v)
defines a pseudo-metric as follows: Assign to every {u, v} ∈ E a length equal to
and let
2
distω be the induced shortest path distance. We will refer to ω as a conformal metric on G (and
sometimes we abuse terminology and refer to distω as a conformal metric as well).
A significant tool will be the study of extremal conformal metrics on a graph G. Unlike
in the edge-weighted case, the family of path distances coming from conformal metrics can be
well-behaved even if G contains arbitrarily large complete graph minors. As a simple example,
let K denote the complete graph on countably many vertices. Every countable metric space is a
shortest-path metric on some edge-weighting of K , and yet every distance arising from a conformal
metric ω : VK → + is bi-Lipschitz to the ultrametric (u, v) 7→ max{ω(u), ω(v)}.
Vertex expansion and observable spread. Fix a graph G (VG , EG ) ∈ rig(G0 ) with n |VG | and
m |EG |. Since the family rig(G0 ) is closed under taking induced subgraphs, a standard
reduction
√
allows us to focus on finding a subset U ⊆ VG with small isoperimetric ratio:
|∂U |
|U |
.
m
n ,
where
∂U : {v ∈ U : EG (v, VG \ U) , ∅} ,
and EG (v, VG \ U) is the set of edges between v and vertices outside U. Also define the interior
U ◦ U \ ∂U.
Let us define the vertex expansion constant of G as
|∂U |
|VG |
φ G : min
: ∅ , U ⊆ VG , |U ◦ | 6
|U |
2
.
(1.2)
In [FHL08], it is shown that this quantity is related to the concentration of Lipschitz functions
on extremal conformal metrics on G. (The study of such properties has a rich history; consider, for
instance, the concentration function in the sense of Lévy and Milman (e.g., [MS86]) and Gromov’s
observable diameter [Gro07].)
For a finite metric space (X, dist), we define the spread of X as the quantity
s (X, dist) :
1 Õ
dist(x, y) .
|X| 2 x,y∈X
Define the observable spread of X by
1 Õ
sobs (X, dist) : sup
|
f
(x)
−
f
(y)|
:
f
is
1-Lipschitz
.
2
|X| x, y∈X
f :X→
4
(1.3)
Remark 1.6. We remark on the terminology: In general, it is difficult to “view” a large metric space
all at once; this holds both conceptually and from an algorithmic standpoint. If one thinks of
Lipschitz maps f : X → as “observations” then the observable spread captures how much of the
spread can be “seen.”
We then define the L1 -extremal observable spread of G as
s̄obs (G) :
sup
ω:VG →+
sobs (VG , distω ) : kωkL1 (VG ) 6 1 ,
(1.4)
where kωkL1 (VG ) : |V1G | v∈VG ω(v).
Such extremal quantities arise naturally in the study of linear programming relaxations of
discrete optimization problems (like finding the smallest balanced vertex separator in a graph).
Related extremal notions are often employed in conformal geometry and its discretizations; see, in
particular, the notions of extremal length employed by Duffin [Duf62] and Cannon [Can94].
In Section 2.1, we recall the proof of the following theorem from [FHL08] that relates expansion
to the observable spread.
Í
Theorem 1.7 ([FHL08]). For every finite graph G,
1
2 s̄obs (G)
1
6 3s̄obs (G) .
φG
6
Example 1.8. If G is the subgraph of the lattice d on the vertex set {0, 1, . . . , L} d , then φ G 1/L
and s̄ (G) L. This can be achieved by taking ω ≡ 1 and defining f : VG → by f (x) x 1 .
In light of Theorem 1.7, to prove Theorem 1.1, it suffices to give a lower bound on s̄obs (G). It is
natural to compare this quantity to the L 1 -extremal spread of G:
(
Õ
1
s̄ (G) : max
distω (u, v) : kωkL1 (VG ) 6 1
|VG | 2 u,v∈V
)
.
(1.5)
G
Let us examine these two notions for planar graphs using the theory of circle packings.
Example 1.9 (Circle packings). Suppose that G is a finite planar graph. The Koebe-AndreevThurston circle packing theorem asserts that G is the tangency graph of a family {Dv : v ∈ VG } of
circles on the unit sphere 2 ⊆ 3 . Let {c v : v ∈ VG } ⊆ 2 and {r v > 0 : v ∈ VG } be the centers
and radii of the circles, respectively. An argument of Spielman and Teng [ST07] (see also Hersch
Í
[Her70] for the analogous result for conformal mappings) shows that one can take v∈VG c v 0.
If we define ω(v) r v for v ∈ VG , then distω > dist2 > dist3 on the centers {c v : v ∈ VG }. (The
latter two distances are the geodesic distance on 2 and the Euclidean distance on 3 , respectively).
Í
Using the fact that v∈VG c v 0, we have
Õ
kc u − c v k22 2n
u,v∈VG
Õ
kc v k 2 2n 2 .
(1.6)
u∈VG
This yields
Õ
distω (u, v) >
u,v∈VG
Õ
kc u − c v k >
u,v∈VG
Moreover,
s
kωkL1 (VG ) 6 kωkL2 (VG )
1 Õ 2
rv 6
n
v∈VG
5
r
n2
.
2
vol(2 )
πn
r
4
.
n
√
n
It follows that s̄ (G) > 4 .
Observe that the three coordinate projections 3 → are all Lipschitz with respect to dist√ω ,
n
and one of them contributes at least a 1/3 fraction to the sum (1.6). We conclude that s̄obs (G) > 12 .
Combined with Theorem 1.7, this yields a proof of the Lipton-Tarjan separator theorem [LT79].
Similar proofs of the separator theorem based on circle packings are known (see [MTTV97]), and
this one is not new (certainly it was known to the authors of [ST07]).
√
We will prove Theorem 1.1 in two steps: By first giving a lower bound s̄ (G) & n/ m and then
establishing s̄obs (G) & s̄ (G).
For the first step, we follow [Mat14, FHL08, BLR10]. The optimization (1.5) is a linear program,
and the dual optimization is a maximum multi-flow problem in G. (See Section 3 for a detailed
discussion and statement of the duality theory.) Matoušek shows that if Ĝ is a string graph, then
a multi-flow with small vertex congestion in Ĝ can be used to construct a multi-flow in a related
planar graph G that has low vertex congestion in the ` 2 sense. This element of the proof is crucial and
ingenious; the reduction from a string graph Ĝ to a planar graph G does not preserve congestion in
the more standard `∞ sense.
Our work in [BLR10] shows that such a multi-flow with small `2 congestion cannot exist, and
thus one concludes that there is no low-congestion flow in Ĝ, providing a lower bound on s̄ (G) via
LP duality. In Section 3, we extend this argument to rigs over K h -minor-free graphs using the flow
crossing framework of [BLR10].
Spread vs. observable spread. Our major departure from [Mat14] comes in the second step:
Rounding a fractional separator to an integral separator by establishing that s̄obs (G) > C h · s̄ (G)
when G is a rig over a K h -minor-free graph. Matoušek used the following result that holds for any
finite metric space. It follows easily from the arguments of [Bou85] or [LR99] (see also [Mat02, Ch.
15]).
Theorem 1.10. For any finite metric space (X, d) with |X| > 2, it holds that
sobs (X, d) >
s (X, d)
.
O(log |X|)
In particular, for any graph G on n > 2 vertices,
s̄obs (G) >
s̄ (G)
O(log n)
.
Instead of using the preceding result, we employ the graph partitioning method of Klein, Plotkin,
and Rao [KPR93]. Those authors present an iterative process for repeatedly partitioning a metric
graph G until the diameter of the remaining components is bounded. If the partitioning process
fails, they construct a K h minor in G.
Since rigs over K h -minor-free graphs do not necessarily exclude any minors, we need to construct
a different sort of forbidden structure. This is the role that Lemma 1.5 plays in Section 4. In order
for the argument to work, it is essential that we construct induced partitions: We remove a subset of
the vertices which induces a partitioning of the remainder into connected components.
After constructing a suitable random partition of G, standard methods from metric embedding
theory allow us to conclude in Theorem 2.4 that if G is a rig over some K h -minor-free graph, then
s̄obs (G) >
6
s̄ (G)
.
O(h 2 )
1.2
Eigenvalues and L 2 -extremal spread
In Section 5.1, we show how the methods presented here can be used to control eigenvalues of the
discrete Laplacian on rigs. Consider the linear space VG { f : VG → }. Let L G : VG → VG be
the symmetric, positive semi-definite linear operator given by
L G f (v)
Õ
( f (v) − f (u)) .
u:{u,v}∈EG
Let 0 λ0 (G) 6 λ1 (G) 6 · · · 6 λ |VG |−1 (G) denote the spectrum of L G .
Define the L p -extremal spread of G as
(
s̄ p (G)
max
ω:VG →+
Õ
1
distω (u, v) : kωkL p (VG ) 6 1
|VG | 2 u,v∈V
)
.
(1.7)
G
In [BLR10], the L 2 -extremal spread is used to give upper bounds on the first non-trivial
eigenvalue of graphs that exclude a fixed minor. In [KLPT11], a stronger property of conformal
metrics is used to bound the higher eigenvalues as well. Roughly speaking, to control the kth
eigenvalue, one requires a conformal metric ω : VG → + such that the spread on every subset of
size > |VG |/k is large. Combining their main theorems with the methods of Section 2 and Section 3,
we prove the following theorem in Section 5.1.
Theorem 1.11. Suppose that G ∈ rig(G0 ) and G0 excludes K h as a minor for some h > 3. If dmax is the
maximum degree of G, then for any k 1, 2, . . . , |VG | − 1, it holds that
2
λ k (G) 6 O(dmax
h 6 log h)
k
.
|VG |
In particular, the bound on λ1√
(G) shows that if dmax (G) 6 O(1), then recursive spectral
partitioning (see [ST07]) finds an O( n)-vertex balanced separator in G.
1.3
Additional applications
Treewidth approximations. Bounding s̄obs (G) for rigs over K h -minor-free graphs leads to some
additional applications. Combined with the rounding algorithm implicit in Theorem 1.7 (and
explicit in [FHL08]), this yields an O(h 2 )-approximation algorithms for the vertex uniform Sparsest
Cut problem. In particular, it follows that if G ∈ rig(G0 ) and G0 excludes K h as a minor, then there is
a polynomial-time algorithm that constructs a tree decomposition of G with treewidth O(h 2 tw(G)),
where tw(G) is the treewidth of G. This result appears new even for string graphs. We refer to
[FHL08].
Lipschitz extension. The padded decomposability result of Section 2.2 combines with the Lipschitz
extension theory of [LN05] to show the following. Suppose that (G, ω) is a conformal graph, where
G is a rig over some K h -minor free graph. Then for every Banach space Z, subset S ⊆ VG , and
L-Lipschitz mapping f : S → Z, there is an O(h 2 L)-Lipschitz extension f˜ : VG → Z with f˜|S f .
See [MM16] for applications to flow and cut sparsifiers in such graphs.
1.4
Preliminaries
We use the notation + [0, ∞) and + ∩ + .
All graphs appearing in the paper are finite and undirected unless stated otherwise. If G is
a graph, we use VG and EG for its edge and vertex sets, respectively. If S ⊆ VG , then G[S] is the
7
induced subgraph on S. For A, B ⊆ VG , we use the notation EG (A, B) for the set of all edges with
one endpoint in A and the other in B. Let NG (A) A ∪ ∂A denote the neighborhood of A in G. We
write GÛ for the graph that arises from G by subdividing every edge of G into a path of length two.
If (X, dist) is a pseudo-metric space and S, T ⊆ X, we write dist(x, S) inf y∈S dist(x, y) and
dist(S, T) infx∈S, y∈T dist(x, y).
Finally, we employ the notation A . B to denote A 6 O(B), which means there exists a universal
(unspecified) constant C > 0 for which A 6 C · B.
2
Vertex separators and conformal graph metrics
The following result is standard. Recall the definition of the vertex expansion (1.2).
Lemma 2.1. Suppose that every induced subgraph H of G satisfies φ H 6
separator of size at most α.
α
|VH | .
Then G has a 23 -balanced
Thus in the remainder of this section, we focus on bounding φ G . We remark on one other basic
fact: Consider a graph G, a partition VG A ∪ B and a subset S ⊆ V. If EG (A \ S, B \ S) ∅, then
|S| > φ G
Indeed, suppose that |A| 6 |B|. Then, φ G 6
2.1
|A| · |B|
.
|VG |
|S|
|A∪S|
6
|S| |VG |
|A| |B|
(2.1)
.
Conformal graphs
A conformal graph is a pair (G, ω) where G is a connected graph and ω : VG → + . Associated
ω(u)+ω(v)
to (G, ω), we define a distance function distω as follows. Assign a length
to every
2
{u, v} ∈ EG ; then distω is the induced shortest-path metric. For U ⊆ VG , we define diamω (U)
supu,v∈U distω (u, v).
The extremal L 1 -spread is a linear programming relaxation of the optimization (1.2) defining
φ G (up to universal constant factors). In Section 3.3, we establish the following result.
Theorem 2.2. If Ĝ is a connected graph and Ĝ ∈ rig(G) for some graph G that excludes a K h minor, then
s̄1 (Ĝ) & p
n
h m log h
,
(2.2)
where n |VĜ | and m |EĜ |.
Recall that Theorem 2.2 completes the first step of our program for exhibiting small separators.
For the second step, we need to relate s̄1 (G) to s̄obs (G). Before that, we restate the proof of Theorem 1.7
from [FHL08] in our language. (The proof presented below is also somewhat simpler, as it does not
employ Menger’s theorem as in [FHL08].)
Theorem 2.3 (Restatement of Theorem 1.7). For any connected graph G, it holds that
1
2 s̄obs (G)
6
1
6 3 s̄obs (G) .
φG
(2.3)
Proof. The right-hand inequality is straightforward: Suppose that U ⊆ VG witnesses φ G |∂U |/|U |.
Let Ū VG \ U and U ◦ U \ ∂U. Let n |VG | and s |∂U |. Partition ∂U S1 ∪ S2 so that
|S1 | ds/2e and S2 bs/2c.
8
Now let ω 1∂U and define two maps
−1
2
−1
2
v ∈ U◦
v ∈ ∂U
v ∈ Ū
f1 (v) 0
1
2
v ∈ S1
v < ∂U
v ∈ S2 .
f2 (v) 0
1
2
Since ∂U separates U ◦ from Ū, the maps f1 , f2 : (VG , distω ) → are 1-Lipschitz, and
Õ
| f1 (u) − f1 (v)| 2|U ◦ | · | Ū | + |∂U |(|U ◦ | + | Ū |)
u,v∈VG
Õ
| f2 (u) − f2 (v)| 2|S1 | · |S2 | + |∂U | |U ◦ | + | Ū | .
u,v∈VG
If |∂U | > |U ◦ |, then the map f2 satisfies
satisfies
1
|VG | 2
Í
x, y∈VG
1
|VG | 2
1
2φ G .
| f1 (x) − f1 (y)| >
Í
x, y∈VG
| f2 (x) − f2 (y)| >
1
3φ G .
Otherwise, the map f1
In either case, we have shown that s̄obs (G) >
1
3φ G .
Left-hand inequality. Now we establish the more interesting bound. Suppose that (G, ω) is a
conformal metric with kωkL1 (VG ) 6 1 and f : (VG , distω ) → is a 1-Lipschitz mapping with
Õ
1
| f (x) − f (y)| s̄obs (G) .
|VG | 2 x, y∈V
G
For θ ∈ , define the three sets
A θ v ∈ VG : f (v) 6 θ
1
S θ v ∈ VG : | f (v) − θ| 6 ω(v)
2
B θ v ∈ VG : f (v) > θ .
Observe that if u ∈ A θ \ S θ and v ∈ B θ \ S θ , then f (u) 6 θ − 21 ω(u) and f (v) > θ + 12 ω(v),
ω(u)+ω(v)
therefore | f (u) − f (v)| >
. But if {u, v} ∈ EG , then since f is 1-Lipschitz, it should hold
2
ω(u)+ω(v)
that | f (u) − f (v)| 6 distω (u, v)
. We conclude that EG (A θ \ S θ , B θ \ S θ ) ∅. Therefore
2
(2.1) yields
φG
|S θ | dθ >
|V
G|
∫
∫
|A θ | · |B θ | dθ
φG Õ
φ G |VG |
| f (x) − f (y)|
s̄obs (G) .
2|VG |
2
x,y∈VG
On the other hand,
∫
|S θ | dθ 6
Õ∫
v∈VG
1B( f (v),ω(v)/2) (θ) dθ
Õ
ω(v) |VG | · kωkL1 (VG ) ,
v∈VG
where we have used the notation B(x, r) { y ∈ : |x − y| 6 r}.
In the next section, we prove the following theorem (though the main technical arguments
appear in Section 4).
Theorem 2.4. If Ĝ ∈ rig(G) and G excludes a K h minor, then
s̄1 (Ĝ) 6 O(h 2 ) s̄obs (Ĝ) .
9
(2.4)
Combining Lemma 2.1 with Theorem 2.2, Theorem 2.3, and Theorem 2.4 yields a proof of
Theorem 1.1. Indeed, suppose that Ĝ ∈ rig(G) and G excludes a K h minor. Let n |VĜ | and
m |EĜ |. Then,
√
(2.3)
(2.4) h 2 (2.2) h 3 m log h
2
φ Ĝ 6
.
.
,
n
s̄obs (Ĝ)
s̄1 (Ĝ)
completing the proof in light of Lemma 2.1.
2.2
Padded partitions and random separators
Let (X, d) be a finite metric space. For x ∈ X and R > 0, define the closed ball
B d (x, R) { y ∈ X : d(x, y) 6 R} .
If P is a partition of X and x ∈ X, we write P(x) for the set of P containing x.
Say that a partition P is ∆-bounded if S ∈ P ⇒ diam(S) 6 ∆.
Definition 2.5. A random partition P of X is (α, ∆)-padded if it is almost surely ∆-bounded and for
every x ∈ X,
B d x, ∆α ⊆ P(x) > 12 .
The following result is essentially contained in [Rab08] (see also [BLR10, Thm 4.4]). We recall
the argument here since the exact statement we need has not appeared.
Lemma 2.6. Let (X, d) be a finite metric space. If (X, d) admits an (α, s (X, d)/4)-padded partition, then
s (X, d) 6 16α · sobs (X, d) .
The proof breaks into two cases whose conjunction yields Lemma 2.6.
Lemma 2.7. If there is an x 0 ∈ X such that |B d (x 0 , s (X, d)/4)| > 21 |X|, then sobs (X, d) > 14 s (X, d).
Proof. Let B B d (x 0 , s (X, d)/4). Define f : X → by f (x) d(x, B). This map is 1-Lipschitz, and
moreover
Õ
| f (x) − f (y)| >
x,y∈X
Õ
x∈B
y <B
|X| Õ
|X| Õ
s (X, d)
d(y, B) >
d(y, B) >
d(y, x 0 ) −
.
2
2
4
y∈X
y∈X
On the other hand,
Õ
x, y∈X
d(x, y) 6
Õ
d(x, x0 ) + d(y, x 0 ) 2|X|
x, y∈X
Õ
d(y, x 0 ) .
y∈X
Combining the two preceding inequalities yields
sobs (X, d) >
Õ
s (X, d)
s (X, d)
1 Õ
1
+
.
|
f
(x)
−
f
(y)|
>
−
d(x, y)
2
2
8
4
|X| x,y∈X
4|X| x, y∈X
Lemma 2.8. If |B d (x, s (X, d)/4)| < 12 |X| for all x ∈ X, then sobs (X, d) >
10
1
16α s (X, d).
Proof. Let P be an (α, s (X, d)/4)-padded partition. Let σ : P → {0, 1} be a map chosen uniformly
at random conditioned on P. Define S {x ∈ X : σ(P(x)) 1} and F : X → by F(x) d(x, S).
Note that F is almost surely 1-Lipschitz. Moreover, observe that
B d (x, s (X, d)/4α) ⊆ P(x) ∧ σ(x) 0 ⇒ d(x, S) >
s (X, d)
4α
.
(2.5)
Therefore if x, y ∈ X satisfy d(x, y) > s (X, d)/4, then (2.5) and independence yields
By assumption,
|F(x) − F(y)| > [σ(P(x)) , σ(P(y))] ·
(x, y) : d(x, y) > s (X, d)/4
sobs (X, d) >
1 s (X, d)
.
2 4α
> 12 |X| 2 , hence
1 Õ
|X| 2 x,y∈X
|F(x) − F(y)| >
s (X, d)
16α
.
In order to produce a padded partition, we will construct an auxiliary random object. Let (G, ω)
be a conformal graph. Define the skinny ball: For c ∈ VG and R > 0,
1
Bω (c, R) v ∈ VG : distω (c, v) < R − ω(v) ,
2
Say that a random subset S ⊆ V is an (α, ∆)-random separator if the following two conditions
hold:
1. For all v ∈ VG and R > 0,
[Bω (v, R) ∩ S ∅] > 1 − α
R
.
∆
2. Almost surely every connected component of G[V \ S] has diameter at most ∆ (in the metric
distω ).
Lemma 2.9. If (G, ω) admits an (α, ∆)-random separator, then (VG , distω ) admits an (8α, ∆)-padded
partition.
Proof. The random partition P is defined by taking all the connected components of G[V \ S], along
with the single sets {{x} : x ∈ S}. The fact that P is almost surely ∆-bounded is immediate.
∆
Set R 2α
and observe that for every v ∈ VG , Bω (v, R) ∩ S ∅ ⇒ Bω (v, R) ⊆ P(x), since
Bω (v, R) is a connected set in G.
Moreover, Bdistω (v, R/4) ⊆ Bω (v, R). To see this, observe that
max
x∈Bdistω (v,R/4)\{v}
and thus
ω(x) 6
R
2
,
x ∈ Bdistω (v, R/4) ⇒ distω (v, x) < R − 12 ω(x) ⇒ x ∈ Bω (v, R) .
It follows that for every v ∈ VG , we have
∆
Bdistω (v G , 8α
) ⊆ P(x) >
1
2
,
completing the proof that P is (8α, ∆)-padded.
This following result is proved in Section 4 (see Corollary 4.3).
11
Theorem 2.10. If Ĝ ∈ rig(G) and G excludes K h as a minor, then for every ∆ > 0, every conformal metric
(G, ω) admits an (α, ∆)-random separator with α 6 O(h 2 ).
We can now prove Theorem 2.4.
Proof of Theorem 2.4. Suppose that Ĝ ∈ rig(G) and G excludes a K h minor. Let (Ĝ, ω) be a conformal
metric such that s (VĜ , distω ) s̄1 (Ĝ). Combining Theorem 2.10 and Lemma 2.9 shows that
(VĜ , distĜ ) admits an (α, s̄1 (Ĝ)/4)-random separator for some α 6 O(h 2 ). Now Lemma 2.6 shows
that s̄obs (Ĝ) > sobs (VĜ , distω ) & h12 s (VĜ , distω ) h12 s̄1 (Ĝ), completing the proof.
Remark 2.11. One advantage to introducing the auxiliary random separator S is that it can be used
to directly relate φ G and s̄ (G) without going through padded partitions. Indeed, this can be done
using the weaker property that for every v ∈ VG ,
[v ∈ S] 6 α
ω(v)
.
∆
(The stronger padding property has a number of additional applications; see Section 5 and
Section 1.3.)
We present the argument. Suppose that (G, ω) is a conformal graph with kωkL1 (VG ) 6 1 and let
s : s (VG , distω ). If there is some vertex v0 ∈ VG for which |Bdistω (v0 , s /4)| > 12 |VG |, then one can
apply Lemma 2.7 and Theorem 2.3 to obtain
φG 6
2
s̄obs (G)
6
2
sobs (VG , distω )
6
8
s
.
Suppose that no such v 0 exists. In that case, every subset U ⊆ VG with diamω (U) 6 s /4 has
|U | < |VG |/2. Let S be an (α, s /4)-random separator. Then every connected component of G[V \ S]
has at most |VG |/2 vertices. (In particular, S is a 23 -balanced separator with probability 1.) Therefore
by linearity of expectation,
8kωkL1 (VG )
8α v∈VG ω(v)
[|S|]
8α
φG 6
6
α
6
.
|VG |/2
|VG |s
s
s
Í
3
Multi-flows, congestion, and crossings
Let G be an undirected graph, and let PG denote the the set of all paths in G. Note that we allow
length-0 paths consisting of a single vertex. For vertices u, v ∈ VG , we use PGuv ⊆ PG for the
subcollection of u-v paths. A multi-flow in G is a map Λ : PG → + . We will use the terms “flow”
and “multi-flow” interchangeably.
Define the congestion map cΛ : VG → + by
Õ
cΛ (v)
Λ(γ) .
γ∈PG :v∈γ
For u, v ∈ VG , we denote the total flow sent between u and v by
Λ[u, v]
Õ
Λ(γ) .
γ∈PGuv
For an undirected graph H, an H-flow in G is a pair (Λ, ϕ) that satisfies the following conditions:
1. Λ is a flow in G
12
2. ϕ : VH → VG
3. For every u, v ∈ VG ,
Λ[u, v] # {x, y} ∈ EH : {ϕ(x), ϕ(y)} {u, v} .
If the map ϕ is injective, we say that (Λ, ϕ) is proper. Say that (Λ, ϕ) is integral if {Λ(γ) : γ ∈ PG } ⊆ + .
3.1
Duality between conformal metrics and multi-flows
For p ∈ [1, ∞], define the ` p -vertex congestion of G by
vp (G) min kcΛ k` p (VG )
Λ
VG
: Λ[u, v] 1 ∀{u, v} ∈
2
,
where the minimum is over all flows in G.
The next theorem follows from the strong duality of convex optimization; see [BLR10, Thm. 2.2]
which employs Slater’s condition for strong duality (see, e.g., [BV04, Ch. 5]).
Theorem 3.1 (Duality theorem). For every G, it holds that if (p, q) is a pair of dual exponents, then
1
vp (G) |VG | 2− q · s̄ q (G) .
We will only require the case p ∞, q 1, except in Section 5.1 where the p q 2 case is
central.
3.2
Crossing congestion and excluded minors
Now we define the crossing congestion of a flow: If (Λ, ϕ) is an H-flow in G, denote
χG (Λ, ϕ)
Õ
Õ
{u,v},{u 0 ,v 0 }∈EH
|{u,v,u 0 ,v 0 }|4
Õ
ϕ(u)ϕ(v)
γ∈PG
Λ(γ)Λ(γ0)1{γ∩γ0,∅} .
ϕ(u0 )ϕ(v 0 )
γ0 ∈PG
It provides a lower bound on the ` 2 congestion of Λ, as clearly
Õ
v∈VG
Define
cΛ (v)2
Õ
Õ
Õ
Λ(γ)Λ(γ0) >
Λ(γ)Λ(γ0)1{γ∩γ0,∅} > χG (Λ, ϕ) .
γ,γ0 ∈PG
v∈VG γ,γ0 ∈PG :
v∈γ∩γ0
∗
χG
(H) inf χG (Λ, ϕ) ,
(Λ,ϕ)
where the infimum is over all H-flows in G. Define also
†
χG
(H) min χG (Λ, ϕ) ,
(Λ,ϕ)
where the infimum is over all integral H-flows in G. The next lemma offers a nice property of
crossing congestion: The infimum is always achieved by integral flows.
Lemma 3.2. For every graph H, it holds that
∗
†
χG
(H) χG
(H) .
13
Proof. Given any H-flow (Λ, ϕ), define a random integral flow Λ† as follows: For every edge
ϕ(u)ϕ(v)
Λ(γ)
{u, v} ∈ EH , independently choose a path γ ∈ PG
with probability Λ[ϕ(u),ϕ(v)] and let Λ† (γ)
be equal to the number of edges of EH that choose the path γ. (For all paths γ not selected in such a
manner, Λ† (γ) 0.) Independence and linearity of expectation yield [χG (Λ† , ϕ)] χG (Λ, ϕ).
The next result relates the topology of a graph to crossing congestion; it appears as [BLR10, Lem
3.2].
Lemma 3.3. If H is a bipartite graph with minimum degree 2 and (Λ, ϕ) is an H-flow in G with χG (Λ, ϕ) 0,
then G has an H-minor.
The preceding lemma allows one to use standard crossing number machinery to arrive at the
following result (see [BLR10, Thm 3.9–3.10]).
Theorem 3.4. For every h > 2, the following holds: If G excludes K h as a minor, then for any N > 4h,
N4
.
h3
√
Moreover, there is a constant K > 0 such that if N > Kh log h, then
∗
χG
(K N ) &
∗
(K N ) &
χG
3.3
N4
.
h 2 log h
Vertex congestion in rigs
We now generalize Matoušek’s argument to prove the following theorem.
Theorem 3.5. For any graph G and Ĝ ∈ rig(G),
∗
χG
(K |VĜ | ) 6 (4|EĜ | + |VĜ |) v∞ (Ĝ)2 .
Before moving to the proof, we state the main result of this section. It follows immediately from
the conjunction of Theorem 3.5 and Theorem 3.4. (One does not require a lower bound on n as in
Theorem 3.4 because the bound v∞ (Ĝ) > |VĜ | always holds.)
Corollary 3.6. Suppose Ĝ is a connected graph and Ĝ ∈ rig(G) for some graph G that excludes K h as a
minor. If n |VĜ | and m |EĜ |, then
r
v∞ (Ĝ) &
∗
(K n )
χG
m
In particular, Theorem 3.1 yields
s̄1 (Ĝ) & p
& p
n2
h m log h
n
h m log h
.
.
Proof of Theorem 3.5. Let {R u : u ∈ VĜ } be a set of regions realizing Ĝ over G. For every path γ in Ĝ,
we specify a path γ̌ in G. For each v ∈ VĜ , fix some distinguished vertex v̌ ∈ R v .
Suppose that γ (v1 , v 2 , . . . , v k ). Let γ̌ be any path γ̌ γ̌1 ◦ γ̌2 ◦ · · · ◦ γ̌k which starts at vˇ1 , ends
at vˇk , and where for each i 1, . . . , k, the entire subpath γ̌i is contained in R v i . This is possible
because each R v i is connected and {v i , v i+1 } ∈ EĜ implies that R v i and R v i+1 share at least one vertex
of G. We describe the path γ̌ as “visiting” the regions R v1 , R v2 , · · · , R v k in order.
Let n |VĜ | and m |EĜ |. Let (Λ, ϕ) be a proper K n -flow in Ĝ achieving kcΛ k∞ v∞ (Ĝ). The
path mapping γ 7→ γ̌ sends (Λ, ϕ) to a (possibly improper) K n -flow (Λ̌, ϕ̌) in G. Establishing the
following claim will complete the proof of Theorem 3.5.
14
Claim 3.7. It holds that
χG (Λ̌, ϕ̌) 6
Õ
cΛ (u)2 +
u∈VĜ
Õ
2
(cΛ (u) + cΛ (v))2 6 (4m + n)kcΛ k∞
.
(3.1)
{u,v}∈EĜ
We prove the claim as follows: If γˇ1 and γˇ2 intersect in G, we charge weight Λ(γ1 )Λ(γ2 ) to
some element in VĜ ∪ EĜ . If γˇ1 visits the regions R u1 , R u2 , · · · , R u k1 and γˇ2 visits the regions
R v1 , R v2 , · · · , R v k2 and γˇ1 ∩ γˇ2 , ∅, then they meet at some vertex x ∈ R u i ∩ R v j . If u i v j , we
charge this crossing to u i ∈ VĜ . Otherwise we charge this crossing to the edge {u i , v j } ∈ EĜ .
If u ∈ VĜ is charged by (γ1 , γ2 ), then u ∈ γ1 ∩ γ2 . Thus the total weight charged to u is at most
Õ
Λ(γ)Λ(γ0) 6 cΛ (u)2 .
γ , γ0 ∈PĜ :u∈γ
Similarly, if {u, v} ∈ EĜ is charged by (γ1 , γ2 ), then {u, v} ⊆ γ1 ∪ γ2 , thus the total weight charged
to {u, v} at most (cΛ (u) + cΛ (v))2 . Since all of the weight contributing to χG (Λ̌, ϕ̌) has been charged,
this yields the desired claim.
4
Careful minors and random separators
For graphs H and G, one says that H is a minor of G if there are pairwise-disjoint connected
subsets {A u ⊆ VG : u ∈ VH } such that {u, v} ∈ EH ⇒ EG (A u , A v ) , ∅. We will sometimes
refer to the sets {A u } as supernodes. Say that H is a strict minor of G if the stronger condition
{u, v} ∈ EH ⇐⇒ EG (A u , A v ) , ∅ holds. Finally, we say that H is a careful minor of G if HÛ is a strict
minor of G.
The next result explains the significance of careful minors for region intersection graphs. We
prove it in the next section.
Lemma 4.1. If Ĝ ∈ rig(G) and Ĝ has a careful H-minor, then G has an H-minor.
We now state the main result of this section; its proof occupies Sections 4.2–4.4.
Theorem 4.2. For any h > 1, the following holds. Suppose that G excludes a careful K h minor. Then there
is a number α 6 O(h 2 ) such that for any ω : VG → + and ∆ > 0, the conformal graph (G, ω) admits an
(α, ∆)-random separator.
Applying Lemma 4.1 immediately yields the following.
Corollary 4.3. Suppose that G excludes a K h minor and Ĝ ∈ rig(G). Then there is a number α 6 O(h 2 )
such that for any ω : VĜ → + and ∆ > 0, the conformal graph (Ĝ, ω) admits an (α, ∆)-random separator.
The proof of Theorem 4.2 is based on a procedure that iteratively removes random sets of
vertices from the graph in rounds. It is modeled after the argument of [FT03] which is itself based
on [KPR93]. For an exposition of the latter argument, one can consult the book [Ost13, Ch. 3].
4.1
Careful minors in rigs
The next lemma clarifies slightly the structure of careful minors.
Lemma 4.4. G has a careful H-minor if and only if there exist pairwise-disjoint connected subsets
Ð
{B u ⊆ VG : u ∈ VH } and distinct vertices W w x y ∈ VG \ u∈VH B u : {x, y} ∈ EH such that
1. EG (B u , B v ) ∅ for u, v ∈ VH with u , v.
15
2. W is an independent set.
3. For every {x, y} ∈ EH , it holds that
EG (w x y , B u ) , ∅ ⇐⇒ u ∈ {x, y} .
(4.1)
Proof. The “only if” direction is straightforward. We now argue the other direction.
Û
Let {A u ⊆ VG : u ∈ VHÛ } witness a strict H-minor
in G. For every {x, y} ∈ EH , there exists a
simple path γx y with one endpoint in A x , one endpoint in A y , and whose internal vertices satisfy
γx◦ y ⊆ A m x y and γx◦ y , ∅, where m x y ∈ VHÛ is the vertex subdividing the edge {x, y}.
Choose some vertex w x y ∈ γx◦ y . Removal of w x y breaks the graph G[A x ∪ A y ∪ γx y ] into two
connected components; define these as B x and B y (so that A x ⊆ B x and A y ⊆ B y ). Property
(1) is verified by strictness of the HÛ minor and the fact that the the non-subdivision vertices
Û Similarly, properties (2) and (3) follow
VHÛ \ m x y : {x, y} ∈ VH form an independent set in H.
Û
from strictness of the H minor and the fact that NHÛ (m x y ) {x, y} for {x, y} ∈ EH .
We now prove that if Ĝ ∈ rig(G), then careful minors in Ĝ yield minors in G.
Proof of Lemma 4.1. Let {R v ⊆ VG : v ∈ VĜ } be a set of regions realizing Ĝ. Assume that Ĝ has a
careful H-minor and let {B u ⊆ VĜ : u ∈ VH } and W {w x y : {x, y} ∈ EH } be the sets guaranteed
by Lemma 4.4.
For u ∈ VH , define
Ø
Au
Rv .
v∈B u
Since B u is connected in Ĝ and the regions {R v : v ∈ VĜ } are each connected in G, it follows that A u
is connected in G.
Let us verify that the sets {A u : u ∈ VH } are pairwise disjoint. If x ∈ A u ∩ A v for u , v, then
there must be regions R a and R b with a ∈ B u , b ∈ B v and x ∈ R a ∩ R b . This would imply {a, b} ∈ EĜ ,
but Lemma 4.4(1) asserts that EĜ (B u , B v ) ∅.
We will show that there exist pairwise vertex-disjoint paths γuv ⊆ VG : {u, v} ∈ EH with
!
γuv ∩
Ø
Ax ⊆ Au ∪ Av ,
(4.2)
x∈VH
and such that γuv connects A u to A v . This will yield the desired H-minor in G.
Fix {u, v} ∈ EH . From Lemma 4.4(3), we know that the connected set R w uv shares a vertex with
A u and also shares a (different) vertex with A v . Thus we can choose γuv as above with γuv ⊆ R w uv .
Note that Lemma 4.4(2) (in particular, (4.1)) also yields R w uv ∩ A x ∅ for any x ∈ VH \ {u, v},
verifying (4.2).
Thus we are left to verify that the sets {R w uv : {u, v} ∈ EH } are pairwise vertex-disjoint. But this
also follows from (4.1), specifically the fact that W {w uv : {u, v} ∈ EH } is an independent set in
Ĝ.
4.2
Chopping trees
Observe that by a trivial approximation argument, it suffices to prove Theorem 4.2 for any conformal
metric ω : VG → (0, ∞), i.e., one that satisfies
ω(v) > 0
∀v ∈ VG
(4.3)
Let us now fix such a conformal metric ω : VG → (0, ∞) on G and a number ∆ > 0. Fix an
arbitrary ordering v 1 , v 2 , . . . , v |VG | of VG in order to break ties in the argument that follows.
16
(a) Illustration of a fat sphere
(b) Chopping a graph into subgraphs
Figure 2: The chopping procedure
We will use I(G) to denote the collection of all connected, induced subgraphs of G. For such a
subgraph H ∈ I(G), we use distH
ω to denote the induced distance coming from the conformal metric
(H, ω|VH ). For c ∈ VH and R > 0, let us define the skinny ball, fat ball, and fat sphere, respectively:
BωH (c, R)
BH
ω (c, R)
v ∈ VH :
v ∈ VH :
distH
ω (c, v)
1
< R − ω(v) ,
2
distH
ω (c, v)
1
6 R + ω(v) ,
2
H
H
1
1
SH
ω (c, R) v ∈ VH : R ∈ distω (c, v) − 2 ω(v), distω (c, v) + 2 ω(v)
H
BH
ω (c, R) \ Bω (c, R) .
See Figure 2(a) for a useful (but non-mathematical) illustration where one imagines a vertex v ∈ VH
as a disk of radius 12 ω(v). Note that BωH (c, R) is the connected component of c in the graph
H[VH \ SH
ω (c, R)]. The next fact requires our assumption (4.3).
Fact 4.5. If γ ⊆ VH is a distH
ω -shortest path emanating from c ∈ VH , then for every R > 0, it holds that
|γ ∩ SH
(c,
R)|
6
1.
ω
For c ∈ VH and τ ∈ [0, ∆], let cut∆ (H, c; τ) k∈+ SH
ω (c, τ + k∆). We define chop∆ (H, c; τ) as
the collection of connected components of the graph H [VH \ cut∆ (H, c; τ)]. See Figure 2(b).
The next lemma is straightforward.
Ð
Lemma 4.6. If τ ∈ [0, ∆] is chosen uniformly at random, then for every v ∈ VH and R > 0,
[BωH (v, R) ∩ cut∆ (H, c; τ) ∅] > 1 −
2R
.
∆
A ∆-chopping tree of (G, ω) is a rooted, graph-theoretic tree T (σ) for some σ : I(G) → [0, ∆]. The
nodes of T (σ) are triples (H, c, j) where H ∈ I(G), c ∈ VH , and j ∈ + . We refer to c as the center of
the node and j as its depth. We now define T (σ) inductively (by depth) as follows.
The root of T T (σ) is (G, v 1 , 0). For a node λ (H, c, j) of T , we let c®T (λ) denote the
sequence of centers encountered on the path from λ to the root of T , not including λ itself. If
chop∆ (H, c; σ(H)) ∅, then λ has no children.
17
Otherwise, if chop∆ (H, c; σ(H)) {Hi : i ∈ I}, the children of (H, c, j) are {(Hi , c i , j + 1)}, where
®T (λ) ∪ {c}) .
c i argmaxx∈VH distG
ω (x, c
i
(4.4)
In other words, c i is chosen as the point of VHi that is furthest from the centers of its ancestors in
the ambient metric distG
ω . For concreteness, if the maximum in (4.4) is not unique, we choose the first
vertex (according to the ordering of VG ) that achieves the maximum.
A final definition: We say that a node λ (H, c, j) of a chopping tree T is β-spaced if the value
of the maximum in (4.4) is at least β, i.e.,
®T (λ)) > β .
distG
ω (c, c
Note that the nodes in each level of T (σ) correspond to the connected components that result
after removing a subset of nodes from G. We state the following consequence.
Lemma 4.7. Suppose that T (σ) is a ∆-chopping tree for some ∆ > 0. Consider an integer k > 0, and let
{Hi : i ∈ I} ⊆ I(G) denote the collection of induced subgraphs occuring
Ð in the
depth-k nodes of T (σ). Then
each Hi is a unique connected component in the induced graph G
i∈I VHi .
We now state the main technical lemma on chopping trees. The proof appears in Section 4.4.
Lemma 4.8. Consider any h > 1 and ∆ > 0. Assume the following conditions hold:
1. maxv∈VG ω(v) 6 ∆
2. T is a ∆-chopping tree of (G, ω).
3. There exists a 21h∆-spaced node of T at depth h − 1.
Then G contains a careful K h minor.
Finally, we have the following analysis of a random chopping tree.
Lemma 4.9. For any k > 1, the following holds. Suppose that σ : I(G) → [0, ∆] is chosen uniformly at
random. Let {Hi : i ∈ I} ⊆ I(G) denote the collection of induced subgraphs occurring in the depth-k nodes of
T (σ). For any v ∈ VG ,
"
#
BωG (v, R) ⊆
Ø
VHi > 1 − 2k
i∈I
R
.
∆
Proof. Note that since BωG (v, R) is a connected set and there are no edges between Hi and H j for
i , j (cf. Lemma 4.7), we have
BωG (v, R) ⊆
Ø
VHi ⇐⇒ ∃i ∈ I s.t. BωG (v, R) ⊆ VHi .
i∈I
The set BωG (v, R) experiences at most k random chops, and the probability it gets removed in any
one of them is bounded in Lemma 4.6. The desired result follows by observing that if H ∈ I(G)
satisfies BωG (v, R) ⊆ VH , then BωH (v, R) BωG (v, R).
18
4.3
The random separator construction
We require an additional tool before proving Theorem 4.2. For nodes that are not well-spaced, we
need to apply one further operation.
If H ∈ I(G) and c® (c 1 , c2 , . . . , c k ) ∈ VGk and τ® (τ1 , τ2 , . . . , τk ) ∈ k+ , we define a subset
®) ⊆ I(G) as follows. Define
shatter∆ (H, c®, τ
®) VH ∩
shards∆ (H, c®, τ
k
Ø
SGω (c i , ∆ + τi ) ,
i1
and let shatter∆ (H, c®, τ®) be the collection of connected components of H[VH \ shards(H, c®, τ®)]. The
next two lemmas are straightforward consequences of this construction.
0
Lemma 4.10. If every v ∈ VH satisfies min{distG
ω (v, c i ) : i 1, . . . , k} 6 ∆, then for every H ∈
®), it holds that
shatter∆ (H, c®, τ
®) .
diamG
ω (VH 0 ) 6 2(∆ + max τ
Lemma 4.11. For any c® ∈ VGk and any ∆0 > 0, if τ® ∈ [0, ∆0]k is chosen uniformly at random, then for any
v ∈ VH and R > 0,
R
[BωH (v, R) ∩ shards∆ (H, c®, τ®) ∅] > 1 − 2k 0 .
∆
Proof of Theorem 4.2. We may assume that if v ∈ VG has ω(v) > ∆, then v ∈ S. Indeed, denote
Q {v ∈ VG : ω(v) > ∆} .
If we can produce an (α, ∆)-separator for each of the connected components of G[VG \ Q], then
taking the union of those separators together with Q yields a (2α, ∆)-separator of G. We may
therefore assume that maxv∈VG ω(v) 6 ∆.
Assume now that G excludes a careful K h minor. Let T T (σ) be the ∆-chopping tree of (G, ω)
with σ : I(G) → [0, ∆] chosen uniformly at random.
Let Dh−1 {λ i (Hi , c i , h) : i ∈ I} be the collection of depth-(h − 1) nodes of T (σ). Let
Ð
S1 VG \ i∈I VHi . By construction, the graphs {Hi } are precisely the connected components of
G[V \ S1 ] (and they occur without repetition, i.e., Hi , H j for i , j).
By Lemma 4.9, for any v ∈ VG and R > 0, we have
[BωG (v, R) ∩ S1 ∅] > 1 − 2h
Define
S2
Ø
R
.
∆
(4.5)
®) ,
shards21h∆ (Hi , c®T (λ i ); τ
i∈I
where τ® ∈ [0, ∆]h is chosen uniformly at random. From Lemma 4.11, for any i ∈ I, v ∈ VHi , and
R > 0, we have
R
[BωHi (v, R) ∩ S2 ∅] > 1 − 2h .
(4.6)
∆
So consider v ∈ VG and R > 0. If BωG (v, R) ∩ S1 ∅, then BωG (v, R) ⊆ VHi for some i ∈ I, and in
that case BωG (v, R) BωHi (v, R). Therefore (4.5) and (4.6) together yield
[BωG (v, R) ∩ (S1 ∪ S2 ) ∅] > 1 − 4h
Moreover, the collection of induced subgraphs
H
Ø
®)
shatter21h∆ (Hi , c®T (λ i ); τ
i∈I
19
R
.
∆
(4.7)
is precisely the set of connected components of G[V \ (S1 ∪ S2 )].
We are thus left to bound diamG
ω (VH ) for every H ∈ H . Consider a node λ i ∈ Dh−1 . Since G
excludes a careful K h minor, Lemma 4.8 implies that λ i is not 21h∆-spaced. It follows that
®T (λ i )) 6 distG
®T (λ i )) 6 21h∆ ,
max distG
ω (v, c
ω (c i , c
v∈VHi
based on how c i is chosen in (4.4). Therefore Lemma 4.10 implies that for every H ∈
®), we have diamG
shatter21h∆ (Hi , c®T (λ i ); τ
ω (VH ) 6 2(21h + 1)∆.
We conclude that every connected component H of G[V \ (S1 ∪ S2 )] has diamG
ω (VH ) 6 (42h + 2)∆.
Combining this with (4.7) shows that S1 ∪S2 is a (4h(42h + 2), (42h + 2)∆)-random separator, yielding
the desired conclusion. (Note that establishing the existence of a (cα, c∆)-random separator for every
∆ > 0 implies the existence of an (α, ∆)-random separator for every ∆ > 0 by homogeneity.)
4.4
A diameter bound for well-spaced subgraphs
Our goal now is to prove Lemma 4.8.
Lemma 4.12 (Restatement of Lemma 4.8). Consider any h > 1 and ∆ > 0. Assume the following
conditions hold:
(A1) maxv∈VG ω(v) 6 ∆
(A2) T is a ∆-chopping tree of (G, ω).
(A3) There is a 21h∆-spaced node of T at depth h − 1.
Then G contains a careful K h minor.
In order to enforce the properties of a careful minor, we will need a way to ensure that there are
no edges between certain vertices. The following simple fact will be the primary mechanism.
Lemma 4.13. Suppose H ∈ I(G) and maxv∈VH ω(v) 6 ∆. If u, v ∈ VH satisfy distH
ω (u, v) > ∆, then
{u, v} < EG .
Proof. Since H is an induced subgraph, {u, v} ∈ EG
ω(u)+ω(v)
distH
6 ∆.
ω (u, v)
2
⇒
{u, v} ∈ EH . And then clearly
Proof of Lemma 4.12. We will construct a careful K h minor inductively. Recall that a careful K h minor
is a strict minor of the subdivision KÛ h . We use the notation {A u } for the supernodes corresponding
to original vertices of K h . The supernodes corresponding to subdivision vertices will be single
nodes of VG which we denote {w uv : {u, v} ∈ EK h }.
Let λ (H, c, h − 1) be a 21h∆-spaced node in T , and denote by λ λ1 , λ2 , . . . , λ h the sequence
of nodes of T on the path from λ to the root of T . For each t 1, . . . , h, write λ t (Ht , c t , h − t).
Observe that since λ is 21h∆-spaced, it also holds that λ t is 21h∆-spaced for each t 1, . . . , h.
(The property only becomes stronger for children in T .) Hence,
distG
ω (c i , c j ) > 21h∆ for all i, j ∈ {1, . . . , h} with i , j.
(4.8)
We will show, by induction on t, that Ht contains a careful K t minor for t 1, . . . , h.
For the sake of the induction, we will need to maintain some additional properties that we now
describe. The first three properties simply ensure that we have found a strict KÛ t minor. Let us use the
numbers {1,
2, . . . , t} to index the vertices of VK t . We will show there exist sets {A tu ⊆ VHt : u ∈ VK t }
and W t w uv ∈ VHt : {u, v} ∈ EK t ⊆ VHt with the following properties:
20
Figure 3: Construction of a careful K h minor.
P1. The sets {A tu : u ∈ VK t } are connected and mutually disjoint, and EG (A tu , A tv ) ∅ for u , v.
P2. The set W t is an independent set in G.
P3. For all {u, v} ∈ EK t it holds that EG (w uv , A tx ) , ∅ ⇐⇒ x ∈ {u, v}.
t
P4. For every u ∈ VK t , there is a representative r ut ∈ A tu such that distG
ω (r u , c u ) 6 10t∆.
t
t
P5. For all u , v ∈ VK t , we have distG
ω (r u , r v ) > 21(h − t)∆ .
P6. For every u ∈ VK t , it holds that
!
t
t
t
distH
ω ru , W ∪
Ø
A tv > 3∆ .
v ,u
In the base case t 1, take A11 {c1 } and r11 c 1 , and W 1 ∅. It is easily checked that these
choices satisfy (P1)–(P6). So now suppose that for some t ∈ {1, 2, . . . , h − 1}, we have objects
satisfying (P1)–(P6). We will establish the existence of objects satisfying (P1)–(P6) for t + 1.
It may help to consult Figure 3 for the inductive step. Recall that, by construction, Ht ∈
chop∆ (Ht+1 , c t+1 ; r) for some r > 0. Note that since λ t is 21h∆-spaced, it holds that distG
ω (c t , c t+1 ) >
21h∆ (cf. (4.8)), and therefore (recalling that maxv∈VG ω(v) 6 ∆)
r > 20h∆ .
(4.9)
t
t+1
t+1
For each i 1, 2, . . . , t, let γi denote a distH
denote
ω -shortest-path from r i to c t+1 . Let r i
Ht+1
the unique element in γi ∩ Sω (c t+1 , r − 4∆), and let w i,t+1 denote the unique element in γi ∩
t+1
SH
ω (c t+1 , r − 8∆). (Recall Fact 4.5.) Define:
t+1
A t+1
A ti ∪ γi \ BH
ω (c t+1 , r − 8∆),
i
A t+1
t+1
BωHt+1 (c t+1 , r
− 8∆) .
21
16i6t
Lemma 4.14. For every i, j ∈ {1, 2, . . . , t} with i , j, the following holds: γi ∩ A tj ∅, γi ∩ W t ∅, and
EG (γi , A tj ) EG (γi , W t ) ∅.
Proof. Observe that lenω (γi ∩ VHt ) 6 ∆ and γi emanates from r it . (P6) implies that
t
t
t
t
distH
ω (r i , A j ∪ W ) > 3∆ ,
and thus γi ∩ (A tj ∪ W t ) ∅ for i , j.
t+1
Note furthermore that SH
ω (c t+1 , r) separates γi \ VHt from VHt in H t+1 . Thus we need only
prove that EG (γi \ BωHt+1 (c t+1 , r), A tj ∪ W t ) ∅. But we have
Ht+1
Ht t
t
t
t
t
t+1
distH
ω (γi \ Bω (c t+1 , r), A j ∪ W ) > distω (r i , A j ∪ W ) − lenω (γi ∩ VHt ) −
1
2
max
H
v∈Sω t+1 (c t+1 ,r)
ω(v)
(A1)
t
t
t
t
> distH
ω (r i , A j ∪ W ) − 3∆/2
(P6)
> ∆.
Therefore Lemma 4.13 yields EG (γi \ BωHt+1 (c t+1 , r), A tj ∪ W t ) ∅, and this suffices to prove
EG (γi , A tj ∪ W t ) ∅ for i , j.
Let us verify the six properties (P1)–(P6) above in order.
1. Consider first the sets A t+1
for i 1, 2, . . . , t. For i , j, we have γi ∩ A tj ∅ and EG (γi , A tj ) ∅
i
from Lemma 4.14. Next consider the sets γi ∩ A t+1
and γ j ∩ A t+1
for i , j and i, j 6 t. By (P5),
i
j
G
t t
t+1
t+1
we have distG
ω (r i , r j ) > 14(h − t)∆, and thus distω (γi ∩ A i , γ j ∩ A j ) > 21(h − t)∆ − 2 · 9∆ > ∆.
Hence γi ∩ A t+1
and γ j ∩ A t+1
are disjoint and by Lemma 4.13, EG (γi ∩ A t+1
, γ j ∩ A t+1
) ∅.
i
j
i
j
t+1
Finally, observe that A t+1
∩A t+1
∅ and EG (A t+1
, A t+1
) ∅ for i 6 t because SH
ω (c t+1 , r −8∆)
t+1
t+1
i
i
Ð
separates A t+1
from i 6 t A t+1
in Ht+1 .
t+1
i
Ht+1
t+1
2. Observe that SH
ω (c t+1 , r −4∆) and S ω (c t+1 , r −8∆) are disjoint because maxv∈VHt+1 ω(v) 6 ∆
t+1
t+1 \ W t from W t in H
by assumption (A1). It follows that SH
t+1 .
ω (c t+1 , r − 4∆) separates W
t+1
t
We thus need only verify that W
\ W is an independent set.
To this end, observe that
t
distG
ω (w i,t+1 , r i ) 6 9∆ + ∆/2 6 9.5∆ ,
(4.10)
where we have again employed (A1). This implies that for i , j,
(P5)
G t t
distG
ω (w i,t+1 , w j,t+1 ) > distω (r i , r j ) − 2 · 9.5∆ > 21(h − t)∆ − 19∆ > ∆ ,
hence Lemma 4.13 implies that Wt+1 \ Wt is indeed an independent set.
, w i,t+1 ) , ∅ and EG (A t+1
, w i,t+1 ) , ∅ for i 6 t both follow immediately
3. The facts that EG (A t+1
t+1
i
from the construction (w i,t+1 is a separator vertex on the path γi connecting r it to c t+1 ).
We are left to verify that for every {u, v} ∈ VK t+1 and x < {u, v}, we have EG (w uv , A t+1
x ) ∅.
We argue this using three cases:
22
• For i 6 t, EG (w i,t+1 , A tx ) ∅:
t+1
This follows because EG (w i,t+1 , VHt ) ∅ since SH
ω (c t+1 , r) separates w i,t+1 from VHt in
Ht+1 .
• For i, j 6 t with i , j: EG (w i,t+1 , γ j ∩ A t+1
) ∅:
j
We have
(4.10)
G t t
t+1
t+1
t+1
distH
ω (w i,t+1 , γ j ∩ A j ) > distω (r i , r j ) − 9.5∆ − lenω (γ j ∩ A j )
> distGω (r it , r tj ) − 18.5∆
(P5)
> 21(h − t)∆ − 18.5∆ > ∆ ,
which implies the desired bound using Lemma 4.13.
• EG (W t ,
Ð
i 6t
γi ∩ A t+1
) ∅: This follows from Lemma 4.14.
i
4. For i 6 t, we have
G t
G t t+1
t+1
distG
ω (r i , c i ) 6 distω (r i , c i ) + distω (r i , r i ) 6 10t∆ + 9.5∆ 6 10(t + 1)∆.
t+1
Moreover, r t+1
c t+1 .
5. Similarly, for i, j 6 t and i , j,
(P5)
G t t
t+1 t+1
distG
ω (r i , r j ) > distω (r i , r j ) − 2 · 9.5∆ > 21(h − t − 1)∆ .
One also has
(P4)
(4.8)
G
G
t+1 t+1
t+1
distG
ω (r t+1 , r i ) distω (c t+1 , r i ) > distω (c t+1 , c i ) − 10t∆ > 21(h − t − 1)∆ .
t+1
6. First, note that if i 6 t, then r it+1 ∈ SH
ω (c t+1 , r − 4∆), so using (A1) gives
t+1
t+1
distH
ω (c t+1 , r i ) 6 r − 4∆ + ∆/2 .
On the other hand, (W t ∪
Ð
j 6t
t+1
A tj ) ∩ BH
ω (c t+1 , r) ∅, hence
t+1
distH
c t+1 , W t ∪
ω
Ð
t
j 6t A j > r .
t+1
t+1
t
It follows from the triangle inequality that distH
ω (r i , W ∪
Ð
j 6t
A tj ) > 3∆.
Next, we have, for i, j 6 t and i , j,
G t+1
t+1
t+1
t+1
t+1
distH
ω (r i , γ j ∩ A j ) > distω (r i , γ j ∩ A j )
> distGω (r it , r tj ) − distGω (r it , r it+1 ) − lenω (γ j ∩ A t+1
j )
> 21(h − t)∆ − 9.5∆ − 5∆ > 3∆ .
t+1
t+1 \ W ⊆ SHt+1 (c
Also note that distH
t
t+1 , r − 8∆)
ω (c t+1 , Wt+1 \ Wt ) 6 r − 8∆ + ∆/2 because W
ω
Ht+1
t+1
t+1
t+1
and distH
(c
,
r
)
>
r
−
4∆
−
∆/2
because
r
∈
S
(c
,
r
−
4∆).
It
follows
that
t+1
t+1
ω
ω
i
i
t+1
t+1
t+1
distH
\ W t ) > 3∆ .
ω (r i , W
23
t+1
t+1
t+1 ∪
We have thus verified that for i 6 t, it holds that distH
ω (r i , W
Ð
j 6 t, j , i
A t+1
) > 3∆.
j
Ht+1
t+1
t+1
t+1
t+1
The fact that distH
ω (r i , A t+1 ) > 3∆ for i 6 t follows similarly since A t+1 Bω (c t+1 , r −8∆)
t+1
and r it+1 ∈ SH
ω (c t+1 , r − 4∆).
t+1
t+1
t+1 ∪
We are left to verify the last case: distH
ω (r t+1 , W
two facts: W t+1 ∪
Ð
Ð
i 6t
i 6t
A t+1
) > 3∆. This follows from the
i
A t+1
∩ BωHt+1 (c t+1 , r − 8∆) ∅ and r > 20∆ from (4.9).
i
We have completed verification of the inductive step, and thus by induction there exists a careful
K h minor in G, completing the proof.
5
5.1
Applications and discussion
Spectral bounds
Say that a conformal graph (G, ω) is (r, ε)-spreading if it holds that for every subset S ⊆ VG with
|S| r, one has
1 Õ
distω (u, v) > εkωk`2 (VG ) .
|S| 2
u,v∈S
Let ε r (G, ω) be the smallest value ε for which (G, ω) is (r, ε)-spreading. The next theorem appears
as [KLPT11, Thm 2.3].
Theorem 5.1. If G is an n-vertex graph with maximum degree dmax , then for k 1, 2, . . . , n − 1, the
following holds: If ω : VG → + satisfies kωk`2 (VG ) 1, and (VG , distω ) admits an (α, ε/2)-padded
partition with ε ε bn/8kc (G, ω), then
λ k (G) . dmax
α2
.
ε2 n
The methods of [KLPT11] also give a way of producing (r, ε)-spreading weights. Consider a
graph G and let µ be a probability measure on subsets of VG . A flow Λ : PG → + is called a µ-flow
if
Λ[u, v] > µ({S : u, v ∈ S}) .
For a number r 6 |VG |, let Fr (G) denote the set of all µ-flows in G with supp(µ) ⊆ VrG (i.e., µ is
supported on subsets of size exactly r). The following is a consequence of the duality theory of
convex programs (see [KLPT11, Thm 2.4]).
Theorem 5.2. For every graph G and r 6 |VG |, it holds that
max
ω:VG →+
ε r (G, ω) : kωk`2 (VG ) 6 1
1
min kcΛ k`2 (VG ) : Λ ∈ Fr (G) .
2
r
We need to extend the notion of H-flows to weighted graphs. Suppose that H is equipped with
a non-negative weight on edges D : EH → + . Then an (H, D)-flow in G is a pair (Λ, ϕ) that satisfies
properties (1) and (2) of an H-flow, but property (3) is replaced by: For every u, v ∈ VG ,
Õ
Λ[u, v]
D({x, y}) .
{x,y}∈EH :{ϕ(x),ϕ(y)}{u,v}
∗
We define the crossing congestion χG
(H, D) as the infimum of χG (Λ, ϕ) over all (H, D)-flows (Λ, f )
V
in G. Given a measure µ on 2 H , let Dµ be defined by
Dµ ({x, y}) µ
S ⊆ VH : x, y ∈ S
24
.
We need the following result which is an immediate consequence of Corollary 3.6 and Corollary 4.2
in [KLPT11].
Theorem 5.3. There is a constant θ0 > 0 such that for every h > 3 and r > θ0 h 2 log h, the following holds.
If G excludes K h as a minor, then for any graph H and any measure µ supported on VrH , it holds that
∗
χG
(H, Dµ ) &
r5
.
|VH |h 2 log h
We can use the preceding theorem combined with the method of Section 3.3 to reach a conclusion
for rigs over K h -minor-free graphs.
Corollary 5.4. Suppose that Ĝ ∈ rig(G) and G excludes K h as a minor for some h > 3. Then for every
r > θ0 h 2 log h and Λ ∈ Fr (Ĝ), it holds that
kcΛ k`22 (V
Ĝ )
&
r5
dmax (Ĝ)|VĜ |h 2 log h
.
Proof. Suppose that Λ ∈ Fr (Ĝ). Let (Λ̌, ϕ̌) be the flow induced in G from the mapping described in
the proof of Theorem 3.5. By Claim 3.7, it holds that
χG (Λ̌, ϕ̌) 6 4dmax (Ĝ)kcΛ k`22 (V ) .
Ĝ
But from Theorem 5.3, we know that
χG (Λ̌, ϕ̌) &
r5
.
|VĜ |
We are now in position to prove Theorem 1.11.
Theorem 5.5 (Restatement of Theorem 1.11). Suppose that G ∈ rig(G0 ) and G0 excludes K h as a minor
for some h > 3. If dmax is the maximum degree of G, then for any k 1, 2, . . . , |VG | − 1, it holds that
2
λ k (G) 6 O(dmax
h 6 log h)
k
.
|VG |
Proof. Let r bn/8kc. We may assume that r > θ0 h 2 log h since the bound λ k (G) 6 2dmax (G)
always holds. From the conjunction of Corollary 5.4 and Theorem 5.2, we know there exists a
conformal metric ω : VG → + with kωk`2 (VG ) 1 and such that
1
ε r (G, ω) & 2
r
s
r5
.
dmax h 2 log h|VĜ |
From Theorem 2.10, we know that (VG , distω ) admits an (α, ∆)-padded partition for every ∆ > 0
with α 6 O(h 2 ). Now applying Theorem 5.1 yields the claimed eigenvalue bound.
5.2
Weighted separators
Throughout the paper, we have equipped graphs with the uniform measure over their vertices.
There are natural extensions to the setting where a graph G is equipped with a non-negative
measure on vertices µ : VG → + . The corresponding definitions naturally replace L p (VG ) by the
weighted space L p (VG , µ). The methods of Section 2 and Section 3 extend in a straightforward way
to this setting (see [FHL08] and, in particular, Section 3.6 there for extensions to a more general
setting with pairs of weights).
As an illustration, we state a weighted version of Theorem 1.1. Suppose that µ is a probability
measure on VG . A 23 -balanced separator in (G, µ) is a subset of nodes S ⊆ VG such that every
connected component of G[V \ S] has µ-measure at most 32 .
25
Theorem 5.6. If G ∈ rig(G0 ) and G0 excludes K h√ as a minor, then for any probability measure µ on VG ,
there is a 23 -balanced separator of weight at most c h
c h 6 O(h 3 log h).
5.3
m
n ,
where m |EG | and n |VG |. One has the estimate
Bi-Lipschitz embedding problems
We state two interesting open metric embedding problems. We state them here only for string
graphs, but the extension to rigs over K h -minor-free graphs is straightforward.
Random embeddings into planar graphs. Let G be a graph and consider a random variable
(F , G0 , len), where F : VG → VG0 , G0 is a (random) planar graph, and len : EG0 → + is an
assignment of lengths to the edges of G0 . We use dist(G0 ,len) denote the induced shortest-path
distance in G0 .
Question 5.7. Is there a constant K > 0 so that the following holds for every finite string graph G?
For every ω : VG → + , there exists a triple (F , G0 , len) such that:
1. (Non-contracting) Almost surely, for every u, v ∈ VG ,
dist(G0 ,len) (F(u), F(v)) > distω (u, v) .
2. (Lipschitz in expectation) For every u, v ∈ VG ,
dist(G0 ,len) (F(u), F(v)) 6 K · distω (u, v) .
A positive answer would clarify the geometry of the conformal metrics on string graphs. In
[CJLV08], the lower bound method of [GNRS04] is generalized to rule out the existence of non-trivial
reductions in the topology of graphs under random embeddings of the above form. But that method
relies on the initial family of graphs being closed under 2-sums, a property which is manifestly
violated for string graphs (since, in particular, string graphs are not closed under subdivision).
Bi-Lipschitz embeddings into L 1 . A well-known open question is whether every planar graph
metric admits an embedding into L 1 ([0, 1]) with bi-Lipschitz distortion at most C (for some
universal constant C); see [GNRS04] for a discussion of the conjecture and its extension to general
excluded-minor families. The following generalization is also natural.
Question 5.8. Do conformal string metrics admit bi-Lipschitz embeddings into L1 ? More precisely,
is there a constant K > 0 such that the following holds for every string graph G? For every
ω : VG → + , there is a mapping ϕ : VG → L 1 ([0, 1]) such that for all u, v ∈ VG ,
distω (u, v) 6 kϕ(u) − ϕ(v)kL1 6 K · distω (u, v) .
Note that, unlike in the case of edge-capacitated flows, a positive resolution does not imply an
O(1) vertex-capacitated multi-flow min/cut theorem for string graphs. See [FHL08] for a discussion
and [LMM15] for stronger types of embeddings that do yield this implication. If Question 5.7 has a
positive resolution, it implies that Question 5.8 is equivalent to the same question for planar graphs.
Acknowledgements
The author thanks Noga Alon, Nati Linial, and Laci Lovász for helpful discussions, Janos Pach for
emphasizing Jirka’s near-optimal bound for separators in string graphs, and the organizers of the
“Mathematics of Jiří Matoušek” conference, where this work was initiated.
26
References
[AST90]
Noga Alon, Paul Seymour, and Robin Thomas. A separator theorem for nonplanar
graphs. J. Amer. Math. Soc., 3(4):801–808, 1990. 2
[BLR10]
Punyashloka Biswal, James R. Lee, and Satish Rao. Eigenvalue bounds, spectral
partitioning, and metrical deformations via flows. J. ACM, 57(3):Art. 13, 23, 2010. Prelim.
version in FOCS 2008. 6, 7, 10, 13, 14
[Bou85]
J. Bourgain. On Lipschitz embedding of finite metric spaces in Hilbert space. Israel J.
Math., 52(1-2):46–52, 1985. 6
[BV04]
Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge University
Press, Cambridge, 2004. 13
[Can94]
James W. Cannon. The combinatorial Riemann mapping theorem. Acta Math., 173(2):155–
234, 1994. 5
[CJLV08]
Amit Chakrabarti, Alexander Jaffe, James R. Lee, and Justin Vincent. Embeddings of
topological graphs: Lossy invariants, linearization, and 2-sums. In IEEE Symposium on
Foundations of Computer Science, 2008. 26
[Duf62]
R. J. Duffin. The extremal length of a network. J. Math. Anal. Appl., 5:200–215, 1962. 5
[FHL08]
Uriel Feige, MohammadTaghi Hajiaghayi, and James R. Lee. Improved approximation
algorithms for minimum weight vertex separators. SIAM J. Comput., 38(2):629–657, 2008.
4, 5, 6, 7, 8, 25, 26
[FP10]
Jacob Fox and János Pach. A separator theorem for string graphs and its applications.
Combin. Probab. Comput., 19(3):371–390, 2010. 2
[FP14]
Jacob Fox and János Pach. Applications of a new separator theorem for string graphs.
Combin. Probab. Comput., 23(1):66–74, 2014. 3
[FPT10]
Jacob Fox, János Pach, and Csaba D. Tóth. A bipartite strengthening of the crossing
lemma. J. Combin. Theory Ser. B, 100(1):23–35, 2010. 3
[FT03]
J. Fakcharoenphol and K. Talwar. An improved decomposition theorem for graphs
excluding a fixed minor. In Proceedings of 6th Workshop on Approximation, Randomization,
and Combinatorial Optimization, volume 2764 of Lecture Notes in Computer Science, pages
36–46. Springer, 2003. 15
[GNRS04] Anupam Gupta, Ilan Newman, Yuri Rabinovich, and Alistair Sinclair. Cuts, trees and
l1 -embeddings of graphs. Combinatorica, 24(2):233–269, 2004. 26
[Gro07]
Misha Gromov. Metric structures for Riemannian and non-Riemannian spaces. Modern
Birkhäuser Classics. Birkhäuser Boston, Inc., Boston, MA, english edition, 2007. Based
on the 1981 French original, With appendices by M. Katz, P. Pansu and S. Semmes,
Translated from the French by Sean Michael Bates. 4
[Her70]
Joseph Hersch. Quatre propriétés isopérimétriques de membranes sphériques homogènes. C. R. Acad. Sci. Paris Sér. A-B, 270:A1645–A1648, 1970. 5
[KLPT11] J. Kelner, J. R. Lee, G. Price, and S.-H. Teng. Metric uniformization and spectral bounds
for graphs. Geom. Funct. Anal., 21(5):1117–1143, 2011. Prelim. version in STOC 2009. 7,
24, 25
27
[KM91]
Jan Kratochvíl and Jiří Matoušek. String graphs requiring exponential representations.
J. Combin. Theory Ser. B, 53(1):1–4, 1991. 2
[Kos82]
A. V. Kostochka. The minimum Hadwiger number for graphs with a given mean degree
of vertices. Metody Diskret. Analiz., (38):37–58, 1982. 2
[KPR93]
Philip N. Klein, Serge A. Plotkin, and Satish Rao. Excluded minors, network decomposition, and multicommodity flow. In Proceedings of the 25th Annual ACM Symposium on
Theory of Computing, pages 682–690, 1993. 6, 15
[Kra91]
Jan Kratochvíl. String graphs. II. Recognizing string graphs is NP-hard. J. Combin.
Theory Ser. B, 52(1):67–78, 1991. 2
[LMM15] James R. Lee, Manor Mendel, and Mohammad Moharrami. A node-capacitated
Okamura-Seymour theorem. Math. Program., 153(2, Ser. A):381–415, 2015. 26
[LN05]
James R. Lee and Assaf Naor. Extending Lipschitz functions via random metric partitions.
Invent. Math., 160(1):59–95, 2005. 7
[LR99]
Tom Leighton and Satish Rao. Multicommodity max-flow min-cut theorems and their
use in designing approximation algorithms. J. ACM, 46(6):787–832, 1999. 6
[LT79]
Richard J. Lipton and Robert Endre Tarjan. A separator theorem for planar graphs.
SIAM J. Appl. Math., 36(2):177–189, 1979. 6
[Mat02]
J. Matoušek. Lectures on discrete geometry, volume 212 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 2002. 6
[Mat14]
Jiří Matoušek. Near-optimal separators in string graphs. Combin. Probab. Comput.,
23(1):135–139, 2014. 2, 6
[Mat15]
Jiří Matoušek. String graphs and separators. In Geometry, structure and randomness in
combinatorics, volume 18 of CRM Series, pages 61–97. Ed. Norm., Pisa, 2015. 2
[MM16]
Konstantin Makarychev and Yury Makarychev. Metric extension operators, vertex
sparsifiers and Lipschitz extendability. Israel J. Math., 212(2):913–959, 2016. 7
[MS86]
Vitali D. Milman and Gideon Schechtman. Asymptotic theory of finite-dimensional normed
spaces, volume 1200 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. With
an appendix by M. Gromov. 4
[MTTV97] Gary L. Miller, Shang-Hua Teng, William Thurston, and Stephen A. Vavasis. Separators
for sphere-packings and nearest neighbor graphs. J. ACM, 44(1):1–29, 1997. 6
[Ost13]
Mikhail I. Ostrovskii. Metric embeddings, volume 49 of De Gruyter Studies in Mathematics.
De Gruyter, Berlin, 2013. Bilipschitz and coarse embeddings into Banach spaces. 15
[Rab08]
Yuri Rabinovich. On average distortion of embedding metrics into the line. Discrete
Comput. Geom., 39(4):720–733, 2008. 10
[SŠ04]
Marcus Schaefer and Daniel Štefankovič. Decidability of string graphs. J. Comput. System
Sci., 68(2):319–334, 2004. 2
[SSŠ03]
Marcus Schaefer, Eric Sedgwick, and Daniel Štefankovič. Recognizing string graphs in
NP. J. Comput. System Sci., 67(2):365–380, 2003. Special issue on STOC2002 (Montreal,
QC). 2
28
[ST07]
Daniel A. Spielman and Shang-Hua Teng. Spectral partitioning works: Planar graphs
and finite element meshes. Linear Algebra and its Applications: Special Issue in honor of
Miroslav Fiedler, 421(2–3):284–305, March 2007. 5, 6, 7
[Tho84]
Andrew Thomason. An extremal function for contractions of graphs. Math. Proc.
Cambridge Philos. Soc., 95(2):261–265, 1984. 2
29
| 8cs.DS
|
arXiv:1107.1171v2 [math.AC] 9 Feb 2014
A NOTE ON THE COMPUTATION OF THE
FROBENIUS NUMBER OF A NUMERICAL SEMIGROUP
JULIO JOSÉ MOYANO-FERNÁNDEZ
A BSTRACT. In this note we observe that the Frobenius number and therefore the conductor of a numerical semigroup can be obtained from the maximal socle degree of the
quotient of the corresponding semigroup algebra by the ideal generated by the biggest
generator of the semigroup.
1. I NTRODUCTION
AND
R EVIEW
Numerical semigroups occur often in many branches of Mathematics. One of the most
challenging problems in this area is the computation of the Frobenius number of the semigroup, i.e., the biggest integer not being an element of the numerical semigroup. In this
paper we describe a method to calculate it based on some fundamental concepts in commutative algebra. For further details and as a general reference on numerical semigroups,
the reader should refer to the works of Rosales and García Sánchez [2], and Ramírez Alfonsín [3]. Much of the notation we will use comes originally from the (in many respects
seminal) work of Herzog and Kunz [1]. Let N denote the set of nonnegative integers.
Let k be an arbitrary field. Let n1 , . . ., nd be positive integer numbers with gcd(n1 , . . ., nd ) =
1. Consider the numerical semigroup
Γ := Nn1 + . . . + Nnd
minimally generated by n1 , . . ., nd . It is well-known the existence of an element c ∈ N
minimal such that c + N ⊆ Γ. This number is called the conductor of Γ, and we will
denote it by c(Γ). The number f (Γ) := c(Γ) − 1 is then the biggest integer not belonging
to Γ, and it is called the Frobenius number of Γ.
Let n be a nonzero element of Γ. The set
Ap(Γ, n) := {h ∈ Γ | h − n ∈
/ Γ}
is called the Apéry set of n in Γ. It is easily checked that (cf. [4])
f (Γ) = max Ap(Γ, n) − n.
(†)
Let I 6= ∅ be a subset of Z satisfying I 6= Z and I + Γ ⊆ I. Such an I is said to be a
fractional Γ-ideal (sometimes also called Γ-semigroup). The Γ-ideal M := {s ∈ Γ | s 6= 0}
2010 Mathematics Subject Classification. Primary: 11D07; Secondary: 14Q05.
Key words and phrases. numerical semigroup; Frobenius problem; graded polynomial ring.
The author was partially supported by the Spanish Government Ministerio de Educación y Ciencia
(MEC) grant MTM2007-64704 in cooperation with the European Union in the framework of the founds
“FEDER”, and by the Deutsche Forschungsgemeinschaft (DFG).
1
2
JULIO JOSÉ MOYANO-FERNÁNDEZ
is the (uniquely determined) maximal ideal of Γ. It will be important in the sequel to
consider also the Γ-ideal
M − := {z ∈ Z | z + M ⊆ Γ}.
Note that M − ⊇ Γ, and since f (Γ) ∈ M − one has indeed M − ) Γ. The inclusion N ⊇ M −
holds precisely when Γ = N, and in this case −1 ∈ M − . The cardinality of the set of
elements in M − \ Γ will be denoted by r(Γ). Note also that
f (Γ) = max{m | m ∈ M − \ Γ}.
Let R := k[X1 , . . ., Xd ] (resp. k[t]) be the polynomial ring over k graded by deg(Xi ) = ni for
every i ∈ {1, . . ., d} (resp. deg(t) = 1). Let π be the graded homomorphism of k-algebras
π : R → k[t] given by Xi 7→ t ni for every i ∈ {1, . . ., d}. The image of π is the semigroup
ring associated with Γ, and it is denoted by k[Γ]. The homogeneous prime ideal p := ker π
is said to be the presentation ideal of k[Γ] (cf. [5]).
Let us consider p′ := πd−1 (p) the image in k[X1, . . . , Xd−1 ] by the epimorphism mapping
Xd onto 0, where πd−1 stands for the projection onto the first d − 1 coordinates, and define
the quotient ring
R′ := k[Γ]/(t nd ).
The following ring isomorphisms are easily checked:
R′ ∼
= k[X1 , . . ., Xd−1 ]/p′ ∼
= k[X 1 , . . . , X d−1 ],
where X i denotes the class of Xi modulo p′ for every i ∈ {1, . . ., d − 1}. Furthermore, the
ring R′ is *local, i.e., it has a unique maximal graded ideal mR′ .
2. T HE M AIN R ESULT
Let us define the trivial submodule (or socle) of R′ as the set of elements in R′ which are
annihilated by the homogeneous maximal ideal mR′ of R′ , namely
Triv(R′ ) := {x ∈ R′ | x · mR′ = (0)}.
This is the largest subspace of R′ having a R′ -module structure of vector space, and can
be identified as Hom(k, R′ ).
Note that the set
∆(Γ) := {λ ∈ N | λ + M ⊆ nd + Γ} \ (nd + Γ),
which is in fact a subset of the semigroup Γ, yields an isomorphism, say ϕ , between the
trivial submodule Triv(R′ ) and the set of formal power series
n
o
ν
r
t
|
r
∈
k
,
∑ ν ν
ν ∈∆(Γ)
whose elements are indeed polynomials. Furthermore, we have a bijection between the
sets ∆(Γ) and M − \ Γ given by mapping every ν ∈ ∆(Γ) to ν − nd ∈ M − \ Γ. This together
with the isomorphism ϕ leads to the equality between the cardinality of M − \ Γ and the
dimension of the socle:
r(Γ) = dimk Triv(R′ ).
A NOTE ON THE COMPUTATION OF THE
FROBENIUS NUMBER OF A NUMERICAL SEMIGROUP 3
This means in particular that the trivial submodule Triv(R′ ) is a finite dimensional vector
space over the field k. Let us then choose a basis B := {b1 , . . ., br(Γ) } and take the element
β ∈ B such that
deg(β ) = max{deg(bi ) | i = 1, 2, . . ., r(Γ)}.
Now it is a simple matter to realise:
Lemma 2.1. The degree deg(β ) is independent of the choice of the basis B of the trivial
submodule of the ring R′ .
We are thus led to the following result:
Theorem 2.2. We have:
f (Γ) = c(Γ) − 1 = deg(β ) − nd .
Proof. The proof is straightforward from the bijection ϕ .
Corollary 2.3. We have:
max Ap(Γ, nd ) = deg(β ).
Proof. The result follows straightforward from the equality (†) at the beginning of the
paper.
Example 2.4. Let us take the monomial curve C given by t 7→ (t 6,t 8 ,t 9). The corresponding numerical semigroup is ΓC = N · 6 + N · 8 + N · 9. The presentation ideal associated
with C is
p = (X13 − X32 , X23 − X1 X32 ),
so we have R′ = k[X1, X2 ]/p′ ∼
= k[X 1 , X 2 ] with p′ = (X13 , X23) and X i ≡ Xi mod p′ for i =
2 2
2 2
1, 2. Therefore we get Triv(R′ ) = k·X 1 X 2 , and it is clearly seen that f (ΓC ) = deg(X 1 X 2 )−
9 = 12 + 16 − 9 = 19; hence c(ΓC ) = 20, as one might have also checked—easily in this
example—straight from the semigroup ΓC .
R EFERENCES
[1] J. Herzog and E. Kunz, Die Wertehalbgruppe eines lokalen Rings der Dimension 1. Sitz. ber. Heidelberg. Akad. Wiss. (1971) 27–67.
[2] J. C. Rosales and P. A. García Sánchez, Numerical semigroups. Developments in Mathematics, Vol. 20.
(Springer, Berlin-Heidelberg-New York, 2010).
[3] J. L. Ramírez Alfonsín, The Diophantine Frobenius Problem. Oxford Lect. Series in Math. and its
Applicat., Vol. 30. (Oxford U.P., New York, 2005).
[4] E. S. Selmer, On a linear Diophantine problem of Frobenius. J. Reine Angew. Math. 293/294 (1977),
1–17.
[5] R. H. Villarreal, Monomial Algebras. (Marcel Dekker, New York–Basel, 2001).
U NIVERSITÄT O SNABRÜCK , FB M ATHEMATIK /I NFORMATIK , 49069 O SNABRÜCK , G ERMANY
E-mail address: [email protected]
| 0math.AC
|
1
Directional Cell Search Delay Analysis for
Cellular Networks with Static Users
arXiv:1709.00779v1 [cs.IT] 4 Sep 2017
Yingzhe Li, François Baccelli, Jeffrey G. Andrews, Jianzhong Charlie Zhang
Abstract
Cell search is the process for a user to detect its neighboring base stations (BSs) and make a
cell selection decision. Due to the importance of beamforming gain in millimeter wave (mmWave)
and massive MIMO cellular networks, the directional cell search delay performance is investigated. A
cellular network with fixed BS and user locations is considered, so that strong temporal correlations
exist for the SINR experienced at each BS and user. For Poisson cellular networks with Rayleigh fading
channels, a closed-form expression for the spatially averaged mean cell search delay of all users is
derived. This mean cell search delay for a noise-limited network (e.g., mmWave network) is proved to
be infinite whenever the non-line-of-sight (NLOS) path loss exponent is larger than 2. For interferencelimited networks, a phase transition for the mean cell search delay is shown to exist in terms of the
number of BS antennas/beams M : the mean cell search delay is infinite when M is smaller than a
threshold and finite otherwise. Beam-sweeping is also demonstrated to be effective in decreasing the
cell search delay, especially for the cell edge users.
I. I NTRODUCTION
Cell search is a critical prerequisite to establish an initial connection between a cellular
user and the cellular network. Specifically, the users will detect their neighboring BSs and
make the cell selection decision during a downlink cell search phase, after which the users
can acquire connections with the network by initiating an uplink random access phase. The
transmissions and receptions during cell search are performed omni-directionally in LTE [1],
but this is unsuitable for mmWave communication [2]–[5] or massive MIMO [6]–[9] due to the
lack of enough directivity gain. By contrast, directional cell search schemes that leverage BS
Y. Li, J. G. Andrews and F. Baccelli are with the Wireless Networking and Communications Group (WNCG), The University
of Texas at Austin (email: [email protected], [email protected], [email protected]). J. Zhang is with
Samsung Research America-Dallas (email: [email protected]). Date revised: September 5, 2017.
2
and/or user beam-sweeping to achieve extra directivity gains, can ensure reasonable cell search
performance [9]–[15]. In this paper, we leverage stochastic geometry [16]–[18] to develop an
analytical framework for the directional cell search delay performance of a fixed cellular network,
where the BS and user locations are fixed over a long period of time (e.g., more than several
minutes). We believe the analytical tools developed in this paper can provide useful insights
into practical fixed cellular networks such as fixed mmWave or massive MIMO broadband
networks [7], [19], [20], or mmWave backhauling networks [2], [21].
A. Related Work
Beam-sweeping is a useful method to improve cell search performance compared to conventional omni-directional cell search for both mmWave and massive MIMO networks. Specifically,
mmWave links generally require high directionality with large antenna gains to overcome the
high isotropic path loss of mmWave propagation. As a result, in mmWave networks, applying
beam-sweeping for cell search not only provides sufficient signal-to-noise ratio (SNR) to create
viable communications, but also facilitates beam alignment between the BS and users [11]–
[13], [22]–[24]. The directional cell search delay performance of mmWave systems has been
investigated in [12], [22], [23] from a link level perspective, and in [13], [24] from a system level
perspective. In particular, [13] and [24] consider the user and mmWave BS locations are fixed
within an initial access cycle, but independently reshuffled across cycles. This block coherent
scenario is fundamentally different from that of a fixed network. For a massive MIMO system,
the BSs can achieve an effective power gain that scales with the number of antennas if the
channel state information (CSI) is known at the BSs [9]. However, since such an array gain is
unavailable for cell search operations due to the lack of CSI, the new users may be unable to join
the system using the traditional omnidirectional cell search [9], [14], [15]. In order to overcome
this issue, [15] has proposed open-loop beamforming to exhaustively sweep through BS beams
for cell search. This design has been implemented and verified on a sub-6 GHz massive MIMO
prototype [15], but the analytical directional cell search performance has not been investigated
for fixed cellular networks from a system level perspective.
Due its analytical tractability for cellular networks [16]–[18], stochastic geometry is a natural
candidate for analyzing the directional cell search delay in such fixed cellular networks. In
particular, stochastic geometry has already been widely used to investigate fixed Poisson network
3
performance through the local delay metric [16], [25]–[28], which characterizes the number of
time slots needed for the SINR to exceed a certain SINR level. In [16], [25], the local delay for
fixed ad hoc networks was found to be infinite under several standard scenarios such as Rayleigh
fading with constant noise. A new phase transition was identified for the interference-limited
case in terms of the mean local delay: the latter is finite when certain parameters are above a
threshold, and infinite otherwise. The local delay for noise-limited and interference-limited fixed
Poisson networks was also investigated in [26]–[28], where it is shown that power control is an
efficient method to ensure a finite mean local delay. These previous works mainly focused on
omni-directional communications.
B. Contributions
In this work, we analyze the cell search delay in fixed cellular networks with a directional cell
search protocol. We consider a time-division duplex (TDD) cellular system, where system time
is divided into different initial access (IA) cycles. Each cycle starts with the cell search period,
wherein BSs apply a synchronous beam-sweeping pattern to broadcast the synchronization
signals. A mathematical framework is developed to derive the exact expression for the mean
cell search delay, which quantifies the spatial average of the individual mean cell search delays
perceived by all users. The main contributions of this paper are summarized as follows.
Beam-sweeping is shown to reduce the number of IA cycles needed to succeed in cell
search. For any arbitrary BS locations and fading distribution, the mean number of initial access
cycles required to succeed in cell search is proved to be decreasing when the number of BS
antennas/beams is multiplied by a factor m > 1.
An exact expression for the mean cell search delay is derived for Poisson point process
(PPP) distributed BSs and Rayleigh fading channels. This expression is given by an infinite
series, based on which the following observations are obtained:
•
Under the noise limited scenario (e.g., mmWave networks), we prove that as long as the
path loss exponent for NLOS path is larger than 2, the mean cell search delay is infinite,
irrespective of the BS transmit power and the BS antenna/beam number.
•
Under the interference limited scenario (e.g., massive MIMO networks in sub-6 GHz bands),
there exists a phase transition for mean cell search delay in terms of the BS antenna/beam
number M. Specifically, the mean cell search delay is infinite when M is smaller than a
4
critical value and finite otherwise. This fact was never observed in the literature to the best
of our knowledge.
Cell search delay distribution is numerically evaluated. The conditional mean cell search
delay of a typical user given its nearest BS distance is derived for PPP distributed BSs and
Rayleigh fading channels. The distribution of this conditional mean cell search delay is also
numerically evaluated, and we observe that the cell search delay distribution is heavy-tailed. We
also show that increasing the number of BS antennas/beams can significantly reduce the cell
search delay for cell edge users.
Overall, this paper has shown that in fixed networks the mean cell search delay could be very
large due to the temporal correlations induced by common randomness. As a result, for fixed
cellular networks, system parameters including the number of BS antennas and/or BS intensity
need to be carefully designed for reasonable cell search delay performance to be achieved.
II. S YSTEM M ODEL
In this work, we consider a cellular system that has carrier frequency fc and total system
bandwidth W . The BS transmit power is denoted by Pb , and the total thermal noise power
is denoted by σ 2 . In the rest of this section, we present the proposed directional cell search
protocol, location models, propagation assumptions, and the performance metrics.
A. Directional Cell Search Protocol
We consider a TDD cellular system as shown in Fig. 1, where system time is divided into
different initial access cycles with period T , and where τ denotes the OFDM symbol period.
Initial access refers to the procedures that establish an initial connection between a user and the
cellular network. It consists of two main steps: cell search on the downlink and random access
(RA) on the uplink. Specifically, by detecting the synchronization signals broadcasted by BSs
during cell search, a user can determine the presence of its neighboring BSs and make the cell
selection decision. Then the user can initiate the random access process to its desired serving BS
by transmitting a RA preamble through the shared random access channel, and it is successfully
connected to the network if the BS can decode the RA preamble without any collision. The
main focus of this work is the cell search performance, while the random access performance
will be incorporated in our future work.
5
Each BS is equipped with a large dimensional antenna array with M to support highly
directional communications. For analytical tractability, the actual antenna pattern is approximated
by a sectorized beam pattern, where the antenna gain is constant within the main lobe. In addition,
we assume a 0 side lobe gain for the BS, which is a reasonable approximation because the BS
uses a large dimensional antenna array with narrow beams, possibly with a front-to-back ratio
larger than 30 dB [29]. Each BS supports analog beamforming with a maximum of M possible
BF vectors, where the m-th (1 ≤ m ≤ M) beamforming (BF) vector corresponds to the mainm
lobe, which has antenna gain M, and covers a sector area with angle [2π m−1
, 2π M
) [30]. Each
M
user is assumed to have a single omni-directional antenna with unit antenna gain [15], [31].
In the cell search phase, each BS sweeps through all M transmit beamforming directions to
broadcast the synchronization signals, and each user is able to detect a BS with sufficiently small
miss detection probability (such as 1%) if the signal-to-interference-plus-noise ratio (SINR) of
the synchronization signal from that BS exceeds Γcs . All BSs transmit synchronously using the
same beam direction during every symbol, and the cell search delay within each IA cycle is
therefore Tcs = M × τ . When every BS transmits using the m-th (1 ≤ m ≤ M) BF direction,
the typical user can only receive from the BSs located inside the “BS sector”
2πm
2π(m − 1)
+ π,
+π ,
S o,
M
M
(1)
where we define the infinite sector domain centered at u ∈ R2 by:
S(u, θ1 , θ2 ) = {x ∈ R2 , s.t., ∠(x − u) ∈ [θ1 , θ2 )}.
(2)
There are M such non-overlapping BS sectors during cell search, with the j-th (1 ≤ j ≤ M)
sector being S(o, 2π(j−1)
, 2πj
). We say a BS sector is detected during cell search if the typical
M
M
user is able to detect the BS that provides the smallest path loss (i.e., the closest BS) inside
this sector, where the path loss can be estimated from the beam reference signals [19]. After
cell search, the typical user selects the BS with the smallest path loss among all the detected
BS sectors as its serving BS. For simplicity, we neglect the scenario where the BS providing
the smallest path loss inside a BS sector is in deep fade and unable to be detected, while some
other BSs can be detected in the same sector. Such a scenario does not change the fundamental
trends regarding the finiteness of the mean cell search that will be detailed in Section III (e.g.
Theorem 3), but the corresponding analysis is significantly more complicated.
6
Initial access cycle n
Initial access cycle n+1
time
......
......
CS period RA period
Data transmission period
CS period RA period
DL beam
pair 2
DL beam
pair 2
...
...
...
...
DL beam
pair 1
τ
Data transmission period
...
...
...
...
DL beam
pair M
DL beam
pair 1
τ
DL beam
pair M
Fig. 1: Illustration of two initial access cycles and the timing structure.
TABLE I: Notation and Simulation Parameters
Symbol
Φ, λ
Φu , λu
Pb , Pu
fc , B
W
M
αL , αN
CL , CN
Rc
Γcs , Γra
τ
T
SM (i)
{xi0 }M
i=1
R0
Lcs (M, λ)
Lcs (R0 .M, λ)
Dcs (M, λ)
B(x, r) (B o (x, r))
Definition
BS PPP and intensity
User PPP and intensity
BS and user transmit power
Carrier frequency and system bandwidth
Total thermal noise power
Number of BS antennas and BF directions supported at each BS
Path loss exponents for dual-slope model
Path loss at close-in reference distance for dual-slope model
Critical distance for dual-slope path loss model
SINR threshold to detect synchronization signal and RA preamble
OFDM symbol period
Initial access cycle period
, 2π Mi )}
i-th BS sector, i.e., SM (i) = {x ∈ R2 , s.t., ∠x ∈ [2π (i−1)
M
BS providing the smallest path loss to the typical user inside SM (i)
Distance from typical user to its nearest BS
Number of IA cycles to succeed in cell search
Mean number of IA cycles to succeed in cell search conditionally on R0
Cell search delay
Closed (open) ball with center x and radius r
Simulation Value
λ = 100 BS/km2
λu = 1000 users/km2
Pb = 30 dBm, Pu = 23 dBm
(fc , B) = (73, 1) GHz, (2, 0.2) GHz
−174 dBm/Hz + 10 log10 (B)
(αL , αN ) = (2.1, 3.3), (2.5, 2.5)
(CL , CN ) = (69.71, 69.71) dB,
(38.46, 38.46) dB
50m
(Γcs , Γra ) = (−4, −4) dB
14.3 µs, 71.4 µs
20 ms, 100 ms
B. Spatial Locations and Propagation Models
The BS locations are assumed to be a realization of a stationary point process Φ = {xi }i with
intensity λ. The user locations are modeled as a realization of a homogeneous PPP with intensity
λu , which is denoted by Φu = {ui }i . In this paper, a fixed network scenario is investigated where
the BS locations are fixed, and the users are either fixed or move with very slow speed such as
a pedestrian speed (e.g., less than 1 km/h) . As a result, the BS and user locations appear to be
fixed across different initial access cycles. This is fundamentally different from the high mobility
scenario investigated in [13], [24], which assumes the BS and user PPPs are independently re-
7
shuffled across every initial access cycles.
Without loss of generality, we can analyze the performance of a typical user u0 located at
the origin. This is guaranteed by Slivnyak’s theorem, which states that the property observed by
′
the typical point of a PPP Φ is the same as that observed by the point at origin in the process
′
Φ ∪ {o} [32], [33].
A dual-slope, non-decreasing path loss function [34] is adopted, where the path loss for a link
with distance r is given by:
C r αL , if r < R ,
L
C
l (r) =
CN r αN , if r ≥ RC .
(3)
The dual slope path loss model captures the dependency of the path loss exponent on the link
distance for various network scenarios, such as ultra-dense [34] and mmWave networks [35].
In particular, (3) is referred to as the LOS ball blockage model for mmWave networks [34],
wherein αL and αN represent the LOS and NLOS path loss exponents, and CL and CN represent
the path loss at a close-in reference distance (e.g., 1 meter). We focus on the scenario where
αN ≥ max(αL , 2). If αL = αN = α and CL = CN = C, the dual slope path loss model reverts
to the standard single-slope path loss model.
Due to the adopted antenna pattern for BSs, the directivity gain between BS and user is M
when the BS beam is aligned with the user, and 0 otherwise. The fading effect for every BSuser link is modeled by an i.i.d. random variable, whose complementary cumulative distribution
function (CCDF) is a decreasing function G(·) with support [0, ∞). In addition, we assume the
IA cycle length is such that the fading random variables for a given link are also i.i.d. across
different cycles.
C. Performance Metrics
The main performance metrics investigated in this work are the number of IA cycles, and
the corresponding cell search delay for the typical user to discover its neighboring BSs and
determine a potential serving BS. Without loss of generality, the IA cycle 1 in Fig. 1 represents
the first IA cycle of the typical user. Denote by eM (n) the success indicator for cell search of
IA cycle n. The number of IA cycles for the typical user to succeed in cell search is therefore:
Lcs (M, λ) = inf{n ≥ 1 : eM (n) = 1}.
(4)
8
Since analog beamforming is adopted at each BS, the cell search delay is defined as follows:
Dcs (M, λ) = (Lcs (M, λ) − 1)T + Mτ.
(5)
Finally, Table I summarizes the notation, the definitions and the system parameters that will
be used in the rest of this paper1 .
III. A NALYSIS
FOR
M EAN C ELL S EARCH D ELAY
In this section, the mean cell search delay performance for the typical user is investigated,
which corresponds to the cell search delay under the Palm expectation with respect to the user
PPP Φu (i.e., E0Φu [Dcs (M, λ)]). In fact, the Palm expectation can also be understood from its
ergodic interpretation, which states that for any user u ∈ Φu with cell search delay Dcs (u, M, λ),
the following relation is true:
X
1
1u∈B(0,n) Dcs (u, M, λ).
n→∞ Φu (B(0, n))
k
E0Φu [Dcs (M, λ)] = lim
(6)
Therefore, the mean cell search delay of the typical user can also be understood as the spatial
average of the individual cell search delays among all the users. For notational simplicity, we
will use E in the rest of this paper to denote the Palm expectation under the user PPP Φu .
A. Cell Search Delay Under General BS Deployment and Fading Assumptions
In this part, we first investigate the cell search delay under a general BS location model (not
necessarily PPP) and fading distribution. According to Section II, the BS and user locations
are fixed, and the fading variables for every link are i.i.d. across IA cycles. Therefore, given
the BS process Φ, the cell search success indicators for different IA cycles {eM (n)} form an
i.i.d. Bernoulli sequence of random variables. The cell search success probability is denoted by
πM (Φ) = E [eM (1)|Φ].
Since each BS sector can be independently detected given Φ, and cell search is successful if
at least one BS sector is detected. Conditionally on Φ, the cell search success probability for
1
For the symbols with two simulation values, the first one is for the noise limited scenario, and the second one is for the
interference limited scenario, which will be detailed in Section IV.
9
every IA cycle is therefore:
πM (Φ) = 1 −
M
Y
[1 − E [êM (i)|Φ]] ,
(7)
i=1
where êM (i) denotes the indicator that the BS providing the smallest path loss inside BS sector
i is detected. Specifically, if we denote by SM (i) , S(o, 2π(i−1)
, 2πi
) the BS sector i, xi0 the
M
M
BS providing the smallest path loss to the typical user in Φ ∩ SM (i), and by {Fji } the fading
random variables from BSs in SM (i) to the typical user, we have:
F0i /l (kxi0 k)
P
> Γcs Φ
E [êM (i)|Φ] = P
Fji /l (kxij k) + W/P M
xij ∈Φ∩SM (i)\{xi0 }
= E G Γcs l (kxi0 k)(
X
Fji /l (kxij k)
+ W/P M)
xij ∈Φ∩SM (i)\{xi0 }
Φ ,
(8)
where the expectation in (8) is taken with respect to the i.i.d. fading random variables {Fji }.
In the following theorem, we derive the mean number of IA cycles for the typical user to
succeed in the cell search under the Palm expectation of the user process.
Theorem 1: The mean number of IA cycles needed for the typical user to succeed in cell
search is given by:
E[Lcs (M, λ)|Φ] =
1
1−
"
E[Lcs (M, λ)] = EΦ
QM
1−
i=1
[1 − E [êM (i)|Φ]]
,
1
QM
i=1
[1 − E [êM (i)|Φ]]
(9)
#
.
(10)
Proof: The first part can be proved by the fact that given Φ, Lcs (M, λ) has a geometric distribution with success probability πM (Φ); while the second part follows by taking the expectation
of (9) with respect to Φ.
Remark 1: Since E [êM (i)|Φ] > 0 according to (8), the conditional mean cell search delay
E[Lcs (M, λ)|Φ] will be finite almost surely. However, the overall spatial averaged mean cell
search delay with respect to (w.r.t.) the BS PPP Φ (i.e. E[Lcs (M, λ)]) could be infinite under
certain network settings. This will be detailed in the next subsection.
A lower bound and an upper bound to E[Lcs (M, λ)] can be immediately obtained from (10),
which are provided in the following remarks.
Remark 2: By applying Jensen’s inequality to the positive random variable X and the function
10
f (x) = x1 , we get that E[ X1 ] ≥
1
.
E[X]
Thus
E[Lcs (M, λ)] ≥
1
1−E
i,
i=1 [1 − E [êM (i)|Φ]]
hQ
M
(11)
where the equality holds when the BS PPP is independently re-shuffled across different IA cycles
from the typical user’s perspective, which coincides with the high mobility scenario considered
in [13], [24].
Remark 3: If we denote by x0 the BS providing the smallest path loss to the typical user, and
Q
∗
i∗ the index for the BS sector that contains x0 , then M
j=1 [1 − E [êM (j)|Φ]] ≤ 1 − E [êM (i )|Φ].
Therefore, an upper bound to E[Lcs (M, λ)] is given by:
1
E[Lcs (M, λ)] ≤ E
.
E [êM (i∗ )|Φ]
(12)
Based on Theorem 1, we can prove the following relation between the BS antenna/beam
number M and the mean cell search delay.
Lemma 1: Given a realization of BS locations Φ, the mean number of IA cycles to succeed
in cell search is such that E[Lcs (M2 , λ)|Φ] < E[Lcs (M1 , λ)|Φ], if M2 = mM1 with m being an
integer larger than 1.
Proof: Since M2 = mM1 , we know that SM1 (i) =
Sm
j=1 SM2 ((i − 1)m + j)
for 1 ≤ i ≤ M1 .
Denote by xi0 the BS providing the smallest path loss to the typical user inside Φ ∩ SM1 (i), and
assume xi0 ∈ Φ ∩ SM2 ((i − 1)m + j0 ) for some j0 ∈ [1, m]. Due to the facts that M2 > M1 ,
SM2 ((i − 1)m + j0 ) ( SM1 (i), and since G(·) is a decreasing function, we get from (8) that
E [êM1 (i)|Φ] < E [êM2 ((i − 1)m + j0 )|Φ]. Also note that E [êM2 ((i − 1)m + j)|Φ] > 0 for ∀j 6=
j0 according to (8), we hence have:
m
Y
[1 − E [êM2 ((i − 1)m + j) |Φ]] < 1 − E [êM2 (i)|Φ] .
j=1
Thus the cell search success probability for the typical IA cycle satisfies:
πM2 (Φ) = 1 −
M2
Y
[1 − E [êM2 (i)|Φ]]
i=1
=1−
M1
m
Y
Y
i=1
j=1
!
1 − E êM2 ((i − 1)m + j) Φ
(13)
11
>1−
M1
Y
[1 − E [êM1 (i)|Φ]] = πM1 (Φ).
(14)
i=1
Finally the proof is concluded by applying Theorem 1.
Lemma 1 shows that for all BS location models and fading distributions, the conditional
number of IA cycles for cell search to succeed decreases when the number of BS antenna/beams
is multiplied by an integer m > 1, or equivalently when the BS beamwidth is divided by m.
This result also implies that E[Lcs (M2 , λ)] ≤ E[Lcs (M1 , λ)] if M2 = mM1 .
Remark 4: In fact, Lemma 1 cannot be further extended. If M2 > M1 but M2 /M1 is not an integer, there will always exist special constructions of BS deployments such that E[Lcs (M2 , λ)|Φ] >
E[Lcs (M1 , λ)|Φ].
For the rest of this section, we investigate the mean cell search delay under several specific
network scenarios.
B. Mean Cell Search Delay in Poisson Networks with Rayleigh Fading
In this part, the BS locations are assumed to form a homogeneous PPP with intensity λ, and
the fading random variables are exponentially distributed with unit mean (i.e., G(x) = exp(−x)).
Due to its high analytical tractability, this network setting has been widely adopted to obtain the
fundamental design insights for conventional macro cellular networks [17], ultra-dense cellular
networks [34], and even mmWave cellular networks2 [35], [36].
Due to the PPP assumption for BSs, and the fact that different BS sectors are non-overlapping,
every BS sector can therefore be detected independently with the same probability. Since the
path loss function l (r) is non-decreasing, the BS that provides the minimum path loss to the
typical user inside the i-th BS sector Φ ∩ SM (i) (i.e., xi0 ) is the closest BS to the origin. The
angle of xi0 is uniformly distributed within [2π(i − 1)/M, 2πi/M), and the CCDF for the norm
of xi0 can be derived as follows:
P(kxi0 k
≥ r) = P
min
x∈Φ∩SM (i)
kxk ≥ r
= exp(−
λπr 2
),
M
(15)
where the second equality follows from the void probability for PPPs. Therefore, the probability
2
The SINR and rate trends for mmWave networks under Rayleigh fading and PPP configured BSs have been shown to be
close to more realistic fading assumptions, such as the Nakagami fading or log-normal shadowing [11].
12
distribution function (PDF) for kxi0 k is given by:
fkxi0 k (r) =
λπr 2
2λπr
exp(−
).
M
M
(16)
By applying Φ ∼ PPP(λ) and G(x) = exp(−x) into (8), the conditional detection probability
for the i-th BS sector is given by:
i
E [êM (i)|Φ] = E exp −Γcs l (kx0 k)
X
Fji /l (kxij k)
+ W/P M
xij ∈Φ∩SM (i)\{xi0 }
W Γcs l (kxi0 k)
= exp −
E
PM
W Γcs l (kxi0 k)
= exp −
PM
Y
exp
xij ∈Φ∩SM (i)\{xi0 }
(a)
Y
xij ∈Φ∩SM (i)\{xi0 }
Φ
−Γcs l (kxi0 k)Fji /l (kxij k)
1
, FM (i, Φ), (17)
1 + Γcs l (kxi0 k)/l (kxij k)
where step (a) is obtained by taking the expectation w.r.t. the fading random variables.
Theorem 2: If Φ ∼ PPP(λ), and the fading variables are exponentially distributed with unit
mean, the mean number of cycles for cell search to succeed is:
E[Lcs (M, λ)] =
∞
X
AM
j ,
(18)
j=0
where Aj = E[(1 − FM (1, Φ))j ] is given by:
Aj =
Z
0
×
j
∞ X
Z
2πλ ∞
1
W kΓcs l (r1 )
j
exp −
1−
rdr
exp −
(−1)
k
P
M
M
(1
+
Γ
k
cs l (r1 )/l (r))
r
1
k=0
k
2λπr1
λπr12
exp(−
)dr1 .
M
M
(19)
Proof: By substituting (17) into Theorem 1, we obtain:
1
E[Lcs (M, λ)] = E
Q
1− M
i=1 [1 − FM (i, Φ)]
X
j
∞
M
Y
(a)
=E
[1 − FM (i, Φ)]
j=0
i=1
Y
j
∞
M
(b) X
=
E
[1 − FM (i, Φ)]
j=0
i=1
13
j M
∞
X
=
E 1 − FM (1, Φ)
,
(c)
(20)
j=0
where step (a) is derived from the fact that
1
1−x
=
P∞
j
j=0 x
for 0 ≤ x < 1, step (b) follows
from the monotone convergence theorem, and step (c) is because the events for BS sectors to
be detected are i.i.d. for PPP distributed BSs. Furthermore, we can compute Aj as follows:
j
E 1 − FM (1, Φ)
j
2λπr1
λπr12
1
=
E 1 − FM (1, Φ) x0 = (r1 , 0)
exp(−
)dr
M
M
0
Z ∞X
j
2λπr1
λπr12
(a)
x10
k
k j
EΦ (FM (1, Φ)) Φ ∩ SM (1) ∩ B(o, r1 ) = 0
=
(−1)
exp(−
)dr
k
M
M
0
k=0
Z ∞X
j
Y
1
W kΓcs l (r1 )
(b)
k j
E exp −
=
(−1)
PM
(1 + Γcs l (r1 )/l (kxij k))k
k
0
i
c
k=0
Z
∞
xj ∈Φ∩SM (i)∩B (o,r1 )
×
2λπr1
λπr12
exp(−
)dr,
M
M
(21)
x1
where EΦ0 [·] in (a) denotes the expectation under the Palm distribution at BS x10 ; and step (b) is
derived from Slivnyak’s theorem. Finally the proof can be concluded by applying the probability
generating functional (PGFL) of PPPs [32] to (21).
Remark 5: Theorem 2 can be interpreted as E[Lcs (M, λ)] =
P∞
j=0
P(Lcs (M, λ) > j), with
AM
j in (18) representing the probability that the BS sectors are not detected within j IA cycles,
i.e., P(Lcs (M, λ) > j).
Theorem 2 provides a series representation of the expected number of IA cycles to succeed
cell search. However, it is unclear from Theorem 2 whether E[Lcs (M)] is finite or not. In the
following, we will investigate the finiteness of E[Lcs (M, λ)] under two representative network
scenarios, namely the noise limited scenario and the interference limited scenario.
1) Noise limited Scenario: In the noise limited scenario, we assume the noise power dominates
the interference power (or interference power is perfectly canceled), such that only noise power
needs to be taken into account. Compared to conventional micro-wave cellular networks that
operate in sub-6 GHz bands, mmWave networks have much higher noise power due to the wider
bandwidth, and the interference power is much smaller due to the high isotropic path loss in
mmWave. As a result, mmWave cellular networks are typically noise limited, especially when
14
the carrier frequency and system bandwidth are high enough (e.g. 73 GHz carrier frequency with
2 GHz bandwidth) [35], [37].
Since the interference power is zero under the noise limited scenario, Theorem 2 becomes:
j
M
∞ Z ∞
X
W Γcs l (r1 )
2λπr1
λπr12
E[Lcs (M, λ)] =
1 − exp −
exp(−
)dr1
.
(22)
PM
M
M
0
j=0
Through the change of variable (v = λr 2 ), (22) becomes
!
p
M
∞ Z ∞
j
X
W Γcs l ( (v/λ))
2π
πv
exp(− )dv
,
1 − exp −
E[Lcs (M, λ)] =
P
M
M
M
0
j=0
(23)
which shows that E[Lcs (M, λ)] is non-increasing as the BS intensity λ increases, i.e., network
densification helps in reducing the number of IA cycles to succeed in cell search.
In the next two lemmas, we prove that the finiteness of E[Lcs (M, λ)] depends on the NLOS
path loss exponent αN , and that a phase transition for E[Lcs (M, λ)] happens when αN = 2.
Theorem 3: Under the noise limited scenario, for any finite number of BS antennas/beams
M and BS intensity λ, E[Lcs (M, λ)] = ∞ whenever the NLOS path loss exponent αN > 2.
Proof: Given the number of BS antennas/beams M and for any arbitrarily large positive
value v0 with v0 > Rc , we can re-write (22) to obtain the following lower bound on E[Lcs (M)]:
j
M
∞ Z ∞
X
2λπr1
λπr12
W Γcs l (r1 )
exp(−
)dr1
1 − exp −
PM
M
M
0
j=0
j
M
∞ Z ∞
(a) X
W Γcs CN r1αN
2λπr1
λπr12
≥
1 − exp −
exp(−
)dr1
P
M
M
M
v
0
j=0
j Z ∞
M
∞
X
λπr12
2λπr1
W Γcs CN v0αN
exp(−
)dr1
>
1 − exp −
PM
M
M
v0
j=0
jM
∞
X
W Γcs CN v0αN
exp(−λπv02 )
=
1 − exp −
P
M
j=0
exp(−λπv02 )
1 − (1 − exp(−W Γcs CN v0αN /P M))M
(b) 1
v0 →∞
≥ exp W Γcs CN v0αN /P M − λπv02 −→
∞,
M
=
(24)
where l (r1 ) = CN r1αN in step (a) because r1 ≥ v0 > Rc . Step (b) follows from the fact that for
any 0 ≤ x ≤ 1 and M ∈ N+ , we have: (1 − x)M + xM ≥ 1, thus
1
1−(1−x)M
≥
1
.
xM
Note that
15
since αN > 2, (24) goes to infinity when v0 goes to infinity, which completes the proof.
According to Lemma 3, the expected cell search delay is infinity whenever αN > 2, which
cannot be alleviated by BS densification (i.e., increase λ), or using a higher number of BS
antennas (i.e., increase M). The reason can be explained from (24), which shows that due to
the PPP-configured BS deployment, the typical user could be located at the “cell edge” with
its closest BS inside every BS sector farther than some arbitrarily large distance v. There is a
exp(−λπv 2 ) fraction of such cell edge users, and the corresponding number of IA cycles required
for them to succeed in cell search is at least exp(Cv αN ) for some C > 0. Therefore, the expected
cell search delay averaged over all the users will ultimately be infinite when αN > 2. From a
system level perspective, this indicates that for noise limited networks with αN > 2, there will
always be a significant fraction of cell edge users requiring a very large number of IA cycle to
succeed cell search, so that the spatial averaged cell search delay perceived by all users will be
determined largely by these cell edge users, which explains why an infinite mean cell search
delay is observed.
Theorem 4: Under the noise limited scenario with NLOS path loss exponent αN = 2, the
expected number of IA cycles to succeed in cell search E[Lcs (M, λ)] = ∞ if the BS density λ
and the BS antenna/beam number M satisfy λM <
Γcs CN W
,
Pπ
Γcs CN W
,
Pπ
and E[Lcs (M, λ)] < ∞ if λM >
Γcs CN W
.
Pπ
Γcs CN W
. In
Pπ
i.e., the phase transition for E[Lcs (M, λ)] happens at (λ∗ , M ∗ ) with λ∗ M ∗ =
Proof: If αN = 2, it is clear from (24) that E[Lcs (M, λ)] = ∞ if λM <
addition, we can simplify the upper bound to E[Lcs (M, λ)] from Remark 3 under the noise
limited scenario, which is given as follows:
E[Lcs (M, λ)]
Z ∞
(a)
W Γcs l (r0 )
λ2πr0 exp(−λπr02 )dr0
≤
exp
PM
0
Z ∞
Z Rc
W Γcs CN r0αN
W Γcs CL r0αL
2
exp
λ2πr0 exp(−λπr0 )dr0 +
λ2πr0 exp(−λπr02 )dr0
exp
=
P
M
P
M
Rc
0
Z ∞
αL
W Γcs CL Rc
W Γcs CN r0αN
2
< exp
exp
1 − exp(−λπRc ) +
λ2πr0 exp(−λπr02 )dr0 ,
PM
P
M
Rc
(25)
where (a) is obtained by applying the noise limited assumption to (17), and noting that the BS
providing the smallest path loss among all the BSs is the closest BS of Φ to the origin. Since
16
αN = 2, it can be observed from (25) that E[Lcs (M)] is guaranteed to have a finite mean if
λM >
Γcs CN W
.
Pπ
We can observe from the proof of Lemma 4 that for any arbitrarily large distance r0 , there
is a fraction exp(−λπr02 ) of cell edge users whose nearest BSs are farther than r0 , and the
number of IA cycles for these edge users to succeed cell search scales as exp(
W Γcs CN r02
).
PM
As
a result, if the BS deployment is too sparse or the number of BS antennas/beams is such that
λM <
Γcs CN W
,
Pπ
the cell search delay averaged over all the users becomes infinity due to cell
edge users. By contrast, with network densification, the fraction of cell edge users with poor
signal power is reduced, and the average cell search delay can be reduced to a finite mean value
whenever λM >
Γcs CN W
Pπ
. A similar behavior happens when the BSs are using more antennas
to increase the SNR for the cell edge users.
To summarize, for the noise limited scenario such as a mmWave network, the mean cell search
delay is infinite whenever the NLOS path loss exponent αN > 2, which is typically the case.
However, for the special case with NLOS path loss exponent αN = 2, the mean cell search
delay could switch from infinity to a finite value through careful network design, such as BS
densification or adopting more BS antennas.
2) Interference limited Scenario: In the interference limited scenario, the noise power is
dominated by the interference power, so that we can assume W = 0. For example, a massive
MIMO network that operates in the sub-6 GHz bands is typically interference limited [6]. In
this part, we investigate the cell search delay in an interference-limited network with a standard
single slope path loss function l (r) = Cr α , which is suitable for networks with sparsely deployed
BSs as opposed to ultra-dense networks [34].
First, we prove that Theorem 2 can be greatly simplified under this interference limited
scenario.
Lemma 2: Under the interference limited scenario, the expected number of initial access
cycles required to succeed in cell search is given by:
E[Lcs (M)] =
j
∞ X
X
j=0
k=0
(−1)k
1+2
R +∞
1
j
k
(1 − (1 + Γcs /r α )−k )rdr
M
.
(26)
Proof: By substituting W = 0 and l (r) = Cr α into (19), Aj defined in (19) can be further
17
simplified as follows:
j
∞ X
Z
2πλ ∞
1
j
2λπr1
λπr12
exp −
1−
rdr
exp(−
)dr1
(−1)
Aj =
α
α k
k
M
(1
+
Γ
M
M
cs r1 /r )
r
0
1
k=0
Z ∞
Z
j
X
2πλr12 ∞
1
λπr12
2λπr1
k j
exp −
(−1)
=
rdr
1−
exp(−
)dr1
k
M
(1 + Γcs /r α )k
M
M
0
1
k=0
j
X
(−1)k kj
,
=
R +∞
α )−k )rdr
1
+
2
(1
−
(1
+
Γ
/r
cs
1
k=0
Z
k
which completes the proof.
Remark 6: We can observe from Lemma 2 that E[Lcs (M)] does not depend on the BS
intensity λ under the interference limited scenario. This is because the increase and decrease of
the signal power can be perfectly counter-effected by the corresponding increase and decrease
of the interference power [17]. Another immediate observation from Lemma 2 is that Aj is
independent of the number of BS antennas M for ∀j. Since Aj ≤ 1 according to its definition
in Theorem 2, E[Lcs (M)] is therefore monotonically non-increasing with respect to M, which
is a stronger observation than Lemma 1.
Remark 7: If the path loss exponent α = 2, it can be proved from Lemma 2 that E[Lcs (M)] =
∞ for ∀M. This is mainly because the interference power will dominate the signal power when
α = 2, so that the coverage probability is 0 for any SINR threshold Γcs .
If α > 2, we can prove that there may exist a phase transition for E[Lcs (M)] in terms of the
BS beam number M. In order to show that, we first apply Remark 3 and obtain a sufficient
condition to guarantee the finiteness for E[Lcs (M)].
Lemma 3: Under the interference limited scenario with path loss exponent α > 2, the expected
number of IA cycles to succeed cell search is such that E[Lcs (M)] < ∞ if the number of BS
beams is such that M >
2Γcs
,
α−2
where Γcs denotes the detection threshold for a BS. In particular,
when M = 1, i.e., the BS is omni-directional, E[Lcs (1)] is finite if and only if α > 2Γcs + 2.
Proof: Denote by x0 the closest BS to the origin among Φ, and SM (i∗ ) the BS sector
containing x0 , we can obtain an upper bound to E[Lcs (M)] by substituting (17) and W = 0 into
Remark 3 as follows:
E[Lcs (M)] ≤ E
Y
1 + Γcs l (kx0 k)/l (kxj k)
xj ∈Φ∩SM (i∗ )\{x0 }
18
(a)
=
Z
∞
0
E
∞
Y
1 + Γcs l (r0 )/l (kxj k)
xj ∈Φ∩SM (i∗ )∩B c (o,r0 )
2λπr0 exp(−λπr02 )dr0
Z
2πλΓcs ∞ l (r0 )r
=
exp
dr 2λπr0 exp(−λπr02 )dr0
M
l
(r)
0
r0
Z ∞
2Γcs
(c)
v dv
=
exp − 1 −
M(α − 2)
0
∞,
cs
if M ≤ 2Γ
,
α−2
=
M
(α−2)
, if M > 2Γcs ,
(b)
Z
M (α−2)−2Γcs
(27)
α−2
where (a) is obtained by noting that x0 is the closest BS to the origin, (b) follows from the
PGFL for the PPP3 , and (c) is derived through change of variables (i.e. v = λπr02 ). It can be
observed that (27) is finite whenever M >
2Γcs
,
α−2
which is a sufficient condition for the finiteness
of E[Lcs (M)]. In particular, the equality holds in the first step of (27) when M = 1. As a result,
E[Lcs (1)] is finite if and only if α > 2Γcs + 2.
According to Lemma 2 and Lemma 3, the number of IA cycles to succeed in cell search (i.e.,
E[Lcs (M)]) may have a phase transition in terms of the number of BS beams M, depending on
the relation between the path loss exponent α and the detection threshold Γcs . This is detailed
in the following theorem.
Theorem 5: The number of IA cycles to succeed in cell search for the interference limited
networks satisfy the following:
•
If α > 2 + 2Γcs , E[Lcs (M)] < ∞ for the omni-directional BS antenna case, i.e., M = 1.
By the monotonicity of E[Lcs (M)] with respect to M, E[Lcs (M)] is guaranteed to be finite
for any M ≥ 1.
•
If α ≤ 2 + 2Γcs , E[Lcs (M)] = ∞ for M = 1, and E[Lcs (M)] < ∞ if M >
2Γcs
.
α−2
Therefore,
cs
according to the monotonicity of E[Lcs (M)], there exists a phase transition at M ∗ ∈ [2, 2Γ
],
α−2
such that E[Lcs (M)] = ∞ for M ≤ M ∗ , and E[Lcs (M)] < ∞ for M > M ∗ . In particular,
E[Lcs (M)] = ∞ for ∀M if α = 2, which means M ∗ = ∞.
The path loss exponent α depends on the propagation environment, and α = 2 corresponds to
a free space LOS scenario; while α increases as the environment becomes relatively more lossy
and scatter-rich, such as urban and suburban areas. In addition, the SINR detection threshold Γcs
3
Note that [38, Theorem 4.9] does not directly apply to the PGFL calculation here since f (x) = 1+Γcs l (r0 )/l (x) is larger than
1. However, we can use dominated convergence theorem
to prove that for PPP Φ with
the PGFL result
R
R
Q intensity measure Λ(·),
still holds if function f (x) satisfies f (x) ≥ 1 and R2 (f (x) − 1)Λ(dx) < ∞, i.e. E[ xi ∈Φ f (xi )] = exp( R2 (f (x) − 1)Λ(dx).
19
depends on the receiver decoding capability, which is typically within −10 dB and 0 dB [12].
Theorem 5 shows that in a lossy environment with α > 2 + 2Γcs , the typical user can detect
a nearby BS in a finite number of IA cycles on average. This is mainly because the relative
strength of the useful signal with respect to the interfering signals is strong enough. However,
when α ≤ 2 + 2Γcs , E[Lcs (M)] could be infinite due to the significant fraction of cell edge
users that have poor SIR coverage and therefore require a very high number of IA cycles to
succeed in cell search. Specifically, when M is very small (e.g., M = 1), the edge user is subject
to many strong nearby interferers inside every BS sector, so that the corresponding cell search
delay averaged over all users becomes infinity. However, as M increases, the BS beam sweeping
will create enough angular separation so that the nearby BSs to the edge user could locate in
different BS sectors. As a result, Lcs (M) is significantly decreased for cell edge users as M
increases, and therefore the phase transition for E[Lcs (M)] happens.
In summary, for an interference-limited network, we can always ensure the network to be in a
desirable condition with finite mean cell search delay by tuning the number of BS beams/antennas
M appropriately.
C. Cell Search Delay Distribution in Poisson Networks with Rayleigh Fading
The previous part is mainly focused on the mean number of IA cycles to succeed in cell search
E[Lcs (M, λ)], or equivalently the mean cell search delay. However, as shown in Theorem 3,
Theorem 4 and Theorem 5, E[Lcs (M, λ)] could be infinite under various settings, and there are
large variations of the performance between cell edge user and cell center user. Therefore, it is
also important to analyze the cell search delay distribution for system design.
Since the cell search delay Dcs (M, λ) depends on the spatial point process model for BSs
and the fading random variables at each IA cycle, its distribution is intractable in general. In
this section, we evaluate the distribution of the conditional mean cell search delay given the
distance from the typical user to its closest BS R0 , which is a random variable with PDF
fR0 (r0 ) = 2πλr0 exp(−λπr02 ). Specifically, we first derive the expected number of IA cycles to
succeed in cell search given R0 , i.e., E[Lcs (M, λ)|R0 ], which is a function of random variable R0
with mean E[Lcs (M, λ)]. For notation simplicity, we denote by Lcs (R0 , M, λ) , E[Lcs (M, λ)|R0 ]
for the rest of the paper. According to (5), we will evaluate the distribution of the following
20
conditional mean cell search delay:
Dcs (R0 , M, λ) , (Lcs (R0 , M, λ) − 1)T + Mτ.
(28)
The main reason to investigate the cell search delay conditionally on R0 is because R0 captures
1
√
2 λ
the location and therefore the signal quality of the typical user. In particular, R0 ≪
corresponds to the cell center user, while R0 ≫
1
√
2 λ
1
√
2 λ
corresponds to the cell edge user, where
represents the mean distance from the typical user to its nearest BS on the PPP Φ.
In order to derive Lcs (R0 , M, λ) in (28), we will first derive E[Lcs (M, λ)|R1 , R2 , ..., RM ],
where Ri denotes the distance from the typical user to its closest BS in the i-th BS sector (i.e.,
Ri = kxi0 k) for 1 ≤ i ≤ M.
Lemma 4: Given the distances from the typical user to its nearest BSs inside every BS sector
R1 , ..., RM , the mean number of IA cycles for cell search is:
E[Lcs (M, λ)|R1 , R2 , ..., RM ] =
M
∞ Y
X
fj (Ri , M, λ),
(29)
j=0 i=1
where fj (Ri ) denotes the probability that xi0 is detected in the first j IA cycles, which is:
j
X
Z
2λπ ∞
1
W kΓcs l (Ri )
j
exp −
)rdr .
(1 −
exp −
(−1)
fj (Ri , M, λ) =
k
P
M
M
(1
+
Γ
k
cs l (Ri )/l (r))
R
i
k=0
k
Proof: We can first prove E[Lcs (M, λ)|R1 , R2 , ..., RM ] = E[E[Lcs (M, λ)|Φ]|R1 , R2 , ..., RM ],
which is due to the tower property for conditional expectations. The rest of the proof follows
steps similar to those of Theorem 2, and therefore we omit the details.
Next we prove the following corollary to derive Lcs (R0 , M, λ) from E[Lcs (M, λ)|R1 , R2 , ..., RM ].
Corollary 1: For all i.i.d. non-negative random variables R1 , R2 ,...,RM with CCDF G(r), and
all functions F : [0, ∞)M → [0, ∞) which are symmetric, the following relation holds true:
E[F (R1 , R2 , ..., RM )| min(R1 , R2 , ..., RM ) = r] =
E[F (r, R2 , ..., RM )1{Rj >r,∀j6=1}]
.
(G(r))M −1
Proof: Denote by R0 = min(R1 , R2 , ..., RM ), then we can obtain (30) as follows:
E[F (R1 , R2 , ..., RM )|R0 = r]
E[F (R1 , R2 , ..., RM ) × 1|R0 −r|<ǫ ]
ǫ→0
P(|R0 − r| < ǫ)
= lim
(30)
21
E[F (R1 , R2 , ..., RM ) × 1({|Ri −r|<ǫ}∩{Rj >Ri ,∀j6=i}) ]
PM
k=1 P({|Rk − r| < ǫ} ∩ {Rj > Rk , ∀j 6= k})
= lim
PM
= lim
PM
ǫ→0
ǫ→0
=
i=1
PM
i=1
E[F (R1 , R2 , ..., RM )1({Rj >Ri ,∀j6=i}) ||Ri − r| < ǫ]
PM
k=1 P({Rj > Rk , ∀j 6= k}||Rk − r| < ǫ)
1
i=1 E[F (R1 , R2 , ..., RM ) ({Rj >Ri ,∀j6=i}) |Ri
PM
k=1 P({Rj > Rk , ∀j 6= k}|Rk = r)
= r]
,
the proof is completed by noting F is symmetric.
By taking F (R1 , R2 , ..., RM ) = E[Lcs (M, λ)|R1 , R2 , ..., RM ] in Corollary 1, E[Lcs (M, λ)|R0 ]
can directly obtained as follows.
Lemma 5: Given the distance from the typical user to the nearest BS R0 , the mean number
of IA cycles to succeed cell search is:
Z ∞
M −1
∞
X
λ2πr
λπr 2
λπ(M − 1)R02
fj (r, M, λ)
Lcs (R0 , M, λ) =
fj (R0 , M, λ)
,
exp(−
)dr exp
M
M
M
R0
j=0
where the function fj (r, M, λ) is defined in Lemma 4.
Lemma 5 provides a method to evaluate the cell search delay distribution under a general
setting. For noise limited networks and interference limited networks, we can obtain the following
simplified results.
Corollary 2: For the noise limited network, Lcs (R0 , M, λ) is given by:
P
α
R∞
Γcs W CN R0 N
∞
Γcs W CN r αN
j
))
{
))j
(1
−
exp(−
(1
−
exp(−
j=0
P
M
PM
R
0
λ2πr
λπr 2
M −1
exp(λπ MM−1 R02 ),
if R0 ≥ Rc ,
× M exp(− M )dr}
α
P∞
α
Γcs W CL R0 L j R ∞
N
Lcs (R0 , M, λ) =
)) { RC (1 − exp(− Γcs WPCMN r ))j
j=0 (1 − exp(−
PM
RR
α
2
× λ2πr
exp(− λπr
)dr + R0C (1 − exp(− Γcs WPCML r L ))j
M
M
2
λ2πr
)dr}M −1 exp(λπ MM−1 R02 ),
if R0 < Rc .
× M exp(− λπr
M
(31)
Corollary 2 can be easily proved from Lemma 5 and the fact that interference power is 0.
Corollary 3: For the interference limited network and the standard single-slope path loss
model with path loss exponent α > 2, Lcs (R0 , M, λ) is given by:
Lcs (R0 , M, λ) =
j
∞ X
X
j=0
2πλR02 H(k, α, Γcs)
j
exp −
(−1)
M
k
k=0
k
22
×
X
j
(−1)k kj
R∞
1
(1 −
2πλR20 H(k,α,Γcs )
M −1
exp −
M
1 + 2H(k, α, Γcs)
k=0
where H(k, α, Γcs) =
,
(32)
1
)rdr.
(1+Γ/r α )k
Proof: Since W = 0 and l (r) = Cr α , fj (Ri , M, λ) in Lemma 5 can be simplified as:
j
X
2πλR02 H(k, α, Γcs)
j
exp −
(−1)
fj (Ri , M, λ) =
.
k
M
k=0
k
(33)
Therefore, we can further obtain that:
Z
∞
R0
j
λπ
2
X
λπr 2
λ2πr
k j exp(− M (1 + 2H(k, α, Γcs ))R0 )
(−1)
exp(−
)dr =
. (34)
fj (r, M, λ)
k
M
M
1 + 2H(k, α, Γcs)
k=0
The proof can be completed by substituting (33) and (34) into Lemma 5.
IV. N UMERICAL E VALUATIONS
In this section, the distribution of the conditional mean cell search delay (28) is numerically
evaluated for both the noise limited scenario and the interference limited scenario. Specifically,
for the noise limited scenario, we consider a cellular network operating in the mmWave band
with carrier frequency fc = 73 GHz, bandwidth B = 2 GHz, and BS intensity λ = 100 BS/km2 .
The path loss exponents for LOS and NLOS links are 2.1 and 3.3 respectively, and the critical
distance is Rc = 50m. In addition, the OFDM symbol period is τ = 14.3 µs, and the IA cycle
length is chosen as T = 20 ms [13], [19]. As for the interference limited scenario, we consider
a cellular network with carrier frequency fc = 2 GHz, BS intensity λ = 100 BS/km2 , and a
standard single slope path loss model with path loss exponent α = 2.5. The OFDM symbol
period is τ = 71.4 µs, and the IA cycle length is T = 100 ms.
A. Conditional Expected Number of Cycles to Succeed in Cell Search
In order to evaluate the distribution of the conditional mean cell search delay, we first illustrate
Lemma 5. Specifically, we have simulated the cellular network with the directional cell search
protocol proposed in Section II-A, given the distance from the user to its nearest BS R0 . As
shown in Lemma 3 and Remark 5, the cell edge users will require a large number of cycles to
succeed in cell search. Therefore, we have set an upper bound for the number of cycles that a
user can try cell search, which is equal to 1500 cycles for the noise limited scenario and 100
23
5
1500
°
M = 4, = 90
°
M = 8, = 45
°
M = 18, = 20
M = 36, = 10 °
4
f = 73 GHz, B = 2 GHz
c
2
= 100 BS/km , Rc = 50m
500
L
= 2.1,
N
M = 4, = 90
M = 8, = 45 °
M = 12, = 30 °
°
M = 18, = 20
4.5
E[Lcs (M, )|R0 ]
1000
cs
0
E[L (M, )|R ]
°
= 3.3
3.5
3
2
= 100 BS/km ,
2.5
= 2.5
2
Lines: theory
Markers: simulation
Lines: theory
Markers: simulation
1.5
0
1
0
50
100
150
0
50
R0 (m)
(a) Noise limited networks
100
150
R0 (m)
(b) Interference limited networks
Fig. 2: Conditional expected number of cycles to succeed in cell search.
cycles for the interference limited scenario. Specifically, the infinite summation in Lemma 3 is
computed up to the 1500-th (100-th) term, and the simulation will treat a user as in outage if it
cannot be connected within 1500 (100) cycles.
Fig. 2 shows a close match between the analytical results and simulation results for both the
noise and interference limited scenarios, which is in line with Lemma 5. In addition, we can also
observe from Fig. 2 that the conditional expected number of cycles to succeed in cell search is
monotonically decreasing as the number of BS antennas/beams M increases, or as the distance
to the nearest BS R0 decreases.
B. Cell Search Delay Distribution in Noise Limited Networks
The cell search delay distribution for noise limited networks is numerically evaluated in this
part. Fig. 3 plots the CCDF of the conditional mean cell search delay Dcs (R0 , M, λ), which is
obtained by generating 106 realizations of R0 and computing the corresponding Dcs (R0 , M, λ)
through Corollary 2. We can observe from Fig. 3 that under the log-log scale, the tail distribution
function of Dcs (R0 , M, λ), i.e., P(Dcs (R0 , M, λ) ≥ t), decreases almost linearly with respect to t.
This indicates that the cell search delay is actually heavy-tailed and of the Pareto type. It can also
be observed from Fig. 3 that the tail distribution function satisfies limt→∞
− log P(Dcs (R0 ,M,λ)≥t)
log t
<1
for M = 4, 8, 18, 36. Therefore, the expected cell search delay is always infinite, which is in
line with Lemma 3.
24
Fig. 3 also shows that as the number of BS antennas M increases, the tail of Dcs (R0 , M, λ)
becomes lighter and thus the cell search delay for edge users is significantly reduced. For
example, the cell search delay for the 10th percentile user is almost 10 times smaller when
M increases from 18 to 36. In fact, increasing M will increase the SNR of cell edge users,
such that the number of IA cycles required for the edge users to succeed in cell search (i.e.,
Lcs (R0 , M, λ)) can be shortened. Since Dcs (R0 , M, λ) , T (Lcs (R0 , M, λ) − 1) + Mτ , and the
IA cycle length T is much larger than the OFDM symbol period τ , the tail distribution of
Dcs (R0 , M, λ) therefore becomes lighter as M increases despite having a higher beam-sweeping
overhead within every IA cycle.
10 0
10 -1
CCDF
fc = 73 GHz, B = 2 GHz
10
-2
= 100 BS/km 2, Rc = 50m
L
10 -3
10 -4
10 -2
= 2.1,
N
M increases
= 3.3
F(x) = 1/x
M = 4, = 90 °
°
M = 8, = 45
°
M = 18, = 20
°
M = 36, = 10
10 -1
10 0
10 1
10 2
10 3
10 4
Cell Search Delay (ms)
Fig. 3: Cell search delay distribution for noise limited networks.
Due to the heavy-tailed nature for the cell search delay distribution, Fig. 3 shows that there
exists an extremely large variation of the cell search delay performance from cell center users
to cell edge users. Fig. 4 plots the cell search delay for the 95th percentile users, as the number
of BS antennas M increases. Since the 95th percentile users are located at the cell center, they
are typically LOS to their serving BSs with sufficiently high isotropic SNR, and thus they can
succeed cell search in the first cycle that they initiates IA. Therefore, Fig. 4 shows that as M
increases, the cell search delay for the 95th percentile users increases almost linearly due to the
increase of the beam-sweeping overhead.
The cell search delay performance for the 50th percentile users, or the median users, is plotted
in Fig. 5. We can observe that in contrast to the mean cell search delay which is infinite, the
median delay is less than 1 ms for various BS antenna number M. When M is small, median
users do not have high enough SNR and thus they will need more than 1 IA cycles to succeed
25
95-th Percentile Cell Search Delay (ms)
0.6
fc = 73 GHz, B = 2 GHz
= 100 BS/km 2, Rc = 50m
0.5
L
= 2.1,
N
= 3.3
0.4
0.3
0.2
0.1
0
5
10
15
20
25
30
35
Number of BS Beams
Fig. 4: 95th percentile cell search delay for noise limited network.
in cell search. As M increases, the cell search delay for median users first decreases due to the
improved SNR and cell search success probability, until the median users could succeed cell
search in the first cycle that they initiates IA. Then the cell search delay will increase as M is
further increased, which is because the beam sweeping overhead becomes more dominant. The
optimal BS antenna number M is 12 (or 30◦ beamwidth) in Fig. 5, which corresponds to a cell
search delay of 0.31 ms.
50-th Percentile Cell Search Delay (ms)
0.55
fc = 73 GHz, B = 2 GHz
0.5
= 100 BS/km 2, Rc = 50m
L
= 2.1,
N
= 3.3
0.45
0.4
0.35
0.3
5
10
15
20
25
30
35
Number of BS Beams
Fig. 5: 50th percentile cell search delay for noise limited network.
C. Cell Search Delay Distribution in Interference Limited Networks
Similar to the noise limited scenario, we have evaluated the CCDF of cell search delay for
the interference limited scenario in Fig. 6 by generating 106 realizations of R0 and computing
the corresponding Dcs (R0 , M, λ) through Corollary 3.
26
Fig. 6 shows that the tail distribution function of Dcs (R0 , M, λ) decreases almost linearly
under the log-log scale, which means the distribution of Dcs (R0 , M, λ) is also heavy-tailed
under the interference limited scenario. However, in contrast to the noise limited scenario where
the overall mean cell search delay is always infinite, the phase transition for mean cell search
delay of the interference limited scenario can be observed from Fig. 6. Specifically, when the
cell search is performed omni-directionally (i.e., M = 1), Fig. 6 shows that the decay rate of the
tail satisfies limt→∞
− log P(Dcs (R0 ,M,λ)≥t)
log t
< 1, which indicates an infinite mean cell search delay.
As M increases to 4, 8, 12, Fig. 6 shows that limt→∞
− log P(Dcs (R0 ,M,λ)≥t)
log t
> 1, which leads to
a finite mean cell search delay. This observation is consistent with Theorem 5, which shows
that for the considered interference limited scenario with path loss exponent α = 2.5 and SINR
detection threshold Γcs = −4 dB, the mean cell search delay is infinite when M = 1, and finite
as long as M > 1.59.
It can also be observed from Fig. 6 that BS beam-sweeping can significantly reduce the cell
search delay for both the median users and edge users in the interference limited networks. For
example, when the number of BS antennas/beams M is 1, 4, 8, and 12, the corresponding cell
search delay for the 50th percentile user is 200 ms, 8.98 ms, 1.18 ms, and 0.9123 ms respectively,
while the corresponding cell search delay for the 10th percentile user is 3720 ms, 53.84 ms, 5.14
ms and 1.35 ms respectively. The main reason for such a performance gain in the interferencelimited network is that as M increases, beam-sweeping creates more angular separations from
the nearby BSs to the user, so that the number of IA cycles to succeed in cell search can be
effectively reduced, especially for edge users.
10 0
CCDF
10
-1
10 -2
M = 1, omni
M = 4, = 90 °
M = 8, = 45 °
M = 12, = 30 °
10 -3
F(x) = 1/x
2
= 100 BS/km , = 2.5
T = 100ms, = 71.4 s
10
-4
10 -1
10 0
10 1
10 2
10 3
10 4
Cell Search Delay (ms)
Fig. 6: Cell search delay distribution for interference limited network.
27
V. C ONCLUSIONS
This paper has proposed a mathematical framework to analyze the directional cell search
delay for fixed cellular networks, where the BS and user locations are static. Conditioned on the
BS locations, we have first derived the conditional expected cell search delay under the Palm
distribution of the user process. By utilizing a Taylor series expansion, we have further derived
the exact expression for the overall mean cell search delay in a Poisson cellular network with
Rayleigh fading channels. Based on this expression, the expected cell search delay in noiselimited network was proved to be infinite when the NLOS path loss exponent is larger than
2. By contrast, a phase transition for the expected cell search delay in the interference-limited
network was identified: the delay is finite when the number of BS beams/antennas is greater than
a threshold, and infinite otherwise. Finally, by investigating the distribution of the conditional
cell search delay given the distance to the nearest BS, the cell search delay for the edge user
was shown to be significantly reduced as the number of BS beams/antennas increases, which
holds true for both the noise and interference limited networks.
The framework developed in this paper provides a tractable approach to handle the spatial
and temporal correlations of user’s SINR process in cellular networks with fixed BS and user
locations. Future work will leverage the proposed framework to derive the random access phase
performance, the overall expected initial access delay, as well as the downlink throughput
performance for such fixed cellular networks. In addition, we will also extend the framework to
incorporate user beamforming or power control.
ACKNOWLEDGMENTS
This work is supported in part by the National Science Foundation under Grant No. NSFCCF-1218338 and an award from the Simons Foundation (#197982), both to the University of
Texas at Austin.
R EFERENCES
[1] E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-advanced for mobile broadband. Elsevier Science, 2011.
[2] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE Communications Magazine,
vol. 49, pp. 101–107, Jun. 2011.
[3] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez,
“Millimeter wave mobile communications for 5G cellular: It will work!,” IEEE Access, vol. 1, pp. 335–349, May 2013.
28
[4] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryanfar, “Millimeter-wave beamforming as an
enabling technology for 5G cellular communications: theoretical feasibility and prototype results,” IEEE Communications
Magazine, vol. 52, pp. 106–113, Feb. 2014.
[5] A. Ghosh, T. Thomas, M. C. Cudak, R. Ratasuk, P. Moorut, F. W. Vook, T. S. Rappaport, G. R. MacCartney, S. Sun,
S. Nie, et al., “Millimeter-wave enhanced local area systems: A high-data-rate approach for future wireless networks,”
IEEE Journal on Selected Areas in Communications, vol. 32, pp. 1152–1163, Jul. 2014.
[6] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Transactions
on Wireless Communications, vol. 9, pp. 3590–3600, Nov. 2010.
[7] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE
Communications Magazine, vol. 52, pp. 186–195, Feb. 2014.
[8] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling up MIMO:
Opportunities and challenges with very large arrays,” IEEE Signal Processing Magazine, vol. 30, pp. 40–60, Jan 2013.
[9] E. Bjornson, E. G. Larsson, and T. L. Marzetta, “Massive MIMO: ten myths and one critical question,” IEEE
Communications Magazine, vol. 54, pp. 114–123, Feb. 2016.
[10] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong, and J. C. Zhang, “What will 5G be?,” IEEE
Journal on Selected Areas in Communications, vol. 32, pp. 1065–1082, Jun. 2014.
[11] J. G. Andrews, T. Bai, M. Kulkarni, A. Alkhateeb, A. Gupta, and R. W. Heath, “Modeling and analyzing millimeter wave
cellular systems,” IEEE Transactions on Communications, vol. 65, pp. 403–430, Jan. 2017.
[12] C. N. Barati, S. A. Hosseini, M. Mezzavilla, T. Korakis, S. S. Panwar, S. Rangan, and M. Zorzi, “Initial access in millimeter
wave cellular systems,” IEEE Transactions on Wireless Communications, vol. 15, pp. 7926–7940, Dec. 2016.
[13] Y. Li, J. G. Andrews, F. Baccelli, T. D. Novlan, and J. C. Zhang, “Design and analysis of initial access in millimeter wave
cellular networks,” IEEE Transactions on Wireless Communications, vol. PP, no. 99, pp. 1–1, 2017.
[14] M. Karlsson and E. G. Larsson, “On the operation of massive MIMO with and without transmitter CSI,” in 2014 IEEE
15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp. 1–5, Jun. 2014.
[15] C. Shepard, A. Javed, and L. Zhong, “Control channel design for many-antenna MU-MIMO,” in Proceedings of the 21st
Annual International Conference on Mobile Computing and Networking, pp. 578–591, Sept. 2015.
[16] F. Baccelli and B. Blaszczyszyn, Stochastic Geometry and Wireless Networks: Volume II-Applications. Now Publishers
Inc, 2010.
[17] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Transactions
on Communications, vol. 59, pp. 3122–3134, Nov. 2011.
[18] M. Haenggi, J. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry and random graphs for the
analysis and design of wireless networks,” IEEE Journal on Selected Areas in Communications, vol. 27, pp. 1029–1046,
Sept. 2009.
[19] TS V5G.213, “Verizon 5G radio access (V5G RA); physical layer procedures,” Jun. 2016.
[20] Z. Pi, J. Choi, and R. Heath, “Millimeter-wave gigabit broadband evolution toward 5G: fixed access and backhaul,” IEEE
Communications Magazine, vol. 54, pp. 138–144, Apr. 2016.
[21] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and A. Ghosh, “Millimeter wave beamforming for wireless
backhaul and access in small cell networks,” IEEE Transactions on Communications, vol. 61, pp. 4391–4403, Oct. 2013.
[22] M. Giordani, M. Mezzavilla, and M. Zorzi, “Initial access in 5G mmwave cellular networks,” IEEE Communications
Magazine, vol. 54, pp. 40–47, Nov. 2016.
29
[23] M. Giordani, M. Mezzavilla, C. Barati, S. Rangan, and M. Zorzi, “Comparative analysis of initial access techniques in 5G
mmWave cellular networks,” in 2016 Annual Conference on Information Science and Systems, pp. 268–273, Mar. 2016.
[24] Y. Li, J. G. Andrews, F. Baccelli, T. D. Novlan, and J. C. Zhang, “Performance analysis of millimeter-wave cellular
networks with two-stage beamforming initial access protocols,” in 2016 50th Asilomar Conference on Signals, Systems
and Computers, pp. 1171–1175, Nov. 2016.
[25] F. Baccelli and B. Blaszczyszyn, “A new phase transition for local delays in MANETs,” in INFOCOM, 2010 Proceedings
IEEE, pp. 1–9, Apr. 2010.
[26] M. Haenggi, “The local delay in Poisson networks,” IEEE Transactions on Information Theory, vol. 59, pp. 1788–1802,
Mar. 2013.
[27] X. Zhang and M. Haenggi, “Delay-optimal power control policies,” IEEE Transactions on Wireless Communications,
vol. 11, pp. 3518–3527, Oct. 2012.
[28] S. K. Iyer and R. Vaze, “Achieving non-zero information velocity in wireless networks,” in 2015 13th International
Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), pp. 584–590, May 2015.
[29] R. B. Waterhouse, D. Novak, A. Nirmalathas, and C. Lim, “Broadband printed sectorized coverage antennas for millimeterwave wireless applications,” IEEE Transactions on Antennas and Propagation, vol. 50, pp. 12–16, Aug. 2002.
[30] A. Alkhateeb, Y. H. Nam, M. S. Rahman, J. Zhang, and R. W. Heath, “Initial beam association in millimeter wave cellular
systems: Analysis and design insights,” IEEE Transactions on Wireless Communications, vol. 16, pp. 2807–2821, May
2017.
[31] M. Hussain and N. Michelusi, “Throughput optimal beam alignment in millimeter wave networks,” arXiv preprint
arXiv:1702.06152, Feb. 2017.
[32] S. N. Chiu, D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry and its applications. John Wiley & Sons, 2013.
[33] F. Baccelli and B. Blaszczyszyn, Stochastic Geometry and Wireless Networks: Volume 1: THEORY. Now Publishers Inc,
2010.
[34] X. Zhang and J. G. Andrews, “Downlink cellular network analysis with multi-slope path loss models,” IEEE Transactions
on Communications, vol. 63, pp. 1881–1894, Mar. 2015.
[35] T. Bai and R. Heath, “Coverage and rate analysis for millimeter-wave cellular networks,” IEEE Transactions on Wireless
Communications, vol. 14, pp. 1100–1114, Feb. 2015.
[36] M. D. Renzo, “Stochastic geometry modeling and analysis of multi-tier millimeter wave cellular networks,” IEEE
Transactions on Wireless Communications, vol. 14, pp. 5038–5057, Sept. 2015.
[37] S. Singh, M. Kulkarni, A. Ghosh, and J. Andrews, “Tractable model for rate in self-backhauled millimeter wave cellular
networks,” IEEE Journal on Selected Areas in Communications, vol. 33, pp. 2196–2211, Oct. 2015.
[38] M. Haenggi, Stochastic geometry for wireless networks. Cambridge University Press, 2013.
| 7cs.IT
|
Gene Expression Time Course Clustering with Countably Infinite
Hidden Markov Models
Matthew J. Beal
Praveen Krishnamurthy
Department of Computer Science and Engineering
State University of New York (SUNY) at Buffalo
Buffalo, NY 14260-2000
{mbeal,pk35}@cse.buffalo.edu
1
Abstract
to learn not just co-expression, but causal relationships
that may help elucidate co-regulation as well.
Most existing approaches to clustering gene
expression time course data treat the different time points as independent dimensions
and are invariant to permutations, such as reversal, of the experimental time course. Approaches utilizing HMMs have been shown
to be helpful in this regard, but are hampered by having to choose model architectures with appropriate complexities. Here we
propose for a clustering application an HMM
with a countably infinite state space; inference in this model is possible by recasting it
in the hierarchical Dirichlet process (HDP)
framework (Teh et al. 2006), and hence we
call it the HDP-HMM. We show that the
infinite model outperforms model selection
methods over finite models, and traditional
time-independent methods, as measured by
a variety of external and internal indices for
clustering on two large publicly available data
sets. Moreover, we show that the infinite
models utilize more hidden states and employ richer architectures (e.g. state-to-state
transitions) without the damaging effects of
overfitting.
Two problematic issues hamper practical methods for
clustering gene expression time course data: first, if
deriving a model-based clustering metric, it is often
unclear what the appropriate model complexity should
be; second, the current clustering algorithms available
cannot handle, and therefore disregard, the temporal
information. This usually occurs when constructing
a metric for the distance between any two such genes.
The common practice for an experiment having T measurements of a gene’s expression over time is to consider the expression as positioned in a T -dimensional
space, and to perform (at worse spherical metric) clustering in that space. The result is that the clustering algorithm is invariant to arbitrary permutations
of the time points, which is highly undesirable since
we would like to take into account the correlations between all the genes’ expression at nearby or adjacent
time points.
Introduction
There are a large number of popular techniques for
clustering gene expression data, with the goal being to
elucidate the many different functional roles of genes
that are players in important biological processes. It is
said that genes that cluster with similar expression—
that is, are co-expressed—serve similar functional roles
in a process (see, for example, Eisen et al. 1998). Bioinformaticians have more recently had access to sets of
time-series measurements of genes’ expression over the
duration of an experiment, and have desired therefore
In i.i.d. data sets, the model complexity issue has been
recently and successfully tackled by using Dirichlet
process mixture models, and in particular countably
infinite Gaussian mixture models (examples can be
found in Rasmussen 2000, Wild et al. 2002, Medvedovic and Sivaganesan 2002, Dubey et al. 2004, Medvedovic et al. 2004). However, these models are not applicable to time series data unless ones uses the unsavory concatenation of time points mentioned above
(for example, in the work of Medvedovic and Sivaganesan (2002), an i.i.d. infinite mixture of Gaussians
is applied to Spellman’s time series data). To address this second issue of time-dependence, parametric models such as state-space models and differential
equation models have been employed, but the modelcomplexity issue still needs to be tackled either with
heuristic model selection criteria (Iyer et al. 1999), or
approximate Bayesian methods such as BIC (Ramoni
et al. 2002) or variational Bayes (Beal et al. 2005).
For continuous time series, modeling spline parame-
ters has been popular (Bar-Joseph et al. 2003, Heard
et al. 2006). For discrete time series, which we consider in this paper, random walk models (Wakefield
et al. 2003) have been proposed. The most advanced
analysis thus far is the mixture of HMMs approach of
Schliep et al. (2005), wherein BIC (Schwarz 1978) and
an entropic criterion are used for model selection. The
approach outlined in this paper also uses an HMM, but
we side-step the question of model selection by using
a flexible nonparametric Bayesian mixture modeling
framework, which allows the model to have a countably infinite number of hidden states (i.e. a countably
infinite by countably infinite transition matrix).
We present a simple but powerful temporal extension
to the recently introduced Hierarchical Dirichlet Process mixture model (HDP) of Teh, Jordan, Beal and
Blei (2004,2005), in the form of a model that is coined
the hierarchical Dirichlet process hidden Markov model
(HDP-HMM). As we describe below, this extension
allows us to address both of the issues noted above.
Moreover, we can still provide a measure of similarity
between any two genes’ time courses, by examining
the probabilistic degree of overlap in the hidden state
trajectories; this is possible despite these state spaces
being countably infinite.
The paper is arranged as follows. In Section 2 we
briefly review the HDP framework, then show how the
countably infinite HMM is a particular (though nontrivial) instantiation within this framework, then describe a straightforward similarity measure for pairs
of sequences. In Section 3 we present results of timecourse clustering experiments on two publicly available
gene data sets, for which ground truth labels are provided, measuring performance with respect to a variety
of so-called external and internal indices. We conclude
in Section 4 by suggesting directions for future work
and expansion of the HDP-HMM.
2
The infinite HDP-HMM
In this section we present a hidden Markov model with
a countably infinite state space, which we call the hierarchical Dirichlet process hidden Markov model (HDPHMM) by way of its relationship to the HDP framework of Teh et al. (2006). We begin with an overview
of the HDP according to this work, and then recast the
infinite HMM into this framework. Previous research
by the first author on infinite HMMs (Beal et al. 2002)
provided an approximate sampling scheme for inference and learning but was unable to prove its correctness. By recasting the infinite HMM as a constrained
HDP, we will show that we can have a functioning
sampling scheme at our disposal, as explained below.
2.1
Hierarchical Dirichlet Processes
The HDP considers problems involving groups of data,
where each observation within a group is a draw from
a mixture model, and where it is desirable to share
mixture components between groups. Consider first a
single group of data, (xi )ni=1 ; if the number of mixture
components is unknown a priori and is to be inferred
from the data, then it is natural to consider a Dirichlet
process mixture model, as depicted in Figure 1(a) (for
an exposition on DP mixtures see Neal 1998). Here,
the well known clustering property of the Dirichlet process provides a nonparametric prior for the number of
mixture components within each group, with the following generative model:
G | α0 , G0 ∼ DP(α0 , G0 )
θi | G ∼ G
xi | θi ∼ F (θi )
(1)
for each i,
for each i,
where G0 is a base measure of the DP, α0 ≥ 0 is a concentration parameter, DP(·, ·) is a Dirichlet process,
and θi is a parameter drawn from G. Each datum xi
is then drawn from a distribution F (·) parameterized
by θi . In our models and experiments hereafter, F (·)
is a Gaussian density. Ferguson (1973) showed that
draws from a DP are discrete with probability one,
and therefore, in a sufficiently large data set, there
will be several i for which the θi are identical; this
gives rise to the natural clustering phenomenon giving
DP mixtures their name.
Now, consider several groups, J, of data, denoted
nj J
((xji )i=1
)j=1 ; in such a setting it is natural to consider
sets of Dirichlet processes, one for each group, and we
still desire to tie the mixture models in the various
groups. We therefore consider a hierarchical model,
specifically one in which each of the child Dirichlet
processes in the set has a base measure that is itself distributed according to a Dirichlet process. Such a base
measure being again discrete with probability one, the
child Dirichlet processes necessarily share atoms, thus
ensuring that the the mixture models in the different groups necessarily share mixture components. The
generative model is given by:
G0 | γ, H
Gj | α0 , G0
θji | Gj
xji | θji
∼ DP(γ, H)
∼ DP(α0 , G0 )
∼ Gj
∼ F (θji )
(2)
for each j,
for each j and i,
for each j and i.
Figure 1(b) shows the graphical model, where there
is a plate not only over data, n, but also over the J
(non-overlapping) groups of data. In this model the
number of mixture components is unknown a priori
for each group, and also for the data as a whole.
H
γ
G0
γ
β
γ
β
G
α0
Gj
α0
πj
α0
πk
xi
θji
xji
n
(a) DP
nj
H
zji
φk
xji
H
nj
J
(b) HDP
v0
J
(c) HDP stick-break
φk
v1
v2
vT
y1
y2
yT
8
θi
8
α0
G0
(d) HMM-HDP stick-breaking
Figure 1: Graphical model descriptions of the HDP-HMM as it compares to the HDP: (a) a DP mixture model; (b) the
original HDP model; (c) the stick-breaking interpretation of the original HDP: mixing proportions for the jth group are
drawn from common weights β, themselves drawn from a stick-breaking process, and then each item xji is drawn from a
mixture model with mixing proportions π j . (d) An unraveled HDP graphical model, wherein the next-state distribution
is a current-state-dependent DP mixture model; each item y t is drawn from a mixture model (having mixture indicator
vt ) with mixing proportions π vt−1 determined by the previous hidden state.
2.2
Recasting the Hidden Markov Model
The most straightforward route to understanding the
connection between a hidden Markov model and the
HDP described above is to first realize the stickbreaking characterization (Sethuraman 1994) of the
HDP, given below and depicted in Figure 1(c),
β | γ ∼ Stick(γ)
(3)
π j | α0 , β ∼ DP(α0 , β)
zji | π j ∼ π j
φk | H ∼ H
xji | zji , (φk )∞
k=1 ∼ F (φzji ) ,
where the stick-breaking construction (Stick) gives rise
to weights β = (βk )∞
k=1 , with
βk0 ∼ Beta(1, γ)
βk = βk0
k−1
Y
(1 − βl0 ) .
(4)
l=1
The advantage of this representation is that it makes
explicit the generation of one (countably infinite) set
of parameters (φk )∞
k=1 ; the jth group has access to various of these parameters should it need them to model
nj
its data (xji )i=1
, depending on the sampled mixing
proportion π j .
Recall that a hidden Markov model (HMM) is a doubly
stochastic Markov chain in which a sequence of multinomial “state” variables (v1 , v2 , . . . , vT ) are linked via
a state transition matrix, and each element yt in a sequence of “observations” (y1 , y2 , . . . , yT ) is drawn independently of the other observations conditional on
vt (Rabiner 1989). This is essentially a dynamic variant of a finite mixture model, in which there is one
mixture component corresponding to each value of the
multinomial state. Note that the HMM involves not
a single mixture model, but rather a set of mixture
models: one for each value of the current state. That
is, the “current state” vt indexes a specific row of the
transition matrix, with the probabilities in this row
serving as the mixing proportions for the choice of the
“next state” vt+1 . Given the next state vt+1 , the observation yt+1 is drawn from the mixture component
indexed by vt+1 .
Thus, to consider a nonparametric variant of the HMM
which allows an unbounded set of states, we must consider a set of DPs, one for each value of the current
state. Moreover, these DPs must be linked, because
we want the same set of “next states” to be reachable
from each of the “current states.” This amounts to
the requirement that the atoms associated with the
state-conditional DPs should be shared—exactly the
framework of the hierarchical DP. Thus, we simply replace the set of conditional finite mixture models underlying the classical HMM with an HDP, and the resulting model, an HDP-HMM, provides an alternative
to methods that place an explicit parametric prior on
the number of states or make use of model selection
methods to select a fixed number of states (e.g. Stolcke
and Omohundro 1993).
While there exist two Gibbs sampling methods for
the HDP—one based on an extension of the Chinese
Restaurant Process (CRP Aldous 1985) called the Chinese Restaurant Franchise (CRF), and the other based
on an auxiliary variable method (both described in
Teh et al. 2006)—only the auxiliary variable method
is straightforward to implement for our HDP-HMM,
and is used in our experiments shown below. In
fact, in work that served as inspiration for the HDPHMM (Beal et al. 2002), a sampler was presented that
resembles the CRF scheme, but was necessarily ap-
proximate to reduce the time complexity of inference.1
The two-level urn model presented in that earlier work
can be related to the HDP framework, by describing
the latter using the stick-breaking formalism. In particular, consider the unraveled hierarchical Dirichlet
process representation shown in Figure 1(d). The parameters in this representation have the following distributions, c.f. (3):
β | γ ∼ Stick(γ)
(5)
∞
π k | α0 , β ∼ DP(α0 , β) vt | vt−1 , (π k )k=1 ∼ π vt−1
φk | H ∼ H
yt | vt , (φk )∞
k=1 ∼ F (φvt ) ,
where we assume for simplicity that there is a distinguished initial state v0 . If we now consider the
CRF representation of this model, it turns out that
the result is equivalent to the coupled urn model of
Beal et al. (2002). The advantage of this representation is that we can use an auxiliary variable sampling
scheme that was designed for the HDP. As described
above, the current instantiation, and most importantly
the order, of the state variables (v1 , . . . , vT ), defines a
grouping of the data (y1 , . . . , yT ) into groups indexed
by the previous-state. Given this grouping, the settings of β and α0 , the group-specific DP mixtures can
be sampled independently. On sampling, the indicators (v1 , . . . , vT ) change, and hence the grouping of
the data changes. Thus, given that we have access to a
countably infinite set of hidden states, the HDP-HMM
can be thought of as an HDP with an ever-shifting,
countably infinite number of groups (metaphor: countably infinite tables in each of countably infinite restaurants, all sharing choices of dishes).
Last, as in the HDP, the HDP-HMM has hyperpriors on the hyperparameters α0 and γ, both gamma
distributed with shape a· and inverse scale b· like so:
α0 ∼ Gamma(aα0 , bα0 ) ), γ ∼ Gamma(aγ , bγ ); we sample over α0 and γ during the auxiliary variable sampling scheme, so these are integrated out of the model.
3
Analysis of Experiments
3.1
Data & Sequence Similarity measure
We used two publicly available sets for the analysis:
i) (Iyer) the gene expression time course data of Iyer
et al. (1999), consisting of 517 genes’ expressions across
12 time points; the expressions are log-normalized and
1
There, bookkeeping for the CRF representation is very
difficult: Adopting the well used CRP metaphor of customers entering a restaurant, in the CRF representation
we have multiple restaurants, and sampling the table that
a customer y t sits at will influence the restaurant in which
the following customer y t+1 must dine, and so on, resulting
in a highly coupled system.
standardized to have log expression 1 at time t = 1 for
all genes. Each gene is labeled belonging to a cluster
1–10 and a further outlier cluster is denoted as cluster ‘-1’. We assume that these labels are the result
of biological expert modification following a preliminary Eisen (simple correlation in T dimensions) analysis (Eisen et al. 1998). ii) (Cho) the second data set is
the expression data described in Cho et al. (1998) consisting of 386 genes’ expression across 17 time points,
similarly normalized, with hand-labeled genes into one
of 5 clusters.
We compare our HDP-HMM model to the standard
HMM, referred to as finite HMM here after. For the
finite HMM, we ran experiments with 7 different seed
values and averaged over the various scores (explained
below), in order to minimize the effects of initialization in EM. We define the (probabilistic measure of)
dissimilarity between two genes’ time courses for finite HMM as the (c, d)th element of a matrix P (e.g.
of size 517×517 for Iyer), which is the probability that
the two time courses of each gene have identical hidden
state trajectories. This can be computed straightforwardly after an E step in the Baum-Welch algorithm.
Denoting the posterior over the hidden state at time
(c) (c)
t of the cth gene sequence by p(vt |y1:T , Θ), where Θ
are the current parameters of the HMM, then log Pcd
is straightforwardly given by
k
T
X
X
(d)
(d)
(c)
(c)
p(vt = r|y1:T , Θ)p(vt = r|y1:T , Θ), (6)
log
t=1
r=1
and therefore Pcd (= Pdc ) measures the probability of
two genes, c and d, having traversed similar entire hidden trajectories. We use − log Pcd as a measure of
divergence (to be thought of as clustering distance)
between genes c and d.2
An analogous measure of divergence, or dissimilarity,
can be computed for the HDP-HMM. In an infinite
model, the posterior distribution over hidden state trajectories is represented as a set of samples, and the
above quantity can be calculate simply from an empirical computation over the samples of trajectories
taken over very long MCMC runs. Since the posterior
samples always consist of represented hidden states, we
do not suffer from the countably infinite state space.
A similar method is used in the (i.i.d.) clustering using infinite mixture of Gaussians work of Rasmussen
(2000) and Wild et al. (2002), but here we have extended to be a measure of similarity over sequences.
2
More thoroughly one should compute a similarity that
involves pairwise marginals over time as well, which would
require a dynamic programming computation of similarity, which for the HDP-HMM is current research. However, we found the above approximation sufficient in our
(c)
experiments. Note that the marginals used in (6), p(vt =
(c)
r|y1:T , Θ), are still obtained using forward-backward.
3.2
Agreement with provided labels
We have used common external and internal indices
to assess the quality of clustering obtained by various
methods. Refer to the Appendix for more details on
the definition of these metrics (note that DB∗ is the
only score for which smaller values are better). We
also use a recently introduced index called purity, as
defined in Wild et al. (2002) which is a tree-based metric. Given a measure of dissimilarity between genes—
provided by simple Eisen correlation, finite HMMs or
HDP-HMMs—we can construct a dendrogram (we use
average linkage), and on this binary-tree based representation we can fix a number of clusters, C, by severing the tree at this point, and force a putative labeling
of each of the genes into each of C possible labels.
3.3
Results
We compare simple correlation (Eisen, no time depedence) analysis to both finite HMMs of sizes varying from k = 1, . . . , 40 and with the HDP-HMM with
several settings of its hyper-hyperparameters. For
the HDP-HMM, the auxiliary variable Gibbs sampling
consisted of 100,000 burnin samples, collecting 250
posterior samples thereafter having a spacing of 750
samples between each.
In Tables 1 and 2, we display a subset of the indices as
we vary C for the Iyer data set only. Here, bγ denotes
the hyper-hyperparameter setting over γ for the HDPHMM, and k denotes the number of hidden states used
in the finite HMM run. Noting the trade-off between
sensitivity and specificity, and also the indifference to
bγ , we decided that for comparative purposes we would
fix C = 11 for Iyer. For the Cho results (not shown here
for space), we used C = 5, for similar reasons.
Table 2 shows a comparison of Eisen, finite HMM and
HDP-HMM for various results of external and internal
indices. A better visualization is given in Figure 2 for
some of the Iyer results (Cho results are more impressive but are omitted due to space) in Figure 2. For
clarity, we highlight entries in each column that perform the best, and reiterate that for DB∗ index lower is
better (we do not consider the result of the finite HMM
with k = 1 since this degenerates to the case where all
genes are in the same class (− log Pcd = 0 ∀c, d).
We note several trends: First, the degree of variation of all indices for different settings of the hyperhyperparameter bγ is very small, suggesting that at
this level of the Bayesian hierarchy the setting of the
priors does not influence the learning of the model.
Moreover, for any index, the variation over k of the
finite HMM is much larger. Second, it is clear that
considering time in the HMM—especially in infinite
Table 1: Choice of number of clusters, C, for Iyer data.
bγ = 1, C = 1
5
10
11
15
20
bγ = 3, C = 1
5
10
11
15
20
rand crand jacc spec sens
0.16
0.00
0.16 0.16 1.00
0.72
0.35
0.33 0.35 0.89
0.73
0.36
0.34 0.36 0.88
0.73
0.36
0.34 0.36 0.88
0.80
0.41
0.36 0.42 0.70
0.82
0.38
0.33 0.43 0.57
0.16
0.00
0.16 0.16 1.00
0.66
0.28
0.29 0.30 0.90
0.73
0.35
0.33 0.35 0.83
0.75
0.36
0.33 0.36 0.82
0.80
0.39
0.35 0.41 0.69
0.80
0.39
0.35 0.42 0.67
models—is advantageous compared to the simple correlation analysis of Eisen (already established Schliep
et al. 2005). Third, there is evidence that the finite
HMM is overfitting in the Iyer data set according to
the Sil and DB indices, and for Cho according to several of the external indices; the HDP-HMM does not fit
any one particular model, but integrates over a countably infinite set of models. Fourth, it is clear from the
Table 2 that the vast majority of highlighted “winners”
are for the HDP-HMM (last 7 rows), which shows a
dramatic success over the finite HMM and Eisen (time
independent) analyses.
Inferred architecture: Generally speaking the finite
models show no improvement in performance beyond
around k = 10 hidden states. It is interesting, therefore, that in Figure 3(a) we find that for a wide range
of bγ settings the HDP-HMM uses in excess of k = 20
represented classes; the reason for this is that the architectures of the countably infinite and finite models
are quite different, as shown in the transition matrices
of Figures 3(b) and 3(c). The HDP-HMM has almost
three times as sparse connectivity as its finite counterpart, with many states having only one or two possible
destination states.
4
Conclusion and further directions
We have described the infinite HMM in the framework of the HDP, in which an auxiliary variable Gibbs
sampling scheme is feasible. The HDP-HMM is in
fact a CRF with a countably infinite number of tables in a countably infinite number of restaurants, all
potentially sharing common dishes. We have shown
for two time course gene expression data sets that the
HDP-HMM performs similarly and, for most scenarios, better than the best finite HMMs found by model
selection. We find that HMMs outperform the standard Eisen analysis based on simple correlation of the
T -dimensional sequence vector, which treats the time
points as independent. We used common measures of
external and internal indices, including a tree-based in-
Table 2: Effect of varying the complexity k of the finite models, and varying the hyper-hyperparameter bγ .
dataset
Iyer et al. (1999) (C = 11)
Cho et al. (1998) (C = 5)
index rand crand jacc sens spec sil dunn DB∗ puri rand crand jacc sens spec sil dunn DB∗ puri
Eisen 0.80
0.38
0.33 0.63 0.41 0.55 1.542 0.70 0.58 0.77
0.43
0.41 0.68 0.51 0.37 1.357 0.78 0.58
k=1
2
3
4
5
6
7
8
14
20
30
40
0.20
0.58
0.75
0.77
0.77
0.75
0.76
0.75
0.71
0.73
0.74
0.75
0.01
0.26
0.34
0.37
0.38
0.36
0.39
0.36
0.34
0.36
0.36
0.38
0.17
0.29
0.32
0.34
0.35
0.34
0.36
0.34
0.33
0.34
0.34
0.35
0.99
0.86
0.72
0.73
0.76
0.78
0.81
0.78
0.84
0.82
0.84
0.82
0.17
0.32
0.37
0.40
0.41
0.38
0.40
0.38
0.36
0.38
0.38
0.39
∞
0.27
0.49
0.53
0.50
0.51
0.54
0.53
0.59
0.57
0.54
0.55
∞
1.358
1.445
1.685
1.329
1.427
1.389
1.518
1.625
1.583
1.449
1.509
0.00
0.76
0.56
0.53
0.62
0.63
0.63
0.62
0.55
0.58
0.60
0.62
0.48
0.42
0.47
0.52
0.52
0.53
0.54
0.53
0.56
0.57
0.55
0.55
0.25
0.73
0.62
0.63
0.65
0.70
0.72
0.72
0.67
0.65
0.65
0.66
0.00
0.33
0.22
0.21
0.27
0.30
0.35
0.33
0.30
0.30
0.30
0.31
0.23
0.35
0.31
0.29
0.33
0.33
0.37
0.35
0.35
0.35
0.35
0.36
0.99
0.63
0.74
0.68
0.73
0.65
0.69
0.64
0.74
0.82
0.82
0.80
0.23
0.44
0.34
0.34
0.38
0.41
0.44
0.44
0.40
0.38
0.38
0.39
∞
0.22
0.44
0.40
0.40
0.41
0.44
0.50
0.56
0.61
0.60
0.63
∞
1.201
1.606
1.610
1.540
1.548
1.381
1.767
1.538
1.343
1.445
1.421
0.00
0.92
0.70
0.70
0.72
0.75
0.72
0.65
0.57
0.55
0.56
0.54
0.47
0.54
0.49
0.48
0.55
0.54
0.56
0.54
0.58
0.60
0.59
0.59
bγ = 0.25
0.5
1
2
3
4
5
0.77
0.78
0.73
0.77
0.75
0.80
0.78
0.37
0.40
0.36
0.38
0.36
0.42
0.40
0.34
0.36
0.34
0.35
0.33
0.37
0.36
0.74
0.80
0.88
0.80
0.82
0.76
0.80
0.38
0.39
0.36
0.38
0.36
0.42
0.39
0.53
0.54
0.62
0.55
0.59
0.54
0.51
1.511
1.510
1.436
1.525
1.788
1.878
1.374
0.57
0.58
0.47
0.55
0.54
0.60
0.63
0.54
0.60
0.52
0.58
0.54
0.56
0.60
0.74
0.80
0.73
0.79
0.73
0.71
0.72
0.40
0.46
0.39
0.46
0.40
0.36
0.38
0.41
0.42
0.40
0.42
0.40
0.38
0.39
0.77
0.63
0.77
0.66
0.78
0.76
0.79
0.46
0.56
0.45
0.54
0.46
0.43
0.44
0.61
0.44
0.62
0.51
0.63
0.46
0.63
1.885
1.578
1.948
1.512
1.853
1.574
2.189
0.59
0.68
0.59
0.60
0.50
0.71
0.51
0.60
0.60
0.61
0.62
0.60
0.59
0.61
0.45
1
0.45
finite
eisen
b=1
b=4
0.95
0.4
0.4
0.35
0.9
0.35
0.3
0.85
0.25
0.3
0.8
0.2
0.75
0.25
0.15
0.7
0.1
0.2
0.15
finite
eisen
b=1
b=4
1
2
3
4
5
6
7
8
9
0.65
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
1
2
3
4
5
6
7
8
(a) spec
9
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
0
1
2
3
4
5
6
7
8
(b) sens
9
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
(c) crand
0.8
0.65
0.7
0.6
0.6
0.55
0.55
0.5
0.45
0.45
0.35
0.5
0.5
0.4
0.3
0.4
0.2
finite
eisen
b=1
b=4
0.1
0
finite
eisen
b=1
b=4
0.05
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
(d) DB∗
finite
eisen
b=1
b=4
0.4
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
(e) purity
finite
eisen
b=1
b=4
0.3
0.25
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 25 30 35 40
(f) silhouette
Figure 2: Relative performance of finite (k = 1, . . . , 40, horizontal axis), infinite (bγ = 1, 4, horizontal solid and dashed
lines), and standard Eisen (correlation only, horizontal thick line) algorithms on a subset of the indices given in Table 2,
for a cluster number C = 11, for the Iyer data set.
dex, purity. We also find that the HDP-HMM models
are learning quite different architectures for the hidden state dynamics. In current work we are examining
more closely the prevalent paths through the hidden
states, which may elucidate interesting regulatory networks at play; we will report the biological significance
of these in a forthcoming article.
In this paper we have not considered a leave-one-out
analysis—either in terms of classification or in terms
of density estimation—of a single gene using a model
trained on the remainder. This is partly because sampling to calculate test likelihoods in the HDP-HMM is
very time-consuming. However, previous preliminary
work (Teh et al. 2005) on a simpler case of learning
450
0.25
0.5
1
2
3
4
5
400
350
300
250
200
150
100
50
0
20
21
22
23
24
25
26
27
28
29
30
(a) represented classes
31
32
(b) HDP-HMM transitions
(c) finite HMM transitions
Figure 3: Analysis of hidden states in HDP-HMM (for Iyer data). (a) Distribution of number of represented classes in the
HDP-HMM models. Shown are stacked values for various values of the (hyper-)hyperparameter bγ = (0.25, 0.5, 1, 2, 3, 4, 5),
demonstrating that the mass over number of represented classes does not shift dramatically even over orders of magnitude
of this high level hyperparameter; (b) Equivalent transition matrix for HDP-HMM (bγ = 0.25, 11.7% non-zero entries):
each row is a source state and its entries are transition probabilities that sum to 1, and brighter squares denote higher
probability; (c) Transition matrix of the same size for the finite model—evidently less sparse (29.4% non-zero entries).
sequences of letters forming sentences on the Alice’s
Adventures in Wonderland showed that perplexity of
test sentences was minimized using the HPD-HMM,
as compared to Maximum Likelihood trained HMMs,
Maximum A Posteriori trained HMMs, and variational
Bayesian HMMs (Beal 2003). Also, Medvedovic and
Sivaganesan (2002) show robustness of their infinite
i.i.d. mixtures of Gaussians model as compared to finite mixtures, which is another reason to expect our
time series analysis to perform well on such analyses.
Finally, there are a host of exciting variants to the
HDP-HMM. There are nested group models, which
may be useful for capturing ontological information,
and we are working on a countably infinite switching state-space model, as well as variants on the DP
mixture model formalism in terms of Pitman-Yor processes that may have attractive properties in the domain of time-series modeling.
Acknowledgements
We acknowledge support from NSF award 0524331,
Yee Whye Teh for HDP code and discussion, and the
helpful comments of anonymous reviewers.
References
D. Aldous. Exchangeability and related topics. In École
d’été de probabilités de Saint-Flour XIII–1983, pages 1–
198. Springer, Berlin, 1985.
Z. Bar-Joseph, G. Gerber, D. Gifford, T. Jaakkola, and
I. Simon. Continuous representations of time series gene
expression data. Journal of Computational Biology, 10
(3–4):241–256, 2003.
M. J. Beal. Variational Algorithms for Approximate
Bayesian Inference. PhD thesis, Gatsby Computational Neuroscience Unit, University College London,
UK, 2003.
M. J. Beal, Z. Ghahramani, and C. E. Rasmussen. The infinite hidden Markov model. In Advances in Neural In-
formation Processing Systems 14, Cambridge, MA, 2002.
MIT Press.
M. J. Beal, F. L. Falciani, Z. Ghahramani, C. Rangel, and
D. Wild. A Bayesian approach to reconstructing genetic
regulatory networks with hidden factors. Bioinformatics, 21:349–356, February 2005.
R. J. Cho, M. J. Campbell, E. A. Winzeler, L. Steinmetz, A. Conway, L. Wodicka, T. G. Wolfsberg, A. E.
Gabrielian, D. Landsman, D. J. Lockhart, and R. W.
Davis. A genome-wide transcriptional analysis of the
mitotic cell cycle. Molecular Cell, 2(1):65–73, 1998.
A. Dubey, S. Hwang, C. Rangel, C. E. Rasmussen,
Z. Ghahramani, and D. L. Wild. Clustering protein sequence and structure space with infinite Gaussian mixture models. In R. B. Altman, A. K. Dunker, L. Hunter,
and T.E. Klein, editors, Pacific Symposium in Biocomputing 2004, pages 399–410. World Scientific Publishing,
Singapore, 2004.
M. B. Eisen, P. T. Spellman, P. O. Brown, and D. Botstein.
Cluster analysis and display of genome-wide expression
patterns. Proceedings of the National Academy of Sciences of the USA, 95(25):14863–14868, December 1998.
T. S. Ferguson. A Bayesian analysis of some nonparametric
problems. Annals of Statistics, 1(2):209–230, 1973.
N. A. Heard, C. C. Holmes, and D. A. Stephens. A quantitative study of gene regulation involved in the immune
response of Anopheline mosquitoes: An application of
Bayesian hierarchical clustering of curves. Journal of the
American Statistical Association, 101(473):18–29, 2006.
V. R. Iyer, M. B. Eisen, D. T. Ross, G. Schuler, T. Moore,
J. C. F. Lee, J. M. Trent, L. M. Staudt, J. Hudson Jr.,
M. S. Boguski, D. Lashkari, D. Shalon, D. Botstein, and
P. O. Brown. The Transcriptional Program in the Response of Human Fibroblasts to Serum. Science, 283
(5398):83–87, 1999.
M. Medvedovic and S. Sivaganesan. Bayesian infinite mixture model based clustering of gene expression profiles.
Bioinformatics, 18(9):1194–1206, 2002.
M. Medvedovic, K. Y. Yeung, and R.E. Bumgarner.
Bayesian mixture model based clustering of replicated
microarray data. Bioinformatics, 20(8):1222–1232, 2004.
|SS|+|DD|
|SS|+|SD|+|DS|+|DD|
.
R. M. Neal. Markov chain sampling methods for Dirichlet
process mixture models. Technical Report 9815, Department of Statistics, University of Toronto, 1998.
1. Rand index: Rand =
L. R. Rabiner. A tutorial on hidden Markov models and
selected applications in speech recognition. Proceedings
of the IEEE, 77, 1989.
3. CRand index; (C)orrected for chance assignments:
CRand =
Pm Pn `nij ´ `N ´−1 Pm `ni ´ Pn `nj ´
− 2
j=1
i=1 2
i=1
j=1 2
2
hP
`nj ´i Pm `ni ´ Pn `nj ´
m `ni ´ Pn
1
−
i=1 2
i=1 2
j=1 2
j=1 2
2
M. F. Ramoni, P. Sebastiani, and I.S. Kohane. Cluster
analysis of gene expression dynamics. Proc. National
Academy of Sciences of the USA, 99(14):9121–6, 2002.
C. E. Rasmussen. The infinite Gaussian mixture model. In
Advances in Neural Information Processing Systems 12,
Cambridge, MA, 2000. MIT Press.
A. Schliep, I. G. Costa, C. Steinhoff, and A. Sch onhuth.
Analyzing gene expression time-courses. IEEE Trans.
Comp. Biology and Bioinformatics, 2(3), 2005.
2. Jaccard coefficient: Jaccard =
|SS|
|SS|+|SD|+|DS|
.
where nij is the number of points in Pi and Cj , ni is
the number of points in the Pi , and nj the number of
points in Cj .
4. The usual definitions for Sensitivity and Specificity:
|SS|
|SS|
Sens. =
,
Spec. =
.
|SS| + |SD|
|SS| + |DS|
G. Schwarz. Estimating the dimension of a model. The
Annals of Statistics, 6:461–464, 1978.
A.2
J. Sethuraman. A constructive definition of Dirichlet priors. Statistica Sinica, 4:639–650, 1994.
Internal indices are computed to quantitatively assess clustering in the absence of provided labels, and attempt to
evaluate cohesion (how similar are points in same clusters),
and separation (how dissimilar are points in different clusters). Usually these indices are computed with a Euclidean
distance metric, but to preserve the integrity of our analysis we used the − log Pcd dissimilarity given by (6).
A. Stolcke and S. Omohundro. Hidden Markov model induction by Bayesian model merging. In S. J. Hanson,
J. D. Cowan, and C. L. Giles, editors, Advances in Neural Information Processing Systems 5, pages 11–18, San
Francisco, CA, 1993. Morgan Kaufmann.
Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei. Sharing clusters among related groups: Hierarchical Dirichlet Processes. In L. K. Saul, Y. Weiss, and L. Bottou,
editors, Advances in Neural Information Processing Systems 17, Cambridge, MA, 2005. MIT Press.
Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei. Hierarchical Dirichlet processes. Journal of the American
Statistical Society, 2006. To appear.
J. Wakefield, C. Zhou, and S. Self. Modelling gene expression over time: curve clustering with informative prior
distributions. In J. M. Bernardo, M. J. Bayarri, J. O.
Berger, A. P. Dawid, D. Heckerman, A. F. M. Smith, and
M. West, editors, Bayesian Statistics 7, Proc. 7th Valencia International Meeting, pages 721–732. OUP, 2003.
D. L. Wild, C. E. Rasmussen, Z. Ghahramani, J. Cregg,
B. J. de la Cruz, C-C. Kan, and K. A. Scanlon. A
Bayesian approach to modelling uncertainty in gene expression clusters, 2002. Extended conference abstract in
3rd Int. Conf. on Systems Biology, Stockholm, Sweden.
Appendices
Cluster validation is usually done by computation of indices, which signiify the quality of clustering, based either
on a comparison with “ground-truth” labels (external) or
without such comparison and relying on inherent qualities
of the dendrogram or putative labels (internal).
A.1
External indices
Let n be the number of data points, F 0 = {F1 , ..., Fm }
be the possible labels, let C 0 = {C1 , ..., Cn } be the clustering obtained by a clustering algorithm. Define two
n × n incidence matrices, F and C, s.t. Fij = 1 if both
the ith point and the j th point belong to same cluster in
F 0 , 0 otherwise. And, Cij = 1, if both the ith point and
the j th point belong to same cluster in C 0 , 0 otherwise.
Defining the following categories: SS = δ(Cij , 1)δ(Fij , 1),
DD = δ(Cij , 0)δ(Fij , 0), SD = δ(Cij , 1)δ(Fij , 0), and
DS = δ(Cij = 0)δ(Fij , 1), we use the following indices:
Internal indices
Silhouette: For a given cluster, Cj (j = 1, ..., m), this
method assigns to each sample of Cj a quality measure,
s(i) (i = 1, ..., n), known as the Silhouette width. The Silhouette width is a confidence indicator on the membership
of the ith sample in cluster Cj , defined as
s(i) = (b(i) − a(i)) / max{a(i), b(i)}
where a(i) is the average distance between the ith sample and all of the samples in Cj , and b(i) is the minimum average distance between the ith sample and all samples in Ck (k = 1, ..., c; k 6= j). For a given cluster, Cj
(j = 1, ...,P
m), it is possible to calculate a cluster Silhouette
Sj = n−1 n
j=1 s(i), which characterizes the heterogeneity
and isolation properties of such
P a cluster, and a Global
Silhouette value, GSu = m−1 m
j=1 Sj .
Dunn’s: This index identifies sets of clusters that are
compact and well separated. For any partition, U, produced by a clustering algorithm, let Ci represent the ith
cluster, the Dunn’s validation index, D, is defined as:
8
9
ff=
<
∆0 (Ci , Cj )
,
D(U ) = min
min
1≤i≤m :1≤i≤m
max1≤k≤m {∆(Ck )} ;
j6=i
0
where ∆ (Ci , Cj ) defines the distance between clusters Ci
and Cj (intercluster distance); ∆(Ck ) represents the intracluster distance of cluster Ck , and m is the number of clusters of partition. The main goal of this measure is to maximize intercluster distances whilst minimizing intra-cluster
distances; large values of D correspond to good clusters.
Davies-Bouldin: This index is defined as:
DB(U ) =
ff
c
∆(Ci ) + ∆(Cj )
1 X
max
,
m i=1 j6=i
δ(Ci , Cj )
where U , δ(Ci , Cj ), ∆(Ci ), ∆(Cj ) and m are defined as in
equation (7). Small values of DB correspond to clusters
that are compact, and whose centers are far away from
each other. Therefore, smaller DB is preferred.
| 5cs.CE
|
POLISH MODELS AND SOFIC ENTROPY
arXiv:1411.1510v4 [math.DS] 31 Dec 2015
BEN HAYES
Abstract. We deduce properties of the Koopman representation of a positive entropy probability measurepreserving action of a countable, discrete, sofic group. Our main result may be regarded as a “representationtheoretic” version of Sinaı̌’s Factor Theorem. We show that probability measure-preserving actions with
completely positive entropy of an infinite sofic group must be mixing and, if the group is nonamenable,
have spectral gap. This implies that if Γ is a nonamenable group and Γ y (X, µ) is a probability measurepreserving action which is not strongly ergodic, then no action orbit equivalent to Γ y (X, µ) has completely
positive entropy. Crucial to these results is a formula for entropy in the presence of a Polish, but a priori
noncompact, model.
Contents
1. Introduction
2. Preliminaries
2.1. Notational Remarks
2.2. Preliminaries on Sofic Groups
3. Definition of Entropy in the Presence of a Polish Model
4. Spectral Consequences of Positive Entropy
4.1. Representation Theoretic Preliminaries
4.2. Proofs of the Main Applications
References
1
6
6
7
9
19
19
21
25
1. Introduction
This paper is concerned with the structural consequences of positive entropy of probability measurepreserving actions of groups. Measure-theoretic entropy for actions of Z is classical and goes back to the
work of Kolmogorov and Sinaı̌. Roughly speaking, it measures the randomness of the action of Z. It was
realized by Kieffer in [27] that one could replace Z with the weaker condition of amenability of the group.
Amenability requires a sequence of non-empty finite subsets of the group over which one can average in
an approximately translation invariant way. Abelian groups, nilpotent groups and solvable groups are all
amenable, whereas the free group on r letters is not if r ≥ 1. Entropy for amenable groups is well established
as a useful quantity in ergodic theory: it can be computed in many interesting cases (although it is not easy
in general), when it is positive it reveals interesting structure on the action, and it has some useful general
and intuitive properties.
The most fundamental examples in entropy theory are Bernoulli shifts. If Γ is a countable, discrete group
and (Ω, ω) is a standard probability space the Bernoulli shift over (Ω, ω) is the probability measure-preserving
action Γ y (Ω, ω)Γ defined by
(gx)(h) = x(g −1 h), x ∈ ΩΓ , g, h ∈ Γ.
Bernoulli shifts over amenable groups are completely classified by their entropy. If Γ is an infinite, amenable
group and Γ y (X, µ) is a free, ergodic, probability measure-preserving action with positive entropy, then
Γ y (X, µ) factors onto a Bernoulli shift. In fact, in this situation, Γ y (X, µ) factors onto any Bernoulli
shift of entropy at most that of Γ y (X, µ). This is known as Sinaı̌’s factor theorem and it was proved for
Γ = Z by Sinaı̌ in [42] and for general amenable groups by Ornstein-Weiss in [35]. Sinaı̌’s factor theorem
Date: January 5, 2016.
2010 Mathematics Subject Classification. 37A35,47A67, 37A55, 46L89,
Key words and phrases. Sofic entropy, Sinaı̌’s factor theorem, noncommutative harmonic analysis.
1
2
BEN HAYES
is a fundamental result in entropy theory, as it shows that Bernoulli factors “capture” all the entropy of a
probability measure-preserving action of an amenable group and in some sense shows that entropy is simply
a measure of the amount of “Bernoulli-like” behavior an action has.
In groundbreaking work, L. Bowen defined entropy for probability measure-preserving actions of sofic
groups assuming the existence of a finite generating partition (see [5]). This assumption was then removed
by Kerr-Li who also defined entropy for actions of sofic groups on compact, metrizable spaces (see [25]).
We refer the reader to see Section 2.2 for the precise definition. Sofic groups form a vastly larger class
of groups than amenable groups, it is known that amenable groups, free groups, residually finite groups
and linear groups are all sofic and soficity is closed under free products with amalgamation over amenable
subgroups (see [16],[15],[17],[36],[38]). Thus sofic entropy is a considerable extension of entropy for actions
of an amenable group defined by Kieffer. Roughly, a group is sofic if it has “almost actions” on finite
sets which are “almost free.” A sequence of such “almost actions” is called a sofic approximation. Entropy
for a probability measure-preserving action Γ y (X, µ) of a sofic group is then defined as the exponential
growth rate of the number of finitary models there are of the action which are compatible with the fixed
sofic approximation. Using sofic entropy, Bowen showed that if two Bernoulli shifts over a sofic group are
isomorphic, then their base entropies are the same. Kerr-Li reproved this result with a more direct proof
when the base space has infinite entropy in [24].
Since the subject is fairly young there are relatively few known structural consequences of positive measure
entropy for actions of arbitrary sofic groups. For example, Sinaı̌’s factor theorem is not known for sofic groups.
Previous consequences of entropy for actions of sofic groups have either been for topological actions or for
specific groups. In [26], Kerr-Li prove that actions with positive topological entropy must exhibit some
chaotic behavior (for example, they must be Li-Yorke chaotic). When the acting group is a free group one
can consider another form of measure entropy, defined by Bowen in [6], called f -invariant entropy. The
f -invariant entropy is roughly a “randomized” version of sofic entropy. Interesting consequences have been
given by Seward in [40] for the case of f -invariant entropy, but those only apply when the group is a free
group. After the appearance of our preprint, Meyerovitch showed in [33] that positive sofic entropy implies
that almost every stabilizer of the action is finite.
In this paper, we will deduce structural consequences of positive measure entropy for actions of arbitrary
sofic groups. Our applications are to spectral properties of such actions. To the best of our knowledge,
aside from the results in our paper, the results of Meyerovitch are the only ones which deduce properties
of an action of a general sofic group assuming only that the action has positive entropy. Recall that if
Γ y (X, µ) is a probability measure-preserving action of a countable discrete group Γ, we have an induced
unitary representation ρΓy(X,µ) : Γ → U(L2 (X, µ)) given by
(ρΓy(X,µ) (g)ξ)(x) = ξ(g −1 x).
We use ρ0Γy(X,µ) for the restriction of ρΓy(X,µ) to L2 (X, µ) ⊖ C1. The representation ρ0Γy(X,µ) is called the
Koopman representation of Γ y (X, µ). Properties of a probability-measure preserving action which only
depend upon the Koopman representation are called spectral properties. The Koopman representation has
played a significant role in ergodic theory since the early days of the subject, being the means to deduce von
Neumann’s Mean Ergodic Theorem which the Ergodic Theorem relies upon as a first step. Additionally,
many other fundamental properties such as compactness, weak mixing, mixing and ergodicity are all spectral
properties.
Our results show that one canonical representation of a group plays a special role in entropy theory. Recall
that the left regular representation of a group λ : Γ → U(ℓ2 (Γ)) is defined by (λ(g)f )(x) = f (g −1 x). We will
sometimes use λΓ to specify the group. We will also use λΓ,R (or λR if the group is clear) for the orthogonal
representation which is the restriction of λΓ to ℓ2 (Γ, R).
Let ρj : Γ → U(Hj ) be two unitary representations of a countable discrete group Γ. We say that ρ1 , ρ2
are singular and write ρ1 ⊥ ρ2 if no nonzero subrepresentation of ρ1 embeds into ρ2 . By subrepresentation,
we mean a restriction of ρj , j = 1, 2 to a closed, Γ-invariant, linear subspace. As is customary, we will often
call a closed, Γ-invariant, linear subspace a subrepresentation as well. The terminology of singularity comes
from the case Γ = Z. If µ is a Borel measure on T = R/Z, then we have a natural unitary representation
ρµ : Z → U(L2 (T, µ))
POLISH MODELS AND SOFIC ENTROPY
3
given by
(ρµ (n)ξ)(θ) = e2πinθ ξ(θ).
It is easy to check that ρµ ⊥ ρν if and only if µ ⊥ ν. Similar analysis can be done for any abelian group
(replacing T by the Pontryagin dual). Thus, singularity of representations is a natural generalization to
noncommutative groups of singularity of measures.
If Γ y (X, µ) is a probability measure-preserving action of a countable, discrete sofic group and Σ is
a sofic approximation of Γ (see Definition 2.1 for the precise definition of a sofic approximation) we use
hΣ,µ (X, Γ) for the entropy of Γ y (X, µ) with respect to Σ as defined by Bowen, Kerr-Li.
Theorem 1.1. Let Γ be a countably infinite discrete sofic group with sofic approximation Σ. Let Γ y
(X, M, µ) be a measure-preserving action where (X, M, µ) is a standard probability space. Suppose that
H ⊆ L2 (X, µ) is a subrepresentation such that M is generated (up to sets of measure zero) by
If ρ0Γy(X,µ)
{f −1 (A) : f ∈ H, A ⊆ C is Borel}.
H
is singular with respect to the left regular representation, then
hΣ,µ (X, Γ) ≤ 0.
It is well-known that if (Ω, ω) is a standard probability space, then for the Bernoulli action Γ y (Ω, ω)Γ
we have
(1)
ρ0Γy(Ω,ω)Γ ∼
= λ⊕∞
Γ .
We mentioned before that Sinaı̌’s factor theorem is not known for sofic groups. Note that if Γ y (X, µ) factors
embeds into ρ0Γy(X,µ) . In this manner, Theorem 1.1 may
onto a Bernoulli shift, then by (1) we have that λ⊕∞
Γ
be regarded as a weak version of Sinaı̌’s Factor Theorem for sofic groups. It shows that at the representationtheoretic level, an action of a sofic group with positive entropy must contain a subrepresentation of the
Koopman representation of a Bernoulli shift. We can thus think of this theorem as a “representationtheoretic version” of Sinaı̌’s Factor Theorem. It is also the first result which indicates that positive entropy
actions of sofic groups must behave in a manner “similar” to Bernoulli shifts.
We can say even more than Theorem 1.1 if we assume a stronger version of positive entropy. Recall that
an action Γ y (X, µ) has completely positive entropy with respect to a sofic approximation Σ if whenever
Γ y (Y, ν) is a factor of Γ y (X, µ) and (Y, ν) is not a one-atom space then hΣ,ν (Y, Γ) > 0. The following
is easy from Theorem 1.1.
Corollary 1.2. Let Γ be a countable discrete sofic group with sofic approximation Σ. Suppose that Γ y (X, µ)
has completely positive entropy with respect to Σ. Then the Koopman representation of Γ y (X, µ) is
embeddable into the infinite direct sum of the left regular representation.
Corollary 1.2 was proved for Γ amenable by Dooley and Golodets in [14]. Here we should mention that
Dooley-Golodets actually prove that if Γ y (X, µ) has completely positive entropy and Γ is amenable,
then the Koopman representation of Γ y (X, µ) is isomorphic to an infinite direct sum of the left regular
representation. This is an easy consequence of Corollary 1.2 and Sinaı̌’s factor theorem. Since Sinaı̌’s factor
theorem is not known for sofic groups, we do not know if ρ0Γy(X,µ) ∼
= λ⊕∞ (instead of just ρ0Γy(X,µ) embeds
⊕∞
into λ ) when Γ is sofic.
From Corollary 1.2, we automatically deduce other important structural properties of completely positive
entropy actions of a sofic group, specifically mixing and spectral gap. A probability measure-preserving
action Γ y (X, µ) is said to:
(i): be mixing if limg→∞ µ(gA ∩ B) = µ(A)µ(B) for all measurable A, B ⊆ X,
(ii): be strongly ergodic if for every sequence An of measurable subsets of X with µ(gAn ∆An ) → 0 for all
g ∈ Γ we have that µ(An )(1 − µ(An )) → 0,
(iii): have spectral gap if for every sequence ξn ∈ L2 (X, µ) with kgξn − ξn k2 → 0 for all g ∈ Γ, it is true
that kξn k2 → 0.
It is easy to see that spectral gap implies strong ergodicity.
4
BEN HAYES
Corollary 1.3. Let Γ be a countable discrete sofic group with sofic approximation Σ. Suppose that Γ y (X, µ)
has completely positive entropy with respect to Σ.
(i): If Γ is infinite, then Γ y (X, µ) is mixing.
(ii): If Λ is any nonamenable subgroup of Γ, then Λ y (X, µ) is strongly ergodic (in fact it has spectral
gap).
There is another approach to nonamenable entropy, called Rokhlin entropy, due to Seward in [41]. It
has the advantage of being easy to define and being defined for all groups, but the disadvantage of being
extremely difficult to compute. Rokhlin entropy is an upper bound for sofic entropy, but there are no known
cases where one can prove an action has positive Rokhlin entropy without using that is has positive sofic
entropy. We mention that completely positive Rokhlin entropy can be defined in a similar manner. After
the appearance of our preprint, Alpeev in [1] proved that actions with completely positive Rohklin entropy
are weakly mixing. His approach is completely elementary. It is easy to see that completely positive sofic
entropy implies completely positive Rokhlin entropy. Whether or not actions of a nonamenable group with
completely positive Rohklin entropy are mixing, strongly ergodic, or have spectral gap all appear to be
open. It seems to be very difficult to deduce properties of the Koopman representation of an action from
the assumption that the action has positive Rokhlin entropy.
Part (ii) of the above Corollary is rather special to nonamenable groups. Recall that two probability
measure-preserving actions Γ y (X, µ), Λ y (Y, ν) of countable discrete groups Γ, Λ are said to be orbit
equivalent if there is a measure space isomorphism θ : (X, µ) → (Y, ν) so that θ takes the Γ-orbits to Λ-orbits,
i.e. θ(Γx) = Λθ(x) for almost every x ∈ X. Orbit equivalence theory is an area of much current interest
relating to operator algebras, ergodic theory and group theory. Strong ergodicity is an invariant of the orbit
equivalence class of the action. Our corollary shows that if Γ y (X, µ) is a probability measure-preserving
action of a nonamenable group and if the action is not strongly ergodic, then no action orbit equivalent to
Γ y (X, µ) has completely positive entropy. Our result indicates that entropy for actions of nonamenable
groups may have nontrivial consequences for orbit equivalence theory. It is a celebrated, and deep, fact that
all ergodic actions of an amenable group are orbit equivalent. This is due to Connes-Feldman-Weiss and
Ornstein-Weiss (see [12],[34]). Thus entropy for actions of amenable groups cannot have any consequences
for orbit equivalence theory.
Spectral gap and strong ergodicity are important properties with many applications. Spectral gap has
connections to expander graphs (see [3],[2],[4], [31]), orbit equivalence rigidity (see [21],[37]) and number
theory (see [39]). Spectral gap is also related to the Banach-Ruziewicz problem which asks if Lebesgue
measure is the unique, finitely-additive, rotation-invariant, probability measure on the sphere defined on all
Lebesgue measurable sets (solved independently by Margulis [32] and Sullivan [43]). It is well known that
no action of an amenable group is strongly ergodic. The consequences entropy has for strong ergodicity,
spectral gap and orbit equivalence indicate that entropy for nonamenable groups may be used to deduce
phenomena not present in the realm of ergodic theory of amenable groups. This reveals the importance of
generalizing entropy to actions of nonamenable groups.
We briefly outline the key differences in our approach and the approach of Dooley-Golodets to prove
Corollary 1.2. Dooley-Golodets first prove Theorem 1.1 when Γ = Z. Fourier analysis reduces Theorem 1.1
for Γ = Z to the fact that if µ, ν are mutually singular, Borel, probability measures on the circle, then there
is a continuous function f on the circle with 0 ≤ f ≤ 1 so that f is “close” to the constant function 1 in
L2 (µ), but is “close” to zero in L2 (ν). This fact is a simple exercise in measure theory. Dooley-Golodets
then deduce Corollary 1.2 for Γ amenable from the case Γ = Z by using that all amenable groups are orbit
equivalent to the integers.
Our approach is to prove Theorem 1.1 for a general sofic group, by simply replacing the harmonic analysis
for Γ = Z with noncommutative harmonic analysis for a general group. The assumption about singularity of
measures is then replaced by singularity of representations of groups. The representation theory of a group
is captured by its universal C ∗ -algebra and so it is natural to replace the algebra of continuous functions on
T with the C ∗ -algebra of the group. One can then characterize singularity of representations of an arbitrary
group in a manner similar to the preceding paragraph. In short, we abstract the harmonic analysis for the
case Γ = Z to noncommutative harmonic analysis for a nonabelian group. This approach uses essentially no
structure of the group and removes the orbit equivalence techniques in the approach of [14], which are only
valid in the case of amenable groups.
POLISH MODELS AND SOFIC ENTROPY
5
Although this approach using noncommutative harmonic analysis may make our methods seem abstract
and esoteric, the proof of Theorem 1.1 is essentially self-contained. After the input of the aforementioned
noncommutative harmonic analysis techniques, as well as basic facts about Borel measures on Polish spaces,
the rest of the proof of Theorem 1.1 is elementary. It only relies on basic consequences of the finitedimensional spectral theorem and simple volume counting estimates. Additionally, the noncommutative
harmonic analysis techniques used lie at the basics of C ∗ -algebra theory. On the other hand, the fact that
every action is amenable groups is orbit equivalent to the integers is fairly deep. So in the amenable case we
have discovered a more direct and elementary proof of Corollary 1.2.
Intuitively, entropy should be some measure of randomness of the system. Theorem 1.1 shows indeed that
positive entropy actions must exhibit some randomness properties. For example, we can view the left regular
representation as the representation which exhibits the perfect amount of mixing. We can also use Theorem
1.1 to show that highly structured actions, like compact actions, must have nonpositive sofic entropy.
Corollary 1.4. Let Γ be a countable discrete sofic group with sofic approximation Σ. Suppose that α : Γ →
Aut(X, µ) has image contained in a compact (for the weak topology) subgroup of Aut(X, µ). Then for the
action given by α,
hΣ,µ (X, Γ) ≤ 0.
In a previous version of this article, we used our techniques to prove that distal measure-theoretic action
have entropy zero. After the appearance of this version, Alpeev proved in [1] that measure distal actions of
an arbitrary group have Rohklin entropy zero. Then, Burton proved in [10] that distal actions have naive
entropy zero if the group contains a copy of Z. Naive entropy zero implies Rohklin entropy zero which in
turn implies sofic entropy zero. These proofs are both elementary, whiles ours is arguably not. We have
elected to remove this section from the current version of this article, as we are actually going to prove more
general results in [19] which will give structural results of a probability measure-preserving action of a sofic
group relative to its Pinsker factor. The results in [1] imply that Rohklin entropy decreases for compact
extensions (and our results in [19] prove the same result for sofic entropy) and it appears that it is unknown
whether or not naive entropy decreases under compact extensions.
An important new tool we use to prove all of these theorems is Polish models. If Γ y (X, µ) is a
probability measure-preserving action of a countable discrete group, a topological model for the action is an
action Γ y (Y, ν) isomorphic to Γ y (X, µ) where Y is a separable, metrizable topological space, ν is a
Γ-invariant Borel probability measure on Y and the action is by homeomorphisms. Roughly one can think of
this as giving a topology on X so that the action is by homeomorphisms. For actions on standard probability
spaces compact models always exist. Moreover, Kerr-Li in [25] show that one can compute entropy in the
presence of a compact model in a manner which uses the topology. Many of the computations for sofic
entropy have used the compact model formalism (see [28],[25],[18]). To prove the above theorems, we give
a definition of sofic entropy in the presence of a Polish model (that is where Y is merely assumed to be
a Polish space, i.e. a completely metrizable separable topological space). We remark here that R. Bowen
defined (see [8]) topological entropy for uniformly continuous automorphisms of a metric space, proving that
it was invariant under uniformly continuous conjugacies. Our approach is slightly different here, we do not
require the homeomorphisms to be uniformly continuous and R. Bowen did not consider measure-preserving
actions.
Since compact models always exist, we should mention why we decided to consider the case of a Polish
model. For this, let us mention a natural way of obtaining Polish models. Given a probability measurepreserving action Γ y (X, M, µ), we say that a family F of measurable functions X → C is generating
if M is the smallest complete, Γ-invariant sigma-algebra of sets in X which makes all the elements of F
measurable. Associated to a family of generators, one can canonically produce a topological model in the
following manner. Define
Φ : X → CF ×Γ
by
Φ(x)(f, g) = f (g −1 x)
and let Γ y CF ×Γ be the shift action given by
(gx)(f, h) = x(f, g −1 h).
6
BEN HAYES
Setting ν = Φ∗ µ, we have that Γ y (CF ×Γ , ν) is a topological model of our action. We can thus think
of topological models as a “presentation-theoretic” approach to ergodic theory analogous to presentation
theory of groups. If F ⊆ L∞ (X, µ), then the topological model one produces above is a compact model. In
general the topological model is Polish. Thus Polish models are a canonical way of dealing with unbounded
generators. This is relevant because all of our assumptions are about functions in L2 (X, µ) and not L∞ (X, µ).
There are of course ways of turning a family of generators which are unbounded into family of generators
which are bounded. One can employ cut-off functions, or compose with injective continuous maps R →
[−1, 1]. We warn the reader that all of these attempts to reduce to a family of bounded generators destroy
our representation-theoretic hypotheses and so we will really need to deal with unbounded generators and,
as a consequence, Polish models.
We remark here that if we take the above topological model associated to family of generators in L∞ (X, µ)
then we essentially recover the operator algebraic approach to sofic entropy by Kerr-Li in [25]. Thus the
Polish model approach to sofic entropy may be regarded as a generalization of the operator algebra approach
given by Kerr-Li in [25] to a family of unbounded generators. The crucial aspect of Polish spaces which
allows us to equate our definition of entropy for a Polish model to that of Bowen and Kerr-Li is tightness
of a single probability measure on a Polish space. Tightness roughly asserts that, up to a small error, the
probability measure the probability measure is supported on a compact set. There are examples of separable
metrizable spaces where not every Borel probability measure is tight (e.g. see the remarks after Theorem 23
of [44]) and this is the reason we need our spaces to be Polish. The assumption of a topological model being
Polish is also natural since the canonical topological model associated to a family of generators is always
Polish.
Acknowledgments Much of this work was done while I was still a PhD student at UCLA. I am very
grateful to the kind hospitality and stimulating environment at UCLA. I would like to thank Lewis Bowen
for suggesting the problem of computing sofic entropy of Gaussian actions in the program “Von Neumann
algebras and ergodic theory of group actions” at Insitut Henri Poincaré in 2011. The solution to this problem
ultimately led to this work. I would like to thank Stephanie Lewkiewicz for many interesting discussion about
Polish models.
2. Preliminaries
2.1. Notational Remarks. We will use ρ for a representation. We will thus have to forego the usual
practice in sofic entropy of using ρ for a metric, and will instead use ∆.
If A, B are sets we use B A for all functions f : A → B. If A = {1, . . . , n} we will use B n instead of
{1,...,n}
B
. If C is another set and φ : B → C is a function, we use φn : B n → C n for the map
φn (f ) = f ◦ φ.
If X is a Polish space, and f ∈ Cb (X) we use kf k for the uniform norm of f. We will use kf kCb (X) if the
space X is not clear from the context.
Let (A, ∆) be a pseudometric space. For subsets C, B of A, and ε > 0 we say that C is ε-contained in
B and write C ⊆ε B if for all c ∈ C, there is a b ∈ B so that ∆(c, b) < ε. We say that S ⊆ A is ε-dense if
A ⊆ε S. We use Sε (A, ∆) for the smallest cardinality of a ε-dense subset of A. If C ⊆δ B are subsets of A,
then
S2(ε+δ) (C, ∆) ≤ Sε (B, ∆).
If A is a finite set we will use uA for the uniform probability measure on A. We will typically write un instead
of u{1,...,n} . For x ∈ Cn , kxk2 will denote the ℓ2 norm with respect to un unless otherwise stated. There are
certain times when we will have to use k · kℓ2 (n) , k · kℓ2 (n,un ) at the same time. In such an instance we will
use notation which specifies which ℓ2 -norm we are using. Additionally, we will use
h·, ·iℓ2 (n) ,
2
2
h·, ·iℓ2 (n,un )
for the inner products on ℓ (n), ℓ (n, un ) when there is potential confusion. If it is not otherwise specified,
then
h·, ·i
refers to the inner product with respect to ℓ2 (n, un ).
POLISH MODELS AND SOFIC ENTROPY
7
We say that N ⊆ A is ε-separated if for every n1 6= n2 in N we have ∆(n1 , n2 ) > ε. We use Nε (A, ∆) for
the smallest cardinality of a ε-separated subset of A. Note that
(2)
N2ε (A, ∆) ≤ Sε (A, ∆) ≤ Nε (A, ∆),
and that if A ⊆ B, then
Nε (A, ∆) ≤ Nε (B, ∆).
2.2. Preliminaries on Sofic Groups. We use Sn for the symmetric group on n letters. If A is a set, we
will use Sym(A) for the set of bijections of A.
Definition 2.1. Let Γ be a countable discrete group. A sofic approximation of Γ is a sequence Σ = (σi : Γ →
Sdi ) of functions (not assumed to be homomorphisms) so that
lim udi ({1 ≤ k ≤ di : σi (g)σi (h)(k) = σi (gh)(k)}) = 1, for all g, h ∈ Γ
i→∞
lim udi ({1 ≤ k ≤ di : σi (g)(k) 6= σi (h)(k)}) = 1, for all g 6= h in Γ.
i→∞
We will call Γ sofic if it has a sofic approximation.
It is known that all amenable groups and residually finite groups are sofic. Also, it is known that soficity
is closed under free products with amalgamation over amenable subgroups (see [17],[36],[16],[15], [38]). Also
graph products of sofic groups are sofic by [11]. Additionally, residually sofic groups and locally sofic groups
are sofic. Thus by Malcev’s Theorem we know all linear groups are sofic. Finally, if Λ is a subgroup of Γ,
and Λ is sofic and Γ y Γ/Λ is amenable (in the sense of having a Γ-invariant mean) then Γ is sofic (this
can be seen by a mild generalization of the argument in Theorem 1 of [17] using e.g. the observation after
Definition 12.2.12 of [9]).
We will need to extend a sofic approximation to certain algebras associated to Γ. Let C(Γ) be the ring of
finite formal linear combinations of elements of Γ with addition defined naturally and multiplication defined
by
!
!
X
X
X X
ag g
bh h =
ah bh−1 g g.
g∈Γ
h∈Γ
g∈Γ
h∈Γ
We will also define a conjugate-linear involution on C(Γ) by
∗
X
X
ag g =
ag−1 g.
g∈Γ
g∈Γ
P
Given a sofic approximation Σ = (σi : Γ → Sdi ) and α = g∈Γ αg g ∈ C(Γ) we define σi (α) ∈ Mdi (C) by
X
σi (α) =
αg σi (g).
g∈Γ
In order to talk about the asymptotic properties of this extended sofic approximation, we will need a more
analytic object associated to Γ.
Let λ : Γ → U(ℓ2 (Γ)) be the left regular representation defined by (λ(g)ξ)(h) = ξ(g −1 h). We will continue
to use λ for the linear extension to C(Γ) → B(ℓ2 (Γ)). The group von Neumann algebra of Γ is defined by
λ(C(Γ))
W OT
where WOT denotes the weak operator topology. We will use L(Γ) to denote the group von Neumann
algebra. Define τ : L(Γ) → C by
τ (x) = hxδe , δe i.
We leave it as an exercise to the reader to verify that τ has the following properties.
1: τ (1) = 1,
2: τ (x∗ x) ≥ 0, with equality if and only if x = 0,
3: τ (xy) = τ (yx), for all x, y ∈ M,
4: τ is weak operator topology continuous.
8
BEN HAYES
We call the third property the tracial property. We will typically view C(Γ) as a subset of L(Γ). In particular,
we will use τ as well for the functional on C(Γ) which is just the restriction of τ on L(Γ).
In order to state our extension of a sofic approximation properly, we shall give a general definition.
Recall that ∗-algebra is a complex algebra equipped with an involution ∗ which is conjugate linear and
antimultiplicative.
Definition 2.2. A tracial ∗-algebra is a pair (A, τ ) where A is a ∗-algebra equipped with a linear functional
τ : A → C so that
1: τ (1) = 1,
2: τ (x∗ x) ≥ 0, with equality if and only if x = 0,
3: τ (xy) = τ (yx), for all x, y ∈ M,
4: For all a ∈ A, there is a Ca > 0 so that for all x ∈ A, |τ (x∗ a∗ ax)| ≤ Ca τ (x∗ x).
For a, b ∈ A we let ha, bi = τ (b∗ a) and we let kak2 = τ (a∗ a)1/2 . We let L2 (A, τ ) be the Hilbert space
completion of A in this inner product. By condition 4 of the definition, we have a representation λ : A →
B(L2 (A, τ )) defined densely by λ(a)x = ax for x ∈ A. We let kak∞ = kλ(a)k.
We make Mn (C) into a tracial ∗-algebra using tr = n1 Tr where Tr is the usual trace. In particular, we
use kAk2 = tr(A∗ A)1/2 and kAk∞ will denote the operator norm.
We let C[X1 , . . . , Xn ] be the free ∗-algebra on n-generators X1 , . . . , Xn . We will call elements of C[X1 , . . . , Xn ]
∗-polynomials in n indeterminates. For a ∗-algebra A, for elements a1 , . . . , an ∈ A, and P ∈ C[X1 , . . . , Xn ]
we use P (a1 , . . . , an ) for the image of P under the unique ∗-homomorphism C[X1 , . . . , Xn ] → A sending Xj
to aj .
Definition 2.3. Let (A, τ ) be a tracial ∗-algebra. An embedding sequence is a sequence Σ = (σi : A →
Mdi (C)) such that
sup kσi (a)k∞ < ∞, for all a ∈ A
i
kP (σi (a1 ), . . . , σi (an )) − σi (P (a1 , . . . , an ))k2 → 0, for all a1 , . . . , an ∈ A and all P ∈ C[X1 , . . . , Xn ]
tr(σi (a)) → τ (a) for all a ∈ A.
We will frequently use the following fact: if x1 , . . . , xn ∈ A and P ∈ C[X1 , . . . , Xn ] then
(3)
kP (σi (x1 ), . . . , σi (xn ))k2 → kP (x1 , . . . , xn )k2 .
To prove this, first note that as
kP (σi (x1 ), . . . , σi (xn )) − σi (P (x1 , . . . , xn ))k2 → 0,
it suffices to handle the case n = 1 and P (X) = X. In this case,
kσi (x)k22 = tr(σi (x)∗ σi (x))
and since kσi (x)∗ σi (x) − σi (x∗ x)k2 → 0 we have
Since
| tr(σi (x)∗ σi (x)) − tr(σi (x∗ x))| → 0.
tr(σi (x∗ x)) → kxk22 ,
we have proved (3).
The proof of the next two propositions will be left to the reader.
Proposition 2.4. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ).
Extend Σ to maps σi : C(Γ) → Mdi (C) linearly. Then Σ is an embedding sequence of (C(Γ), τ ).
Proposition 2.5. Let (A, τ ) be a tracial ∗-algebra and Σ = (σi : A → Mdi (C)) be an embedding sequence.
If Σ′ = (σi′ : A → Mdi (C)) is another sequence of functions so that
sup kσi′ (a)k∞ < ∞, for all a ∈ A,
i
then Σ′ is an embedding sequence.
kσi (a) − σi′ (a)k2 → 0, for all a ∈ A,
POLISH MODELS AND SOFIC ENTROPY
9
We will in fact need to extend our sofic approximation to the group von Neumann algebra. For this, we
use the following.
Lemma 2.6 (Lemma 5.5 in [20]). Let Γ be a countable discrete group. Then any embedding sequence for
C(Γ) extends to one for L(Γ).
We will use the preceding lemma when Γ is sofic, in combination with Proposition 2.4. We will often need
the following volume-packing estimate.
Lemma 2.7. Let n ∈ N, and p ∈ B(ℓ2 (n, un )) an orthogonal projection. For any M, ε > 0 we have
2 tr(p)n
3M + ε
.
Sε (M p Ball(ℓ2 (n, un ))) ≤
3
Proof. Let A ⊆ M p Ball(ℓ2 (n, un )) be a maximal ε-separated subset of M p Ball(ℓ2 (n, un )). Thus A is ε-dense.
Further,
[
(M + ε/3)p Ball(ℓ2 (n, un )) ⊇
x + (ε/3)p Ball(ℓ2 (n, un )),
x∈A
and the right-hand side is a disjoint union. By linear algebra, the real dimension of the image of p is 2 tr(p)n.
Thus computing volumes:
(M + ε/3)2 tr(p)n vol(p Ball(ℓ2 (n, un )) ≥ |A|(ε/3)2 tr(p)n vol(p Ball(ℓ2 (n, un ))).
Thus
|A| ≤
3M + ε
ε
2 tr(p)n
.
3. Definition of Entropy in the Presence of a Polish Model
Our definition will follow the ideas in [23] and will use dynamically generating pseudometrics. We need
to state the definition so that it works for actions on Polish spaces. We will need to assume that the
pseudometrics are bounded as this is no longer automatic in the noncompact case.
Definition 3.1. Let Γ be a countable discrete group and X a Polish space with Γ y X by homeomorphisms.
A bounded continuous pseudometric ∆ on X is said to be dynamically generating if for any x ∈ X and any
open neighborhood U of x in X, there is a δ > 0 and a F ⊆ Γ finite so that if x ∈ X and maxg∈F d(gx, gy) < δ,
then y ∈ U.
Recall that if X is compact, and Γ is a countable discrete group acting on X by homeomorphisms and ∆
is a continuous pseudometric on X, then ∆ is said to be dynamically generating if
sup ∆(gx, gy) > 0 whenever x 6= y in X
g∈Γ
(see e.g. [29] Section 4). The fact that this is equivalent to our definition is an easy exercise using compactness
of X.
When X is Polish, as we shall see in the proof of Lemma 3.9 it really is necessary to require the existence
of U, δ, F as in the preceding definition instead of just
sup ∆(gx, gy) > 0 whenever x 6= y in X.
g∈Γ
One way to realize why this is the correct definition is as follows: let (X, ∆, Γ) be as in the preceding
definition and let Y be X modded out by the equivalence relation of a ∼ b if ∆(a, b) = 0. Give Y the metric
∆([a], [b]) = ∆(a, b).
Consider the continuous map
given by
Φ: X → Y Γ
Φ(x)(g) = [gx],
10
BEN HAYES
where we use [a] for the equivalence class of a ∈ X. The existence of U, δ as in the definition is the precise
requirement one needs to guarantee that Φ is a homeomorphism onto its image. This will be explicitly
proven in Lemma 3.9.
If (X, ∆) is a pseudometric space, we let ∆2 be the pseudometric on X n defined by
n
1X
∆(x(j), y(j))2 .
∆2 (x, y) =
n j=1
2
Definition 3.2. Let Γ be a countable discrete group and X a Polish space with Γ y X by homeomorphisms.
Let ∆ be a bounded pseudometric on X. For a function σ : Γ → Sd , for some d ∈ N, a finite F ⊆ Γ, and a
δ > 0 we let Map(∆, F, δ, σ) be all functions φ : {1, . . . , d} → X so that
max ∆2 (φ ◦ σ(g), gφ) < δ,
g∈F
We caution the reader that though we shall typically require X to be Polish, we will not require our
pseudometrics to be complete. We will typically only need to care about the topological consequences of
being Polish and not any metric properties. Note that Map(∆, F, δ, σ) does not account for the measuretheoretic structure of X. Given a Polish space X, a finite L ⊆ Cb (X), a δ > 0, and µ ∈ Prob(X) let
Z
Z
\
UL,δ (µ) =
ν ∈ Prob(X) :
f dν − f dµ < δ .
f ∈L
Then UL,δ (µ) form a basis of neighborhoods of µ for the weak topology. Here Cb (X) is the space of bounded
continuous functions on X. Recall that ud denotes the uniform probability measure on {1, . . . , d}.
Definition 3.3. Suppose that µ is a Γ-invariant Borel probability measure on X. For F ⊆ Γ finite, δ > 0
and L ⊆ Cb (X) finite, and σ : Γ → Sd for some d ∈ N we let Mapµ (∆, F, δ, L, σ) be the set of all φ ∈
Map(∆, F, δ, σ) so that
φ∗ (ud ) ∈ UL,δ (µ).
for all f ∈ L.
Definition 3.4. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let
X be a Polish space with Γ y X by homeomorphisms and µ a Γ-invariant, Borel, probability measure on
X. We define the entropy of Γ y (X, µ) by
hΣ,µ (∆, ε, F, δ, L) = lim sup
i→∞
hΣ,µ (∆, ε) =
1
log Nε (Mapµ (∆, F, δ, L, σi ), ∆2 )
di
inf
F ⊆Γfinite,
δ>0,
L⊆Cb (X)finite
hΣ,µ (∆, F, δ, L, ε)
hΣ,µ (∆) = sup hΣ,µ (∆, ε).
ε>0
By (2) we know that hΣ,µ (∆) is unchanged if we replace Nε with Sε . Because we use Nε instead of Sε ,
we have hΣ,µ (∆, ε, F, δ, L) ≤ hΣ,µ (∆, ε, F ′ , δ ′ , L′ ) if F ⊇ F ′ , δ ≤ δ ′ , L ⊇ L′ .
The reader may be concerned about finiteness of the expression Nε (Mapµ (∆, F, δ, L, σi ), ∆2 ), since X is
not compact. We note that if ε > 0 is given, there is a finite L ⊆ Cb (X), and δ > 0 so that for any finite
F ⊆ Γ we have
Nε (Mapµ (∆, F, δ, L, σi ), ∆2 ) < ∞.
To see this, note that by Prokhorov’s Theorem we may choose a K ⊆ X compact so that
µ(K) ≥ 1 − ε.
It is not hard to see that if L ⊆ Cb (X) is sufficiently large and δ > 0 is sufficiently small, then for any
φ ∈ Mapµ (∆, F, δ, L, σi ),
1
|{j : φ(j) ∈
/ K}| = φ∗ (ud )(K c ) ≤ 2ε.
d
POLISH MODELS AND SOFIC ENTROPY
11
Suppose that S is a finite ε-dense subset of K. Let M be the diameter of (X, ∆) and fix x0 ∈ X. If we
set A = φ−1 (K), then we find that udi (A) ≥ 1 − ε. We can find a ψ ∈ S A so that if ψe : {1, . . . , di } → X is
defined by
(
e = ψ(j), if j ∈ A,
ψ(j)
x0 ,
if j ∈ {1, . . . , di } \ A.
Then
e 2 ≤ 2M 2 ε + ε2 .
∆2 (φ, ψ)
Thus
N2(2M 2 ε+ε2 )1/2 (Mapµ (∆, F, δ, L, σi ), ∆2 ) ≤ S(2M 2 ε+ε2 )1/2 (Mapµ (∆, F, δ, L, σi ), ∆2 ) ≤ Sε (K) < ∞.
The main goal of this section is to show that hΣ,µ (∆) is the same as the measure entropy of Γ y (X, µ)
as defined by Bowen and extended by Kerr-Li. Throughout we use hΣ,µ (X, Γ) for sofic measure entropy as
defined by [5],[25]. We will use the formulation of sofic entropy in terms of partitions due to Kerr in [22].
However, we will use the terminology of observables as Bowen did in [6].
Definition 3.5. Let (X, M, µ) be a standard probability space. Let S be a subalgebra of M (here S is
not necessarily a σ-algebra). A finite S-measurable observable is a measurable map α : X → A where A
is a finite set and α−1 ({a}) ∈ S for all a ∈ A. If S = M we simply call α a finite observable. Another
finite S-measurable observable β : X → B is said to refine α, written α ≤ β, if there is a π : B → A so that
π(β(x)) = α(x) for almost every x ∈ X. If Γ is a countable discrete group and Γ y (X, M, µ) by measurepreserving transformations we say that S is generating if M is the σ-algebra generated by {gA : A ∈ S} (up
to sets of measure zero).
For the next definition we need to set up some notation. Given a standard probability space (X, M, µ),
a countable discrete group with Γ y (X, M, µ) by measure-preserving transformations, a finite observable
α : X → A and F ⊆ Γ finite, we let
α
eF : X → AF
be defined by
α
eF (x)(g) = α(g −1 x).
Definition 3.6. Let Γ be a countable discrete group and σ ∈ SdΓ for some d ∈ N. Let (X, M, µ) be a standard
probability space and let S ⊆ M be a subalgebra. Let α : X → A be a finite S measurable-observable. Given
F ⊆ Γ finite, and δ > 0, we let AP(α, F, δ, σ) be the set of all φ : {1, . . . , d} → AF so that
X
ud ((φ−1 ({a})) − µ((e
αF )−1 ({a})) < δ.
a∈AF
udi ({j : φ(j)(g) = φ(σi (g)−1 (j))(e)}) < δ, for all g ∈ F.
We now give Kerr’s definition of sofic measure entropy in [22].
Definition 3.7. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let
(X, M, µ) be a standard probability space and Γ y (X, M, µ) by measure-preserving transformations. Let
S be a subalgebra of M. Let α : X → A be a finite S-measurable observable, and let β : X → B refine α,
and π : B → A as in the definition of α ≤ β. For any F ⊆ Γ finite, we use
π∗ : B F → A
for
π ∗ (b) = π(b(e)).
We set
1
log (π ∗ )di (AP(β, F, δ, σi ))
di
hΣ,µ (α; β, F, δ).
inf
hΣ,µ (α; β, F, δ) = lim sup
i→∞
hΣ,µ (α; β) =
F ⊆Γ finite,
δ>0
We then set
hΣ,µ (α; S) = inf hΣ,µ (α; β)
α≤β
12
BEN HAYES
hΣ,µ (S) = sup hΣ,µ (α; β)
α
where the last infimum and supremum are over all S-measurable observables.
We need the following result of Kerr.
Theorem 3.8. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let (X, M, µ) be a
standard probability space with Γ y (X, M, µ) by measure-preserving transformations. Let S ⊆ M be a
generating subalgebra. Then
hΣ,µ (S) = hΣ,µ (X, Γ).
Additionally, one can show that
hΣ,µ (α; S)
is independent of S if S generates M (up to sets of measure zero). In this case, we set
hΣ,µ (α; S) = hΣ,µ (α; X, Γ).
We now proceed to prove that our definition of sofic entropy with respect to a Polish model recovers
measure-theoretic entropy with respect to a sofic approximation. Let us briefly outline the proof. First we
show that for any dynamically generating pseudometric ∆ on X, there is a compatible metric ∆′ so that
hΣ,µ (∆) = hΣ,µ (∆′ )
(see Lemma 3.9). Thus we may assume that our metrics our compatible. We then use Kerr’s version of
measure entropy, using the subalgebra of sets which from the point of view of the measure “appear” to be
open and closed (in a sense to be made precise later). We then show that both the topological version and
the observable version of microstates produce roughly the same space (see Lemma 3.11). The essential fact
for proving the last step will be tightness of a single probability measure on a Polish space. The theorem
will follow without too much difficulty from these preliminary lemmas. The following proof is a minor
modification of the argument in Lemma 4.4 of [30], as well as Lemma 6.12 of [7]. We have decided to include
the proof to alleviate any concerns that may arise from working in the noncompact case, as well as address
the necessary modifications that occur in the definition of a dynamically generating pseudometric in the
Polish case.
Lemma 3.9. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let
X be a Polish space with Γ y X by homeomorphisms and µ a Γ-invariant probability measure. Given a
dynamically generating pseudometric ∆ on X, there is a bounded compatible metric ∆′ on X so that
hΣ,µ (∆) = hΣ,µ (∆′ ).
Proof. Let M be the diameter of (X, ∆). Since Γ is countable, we may find positive real numbers {αg : g ∈ Γ}
with αe ≥ 1/2 and
X
αg = 1.
g
Set
∆′ (x, y) =
X
αg ∆(gx, gy).
g∈Γ
We prove the lemma with this ∆′ . The lemma is proved in several steps.
Step 1: We show that ∆′ is a compatible metric. For this let Y be X modded out by the equivalence
relation a ∼ b if ∆(a, b) = 0. For a ∈ X, let [a] be the equivalence class of a. Make Y a metric space with
metric ∆ given by
∆([a], [b]) = ∆(a, b).
Observe that ∆ is well-defined because ∆ satisfies the triangle inequality. Then Y Γ has a compatible metric
given by
X
∆Γ (a, b) =
αg ∆(a(g), b(g)).
g∈Γ
Moreover we have an injective map Φ : X → Y Γ by
Φ(x)(g) = [gx],
POLISH MODELS AND SOFIC ENTROPY
13
as
∆Γ (Φ(x), Φ(y)) = ∆′ (x, y),
it is enough to show that Φ is a homeomorphism onto its image. It is clear that Φ is continuous. Suppose
that x ∈ X and U is a neighborhood of x ∈ X. By the definition of dynamically generating, we may choose
a finite F ⊆ Γ and a δ > 0 so that if y ∈ X and
max ∆(gx, gy) < δ,
g∈F
then y ∈ U. If we let
V = {y ∈ Y Γ : ∆([gx], y(g)) < δ for all g ∈ F },
then V is a neighborhood of Φ(x) in Y Γ and if y ∈ X has Φ(y) ∈ V, then y ∈ U. Thus Φ is a homeomorphism
onto its image.
Step 2: We show that hΣ,µ (∆′ ) ≤ hΣ,µ (∆). Let ε > 0 and choose a finite E ⊆ Γ sufficiently large so that
X
αg < ε.
g∈Γ\E
Let F ⊆ Γ finite, δ > 0 and L ⊆ Cb (X) finite be given. We will assume that F ⊇ E. We will choose δ
sufficiently small in a manner depending upon ε to be determined later. Since αe ≥ 1/2 we have
∆(x, y) ≤ 2∆′ (x, y).
So by Minkowski’s inequality
Mapµ (∆′ , F, δ, L, σi ) ⊆ Mapµ (∆, F, 4δ, L, σi ).
For φ, ψ ∈ Mapµ (∆, F, 2δ, L, σi ) we have by Minkowski’s inequality,
X
∆′2 (φ, ψ) ≤
αg ∆2 (gφ, gψ)
g∈Γ
≤ εM +
X
αg ∆2 (gφ, gψ)
g∈E
≤ εM + 2δ +
X
g∈E
αg ∆2 (φ ◦ σi (g), ∆ ◦ σi (g))
≤ εM + 2δ + ∆2 (φ, ψ)
where in the last line we use that ∆2 (φ ◦ σi (g), ψ ◦ σi (g)) = ∆2 (φ, ψ). Thus for all sufficiently small δ
S3ε(M+2) (Mapµ (∆′ , F, δ, L, σi ), ∆′2 ) ≤ Sε (Mapµ (∆, F, 4δ, L, σi )).
Thus by (2) we have
hΣ,µ (∆′ , 6ε(M + 2)) ≤ hΣ,µ (∆, ε, F, 4δ, L).
As hΣ,µ (∆, ε, F, 4δ, L) is monotone in (F, 4δ, L) we let F, L increase to Γ, Cb (X) and take δ → 0 to find that
hΣ,µ (∆′ , 6ε(M + 2)) ≤ hΣ,µ (∆′ , ε) ≤ hΣ,µ (∆).
Now letting ε → 0 completes the proof of Step 2.
Step 3: We show that hΣ,µ (∆) ≤ hΣ,µ (∆′ ). Let ε > 0. Suppose we are given finite F ′ ⊆ Γ, L′ ⊆ Cb (X)
and δ ′ > 0. Set L = L′ and let F ⊆ Γ be a sufficiently large finite set depending upon F ′ , δ ′ in a manner to
determined later and set δ = δ ′ . Choose a finite E ⊆ Γ so that
X
αg < δ ′ .
g∈Γ\E
14
BEN HAYES
If φ ∈ Mapµ (Γ, F, δ, L, σi ) and h ∈ F ′ , we have by Minkowski’s inequality
X
∆′2 (hφ, φ ◦ σi (h)) ≤
αg ∆2 (ghφ, gφ ◦ σi (h))
g∈Γ
≤ δ′ +
+
X
g∈F
X
g∈F
g∈F
αg ∆2 (ghφ, φ ◦ σi (gh)) +
X
g∈F
αg ∆2 (φ ◦ σi (gh), φ ◦ σi (g)σi (h))
αg ∆2 (φ ◦ σi (g)σi (h), gφ ◦ σi (h))
= δ′ +
+
X
X
g∈F
αg ∆2 (ghφ, φ ◦ σi (gh)) +
X
g∈F
αg ∆2 (φ ◦ σi (gh), φ ◦ σi (g)σi (h))
αg ∆2 (φ ◦ σi (g), gφ)
< δ ′ + 2δ +
X
g∈E
αg ∆2 (φ ◦ σi (g)σi (h), φ ◦ σi (gh)),
if we force F ⊇ EF ′ ∪ E. Set δ = δ ′ , as
∆2 (φ ◦ σi (g)σi (h), φ ◦ σi (gh))2 ≤ M udi ({1 ≤ j ≤ di : σi (g)σi (h)(j) 6= σi (gh)(j)}) → 0
for g, h ∈ Γ we find that
for all large i. As
we find that
Mapµ (∆, F, δ, L, σi ) ⊆ Mapµ (∆′ , F ′ , 4δ ′ , σi ),
∆2 (φ, ψ) ≤ 2∆′2 (φ, ψ),
hΣ,µ (∆, 6ε) ≤ hΣ,µ (∆, 6ε, F, δ, L) ≤ hΣ,µ (∆′ , ε, F ′ , δ ′ , L′ ).
Taking the infimum over F ′ , δ ′ , L′ we find that
hΣ,µ (∆, 6ε) ≤ hΣ,µ (∆′ , ε) ≤ hΣ,µ (∆′ ).
Letting ε → 0 proves Step 3.
To prove the theorem, we need to single out a nice subalgebra of measurable sets. Let X be a Polish
space and µ a Borel probability measure on X. We let COµ be the set of Borel sets E so that
µ(int E) = µ(E).
Note that COµ is an algebra of sets. These sets are often called continuity sets in the literature.
For the next lemma we need some notation. Given a metric space (X, ∆), E ⊆ X and ε > 0 we let
Eε = {x ∈ E : ∆(x, E c ) ≥ ε}
Oε (E) = {x ∈ X : ∆(x, E) < ε}.
Lemma 3.10. Let X be a Polish space and µ a Borel probability measure on X. Let ∆ be a compatible
metric on X. Given E ∈ COµ , and η > 0 there is a neighborhood U of µ in the weak topology and a κ > 0
so that if ν ∈ U ∩ Prob(X) then
|ν(E) − µ(E)| < η
ν(Oκ (E) \ Eκ ) < η.
Proof. It is a consequence of the Portmanteau Theorem that a sequence of probability measure µn on X
converges weakly to µ if and only if µn (A) → µ(A) for every continuity set A of µ. Thus we can choose a
neighborhood U1 of µ so that
|ν(E) − µ(E)| < η
for all ν ∈ U1 . To obtain the second estimate, again by the Portmanteau Theorem we can choose a neighborhood U2 of µ so that
η
ν(O2κ (E) \ E2κ ) ≤ µ(O2κ (E) \ E2κ ) +
2
POLISH MODELS AND SOFIC ENTROPY
15
for all ν ∈ U2 . We have Oκ (E) \ Eκ ⊆ O2κ (E) \ E2κ , so
η
ν(Oκ (E) \ Eκ ) ≤ µ(O2κ (E) \ E2κ ) + .
2
Since E is a continuity set of µ we can choose κ small enough so that
η
µ(O2κ (E) \ E2κ ) < .
2
The lemma is now completed by setting U = U1 ∩ U2 .
d
F
F
Given σ : Γ → Sd for some d ∈ N, and φ ∈ A we shall define φ : {1, . . . , d} → A
φF
σ (j)(g)
−1
= φ(σ(g)
by’
(j)), for g ∈ F
Lemma 3.11. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let X
be a Polish space with Γ y X by homeomorphisms and µ a Γ-invariant Borel probability measure on X. Let
∆ be a bounded compatible metric on X.
(i): Let β : X → B a finite COµ -measurable observable. Given F ⊆ Γ finite, δ > 0 there are finite
F ′ ⊆ Γ, L′ ⊆ Cb (X), and δ ′ > 0 so that if φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ) then (β ◦ φ)F
σi ∈ AP(β, F, δ, σi ).
(ii): Given finite F ′ ⊆ Γ, L′ ⊆ Cb (X) and δ ′ > 0, there are finite F ⊆ Γ and a δ > 0 and a finite
COµ -measurable observable β : X → B so that if βeF ◦ φ ∈ AP(β, F, δ, σi ) then φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ).
Proof. (i): Let η > 0 be sufficiently small depending upon δ in a manner to be determined later. By the
preceding lemma, we may find a L′ ⊆ Cb (X) finite, and a δ ′ > 0 so that if ν ∈ Prob(X) and
Z
Z
f dν < δ ′ for all f ∈ L′
f dµ −
X
X
then for all E ⊆ F and for all b ∈ B E ,
|µ((βeE )−1 ({b})) − ν((βeE )−1 ({b}))| < η,
ν(O√δ′ (βeE )−1 ({b})) \ (βeE )−1 ({b})√δ′ ) ≤ η.
Set F = F. Suppose that φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ). Note that if we choose η sufficiently small, then we
have forced
X
−1
({b}))| < δ
|µ((βeF )−1 ({b})) − udi (((β ◦ φ)F
σi )
′
b∈B F
, L′ , σi ).
F
−1
(j))(e), this means (β ◦ φ)F
for all φ ∈ Map(∆, F , δ
Since (β ◦ φ)F
σi ∈
σi (j)(g) = (β ◦ φ) (σi (g)
AP(β, F, δ, σi ).
(ii): Let M > 0 be the diameter of (X, ∆). Let κ > 0 be sufficiently small depending upon F ′ , δ ′ in a
manner to be determined later. Since X is Polish, Prokhorov’s Theorem applied to {µ} implies that we can
find a compact set K ⊆ X so that
µ(X \ K) ≤ κ.
Since K is compact, we can find points x1 , . . . , xn ∈ K, and numbers δj ∈ (0, κ), j = 1, . . . , n so that if
B(x, α) is the ball in X of radius α with respect to ∆ then
′
′
sup
µ(B(xj , δj ) \ B(xj , δj )) = 0,
x,y∈B(xj ,δj )
|f (x) − f (y)| < δ ′ for all f ∈ L′ ,
K⊆
Let
n
[
E=X\
and define β : X → {0, 1}n+1 by
β(x)(k) =
(
B(xj , δj ).
j=1
n
[
B(xj , δj )
j=1
χB(xk ,δk ) (x),
χE (x),
if 1 ≤ k ≤ n
if k = n + 1.
16
BEN HAYES
Note that β is COµ - measurable. Set F = F ′ ∪ {e} ∪ (F ′ )−1 , and let δ > 0 be sufficiently small in a manner
to depend upon L′ , δ ′ , F ′ to be determined later. We will assume that
δ < min{δj : j = 1, . . . , n}.
Suppose that βeF ◦ φ ∈ AP(β, F, δ, σi ). Given g ∈ F, let
Cg = {j : β(gφ(j)) = β(φ(σi (g)(j)))}.
Since
β(gφ(j)) = (βeF ◦ φ)(j)(g −1 ),
β(φ(σi (g)(j))) = (βeF ◦ φ)(σi (g))(j)(e),
and
udi ({j : σi (g −1 )−1 (j) = σi (g)(j)}) →i→∞ 1
we see that for all large i,
udi (Cg ) ≥ 1 − 2δ.
For j ∈ Cg , we necessarily have
∆(φ(σi (g)(j)), gφ(j)) < κ.
Thus
∆2 (φ ◦ σi (g), gφ)2 ≤ κ2 + M 2 δ.
So if we choose κ < δ, and then δ sufficiently small, we have forced
∆2 (φ ◦ σi (g), gφ) < 2δ ′ .
We will want to force κ, δ ′ to be even smaller later. Using kf k for the uniform norm of f ∈ Cb (X), we have
for all f ∈ L′ ,
Z
Z
Z
n Z
X
f dφ∗ (udi )
f dµ −
f dφ∗ (udi ) ≤ κkf k + udi (φ−1 (E))kf k +
f dµ −
X
X
j=1
≤ κkf k + 2κkf k + δ ′ +
n
X
+
n
X
j=1
f (xj )µ(B(xj , δj )) −
j=1
≤ κkf k + 2κkf k + δ ′ + δ ′
n
X
j=1
f (xj )µ(B(xj , δj )) −
≤ κkf k + 2κkf k + 2δ ′ +
B(xj ,δj )
B(xj ,δj )
1
di
n
X
j=1
1
di
X
f (φ(k))
k:φ(k)∈B(xj ,δj )
φ∗ (udi )(B(xj , δj ))
X
f (xj )
k:φ(k)∈B(xj ,δj )
|f (xj )||µ(B(xj , δj ) − φ∗ (udi )(B(xj , δj ))|
≤ κkf k + 2κkf k + 2δ ′ + nkf kδ.
′
We may choose κ < kfδ k . This forces n on us, but we may then choose δ sufficiently small so that φ ∈
Mapµ (∆, F ′ , 6δ ′ , L′ , σi ). As δ ′ is arbitrary this completes the proof.
We are now ready to show that our definition of entropy in the case of a Polish model agrees with the
usual measure entropy.
Theorem 3.12. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let X be a Polish space
with Γ y X by homeomorphisms and µ a Γ-invariant, Borel, probability measure on X. For any dynamically
generating pseudometric ∆ on X we have
hΣ,µ (∆) = hΣ,µ (X, Γ).
POLISH MODELS AND SOFIC ENTROPY
17
Proof. Let Σ = (σi : Γ → Sdi ). By Lemma 3.9 we may assume that ∆ is a bounded compatible metric on X.
Let M be the diameter of (X, ρ). We will apply Theorem 3.8 with S = COµ . We leave it as an exercise to
show that for all x ∈ X we have
µ(B(x, r)) = µ(B(x, r))
for all but countably many r ∈ (0, ∞). Thus COµ generates the σ-algebra of Borel subsets of X. We first
show that
hΣ,µ (∆) ≤ hΣ,µ (COµ ).
Let ε > 0. Since X is Polish, we may apply Prokhorov’s Theorem to find a compact K ⊆ X so that
µ(X \ K) < ε.
By compactness of K, we find x1 , . . . , xn ∈ K, and ε > δ1 , . . . , δn > 0 so that
n
[
B(xj , δj ),
K⊆
j=1
B(xj , δj ) ∈ COµ .
Set
E=X\
Define
n
[
B(xj , δj ).
j=1
α : X → {0, 1}n+1
by
α(x)(k) =
(
χB(xk ,δk ) (x),
χE (x),
if 1 ≤ k ≤ n
if k = n + 1.
Let β : X → B be any finite COµ -measurable observable refining α and let π : B → A be as in the definition
of β ≥ α. Suppose we are given a finite F ⊆ Γ and a δ > 0. By the preceding Lemma, we may find a finite
F ′ ⊆ Γ, L′ ⊆ Cb (X), and a δ ′ > 0 so that
βσFi (Mapµ (∆, F ′ , δ ′ , L′ , σi )) ⊆ AP(β, F, δ, σi ).
By Lemma 3.10, we may assume that L′ is sufficiently large so that
φ∗ (udi )(X \ E) ≤ 2ε
for all φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ). Choose elements {φs }s∈S with φs ∈ Mapµ (∆′ , F ′ , L′ , δ ′ , σi ) where S is
some index set so that
{α ◦ φs : s ∈ S} = αdi (Mapµ (∆, F ′ , δ ′ , L′ , σi ))
and so that
α ◦ φs 6= α ◦ φs′ for s 6= s′ in S.
Then
|S| ≤ |(π ∗ )di (AP(β, F, δ, σi ))|.
Let φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ) and let s ∈ S be such that α ◦ φ = α ◦ φs . Then
X
1
∆2 (φ, φs )2 ≤ 4M 2 ε +
∆(φ(j), φs (j))2 .
di
j:φ(j),φs (j)∈E
/
If φ(j) and φs (j) are not in E, then the fact that α(φ(j)) = α(φs (j)) implies ∆(φ(j), φs (j)) < ε, so
∆2 (φ, φs )2 < 4M 2 ε + ε2 .
Thus by (2),
hΣ,µ (∆, 2(4M ε + ε2 )1/2 ) ≤ hΣ,µ (α; β, F, δ).
Taking the infimum over all β, F, δ we find that
hΣ,µ (∆, 2(4M ε + ε2 )1/2 ) ≤ hΣ,µ (α) ≤ hΣ,µ (X, Γ),
and letting ε → 0 implies that
hΣ,µ (∆) ≤ hΣ,µ (X, Γ).
18
BEN HAYES
For the reverse inequality, let α : X → A be a COµ -measurable finite observable. Fix κ > 0, and let κ′ > 0
depend upon κ in a manner to be determined later. By Lemma 3.10 we may choose η > 0 and L0 ⊆ Cb (X)
finite so that if ν ∈ Prob(X) and
Z
Z
f dν < η,
f dµ −
X
for all f ∈ L0 then
X
ν(α−1 ({a})) − µ(α−1 ({a})) < κ′ ,
ν(Oη (α−1 ({a}) \ α−1 ({a})η ) < κ′ .
Let F ′ ⊆ Γ, L′ ⊆ Cb (X) be given finite sets and δ ′ > 0 be given. We may assume that L′ ⊇ L0 . By the
preceding Lemma, we may choose a refinement β : X → B of α, a finite F ⊆ Γ, and a δ > 0 so that if
φ ∈ X di and βeF ◦ φ ∈ AP(β, F, δ, σi ), then φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ). Choose π : B → A so that α = π ◦ β
and choose a map s : B F → X so that Id = βeF ◦ s. By construction if φ ∈ AP(β, F ′ , δ ′ , σi ) we have
s ◦ φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ).
Let ε > 0 be sufficiently small depending upon η to be determined later. Let {φt : t ∈ T } be such that
φt ∈ AP(β, F, δ, σi ) for all t ∈ T ,
{s ◦ φt : t ∈ T } is ε-dense in{s ◦ φ : φ ∈ AP(β, F, δ, σi )},
We may (and will) choose T so that
Note that
φt 6= φt′ for t 6= t′ .
|T | ≤ Sε/2 (Mapµ (∆, F ′ , δ ′ , L′ , σi )) ≤ Nε/2 (Mapµ (∆, F ′ , δ ′ , L′ , σi )).
(π ∗ )di (AP(β, F, δ, σi )) ⊆
[
t∈T
αdi (B∆2 (s ◦ φt , ε) ∩ Mapµ (∆, F ′ , δ ′ , L′ , σi )).
We thus have to bound |α (B∆2 (φt , ε) ∩ Mapµ (∆, F ′ , δ ′ , L′ , σi ))| from above. Fix t ∈ T, Suppose that
φ ∈ Mapµ (∆, F ′ , δ ′ , L′ , σi ) and that
∆2 (φ, s ◦ φt ) < ε.
Let
[
C=
{1 ≤ j ≤ di : φ(j) ∈ Oη (α−1 ({a}))\α−1 ({a})η }∪{1 ≤ j ≤ di : s◦φt (j) ∈ Oη (α−1 ({a}))\α−1 ({a})η }.
di
a∈A
By choosing κ′ sufficiently small, we may assume that
udi (C) ≤ κ.
Let
D = {1 ≤ j ≤ di : ∆(φ(j), s ◦ φt (j)) ≥
Thus
√
ε}.
√
udi (D) ≤ ε.
For j ∈ {1, . . . , di } \ (C ∪ D), and a = α(φ(j)), we have that s ◦ φt (j) ∈ O√ε (α−1 ({a})). If we choose
√
ε < η, then√α(s ◦√φt (j)) = a for all j ∈ {1, . . . , di } \ (C ∪ D). Thus we can find V ⊆ {1, . . . , di } with
udi (V ) ≥ 1 − κ − ε and α(φ(j)) = α(s ◦ φt (j)) for j ∈ V. Thus
X
|αdi (B∆2 (s ◦ φt , ε) ∩ Mapµ (∆, F ′ , δ ′ , L′ , σi ))| ≤
|A||V|
≤
If κ +
√
ε < 1/2 then for all large i we have
di
di
p
≤
.
⌊κ + ε⌋di
l
V⊆{1,...,d
√ i },
|V|≤(κ+ ε)di
√
⌊(κ+ ε)di ⌋
X
l=1
di
|A|l .
l
POLISH MODELS AND SOFIC ENTROPY
19
So by Stirling’s Formula the above sum is at most
√
√
R(κ + ε)di exp(di H(κ + ε))|A|κdi
for some constant R > 0, where
H(t) = −t log t − (1 − t) log(1 − t) for 0 ≤ t ≤ 1.
Thus
hΣ,µ (α) ≤ hΣ,µ (α; β, F, δ, σi ) ≤ H(κ +
√
ε) + κ log |A| + hΣ,µ (∆, ε/2, F ′ , δ ′ , L′ ).
Taking the infimum over all F ′ , δ ′ , L′ and let κ → 0 we have
√
hΣ,µ (α) ≤ hΣ,µ (∆, ε/2) + H( ε).
Letting ε → 0 and then taking the supremum over all α we have
hΣ,µ (X, Γ) ≤ hΣ,µ (∆).
4. Spectral Consequences of Positive Entropy
Let Γ y (X, µ) be a probability measure-preserving action of a countable discrete group. Associated to
this action we have a natural representation
ρΓy(X,µ) : Γ → U (L2 (X, µ))
by
(ρΓy(X,µ) (g)f )(x) = f (g −1 x).
The space C1 inside L2 (X, µ) is clearly Γ-invariant, so we can consider the representation ρ0Γy(X,µ) obtained
by restricting ρΓy(X,µ) to L2 (X, µ) ⊖ C1. The representation ρ0Γy(X,µ) is called the Koopman representation.
Properties of a probability measure-preserving action are called spectral when they only depend upon the
Koopman representation. In this section, we deduce spectral properties of an action from assumptions of
positive entropy.
4.1. Representation Theoretic Preliminaries. We will need to apply the theory of representations of
∗-algebras. For this paper, we will only need unitary representations of groups, but later work will need this
generality. For notation, if A is a ∗-algebra and ρ : A → B(H) is a ∗-representation and E is a set we use
ρ⊕E for the ∗-representation of A on ℓ2 (E, H) given by
(ρ⊕E (a)ξ)(x) = ρ(a)ξ(x).
Let us mention how the theory for ∗-algebras generalizes
P that of groups. If Γ is a countable discrete group
and ρ : Γ → U(H) is a unitary representation, for f = g∈Γ fg g ∈ C(Γ) we define
X
ρ(f ) =
fg ρ(g).
g∈Γ
We use the conjugate linear, antimultiplicative map ∗ on C(Γ) given by
X
fg−1 g.
f∗ =
g∈Γ
∗
Under these operations C(Γ) is ∗-algebra and ρ(f ) = ρ(f )∗ for any unitary representation of Γ.
If A is a ∗-algebra and ρj : Γ → U(Hj ) are two ∗-representations we write
HomA (ρ1 , ρ2 )
for the space of bounded linear A-equivariant maps H1 → H2 .
Definition 4.1. Let A be a ∗-algebra and ρj : A → B(Hj ), j = 1, 2 be ∗-representations. We say that ρ1
and ρ2 are mutually singular, written ρ1 ⊥ ρ2 if for every pair of nonzero subrepresentations ρ′j of ρj we
have that ρ′1 is not isomorphic to ρ′2 . We say that ρ1 is absolutely continuous with respect to ρ2 , and write
ρ1 ≪ ρ2 if ρ1 is embeddable in ρ⊕E
for some set E.
2
20
BEN HAYES
The terminology is motivated by measure theory. For intuition, suppose that A = C(X) and that
ρj : A → B(Hj ), j = 1, 2 are ∗-homomorphisms. Then we can find spectral measures (in the sense of [13]
IX.1) Ej on X so that
Z
ρj (f ) =
f (x) dEj (x).
X
We leave it as an exercise to the reader to check that ρ1 ⊥ ρ2 if and only if E1 ⊥ E2 , and similarly that
ρ1 ≪ ρ2 if and only if E1 ≪ E2 (the definitions of absolute continuity and singularity of spectral measures
is the same as for usual measures).
We will need the following equivalent conditions on singularity of representations. The following must be
well known, but we include a proof for completeness. Throughout the proof, we shall use functional calculus.
See [13] Chapter VII, IX for background on functional calculus.
Proposition 4.2. Let A be a unital ∗-algebra and ρj : A → B(Hj ) be two unitary representations and
suppose that H1 , H2 are separable. The following are equivalent:
(i): ρ1 ⊥ ρ2
(ii): HomA (ρ1 , ρ2 ) = {0}
(iii): HomA (ρ2 , ρ1 ) = {0}
(iv): There is a sequence xn ∈ A so that max(kρ1 (xn )k, kρ2 (xn )k) ≤ 1 and
ρ1 (x∗n xn ) → IdH1 in the strong operator topology
ρ2 (x∗n xn ) → 0 in the strong operator topology.
Proof. The equivalence of (ii) and (iii) is proved by taking adjoints. To prove that (ii) implies (i) suppose
that Kj , j = 1, 2 are closed, A-invariant, linear subspaces of Hj , j = 1, 2 and that φ : K1 → K2 is an
isomorphism. Define T : H1 → H2 by
T (ξ) = Φ(projK1 (ξ)).
By (ii) we know that T = 0 which implies that Kj , j = 1, 2 are zero.
To see that (i) implies (ii) suppose that T ∈ HomA (ρ1 , ρ2 ). Let T = U |T | be the polar decomposition (see
[13] VIII.3.11) . The fact that T is equivariant implies that T ∗ T is equivariant, and hence that |T | = (T ∗ T )1/2
is, by approximating the square root function by polynomials. Since
U = SOT − lim T (|T | + ε)−1
ε→0
we see that U is equivariant. Thus U gives an A-equivariant isomorphism (ker T )⊥ → im T . Since ρ1 ⊥ ρ2
we find that (ker T )⊥ = 0, and hence that T = 0.
To prove that (iv) implies (ii) let T ∈ HomA (ρ1 , ρ2 ). Let xn be as in (iv). Then, for any ξ ∈ H1
T (ξ) = lim T (ρ1 (x∗n xn )ξ) = lim ρ2 (x∗n xn )T (ξ) = 0,
n→∞
n→∞
so T = 0.
Suppose that (iii) and (ii) hold, we wish to prove (iv). Recall that if H is a Hilbert space and E ⊆ B(H)
then E ′ denotes the commutant of E, i.e. E ′ = {T ∈ B(H) : T S = ST for all S ∈ E}. Suppose that
T ∈ (ρ1 ⊕ ρ2 )(A)′ ,
Then we can regard T as a matrix
T
T = 11
T21
T12
T22
where Tij ∈ B(Hi , Hj ). Since T ∈ (ρ1 ⊕ ρ2 )(A)′ we see that Tij ∈ HomA (ρi , ρj ). Thus T12 , T21 are 0. We
thus see that
SOT
IdH1 0
∈ (ρ1 ⊕ ρ2 )(A)′′ = (ρ1 ⊕ ρ2 )(A)
0
0
where the last equality follows from von Neumann’s Double Commutant Theorem. We now prove (iv) by
using Kaplansky’s Density Theorem.
We need an analogue of the Lebesgue decomposition.
POLISH MODELS AND SOFIC ENTROPY
21
Proposition 4.3. Let A be a unital ∗-algebra and ρj : A → B(Hj ), j = 1, 2 be two ∗-representations. Then
ρ1 = ρ1,s ⊕ ρ1,c
where ρ1,s ⊥ ρ2 , and ρ1,c ≪ ρ2 .
Proof. By Zorn’s Lemma, we can find a maximal family (Kβ )β∈B of pairwise orthogonal, A-invariant, closed,
linear subspaces of H1 so that ρ1 K embeds into ρ2 . Let
β
⊥
M
H1,s = H1 ∩
Kβ .
β∈B
By maximality ρ1
H1,s
is singular with respect to ρ2 . Setting
M
H1,c =
Kβ
β∈B
and defining ρ1,s , ρ1,c by restricting ρ1 to H1,s , H1,c completes the proof.
4.2. Proofs of the Main Applications.
Theorem 4.4. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let (X, M, µ) be a
standard probability space with Γ y (X, M, µ) by measure-preserving transformations. Let H ⊆ L2 (X, µ) be
a closed, linear, Γ-invariant subspace such that M is generated by
Suppose that
{gf −1 (A) : g ∈ Γ, f ∈ H, A ⊆ C is Borel}.
ρΓy(X,µ)
Then
H
⊥ λΓ .
hΣ,µ (X, Γ) ≤ 0.
Proof. Let (fn )∞
n=1 be a k · k2 -dense subset of H. Let
Φ : X → CN×Γ
be defined by
Φ(x)(n, g) = fn (g −1 x)
and let
ν = Φ∗ (µ).
Let Γ y CN×Γ by shifts. Since {gfn−1 (A) : g ∈ Γ, n ∈ N, A ⊆ C is Borel} generates M we see that Φ induces
a Γ-invariant isomorphism (X, µ) ∼
= (CN×Γ , ν). Thus
hΣ,µ (X, Γ) = hΣ,ν (CN×Γ , Γ).
For n ∈ N let Zn : CN×Γ → C be defined by
then Zn ◦ Φ = fn . Thus
ρΓy(CN×Γ ,ν)
Zn (y) = z(n, e),
Span{gZn :g∈Γ,n∈N}
∼
= ρΓy(X,µ)
H
⊥ λΓ .
To simplify notation, we will use ρ for ρΓy(CN×Γ ,ν) . Let ∆ be the dynamically generating pseudometric on
CN×Γ defined by
∞
X
|z(n, e) − w(n, e)|
∆(z, w) =
2−n
.
1
+
|z(n, e) − w(n, e)|
n=1
As CN×Γ is clearly Polish, we can use CN×Γ , ∆, ν to do our computation. Let ε > 0 be arbitrary, and let
0 < η < ε be arbitrary. Let δ > 0 be sufficiently small and F ⊆ Γ be a sufficiently large finite set which will
depend upon ε, η in a manner to be determined later. Given φ ∈ Map(∆, F, δ, σi ) for n ∈ N define zφ,n ∈ Cdi
by
zφ,n (j) = φ(j)(n)(e)
22
BEN HAYES
and define zφ ∈ (CN )di by
zφ (j)(n) = zφ,n (j).
N di
Conversely, given z ∈ (C )
define ψz : {1, . . . , di } → CN×Γ by
ψz (j)(n, g) = z(σi (g)−1 (j))(n)
If F, δ are chosen carefully, then for all sufficiently large i we have that
∆2 (ψzφ , φ) < ε.
Choose N ∈ N so that
2−N < ε.
Since
Γ y Span{gZn : g ∈ Γ, n ∈ N}} ⊥ λΓ ,
by Proposition 4.2 we may find a α ∈ C(Γ) so that max(kλ(α∗ α)k, kρ(α∗ α)k) ≤ 1 and
kρ(α∗ α)Zn − Zn k2 < η, 1 ≤ n ≤ N
kλ(α∗ α)δe k2 < η.
Let M > 0 be sufficiently large depending upon ε in a manner to be determined later. We will assume that
M is large enough so that there exists G ∈ Cc (C) with kGk∞ ≤ M and G(z) = z for |z| ≤ M and
kG ◦ Zn − Zn k2 < ε for 1 ≤ n ≤ N .
Note that G may be chosen independent of η. As
kρ(α∗ α)k ≤ 1
we then have
kρ(α∗ α)G ◦ Zn − G ◦ Zn k2 < 3ε, for 1 ≤ n ≤ N .
(4)
Let L ⊆ Cb (CN×Γ ) be sufficiently large in a manner to be determined later. We will assume that
L ⊇ {G ◦ Z1 , . . . , G ◦ ZN }.
We will use k · k2 for the ℓ2 norm on {1, . . . , di } with respect to the uniform probability measure. By (4) if
F ⊆ Γ, L ⊆ Cb (CN×Γ ) are sufficiently large and δ > 0 is sufficiently small then
kσi (α)∗ σi (α)G ◦ zφ,n − G ◦ zφ,n k2 < 6ε for 1 ≤ n ≤ N ,
for all φ ∈ Mapµ (ρ, F, L, δ, σi ) (with the notational conventions introduced after Definition 2.1). Let p =
χ[1−√ε,1+√ε] (σi (α)∗ σi (α)) (this expression should be interpreted in the sense of functional calculus) then for
all φ ∈ Mapµ (ρ, F, L, δ, σi ),
kpG ◦ zφ,n − G ◦ zφ,n k22 = kχ(√ε,∞) (|σi (α)∗ σi (α) − 1|)G ◦ zφ,n k22
= hχ(√ε,∞) (|σi (α)∗ σi (α) − 1|)G ◦ zφ,n , G ◦ zφ,n i
1
≤ h(σi (α)∗ σi (α) − 1)2 G ◦ zφ,n , G ◦ zφ,n i
ε
1
= k(σi (α)∗ σi (α) − 1)(G ◦ zφ,n )k22
ε
< 36ε
and
tr(p) ≤
Since
1
√ tr(σi (α)∗ σi (α)).
1− ε
lim tr(σi (α)∗ σi (α)2 ) = τ (λ(α)∗ λ(α)) = hλ(α)∗ λ(α)δe , δe i ≤ kλ(α)∗ λ(α)δe k2 < η
i→∞
we see that for all large i we have
tr(p) < 2η
POLISH MODELS AND SOFIC ENTROPY
23
1
provided ε < (2015)!
. Since p is an orthogonal projection, we know by Lemma 2.7 we may choose an ε-dense
subset S of M p Ball(ℓ2 (di , udi )) with
4ηdi
3M + ε
.
|S| ≤
ε
If L and M are sufficiently large, then
Ci = {1 ≤ j ≤ di : |zφ,n (j)| ≤ M for 1 ≤ n ≤ N }
has
|Ci | ≥ (1 − ε)di .
Define wφ ∈ (CN )di by
wφ (j)(n) = G(zφ,n (j)).
For these values of F, δ, L, M we have
2
∆2 (φ, ψwφ ) ≤ ε +
≤ 2ε
di X
N
X
j=1 n=1
2−n
|zφ (n, j) − wφ (n, j)|
1 + |zφ (n, j) − wφ (n, j)|
as wφ (n, j) = zφ (n, j) for j ∈ {1, . . . , di } \ Ci . For 1 ≤ n ≤ N choose ξφ,n ∈ S so that
N di
and define ξφ ∈ (C )
kpG ◦ zφ,n − ξφ,n k2 < ε
by
ξφ (j)(n) = χ{1,...,N } (n)ξφ,n (j).
Then
∆2 (ψwφ , ψξφ )2 ≤ ε +
di
N
X
2−n X
|G(zφ,n (j)) − ξφ,n (j)|
di j=1
n=1
≤ ε + max kG ◦ zφ,n − ξφ,n k2
1≤n≤N
√
≤ 2ε + 6 ε
the second line following from the Cauchy-Schwartz inequality.
Thus
√ √
√
∆2 (φ, ψξφ ) ≤ 2 ε + (2ε + 6 ε)1/2 ,
Hence for all large i,
S2√2√ε+2(2ε+6√ε)1/2 (Mapµ (∆, F, δ, L, σi ), ∆2 ) ≤ |S|N ≤
Thus
√ √
√
hΣ,µ (∆, 4 2 ε + 4(ε + 6 ε)1/2 ) ≤ 4N η log
3M + ε
ε
3M + ε
ε
Note that η can be any number in (0, ε). Thus we can let η → 0 to find that
√ √
√
hΣ,µ (∆, 4 2 ε + 4(ε + 6 ε)1/2 ) ≤ 0
4N ηdi
.
.
since ε > 0 is arbitrary we find that
hΣ,µ (X, Γ) ≤ 0.
The space H can be much smaller than L2 (X, µ). For example consider the case that (X, µ) = (B, η)Γ
and the action is Bernoulli. For f ∈ L2 (B, η) and g ∈ Γ, let fg ∈ L2 (X, µ) be defined by
fg (x) = f (x(g)).
Then we can take
and one can show that
H = Span{fg : g ∈ Γ, f ∈ L2 (B, η)}
Γ y L2 (X, µ) ⊖ H ∼
= ℓ2 (Γ)⊕∞ .
24
BEN HAYES
So indeed the space H is much smaller than L2 (X, µ).
For the next application, recall that the weak topology on Aut(X, µ) is defined by saying that a basic
neighborhood of α is given by UA1 ,...,An ,ε (α) where A1 , . . . , An are measurable subsets of X and ε > 0 and
UA1 ,...,An ,ε (α) =
n
\
j=1
{β ∈ Aut(X, µ) : |µ(β −1 (Aj )∆α−1 (Aj ))| < ε}.
An action Γ y (X, µ) is compact if there is a compact subgroup K ⊆ Aut(X, µ) (for the weak topology) and
a homomorphism π : Γ → K so that
gx = π(g)(x)
for almost every x ∈ X.
Corollary 4.5. Let Γ be a countably infinite discrete sofic group with sofic approximation Σ. Suppose that
Γ y (X, µ) is a compact action. Then hΣ,µ (X, Γ) ≤ 0.
Proof. Recall that a unitary representation ρ : Γ → U (H) is called weakly mixing if
0 ∈ ρ(Γ)
and is compact if
W OT
,
W OT
ρ(Γ)
⊆ U (H).
It is clear that a compact representation has no nontrivial weakly mixing subrepresentations. It is also
well-known that the left regular representation is weakly mixing. Thus if Γ y (X, µ) is compact, then
ρ0Γy(X,µ) ⊥ λΓ
and we may apply Theorem 4.4.
In particular, note that if K is a compact group and φ : Γ → K is a homomorphism, then the action α of
Γ on K given by
α(g)(x) = φ(g) · x
has entropy at most zero with respect to any sofic approximation.
Definition 4.6. Let Γ be a countable discrete sofic group with sofic approximation Σ. We say that a
probability measure-preserving action Γ y (X, µ) has completely positive entropy (with respect to Σ) if
whenever Γ y (Y, ν) is a factor of Γ y (X, µ) and Y is not a one-atom space then hΣ,ν (Y, Γ) > 0.
Corollary 4.7. Let Γ be a countable discrete sofic group with sofic approximation Σ. Suppose that Γ y (X, µ)
is a probability measure-preserving action which has completely positive entropy with respect to Σ. Then
ρ0Γy(X,µ) ≪ λΓ .
Proof. By Proposition 4.3, we can write
where ρΓy(X,µ)
H1
L2 (X, µ) = H1 ⊕ H2
⊥ λΓ and
ρΓy(X,µ)
Suppose that f ∈ H1 . Define
H2
≪ λΓ .
Φ : X → CΓ
by
Φ(x)(g) = f (g −1 x)
and let ν = Φ∗ µ, and let Γ y C be the Bernoulli action. Then Γ y (CΓ , ν) is a factor of Γ y (X, µ). Set
Γ
K = Span{g(f ◦ Φ) : g ∈ Γ}
then
{gξ −1 (A) : ξ ∈ K, BorelA ⊆ C, g ∈ Γ}
generates the Borel subsets of C up to ν-measure zero. Tautologically,
K∼
= Span{gf : g ∈ Γ} ⊥ λΓ .
Γ
POLISH MODELS AND SOFIC ENTROPY
25
Hence by Theorem 4.4 we know that
hΣ,ν (CΓ , Γ) ≤ 0.
Since Γ y (X, µ) has completely positive entropy this implies that (CΓ , Γ) is a one-atom space. But this is
only possible if f is constant. Thus H1 = C1 and
ρ0Γy(X,µ) = ρ
H2
≪ λΓ .
Again the above corollary illustrates the utility in not assuming that H = L2 (X, µ) in Theorem 4.4 but
instead just assuming that H generates (X, µ). If (CΓ , ν) is as in the above proof we do not a priori know
that
Γ y L2 (CΓ , ν) ⊖ C1 ⊥ λΓ .
References
[1] A. Alpeev. On Pinsker factors for Rokhlin entropy. In Representation theory, dynamical systems, combinatorial methods.
Part XXIV, volume 432 of Zap. Nauchn. Sem. POMI, pages 30–35, St. Petersburg, 2015. POMI.
[2] J. Bourgain and A. Gamburd. On the spectral gap for finitely-generated subgroups of SU(2). Invent. Math, 171(1):83–121,
2008.
[3] J. Bourgain and A. Gamburd. Uniform expansion bounds for cayley graphs of SL2 (Fp ). Ann. of Math., 167(2):625–642,
2008.
[4] J. Bourgain and A. Yehudayof. Expansion in SL2 (R) and monotone expanders. Geom. Funct. Anal., 23(1):1–41, 2013.
[5] L. Bowen. Measure conjugacy invariants for actions of countable sofic groups. J. Amer. Math. Soc, 23:217–245, 2010.
[6] L. Bowen. A new measure conjugacy invariant for actions of free groups. Ann. of Math., 171(2):1387–1400, 2010.
[7] L. Bowen. Entropy theory for sofic groupoids I: the foundations. J. Anal. Math., 124(1):149–233, 2014.
[8] R. Bowen. Entropy for group endomorphisms. Trans. Amer. Math. Soc., 153(401-414), 1971.
[9] N. Brown and N. Ozawa. C ∗ -Algebras and Finite-Dimensional Approximations. Cambridge University Press, 1994.
[10] P. Burton. Naive entropy of dynamical systems. arXiv:1503.06360.
[11] L. Ciobanu, D. Holt, and S. Rees. Sofic groups: graph products and graphs of groups. Pacific Journal of Mathematics,
271(1):53–64, November 2014.
[12] A. Connes, J. Feldman, and B. Weiss. An amenable equivalence relations is generated by a single transformation. Ergodic
Theory Dynam. Systems, 10(4):431–450, 1983.
[13] J. Conway. A Course in Functional Analysis. Graduate Texts in Mathematics. Springer, New York, NY, second edition,
2000.
[14] A. Dooley and V. Golodets. The spectrum of completely positive entropy actions of countable amenable groups. J. Funct.
Anal., 196(1):1–18, 2002.
[15] K. Dykema, D. Kerr, and M. Pichot. Orbit equivalence and sofic approximation. arXiv:1102.2556.
[16] K. Dykema, D. Kerr, and M. Pichot. Sofic dimension for discrete measurable groupoids. Trans. Amer. Math. Soc,
366(2):707–748, 2013.
[17] G. Elek and E. Szabo. On sofic groups. J. Group Theory, 9(2):161–171, 2006.
[18] B. Hayes. Fuglede-Kadison determinants and sofic entropy. arXiv:1402.1135.
[19] B. Hayes. Mixing and spectral gap relative to pinkser factors for sofic groups. arXiv:1509.07839
[20] B. Hayes. An lp -version of von neumann dimension for Banach space representations of sofic groups. J. Funct. Anal.,
266(2):989–1040, 2014.
[21] A. Ioana. Orbit equivalence and borel reducibility rigidity for profinite actions with spectral gap. arXiv:1309.3026.
[22] D. Kerr. Sofic measure entropy via finite partitions. Groups Geom. Dyn, 7(617-632), 2013.
[23] D. Kerr and H.Li. Soficity, amenability, and dynamical entropy. Amer. J. Math, 135(3):721–761, 2013.
[24] D. Kerr and H. Li. Bernoulli actions and infinite entropy. Groups Geom. Dyn., 5:663–672, 2011.
[25] D. Kerr and H. Li. Topological entropy and the variational principle for actions of sofic groups. Invent. Math, 186:501–558,
2011.
[26] D. Kerr and H. Li. Combinatorial independence and sofic entropy. Comm. Math. Stat., 1(2):213–257, 2014.
[27] J. Kieffer. A generalized Shannon-McMillan theorem for the action of an amenable group on a probability space. Ann.Prob,
3(6):1031–1037, 1975.
[28] L.Bowen and H.Li. Harmonic models and spanning forests of residually finite groups. J. Funct. Anal., 263(7):1769–1808,
2012.
[29] H. Li. Compact group automorphisms, addition formulas and Fuglede-Kadison determinants. Ann. of Math., 176(1):303–
347, 2012.
[30] H. Li. Sofic mean dimension. Adv. Math, 244:570–604, 2014.
[31] A. Lubotzky, R. Phillips, and P. Sarnak. Ramanujan graphs. Combinatorica, 8(3):261–277, 1988.
[32] G. Margulis. Some remarks on invariant means. Monatsh. Math., 90(3):233–234, 1980.
[33] T. Meyerovitch. Positive sofic entropy implies finite stabilizer. arXiv:1504.08137.
[34] D. Ornstein and B. Weiss. Ergodic theory of amenable groups I. The Rokhlin lemma. Bull. Amer. Math. Soc.(N.S.),
2(1):161–165, 1980.
26
BEN HAYES
[35] D. Ornstein and B. Weiss. Entropy and isomorphism theorems for actions of amenable groups. J. Anal. Math, 48:1–141,
1987.
[36] L. Paunescu. On sofic actions and equivalence relations. J. Funct. Anal., 261(9):2461–2485, November 2011.
[37] S. Popa. On the superrigidity of malleable actions with spectral gap. J. Amer. Math. Soc., 21(4):981–1000, 2008.
[38] S. Popa. Independence properties in sublagebras of ultraproduct II1 factors. J. Funct. Anal., 266(9):5818–5846, 2014.
[39] A. Selberg. On the estimation of Fourier coefficients of modular forms. In Proc. Sympos. Pure Math, volume 8, pages 1–15,
Providence, RI, 1965. American Mathematical Society.
[40] B. Seward. Finite entropy actions of free groups, rigidity of stabilizers, and a Howe–Moore type phenomenon. To appear
in J. Anal. Math.
[41] B. Seward. Krieger’s finite generator theorem for ergodic actions of countable groups I. arXiv:1405.3604.
[42] J. Sinaı̌. On a weak isomorphism of transformations with invariant measure. Mat. Sb. (N.S.), 63(105):23–42, 1964.
[43] D. Sullivan. For n > 3 there is only one finitely additive rotationally invariant measure on the n-sphere on all Lebesgue
measurable sets. Bull. of the AMS, 4(1):121–123, 1981.
[44] V. Varadarajan. Measure on topological spaces. Mat. Sb. (N.S)., 55(97):35–100, 1961.
Stevenson Center, Nashville, TN 37240
E-mail address: [email protected]
| 4math.GR
|
Eliminating Network Protocol Vulnerabilities
Through Abstraction and Systems Language Design
C. Jasson Casey∗ , Andrew Sutton† , Gabriel Dos Reis† , Alex Sprintson∗
∗ Department
of Electrical and Computer Engineering, Texas A&M University
of Computer Science, Texas A&M University
arXiv:1311.3336v1 [cs.NI] 13 Nov 2013
† Department
Abstract—Incorrect implementations of network protocol message specifications affect the stability, security, and cost of
network system development. Most implementation defects fall
into one of three categories of well defined message constraints.
However, the general process of constructing network protocol
stacks and systems does not capture these categorical constraints. We introduce a systems programming language with
new abstractions that capture these constraints. Safe and efficient
implementations of standard message handling operations are
synthesized by our compiler, and whole-program analysis is used
to ensure constraints are never violated. We present language
examples using the OpenFlow protocol.
I. I NTRODUCTION
The message handling layer of any network protocol is
notoriously difficult to implement correctly. Common errors
include: accepting or allowing the creation of malformed
messages, using incorrect byte ordering or byte alignment,
using undefined values, etc. (§ II). These defects lead to problems in stability, security, performance, and cost for network
systems. Table I, the result of a survey of the US-CERT
Vulnerability Database [7], demonstrates that sophisticated
organizations implementing mature protocols commit these
errors. The persistent introduction of these defects is not the
sign of an engineering problem, but a failure to use the correct
levels of abstraction when working with network protocols.
Message handling has been the focus of several research
efforts. When the wire-format of the message is not important,
serialization solutions can be used [15], [6]. However, with
network protocols, because of interoperability requirements,
adherence to the specific wire-format is necessary. As a result,
a series of Domain Specific Languages (DSLs) that allow
programmer control over the wire-format have been designed.
These approaches synthesize data structures to hold messages
and the typical operations necessary to manipulate them using
correct by construction techniques [10], [2], [12]. Language
researchers have improved upon these DSLs with rich type
systems that can prove certain safety properties, and address
some of the problems mentioned previously [4], [5]. Other
work developed static analysis techniques, that require no
domain knowledge, to survey existing code bases and find
occurrences of some of the previously mentioned defects [3].
This material is based upon work partially supported by by the AFOSR
under contract No. FA9550-13-1-0008, and by the NSF under grants CCF1150055 and ASI-1148461.
978-1-4799-1270-4/13/$31.00 c 2013 IEEE
Systematically eliminating the categories of message related
defects requires rich type systems and whole program analysis, which is not supported by existing declarative DSLs.
Invariants and semantic information produced by the DSL is
not incorporated or used in program analysis by the target
language. Furthermore, finding all occurrences of message
related defects require some level of domain knowledge during
static analysis. This must either be built in to the language or
programmer specified in a way that resembles existing network
protocol specifications. Analysis by formal methods should
be a by-product of compiling the network program, and not
require any specialized knowledge by the programmer. Our
work-in-progress develops a systems programming language
to address these issues. This allows for full program analysis,
providing stronger safety guarantees and offering domain
specific optimization that is exceedingly difficult to accomplish
by hand or impossible with a DSL.
In this paper we clearly identify categories of message
related vulnerabilities by their structural and semantic constraints. We show that these categories are responsible for
known vulnerabilities, and using our tools we show that even
some live Internet traffic violates these constraints. We then
introduce a systems programming language that allows programmers to capture network protocol message structure and
constraints. The constraints allow the compiler to reason over
entire programs, identifying and eliminating the categories of
vulnerabilities mentioned before. Additionally, the choice of
a systems programming language allows for efficient code
generation. Throughout the paper we use OpenFlow [11] as
our reference protocol. Our contributions are:
• the identificaiton of three categories of network protocol vulnerabilities common in network programs and in
network traffic (in § II),
• the development of abstractions that prevent the construction of messages that lead to these vulnerabilities, and
unsafe access of conditional fields (in § II and § III),
TABLE I
M ESSAGE R ELATED V ULNERABILITIES
Proto.
802.11i
OSPFv2
NTPD
ICMP
VTP
Bootp
Age
2004
1998
1985
1981
1996
1985
Bug Date
2012
2012
2009
2007
2006
2006
Vendor
Broadcom
Quagga
GNU
Cisco
Cisco
Apple
Error
semantic
struct
struct
both
semantic
struct
CERT #
160027
551715
853097
341288
821420
776628
•
•
the design of a systems programming language that
eliminates unsafe protocol implementations through type
checking (in § III, § IV, and § V), and
the implementation of a compiler and library supporting
the language (in § VI).
II. M ESSAGE V ULNERABILITIES
There are three categories of message vulnerabilities that we
address: structural constraint violation, semantic constraint violation, and unsafe access of conditional fields. In this section
we describe these categories in detail with examples using
the OpenFlow v1.0 protocol [11]. We first briefly describe the
protocol and then show vulnerability examples.
Version(s)
1 - 1.0
2 - 1.1
3 - 1.2
4 - 1.3
0x00
0x04
0x08
...
version
Type(s)
0 - HELLO
…
14 - FLOW_MOD
...
32 bits
type
xid
length
Payload(type)
Fig. 1.
HELLO
0
14
...
FLOW_MOD
...
OpenFlow Message Format
Figure 1 summarizes the message format, or representation,
of an OpenFlow message. A message consists of a fixed
8 byte header followed by a variable sized payload. The
header indicates the version of the OpenFlow protocol, the
type of the payload, the length of the entire message, and a
transaction identifier used to match response messages to their
requests. The payload can be one of 22 types in version 1.0.
This protocol was designed to operate over stream-oriented
transports, which are more difficult to handle than datagram
or message oriented transports. Streams have no concept of
message boundaries and can give the program anything from
a single byte to several messages in a single read. It is the
programmer’s responsibility, not the transport’s, to determine
where the payload ends and the next message begins.
Semantic constraints ensure that a message field’s value
has a defined meaning in the protocol. For instance, in OpenFlow 1.0 the domain of the version field is [1, 1], the domain
of the type field is [0, 21], the domain of the length field is
[8, 216 − 1], and the domain of the xid field is [0, 232 − 1].
Any value that is not in the domain of its field is semantically
invalid; it has no meaning in the protocol definition. Violating
a semantic constraint is similar to using undefined behavior
in a programming language. As Table I shows, constraint
violations lead to vulnerabilities.
Structural constraints address how messages are constructed and used by a program. Messages are constructed
in two ways: either by a program to send over a network, or
from the network to send to the program. In both cases it is
important to construct only structurally valid messages.
These constraints deal with the number of bytes a message
occupies in buffers used to communicate with the network.
Structural constraint testing is the process of ensuring there
are enough bytes to complete the operation of constructing a
message from a buffer, or filling a buffer with a well-formed
message. For example, any buffer containing less than 8 bytes
cannot possibly represent a valid OpenFlow message and any
attempt to interpret that buffer as a message, would be an error.
Semantic constraint violations can lead to structural constraint violations. Reading from a stream can produce a buffer
containing several messages. The header of the first message
must be contained in the first 8 bytes, and the payload must
end at the position where the header’s length field indicates.
In the header, the length field can be semantically invalid, but
because it is used to constrain the payload’s size, it becomes a
structural constraint violation as well. A similar problem arises
with a semantically invalid type field that is used to choose
the payload.
Safe access ensures that fields with run-time dependency
are always validated before use. Many fields in OpenFlow are
dependent; their meaning is determined by the values of previously encountered fields. For example, there is a dependency
between the payload of a message and the header’s type field.
A structurally and semantically valid Hello message cannot
have its payload treated as the FlowMod type without invoking
undesired behavior.
III. L ANGUAGE
This work builds on fundamental notions from system programming [8], [14], from structured generic programming, and
mathematical programming languages such as AXIOM [9],
and Liz [13]. This work, heavily inspired by the Liz language,
aims to support simple, safe, and efficient handling of network
protocol messages. The core of our language supports values,
references, constants, functions, records, a minimal set of
expressions, and follows Call-By-Value semantics. We do not
expose pointers to the users of the language and drastically
restrict heap allocation to certain language built-in types.
Figure 2 shows the abstract syntax of the language. The
dependent types ω, are a primary contribution of this paper.
Our language captures the structural and semantic constraints of a message with user-defined type and variable declarations. We then enforce these constraints through the process
of object and symbolic construction. If object construction
completes successfully, then structural constraints are upheld,
if symbolic construction completes successfully, then semantic
constraints hold. Using construction to establish invariants is
a common way to reason about program behavior.
An object, or instance of a type, must be constructed before
use. The process of object construction involves allocating
space where the object will live, and initializing its values to
establish its invariant. Symbolic construction extends object
construction to include ensuring the value of the object is
consistent with its symbolic constructor. Upon completion of
construction, an object is well-formed and its invariant has
been established.
The ω types, see Figure 2, allow user definition of
precise structural and semantic constraints. Structural constraints can be explicitly stated by the specifiers: bits(e), and
constraint(e). Structural constraints are otherwise implicit
in the type, which will be explained below. Semantic con-
τ := unit | bool | byte | char | int
| uint | string | const(τ)
| ref(τ) | τ → τ | type | Buffer | View | ω
spec := bits(e) | constraint(e)
xform := msbf | lsbf
ω := uint(spec[, xform]) | array(τ, e)
| vector(τ) | record{decl}
| variant{τ if e}
e := c | x | e.x | e‘[’e‘]’ | x(e)) | return(e)
| if(e)then(e)else(e) | while(e)e | block
block := {decl|e}
define Hdr:type = record {
vrsn: uint(bits(8));
type: uint(bits(8));
len: uint(bits(16), msbf);
xid: uint(bits(32), msbf);
}
define Pld(x:uint):type = variant {
Hello if x == 0;
...
FlowMod if x == 14;
...
}
define valid_hdr(h:cref(Hdr)):bool = {
return h.vrsn == 1
and h.len >= bytes(h);
}
define Msg:type = record {
hdr: Hdr | valid_hdr;
pld: Pld(hdr.type,
constraint(hdr.len - bytes(hdr)));
}
func := x(decl) : τ = block
decl := x : τ[‘|’e][= e]
toplevel := decl|func
Fig. 2.
Core Language Syntax
straints are introduced through the declaration syntax. Any
declaration can impose a semantic constraint through the
use of the bar, |, followed by a guard expression. These
declarations provide constraint information to the compiler
that is used in full program analysis. This allows the compiler
to reason about both constraint satisfaction and safe usage.
The following is a short summary of the ω types:
• uint(spec[, xform]) defines an unsigned integer that has
a precise bit width specified by its structural constraint
(spec). Optionally, the type also takes a transform parameter (xform) that allows for non-native representation
of the data. For instance, a protocol may specify that
a value’s representation is Most Significant Byte First
(MSBF) and 1’s complement.
• array(τ, e) and vector(τ) types allow for sequences of
objects of type τ. The array is statically sized to contain
e elements, while the vector has a dynamic size.
• record{decl} is a sequence of declarations, whose objects can be accessed by field name. There is no padding,
alignment, or meta-data applied to the object by the
compiler. If padding, alignment, or meta-data is desired,
the programmer can explicitly declare them as fields.
Additionally, the order of fields is preserved.
• variant{τ if e} is a union of types where each type is
guarded by a predicate. A variant is constructed by evaluating the predicate set and invoking the constructor of
the corresponding true predicate. A variant is uninitialized
until evaluation of the predicate set.
Figure 3 demonstrates how to declare types corresponding
to the OpenFlow header, payload, and message. The header is
a simple record of four fields, all of which have constant specifiers, and follow MSBF ordering. The payload type is a unique
choice of types based on a type parameter. For the values in
the header to have semantic meaning in version 1.0 of the
Fig. 3.
OpenFlow v1.0 Message Declaration
OpenFlow protocol we have to constrain their values. This is
achieved with the semantic constraint valid_hdr, which must
first be defined as a function that takes a constant reference
to a header and returns bool. Message is defined as a record
including a header with version 1.0 semantic constraints, and a
payload that is parameterized over the header’s type field and
constraint. All type constructors, with the exception of uint
and array allow for an optional specifier constraint. In this
particular case, construction of Pld is not to exceed the result
of the constraint hdr.len − bytes(hdr).
Buffer & View are abstractions over the underlying machine architecture that help the compiler ensure structural constraints are never violated. Reading from a file or socket results
in a buffer with begin and end boundaries surrounding the
bytes received. View is a mechanism that restricts visibility
into a Buffer. The set of operations defined for Buffer and
View are:
• view: returns a view of an entire buffer
• available: returns the byte size of a view
• advance: returns a view with an advanced head
• constrain: returns a view with a constrained tail
• put: writes a value to a view
• get: reads a value from a view
View
View
msg
msg
Buffer
msg
constrain
Fig. 4.
Constraining a View limits the number of bytes available for access
Figure 4 illustrates a buffer returned from a read system call
against a TCP socket. A single read has resulted in more than
one protocol message. The initial view wraps all of the data;
however, the first few bytes of the view contain a protocol
header, which provides the length of the first message. This
length is then used to constrain the visibility to precisely one
message. The constrain operation supports use of datagram
and stream oriented transports, while also providing safety
boundaries for object construction.
IV. C OMPILER S YNTHESIS
Programmers continue to make mistakes implementing message operations that are necessary in all protocols such as those
mentioned in § I and § VI. Our strategy is to eliminate the
need to write these common operations by having the compiler
synthesize safe and efficient versions. All user-defined type
definitions contain structural and semantic constraints. This
information is sufficient to synthesize the following operations:
• construction: constructs an object from expressions
• copy_construction: constructs an object from another
• assignment: copies an object’s state into another
• bytes: returns the number of bytes of an object
• to_view: writes an object to a view
• from_view: constructs an object from a view
• equal, not_equal: compare objects for equivalence
• to_string: returns a string representation of the object
The remainder of this section will describe the synthesis
process for a small subset of the above operations focusing on
synthesis and constraint validation for bytes and from_view.
Bytes is the name of the operation for determining the byte
size of any object. The operation bytes is synthesized for each
declared ω type in a program. For uint it returns the number
of bytes indicated by the specifier. For array it returns the
result of sizeof(T )∗elements, where T is the type contained
by the array. For objects of type Vector, bytes is not a
constant expression. It has run-time dependencies and returns
the accumulation of calling bytes over its elements. Record
returns the sum of calling bytes over its constituent fields,
and is only a constant if all of its fields are also constant.
Calling bytes over a variant returns 0 if the variant is
uninitialized, or it proxies the call to the contained object.
Figure 5 illustrates the process just described in pseudo code.
Several other synthesized operations depend on bytes.
Structural constraint violations must be prevented during
object construction. There are only three ways to violate a
structural constraint: overflowing a view, underflowing a view,
or constructing an invalid variant. Overflowing a view involves
advancing the view beyond the number of bytes contained.
Underflowing a view involves constraining a view by more
bytes than contained. Constructing an invalid variant is caused
when none of the variant’s contained type predicates evaluate
to true.
In order to construct the OpenFlow header the size of
the view must be at least as large as the number of bytes
of the header, or 8 bytes. During object construction the
view is always advanced by the size in bytes of the object.
Constructing the message object from a view of 7 or less bytes
would result in a view overflow.
The OpenFlow protocol indicates the length of a message
with the header length field. This value is used to constrain the
view for payload construction. It is possible, through accident
or malicious intent, for the length field to be inconsistent with
define bytes(u : uint(bits(v), f)) : uint = {
return v/8;
}
define bytes(a : cref array(T,x)) : uint = {
return sizeof(T) * x;
}
define bytes(v : cref vector(T) : uint = {
accum : uint = 0;
foreach(item in v) {
accum += bytes(v);
}
return accum;
}
define bytes(r : cref record) : uint = {
return bytes(r.x1)+bytes(r.x2)+...+bytes(r.xN);
}
define bytes(v : cref variant) : uint = {
if(not init(v))
return 0;
switch(v.kind) {
case variant::K1 => return bytes(T1(v.value));
case variant::K2 => return bytes(T2(v.value));
...
case variant::KN => return bytes(TN(v.value));
}
}
Fig. 5.
Synthesis rules for the bytes operation
the amount of data actually sent. If the field indicated less than
8 bytes of payload it could be possible to underflow the view.
The value used to construct the variant payload is in the
header. Again, either through accident or malicious intent, it is
possible for the header to indicate a type which will not result
in valid variant construction. Using a variant in an invalid way
will result in undefined behavior.
Object construction is possible with either a constructor that operates over expressions or using the from_view
operation. Figure 6 illustrates pseudo code for synthesizing
from_view. The operation returns false when a structural
constraint has been violated. Failure indicates a partially
constructed object. For simple types, such as uint and array,
the structural constraints are always checked. If there are
not enough bytes in the view to complete the operation, the
operation fails. Otherwise, the object’s value is constructed by
reading from the view. If a xform is present the object’s value
is updated using the specified transform. Finally, the view is
advanced by the size of the object just constructed.
The vector version of from_view operates in a greedy
fashion, it will consume the entire view. If this behavior is
not desired the view must be constrained before construction.
As long as there are bytes in the view the vector will attempt to
construct an object. Upon success, the object is inserted into
the vector and the process repeats. The record version will
attempt to construct its constituent fields and either return at
the first failure or succeed. The variant must guard against
the third type of structural constraint; it must be initialized
to a valid type. If this is not true then from_view will fail,
otherwise from_view is called over the appropriate type.
Symbolic construction ensures that all run-time type dependencies must be propagated and semantic constraints are
inserted into synthesized code. The OpenFlow message from
Figure 3 has semantic constraints, a run-time type dependency,
and a constrained view. The semantic constraint turns into a
predicate check immediately after the call to from_view of
the header. If the check fails, the operation immediately returns
define from_view(v:ref(view), u:uint(bits(v),f)):bool={
if (available(v) < bytes(u))
return false;
get(v, u);
u = f(u);
advance(v, bytes(u));
return true;
}
define from_view(v:ref(view), a:ref(array(T,e))):bool={
if (available(v) < bytes(a))
return false;
foreach(x in a) {
from_view(v, x);
}
return true;
}
define from_view(v:ref(view), vc:ref(vector(T))):bool={
while(available(v) > 0) {
t:T;
if (not from_view(v, t))
return false;
push(vc, t);
}
return true;
}
define from_view(v:ref(view), r:ref(record)):bool={
if (not from_view(v, r.x1))
return false;
if (not from_view(v, r.x2))
return false;
...
return from_view(v, r.xN);
}
define from_view(v:ref(view), vr:ref(variant)):bool={
if(not init(vr))
return false;
switch(vr.kind) {
case vr::K1 => return from_view(v, T1(vr.value));
case vr::K2 => return from_view(v, T2(vr.value));
...
case vr::KN => return from_view(v, TN(vr.value));
}
}
Fig. 6.
Synthesis rules for the from_view operation
define from_view(v:ref(view), m:Msg):bool = {
if(not from_view(v,m.hdr))
return false;
if(not valid_hdr(m.hdr))
return false;
if(not construct(m.pld, m.hdr)) return false;
return from_view(constrain(v,m.hdr.len-bytes(m.hdr)),
m.pld);
}
Fig. 7.
the compilation of the program with useful error messages.
The goal of the compiler is to synthesize safe and efficient
code. The synthesis algorithm described previously will produce safe but inefficient code. The operations to_view and
from_view contain fine-grained guards that protect against
view underflow and overflow. These fine-grained guards are
the source of inefficiency. For example, it is better to have a
single guard over a sequence of objects of constant size, then
to have a guard over each object. We call this guard fusing, it
is analogous to fusing basic blocks to form larger basic blocks.
It is also useful to lift guards from inside the called function
to the call site. By lifting guards to the call site, potential
guard fusing optimization becomes possible. The optimization
strategy that we follow is a set of fusing and lifting rounds
that reduce the number of guards and form the largest possible
object construction basic blocks for to_view and from_view.
All synthesized operations have a similar Call Sequence
Graph (CSG). Figure 8 illustrates the generalized CSG for a
Flow Modification message. Each node represents a function
in the call sequence, the function types are indicated by
the node shape in the figure. This CSG is used to both
synthesize operation definitions and analyze safe usage of
messages. Guards start at the leaves of the graph and are lifted
to their parent nodes. If all guards can be fused within an
interior node, then the new guard is lifted and the process
is repeated. This process ensures guards covering the largest
possible constant sized objects are performed, additionally this
process is unaware of protocols and will optimize across layer
boundaries. Sometimes a node can have more than one parent
node, where the parents have differing behaviors. In this case,
we split the node into two versions where we lift the guard
when it is contained within a constant structure, or leave it in
place otherwise.
Message
Header
record
variant
vector
uint
Synthesis of from_view for Msg
Flow_Mod
with failure. Next, the type parameter must be checked and
initialized before the call to from_view over the payload. The
check ensures that failure happens if the value is undefined,
and if upon success initializes the payload’s kind. Finally, the
constrained view operation is propagated. The result of this
final step is shown in Figure 7.
Match
Action
Action_Hdr
V. S AFETY A ND O PTIMIZATION
Action_Payload
There are only three ways to violate structural constraints:
view underflow, view overflow, and reading or writing from
an uninitialized variant. These three categories of mistakes
can be identified with two simple invariants. If the bytes
of a view are always non-negative then it is impossible to
underflow or overflow the view. If a read or write of a variant is
always preceded by a valid initialization then the third category
is also impossible. The compiler uses a dataflow analysis
framework to prove that these invariants always hold or fail
Fig. 8.
Flow Modification Control Flow Graph
Code generation takes place after the optimization phase.
We currently support C++11 as our target language. All
message related type definitions and synthesized operations
will be written to a single set of .cpp and .hpp files, the
program itself is written to main.cpp.
VI. I MPLEMENTATION AND E VALUATION
Protocol implementations that send messages that are either
structurally or semantically invalid exist. Applying this work
we were able to discover structural constraint violations in
packet traces from core Internet routers. Furthermore, we
were able to define three categories of message constraints,
and show violation of these constraints can lead to high
profile vulnerabilities. Network programs must always handle
messages in a safe manner. To this aim, we developed a
systems programming language and library for writing safe
efficient network programs.
The language implementation was originally developed
as a C++11 library. The library was used to test ideas and
guide the language design. However, in order to enforce safety
guarantees with optimization a compiler was necessary. We
experimented with the language using two types of network
programs: a protocol analyzer, and an OpenFlow stack. This
set of applications provided good coverage over the diversity
of protocol formats and their constraints.
Core Internet traces were obtained from Caida [1] as test
data for packet analyzers written in our new language facilities.
The traffic was recorded from high speed interfaces, OC-192
(up to 10Gbps). Only the Layer 3 and Layer 4 headers, timing,
and summary information are present. The layer 3 addresses
were randomized and the layer 4 payloads were removed for
anonymization purposes. Each trace has between 500 MB and
1 GB of data and was timestamped in minute intervals. We
analyzed a 10 minute segment of traces from October of 2012.
We focused on looking for structural and value constraints
violations within IPv4, IPv6, TCP, and UDP. Table II shows
that we found structural constraint violations in all but one
protocol; no semantic constraint violations were found. IPv4
violated its structural constraint with regards to IPv4 Options.
The values in the Internet Header Length (IHL) field indicated
a number of Options that should be constructed; however, this
packet would overflow its view, the received block of data was
too small. The TCP and UDP structural constraint violations
were of the same nature; they violated the basic constraint of
a minimum sized header.
TABLE II
C AIDA T RACES
Desc.
Count
CDF
Struct
IPv4
247,849,217
99.95%
16
IPv6
130,760
0.05%
0
TCP
221,243,574
89.22%
84,274
UDP #
23,633,921
9.53%
86,123
The source of these structural constraint violations is not
currently known. It could be evidence of unintentional errors
in sending devices, it could be maliciously crafted packets,
or it could be due to the collection process of the trace data.
However, regardless of the source, structural constraints have
been violated and these packets should not be admitted to safe
network programs.
VII. C ONCLUSION A ND F UTURE W ORK
Incorrect implementations of protocol message specifications affect the stability of network systems and potentially
lead to vulnerabilities. In this paper we identified three categories of constraints that can be used either to test whether a
message is well-formed or to generate safe code. We developed
a systems programming language that allowed user-defined
types to capture these constraints as well as a reasoning framework to ensure these constraints are always upheld within the
users program. We presented example type definitions and
compiler synthesized code using the OpenFlow 1.0 protocol.
The next steps for this work fall into two categories:
extending the ω types and formalizing the meta-system.
Extending the type system will allow for the support of more
protocols. Vectors will be extended to support termination
predicates as a structural constraint parameter, this will allow for self-terminating sequences such as null-terminated
character strings. Generalized enumerations will be added as
an easier mechanism for restricting values used in message
construction. Finally, work on the meta-system is focused on
generating proof certificates that can be used for mechanical
verification of safety.
R EFERENCES
[1] The CAIDA Anonymized 2012 Internet Traces - 20120119-125903, kc
claffy, Dan Andersen, Paul Hick.
[2] G. Back. Datascript- a specification and scripting language for binary
data, 2002.
[3] J. Condit, M. Harren, Z. Anderson, D. Gay, and G. Necula. Dependent
types for low-level programming. Programming Languages and Systems,
pages 520–535, 2007.
[4] K. Fisher and R. Gruber. PADS: a domain-specific language for
processing ad hoc data. SIGPLAN Not., 40(6):295–304, June 2005.
[5] K. Fisher, Y. Mandelbaum, and D. Walker. The next 700 data description
languages. SIGPLAN Not., 41(1):2–15, January 2006.
[6] Google. Protocol buffers, 2012. http://code.google.com/p/protobuf/.
[7] US Government. United States Computer Emergency Readiness Team,
2012.
[8] International Organization for Standards.
International Standard
ISO/IEC 14882. Programming Languages — C++, 2nd edition, 2003.
[9] R. D. Jenks and R. S. Sutor. AXIOM: The Scientific Computation System.
Springer-Verlag, 1992.
[10] P. J. McCann and S. Chandra. Packet types: abstract specification
of network protocol messages. In Proceedings of the conference on
Applications, Technologies, Architectures, and Protocols for Computer
Communication, SIGCOMM ’00, pages 321–333, New York, NY, USA,
2000. ACM.
[11] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
J. Rexford, S. Shenker, and J. Turner. Openflow: enabling innovation
in campus networks. SIGCOMM Comput. Commun. Rev., 38(2):69–74,
March 2008.
[12] R. Pang, V. Paxson, R. Sommer, and L. Peterson. binpac: a yacc for
writing application protocol parsers. In Proceedings of the 6th ACM
SIGCOMM conference on Internet measurement, IMC ’06, pages 289–
300, New York, NY, USA, 2006. ACM.
[13] G. Dos Reis. A System for Axiomatic Programming. In AISC/MKM/Calculemus, pages 295–309, 2012.
[14] Alexander A Stepanov and Paul McJones. Elements of programming.
Addison-Wesley Professional, 2009.
[15] International Telecommunication Union.
Abstract syntax
notation one (ASN.1). Technical report, 2002. Available from:
http://www.itu.int/ITU-T/studygroups/com17/languages/X.6800207.pdf.
| 6cs.PL
|
Ping-pong configurations and circular orders on free groups
arXiv:1709.02348v2 [math.GR] 14 Nov 2017
Dominique Malicet
∗
Kathryn Mann
†
Cristóbal Rivas
‡
Michele Triestino
§
Abstract
We discuss actions of free groups on the circle with “ping-pong” dynamics; these are dynamics
determined by a finite amount of combinatorial data, analogous to Schottky domains or Markov
partitions. Using this, we show that the free group Fn admits an isolated circular order if and
only if n is even, in stark contrast with the case for linear orders. This answers a question from
[21]. Inspired by work in [2], we also exhibit examples of “exotic” isolated points in the space of
all circular orders on F2 . Analogous results are obtained for linear orders on the groups Fn × Z.1
1
Introduction
Let G be a group. A (left-invariant) linear order, often called a left order on G is a total order
invariant under left multiplication. Left-invariance directly implies that the order is determined by
the set of elements greater than the identity, called the positive cone. It is often far from obvious
whether a given order can be determined by only finitely many inequalities, or whether a given
group admits such a finitely-determined order. This latter question turns out to be quite natural
from an algebraic perspective, and can be traced back to Arora and McCleary [3] for the special
case of free groups. McCleary answered this question for free groups shortly afterwards, showing
that Fn has no finitely determined orders [25].
The question of finite determination gained a topological interpretation following Sikora’s
definition of the space of linear orders on G in [29]. This space, denoted LO(G), is the set of all
linear orders on G endowed with the topology generated by open sets
U(,X) := {0 | x 0 y iff x y for all x, y ∈ X}
as X ranges over all finite sets of G. Finitely determined linear orders on G are precisely the isolated
points of LO(G); going forward, we will refer to these as isolated orders. This correspondence between
isolated points and finitely determined orders is perhaps the simplest instance of the general theme
that topological properties of LO(G) should reflect algebraic properties of G.
Presently, several families of groups are known to either admit or fail to admit isolated orders,
with proofs that use both purely algebraic and dynamical methods. Some examples of groups that
do not admit isolated orders include free abelian groups [29], free groups [25, 26], free products of
arbitrary linearly orderable groups [28], and some amalgamated free products such as fundamental
groups of orientable closed surfaces [1]. Large families of groups which do have isolated orders
∗
LAMA, Université Paris-Est Marne-la-Vallée, CNRS UMR 8050, 5 bd. Descartes, 77454 Champs sur Marne,
France.
†
Dept. Mathematics, Brown University, 151 Thayer Street, Providence, RI 02912.
‡
Dpto. Matemática y C.C. Universidad de Santiago de Chile, Alameda 3363, Estación Central, Santiago, Chile.
§
IMB, Université Bourgogne Franche-Comté, CNRS UMR 5584, 9 av. Alain Savary, 21000 Dijon, France.
1
MSC2010: Primary 20F60, 57M60. Secondary 20E05, 37C85, 37E05, 37E10, 57M60.
1
include braid groups [11, 15], groups of the form hx, y | xn = y m i (n, m ∈ Z) [19, 27], and groups
with triangular presentations [10]. (In fact, all of these latter examples have orders for which the
positive cone is finitely generated as a semi-group, a strictly stronger condition.) As a consequence
of our work here, we give a family of groups where, interestingly, both behaviors occur.
Theorem 1.1. Let Fn denote the free group of n generators. The group Fn × Z has isolated linear
orders if and only if n is even.
This result appears to give the first examples of any group G with a finite index subgroup H (in
this case Fn × Z ⊂ Fm × Z, for n odd and m even) such that LO(G) and LO(H) are both infinite,
but only G contains isolated points.
Theorem 1.1 also has an interesting consequence regarding the space of marked groups. As shown
in [8, Prop. 2.13], the set of left-orderable groups is a closed subset of the space of marked groups
on n generators. However, Theorem 1.1 implies that this is not the case either for the subset of
groups admitting isolated linear orders (or its complement): one may take a sequence of markings of
F2 × Z so as to approach F3 × Z, and similarly, a sequence of markings on F3 × Z can be chosen to
approach F4 × Z (see [8, §2.4]). Thus, Theorem 1.1 immediately gives the following.
Corollary 1.2. In the space of finitely generated marked groups, having an isolated linear order is
neither a closed, nor an open property.
The main tool for Theorem 1.1, and main focus of this work, is the study of circular orders on Fn
and the dynamics of their corresponding actions of Fn on S1 . It is well known that, for countable G,
admitting a linear order is equivalent to acting faithfully by orientation-preserving homeomorphisms
on the line. In the same vein, a circular order on G is an algebraic condition which, for countable
groups, is equivalent to acting faithfully by orientation-preserving homeomorphisms on S1 . We recall
the definition and basic properties in Section 3. Any action of G on S1 lifts to an action of a central
extension of G by Z on the line, giving us a way to pass between circular and linear orders on these
groups, and giving us many dynamical tools for their study.
Analogous to LO(G), one can define a space of circular orders CO(G). In [21], the second and
third authors showed that a circular order on Fn is isolated if and only if the corresponding action
on the circle has what they called ping-pong dynamics. They gave examples of isolated circular
orders on free groups of even rank, but the odd rank case was left as an open problem. Here we
answer this question in the negative:
Theorem 1.3. Fn admits an isolated circular order if and only if n is even.
Similarly to Corollary 1.2, one can also prove that the set of groups admitting isolated circular
orders is neither closed nor open in the space of marked groups.
We prove Theorem 1.3 by developing a combinatorial tool for the study of actions on S1 with
ping-pong dynamics (similar to actions admitting Markov partitions), inspired by the work in [13]
and [2]. We expect these to have applications beyond the study of linear and circular orders; one
such statement is given in Theorem 3.9. The notion of ping-pong dynamics is defined and motivated
in the next section. Sections 3 and 4 give the application to the study of circular and linear orders,
respectively, and the proofs of Theorems 1.1 and 1.3.
2
Ping-pong actions and configurations
Definition 2.1. Let G = Fn be the free group of rank n, freely generated by S = {a1 , . . . , an }.
A ping-pong action of (G, S) on S1 is a representation ρ : G → Homeo+ (S1 ) such that there exist
2
D(b)
ρ(a) S1 \ D(a−1 )
D(a)
D(a−1 )
D(b−1 )
Figure 2.1: Classical ping-pong on two generators
pairwise disjoint open sets D(a) ⊂ S1 , a ∈ S
∪ S −1 , each of which has finitely many connected
components, and such that ρ(a) S1 \ D(a−1 ) ⊂ D(a). We further assume that if I and J are any
connected components of D(a), then I¯ ∩ J¯ = ∅.
We call the sets D(a) the ping-pong domains for ρ.
A similar definition is given in [21], with the additional requirement that ping-pong domains be
closed. The above, more general definition is more natural for our purposes, although we will later
introduce Convention 3.7 to reconcile the two. The reader may notice that, for a given ping-pong
action ρ of (G, S), there can be many choices of sets D(a) satisfying the property in Definition 2.1.
S
For instance, if ρ is a ping-pong action such that a∈S∪S −1 D(a) 6= S1 , then one may choose an
S
arbitrary open set I disjoint from a∈S∪S −1 D(a) and replace D(a1 ) withD(a1 ) ∪ I, leaving the
other domains unchanged. These new domains still satisfy ρ(a) S1 \ D(a−1 ) ⊂ D(a). Later we will
adopt a convention to avoid this kind of ambiguity.
Motivation: why ping-pong actions? The classical ping-pong lemma implies that ping-pong
actions are always faithful, and a little more work shows that the action is determined up to
semi-conjugacy by a finite amount of combinatorial data coming from the cyclic ordering and the
images of the connected components of the sets D(a) (see Definition 2.4 and Lemma 3.4 below, or
[23, Thm. 4.7]). In particular, one can think of ping-pong actions as the family of “simplest possible”
faithful actions of Fn on S1 , and it is very easy to produce a diverse array of examples. Perhaps
the best-known examples are the actions of discrete, free subgroups of PSL(2, R) on RP1 . For these
actions, one can choose domains D(a) with a single connected component. Figure 2.1 shows an
example of the dynamics of such an action of F2 = ha, bi.
Despite their simplicity, ping-pong actions are quite useful. For instance, in [2] ping-pong
actions were used to construct the first known examples of discrete groups of real-analytic circle
diffeomorphisms acting minimally, but not conjugate to a subgroup of a finite central extension of
PSL(2, R). This was a by-product of a series of papers concerning longstanding open conjectures of
Hector, Ghys and Sullivan on the relationship between minimality and ergodicity of a codimensionone foliation (see for instance [12, 13, 16]). In general, it is quite tractable to study the dynamic and
ergodic properties of a ping-pong action (or a Markov system), and this program has been carried
out by many authors [6, 7, 18, 20, 22].
3
2.1
Basic properties
Lemma 2.2. Given a ping-pong
action of (G, S), there exists a choice of ping-pong domains D(a)
1
such that ρ(a) S \ D(a−1 ) = D(a) holds for all a ∈ S ∪ S −1 .
Proof. Let ρ be a ping-pong action with sets D(a) given. We will modify these domains to satisfy
the requirements of the lemma. For each generator a ∈ S (recall this
is the free,not the symmetric,
0
generating set), we shrink the domain D(a), setting D (a) := ρ(a) S1 \ D(a−1 ) . Applying a−1 to
both sides of the above expression gives ρ(a−1 )(D0 (a)) = S1 \ D(a−1 ). Moreover, since the connected
components of D(a)have disjoint
closures, and the same holds for D(a−1 ), hence
also for D0 (a),
we also have ρ(a−1 ) D0 (a) = S1 \ D(a−1 ); or equivalently ρ(a−1 ) S1 \ D0 (a) = D(a−1 ). This is
what we needed to show.
Convention 2.3. From now on, we assume all choices of domains D(a) for every ping-pong action
are as in Lemma 2.2. In particular, this means that, for each a ∈ S, the sets of connected components
π0 (D(a)) and π0 (D(a−1 )) have the same cardinality, and ρ(a) induces a bijection between the
connected components of S1 \ D(a−1 ) and connected components of D(a).
Definition 2.4. Let ρ be a ping-pong action of (G, S). The ping-pong configuration of ρ is the data
consisting of
1. the cyclic order of the connected components of
S
a∈S∪S −1
D(a) in S1 , and
2. for each a ∈ S ∪ S −1 , the assignment of connected components
λa : π0
[
D(b) → π0 (D(a))
b∈S∪S −1 \{a−1 }
induced by the action.
Note that not every abstract assignment λa as in the definition above can be realized by an
action F2 → Homeo+ (S1 ). The following construction gives one way to produce some large families
of examples.
Example 2.5 (An easy construction of ping-pong actions). For a ∈ S, let Xa and Ya ⊂ S1 be disjoint
sets each of cardinality k(a), for some integer k(a) ≥ 1, such that every two points of Xa are
separated by exactly one point of Ya . Choose these so that all the sets Xa ∪ Ya are pairwise disjoint
as a ranges over S. Let D(a) and D(a−1 ) be neighborhoods of Xa and Ya , respectively, chosen small
enough so that all these sets remain pairwise disjoint. Now one can easily construct a piecewise
linear homeomorphism (or even a smooth diffeomorphism) ρ(a) with Xa as its set of attracting
periodic points, and Ya as the set of repelling periodic points such that ρ(a) S1 \ D(a−1 ) = D(a).
The assignments λa are now dictated by the period of ρ(a) and the cyclic order of the sets Xa and
Ya .
While the reader should keep the construction above in mind as a source of examples, we will
show in Example 3.10 that not every ping-pong configuration can be obtained in this manner.
However, the regularity (PL or smooth) in the construction is attainable in general. The following
construction gives one possibility for a PL realization that will be useful later in the text. We leave
the modifications for the smooth case as an easy exercise.
4
Lemma 2.6. Given a ping-pong action ρ0 of (G, S) with domains {D0 (a)}a∈S∪S −1 following Convention 2.3, one can find another ping-pong action ρ of (G, S) with domains {D(a)}a∈S∪S −1 such
that:
1. the action ρ is piecewise linear,
2. there exists µ > 1 such that for any a ∈ S ∪ S −1 , one has ρ(a)0 |D(a−1 ) ≥ µ, and
3. the actions ρ0 and ρ have the same ping-pong configuration.
Proof. Let ρ0 be as in the statement of the Lemma. For each a ∈ S ∪ S −1 , replace the original
domains by smaller domains D(a) ⊂ D0 (a), chosen small enough so that the largest connected
component of D(a) is at most half the length of the smallest connected component of S1 \ D(a−1 ).
We require also that D(a) has exactly one connected component in each connected component of
D0 (a). Now define ρ(a) as a piecewise linear homeomorphism that maps connected components of
S1 \ D(a−1 ) onto connected components of D(a) linearly following the assignment λa .
The next definition and proposition give a further means of encoding the combinatorial data of a
ping-pong action. This will be used later in the proof of Theorem 1.3.
Definition 2.7. Let ρ be a ping-pong action of (G, S) with domains D(a). For each a ∈ S, we
define an oriented bipartite graph Γa with vertex set equal to π0 (D(a)) ∪ π0 (D(a−1 )), and edges
defined as follows:
• For I + ∈ π0 (D(a)), let J + denote the connected component of S1 \ D(a) adjacent to I +
on the right. Put an oriented edge from I + to an interval I − ∈ π0 (D(a−1 )) if and only if
ρ(a−1 )(J + ) = I − .
• Similarly, for I − ∈ π0 (D(a)), with J − the adjacent interval of S1 \ D(a−1 ) on the right, put
an oriented edge from I − ∈ π0 (D(a−1 )) to I + ∈ π0 (D(a)) if and only if ρ(a)(J − ) = I + .
Proposition 2.8. Let ρ : G → Homeo+ (S1 ) be a ping-pong action of a free group (G, S). Then, for
each generator a ∈ S, there exists k(a) ∈ N such that the graph Γa is an oriented 2k(a)-cycle.
Proof. First, the construction of the graph ensures that it is bipartite, and that each vertex has at
most one outgoing edge. As a consequence of Convention 2.3, for any s ∈ S ∪ S −1 and connected
component J of S1 \ D(s−1 ), there exists I ∈ π0 (D(s)) such that ρ(s)(J) = I, so each vertex
does indeed have an outgoing edge. Moreover, if J 0 6= J is a different connected component, then
ρ(s)(J) ∩ ρ(s)(J 0 ) = ∅, so each vertex I has a unique incoming edge. This shows that Γa is a union
of disjoint cycles, and it remains only to prove that the graph is connected.
To show connectivity, let I − be a connected component of D(a−1 ) and consider the connected
component J + of S1 \ D(a) such that ρ(a−1 )(J + ) = I − . Let I1+ and I2+ be the connected components
of D(a) (possibly the same) which are adjacent to J + on either side. By definition of the graph Γa ,
the intervals I1+ , I − , I2+ are consecutive vertices in the same cycle of the graph. And vice versa: if
three intervals I1+ , I − , I2+ are consecutive vertices, then J + := ρ(a)(I − ) is the connected component
of S1 \ D(a) adjacent to both I1+ and I2+ .
This proves that if I1+ and I2+ are consecutive connected components of D(a) in S1 , then they
belong to the same cycle in Γa . Hence we easily deduce that all connected components of D(a)
are in the same cycle in Γa . The same also holds for the components of D(a−1 ), and the graph is
connected.
5
3
Left-invariant circular orders
We begin by quickly recalling standard definitions and properties. A reader familiar with circular
orders may skip to Section 3.1.
Definition 3.1. Let G be a group. A left-invariant circular order is a function c : G × G × G →
{0, ±1} such that
1. c is homogeneous: c(γg0 , γg1 , γg2 ) = c(g0 , g1 , g2 ) for any γ, g0 , g1 , g2 ∈ G;
2. c is a 2-cocycle on G:
c(g1 , g2 , g3 ) − c(g0 , g2 , g3 ) + c(g0 , g1 , g3 ) − c(g0 , g1 , g2 ) = 0 for any g0 , g1 , g2 , g3 ∈ G;
3. c is non-degenerate: c(g0 , g1 , g2 ) = 0 if and only if gi = gj for some i 6= j.
The space of all left-invariant circular orders on G, denoted CO(G), is the set of all such functions,
endowed with the subset topology from {0, ±1}G×G×G (with the natural product topology).
Although spaces of left-invariant linear orders have been well-studied, there are very few cases
where the topology of CO(G) is completely understood. Other than a few sporadic examples, the only
complete description of spaces of circular orders known to the authors comes from [9], which gives a
classification of all groups such that CO(G) is finite, and also a proof that CO(A) is homeomorphic
to a Cantor set for any Abelian group A.
Given that left-orders on free groups are well understood, a natural next case of circular orders
to study is CO(Fn ). Our main tool for this purpose is the following classical relationship between
circular orders and actions on S1 (see [5, 21]).
Proposition 3.2. Given a left-invariant circular order c on a countable group G, there is an action
ρc : G → Homeo+ (S1 ) such that c(g0 , g1 , g2 ) = ord (ρc (g0 )(x), ρc (g1 )(x), ρc (g2 )(x)) for some x ∈ S1 ,
where ord denotes cyclic orientation.
Moreover, there is a canonical procedure for producing ρc which gives a well-defined conjugacy
class of action. This conjugacy class is called the dynamical realization of c with basepoint x.
A description of this procedure is given in [21], modeled on the analogous linear case (see
e.g. [17]). Note that modifying a dynamical realization by blowing up the orbit of some point
y ∈
/ ρ(G)(x) may result in a non-conjugate action that still satisfies the property c(g0 , g1 , g2 ) =
ord (ρc (g0 )(x), ρc (g1 )(x), ρc (g2 )(x)). However, this non-conjugate action cannot be obtained through
the canonical procedure.
Remark 3.3. The converse to the above proposition is also true: if G is a countable subgroup of
Homeo+ (S1 ), then G admits a circular order. A proof is given in [5, Thm. 2.2.14]2 . As a special
case, if ρ : G → Homeo+ (S1 ) is such that some point x has trivial stabilizer, then we may define an
induced order on G by
c(g1 , g2 , g3 ) := ord(ρ(g1 )(x), ρ(g2 )(x), ρ(g3 )(x)).
2
In [4, Prop. 2.4] the authors propose an alternative way of inducing an ordering of G, different from that in [5].
However their method is incorrect, as the following example shows: suppose to have three distinct homeomorphisms
f, g, h, with f coinciding with g on one half circle and with h on the other half. Then for any point x ∈ S1 , there are
always two equal points in the triple (f (x), g(x), h(x)).
6
While one cannot expect in general to find a point with trivial stabilizer, this does hold for
ping-pong actions by the following lemma.
Lemma 3.4. Suppose that ρ is a ping-pong action of (G, S) with domains D(a). If x0 ∈ S1 \
a∈S∪S −1 D(a), then the orbitSof x0 is free and its cyclic order is completely determined by the cyclic
order of the elements of {π0 ( a∈S∪S −1 D(a)) , {x0 }} and the assignments λa .
S
The proof is obtained by a careful reading of the standard proof of the classical ping-pong lemma.
Details are given in [21, Lemma 4.2].
3.1
Isolated circular orders on free groups
In this section we will use ping-pong actions to prove Theorem 1.3 from the introduction. As
this builds on the framework of [21], we start by introducing two results obtained there.
Let G be any group, and ρ : G → Homeo+ (S1 ). Recall that, if ρ(G) does not have a finite
orbit, then there is a unique closed, ρ(G)-invariant set contained in the closure of every orbit, called
the minimal set of ρ(G). We denote this set by Λ(ρ). If Λ(ρ) = S1 , the action is called minimal.
Otherwise, Λ(ρ) is homeomorphic to a Cantor set and ρ permutes the connected components of
S1 \ Λ(ρ). While, for many examples of actions, the permutation will have many disjoint cycles, the
next lemma states that this is not the case for dynamical realizations.
Lemma 3.5 ([21] Lemma 3.21 and Cor. 3.24). Let ρ : G → Homeo+ (S1 ) be a dynamical realization of
a circular order c. Suppose that ρ has a minimal invariant Cantor set Λ(ρ). Then ρ acts transitively
on the set of connected components of S1 \ Λ(ρ).
Since a ping-pong action of a free group of rank at least 2 cannot have finite orbits, invariance
S
of the minimal set immediately implies that Λ(ρ) ⊂ a∈S∪S −1 D(a). If additionally, for each
s 6= t ∈ S ∪ S −1 , one has D(s) ∩ D(t) = ∅, then invariance of Λ(ρ) and the definition of ping-pong
S
implies that in fact Λ(ρ) ⊂ a∈S∪S −1 D(a). Going forward, it will be convenient to have the this
stronger condition, which is given by following lemma.
Lemma 3.6. Let ρ0 be a ping-pong action of (G, S) with domains D0 (a) for a ∈ S ∪ S −1 . Then
there exists an action ρ with the same ping-pong configuration as ρ0 and with domains D(a) satisfying
D(s) ∩ D(t) = ∅ whenever s 6= t.
Proof. Let ρ0 be a ping-pong action. There are only finitely many points x contained in sets of the
form D0 (s) ∩ D0 (t) for t 6= s ∈ S ∪ S −1 . For each such point x, blow up its orbit, replacing each
point y ∈ ρ0 (G)(x) with an interval Iy ; if lengths of the Iy are chosen so their sum converges, then
we obtain a new circle, say Ŝ1 , with a natural continuous, degree one map h : Ŝ1 → S1 given by
collapsing each Iy to the point y. For each s ∈ S ∪ S −1 , let D1 (s) ⊂ Ŝ1 be the preimage of D0 (S)
under h.
Since x has trivial stabilizer and G is free, we may extend the action of G to this new circle by
allowing a ∈ S to act as any orientation-preserving map from Iy to Ia(y) . We now show that we may
choose maps in such a way as to achieve a ping-pong action with the desired properties.
For each inserted interval I = [p, q] that is adjacent to a set of the form D1 (s) on the left and
D1 (t) on the right (where s, t ∈ S ∪ S −1 ), fix points ps < pt in the interior of I and extend D1 (s)
into I to include [p, ps ) and D1 (t) to include (pt , q]. Having done this on each such interval, let
D(s) denote the new extended domains, and note that these have disjoint closures. Now for a ∈ S,
define the action of a on such an interval I = Iy as follows. If ρ0 (a)(y) ∈ D0 (a), the restriction
of ρ to I may be any orientation-preserving homomorphism between Iy and Ia(y) . Otherwise, Iy
7
is adjacent to D1 (a−1 ) either on the right or the left, and we define ρ(a) on Iy to map the chosen
point pa−1 ∈ Iy to the point pa in Ia(y) . This ensures that ρ(a)(S 1 \ D(a−1 )) = D(a), so that these
are indeed ping-pong domains for the action. Finally, note that by construction, the ping-pong
configuration has not changed.
Convention 3.7. In a ping-pong action of (G, S), we assume from now on that the domains D(s)
satisfy D(s) ∩ D(t) = ∅ whenever s 6= t ∈ S ∪ S −1 .
It follows easily from invariance of Λ(ρ) and the definition of ping-pong that, for actions as in
S
Convention 3.7, we have the inclusion Λ(ρ) ⊆ a∈S∪S −1 D(a). The following theorem from [21]
relates circular orders and ping-pong actions.
Theorem 3.8 ([21] Thm. 1.5). Let G = Fn be a free group. A circular order c ∈ CO(G) is isolated
if and only if its dynamical realization ρc : G → Homeo+ (S1 ) is a ping-pong action satisfying
Convention 3.7.
With these tools, we proceed to the main goal of this section.
Proof of Theorem 1.3. The case where n is even is covered in [21]. As explained there, the representation of G into PSL(2, R) coming from a hyperbolic structure on a genus n/2 surface with one
boundary component gives an isolated circular order. (In fact, by taking lifts to cyclic covers, one
can obtain infinitely many isolated circular orders in distinct equivalence classes under the action of
Aut(Fn ) on CO(G).)
To show that Fn does not admit an isolated circular order when n is odd, we need more work.
We begin with some generalities, applicable to free groups of any rank (even or odd). Suppose
that ρ : Fn → Homeo+ (S1 ) is a dynamical realization of an isolated circular order, and fix a free
generating set S = {a1 , . . . , an } for Fn . By Theorem 3.8 and Lemma 3.5, ρ is a ping-pong action
with domains satisfying Convention 3.7 and the connected components of S1 \ Λ(ρ) form a unique
orbit. Let c0 , . . . , cr be the (finitely many) connected components of S1 \ Λ(ρ) that are not contained
in any domain D(s).
Suppose that ci has endpoints in D(s) and D(t), for some s 6= t. Then, for any generator
u∈
/ {s−1 , t−1 }, we have that ρ(u)(ci ) ∈ D(u). In addition, we have that ρ(s−1 )(ci ) and ρ(t−1 )(ci )
belong to {c0 , . . . , cr }: indeed, the intersection ci ∩D(s) is nonempty; its image by ρ(s−1 ) is contained
in S1 \ (Λ ∪ D(s−1 )) and is adjacent to D(s−1 ) because of Convention 2.3; moreover, it has to intersect
some cj because of Convention 3.7. Then we must have ρ(s−1 )(ci ) = cj . The same holds for t−1 . This
implies that ci and cj are in the same orbit if and only if they are equivalent under the equivalence
relation ∼ on {c0 , . . . , cr } generated by
ci ∼ cj
if there exists t ∈ S ∪ S −1 such that ci = ρ(t)(cj ) and ci ∩ D(t) 6= ∅
We will now argue that the number of equivalence classes under this relation can be 1 only if n is
even. This is done by using the combinatorial data of the graphs from Definition 2.7 to build a
surface with boundary using the disc, and then making an Euler characteristic argument.
For each generator a ∈ S, let k(a) be the integer given by Proposition 2.8. Let Pa be a 4k(a)-gon
(topologically a disc) with cyclically ordered vertices v1 , v2 , . . . , v4k(a) . Choose a connected component
I = [x1 , y1 ] of D(a) and glue the oriented edge v1 v2 to I so as to agree with the orientation of
I ⊂ S1 . Then glue the edge v3 v4 to the connected component of D(a−1 ) containing ρ(a−1 )(x1 ),
according to the orientation in S1 . Let y2 denote the other endpoint of this connected component,
and glue v5 v6 to the connected component of D(a) containing ρ(a)(y2 ). Iterate this process until all
edges v2j−1 v2j have been glued to S1 = ∂D. Our convention to follow the orientation of S1 implies
8
that the resulting surface with boundary is orientable. Note that the remaining (unglued) edges of
Pa correspond exactly to the edges of the graph Γa from Definition 2.7; precisely, collapsing each
connected component of D(a) and of D(a−1 ) to a point representing a vertex recovers the cycle Γa .
Now repeat this procedure for each generator in S, to obtain an orientable surface with boundary,
which we will denote by Σ. A cartoon of the result of this procedure for the ping-pong action of
Example 3.10 is shown in Figure 3.1, and may be helpful to the reader.
We claim that the number of boundary components of the surface Σ is exactly the number
of equivalence classes of the relation ∼. To see this, we proceed as follows. By construction, the
connected components of ∂Σ ∩ {c0 , . . . , cr } are exactly the intervals ci . If some interval ci has
endpoints in D(s) and D(t), then ∂Σ ∩ ci is joined to ρ(t−1 )ci ∩ ∂Σ and ρ(s−1 )ci ∩ ∂Σ by edges of
Ps and Pt respectively. Thus, ci ∼ cj implies that ci and cj lie in the same boundary component of
Σ, and the intersection of that boundary component with {c0 , . . . , cr } defines an equivalence class.
This proves the claim.
We now compute the Euler characteristic of Σ and conclude the proof. Proposition 2.8 implies
that the gluing of Pa described in our procedure adds one face and 2k(a) edges to the existing
surface. Therefore after all the polygons Pa (as a ranges over elements of S) have been glued, the
surface Σ obtained has χ(Σ) ≡ n + 1 mod 2. Since Σ is orientable, χ(Σ) agrees mod 2 with the
number of boundary components of Σ, which, by our claim proved above, agrees with the number of
equivalence classes of ci . As discussed above, if ρ is the dynamical realization of an isolated order
then this number is equal to 1, hence n + 1 ≡ 1 mod 2, and n must be even.
The proof above can be improved to give a statement about general ping-pong actions:
Theorem 3.9. Let G = Fn be the free group of rank n with free generating set S. Consider a
ping-pong action ρ of (G, S) satisfying Conventions 2.3 and 3.7. Let Λ(ρ) be the minimal invariant
Cantor set for the action. Then the number of orbits of connected components of the complement
S1 \ Λ(ρ) is congruent to n + 1 mod 2.
Proof. As in the previous proof, let c0 , c1 , . . . , cr be the connected components of S1 \ Λ(ρ) that
are not contained in any domain D(s), and recall that G permutes the connected components of
S1 \ Λ(ρ). We claim that each cycle of this permutation contains at least one of the ci . Given this
claim, we may construct an orientable surface Σ as in the proof of Theorem 1.3, whose boundary
components count the number of cycles. Computing Euler characteristic as above shows that the
number of cycles is congruent to n + 1 mod 2.
We now prove the claim. Suppose that I is a connected component of S1 \ Λ(ρ) contained in
some D(s). By Lemma 2.6, we can take ρ(s) to be piecewise linear, and such that each ρ(s−1 )
expands D(s) uniformly, increasing the length of each connected component by a factor of some
µ > 1, independent of s. Iteratively, assuming that ρ(sk sk−1 · · · s1 )(I) ⊂ D(s−1
k+1 ), then the length
k+1
of ρ(sk+1 sk · · · s1 )(I) is at least µ
length(I). This process cannot continue indefinitely, so some
image of I is not contained in a ping-pong domain.
3.2
Exotic examples
To indicate some of the potential difficulty of the problem of classifying all isolated orders on Fn ,
we give an example of a ping-pong configuration for F2 that, even after applying an automorphism
of F2 , cannot arise from the construction in Example 2.5.
Example 3.10. Let F2 = ha, bi and consider a ping-pong action where ρ(b) is as defined by the graph
in Figure 3.2, and ρ(a) is a hyperbolic element of SL(2, R) chosen so that the connected components
9
Pb
Pa
D
Figure 3.1: The surface associated with the exotic example (left), and its boundary component
(right).
ρ(b) S1 \ D(b−1 )
D(b−1 )
D(b)
D(b−1 )
ρ(b) S1 \ D(b−1 )
D(b)
D(b)
−1
D(b
)
D(b)
D(b−1 )
D(b)
D(b)
D(b−1 )
D(b−1 )
Figure 3.2: The ping-pong domains for ρ(b) (left) and its graph (right). The circle is oriented
counterclockwise.
10
of the domains for the ping-pong action are in cyclic order as follows:
D(b−1 ), D(a−1 ), D(b−1 ), D(a), D(b), D(a−1 ), D(b), D(a)
(we are abusing notation slightly here, using each appearance of D(s) to stand for a connected
component of D(s)). See Figure 3.1 (left) for an illustration of the domains and the surface Σ
constructed in the proof of Theorem 1.3.
Since ρ(b) has two hyperbolic fixed points, and ρ(a) has four, this example is not realized by a
ping-pong action in PSL(2, R), nor in any finite extension of it. In fact, the ping-pong configuration
for ρ(b) alone is atypical, in the sense that it is not the classical ping-pong configuration for a
hyperbolic element in PSL(2, R) – ρ(b) has a “slow” contraction on the left half of the circle, as two
iterations of ρ(b) are needed in order to bring the external gaps of D(b) ∪ D(b−1 ) into the component
of D(b) with the attracting fixed point.
However, the surface Σ from this construction has one boundary component, as shown in
Figure 3.1 (right), so it corresponds to an isolated circular order in CO(F2 ).
Observe that one can create several examples of this kind, by choosing ρ(b) to have two hyperbolic
fixed points, but with an arbitrarily slow contraction (i.e. with N connected components for D(b),
N ∈ N arbitrary) and then choosing ρ(a) to be a N -fold lift of a hyperbolic element in PSL(2, R).
4
Left-invariant linear orders on Fn × Z
The purpose of this section is to prove Theorem 1.1, stating that Fn × Z admits an isolated
linear order if and only if n is even.
4.1
Preliminaries on linear orders
Before proceeding to the proof of Theorem 1.1, we recall some standard tools. As for circular orders, linear orders on countable groups have a dynamical realization (see for instance [14, Prop. 1.1.8]).
One quick way of seeing this given what we have already described, is by thinking of a linear order
as a special case of a circular order. Indeed, given a linear order on a group G, one defines the
cocycle c by setting, for distinct g1 , g2 , g3 ∈ G
c (g1 , g2 , g3 ) = sign(σ),
where σ is the permutation of the indices such that gσ(1) ≺ gσ(2) ≺ gσ(3) . Thus, the construction
of the dynamical realization sketched in the proof of Proposition 3.2 may be performed also for a
linear order. The result is an action on the circle with a single one global fixed point, which one can
view as an action on the line with no global fixed point. Conversely, a faithful action on the real line
ρ : G → Homeo+ (R) can be viewed as a faithful action on the circle with a single fixed point, and
the circular orders produced as in Remark 3.3 will be linear orders on G.
Next, we recall the notion of convex subgroups, their dynamical interpretation, and their
relationship to isolated orders.
Definition 4.1. A subgroup C in a linearly-ordered group (G, ) is convex if for any two elements
h, k ∈ C, and for any g ∈ G, the condition h g k implies g ∈ C.
Lemma 4.2 (see [14], Prop. 2.1.3). Let G be a countable left-ordered group and consider a dynamical
realization ρ of (G, ) with basepoint x such that C is a convex subgroup. Let I be the interval
bounded by inf h∈C ρ(h)(x) and suph∈C ρ(h)(x). Then I has the following property:
for any g ∈ G, either ρ(g)(I) = I, or ρ(g)(I) ∩ I = ∅.
11
(4.1)
Moreover, the stabilizer of I is precisely C.
Conversely, given a faithful action on the real line ρ : G → Homeo+ (R), if an interval I has the
property (4.1), then the stabilizer C = StabG (I) is convex in any induced order with basepoint x ∈ I.
It is easy to see that the family of convex subgroups of a linearly ordered group (G, ) forms a
chain: if C1 , C2 are two convex subgroups of (G, ), then either C1 ⊂ C2 or C2 ⊂ C1 . Moreover,
for any convex subgroup C ⊂ G, the group G acts on the ordered coset space (G/C, C ) by orderpreserving transformation. (The induced order on the coset space is given by f C <C gC if and only
if f c < gc0 for every c, c0 ∈ C, which makes sense because C is convex.) In particular, this implies
that if C is convex in (G, ), then any linear order C on C may be extended to a (new) order 0
on G by declaring
(
C C gC if g ∈
/ C,
id 0 g ⇔
0
id g
if g ∈ C.
Elaborating on this, one can show the following lemma (see [14, Prop. 3.2.53] or [24, Thm. 2] for
details).
Lemma 4.3. If (G, ) has an infinite chain of convex subgroups, then is non-isolated in LO(G).
Let us also introduce a dynamical property that implies that an order is non-isolated. Recall
that two representations ρ1 and ρ2 : G → Homeo+ (R) are semi-conjugate if there is a proper,
non-decreasing map f : R → R such that f ◦ ρ1 = ρ2 ◦ f .
Definition 4.4. Let G be a discrete group. Let Rep(G, Homeo+ (R)) denote the space of representations (homomorphisms) G → Homeo+ (R), endowed with the compact-open topology; let
Rep# (G, Homeo+ (R)) be the subspace of representations with no global fixed points.
A representation ρ ∈ Rep# (G, Homeo+ (R)) is said to be flexible if every open neighborhood of ρ
in Rep# (G, Homeo+ (R)) contains a representation that is not semi-conjugate to ρ.
The following lemma is implicit in work of Navas [26] as well as in [28]. An explicit proof can be
found in [1, Prop. 2.8].
Lemma 4.5. Let G be a discrete, countable group and let ρ0 be the dynamical realization of an
order with basepoint x ∈ R. If ρ0 is flexible, then is non-isolated in LO(G).
Remark 4.6. Though not needed for our work here, we note that a precise characterization of isolated
circular and linear orders in terms of a strong form of rigidity (i.e. strong non-flexibility) of their
dynamical realizations is given in [21].
As mentioned in the introduction, in order to prove Theorem 1.1 we use the relationship between
circular orders on groups and linear orders on their central extensions by Z. For this purpose, we
need the notion of cofinal elements.
Definition 4.7. An element h in a linearly-ordered group (G, ) is called cofinal if
for all g ∈ G, there exist m, n ∈ Z such that hm g hn .
(4.2)
Remark 4.8. Cofinal elements also have a characterization in terms of the dynamical realization: If
ρ is a dynamical realization of with basepoint x, then h ∈ G is cofinal if and only if ρ(h) has no
fixed point. Indeed, if h is not cofinal, then the point inf{ρ(g)(x) | hn g for every n ∈ Z} is fixed
by ρ(h). Conversely, if h satisfies (4.2), then the orbit of x under ρ(h) is clearly unbounded on both
sides.
12
ρ(g)
ρ(f )
u
w
v
Figure 4.1: A crossing in the dynamical realization.
Given a group G with a circular order c, there is a natural procedure to lift c to a linear order c
on a central extension of G by Z such that any generator of the central Z subgroup is cofinal for c
[21, 30]. The following statement appears as Proposition 5.4 in [21].
Proposition 4.9. Assume that G is finitely generated and c is an isolated circular order on G. If
b of G by Z, then the induced linear order c is isolated in
c is the lift of c to a central extension G
b
LO(G).
Finally, we recall the definition of crossings.
Definition 4.10. Let G be a group acting on a totally ordered space (Ω, ≤). The action has
crossings if there exist f, g ∈ G and u, v, w ∈ Ω such that:
1. u < w < v.
2. g n u < v and f n v > u for every n ∈ N, and
3. there exist M, N in N such that f N v < w < g M u.
In this case, we say that f and g are crossed.
If f and g are crossed, then the graph of ρ(f ) and ρ(g) in the dynamical realization is locally
given by the picture in Figure 4.1.
Our application of the notion of crossings will be through the following lemma.
Lemma 4.11 ([14] Cor. 3.2.28). Let C be a convex subgroup of (G, ) and suppose that the (natural)
action of G on (G/C, C ) has no crossings. Then there exists a homomorphism τ : G → R with
C in its kernel. Moreover, if C is the maximal convex subgroup of (G, ), then C agrees with the
kernel of τ .
4.2
Isolated linear orders on Fn × Z
We now turn to our main goal of describing isolated linear orders on Fn × Z and proving
Theorem 1.1. We begin by reducing the proof to the statement of Proposition 4.12 below.
13
Since every central extension of Fn by Z splits, Proposition 4.9 tell us that F2n × Z admits
isolated linear orders – more precisely, any lift of an isolated order on F2n to F2n × Z will be isolated.
Furthermore, if
π
0 → Z → Ĝ → G → 1
is a central extension of G by Z then any linear order on Ĝ in which Z is cofinal gives a canonical
circular order on G as follows. Let z be the generator of Z such that z id. Since z is cofinal, for
each g ∈ G, there exists a unique representative ĝ ∈ π −1 (g) such that id ĝ ≺ z. Given distinct
elements g1 , g2 , g3 ∈ G, let σ be the permutation such that
id ĝσ(1) ≺ ĝσ(2) ≺ ĝσ(3) ≺ z.
Define π ∗ ()(g1 , g2 , g3 ) := sign(σ). One checks that this is a well defined circular order on G. In the
proof of Proposition 5.4 of [21], it is shown that π ∗ is continuous and is locally injective when G
is finitely generated, which implies that an isolated linear order of F2n+1 × Z with cofinal center
induces an isolated circular order of F2n+1 by this procedure. Since F2n+1 has no isolated circular
orders, by Theorem 1.3, to finish the proof of Theorem 1.1 it is enough to show the following:
Proposition 4.12. Let F be a free group, and a linear order on G = F × Z in which the central
factor is not cofinal. Then is non-isolated.
As a warm-up, as well as tool to be used in the proof, we start with a short proof of a special
case.
Lemma 4.13. Let F be a free group of infinite rank and G = F × Z. Then no order in LO(G) is
isolated.
Proof. Let f1 , f2 , . . . be a set of free generators of the free factor F and g the generator of the central
factor Z. Let be any order on G and ρ0 a dynamical realization with basepoint x. For any fixed
n ∈ N, we can define a representation ρn : G → Homeo+ (R) by setting
(
ρn (g) = ρ0 (g),
ρn (fk ) =
ρ0 (fk )
ρ0 (fn )−1
if k 6= n,
if k = n.
It is easy to see that the orbit of x is free for all the actions ρn , and that no two distinct representations
ρn and ρm are semi-conjugate one to another. Thus, they determine distinct orders n ; and these
orders converge to in LO(G) as n → ∞.
Proof of Proposition 4.12. We have already eliminated the case where F has infinite rank. If F has
rank one, then F × Z is abelian, and so admits no isolated orders (see [29]). So from now on we
assume that the rank of F is finite and at least 2.
Looking for a contradiction, suppose that is a linear order on G = F × Z which is isolated,
and in which the center is not cofinal. Let ρ be its dynamical realization, and let z be a generator of
the central Z subgroup. By Remark 4.8, ρ(z) acts with fixed points. Moreover since Z is central, the
set of fixed points of ρ(z) is ρ(G)-invariant. Since ρ(G) has no global fixed point, this implies that
ρ(z) has fixed points in every neighborhood of +∞ and of −∞.
We now find a convex subgroup in which z is cofinal. Let I denote the connected component
of R \ Fix(ρ(z)) that contains the basepoint x0 of ρ, so in particular I is a bounded interval. Let
C = StabG (I).
Claim 1. C is a convex subgroup of (G, ) and z is cofinal in (C, ).
14
Proof of Claim. If h, k ∈ C and g ∈ G satisfy ρ(h)(x0 ) < ρ(g)(x0 ) < ρ(k)(x0 ), then for any n we
also have ρ(z n h)(x0 ) < ρ(z n g)(x0 ) < ρ(z n k)(x0 ). Since z is central, this implies
ρ(hz n )(x0 ) < ρ(gz n )(x0 ) < ρ(kz n )(x0 ).
(4.3)
Up to replacing z with z −1 , without loss of generality we may assume that ρ(z)(x0 ) > x0 . Thus, as
n → ∞, the sequence of points ρ(z n )(x0 ) converges to the rightmost point of I, which is fixed by
both ρ(h) and ρ(k). We deduce from (4.3) that ρ(g) also fixes this point. Similarly, considering the
limit in (4.3) as n → −∞ shows that the leftmost point of I is fixed by ρ(g). Hence g ∈ C, which
shows C is convex. Finally, by Remark 4.8, the fact that ρ(z) has no fixed points in I implies that z
is cofinal in (C, ).
Since (G, ) is isolated and C is convex, the restriction of to C is also an isolated order on C.
Additionally, the fact that is isolated implies, by Lemma 4.3, that the chain of convex subgroups
of G is finite. Let G0 denote the smallest convex subgroup properly containing C. Since Z ⊂ C,
we have that G0 is also a direct product of Z and a free group (a subgroup of F ), and again, our
assumptions on imply that (the restriction of) is an isolated left-order on G0 in which Z is not
cofinal. Thus, we may work from now on with G0 instead of G. Equivalently – and, for notational
convenience, this is how we will proceed – we may assume that C is the maximal convex subgroup
of G.
For our next claim observe that this maximal convex subgroup C also admits a decomposition of
the form F ∗ × Z, where F ∗ is a subgroup of F .
Claim 2. F ∗ is a non trivial free group of even rank.
Proof of Claim. Since the restriction of to C = F ∗ × Z is isolated, as before, Lemma 4.13 implies
that F ∗ cannot have infinite rank. If F ∗ were trivial, then the action of G would be semi-conjugate
to an action of F , thus making very easy to perturb the action of F × Z and thus the order (recall
that free groups have no isolated orders [25]). Thus, F ∗ is a nontrivial free group of finite rank,
and as z is cofinal in the ordering in C, its rank must be even (c.f. the remarks at the beginning of
Section 4.2).
Claim 3. C has infinite index in G.
Proof of Claim. If C had finite index, then the G-orbit of the interval I would be bounded. This
would imply that the dynamical realization has a global fixed point, which is absurd.
Since every nontrivial normal, infinite index subgroup of F has infinite rank, we conclude from
Claims 2 and 3 that F ∗ (and thus C) is not a normal subgroup of G. Lemma 4.11 thus implies that
the action of G on (G/C, C ) has crossings, as otherwise C would be normal. In particular, if we
collapse I and its G-orbit, we obtain a semi-conjugate action ρ̄ : G → Homeo+ (R) which is minimal
and has crossings. Using this observation, we now prove the following claim.
Claim 4. For any compact set K ⊂ R, there exists ρ0 agreeing with ρ on K, but not semi-conjugate
to ρ.
Proof of Claim. Fix a compact set K. We will modify the action of F outside K to produce an
action of G that is not semi-conjugate to ρ.
Suppose as an initial case that there is a primitive element (i.e. a generator in some free generating
set) a of F such that ρ(a) has a fixed point p ∈
/ K. Without loss of generality, assume p is to the
right of K, the other case is completely analogous. Since Fix(ρ(a)) is ρ(z)-invariant, and ρ(z n )(p) is
15
bounded and accumulates at a fixed point of ρ(z), we may also assume without loss of generality
that we have chosen p to be a common fixed point of ρ(z) and ρ(a).
We now define a+ and a− ∈ Homeo+ (R) which commute with ρ(z), and have the property that
a+ (x) ≥ x for all x ≥ p, and a− ◦ ρ(a)(x) ≤ x for all x ≥ p. For this, let J be any connected
component of (R\Fix(ρ(z)))∩[p, ∞). Suppose first that J contains a point of Fix(ρ(a)). The fact that
a and z commute means that the endpoints of J are preserved by ρ(a). Then define the restriction
of a+ to J to agree with ρ(z) if ρ(z)(x) > x on J, or with ρ(z −1 ) if ρ(z)(x) < x on J. If J contains
no point of Fix(ρ(a)), then J ⊆ J 0 where J 0 is a connected component of R \ Fix(ρ(a)) ∩ [p, ∞), and
we may define a+ to agree with ρ(a) or ρ(a−1 ) there, so as to satisfy a+ (x) > x for x ∈ J. Lastly, set
a+ (x) = x for any x ∈ Fix(ρ(z)) ∩ [p, ∞). The definition of a− is analogous. Let ρ± be the actions
obtained by replacing the action of ρ(a) by that of a± on [p, ∞) and leaving the other generators
unchanged. Since a+ and a− commute with ρ(z), this defines representations of G, and clearly ρ+
and ρ− are not semi-conjugate.
We are left to deal with the case where no primitive element of F has a fixed point outside K.
In this case, we will perturb the action ρ to obtain a primitive element with a fixed point outside K,
and hence a non-semi-conjugate action. To do this, we use the fact that the semi-conjugate action ρ̄
has crossings and is minimal. Minimality implies that crossings can be found outside any compact
set and, thus, for any compact K ⊂ R there is g ∈ G such that R \ Fix(ρ(g)) has a component outside
(and on the right of) K. Let J = (j0 , j1 ) denote one of those components.
Notice that some primitive element a ∈ F has the property that ρ(a)(j0 ) ∈ J, but ρ(a)(j1 ) ∈
/ J.
For if this was not the case, J would satisfy property (4.1) and as observed in Remark 4.2, G would
then have a convex subgroup properly containing a conjugate of C. Since C was assumed maximal,
this is impossible.
Fix a primitive element a with the property above, and let ḡ be the homeomorphism defined as
the identity outside J and agreeing with ρ(g) on J. Define ρḡ by
ρḡ (a) = ḡρ(a), and ρḡ (b) = ρ(b) for any other generator of F , and ρḡ (z) = ρ(z).
Since ḡ commutes with ρ(z) = ρḡ (z), the new action ρḡ is a representation of G. Moreover, by
changing ḡ by some power if necessary, we have that ρḡ (a) has a fixed point in J. This ends the
proof of Claim 4.
To finish the proof of Proposition 4.12 (and thus that of Theorem 1.1), we note that the flexibility
of ρ from Claim 4 together with the statement of Lemma 4.5 implies that the order is non-isolated,
giving the desired contradiction.
Acknowledgments
The authors thank Yago Antolín for suggesting Corollary 1.2. K.M. was partially supported
by NSF grant DMS-1606254. C.R. was partially supported by FONDECYT 1150691. M.T. was
partially supported by PEPS – Jeunes Chercheur-e-s – 2017 (CNRS), Projet “Jeunes Géométres” of
F. Labourie (financed by the Louis D. Foundation) and the Réseau France-Brésil en Mathématiques.
References
[1] J. Alonso, J. Brum, and C. Rivas, Orderings and flexibility of some subgroups of Homeo+ (R), J. Lond. Math. Soc.
(2) 95 (2017), no. 3, 919–941. MR3664524
16
[2] S. Alvarez, P. Barrientos, D. Filimonov, V. Kleptsyn, D. Malicet, C. Meniño, and M. Triestino, Maskit partitions
and locally discrete groups of real-analytic circle diffeomorphisms. In preparation.
[3] A.K. Arora and S.H. McCleary, Centralizers in free lattice-ordered groups, Houston J. Math. 12 (1986), no. 4,
455–482. MR873641
[4] H. Baik and E. Samperton, Spaces of invariant circular orders of groups, Groups Geom. Dyn. To appear.
[5] D. Calegari, Circular groups, planar groups, and the Euler class, Proceedings of the Casson Fest, 2004, pp. 431–491.
MR2172491
[6] J. Cantwell and L. Conlon, Foliations and subshifts, Tohoku Math. J. (2) 40 (1988), no. 2, 165–187. MR943817
[7]
, Leaves of Markov local minimal sets in foliations of codimension one, Publ. Mat. 33 (1989), no. 3, 461–484.
MR1038484
[8] C. Champetier and V. Guirardel, Limit groups as limits of free groups, Israel J. Math. 146 (2005), 1–75. MR2151593
[9] A. Clay, K. Mann, and C. Rivas, On the number of circular orders on a group, ArXiv e-prints (2017), available at
1704.06242.
[10] P. Dehornoy, Monoids of O-type, subword reversing, and ordered groups, J. Group Theory 17 (2014), no. 3,
465–524. MR3200370
[11] P. Dehornoy, I. Dynnikov, D. Rolfsen, and B. Wiest, Ordering braids, Mathematical Surveys and Monographs,
vol. 148, American Mathematical Society, Providence, RI, 2008. MR2463428
[12] B. Deroin, V.A. Kleptsyn, and A. Navas, On the question of ergodicity for minimal group actions on the circle,
Mosc. Math. J. 9 (2009), no. 2, 263–303. MR2568439
[13]
, On the ergodic theory of free group actions by real-analytic circle diffeomorphisms, Invent. Math. (2014).
To appear.
[14] B. Deroin, A. Navas, and C. Rivas, Groups, orders and dynamics, Arxiv e-print (2016), available at 1408.5805.
[15] T. V. Dubrovina and N. I. Dubrovin, On braid groups, Mat. Sb. 192 (2001), no. 5, 53–64. MR1859702
[16] D.A. Filimonov and V.A. Kleptsyn, Structure of groups of circle diffeomorphisms with the property of fixing
nonexpandable points, Funct. Anal. Appl. 46 (2012), no. 3, 191–209. MR3075039
[17] É. Ghys, Groups acting on the circle, Enseign. Math. (2) 47 (2001), no. 3-4, 329–407. MR1876932
[18] T. Inaba and S. Matsumoto, Resilient leaves in transversely projective foliations, J. Fac. Sci. Univ. Tokyo Sect. IA
Math. 37 (1990), no. 1, 89–101. MR1049020
[19] T. Ito, Dehornoy-like left orderings and isolated left orderings, J. Algebra 374 (2013), 42–58. MR2998793
[20] R. Mañé, Hyperbolicity, sinks and measure in one-dimensional dynamics, Comm. Math. Phys. 100 (1985), no. 4,
495–524. MR806250
[21] K. Mann and C. Rivas, Group orderings, dynamics, and rigidity, Ann. Inst. Fourier (Grenoble) (2016). To appear.
[22] S. Matsumoto, Measure of exceptional minimal sets of codimension one foliations, A fête of topology, 1988,
pp. 81–94. MR928398
[23]
[24]
, Basic partitions and combinations of group actions on the circle: a new approach to a theorem of Kathryn
Mann, Enseign. Math. 62 (2016), no. 1-2, 15–47. MR3605808
, Dynamics of isolated left orders, ArXiv e-prints (2017), available at 1701.05987.
[25] S.H. McCleary, Free lattice-ordered groups represented as o-2 transitive l-permutation groups, Trans. Amer. Math.
Soc. 290 (1985), no. 1, 69–79. MR787955
[26] A. Navas, On the dynamics of (left) orderable groups, Ann. Inst. Fourier (Grenoble) 60 (2010), no. 5, 1685–1740.
MR2766228
[27]
, A remarkable family of left-ordered groups: central extensions of Hecke groups, J. Algebra 328 (2011),
31–42. MR2745552
[28] C. Rivas, Left-orderings on free products of groups, J. Algebra 350 (2012), 318–329. MR2859890
[29] A.S. Sikora, Topology on the spaces of orderings of groups, Bull. London Math. Soc. 36 (2004), no. 4, 519–526.
MR2069015
[30] S.D. Želeva, Cyclically ordered groups, Sibirsk. Mat. Ž. 17 (1976), no. 5, 1046–1051, 1197. MR0422106
17
| 4math.GR
|
Computing the Convolution of Analog and Discrete Time
Exponential Signals Algebraically
Francisco Mota
arXiv:1606.08072v1 [cs.SY] 26 Jun 2016
Departamento de Engenharia de Computação e Automação
Universidade Federal do Rio Grande do Norte – Brasil
e-mail:[email protected]
November 17, 2017
Abstract
We present a procedure for computing the convolution of exponential signals without the need of
solving integrals or summations. The procedure requires the resolution of a system of linear equations
involving Vandermonde matrices. We apply the method to solve ordinary differential/difference equations
with constant coefficients.
1
Notation and Definitions
Below we introduce the definitions and notation to be used along the paper:
• Z, R and C are, respectively, the set of integers, real and complex numbers;
• An analog time signal is defined as a complex valued function f : R → C , and a discrete time signal
t 7→ f (t)
is a complex valued function f : Z → C . In this paper we are mainly concerned with exponential
rt
k 7→ f (k)
k
signals, that is, f (t) = e , or f (k) = r , where r ∈ C. Two basic signals will be necessary in our
development, namely, the unit step signal (σ) and the unit impulse (generalized) signal (δ), both in
analog or discrete time setting. The unit step is defined as
(
(
0, t < 0
0, k < 0
σ(t) =
(analog) and σ(k) =
(discrete time)
1, t > 0
1, k ≥ 0
In analog time context we define the unit impulse as δ = σ̇, where the derivative is supposed to be
defined in the generalized sense, since σ has a “jump” discontinuity at t = 0, and this is why we denote
δ as a “generalized” signal [1]. If f is an analog signal continuous at t = 0, the product “f σ” is given
by
(
0,
t<0
(f σ)(t) = f (t)σ(t) =
f (t), t > 0
and then, if f (0) 6= 0, the module of f σ also has a “jump” discontinuity at t = 0, in fact (f σ)(0− ) = 0
while (f σ)(0+ ) = f (0). Additionally, using the generalized signal δ, we can obtain the derivative of f σ
as
(f˙σ) = f˙σ + f σ̇ = f˙σ + f (0)δ
(1)
In discrete time context, time shifting is a fundamental operation. We denote by [f ]n the shifting of
signal f by n units in time, that is, [f ]n (k) = f (k − n). Using this notation, the discrete time impulse
δ can be written as δ = σ − [σ]1 or δ(k) = σ(k) − σ(k − 1).
• The convolution between two signals f and g, represented by f ∗ g, is the binary operation defined as
[2]:
Z ∞
(f ∗ g)(t) =
f (τ )g(t − τ )dτ, for analog signals, or
−∞
(f ∗ g)(k) =
∞
X
j=−∞
(2)
f (j)g(k − j),
1
for discrete time signals
Additionally if we have f (t) = g(t) = 0 for t < 0 and f (k) = g(k) = 0 for k < 0, we get from (2) that:
k<0
0,
t<0
0,
Z t
k
(f ∗ g)(t) =
(3)
and (f ∗ g)(k) = X
f (j)g(k − j), k ≥ 0
f (τ )g(t − τ )dτ, t > 0
0
j=0
Convolution is commutative, associative and the unity of the operation is the signal δ, that is, f ∗ δ =
δ ∗ f = f for any signal f . Other important properties of convolution are related with derivation and
time shifting:
(f ∗˙ g) = f˙ ∗ g = f ∗ ġ
f ∗ [δ]n = [f ]n
2
Introduction
Convolution between signals is a fundamental operation in the theory of linear time invariant (LTI) systems1
and its importance comes mainly from the fact that a LTI operator H, which represents a LTI system in
analog or discrete time context, satisfies the following property involving signals convolution [3]:
H(u ∗ v) = H(u) ∗ v = u ∗ H(v)
(4)
for any signals u and v, analog or discrete time defined. Since u = u ∗ δ for any signal u, taking in particular
v = δ in (4), we get:
H(u) = u ∗ H(δ) = H(δ) ∗ u, for any signal u.
(5)
Equation (5) above implies that the signal h = H(δ) (denominated impulse response) characterizes the
operator H, or the LTI system, in the sense that the system output due to any input signal u, that is H(u),
is given by the convolution between u itself and the system impulse response h. This is pretty much similar
to the fact that a linear function, e.g. f (x) = ax, is characterized by its value at x = 1, or f (x) = f (1)x.
Maybe the most important class of LTI systems (in analog or discrete time context) are the ones modeled
by a n order ordinary differential/difference equation with constant coefficients, as shown bellow:
y (n) (t) + an−1 y (n−1) (t) + · · · + a2 ÿ(t) + a1 ẏ(t) + a0 y(t)
= u(t),
y(k + n) + an−1 y(k + n − 1) + · · · + a2 y(k + 2) + a1 y(k + 1) + a0 y(k) = u(k),
analog, or
(6)
discrete time
where y represents the system output signal and u is the system input signal. For this class of systems, it
can be shown that the impulse response h can be written as a convolution between n exponential signals
which are defined from the system model (6); more specifically [3]:
(
hi (t) = eri t σ(t)
analog, or
h = h1 ∗ h2 ∗ · · · ∗ hn ,
(7)
hi (k) = rik−1 σ(k − 1) discrete time
where r1 , r2 , . . . , rn , ri ∈ C, are the roots of the characteristic equation xn +an−1 xn−1 +· · ·+a2 x2 +a1 x+a0 =
0, associated to the model (6).
The result in Equation (7) above motivate us to find a procedure to compute the convolution between
exponential signals. In most text books this question is generally dealt in the domain of Laplace or Z
transforms, where time domain convolution, under certain circumstances, becomes the usual product (see
e.g. [4]). In next sections, on the other hand, we show that convolution between exponential signals can
be evaluated directly in time domain, without having to solve integrals or summations, just by solving an
algebraic system of linear equations involving Vandermonde matrices. This approach is very adequate to be
implemented computationally in software packages like Scilab [5]. We should note that, since this is a quite
old question, equivalent results scattered in literature may exists (see e.g. [6, 7] for results obtained in the
context of probability theory); but we believe that our approach to this problem is new. Additionally to find
the system impulse response (7), we also use the same technique to compute the complete solution of the
differential/difference equation (6) for a given signal u.
1 In another important context, convolution can also be used to compute the probability density function of a sum of
independent random variables [8, 9].
2
3
Convolution between analog exponential signals
Consider the analog time signal h : R → C defined by:
rt
h(t) = e σ(t),
r ∈ C, and σ(t) =
(
0, t < 0
1, t > 0
(8)
which is well known to appear as the impulse response of (causal) linear time invariant systems (LTI)
modeled by a first order ordinary differential equation. We note that h, as defined in (8), has two simple,
and important, properties:
1. Its module has a jump discontinuity of amplitude one at t = 0, more precisely, h(0− ) = 0 and h(0+ ) = 1;
2. Its derivative (ḣ) satisfies the (first order differential) equation ḣ = rh + δ, as can be deduced from (1).
Now lets consider the convolution between two signals of this kind, that is, let be h1 (t) = er1 t σ(t) and
h2 (t) = er2 t σ(t). Since both of them are zero for t < 0, we get from (3) that (h1 ∗ h2 )(t) = 0 for t < 0 and,
for t > 0 we have:
Z
Z
t
(h1 ∗ h2 )(t) =
t
er1 τ er2 (t−τ ) dτ = er2 t
0
e(r1 −r2 )τ dτ,
(9)
0
and, before solving this integral, we note that the convolution h1 ∗ h2 satisfies the properties below:
1. h1 ∗ h2 is continuous at t = 0, more precisely, (h1 ∗ h2 )(0− ) = (h1 ∗ h2 )(0+ ) = 0, since by (9) we have:
+
r2 0+
(h1 ∗ h2 )(0 ) = e| {z }
=1
Z
|
0+
0
e(r1 −r2 )τ dτ = 0;
{z
}
(10)
=0
and, of course, the integral above is zero because we have an integration of an exponential function
over an infinitesimal interval.
2. The derivative of (h1 ∗ h2 ), that is (h1 ∗˙ h2 ), is such that (h1 ∗˙ h2 )(0− ) = 0 and (h1 ∗˙ h2 )(0+ ) = 1. In
fact:
(h1 ∗˙ h2 ) = h1 ∗ ḣ2 = h1 ∗ (r2 h2 + δ) = r2 (h1 ∗ h2 ) + h1 ∗ δ = r2 (h1 ∗ h2 ) + h1
and then
(h1 ∗˙ h2 )(0+ ) = r2 (h1 ∗ h2 )(0+ ) + h1 (0+ ) = 1
|
{z
} | {z }
=0
(11)
=1
Now we return to analyse the integral in (9), by considering two cases:
1. r1 6= r2 (or h1 6= h2 ):
(h1 ∗ h2 )(t)
=
(h1 ∗ h2 )(t)
=
1
1
er1 t σ(t) +
er2 t σ(t), or
r1 − r2
r2 − r1
1
1
A1 h1 (t) + A2 h2 (t), A1 =
and A2 =
r1 − r2
r2 − r1
(12)
(13)
Remark 3.0.1. Note that in case where r1 and r2 is a complex conjugate pair, represented by α ± jω,
we get from (12) that (h1 ∗ h2 )(t) = (eαt /ω) sin(ωt) for t ≥ 0.
From Equation (13) we see that, in case that r1 6= r2 , the convolution h1 ∗ h2 can be written as a linear
combination of the signals h1 and h2 , and this fact, along with conditions (10) and (11), can be used to
find the scalars A1 and A2 , without the need of solving the convolution integral (9), as shown bellow:
(h1 ∗ h2 )(0+ ) =
(h1 ∗˙ h2 )(0+ ) =
or:
1
r1
1
r2
A1 h1 (0+ ) + A2 h2 (0+ ) = A1 + A2 = 0
A1 ḣ1 (0+ ) + A2 ḣ2 (0+ ) = A1 r1 + A2 r2 = 1
0
1
A1
A1
=
=
=⇒
1
r1
A2
A2
Solving (14) we get A1 and A2 as shown in (13).
3
1
r2
−1
0
.
1
(14)
2. r1 = r2 = r (or h1 = h2 = h):
(h ∗ h)(t) = tert σ(t) = th(t)
(15)
Now we consider a generalization of the results above for a convolution of n ≥ 2 exponential signals as
shown in (8). We start by finding a generalization for the conditions (10) and (11):
Theorem 3.1. Consider the convolution of n ≥ 2 signals {h1 , h2 , . . . , hn } with hj (t) = erj t σ(t) and rj ∈ C.
The i-th derivative of (h1 ∗ h2 ∗ · · · ∗ hn ), represented by (h1 ∗ h2 ∗ · · · ∗ hn )(i) , evaluated at t = 0+ is given by:
(
0, i = 0, 1, . . . , n − 2
(i) +
(h1 ∗ h2 ∗ · · · ∗ hn ) (0 ) =
1, i = n − 1
and we consider (h1 ∗ h2 ∗ · · · ∗ hn )(0) = h1 ∗ h2 ∗ · · · ∗ hn .
Proof. We note that (h1 ∗ h2 ∗ · · · ∗ hn )(0+ ) = 0 if n ≥ 2, since this involves an integration of exponentials
over an infinitesimal interval; this proves that (h1 ∗ h2 ∗ · · ·∗ hn )(0) (0+ ) = 0. Now consider (h1 ∗ h2 ∗ · · ·∗ hn )(i)
for 1 ≤ i ≤ n − 2, then:
(h1 ∗ h2 ∗ · · · ∗ hn )(i)
= (ḣ1 ∗ ḣ2 ∗ · · · ∗ ḣi ) ∗ (hi+1 ∗ · · · ∗ hn−1 ∗ hn )
{z
}
|
at least two terms
= [(r1 h1 + δ) ∗ (r2 h2 + δ) ∗ · · · ∗ (ri hi + δ)] ∗ (hi+1 ∗ · · · ∗ hn−1 ∗ hn )
= (f + δ) ∗ (hi+1 ∗ · · · ∗ hn−1 ∗ hn )
= f ∗ (hi+1 ∗ · · · ∗ hn−1 ∗ hn ) + (hi+1 ∗ · · · ∗ hn−1 ∗ hn )
(16)
Since the two terms in (16) are composed by a convolution of at least two signals, we conclude that (h1 ∗
h2 ∗ · · · ∗ hn )(i) (0+ ) is equals to zero. Now, considering i = n − 1, we have:
(h1 ∗ h2 ∗ · · · ∗ hn )(n−1)
=
=
=
=
(ḣ1 ∗ ḣ2 ∗ · · · ∗ ḣn−1 ) ∗ hn
(r1 h1 + δ) ∗ (r2 h2 + δ) ∗ · · · ∗ (rn−1 hn−1 + δ) ∗ hn
(f + δ) ∗ hn = f ∗ hn + δ ∗ hn
f ∗ hn + hn
(17)
Then from (17), since f ∗ hn is a sum of (at least) two signals convolution, we have that (f ∗ hn )(0+ ) = 0
and consequently (h1 ∗ h2 ∗ · · · ∗ hn )(n−1) (0+ ) = hn (0+ ) = 1
In the following we will find a procedure for computing the convolution h1 ∗ h2 ∗ · · · ∗ hn for n ≥ 2 and
hj (t) = erj t σ(t) with rj ∈ C without the need of solving integrals. To begin with, we consider the case where
hi 6= hj for i 6= j, which implies ri 6= rj for i 6= j, and it is just a generalization of (14):
Theorem 3.2. The convolution between n ≥ 2 exponentials signals {h1 , h2 , . . . , hn }, with hj (t) = erj t σ(t),
rj ∈ C and hi 6= hj for i 6= j, is given by
h1 ∗ h2 ∗ · · · ∗ hn = A1 h1 + A2 h2 + · · · + An hn ,
(18)
where Aj ∈ C are scalars that can be computed by solving a linear system V A = B where V is the
n × n (nonsingular) Vandermonde matrix defined by Vij = rji−1 , A and B are the n-column vectors A =
(A1 , A2 , . . . , An ) and B = (0, 0, . . . , 1), that is:
1
1
···
1
0
A1
r1
r
·
·
·
r
A
2
n 2
0
2
2
2
r1
r
·
·
·
r
0
A
2
n 3 =
(19)
..
..
..
.. .. ..
.
.
.
.
.
.
r1n−1
r2n−1
rnn−1
···
An
1
So, vector A is the last (n-th) column of the inverse of V .
Proof. We use induction on n to prove (18), which is valid for n = 2, as shown in (13). Suppose (18) is valid
for n = k, and we prove it for n = k + 1:
h1 ∗ h2 ∗ · · · ∗ hk ∗ hk+1
=
=
=
=
=
(h1 ∗ h2 ∗ · · · ∗ hk ) ∗ hk+1
(A1 h1 + A2 h2 + · · · + Ak hk ) ∗ hk+1
A1 (h1 ∗ hk+1 ) + A2 (h2 ∗ hk+1 ) + · · · + Ak (hk ∗ hk+1 )
A1 (B1 h1 + C1 hk+1 ) + A2 (B2 h2 + C2 hk+1 ) + · · · + Ak (Bk hk + Ck hk+1 )
(A1 B1 )h1 + (A2 B2 )h2 + · · · + (Ak Bk )hk + (A1 C1 + · · · + Ak Ck )hk+1 ,
4
and then (18) is proved. To prove (19) we take the i-th derivative at t = 0+ on both sides of (18) to get:
(i)
(i)
(h1 ∗ h2 ∗ · · · ∗ hn )(i) (0+ ) = A1 h1 (0+ ) + A2 h2 (0+ ) + · · · + An hn(i) (0+ ),
i = 0, 1, 2, . . . , n − 1.
(i)
Applying Theorem 3.1 to left side of equation above and using the fact that hj (0+ ) = rji we get (19).
Now we consider the more general convolution h1 ∗h2 ∗· · ·∗hn , n ≥ 2, where there is the possibility of some
hi to be repeated in the convolution, that is hi = hj for some i 6= j. We initially consider some facts about
the so-called “convolution power” (or “n-fold” convolution [8, 9]) of exponentials, that is, the convolution of
h, as defined in (8), repeated between itself n times, and we represent it by h∗n (in Equation (15) we have
a formula for h∗2 ).
Lemma 3.2.1. The convolution power of n exponentials h(t) = ert σ(t), denoted by h∗n , is given by
1
tn−1 h(t),
h∗n (t) = (h ∗ h ∗ · · · ∗ h)(t) =
|
{z
}
(n − 1)!
n≥1
n terms
Proof. By induction on n. It is trivially true for n = 1 and suppose it is valid for n = k, then:
Z t
1
h∗(k+1) (t) = (h∗k ∗ h)(t) =
τ k−1 erτ er(t−τ ) dτ, t > 0
(k
−
1)!
0
Z t
ert
τ k−1 dτ
=
(k − 1)! 0
1
1
=
tk ert = tk h(t).
k(k − 1)!
k!
The Lemma bellow shows a generalization of Theorem 3.1 applied to the convolution power of h:
Lemma 3.2.2. Let be h(t) = ert σ(t), then i-th derivative of h∗n , for n ≥ 2, computed at t = 0+ and
represented by (h∗n )(i) (0+ ), is given by:
i = 0, 1, . . . , n − 2
0,
(20)
(h∗n )(i) (0+ ) = i
ri−n+1 , i ≥ n − 1
n−1
Proof. Equation (20) follows from Lemma 3.2.1 by setting k = n − 1 in the well-known formula:
i = 0, 1, . . . , k − 1
k
0,
i
d
t rt
=
e
i i−k
dti k!
t=0
r , i≥k
k
r1 t
1
2
Now we analyse how it would be like the convolution h∗n
∗ h∗n
σ(t) and h2 (t) =
1
2 , where h1 (t) = e
e σ(t), with r1 6= r2 , that is the convolution between the “n1 -power” convolution of h1 with the “n2 -power”
convolution of h2 when h1 6= h2 .
r2 t
Lemma 3.2.3. Let be h1 (t) = er1 t σ(t) and h2 (t) = er2 t σ(t), with r1 6= r2 , the convolution between the
1
2
n1 -power convolution of h1 and the n2 -power convolution of h2 , denoted by h∗n
∗ h∗n
1
2 , is given by:
2
1
∗ h∗n
h∗n
2
1
=
(h1 ∗ h1 ∗ · · · ∗ h1 ) ∗ (h2 ∗ h2 ∗ · · · ∗ h2 )
|
{z
} |
{z
}
n1 terms
=
(A1 h1 +
A2 h∗2
1
+
n2 terms
1
· · · An1 h∗n
1 ) + (B1 h2
∗n2
+ B2 h∗2
2 + · · · Bn2 h1 )
Proof. We prove by induction on (n1 , n2 ). It is true for (n1 , n2 ) = (1, 1) as shown in (13).
5
1. Induction on n1 : Valid for n1 = k and n2 = 1. Let it be n1 = k + 1:
∗(k+1)
h1
∗ h2 = h1 ∗ (h∗k
1 ∗ h2 ) =
∗k
h1 ∗ (A1 h1 + A2 h∗2
1 + · · · Ak h1 + B1 h2 )
∗(k+1)
∗3
A1 h∗2
1 + A2 h1 + · · · Ak+1 h1
=
=
A1 h∗2
1
=
(B1 C1 )h1 +
+ B1 (h1 ∗ h2 )
∗(k+1)
+ · · · Ak+1 h1
+ B1 (C1 h1 + C2 h2 )
∗(k+1)
∗2
∗3
+ (B1 C2 )h2
A1 h1 + A2 h1 + · · · Ak+1 h1
A2 h∗3
1
+
∗(k+1)
1
2. Induction on n2 : Valid for generic n1 and n2 = k. Let it be n2 = k + 1: Since h∗n
∗ h2
1
∗n1
∗k
(h1 ∗ h2 ) ∗ h2 , then:
1
∗ h∗k
(h∗n
2 ) ∗ h2
1
∗n1
∗2
∗k
= [(A1 h1 + A2 h∗2
1 + · · · + An1 h1 ) + (B1 h2 + B2 h2 + · · · + Bk h2 )] ∗ h2
=
∗(k+1)
∗n1
∗3
∗ h2 ) +B1 h∗2
= A1 (h1 ∗ h2 ) + A2 (h∗2
2 + B2 h 2 + · · · + Bk h 2
1 ∗ h2 ) + · · · + An1 (h1
{z
}
|
∗n1
Rearranged as (C1 h1 +C2 h∗2
1 +···+Cn1 h1
= (C1 h1 +
C2 h∗2
1
+ ···+
1
Cn1 h∗n
1 )
+Dh2 )
∗(k+1)
∗3
+ (Dh2 + B1 h∗2
2 + B2 h 2 + · · · + Bk h 2
)
We now prove the general result about the power convolution of n exponential signals as show in (8)
which is a generalization of Theorem 3.2:
Theorem 3.3. The convolution between n ≥ 2 exponentials signals {h1 , h2 , . . . , hn }, with hi (t) = eri t σ(t),
ri ∈ C and q distinct hs , each of them repeated ns times, so that n1 + n2 + · · · nq = n, is given by
1
2
q
h∗n
∗ h∗n
=
∗ · · · ∗ h∗n
q
1
2
n1
X
A1j h∗j
1 +
n2
X
j=1
j=1
A2j h∗j
2 + ···+
nq
X
Aqj h∗j
q ,
(21)
j=1
where Asj ∈ C are scalars that can be computed by solving a linear system V A = B where V is the n × n
nonsingular confluent (or generalized) Vandermonde matrix defined by V = V1 V2 · · · Vq , where each
block Vs is the n × ns matrix whose entries are defined by
i<j
0,
(Vs )ij =
i − 1 i−j
r , i≥j
j−1 s
A and B are the n-column vectors A = (A1 , A2 , . . . , Aq ), each As is a ns -column vector, and B =
(01 , 02 , . . . , Bq ), where 0s are ns -column zero vectors and Bq is the nq -column vector (0, 0, · · · , 1) that
is:
A1
01
A2 02
V1 V2 · · · Vq A3 = 03
(22)
.. ..
. .
Aq
Bq
So, vector A is the last (n-th) column of the inverse of V . Alternatively, using Lemma 3.2.1, we can rewrite
(21) as
1
2
q
h∗n
∗ h∗n
= p 1 h1 + p 2 h2 + · · · + p q hq
(23)
∗ · · · ∗ h∗n
q
1
2
where each ps , s = 1, . . . , q, is a polynomial defined as
ps (t) =
ns
X
Asj
j=1
tj−1
(j − 1)!
Proof. We use induction on q to prove (21), which is valid for q = 2, as shown in Lemma 3.2.3. Suppose
6
(21) is valid for q = k, and we prove it for q = k + 1:
∗n
k+1
1
2
k
h∗n
∗ h∗n
∗ · · · ∗ h∗n
∗ hk+1
1
2
k
=
=
∗n
k+1
1
2
k
(h∗n
∗ h∗n
∗ · · · ∗ h∗n
1
2
k ) ∗ hk+1
nk
n2
n1
X
X
X
∗ h∗nk+1
Akj h∗j
A2j h∗j
A1j h∗j
2 + ··· +
1 +
k
k+1
=
j=1
=
n1
X
∗n
k+1
A1j (h∗j
1 ∗ hk+1 ) +
B1j h∗j
1 +
j=1
j=1
j=1
j=1
n1
X
n2
X
j=1
n2
X
j=1
∗n
k+1
A2j (h∗j
2 ∗ hk+1 ) + · · · +
B2j h∗j
2 + ···+
nk
X
nk+1
Bkj h∗j
k +
X
nk
X
j=1
∗n
k+1
Akj (h∗j
k ∗ hk+1 )
B(k+1)j h∗j
k+1
j=1
j=1
and the (21) is proved. To prove (22) we take the i-th derivative at t = 0+ on both sides of (21) to get:
+
1
2
q (i)
(h∗n
∗ h∗n
∗ · · · ∗ h∗n
q ) (0 ) =
1
2
n1
X
(i) +
A1j (h∗j
1 ) (0 ) +
j=1
j=1
nq
+
X
n2
X
(i) +
Aqj (h∗j
q ) (0 ),
j=1
(i) +
A2j (h∗j
2 ) (0 ) + · · ·
i = 0, 1, 2, . . . , n − 1.
(i)
(i) +
+
i
Applying Theorem 3.1 to left side of equation above and using the fact that (h∗1
k ) (0 ) = hk (0 ) = rk
along with Lemma 3.2.2, i.e., for j ≥ 2:
i = 0, 1, . . . , j − 2
0,
(i) +
(h∗j
)
(0
)
=
s
i
ri−j+1 , i ≥ j − 1
j−1 s
we get (22).
3.1
Solution of ordinary differential equations with constant coefficients
Consider the ordinary differential equation
y (n) + an−1 y (n−1) + · · · + a1 ẏ + a0 y = u,
ai ∈ R
(24)
which models an n order (causal) linear time invariant (LIT) system with input signal u and output signal
y. The impulse response (h) for this system is given by the convolution [3]:
h = h1 ∗ h2 ∗ · · · ∗ hn ,
hi (t) = eri t σ(t),
ri ∈ C
and r1 , r2 , . . . , rn are the roots of the characteristic equation xn + an−1 xn−1 + · · · + a1 x + a0 = 0 associated
to (24). Supposing that the characteristic equation has q distinct roots rs , each one repeated ns times, so
that n1 + n2 + · · · + nq = n, then we can obtain the impulse response h by using Theorem 3.3, Equation (23),
that is
ns
X
tj−1
Asj
h = p1 h1 + p2 h2 + · · · + pq hq , hs (t) = ers t , ps (t) =
, t>0
(25)
(j − 1)!
j=1
where Asj , j = 1, . . . , ns and s = 1, . . . , q are calculated by solving the Vandermonde system (22).
The complete solution of (24) is generally written as
y = yh + yp
(26)
where yh is the homogeneous (or zero input) solution and yp is a particular solution, i.e., it depends on input
signal u. When solving (24) for t ≥ 0, the particular solution yp can be written as
(
Z t
0,
t<0
u(τ )h(t − τ )dτ = [(uσ) ∗ h](t), where (uσ)(t) =
yp (t) =
(27)
u(t), t > 0
0
The homogeneous solution (yh ) has the same format of (25), that is
yh = p̄1 h1 + p̄2 h2 + · · · + p̄q hq ,
hs (t) = ers t , and p̄s (t) =
ns
X
j=1
7
Āsj
tj−1
.
(j − 1)!
(28)
Therefore to solve (24) we need to obtain yh , which is equivalent to obtain the constants Āsj in (28), and
then compute yp , by evaluating the convolution “(uσ) ∗ h” as showed in (27). To find yh we use the fact
that the particular solution yp is a convolution between n + 1 signals, namely, “(uσ) ∗ h1 ∗ h2 ∗ · · · ∗ hn ” ,
and conclude, by using Theorem 3.1, that:
yp (0+ ) = ẏp (0+ ) = ÿp (0+ ) = · · · = yp(n−1) (0+ ) = 0
and so, using these conditions in (26), we get:
y(0+ ) = yh (0+ ),
ẏ(0+ ) = ẏh (0+ ),
ÿ(0+ ) = ÿh (0+ ),
···
(n−1)
y (n−1) (0+ ) = yh
(0+ ).
This set of conditions on yh can be used to find the constants Āsj in (28) since the “initial values”
y(0), ẏ(0), ÿ(0), . . . , y (n−1) (0) are generally known when solving (24) for t ≥ 0. This implies that the constants Āsj , s = 1, . . . q and j = 1, . . . , ns , can be computed by solving a Vandermonde system like the one
showed in Theorem 3.3, that is V Ā = B̄, where the Vandermonde matrix V is the same one used to compute
the impulse response h, Ā is the n × 1 vector composed by the Āsj ’s and the vector B̄, differently from the
one used to compute h, it is now defined as B̄ = (y(0), ẏ(0), ÿ(0), · · · , y n−1 (0)).
Finally, in order to obtain the complete solution y for (24) as shown in (26), we need to compute the
particular solution “yp = (uσ) ∗ h”, that is the convolution between the input signal uσ and the impulse
response h, and to avoid solving a convolution integral we can use the result of Theorem 3.3, by writing,
if possible, the signal “uσ” as a convolution (or a finite sum) of exponential signals of type “ert σ(t)”, for
some r ∈ C. In this situation, as shown in examples in Section 5.1 bellow, we increase the order of the
Vandermonde matrix, as defined in Theorem 3.3, depending on how many “exponential modes” exists in the
input signal “uσ”.
4
Convolution between discrete time exponential signals
In the context of discrete time signals we consider the exponential signal e : Z → C defined as
(
0, k < 0
e(k) = rk σ(k), r 6= 0 ∈ C, σ(k) =
1, k ≥ 0
(29)
And also consider the signal defined as a right shift of “e” by one unit, that is h = [e]1 , or:
h(k) = rk−1 σ(k − 1),
(30)
which is well known to appear as the impulse response of (causal) linear time invariant systems (LTI)
modeled by a first order difference equation, since it satisfies the relationship h(k + 1) = rh(k) + δ(k). Now
lets consider the convolution between two signals of this kind, that is, let be h1 (k) = r1k−1 σ(k − 1) and
h2 (k) = r2k−1 σ(k − 1), with r1 6= 0 and r2 6= 0. Since both of them are time shift of exponentials as defined
in (29), we can write h1 = [e1 ]1 and h2 = [e2 ]1 , where e1 (k) = r1k σ(k) and e2 = r2k σ(k), and then:
h1 ∗ h2 = [e1 ]1 ∗ [e2 ]1 = (e1 ∗ [δ]1 ) ∗ (e2 ∗ [δ]1 ) = (e2 ∗ e2 ) ∗ ([δ]1 ∗ [δ]1 ) = (e1 ∗ e2 ) ∗ [δ]2 = [e1 ∗ e2 ]2
therefore, h1 ∗ h2 can be obtained by a right time shift of e1 ∗ e2 by two units. We develop e1 ∗ e2 instead,
noting that (e1 ∗ e2 )(k) = 0 for k < 0, since both e1 (k) and e2 (k) are null for k < 0 and
(e1 ∗ e2 )(k) =
k
X
r1j r2k−j ,
j=0
for k ≥ 0
(31)
Additionally we also have that (e1 ∗ e2 )(0) = r10 r20 = 1. Then, before solving this summation, we note that
the convolution h1 ∗ h2 is such that (h1 ∗ h2 )(k) = 0 for k ≤ 0, and, more importantly:
(h1 ∗ h2 )(1)
(h1 ∗ h2 )(2)
= 0
= 1
since h1 ∗ h2 is a right shift of e1 ∗ e2 by two units.
We now develop the summation in (31) by considering two cases:
8
(32)
(33)
1. r1 6= r2 (or e1 6= e2 ):
= r2k [1 + (r1 /r2 ) + (r1 /r2 )2 + · · · + (r1 /r2 )k ]
(e1 ∗ e2 )(k)
= r2k
=
(r1k+1 /r2k+1 ) − 1
(r1 /r2 ) − 1
r1k+1 − r2k+1
r1 − r2
and since h1 ∗ h2 = [e1 ∗ e2 ]2 , then (h1 ∗ h2 )(k) = (e1 ∗ e2 )(k − 2) or:
1
1
r1k−1 σ(k − 1) +
rk−1 σ(k − 1), or
r1 − r2
r2 − r1 2
1
1
A1 h1 (k) + A2 h2 (k), A1 =
and A2 =
r1 − r2
r2 − r1
(h1 ∗ h2 )(k) =
(h1 ∗ h2 )(k) =
(34)
(35)
Remark 4.0.1. Note that in case where r1 and r2 is a complex conjugate pair, represented by α±jω =
Re±jφ , we get from (34) that (h1 ∗ h2 )(k) = (Rk−1 /ω) sin[(k − 1)φ], for k ≥ 1.
From Equation (35) we see that, in case that r1 6= r2 , the convolution h1 ∗ h2 can be written as a linear
combination of signals h1 and h2 , and this fact, along with conditions (32) and (33), can be used to
find the scalars A1 and A2 , without the need of solving the convolution sum (31), as shown bellow:
(h1 ∗ h2 )(1)
(h1 ∗ h2 )(2)
And then:
1
r1
1
r2
= A1 h1 (1) + A2 h2 (1) = A1 + A2 = 0
= A1 h1 (2) + A2 h2 (2) = A1 r1 + A2 r2 = 1
1
0
A1
A1
=
=
=⇒
r1
A2
1
A2
1
r2
−1
0
.
1
(36)
Solving (36) we get A1 and A2 as shown in (35).
2. r1 = r2 = r (or e1 = e2 = e):
(e ∗ e)(k) = rk
k
X
j=0
rj ∗ r−j = (k + 1)rk ,
k≥0
and then, since h1 = h2 = h = [e]1 , (h ∗ h)(k) = (e ∗ e)(k − 2) is given by
(
0,
k≤1
(h ∗ h)(k) =
k−2
(k − 1)r
, k≥2
(37)
Now we consider a generalization of the results above for a convolution of n ≥ 2 exponential signals as
shown in (30). We start by finding a generalization for conditions (32) and (33) applied to the convolution
h1 ∗ h2 ∗ · · · ∗ hn , with hi (k) = rik−1 σ(k − 1) and n ≥ 2:
Theorem 4.1. Consider the convolution h1 ∗ h2 ∗ · · · ∗ hn , n ≥ 2 and each hi (k) = rik−1 σ(k − 1), ri 6= 0 ∈ C.
Then we have
(
0, k ≤ n − 1
(h1 ∗ h2 ∗ · · · ∗ hn )(k) =
1, k = n
Proof. Defining ei (k) = rik σ(k), we note that hi = [ei ]1 and then
(h1 ∗ h2 ∗ · · · ∗ hn ) = ([e1 ]1 ∗ [e2 ]1 ∗ · · · ∗ [en ]1 ) = [e1 ∗ e2 ∗ · · · ∗ en ]n
that is, h1 ∗ h2 ∗ · · · ∗ hn is a time shift right of e1 ∗ e2 ∗ · · · ∗ en by n units, and since (e1 ∗ e2 ∗ · · · ∗ en )(k) = 0
for k < 0 and (e1 ∗ e2 ∗ · · · ∗ en )(0) = 1 the result is proved.
In the following we will find a formula for computing the convolution h1 ∗ h2 ∗ · · · ∗ hn for n ≥ 2 and
hj (k) = rjk−1 σ(k − 1) with rj ∈ C. To begin with, we consider the case where hi 6= hj for i 6= j, which
implies ri 6= rj for i 6= j, and it is just a generalization of Equation (36):
9
Theorem 4.2. The convolution between n ≥ 2 exponentials signals hj (k) = rjk−1 σ(k − 1), j = 1, 2, . . . , n,
with rj 6= 0 ∈ C and hi 6= hj for i 6= j, is given by
h1 ∗ h2 ∗ · · · ∗ hn = A1 h1 + A2 h2 + · · · + An hn ,
(38)
where Aj ∈ C are scalars that can be computed by solving a linear system V A = B where V is the
n × n (nonsingular) Vandermonde matrix defined by Vij = rji−1 , A and B are the n-column vectors A =
(A1 , A2 , . . . , An ) and B = (0, 0, . . . , 1), that is:
1
1
···
1
0
A1
r1
r
·
·
·
r
A
2
n 2
2
0
2
2
r1
r
·
·
·
r
A
0
2
n 3 =
(39)
..
..
..
.. .. ..
.
.
.
.
.
.
r1n−1
r2n−1
···
rnn−1
An
1
So, vector A is the last (n-th) column of the inverse of V .
Proof. We use induction on n to prove (38), which is valid for n = 2, as shown in (35). Suppose (38) is valid
for n = k, and we prove it for n = k + 1 following the same reasoning we used to prove (18) in Theorem 3.2.
To prove (39) we apply the result of Theorem 4.1 to Equation (38). Taking the value at k = i on both sides
of (38) we have:
(h1 ∗ h2 ∗ · · · ∗ hn )(i) = A1 h1 (i) + A2 h2 (i) + · · · + An hn (i),
i = 1, 2, . . . , n.
Using Theorem 4.1 and the fact that hj (i) = rji−1 we get (39).
Now we consider the more general convolution h1 ∗ h2 ∗ · · · ∗ hn , n ≥ 2, where there is the possibility of
some hi to be repeated in the convolution, that is hi = hj for some i 6= j. To begin with, we consider some
facts about “n-power” convolution of discrete time exponentials, that is, the convolution of h, as defined in
(30), repeated between itself n times, that we represent it by h∗n (in Equation (37) we have a formula for
h∗2 ). The Lemma bellow shows a generalization of Theorem 4.1 applied to the “n-power” convolution of the
exponential signal:
Lemma 4.2.1. The power convolution of n ≥ 1 exponentials e(k) = rk σ(k), r 6= 0 ∈ C, denoted by e∗n , is
given by
k<0
0,
∗n
e (k) = (e ∗ e ∗ · · · ∗ e)(k) =
n−1+k k
{z
}
|
r , k≥0
n terms
n−1
or, in a more compact notation
n−1+k k
n−1+k
e (k) =
r σ(k) =
e(k)
n−1
n−1
∗n
Proof. By induction on n. It is trivially true for n = 1 and suppose it is valid for n = p then
p−1+k k
e∗p (k) =
r σ(k)
p−1
Obviously e∗(p+1) (k) = 0 for k < 0 since e(k) = 0 for k < 0; for k ≥ 0 we have:
e∗(p+1) (k) = (e∗p ∗ e)(k) =
k
X
j=0
e∗p (j)e(k − j)
=
k
X
p − 1 + j j (k−j)
r r
p−1
j=0
=
rk
=
10
k
X
p−1+j
p−1
j=0
p+k
rk
p
(40)
In the last step of the proof above we used the following well-known fact about sum of binomial coefficients
[8]:
k
X
p−1+j
p−1
p
p+1
p−1+k
p+k
=
+
+
+ ···+
=
p−1
p−1
p−1
p−1
p−1
p
j=0
Corollary 4.2.1. If we consider the n-power convolution of exponentials h(k) = rk−1 σ(k − 1), that is
h = [e]1 , we have:
k ≤n−1
0,
∗n
(41)
h (k) =
k − 1 k−n
r
, k≥n
n−1
equivalently
h∗n (k) =
k−1
n−1
since it is assumed that
Proof. Since h = [e]1 , then
1
rn−1
k−1
h(k),
n−1
n≥1
(42)
= 0 for k = 1, 2, . . . , n − 1.
h∗n
=
(h ∗ h ∗ · · · ∗ h)
{z
}
|
n terms
=
=
=
[e]1 ∗ [e]1 ∗ · · · ∗ [e]1
[e ∗ e ∗ · · · ∗ e]n
[en ]n
that is, h∗n is e∗n (right) shifted n units. Then we have by setting k := k − n in (40):
k − 1 k−n
h∗n (k) =
r
σ(k − n)
(43)
n−1
k−1
which is equivalent to (41). To obtain (42), we note that n−1
= 0, for k = 1, 2, . . . n − 1, and so (43) can
be rewriten as
1
k − 1 k−1
k−1
1
k − 1 k−n
∗n
r
σ(k − 1) = n−1
h(k)
h (k) =
r
σ(k − 1) = n−1
n−1
n−1
n−1
r
r
k−1
1
2
Now we analyse how it would be like the convolution h∗n
∗ h∗n
σ(k − 1) and
1
2 , where h1 (k) = r1
k−1
h2 (k) = r2 σ(k − 1), with r1 6= r2 :
Lemma 4.2.2. Let be h1 (t) = r1k−1 σ(k − 1) and h2 (k) = r2k−1 σ(k − 1), with r1 6= r2 , the convolution
1
2
between the n1 -power convolution of h1 and the n2 -power convolution of h2 , denoted by h∗n
∗ h∗n
1
2 , is given
by:
1
2
h∗n
∗ h∗n
1
2
=
(h1 ∗ h1 ∗ · · · ∗ h1 ) ∗ (h2 ∗ h2 ∗ · · · ∗ h2 )
{z
} |
{z
}
|
n1 terms
=
(A1 h1 +
A2 h∗2
1
+
n2 terms
1
· · · An1 h∗n
1 ) + (B1 h2
∗n2
+ B2 h∗2
2 + · · · Bn2 h1 )
Proof. We prove by induction on (n1 , n2 ). It is true for (n1 , n2 ) = (1, 1) as shown in (35). The inductive
step is the same one used in the proof of Lemma 3.2.3 for the analog time case.
In the following we prove the general result about the convolution of n exponential signals as show in
(30) which is a generalization of Theorem 4.2:
Theorem 4.3. The convolution between n ≥ 2 exponentials signals hi (k) = rik−1 σ(k − 1), i = 1, 2, . . . , n,
with ri 6= 0 ∈ C, and q distinct hs , each of them repeated ns times, so that n1 + n2 + · · · nq = n, is given by
1
h∗n
1
∗
2
h∗n
2
∗ ···∗
q
h∗n
q
=
n1
X
A1j h∗j
1
j=1
+
n2
X
j=1
11
A2j h∗j
2
+ ···+
nq
X
j=1
Aqj h∗j
q ,
(44)
where Asj ∈ C are scalars that can be computed by solving a linear system V A = B where V is the
n×n
(nonsingular) confluent (or generalized) Vandermonde matrix defined by V = V1 V2 · · · Vq , where
each block Vs is the n × ns matrix whose entries are defined by
i<j
0,
(Vs )ij =
i − 1 i−j
r , i≥j
j−1 s
A and B are the n-column vectors A = (A1 , A2 , . . . , Aq ), each As is a ns -column vector, and B =
(01 , 02 , . . . , Bq ), where 0s are ns -column zero vectors and Bq is the nq -column vector (0, 0, · · · , 1) that
is:
01
A1
A2 02
V1 V2 · · · Vq A3 = 03
(45)
.. ..
. .
Bq
Aq
So, vector A is the last (n-th) column of the inverse of V . Alternatively, using Equation (42), we can rewrite
(44) as
1
2
q
h∗n
∗ h∗n
= p 1 h1 + p 2 h2 + · · · + p q hq
(46)
∗ · · · ∗ h∗n
q
1
2
where each ps , s = 1, . . . , q, is a polynomial defined as
ps (k) =
ns
X
Asj
j=1
k−1
,
rsj−1 j − 1
1
k≥1
Proof. We use induction on q to prove (44), which is valid for q = 2, as shown in Lemma 4.2.2. The inductive
step follows in the same way we did in the proof of Theorem 3.3. To prove (45) we evaluate Equation (44)
at k = i to obtain:
1
2
q
(h∗n
∗ h∗n
∗ · · · ∗ h∗n
q )(i) =
1
2
n1
X
A1j h∗j
1 (i) +
j=1
nq
+
X
n2
X
j=1
Aqj h∗j
q (i),
A2j h∗j
2 (i) + · · ·
i = 1, 2, . . . , n.
j=1
Applying the result of Theorem 4.1 to the left side of this equation and using Lemma 4.2.1, Equation (41),
that is for j ≥ 1:
0,
i≤j−1
h∗j
(i)
=
s
i − 1 i−j
r , i≥j
j−1 s
we get (45).
4.1
Solution of difference equations with constant coefficients
Consider the “n order” difference equation
y(k + n) + an−1 y(k + n − 1) + · · · + a1 y(k + 1) + a0 y(k) = u(k)
(47)
which models an n order discrete time (causal) linear time invariant (LIT) system with input signal u and
output signal y. The impulse response (h) for this system is given by the convolution [3]:
h = h1 ∗ h2 ∗ · · · ∗ hn ,
hi (k) = rik−1 σ(k − 1),
ri 6= 0 ∈ C
and r1 , r2 , . . . , rn are the roots of the characteristic equation xn + an−1 xn−1 + · · · + a1 x + a0 = 0 associated
to (47), which all are assumed to be non-zero.2 Supposing that the characteristic equation has q distinct
2 Zero roots are discarded and order of the difference equation reduced by the amount of discarded roots. The final solution
is then the solution of the reduced order equation right-shifted as many units as the number of zero roots of the characteristic
equation (see examples in Section 5.2).
12
non-zero roots rs , each one repeated ns times, so that n1 + n2 + · · · + nq = n, then we can obtain h by using
Theorem 4.3, Equation (46), that is
h = p 1 h1 + p 2 h2 + · · · + p q hp ,
hs (k) =
rsk−1 ,
ns
X
k−1
,
Asj j−1
ps (k) =
j−1
rs
j=1
1
k≥1
(48)
where Asj , j = 1, . . . , ns and s = 1, . . . , q are calculated by solving the Vandermonde system (45).
The solution of (47) for k ≥ 0 can be written as:
y = yh + yp
(49)
where yh is the homogeneous (or zero input) solution and yp is a particular solution, i.e., it depends on the
input signal u. When solving (49) for k ≥ 0, the particular solution can be written as
(
k
X
0,
k<0
u(j)h(k − j) = [(uσ) ∗ h](k), where (uσ)(k) =
(50)
yp (k) =
u(k), k ≥ 0
j=0
The homogeneous solution has the same format of (48), that is
yh = p̄1 h̄1 + p̄2 h̄2 + · · · + p̄q h̄q ,
h̄s (k) =
rsk ,
and p̄s (k) =
nX
s −1
j=0
k
Āsj j
,
rs j
1
k≥0
(51)
Therefore to solve (47) we need to obtain yh , which is equivalent obtain the constants Āsj in (51), and then
obtain yp , by evaluating the convolution “(uσ) ∗ h” as shown in (50). Since the particular solution yp is,
in fact, a convolution between n + 1 signals, namely, “(uσ) ∗ h1 ∗ h2 ∗ · · · ∗ hn ” , we conclude, by using
Theorem 3.1, that:
yp (0) = yp (1) = yp (2) = · · · = yp (n − 1) = 0
and so, by (49), we have that:
y(0) = yh (0),
y(1) = yh (1),
y(2) = yh (2),
···
y(n − 1) = yh (n − 1)
which can be used in (51) to find the constants Āsj , j = 1, . . . ns and s = 1, . . . q, since the “initial values”
y(0), y(1), y(2), . . . , y(n − 1) are generally known when solving (47) for k ≥ 0. In fact, constants Āsj are
computed by solving a Vandermonde system like the one showed in Theorem 4.3, that is V Ā = B̄, where
the Vandermonde matrix V is the same one used to compute the impulse response h, Ā is the n × 1 vector
composed by the Āsj ’s and the vector B̄, differently from the one used to compute h, it is defined as
B̄ = (y(0), y(1), y(2), · · · , y(n − 1)).
Finally, in order to obtain the complete solution y for (47) as shown in (49), we need to compute
the particular solution “yp = (uσ) ∗ h”, that is the convolution between the input signal uσ and the inpulse
response h, and this can be done by the result of Theorem 4.3 if we can write the signal “uσ” as a convolution
(or a sum) of exponential signals of type “rk σ(k)”, for some r 6= 0 ∈ C. In this situation, as shown in examples
bellow, we increase the order of the Vandermonde matrix, as defined in Theorem 4.3, depending on how
many “exponential modes” exists in the input signal “uσ”. In Section 5.2 we apply these results to the
resolution of some specific difference equations.
5
Examples
Bellow we apply the results discussed in previous sections to the solution to some specific differential/difference
equations.
5.1
Differential Equations
Example 5.1.1. Let be the second order initial value problem (IVP):
ÿ + 3ẏ + 2y = 1,
with y(0) = −1 and ẏ(0) = 2.
(52)
To find the solution y, we consider the characteristic equation is x2 + 3x + 2 = 0 whose roots as r1 = −1
and r2 = −2.
13
(a) Impulse response: h(t) = A1 e−t + A2 e−2t , where A1 and A2 are computed as
1
0
A1
1
1
A1
=
=
=⇒
−1
A2
1
−1 −2 A2
which implies h(t) = e−t − e−2t .
(b) Homogeneous solution: yh (t) = B1 e−t + B2 e−2t , where B1 and B2 are computed as:
1
1
B1
y(0)
−1
B1
0
=
=
=⇒
=
−1 −2 B2
ẏ(0)
2
−1
B2
which implies yh (t) = −e−2t .
(c) Particular solution: yp = (uσ) ∗ h, and (uσ)(t) = 1.σ(t) = e0t σ(t), then
yp = (uσ) ∗ h = h ∗ (uσ) = h1 ∗ h2 ∗ h3
where h1 (t) = e−t σ(t), h2 (t) = e−2t σ(t) and h3 (t) = e0t σ(t), or:
yp (t) = C1 e−t + C2 e−2t + C3 e0t
where C1 , C2 and C3 are compute as
1
1
−1 −2
1
4
the solution of the “augmented” Vandermonde system:
1 C1
0
C1
−1
0 C2 = 0 =⇒ C2 = 0.5
0 C3
1
C3
0.5
which implies yp (t) = −e−t + 0.5e−2t + 0.5.
Finally, the solution for the IVP (52) is y = yh + yp or
y(t) = −e−t − 0.5e−2t + 0.5
Example 5.1.2. Let be the following third order IVP
...
y + 7ÿ + 20ẏ + 24y = sin 2t, y(0) = 0, ẏ(0) = 1, ÿ(0) = −3
(53)
The characteristic equation is x3 + 7x + 20x + 24 = 0 whose roots are r1 = −3, r2 = −2 + 2i and r3 = −2 − 2i.
(a) Impulse response: h(t) = A1 e−3t + A2 e(−2+2i)t + A3 e(−2−2i)t , and
0
1
1
1
0.2
A1
A1
−3 −2 + 2i −2 − 2i A2 = 0 =⇒ A2 = −0.1 − 0.05i
1
9
−8i
8i
A3
−0.1 + 0.05i
A3
and then
h(t) = 0.2e−3t + (−0.1 − 0.05i)e(−2+2i)t + (−0.1 + 0.05i)e(−2−2i)t
which (optionally) can be simplified to
h(t) = 0.2e−3t − 0.2e−2t cos 2t + 0.1e−2t sin 2t.
(b) Homogeneous solution: yh (t) = B1 e−3t + B2 e(−2+2i)t + B3 e(−2−2i)t , and
1
1
1
B1
0
B1
0.2
−3 −2 + 2i −2 − 2i B2 = 1 =⇒ B2 = −0.1 − 0.3i
9
−8i
8i
B3
−3
B3
−0.1 + 0.3i
and then
yh (t) = 0.2e−3t + (−0.1 − 0.3i)e(−2+2i)t + (−0.1 + 0.3i)e(−2−2i)t
or
yh (t) = 0.2e−3t − 0.2e−2t cos 2t + 0.6e−2t sin 2t.
14
(c) Particular solution: yp = (uσ) ∗ h, since u(t) = sin 2t we have two possibilites:
uσ = 2(h4 ∗ h5 ) or uσ =
using uσ = 2(h4 ∗ h5 ) we have
h4 + h5
,
2i
where h4 (t) = e2it σ(t) and h5 (t) = e−2it σ(t)
yp = 2(h1 ∗ h2 ∗ h3 ∗ h4 ∗ h5 ),
where
h1 (t) = e−3t σ(t), h2 (t) = e(−2+2i)t σ(t), h3 (t) = e(−2−2i)t σ(t), h4 (t) = e2it σ(t), h5 (t) = e−2it σ(t).
So, to compute h1 ∗ h2 ∗ h3 ∗ h4 ∗ h5 , we have the following (augmented) Vandemonde system:
1
1
1
1
1
C1
0
C1
0.0307692
−3 −2 + 2i −2 − 2i 2i −2i C2 0
C2
−0.025i
9
C3 = 0 =⇒ (2) × C3 =
−8i
8i
−4
−4
0.025i
−27 16 + 16i 16 − 16i −8i 8i C4 0
C4 −0.0153846 + 0.0019231i
C5
81
−64
−64
16
16
1
C5
−0.0153846 − 0.0019231i
Then the particular solution is
yp (t) = 0.0307692e−3t + 0.05e−2t sin 2t − 0.0307692 cos 2t − 0.0038462 sin 2t
Finally, the solution y = yh + yp for the IVP (53) is given by:
y(t) = 0.2307692e−3t − 0.2e−2t cos 2t + 0.65e−2t sin 2t − 0.0307692 cos 2t − 0.0038462 sin 2t
Example 5.1.3. Let be the following IVP
ÿ + 4y = t cos 2t,
y(0) = −2,
ẏ(0) = 4
whose characteristic equation is x2 + 4 = 0 which implies r1 = e2it and r2 = e−2it .
(a) Impulse response: h(t) = A1 e2it + A2 e−2it , and
−0.25i
0
1
1
A1
A1
=
=
=⇒
0.25i
1
2i −2i A2
A2
Then
h(t) = −025ie2it + 0.25ie−2it = 0.5 sin 2t.
(b) Homogeneous solution: yh (t) = B1 e2it + B2 e−2it , and
1
1
B1
−2
B1
−1 − i
=
=⇒
=
2i −2i B2
4
−1 + i
B2
Then
yh (t) = (−1 − i)e2it + (−1 + i)e−2it = −2 cos 2t + 2 sin 2t
(c) Particular solution: yp = (uσ) ∗ h, and u(t) = t cos 2t = t(e2it + e−2it )/2, or:
u = 0.5(u1 + u2 ),
u1 (t) = te2it ,
u2 (t) = te−2it
and so, yp = 0.5(u1 σ) ∗ h + 0.5(u2 σ) ∗ h. Since u1 (t) = te2it and u2 (t) = te−2it , we have
u1 σ
u2 σ
=
=
h3 ∗ h3 ,
h4 ∗ h4 ,
h3 (t) = e2it σ(t)
h4 (t) = e−2it σ(t)
Therefore
(u1 σ) ∗ h
(u2 σ) ∗ h
=
=
h1 ∗ h2 ∗ h3 ∗ h3 ,
h1 ∗ h2 ∗ h4 ∗ h4 ,
h1 (t) = h3 (t) = e2it σ(t), h2 (t) = e−2it σ(t)
h1 (t) = e2it σ(t), h2 (t) = h4 (t) = e−2it σ(t)
15
(54)
and then
= C0 e−2it + p(t)e2it ,
= D0 e2it + q(t)e−2it ,
((u1 σ) ∗ h)(t)
((u2 σ) ∗ h)(t)
p(t) = C1 + C2 t + C3 t2 /2
q(t) = D1 + D2 t + D3 t2 /2
where
1
1
0
0
C0
−2i 2i
C1
1
0
=
−4 −4
4i
1 C2
8i −8i −12 6i C3
1
1
0
0
D0
2i −2i
D1
1
0
=
−4 −4 −4i
1 D2
−8i 8i −12 −6i D3
0
C0
−0.015625i
0
=⇒ C1 = 0.015625i
0
C2 0.0625
1
C3
−0.25i
0
D0
0.015625i
0
=⇒ D1 = −0.015625i
0
D2 0.0625
1
D3
0.25i
Since yp = 0.5(u1 σ) ∗ h + 0.5(u2 σ) ∗ h we have, after regrouping the terms
yp (t) = −0.03125 sin 2t + 0.0625t cos 2t + 0.125t2 sin 2t
and the solution y = yh + yp will be given by
y(t) = 1.96875 sin 2t − 2 cos 2t + 0.0625t cos 2t + 0.125t2 sin 2t
5.2
Difference Equations
Example 5.2.1. Let be the third order initial value problem (IVP):
y(k + 3) − 1.5y(k + 2) + 0.75y(k + 1) − 0.125y(k) = 1,
y(0) = −1, y(1) = 2, y(2) = 0.8
(55)
To find the solution y, we consider its characteristic equation z 3 − 1.5z 2 + 0.75z − 0.125 = 0 whose roots as
r1 = r2 = r3 = 0.5.
(a) Impulse response (k ≥ 1): h(k) = p(k)(0.5)k−1 where p(k) = A1 + (k − 1)A2 /0.5 + 0.5(k − 1)(k −
2)A3 /(0.5)2 with A1 , A2 and A3 being computed as
1
0 0 A1
0
A1
0
0.5 1 0 A2 = 0 =⇒ A2 = 0
0.25 1 1 A3
1
A3
1
and then
h(k) = (k − 1)(k − 2)(0.5)k−2 ,
k ≥ 1.
(b) Homogeneous solution (k ≥ 0): yh (k) = p(k)(0.5)k where p(k) = B0 + kB1 /0.5 + 0.5k(k − 1)B2 /(0.5)2
with B0 , B1 and B2 being computed as:
1
0 0 B0
−1
B0
−1
0.5 1 0 B1 = 2 =⇒ B1 = 2.5
0.25 1 1 B2
0.8
B2
−1.45
and then
yh (k) = −(0.5)k + 2.5k(0.5)k−1 − 1.45k(k − 1)(0.5)k−1 ,
k ≥ 0.
(c) Particular solution (k ≥ 0): yp = (uσ) ∗ h, and (uσ) = (1σ) = σ then
yp = (uσ) ∗ h = h ∗ (uσ) = h1 ∗ h2 ∗ h3 ∗ σ
where h1 (k) = h2 (k) = h3 (k) = (0.5)k−1 σ(k − 1). Since σ(0) = 1 we will first compute ȳp = h1 ∗
h2 ∗ h3 ∗ [σ]1 , where [σ]1 (k) = σ(k − 1), in order we have a convolution in the format as required in
Theorem 4.3; at the end we take yp (k) = ȳp (k + 1). Then:
ȳp (k) = q(k)h1 (k) + C4 σ(k − 1),
q(k) = C1 + (k − 1)C2 /(0.5) + 0.5(k − 1)(k − 2)C3 /(0.5)2 ,
16
k≥1
where C1 , C2 , C3 and C4 are compute as
1
0
0
0.5
1
0
2
0.5
1
1
0.53 0.75 1.5
Then
the solution of the “augmented” Vandermonde system:
C1
−8
0
1 C1
1
C2 = 0 =⇒ C2 = −4
C3 −2
1 C3 0
C4
8
1
1 C4
ȳp (k) = −8(0.5)k−1 − 4(k − 1)(0.5)k−2 − (k − 1)(k − 2)(0.5)k−3 + 8,
k≥1
and since yp (k) = ȳp (k + 1) we have
yp (k) = −8(0.5)k − 4k(0.5)k−1 − k(k − 1)(0.5)k−2 + 8,
k≥0
Finally, the solution for the IVP (55) is y = yh + yp , or
y(k) = −9(0.5)k − 1.5k(0.5)k−1 − 3.45k(k − 1)(0.5)k−1 + 8,
k≥0
Example 5.2.2. Let be the third order IVP:
y(k + 3) − 1.4y(k + 2) + 0.9y(k + 1) − 0.2y(k) = k,
y(0) = 2, y(1) = −3, y(2) = 0.5
(56)
whose characteristic polynomial is −0.2 + 0.9z 2 − 1.4z 2 + z 3 = (z − 0.4)[(z − 0.5)2 + 0.5].
√
√
(a) Impulse response (k ≥ 1): h(k) = A1 (0.4)k−1 + A2 (1/ 2)k−1 ej(k−1)π/4 + A3 (1/ 2)k−1 e−j(k−1)π/4
where A1 , A2 and A3 are computed as
1
1
1
A1
0
A1
3.8461538
0.4 0.5 + 0.5i 0.5 − 0.5i A2 = 0 =⇒ A2 = −1.9230769 − 0.3846154i
0.16
0.5i
−0.5i
1
A3
−1.9230769 + 0.3846154i
A3
and then
h(k) =
√
3.8461538(0.4)k−1 + (−1.9230769 − 0.3846154i)(1/ 2)k−1 ej(k−1)π/4 +
√
(−1.9230769 + 0.3846154i)(1/ 2)k−1 e−j(k−1)π/4
or
h(k) =
√
3.8461538(0.4)k−1 − 3.8461538(1/ 2)k−1 cos[(k − 1)π/4] +
√
0.7692308(1/ 2)k−1 sin[(k − 1)π/4], k ≥ 1
√
√
(b) Homogeneous solution (k ≥ 0): yh (k) = B0 (0.4)k + B1 (1/ 2)k ejkπ/4 + B2 (1/ 2)k e−jkπ/4 with B0 , B1
and B2 being computed as:
1
1
1
B0
2
B0
17.307692
0.4 0.5 + 0.5i 0.5 − 0.5i B1 = −3 =⇒ B1 = −7.6538462 + 2.2692308i
0.16
0.5i
−0.5i
B2
0.5
B2
−7.6538462 − 2.2692308i
and then
√
√
yh (k) = 17.307692(0.4)k − 15.307692(1/ 2)k cos(kπ/4) − 4.5384615(1/ 2)k sin(kπ/4)
(c) Particular solution (k ≥ 0): yp = (uσ) ∗ h where u(k) = k, and then we need to write kσ(k) as a sum
of convolution of signals. From Remark 4.2.1 we have that (σ ∗ σ)(k) = k + 1, then we easily get
kσ(k) = (σ ∗ σ)(k) − σ(k),
or (uσ) = (σ ∗ σ) − σ
then
yp = (uσ) ∗ h = h ∗ (uσ) = h1 ∗ h2 ∗ h3 ∗ σ ∗ σ − h1 ∗ h2 ∗ h3 ∗ σ
|
{z
} |
{z
}
yp1
yp2
where h1 (k) = (0.4)k−1 σ(k − 1), h2 (k) = (0.5 + 0.5i)k−1 σ(k − 1) and h3 (k) = (0.5 − 0.5i)k−1 σ(k − 1)
and yp1 and yp2 can be calculated as:
17
(c.1) yp1 (k) = ȳp1 (k + 2), where ȳp1 = h1 ∗ h2 ∗ h3 ∗ [σ]1 ∗ [σ]1 , or
√
√
ȳp1 (k) = C1 (0.4)k−1 + C2 (1/ 2)k−1 ej(k−1)π/4 + C3 (1/ 2)k−1 e−j(k−1)π/4 + C4 + C5 (k − 1),
with C1 , C2 , C3 , C4 and C5 computed by
1
1
1
0.4
0.5
+
0.5i
0.5
−
0.5i
0.16
0.5i
−0.5i
0.064 −0.25 + 0.25i −0.25 − 0.25i
0.0256
−0.25
−0.25
and so
ȳp1 (k) =
solving
1 0 C1
0
C1
10.683761
C2 0.7692308 − 3.8461538i
1 1
C2 0
1 2 C3 = 0 =⇒
C3 = 0.7692308 + 3.8461538i
1 3 C4
0
C4
−12.222222
1 4 C5
1
C5
3.3333333
√
10.683761(0.4)k−1 + 1.5384615(1/ 2)k−1 cos[(k − 1)π/4] +
√
7.6923077(1/ 2)k−1 sin[(k − 1)π/4] − 12.222222 + 3.3333333(k − 1)
then yp1 (k) = ȳp1 (k + 2) is given by
yp1 (k) =
√
10.683761(0.4)k+1 + 1.5384615(1/ 2)k+1 cos[(k + 1)π/4] +
√
7.6923077(1/ 2)k+1 sin[(k + 1)π/4] − 12.222222 + 3.3333333(k + 1)
(c.2) yp2 (k) = ȳp2 (k + 1), where ȳp2 = h1 ∗ h2 ∗ h3 ∗ [σ]1 , or
√
√
ȳp2 (k) = D1 (0.4)k−1 + D2 (1/ 2)k−1 ej(k−1)π/4 + D3 (1/ 2)k−1 e−j(k−1)π/4 + D4 ,
with D1 , D2 , D3 and D4 computed by solving
D1
−6.4102564
0
1
1
1
1 D1
0.4
0.5 + 0.5i
0.5 − 0.5i
1
D2 = 0 =⇒ D2 = 1.5384615 + 2.3076923i
0.16
D3
1.5384615 − 2.3076923i
0
0.5i
−0.5i
1 D3
D4
3.3333333
1
0.064 −0.25 + 0.25i −0.25 − 0.25i 1 D4
which implies
ȳp2 (k)
√
= −6.4102564(0.4)k−1 + 3.0769231(1/ 2)k−1 cos[(k − 1)π/4] −
√
4.6153846(1/ 2)k−1 sin[(k − 1)π/4] + 3.3333333
and then
√
√
yp2 (k) = −6.4102564(0.4)k + 3.0769231(1/ 2)k cos[kπ/4] − 4.6153846(1/ 2)k sin[kπ/4] + 3.3333333
Therefore yp = yp1 + yp2 is given by
yp (k) =
√
10.683761(0.4)k+1 + 1.5384615(1/ 2)k+1 cos[(k + 1)π/4] +
√
7.6923077(1/ 2)k+1 sin[(k + 1)π/4] − 12.222222 + 3.3333333(k + 1) − 6.4102564(0.4)k +
√
√
3.0769231(1/ 2)k cos[kπ/4] − 4.6153846(1/ 2)k sin[kπ/4] + 3.3333333
Finally, the solution for the IVP (55) is y = yh + yp , or
√
√
y(k) = 17.307692(0.4)k − 15.307692(1/ 2)k cos(kπ/4) − 4.5384615(1/ 2)k sin(kπ/4) +
√
10.683761(0.4)k+1 + 1.5384615(1/ 2)k+1 cos[(k + 1)π/4] +
√
7.6923077(1/ 2)k+1 sin[(k + 1)π/4] − 12.2222222 + 3.3333333(k + 1) −
√
√
6.4102564(0.4)k + 3.0769231(1/ 2)k cos[kπ/4] − 4.6153846(1/ 2)k sin[kπ/4] +
3.3333333
which, in turn, can be simplified to
√
√
y(k) = 15.17094(0.4)k − 7.6153843(1/ 2)k cos(kπ/4) − 6.076923(1/ 2)k sin(kπ/4) + 3.3333333k − 5.5555556
18
Example 5.2.3. Let be the third order IVP:
y(k + 3) + y(k + 1) = sin(kπ/2),
y(0) = 1, y(1) = y(2) = 0
(57)
whose characteristic equation is z 3 + z = z(z 2 + 1) = 0, and r1 = j, r2 = −j and r3 = 0. We discard r3 = 0
and solve a second order equation and at the end shift the solution by one unity to the right.
(a) Impulse response (k ≥ 1): h(k) = A1 (j)k−1 + A2 (−j)k−1 where A1 and A2 are computed as
1 1
A1
0
A1
−j/2
=
=⇒
=
j −j A2
1
j/2
A2
which implies h(k) = −j/2(j)k−1 + j/2(−j)k−1 or h(k) = − cos(kπ/2), k ≥ 1
(b) Homogeneous solution (k ≥ 0): yh (k) = B0 (j)k + B1 (−j)k with B0 and B1 being computed as:
1 1
0
0
B0
B0
=
=
=⇒
j −j B1
0
0
B1
and then yh (k) = 0.
(c) Particular solution (k ≥ 0): yp = (uσ) ∗ h where u(k) = sin(kπ/2) = (1/2j)(j)k − (1/2j)(−j)k . Then
yp = (1/2j) (u1 ∗ h1 ∗ h2 ) −(1/2j) (u2 ∗ h1 ∗ h2 )
|
{z
}
{z
}
|
yp1
yp2
where u1 (k) = (j)k σ(k), u2 (k) = (−j)k σ(k), h1 (k) = (j)k−1 σ(k − 1), h2 (k) = (−j)k−1 σ(k − 1) and
yp1 (k) = ȳp1 (k + 1) and yp2 (k) = ȳp2 (k + 1):
(c.1) ȳp1 (k) = p(k)h1 + C3 h2 (k), where p(k) = C1 + (k − 1)C2 /j, and
1 0
1
0
C1
1/4
C1
j
1
−j C2 = 0 =⇒ C2 = −j/2
j 2 2j (−j)2
C3
1
C3
−1/4
Then ȳp1 (k) = 1/4(j)k−1 − 1/2(k − 1)(j)k−1 − 1/4(−j)k−1 and so
yp1 (k) = ȳp1 (k + 1) = 1/4(j)k − 1/2k(j)k − 1/4(−j)k ,
k≥0
(c.2) ȳp2 (k) = D1 h1 + q(k)h2 (k), where q(k) = D2 + (k − 1)D3 /(−j), and
1
1
0
D1
0
D1
−1/4
j
−j
1 D2 = 0 =⇒ D2 = 1/4
2
2
j
(−j) −2j
D3
1
D3
j/2
Then ȳp2 (k) = −1/4(j)k−1 + 1/4(−j)k−1 − 1/2(k − 1)(−j)k−1 and so
yp2 (k) = ȳp2 (k + 1) = −1/4(j)k + 1/4(−j)k − 1/2k(−j)k ,
k≥0
Now, since yp = (1/2j)yp1 − (1/2j)yp2 , we have
yp (k) = (1/2j)[1/4(j)k − 1/2k(j)k − 1/4(−j)k ] − (1/2j)[−1/4(j)k + 1/4(−j)k − 1/2k(−j)k ]
which can (optionally) be simplified to yp (k) = (1/2) sin(kπ/2) − (k/2) sin(kπ/2). Finally, to contemplate the zero root of the characteristic equation and the initial condition y(0) = 1, we have:
y(k) = 2δ(k) + (1/2) sin[(k − 1)π/2] − [(k − 1)/2] sin[(k − 1)π/2],
or
y(k) = 2δ(k) + (k/2) cos(kπ/2) − cos(kπ/2),
19
k≥0
k≥0
6
Conclusions
We showed in this paper a technique for computing the convolution of exponential signals, in analog and
discrete time context, that avoids the resolution of integrals and summations. The method is essentially
algebraic and requires the resolution of Vandermonde systems, which is a well-known and extensively discussed problem in literature (see e.g. [10, 11] and references therein). While the question of computing
convolution of exponentials have been discussed previously in literature ([6, 7]), the proposed approach is
apparently different from the previous ones, and additionally is quite simple and suitable to be implemented
computationally. Finally, we use the proposed approach to solve a n order differential/difference equation
with constant coefficients.
References
[1] G. Grubb. Distributions and Operators. Graduate Texts in Mathematics (Book 252). Springer Science+Business Media, LLC, NY, 2009.
[2] Wikipedia: The Free Encyclopedia. Wikimedia Foundation, Inc. 7 June 2016. Web. 21 June, 2016.
Available at http://en.wikipedia.org/wiki/Convolution.
[3] F. Mota. Signals, Systems and Transforms. Lecture Notes, 2015.
[4] J. D’Azzo and C. Houpis. Linear Control System Analysis and Design. Second Edition. MacGraw-Hill
Kogakusha, Ltd., 1981.
[5] Scilab Enterprises. Scilab: Free and Open Source Software for Numerical Computation. Orsay, France,
2012. Available at http://www.scilab.org.
[6] M. Akkouchi. On The Convolution of Exponential Distributions. Journal of The Chungcheong Mathematical Society, Vol. 21, No. 4, December 2008.
[7] N.-Y. Ma, F. Liu. A Novel Analytical Scheme to Compute the n-Fold Convolution of Exponential-Sum
Distribution Functions. Applied Mathematics and Computation, 158 (2004) 225–235.
[8] K. L. Chung. Elementary Probability Theory with Stochastic Processes. Third Edition. Springer-Verlag
NY Inc., 1979.
[9] Wikipedia: The Free Encyclopedia. Wikimedia Foundation, Inc. 4 June 2016. Web. 21 June, 2016.
Available at http://en.wikipedia.org/wiki/Convolution power.
[10] G. H. Golub and C. F. Van Loan. Matrix Computations. Second Edition. The Johns Hopkins Univ.
Press, 1989.
[11] S.-H. Hou and W.-K. Pang. Inversion of Confluent Vandermonde Matrices. Computers and Mathematics
with Applications 43 (2002) 1539-1547.
20
| 3cs.SY
|
arXiv:1711.06467v1 [cs.PL] 17 Nov 2017
W YS? : A Verified Language Extension for Secure Multi-party Computations
Aseem Rastogi
Nikhil Swamy
Michael Hicks
University of Maryland, Microsoft Research
Microsoft Research
University of Maryland
Abstract—Secure multi-party computation (MPC) enables a
set of mutually distrusting parties to cooperatively compute,
using a cryptographic protocol, a function over their private data. This paper presents W YS? , a new domain-specific
language (DSL) implementation for writing MPCs. W YS?
is a Verified, Domain-Specific Integrated Language Extension
(VDSILE), a new kind of embedded DSL hosted in F? , a fullfeatured, verification-oriented programming language. W YS?
source programs are essentially F? programs written against
an MPC library, meaning that programmers can use F? ’s
logic to verify the correctness and security properties of their
programs. To reason about the distributed semantics of these
programs, we formalize a deep embedding of W YS? , also in
F? . We mechanize the necessary metatheory to prove that the
properties verified for the W YS? source programs carry over
to the distributed, multi-party semantics. Finally, we use F? ’s
extraction mechanism to extract an interpreter that we have
proved matches this semantics, yielding a verified implementation. W YS? is the first DSL to enable formal verification of
source MPC programs, and also the first MPC DSL to provide
a verified implementation. With W YS? we have implemented
several MPC protocols, including private set intersection, joint
median, and an MPC-based card dealing application, and have
verified their security and correctness.
1. Introduction
Secure multi-party computation (MPC) is a framework
that enables two or more parties to compute a function f
over their private inputs x1 , ..., xn so that no party sees
any of the others’ inputs, but rather only sees the output
f (x1 , ..., xn ). Utilizing a trusted third party to compute
f would achieve this goal, but in fact we can achieve it
using one of a variety of cryptographic protocols carried
out only among the participants [1]–[4]. One example use
of MPC is private set intersection (PSI): the xi could be
individuals’ personal interests, and the function f computes
their intersection, revealing which interests the group has in
common, but not any interests that they don’t. Among other
applications, MPC has been used for auctions [5], detecting
tax fraud [6], managing supply chains [7], and performing
privacy preserving statistical analysis [8].
Typically, cryptographic protocols expect f to be specified as a boolean or arithmetic circuit. Programming directly
with circuits and cryptography via a host-language API
is painful, so starting with the Fairplay project [9] many
researchers have designed higher-level domain-specific languages (DSLs) in which to program MPCs [10]–[21]. These
DSLs compile source code to circuits which are then given
to the underlying protocol. While doing this undoubtedly
makes it easier to program MPCs, these languages still have
several drawbacks regarding both security and usability.
First, MPC participants should be able to reason that
f is sufficiently privacy preserving, i.e., that its output will
not reveal too much information about the inputs [22]. The
goal of an MPC DSL is secure computations, and such
reasoning gives assurance that this goal is being achieved.1
Yet, only a few DSLs (Sharemind DSL [21], Wysteria [19],
and SCVM [20]) have a mathematical semantics that can
serve as a basis for formal reasoning.
Second, those languages that do have a semantics lack
support for mechanized reasoning about MPC programs:
only by-hand proofs are possible, which provide less assurance than machine-checked proofs. A middle ground might
be a mechanization of the semantics and its metatheory [23]
(e.g., in a system like Coq, Agda, or Isabelle), which adds
greater assurance that it is correct [24], but so far no MPC
DSL has even had a mechanized semantics.
Third, there is a gap between the semantics, if there is
one, and the actual implementation. Within that gap is the
potential for security holes. Formal verification of the MPC
DSL’s toolchain can significantly reduce the occurrence of
security-threatening bugs [25]–[29], but no existing MPC
DSL implementation has been (even partially) formally
verified—this should not be surprising since, as mentioned
earlier, these DSLs have lacked a formal semantics on which
to base a verification effort.
Finally, there is the practical problem that existing DSLs
do not scale up, because they lack the infrastructure of a
full-featured language. Adding more features (to both the
language and the formalization) would help, but doing so
quickly becomes unwieldy and frustrating, especially when
the added features are “standard” and do not have much to
do with MPC. We want access to libraries and frameworks
for I/O, GUIs, etc. in a way that easily adds to functionality
without adding complexity or compromising security.
This paper presents W YS? , a new MPC DSL that addresses these problems. Unlike most previous MPC DSLs,
1. Our attacker model is the “honest-but-curious” model where the
attackers are the participants in the protocol themselves. That is, we assume
that the participants in the protocol play their roles faithfully, but they are
motivated to deduce as much as they can about the other participants’
secrets by observing the protocol.
W YS? is not a standalone language, but is rather what
we call a Verified, Domain-Specific Integrated Language
Extension (VDSILE), a new kind of embedded DSL that can
be hosted by F? [30], a full-featured, verification-oriented
programming language.
W YS? has the following distinguishing elements:
• Integrated language extension (Section 3). Programmers can write W YS? MPC source programs in what is essentially an extended dialect of F? . W YS? inherits the basic
programming model from Wysteria [19] (see Section 2 for
comparison with Wysteria). Like so-called shallow domainspecific language embeddings, W YS? embeds the Wysteriaspecific combinators in normal F? syntax, with prescriptions
on their correct use expressed with F? ’s dependent type-andeffect system. This arrangement has two benefits. Firstly,
W YS? programs can use, with no extra effort, standard
constructs such as datatypes and libraries directly from F? .
Secondly, programmers can formally verify properties, such
as those related to correctness and security, about their MPC
program using F? ’s semi-automated verification facilities.
W YS? is the first DSL to enable formal verification of source
MPC programs.
• Deep embedding of domain-specific semantics (Section 4). A shallow embedding implements the semantics of
a DSL using the abstraction facilities of the host language,
e.g., as a kind of library. However, for W YS? this is impossible because its core semantics cannot be directly encoded in
F? ’s semantics. A W YS? program is like a SIMD program
whose execution alternates between in-parallel computations
at each party, privately, and joint computations involving all
parties, securely. While such a program can be conceptually
viewed as having a single thread of control, it is not directly
implemented that way. As such, we take the approach of a
typical deep embedding: We define an interpreter in F? that
operates over W YS? abstract syntax trees (ASTs), defined
as an F? data type; these trees are produced by running
the F? compiler (in a special mode) on the extended source
program. Importantly, the W YS? AST (and hence the interpreter) does not “bake in” standard F? constructs like
numbers and lists; rather, inherited language features appear
abstractly in the AST, and their semantics is handled by a
novel foreign function interface (FFI) that is easy to use,
both when programming and when verifying.
• Verified implementation, modulo a trusted crypto library (Sections 4 and 5). Within F? , we mechanize two
operational semantics for deep-embedded W YS? ASTs: a
conceptual single-threaded semantics that formalizes the
SIMD view and a distributed semantics that formalizes the
actual multi-party runs of the programs. We prove that
(a) the conceptual single-threaded semantics is sound with
respect to the actual distributed semantics (including the
semantics of the FFI calls), and (b) the distributed semantics
is correctly implemented by our interpreter. These proofs
are checked by F? ’s proof checking algorithm. As a result,
we have verified that the properties formally proven about
the W YS? source programs carry over when these programs
are run by multiple parties in a distributed manner. There
is an important caveat though: Our interpreter makes use
of a circuit library to compile ASTs to circuits and then
execute them using the Goldreich, Micali and Wigderson
(GMW) multi-party computation protocol [3], but at present
this library is not formally verified. Formal verification of
GMW (which is, at present, an open problem) would add
even greater assurance.
Using W YS? we have implemented several programs,
including PSI, joint median, and an MPC-based card dealing
application (Section 6). For PSI and joint median we implement two versions: a straightforward one and a version that
composes several small MPCs, which improves performance
but increases the number of visible outputs. We formally
prove, for both PSI and median, that the optimized and
unoptimized versions are equivalent, both functionally and
with respect to the privacy of parties’ inputs. In particular,
W YS? enhances the Wysteria target semantics to “instrument” it with a trace of observations, and we prove that
the visible events in the optimized versions’ traces provide
neither participant with any additional information about
the other’s secrets. Performance experiments confirm that
the optimized versions do indeed perform better. Our card
dealing application relies on W YS? ’s support for secret
shares [31]. We formally prove that our card dealing algorithm always deals a fresh card.
In summary, this paper’s main contribution is W YS? ,
a new Verified, Domain-Specific Integrated Language Extension (VDSILE) for supporting secure multiparty computation. W YS? is unique in its use of formal methods
to ensure that both its underlying implementation and the
programs written in it behave according to important correctness and security properties. The W YS? implementation,
example programs, and proofs are publicly available online
on github, https://github.com/FStarLang/FStar/tree/stratified last/
examples/wysteria.
2. Related work
W YS? ’s computational model is based on programming
abstractions of a previous domain-specific language, Wysteria [19]. But W YS? offers several new contributions. First,
while Wysteria is a standalone language, W YS? is implemented as an extension to the language F? . As such, W YS?
programs can freely use datatypes and libraries from the host
language via a novel verification-friendly FFI mechanism
(outlined in Section 5.2). This architecture allows programs
to scale more easily (rather than requiring constant extension
and reimplementation of the standalone language, and its
verification results). Second, relying on F? ’s verification
features, the W YS? implementation has also been verified,
except for its cryptographic libraries. Third, since W YS?
is also an embedded DSL in F? , MPC programs can be
formally verified to satisfy correctness and security properties. In support of security verification, W YS? ’s semantics
includes a notion of observable traces, which we use to state
and prove information flow properties. Although it is our
point of departure, Wysteria enjoys none of these benefits.
MPC DSLs and DSL extensions. In addition to Wysteria,
several other MPC DSLs have been proposed in the literature [10]–[21]. Most of these languages have standalone
implementations, and the drawbacks that come with these.
Like W YS? , a few are implemented as language extensions.
Launchbury et al. [32] describe a Haskell-embedded DSL
for writing low-level “share protocols” on a multi-server
“SMC machine”. OblivC [33] is an extension to C for twoparty MPC that annotates variables and conditionals with
an obliv qualifier to identify private inputs; these programs
are compiled by source-to-source translation. The former is
essentially a shallow embedding, and the latter is compilerbased; W YS? is unique in its use of the VDSILE strategy,
that combines a shallow embedding to support source program verification and a deep embedding to support a nonstandard target semantics.
Mechanized metatheory. Our verification results are different from a typical verification result that might either
“mechanize metatheory” using a proof assistant for an
idealized language [23], or might prove an interpreter or
compiler correct w.r.t. a formal semantics [26]—we do both.
We mechanize the metatheory of W YS? establishing the
soundness of the conceptual single-threaded semantics w.r.t.
the actual distributed semantics, and we also mechanize
the proof that the W YS? interpreter implements the correct
W YS? semantics.
Source MPC verification. While the verification of the underlying crypto protocols has received some attention [34],
the verification of the MPC source programs has remained
largely unexplored. The only previous work that we know
of is Backes et. al. [35] who devise an applied pi-calculus
based abstraction for MPC, and use it for formal verification.
For an auction protocol that computes the min function, their
abstraction comprises about 1400 lines of code. W YS? , on
the other hand, enables direct verification of the higher-level
MPC source programs, and in addition provides a verified
toolchain.
General DSL implementation strategies. DSLs (for MPC
or other purposes) are implemented in various ways, such
as by developing a standalone compiler/interpreter, or by
embedding the DSL (shallowly or deeply) in a host language. VDSILE’s language-integrated syntax bears relation
to the approach taken in LINQ [36], which embeds a query
language in normal C# programs, and implements these
programs by extracting the query syntax tree and passing it
to a provider to implement for a particular backend. Other
researchers have embedded DSLs in verification-oriented
host languages (e.g., Bedrock [37] in Coq [38]) to permit
formal proofs of DSL programs. F? provides some advantage as a host language since it is both higher-order and
effectful, making it easier to write DSL combinators for
effectful languages while still proving that DSL programs
have good properties, and being able to (easily) extract those
programs to runnable code.
Figure 1. Architecture of an W YS? deployment
In sum, W YS? is the first DSL to enable formal verification of efficient source MPC programs as written in a
full-featured host programming language. W YS? is also the
first MPC DSL to provide a (partially) verified interpreter.
3. Verified programming in W YS?
Consider a dating application that enables its users to
compute their common interests without revealing all their
private interests to one another. This is an instance of the
private set intersection (PSI) problem. We illustrate the
main concepts of W YS? by showing, in several stages, how
to program, optimize, verify and deploy this application—
Figure 1 provides an overview.
3.1. Secure computations with
as sec
In W YS? , an MPC is written as a single specification
which executes in one of two computation modes. The
primary mode is called sec mode, and it specifies a secure
computation to be carried out among multiple parties. Here
is the private set intersection example written in W YS? :
let psi a b input a input b l a l b =
as sec {a,b} (fun () →
let r = List.intersect (reveal input a) (reveal input b) l a l b
give a r ++ give b r)
The six arguments to psi are, respectively, principal identifiers for Alice and Bob, Alice and Bob’s secret inputs,
expressed as lists, and their (public) lengths. The as sec ps f
construct indicates that thunk f should be run in sec mode.
In this mode, the code may jointly access the secrets of the
principals ps. In this case, we jointly intersect input a and
input b, the inputs of a and b, and then return the same result
r to both a and b. Outside of sec mode, Alice would not be
permitted to see Bob’s secret input, and vice versa, but inside
both can be made visible using the reveal coercion. Finally,
the code constructs a map, associating a result for each
principal (which in this case is the same)—give p v builds
a singleton map p 7→ v and ++ concatenates disjoint maps.
Running this code requires the following steps. First,
we run the F? compiler in a special mode that extracts the
above code, psi.fst, into the W YS? AST as a data structure
in psi.ml. W YS? has only a few constructs of its own, like
as sec (the full syntax is in Figure 2 in Section 4), and
these are extracted to Wysteria-specific nodes. The rest of
a program’s code is extracted into FFI nodes that indicate
the use of, or calls into, functionality provided by F? itself.
The next step is for each party, Alice and Bob, to run
the extracted program using the W YS? interpreter. This
interpreter is written in F? and provably implements a deep
embedding of the F? semantics, also specified in F? (shown
in Figures 4, 5, and 6 in Section 4). This interpreter is
extracted to OCaml code by a standard F? process. When
each party reaches as sec ps f, the interpreter’s back-end
compiles f, on-the-fly, for particular values of the secrets
in f’s environment, to a boolean circuit. First-order, loopfree code can be compiled to a circuit; W YS? provides
specialized support for several common combinators (e.g.,
List.intersect, List.mem, List.nth etc.). In the current example,
the lengths of the input lists are required to be public in
order for Alice and Bob to be able to create boolean circuits.
The circuit is handed to a library by Choi et al. [39] that
implements the GMW [3] multi-party computation protocol.
Running the protocol at each party starts by confirming that
they wish to run the same circuit, and then proceeds by
generating (XOR-based) secret shares [31] for each party’s
secret inputs. Running the GMW protocol involves evaluating the boolean circuit for f over the secret shares, involving
communication between the parties for each AND-gate.
One obvious question is how both parties are able to
get this process off the ground, running this program of six
inputs, when only five of the inputs are known to them (the
principals, their own inputs, and the size of other party’s
input). In W YS? , values specific to each principal are sealed
with the principal’s name (which appears in the sealed container’s type). As such, the types of input a and input b are,
respectively, list (sealed {a} int) and list (sealed {b} int). When
the program is run on Alice’s host, the former will be a list
of l a of Alice’s values, whereas the latter will be a list of
l b garbage values (which we denote as •). The reverse will
be true on Bob’s host. When the circuit is constructed, each
principal links their non-garbage values to the relevant input
wires of the circuit. Likewise, the output map component of
each party is derived from their output wires in the circuit,
thus, each party only gets to see their own output.
We would like MPC’s like psi to be called from normal
F? programs. For example, we would like the logic for a
dating application, which involves reading inputs, displaying
results, etc. to be able to call into psi to compute common
interests. To achieve this, W YS? provides a way to compute
a “single-party projection” of multi-party functions, i.e., a
version of psi that can be called with just a single party’s
inputs. The other party’s inputs are filled in with sealed
garbage values, as described above. Calling this function
from F? code also kicks off the W YS? interpreter, so that
it can run psi as described above. When the interpreter
completes, the result is returned and the F? program can
continue.
3.2. Optimizing PSI with
as par
Although psi gets the job done, it turns out to be
inefficient for some cases (as shown in §6). Better implementations of PSI for such cases involve performing a
mixed-mode computation, where each participant evaluates
some local computations in parallel (e.g., iterating over
the elements of their sets) interleaved with small amounts
of jointly evaluated, cryptographically secure computations.
W YS? ’s second computation mode, called par mode, supports such mixed-mode computation. In particular, the construct as par ps f states that each principal in ps should
locally execute the thunk f, simultaneously (any principal
not in the set ps simply skips the expression). Within f,
principals may engage in secure computations via as sec.
Below is an optimized version of PSI, based on an
algorithm by [40], which uses as par. The function psi opt
(line 12) begins by using as par involving Alice and Bob. In
the provided thunk, each principal calls for each alice la lb,
which in turn calls check each bob a lb, for each element a
of Alice’s list la. Secure computation occurs at the use of
as sec at line 8. Within the circuit, Alice and Bob securely
compare their values ax and bx, and gather a list (list bool).
There is one outer list for each of Alice’s elements, the
ith inner list contains comparisons of Alice’s ith value
with some of Bob’s values—rather than comparing each of
Alice’s elements with all of Bob’s, the code is optimized
(as described below) to omit redundant comparisons. At
line 13, both parties build a matrix of comparisons from the
boolean lists. Alice inspects the rows of the matrix (line 14)
to determine which of her elements are in the intersection;
Bob inspects the columns (line 15); and the joint function
gives a result to each principal (line 16).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
let rec for each alice a b la lb =
if la=[] then []
else let lb, r = check each bob a b (List.hd la) lb in
r::for each alice a b (List.tl la) lb
and check each bob a b ax lb =
if lb=[] then [], []
else let bx = List.hd lb in
let r = as sec {a,b} (fun () → reveal ax = reveal bx) in
if r then List.tl lb, [r]
else let lb’, r’ = check each bob a b ax (List.tl lb) in
bx::lb’, r::r’
let psi opt a b la lb = as par {a,b} (fun () →
let bs = build matrix (for each alice la lb) in
let ia = as par {a} (fun () → filteri (contains true ◦ row bs) la) in
let ib = as par {b} (fun () → filteri (contains true ◦ col bs) lb) in
give a ia ++ give b ib)
The optimizations are at line 9. Once we detect that
element ax is in the intersection, we return immediately
instead of comparing ax against the remaining elements of
lb. Furthermore, we remove bx from lb, excluding it from
any future comparisons with other elements of Alice’s set
la. Since la and lb are representations of sets (no repeats),
all the excluded comparisons are guaranteed to be false.
One might wonder whether we could have programmed
most of this code in normal F? , relying on just sec mode for
the circuit evaluation. However, recalling that our goal is to
formally reason about the code and prove it correct and secure, par mode provides significant benefits. In particular, the
SIMD model provided by W YS? enables us to capture many
invariants for free. For example, proving the correctness
of psi opt requires reasoning that both participants iterate
their loops in lock step—W YS? assures this by construction.
Besides, the code would be harder to write (and read) if it
were split across multiple functions or files. As a general
guideline, we use F? for code written from the view of
a single principal, and W YS? when programming for all
principals at once, and rely on the FFI to mediate between
the two.
3.3. Embedding a type system for W YS? in F?
Using the abstractions provided by Wysteria, designing
various high-level, multi-party computation protocols is relatively easy. However, before deploying such protocols, three
important questions arise.
1)
2)
3)
Is the protocol realizable? For example, does a
computation that is claimed to be executed only
by some principals ps (e.g., using an as par ps or
an as sec ps) only ever access data belonging to ps?
Does the protocol correctly implement the desired
functionality? For example, does it correctly compute the intersection of Alice and Bob’s sets?
Is the protocol secure? For example, do the optimizations of the previous section that omit certain
comparisons inadvertently also release information
besides the final answer?
?
?
By embedding W YS in F and leveraging its type system, we address each of these three questions. Our strategy
is to make use of F? ’s extensible, monadic dependent typeand-effect system to define a new indexed monad (called
Wys) and use it to describe precise trace properties of Wysteria multi-party computations. Additionally, we make use of
an abstract type, sealed ps t, representing a value accessible
only to the principals in ps. Combining the Wys monad
with the sealed type, we encode a form of information-flow
control to ensure that protocols are realizable.
The Wys monad. The Wys monad provides several features.
First, all DSL code is typed in this monad, encapsulating
it from the rest of F? . Within the monad, computations and
their specifications can make use of two kinds of ghost state:
modes and traces. The mode of a computation indicates
whether the computation is running in an as par or in an
as sec context. The trace of a computation records the
sequence and nesting structure of messages exchanged between parties as they jointly execute as sec expressions—the
result of a computation and its trace constitute its observable
behavior. The Wys monad is, in essence, the product of a
reader monad on modes and a writer monad on traces.
Formally, we define the following types for modes and
traces. A mode Mode m ps is pair of a mode tag (either Par
or Sec) and a set of principals ps. A trace is a forest of trace
element (telt) trees. The leaves of the trees record messages
TMsg x that are received as the result of executing an as sec
block. The tree structure represented by the TScope ps t
nodes record the set of principals that are able to observe
the messages in the trace t.
type mtag = Par | Sec
type mode = Mode: m:mtag → ps:prins → mode
type telt =
| TMsg : x:α → telt
| TScope: ps:prins → t:list telt → telt
type trace = list telt
Every W YS? computation e has a monadic computation
type Wys t pre post. The type indicates that e is in the Wys
monad (so it may perform multi-party computations); t is
its result type; pre is a pre-condition on the mode in which
e may be executed; and post is a post-condition relating
the computation’s mode, its result value, and its trace of
observable events. When run in a context with mode m
satisfying the pre-condition predicate pre m, e may send
and receive message according to some trace tr, and if and
when it returns, the result is a t-typed value v validating the
post-condition predicate post m v tr. The style of indexing a
monad with a computation’s pre- and post-condition is a
standard technique [30], [41], [42]—we defer the definition
of the monad’s bind and return to the actual implementation
and focus instead on specifications of combinators specific
to W YS? .
We now describe two of the Wysteria-specific combinators in W YS? , as sec and reveal, and how we give them
types in F? .
Defining
as sec
in W YS? .
1 val as sec: ps:prins → f:(unit → Wys a pre post) → Wys a
2 (requires (fun m → m=Mode Par ps ∧ pre (Mode Sec ps)))
3 (ensures (fun m r tr → tr=[TMsg r] ∧ post (Mode Sec ps) r [])))
The type of as sec is dependent on the first parameter,
Its second argument f is the thunk to be evaluated
in as sec mode. The result’s computation type has the
form Wys a (requires φ) (ensures ψ), for some pre-condition
and post-condition predicates φ and ψ , respectively. The
free variables in the type (a, pre and post) are implicitly
universally quantified (at the front); we use the requires and
ensures keywords for readability—they are not semantically
significant.
The pre-condition of as sec is a predicate on the mode
m of the computation in whose context as sec ps f is called.
For all the ps to jointly execute f, we require all of them
to transition to perform the as sec ps f call simultaneously,
i.e., the current mode must be Mode Par ps. We also require
the pre-condition pre of f to be valid once the mode has
transitioned to Mode Sec ps—line 2 says just this.
The post-condition of as sec is a predicate relating the
initial mode m, the result r:a, and the trace tr of the computation. Line 3 states that the trace of a secure computation
as sec ps f is just a singleton [TMsg r], reflecting that its execution reveals only result r.2 Additionally, it ensures that the
result r is related to the mode in which f is run (Mode Sec ps)
ps.
2. This is the “ideal functionality” ensured by the backend, e.g., GMW.
and the empty trace [] (since f has no observables) according
to post, the post-condition of f.
Defining reveal in W YS? . As discussed earlier, a value v of
type sealed ps t encapsulates a t value that can be accessed by
calling reveal v. This call should only succeed under certain
circumstances. For example, in par mode, Bob should not
be able to reveal a value of type sealed {Alice} int. The type
of reveal makes the access control rules clear:
val unseal: #ps:prins → sealed ps α → Ghost α
val reveal: #ps:prins → x:sealed ps α → Wys α
(requires (fun m → m.mode=Par =⇒ m.ps ⊆ ps
∧ m.mode=Sec =⇒ m.ps ∩ ps =
6 ∅))
(ensures (fun m r tr → r=unseal x ∧ tr=[]))
The unseal function is a Ghost function, meaning that it
can only be used in specifications for reasoning purposes.
On the other hand, reveal can be called in the concrete
W YS? programs. Its precondition says that when executing
in Mode Par ps’, all current participants must be listed in the
seal, i.e., ps’ ⊆ ps. However, when executing in Mode Sec ps’,
only a subset of current participants is required: ps’ ∩ ps 6= ∅.
This is because the secure computation is executed jointly
by all of ps’, so it can access any of their individual data. The
postcondition of reveal relates the result r to the argument x
using the unseal function.
3.4. Correctness and security verification
Using the Wys monad and the sealed type, we can write
down precise types for our psi program, proving various useful properties. For lack of space, we discuss the statements
of the main lemmas we prove and the proof structure—
the details of their machine-checked proofs are left to the
actual implementation. By programming the protocols using
the high-level abstractions provided by W YS? , our proofs
are relatively straightforward. In particular, we rely heavily
on the view that both parties execute (different fragments
of) the same code. In contrast, reasoning directly against
the low-level message passing semantics would be much
more unwieldy. In Section 4, by formalizing the connection
between the high- and low-level semantics, we justify our
source-level reasoning.
We present the structure of the security and correctness
proof for psi opt by showing the top-level specification for
psi opt:
val psi opt: a:prin → b:prin
→ la:list (sealed {a} int) → lb:list (sealed {b} int)
→ Wys (map {a, b} (list int))
(requires (fun m → m=Mode Par {a, b}
∧ no dups la ∧ no dups lb))
(ensures (fun m r tr →
let ia = as set (Map.get r a) in
let ib = as set (Map.get r b) in
ia = ib ∧ ia = (as set la ∩ as set lb) ∧ tr=psi opt trace la lb))
The signature above establishes that when Alice and Bob
simultaneously execute psi opt (they start together in Par
mode), with lists la and lb containing their secrets (without
any duplicates), then if and when the protocol terminates,
they both obtain that same results ia and ib corresponding to
the intersection of their sets, i.e., the protocol is functionally
correct.
To prove properties beyond functional correctness, we
also prove that the trace of observable events from a run of
psi la lb is described by the function psi opt trace la lb. This is
a purely specificational function that, in effect, records each
of the boolean results of every as sec comparison performed
during a run of psi—it has the same structure as for each alice
and check each bob.
Given a full characterization of the observable behavior
of psi opt trace la lb in terms of its inputs, we can prove
optimizations correct using relational reasoning [43] and
we can also prove security hyperproperties [44] by relating
traces from multiple runs of the protocol.
Our goal is to prove a noninterference with delimited
release [45] property for psi opt. Since our attacker model is
the “honest-but-curious” model (as mentioned in Section 1),
we do not aim to prove security properties against a thirdparty network adversary.
For psi, from the perspective of Alice as the attacker,
we aim prove that for two runs of the protocol in which
Alice’s input is constant but Bob’s varies, Alice learns no
more by observing the the protocol trace than what she is
allowed to. Covering Bob’s perspective symmetrically, we
show that in two runs of psi la0 lb0 and psi la1 lb1 that satisfy
formula Ψ below, the traces observed by Alice and Bob are
indistinguishable, up to permutation, where la0 , la1 , lb0 , lb1
have type lset int, the type of integer sets represented as lists.
Ψ la0 la1 lb0 lb1 =
intersect la0 lb0 = intersect la1 lb1
∧ length la0 = length la1 ∧ length lb0 = length lb1
In other words, Alice and Bob learn no more than the
intersection of their sets and the size of the other’s set; Ψ
is the predicate that delimits the information released by
the protocol. As far as we are aware, this is the first formal
proof of correctness and security of Huang et al.’s optimized,
private set-intersection protocol.3
The proof is in the style of a step-wise refinement, via
psi, an inefficient variant of the psi opt program. Running
psi la lb always involves doing exactly length la ∗ length lb
comparisons in two nested loops. We prove the following relational security property for psi, relating the traces
trace psi la0 lb0 and trace psi la1 lb1 —the formal statement of
the lemma we prove in F? is shown below.
val psi is secure: la0 : → lb0 : → la1 : → lb1 : → Lemma
(requires (Ψ la0 la1 lb0 lb1 ))
(ensures (permutation (trace psi la0 lb0 )
(trace psi la1 lb1 )))
We reason about the traces of psi only up to permutation.
Given that Alice has no prior knowledge of the choice of
representation of Bob’s set (Bob can shuffle his list), the
3. In carrying out this proof, it becomes evident that Alice and Bob
learn the size of each other’s sets. One can compose psi opt with other
protocols to partially hide the size—W YS? makes it easy to compose
protocols simply by composing their functions.
Principal
Principal set
FFI const
Constant
Expression
p
s
c, f
c
e
Map m
Value v
::=
::=
|
|
|
|
|
p | s | () | true | false | c
as par e1 e2 | as sec e1 e2
seal e1 e2 | reveal e | ffi f ē
mkmap e1 e2 | project e1 e2
concat e1 e2
c | x | let x = e1 in e2 | λx.e | e1 e2
fix f.λx.e | if e1 then e2 else e3
Figure 2. W YS? syntax
traces Alice observes are equivalent up to permutation—
we can formalize this observation using a probabilistic,
relational variant of F? [46], but have yet to do so.
As a next step, we prove that optimizing psi to psi opt
is secure by showing that there exists a function f, such
that for any trace tr=trace psi la lb, the trace of psi opt,
trace psi opt la lb, can be computed by f (length la) tr. In other
words, the trace produced psi opt la lb can be computed using
a function of information already available to Alice (or Bob)
when she (or he) observes a run of the secure, unoptimized
version psi la lb. As such, the optimizations do not reveal
further information.
We present more examples and their verification details
in Section 6.
4. Formalizing W YS?
In the previous section, we presented examples of verifying properties about W YS? programs using F? ’s logic.
However, these programs are not executed using the F?
(single-threaded) semantics; they have a distributed semantics carried out by multiple parties. So, how do the properties
that we verify using F? carry over to the actual runs?
In this section, we present the metatheory that answers
this question. First, we formalize the W YS? single-threaded
(ST) semantics, arguing that it faithfully realizes the F?
semantics, including the W YS? API presented in Section 3.
Next, we formalize the distributed (DS) semantics that the
multiple parties use to run W YS? programs. Our theorems
establish the correspondence between the two semantics,
thereby ensuring that the properties that we verify using F?
carry over to the actual protocol runs. We have mechanized
all the metatheory presented in this section in F? .
4.1. Syntax
Figure 2 shows the complete syntax of W YS? . Principal
and principal sets are first-class values, and are denoted by p
and s respectively. Constants in the language also include ()
(unit), booleans, and FFI constants c. Expressions e include
the regular forms for functions, applications, let bindings,
etc. and the W YS? -specific constructs. Among the ones that
we have not seen in Section 3, expression mkmap e1 e2
Mode
Context
Frame
Stack
Environment
Trace element
Trace
Configuration
M
E
F
X
L
t
T
C
Par component P
Sec component S
Protocol π
::=
::=
|
|
::=
::=
::=
::=
::=
::=
::=
::=
· | m[p 7→ v]
p | s | () | true | false | sealed s v
m | v | (L, λx.e) | (L, fix f.λx.e)
•
Par s | Sec s
hi | as par hi e | as par v hi | . . .
(M, L, E, T )
· | F, X
· | L[x 7→ v]
TMsg v | TScope s T
· | t, T
M ; X; L; T ; e
::= · | P [p 7→ C]
::= · | S[s 7→ C]
::= P ; S
Figure 3. Runtime configuration syntax
creates a map from principals in e1 (which is a principal
set) to the value computed by e2 . project e1 e2 projects the
value of principal e1 from the map e2 , and concat e1 e2
concatenates the two maps.
Host language (i.e., F? ) constructs are also part of the
syntax of W YS? , including constants c include strings,
integers, lists, tuples, etc. Likewise, host language functions/primitives can be called from W YS? —ffi f ē is the
invocation of a host-language function f with arguments ē.
The FFI confers two benefits. First, it simplifies the core
language while still allowing full consideration of security
relevant properties. Second, it helps the language scale by
incorporating many of the standard features, libraries, etc.
from the host language.
4.2. Single-threaded semantics
The ST semantics is a model of the F? semantics and the
W YS? API. The ST semantics defines a judgment C → C 0
that represents a single step of an abstract machine. Here,
C is a configuration M ; X; L; T ; e. This five-tuple consists
of a mode M , a stack X , an environment L, a trace T ,
and an expression e. The syntax for these elements is given
in Figure 3. The value form v represents the host language
(FFI) values. The stack and environment are standard; trace
T and mode M were discussed in the previous section.
The ST semantics is formalized in the style of Hieb
and Felleisen [47], where the redex is chosen by (standard)
evaluation contexts E , which prescribe left-to-right, callby-value evaluation order. A few of the core rules are
given in Figure 4. In essence, the semantics extends a
standard reduction machinery for a call-by-value, lambda
calculus (in direct correspondence with a pure fragment of
F? ), with several Wysteria-specific constructs. We argue, by
inspection, that the Wysteria-specific constructs are in 1-1
correspondence with their specifications in the Wys monad.
Despite the “eyeball closeness”, there is room for formal
discrepancy between the ST semantics and its static model
S- LET
X1 = (M ; L; let x = hi in e2 ; T ), X
M ; X; L; T ; let x = e1 in e2 → M ; X1 ; L; ·; e1
M ; X; L; T ; (L1 , λx.e) e1 → M ; X; L1 [x 7→ e1 ]; T ; e
S- APP
S- ASPARRET
S- ASPAR
M = Par s1 s ⊆ s1 X1 = (M ; L; seal s hi; T ), X
M ; X; L; T ; as par s (L1 , λx.e) → Par s; X1 ; L1 [x 7→ ()]; ·; e
can seal s v
X = (M1 ; L1 ; seal s hi; T1 ), X1 T2 = append T1 [TScope s T ]
M ; X; L; T ; v → M1 ; X1 ; L1 ; T2 ; sealed s v
S- ASSECRET
S- ASSEC
M = Par s X1 = (M ; L; hi T ), X
M ; X; L; T ; as sec s (L1 , λx.e) → Sec s; X1 ; L1 [x 7→ ()]; ·; e
S- SEAL
M = s1 s ⊆ s1
M ; X; L; T ; seal s v → M ; X; L; T ; sealed s v
is sec M
X = (M1 ; L1 ; hi; T ), X1 T1 = append T [TMsg v]
M ; X; L; ·; v → M1 ; X1 ; L1 ; T1 ; v
M = Par s1 ⇒ s1 ⊆ s
M = Sec s1 ⇒ s1 ∩ s 6= φ
S- REVEAL
M ; X; L; T ; reveal (sealed s v) → M ; X; L; T ; v
S- MKMAP
M = Par s1 ⇒ v = sealed s2 v2 ∧ s ⊆ s1 ∧ s ⊆ s2
M = Sec s1 ⇒ s ⊆ s1 ∧ v2 = v
M ; X; L; T ; mkmap s v → M ; X; L; T ; [s 7→ v2 ]
S- CONCAT
S- PROJ
M = Par s ⇒ s = singleton p M = Sec s ⇒ p ∈ s m[p] = v
M ; X; L; T ; project m p → M ; X; L; T ; v
dom(m1 ) ∩ dom(m1 ) = φ
M ; X; L; T ; concat m1 m2 → M ; X; L; T ; m1 ] m2
v = exec ffi f v̄
S- FFI
M ; X; L; T ; ffi f v̄ → M ; X; L; T ; v
Figure 4. W YS? ST semantics (selected rules)
within F? ’s Wys monad. We leave to future work formally
proving a correspondence between the ST semantics and
µF? , the official semantics of F? in F? [30].
The standard constructs such as let bindings (let x =
e1 in e2 ), applications (e1 e2 ), etc. evaluate as usual (see
rules S- LET and S- APP), where the mode and traces play
no role. Rules S- ASPAR and S- ASPARRET reduce an as par
expression once its arguments are fully evaluated. S- ASPAR
first checks that the current mode is Par and contains all the
principals from the set s. It then pushes a seal s hi frame
on the stack, and starts evaluating e. The rule S- ASPARRET
pops the frame and seals the result, so that it is accessible
only to the principals in s. The rule also creates a trace
element TScope s T , essentially making observations during
the reduction of e (i.e., T ) visible only to the principals in
s.
To see that these rules faithfully model the F? API,
consider the F? type of as par, shown below.
1 val as par: ps:prins → (unit → Wys a pre post) →
Wys (sealed ps a)
2 (requires (fun m → m.mode=Par ∧ ps ⊆ m.ps ∧
3
can seal ps a ∧ pre (Mode Par ps)))
4 (ensures (fun m r tr → ∃t. tr=[TScope ps t] ∧
5
post (Mode Par ps) (unseal r) t)))
Rule S- ASPAR implements the pre-condition on line 2.
For the pre-condition on line 3, rule S- ASPARRET checks
that the returned value can be sealed.4 The rule also generates a trace element TScope s T , as per the post-condition
4. For technical reasons, function closures may not be sealed; see the
end of Section 4.3 for details.
on line 4, and returns the sealed value, as per the return type
of the API and the post-condition on line 5.
Next consider the rules S- ASSEC and S- ASSECRET.
Again, we can see that the rules implement the type of
as sec (shown in §3). The rule S- ASSEC checks the precondition of the API, and the rule S- ASSECRET generates
a trace observation TMsg v , as per the postcondition of the
API.
In a similar manner, we can easily see that the rule SREVEAL implements the corresponding pre- and postconditions as given in Section 3.
Rules S- MKMAP, S- PROJ, and S- CONCAT implement
map creation, projection, and concatenation respectively. For
map creation, if the current mode is Par, the rule ensures that
the parties in s can access the value v by requiring v to be
a sealed value that all parties in s can reveal (s ⊆ s2 ). The
rule also requires (for both Par and Sec mode) that all the
parties in the map domain are present in the current mode
(s ⊆ s1 ). In rule S- PROJ, if the current mode is Par then
the current party set must be a singleton equal to the index
of the map projection, whereas if the current mode is Sec,
then the index of the map projection must be present in the
current party set. Rule S- CONCAT simply checks that the
two maps have disjoint range, and returns the disjoint union
(]) of the two maps.
The rule S- FFI implements the FFI call by calling a
host-language function exec ffi. As expected, calling a hostlanguage function has no effect on the W YS? -specific state.
Concretely, this is enforced by F? ’s monadic encapsulation
of effects. We present more details of exec ffi in Section 5.2.
The remaining rules are straightforward.
4.3. Distributed semantics
The DS semantics implements judgments of the form
π −→ π 0 , where a protocol π is a tuple (P ; S) such that P
maps each principal to its local configuration and S maps a
set of principals to the configuration of an ongoing, secure
computation. Both kinds of configurations (local and secure)
have the form C (per Figure 3).
In the DS semantics, principals evaluate the same program locally and asynchronously until they reach a secure
computation, at which point they synchronize to jointly
perform the computation. This semantics is expressed with
four rules, given in Figure 6, which state that either: (1)
a principal can take a step in their local configuration, (2)
a secure computation can take a step, (3) some principals
can enter a new secure computation, and finally, (4) a
secure computation can return the result to the (waiting)
participants.
The first case is covered by rule P- PAR, which (nondeterministically) chooses a principal’s configuration and
evaluates it according to the local evaluation judgment
C
C 0 , which is given in Figure 5 (discussed below). The
second case is covered by P- SEC, which evaluates using the
ST semantics. The last two cases are covered by P- ENTER
and P- EXIT, also discussed below.
Local evaluation. The rules in Figure 5 present the local
evaluation semantics. These express how a single principal
behaves while in par mode; as such, mode M will always
be Par {p}. Local evaluation agrees with the ST semantics
for the standard language constructs (e.g. rules L- LET and
L- APP) and differs for W YS? -specific constructs.5
For an as par expression, a principal either participates
in the computation, or skips it. Rules L- ASPAR 1 and LASPARRET handle the case when p ∈ s, and so, the principal
p participates in the computation. The rules closely mirror
the corresponding ST semantics rules. One difference in the
rule L- ASPARRET is that the trace T is not scoped. In the DS
semantics, traces only contain TMsg elements; i.e., a trace
is the (flat) list of secure computation outputs observed by
that active principal. If p 6∈ s, then the principal skips the
computation with the result being a sealed value containing
garbage • (rule L- ASPAR 2). The contents of the sealed value
do not matter, since the principal will not be allowed to
unseal the value anyway.
Rule L- SEAL has the same intuition as above. Rule LREVEAL allows principal p to reveal the value sealed s v ,
only if p ∈ s. Rule L- MKMAP requires value v to be a sealed
value. In case the current principal p is in the set s, the rule
requires that p can access the contents of the sealed value
(p ∈ s2 ) and creates a singleton map that maps p to the
contents of the sealed value. In case the current principal is
not in the set s, the rule simply creates an empty map. Rule
5. Our formal development actually shares the code for both sets of rules,
using an extra flag to indicate whether a rule is “local” or “joint”.
L- PROJ projects the current principal’s mapping from the
map m. The rule for map concatenation is straightforward.
As should be the case, there are no local rules for
as sec—to perform a secure computation parties need to
combine their data and jointly do the computation.
Entering/exiting secure computations. Returning to Figure 6, Rule P- ENTER handles the case when principals enter
a secure computation. It requires that all the principals p ∈ s
must have the expression form as sec s (Lp , λx.e), where
Lp is their local environment associated with the closure.
Each party’s local environment contains its secret values (in
addition to some public values). Conceptually, a secure computation combines these environments, thereby producing a
joint view, and evaluates e under the combination. We define
an auxiliary combine v function on values as follows:
combine
combine
combine
combine
...
v
v
v
v
(•, v) = v
(v, •) = v
(p, p) = p
(sealed s v1 , sealed s v2 ) = sealed s (combine v v1 v2 )
The first two rules handle the case when one of the
values is garbage; in these cases, the function picks the other
value. For sealed values, if the set s is the same, the function
recursively combines the contents. The combine function
for the environments combines the mappings pointwise.
The combine functions for n values and environments is
a folding of the corresponding function.
So now, consider the following code:
let x = as par alice (fun x → 2) in
let y = as par bob (fun x → 3) in
let z = as sec (alice, bob) (fun z → (unseal x) + (unseal y)) in ...
In alice’s environment x will be mapped to sealed alice
2, whereas in bob’s environment it will be mapped to sealed
alice •. Similarly, in alice’s environment y will be mapped
to sealed bob •, whereas in bob’s environment it will be
mapped to sealed bob 3. Before the secure computation, their
environments will be combined, producing an environment
with x mapped to sealed alice 2 and y mapped to sealed bob 3,
and then, the secure computation function will be evaluated
in this new environment.
Although the combine v function as written is a partial
function, our metatheory guarantees that at runtime, the
function always succeeds. Since the principals are computing the same program over their view of the data, these views
are structurally similar.
So, the rule P- ENTER combines the principals’ environments, and creates a new entry in the S map. The principals
are now waiting for the secure computation to finish.
The rule P- EXIT applies when a secure computation has
terminated and returns results to the waiting principals. If the
secure computation terminates with value v , each principal
gets the value slice v p v . The slice v function is analogous
to combine v, but in the opposite direction—it strips off the
parts of v that are not accessible to p. Some cases for the
slice v function are:
slice v p p’ = p’
L- LET
X1 = (M ; L; let x = hi in e2 ; T ), X
M ; X; L; T ; let x = e1 in e2 → M ; X1 ; L; ·; e1
M ; X; L; T ; (L1 , λx.e) e1 → M ; X; L1 [x 7→ e1 ]; T ; e
L- APP
L- ASPAR 1
L- ASPARRET
p ∈ s X1 = (M ; L; seal s hi; T ), X
M ; X; L; T ; as par s (L1 , λx.e)
M ; X1 ; L1 [x 7→ ()]; ·; e
X = (M ; L1 ; seal s hi; T1 ), X1 T2 = append T1 T
M ; X; L; T ; v
M ; X1 ; L1 ; T2 ; sealed s v
L- ASPAR 2
L- SEAL
p 6∈ s
M ; X; L; T ; as par s (L1 , λx.e)
p ∈ s ⇒ v1 = seal s v p 6∈ s ⇒ v1 = seal s •
M ; X; L; T ; seal s v
M ; X; L; T ; v1
M ; X; L; T ; sealed s •
p∈s
M ; X; L; T ; reveal (sealed s v)
M ; X; L; T ; v
L- REVEAL
L- MKMAP
v = sealed s2 v2 p ∈ s ⇒ p ∈ s2 ∧ m = [p 7→ v] p 6∈ s ⇒ m = ·
M ; X; L; T ; mkmap s v → M ; X; L; T ; m
L- CONCAT
p = p1 m = [p 7→ v]
L- PROJ
M ; X; L; T ; project m p1 → M ; X; L; T ; v
dom(m1 ) ∩ dom(m1 ) = φ
M ; X; L; T ; concat m1 m2 → M ; X; L; T ; m1 ] m2
Figure 5. Distributed semantics, selected local rules (the mode M is always Par (singleton p))
P- PAR
C
C0
P [p 7→ C]; S −→ P [p 7→ C 0 ]; S
P- SEC
C → C0
P ; S[s 7→ C] −→ P ; S[s 7→ C 0 ]
∀p ∈ s. P [p].e = as sec s (Lp , λx.e) s 6∈ dom(S) L = combine L̄p
P ; S −→ P ; S[s 7→ Sec s; ·; L[x 7→ ()]; ·; e]
P- EXIT
S[s] = Sec s; ·; L; T ; v
P 0 = ∀p ∈ s. P [p 7→ P [p] / (slice v p v)]
P ; S −→ P 0 ; S 0
P- ENTER
S0 = S \ s
Figure 6. Distributed semantics, multi-party rules
slice v p (sealed s v) = sealed s •, if p ∈
6 s
slice v p (sealed s v) = sealed s (slice v p v), if p ∈ s
As an example, consider the following code:
let x = as sec (alice, bob) (fun x → let y = ... in seal alice y)
Since the return value of the secure computation is
sealed for alice, bob will get a sealed alice •, produced using
the slice v function on the result of seal alice y.
In the rule P- EXIT, the / notation is defined as:
M ; X; L; T ; / v = M ; X; L; append T [TMsg v]; v
That is, the returned value is also added to the principal’s
trace to note their observation of the value.
We now return to the point of not allowing closures to
be sealed. Consider the following example:
secure block, alice’s environment would map x to seal alice 2
and y to seal bob •, whereas bob’s environment would map
x to seal alice •, and y to the closure (L, fun z → x), where
L maps x to seal alice •. Now, as per the target semantics,
their environments are combined, and in the combined environment y gets the value from bob’s environment, but the
closure for y has a garbage value for x. Thus, running this
program in the target semantics fails to make progress.
We found this problem during our effort of mechanizing
the semantics. For now, we do not allow closures to be
boxed. We plan to fix the problem in future.
4.4. Metatheory
let x = as par alice 2 in
let y = as par bob (fun z → x) in
let z = as sec ab (fun z →
let a = reveal y in
let b = reveal (a ()) in
b) in
z
Our goal is to show that the ST semantics faithfully
represents the semantics of W YS? programs as they are
executed by multiple parties, i.e., according to the DS
semantics. We do this by proving simulation of the ST
semantics by the DS semantics, and by proving confluence
of the DS semantics. Our F? development mechanizes all
the metatheory presented in this section.
In the source semantics, the program returns 2 to both
the parties. In the target semantics, just before the call to the
Simulation. We define a slice s C function that returns the
corresponding protocol πC for an ST configuration C . In
the P component of πC , each principal p ∈ s is mapped
to their slice of the protocol. For slicing values, we use the
same slice v function as before. Traces are sliced as follows:
slice tr p (TMsg v) = [TMsg (slice v p v)]
slice tr p (TScope s T) = slice tr p T, if p ∈ s
slice tr p (TScope s T) = [], if p 6∈ s
The slice of an expression (e.g., the source program) is
itself. For all other components of C , slice functions are
defined analogously.
We say that C is terminal if it is in Par mode and is
fully reduced to a value (i.e., C.e is a value and C.X is
empty). Similarly, a protocol π = (P, S) is terminal if S
is empty and all the local configurations in P are terminal.
The simulation theorem is then the following:
Theorem 1 (Simulation of ST by DS). Let s be the set of all
principals. If C1 →∗ C2 , and C2 is terminal, then there
exists some derivation (slice s C1 ) −→∗ (slice s C2 )
such that (slice s C2 ) is terminal.
Notably, each principal’s value and trace in protocol
(slice s C2 ) is the slice of the value and trace in C2 .
Confluence. To state the confluence theorem, we first define
the notion of strong termination.
Definition 1 (Strong termination). A protocol π strongly
terminates in the terminal protocol πt , written as π ⇓ πt ,
if all possible runs of π terminate in some number of
steps in πt .
Our confluence result then says:
Theorem 2 (Confluence of DS). If π −→∗ πt and πt is
terminal, then π ⇓ πt .
Combining the two theorems, we get a corollary that
establishes the soundness of the ST semantics w.r.t. the DS
semantics:
Corollary 1 (Soundness of ST semantics). Let s be the set
of all principals. If C1 →∗ C2 , and C2 is terminal, then
(slice s C1 ) ⇓ (slice s C2 ).
Now suppose that for a W YS? source program, we prove
in F? a post-condition that the result is sealed alice n, for
some n > 0. By the soundness of the ST semantics, we can
conclude that when the program is run in the DS semantics,
it may diverge, but if it terminates, alice’s output will also
be sealed alice n, and for all other principals their outputs
will be sealed alice •. Aside from the correspondence on
results, our semantics also covers correspondence on traces.
Thus, via our VDSILE embedding of Wysteria in F? , the
correctness and security properties that we prove about a
W YS? program using F? ’s logic, hold for the program that
actually runs.
Of course, this statement is caveated by how we produce
an actual implementation from the DS semantics; details are
presented in the next section.
5. Implementation
This section describes our W YS? implementation. We
begin by describing our interpreter; we have proved that its
core implements our formal semantics, adding confidence
that bugs have not been introduced in the translation from
formalism to implementation. Then we describe our novel
FFI by which W YS? programs can easily take advantage of
features and libraries of W YS? ’s host language, F? .
5.1. W YS? interpreter
The formal semantics presented in the prior section is
mechanized as an inductive type in F? . This style is useful
for proving properties, but does not directly translate to an
implementation. Therefore, we implement an interpretation
function step in F? and prove that it corresponds to the
rules; i.e., that for all input configurations C , step(C) = C 0
implies that C
C 0 according to the semantics. Then,
the core of each principal’s implementation is an F? stub
function tstep that repeatedly invokes step on the AST of
the source program (produced by the F? extractor run in a
custom mode), unless the AST is an as sec node. Functions
step and tstep are extracted to OCaml by the standard F?
extraction process.
Local evaluation is not defined for as sec, so the stub
implements what amounts to P- ENTER and P- EXIT from
Figure 6. When the stub notices the program has reached
an as sec expression, it calls into a circuit library we have
written that converts the AST of the second argument of
as sec to a boolean circuit. This circuit and the encoded
inputs are communicated to a co-located server, written
using a library due to Choi et. al. [39] that implements
the GMW MPC protocol. The server evaluates the circuit,
coordinating with the GMW servers of the other principals,
and sends back the result. The circuit library decodes the
result and returns it to the stub. The stub then carries on
with the local evaluation.
Our F? formalization of the W YS? semantics, including
the AST specification, is 1900 lines of code. This formalization is used both by the metatheory as well as by
the (executable) interpreter. The metatheory that connects
the ST and DS semantics (Section 4.4) is 3000 lines. The
interpreter and its correctness proof are another 290 lines
of F? code. The interpreter step function is essentially a
big switch-case on the current expression, that calls into the
functions from the semantics specification. The tstep stub is
another 15 lines. The size of the circuit library, not including
the GMW implementation, is 836 lines.
The stub, the implementation of GMW, the circuit library, and the F? extractor (including our custom W YS?
mode for it) are part of our trusted computing base. As
such, bugs in them could constitute security holes. Verifying
these components as well (especially the circuit library and
the GMW implementation, which are open problems to our
knowledge) is interesting future work.
5.2. FFI
When writing a source W YS? program (in F? ), the
programmer can access definitions from an FFI module,6
which exports F? datatypes and library functions to W YS?
programs. Below we explain W YS? ’s extensible FFI mechanism, which enables programmers to add new datatypes and
functions to the FFI module, while ensuring that the W YS?
metatheory remains applicable.
Source programming with FFI. Datatypes and functions
are added to the FFI module as usual:
type list α =
| Nil: list α
| Cons: hd:α → tl:list α → list α
val append: l1 :list α → l2 :list α → Tot (list α)
let rec append l1 l2 = match l1 with
| [] → l2
| hd::tl → hd::(append tl l2 )
In addition to these usual definitions, W YS? requires the
programmer to define the corresponding slice (and combine)
functions (Section 4.3) for the new datatype, e.g.:
val slice list: (prin → α → Tot α) → prin → list α → Tot (list α)
let rec slice list f p l = match l with
| [] → []
| hd::tl → (f p hd)::(slice list f p tl)
Once the datatype is defined, an W YS? program is free
to use it by importing the FFI module, and can prove
properties with it using standard F? reasoning. F? ’s monadic
encapsulation of effects ensures that the FFI functions do
not interfere with W YS? -specific state, i.e., the mode and
trace. For example, the effect annotation Tot in append above
indicates that append is a pure function, and hence does not
modify or use the the Wys monad’s state.
FFI metatheory. Recall from Section 4 that we formalize
FFI calls using the expression form ffi, FFI values using
the value form v, and the semantics of FFI calls using the
meta function exec ffi. The W YS? metatheory (Section 4.4)
needs to relate the ST and DS semantics of FFI calls. For
this purpose, the FFI programmer must prove that the added
definitions meet certain obligations. For example, one such
obligation is that the slice of the value returned from an FFI
call must match the return value of the FFI call on the slice
of the arguments. Formally, for an FFI function f ,
∀p, v̄. slice p (exec ffi f v̄) = exec ffi f slice p v
To fulfill this obligation for the append function, for
example, the programmer is required to prove the following
lemma:
val slice append lemma:
f:(prin → α → Tot α) → p:prin → l1 :list α → l2 :list α → Lemma
(ensures (slice list f p (append l1 l2 ) =
append (slice list f p l1 ) (slice list f p l2 )))
6. How F? programs call into W YS? functions was described in Section 3.1.
This lemma is easily provable in F? .
Once all obligations are proven, W YS? metatheory guarantees that the theorems of Section 4.4 extend to the new
datatypes and functions in the FFI.
FFI implementation. The FFI module itself is extracted
to OCaml via regular F? extraction. We rely on the F?
metatheory to conclude that the extracted code executes per
its specification.
The custom F? extraction mode that we have implemented for W YS? , identifies integer, string, and other constants in a W YS? source program, and extracts them to the
V ffi AST form, which is part of the AST value type. The FFI
calls, that can take as arguments and return F? datatypes,
are extracted to an E ffi AST form, which is part of the AST
expression type:
type exp = . . .
| E ffi: f:α → args:list exp → inj:β → exp
The f argument is extracted to be the name of the FFI
function, that links to the extracted OCaml function. We
explain the inj argument shortly.
When evaluating the W YS? AST, the interpreter may
reach an E ffi f args inj node. As we saw in Section 4.2, the
interpreter calls a library function exec ffi with the list of
values; in addition, it passes a inj argument. The exec ffi
function first un-embeds any embedded host-language arguments. The un-embedding function is straightforward (the
values shown below are from the value AST form):
unembed V unit = ()
unembed (V ffi v) = v (∗ values in the host language ∗)
unembed (V seal s v) = V seal s v
Interpreter-specific values, such as V seal, are passed as
is. The FFI module does not have access to the W YS? API
in F? , and hence it must treat these values abstractly. exec ffi
then calls the OCaml function f with the un-embedded
arguments. The OCaml function returns some result which
needs to be embedded back into the AST. F? programs
cannot inspect the type of values at run-time, but fortunately
the compiler has enough information during extraction to
know how to re-embed the result.
In particular, when the extractor compiles an FFI call
in the source program to an E ffi node, it has the type
information for the return value of the FFI call. Using this
information, it instruments the E ffi node with an injection,
a function that can be used at runtime to embed the FFI call
result back to the AST. For example, if the result is (), the
injection is (an OCaml function) fun x → V unit. If the return
value is an interpreter value (e.g. V seal), the injection is the
identity. If the return value is some host F? value (such as a
list, tuple, or int), the injection creates an V ffi node. exec ffi
uses the injection to embed the result back to the AST, and
returns it to the interpreter.
Our interface essentially provides a form of monomorphic, first-order interoperability between the (dynamically
typed) interpreter and the host language. We do not foresee
any problems extending this approach to be higher order by
using coercions [48].
Here, pre captures the preconditions on the inputs as
mentioned above, and median spec is an idealized median
specification. For both the versions, F? is able to prove
the specification without any extra hints in the code. For
example, the monolithic version is programmed as:
let median x y =
let g: unit → Wys int
(requires (fun m → m = Mode Sec {Alice, Bob}))
(ensures (fun r → (pre (unseal x) (unseal y) =⇒
r = median spec (unseal x) (unseal y))))
= fun → ... in //monolithic median algorithm
as sec {Alice, Bob} g
Figure 7. Time to run (in secs) normal and optimized PSI for varying
per-party set sizes and intersection densities.
6. Applications
Private set intersection. We evaluate the performance of the
psi (computing intersection in a single secure computation),
and the psi opt (the optimized version) algorithms from
Section 3. The programs that we benchmark are slightly
different than the ones presented there, in that the local col
and row functions are not the verified ones. The results are
shown in Figure 7. We measure the time (in seconds) for per
party set sizes 96, 128, and 256, and intersection densities
(i.e. the fraction of elements that are common) 0.5, 0.75,
and 0.9.
The time taken by the unoptimized version is independent of the intersection density since it always compares
all pairs of values. However, as the intersection density
increases, the optimized version performs far better – it
is able to skip many comparisons. For lower densities (<
0.35), the optimization does not improve performance, as
the algorithm essentially becomes quadratic, and the setup
cost for each secure computation takes over.
We note that a similar performance profile was also
noted by Rastogi et. al. [19], although they did not experiment with set size 256, like we did.
Joint median. We program unoptimized and optimized versions of the two-party joint median [49]. The programs take
two distinct, sorted inputs from alice, x1 and x2 , and two
distinct, sorted inputs from bob, y1 and y2 and return the
median of all four. In the unoptimized version, the whole
computation takes place as a monolithic secure computation,
whereas the optimized version breaks up the computation,
revealing some intermediate results and off-loading some
parts to the local hosts (much like PSI). We refer the reader
to [49] for more details of the algorithms.
For both the versions, we prove functional correctness:
val median:
x:sealed Alice (int ∗ int) → y:sealed Bob (int ∗ int) → Wys int
(requires (fun m → m = Mode Par {Alice, Bob}))
(ensures (fun r t → (pre (unseal x) (unseal x) =⇒
r = median spec (unseal x) (unseal y))))
For proving the security properties, for the unoptimized
version, we prove that the observable trace is [TMsg r], where
r is the result of the computation, basically reflecting that
both the parties only see the final result. For the optimized version, we first prove that the observable trace is
opt trace (reveal x) (reveal y), where opt trace is a purely specificational function that takes as arguments Alice’s (reveal x)
and Bob’s (reveal y) inputs, and returns the trace generated
by the optimized median algorithm. Then, we prove that the
trace does not reveal more than the final output, by proving
the following relational lemma:
val optimized median is secure for alice:
x1:(int ∗ int) → x2:(int ∗ int) → y:(int ∗ int)
→ Lemma (requires (pre x1 y ∧ pre x2 y ∧
median spec x1 y = median spec x2 y))
(ensures (opt trace x1 y = opt trace x2 y))
The lemma says that for two runs of the optimized median with arbitrarily different inputs from Alice (x1 and x2), if Bob’s input and output are the same,
i.e. median spec x1 y = median spec x2 y, then the observable
trace is also same. Essentially, the trace does not reveal more
about Alice’s inputs, beyond what is already revealed by the
output. We also prove a symmetrical lemma for Bob where
we vary Bob’s inputs and keep Alice’s input and output the
same. Once again, for both the proofs, F? is able to prove
them automatically.
Card dealing. We have implemented an MPC-based card
dealing application in W YS? . Such an application can play
the role of the dealer in a game of online poker, thereby
eliminating the need to trust the game portal for card dealing. The application relies on W YS? ’s support for secret
shares [31]. Using secret shares, the participating parties can
share a value in a way that none of the parties can observe
the actual value individually (each party’s share consists of
some random-looking bytes), but they can recover the value
by combining their shares in a secure block.
In the application, the parties maintain a list of secret
shares of already dealt cards (the number of already dealt
cards is public information). To deal a new card, each party
first generates a random number locally. The parties then
perform a secure computation to compute the sum of their
random numbers modulo 52, let’s call it n. The output of
the secure block is secret shares of n. Before declaring n
as the newly dealt card, the parties needs to ensure that the
card n has not already been dealt. To do so, they iterate over
the list of secret shares of already dealt cards, and for each
element of the list, check that it is different from n. The
check is performed in a secure block that simply combines
the shares of n, combines the shares of the list element, and
checks the equality of the two values. If n is different from
all the previously dealt cards, it is declared to be the new
card, else the parties repeat the protocol by again generating
a fresh random number each.
W YS? exports the following API for secret shares:
location sharing application. At the moment, we have only
run these applications using a secure server backend. In
this backend, as sec works by literally sending code and
inputs to a separate server that implements the ST semantics
directly. The server returns the result with a cryptographic
proof of correctness to each party (we have verified the
use of cryptography using a technique similar to [50]). We
conjecture that such a server could be useful for a trusted
hardware based deployment scenario.
type Sh: Type → Type
type can sh: Type → Type
assume Cansh int: can sh int
7. Conclusions
val v of sh: #a:Type → sh:Sh a → Ghost a
val ps of sh: #a:Type → sh:Sh a → Ghost prins
val mk sh: #a:Type → x:a → Wys (Sh a)
(requires (fun m → m.mode = Sec ∧ can sh a))
(ensures (fun m r tr → v of sh r = x ∧ ps of sh r = m.ps ∧
tr = [])
val comb sh: #a:Type → x:Sh a → Wys a
(requires (fun m → m.mode = Sec ∧ ps of sh x = m.ps))
(ensures (fun m r tr → v of sh x = r ∧ tr = [])
Type Sh a types the shares of values of type a. Our
implementation currently supports shares of int values only;
the can sh predicate enforces this restriction on the source
programs. Extending secret shares support to other types
(such as pairs) should be straightforward. Functions v of sh
and ps of sh are marked Ghost, meaning that they can only
be used in specifications for reasoning purposes. In the
concrete code, shares are created and combined using the
mk sh and comb sh functions. Together, the specifications
of these functions enforce that the shares are created and
combined by the same set of parties (through ps of sh), and
that comb sh recovers the original value (through v of sh).
The W YS? interpreter transparently handles the low-level
details of extracting shares from the GMW implementation
of Choi et al. (mk sh), and reconstituting the shares back
(comb sh).
In addition to implementing the card dealing application
in W YS? , we have formally verified that the returned card is
fresh. The signature of the function that checks for freshness
of the newly dealt card is as follows (abc is the set of
parties):
val check fresh:
l:list (Sh int){∀ s’. mem s’ l =⇒ ps of sh s’ = abc}
→ s:Sh int{ps of sh s = abc}
→ Wys bool (requires (fun m → m = Mode Par abc))
(ensures (fun r → r ⇐⇒ (∀ s’. mem s’ l =⇒
not (v of sh s’ = v of sh s))))
The specification says that the function takes two arguments: l is the list of secret shares of already dealt cards, and
s is the secret shares of the newly dealt card. The function
returns a boolean r that is true iff the concrete value (v of sh)
of s is different from the concrete values of all the elements
of the list l. Using F? , we verify that the implementation of
check fresh meets this specification.
Other applications and secure server. We have implemented some more applications in W YS? , including a geo-
This paper has proposed Verified, Domain-Specific Integrated Language Extensions (VDSILE) as a new way to
implement a domain-specific language. The paper specifically applies the idea to design and implement W YS? , a new
MPC DSL that is hosted in F? . W YS? inherits the basic
programming model from Wysteria. However, by virtue
of being implemented as a VDSILE, it provides several
novel capabilities missing from all previous MPC DSLs,
including Wysteria. W YS? is the first DSL to enable formal
verification of the source MPC programs. W YS? is also
the first MPC DSL to provide a (partially) verified interpreter. Furthermore, W YS? programs can freely use standard
constructs such as datatypes and libraries directly from
F? , thereby making it more scalable and usable. All these
capabilities constitute a significant step towards making
MPC more practical and trustworthy. The paper has reported
on several MPC applications programmed in W YS? , and
verified for correctness and security.
References
[1]
A. Shamir, R. L. Rivest, and L. M. Adleman, Mental poker. Springer,
1980.
[2]
A. C.-C. Yao, “How to generate and exchange secrets,” in FOCS,
1986.
[3]
O. Goldreich, S. Micali, and A. Wigderson, “How to play ANY
mental game,” in STOC, 1987.
[4]
D. Beaver, S. Micali, and P. Rogaway, “The round complexity of
secure protocols,” in STOC, 1990.
[5]
P. Bogetoft, D. L. Christensen, I. Damgård, M. Geisler, T. Jakobsen,
M. Krøigaard, J. D. Nielsen, J. B. Nielsen, K. Nielsen, J. Pagter,
M. Schwartzbach, and T. Toft, “Financial cryptography and data
security,” 2009, ch. Secure Multiparty Computation Goes Live.
[6]
D. Bogdanov, M. Jemets, S. Siim, and M. Vaht, “How the estonian
tax and customs board evaluated a tax fraud detection system based
on secure multi-party computation,” in Financial Cryptography and
Data Security. Springer Berlin Heidelberg, 2015.
[7]
F. Kerschbaum, A. Schroepfer, A. Zilli, R. Pibernik, O. Catrina,
S. de Hoogh, B. Schoenmakers, S. Cimato, and E. Damiani, “Secure
collaborative supply-chain management,” Computer, 2011.
[8]
L. Kamm, “Privacy-preserving statistical analysis using secure multiparty computation,” Ph.D. dissertation, University of Tartu, 2015.
[9]
D. Malkhi, N. Nisan, B. Pinkas, and Y. Sella, “Fairplay: a secure
two-party computation system,” in USENIX Security, 2004.
[10] Y. Huang, D. Evans, J. Katz, and L. Malka, “Faster secure two-party
computation using garbled circuits,” in USENIX, 2011.
[11] “VIFF, the virtual ideal functionality framework,” http://viff.dk/.
[12] L. Malka, “Vmcrypt: modular software architecture for scalable secure computation,” in CCS, 2011.
[13] A. Ben-David, N. Nisan, and B. Pinkas, “FairplayMP: a system for
secure multi-party computation,” in CCS, 2008.
[14] A. Holzer, M. Franz, S. Katzenbeisser, and H. Veith, “Secure twoparty computations in ANSI C,” in CCS, 2012.
[15] J. D. Nielsen and M. I. Schwartzbach, “A domain-specific programming language for secure multiparty computation,” in PLAS, 2007.
[16] J. D. Nielsen, “Languages for secure multiparty computation and
towards strongly typed macros,” Ph.D. dissertation, 2009.
[17] D. Bogdanov, S. Laur, and J. Willemson, “Sharemind: A framework
for fast privacy-preserving computations,” in Computer Security ESORICS 2008, 2008.
[18] A. Schropfer, F. Kerschbaum, and G. Muller, “L1 - an intermediate
language for mixed-protocol secure computation,” in COMPSAC,
2011.
[19] A. Rastogi, M. A. Hammer, and M. Hicks, “Wysteria: A programming language for generic, mixed-mode multiparty computations,” in
Proceedings of the 2014 IEEE Symposium on Security and Privacy,
2014.
[20] C. Liu, Y. Huang, E. Shi, J. Katz, and M. Hicks, “Automating efficient
ram-model secure computation,” in IEEE Symposium on Security and
Privacy (Oakland), 2014.
[21] P. Laud and J. Randmets, “A domain-specific language for low-level
secure multiparty computation protocols,” in Proceedings of the
22Nd ACM SIGSAC Conference on Computer and Communications
Security, ser. CCS ’15. New York, NY, USA: ACM, 2015, pp.
1492–1503. [Online]. Available: http://doi.acm.org/10.1145/2810103.
2813664
[22] P. Mardziel, M. Hicks, J. Katz, M. Hammer, A. Rastogi, and M. Srivatsa, “Knowledge inference for optimizing and enforcing secure
computations,” in Proceedings of the Annual Meeting of the US/UK
International Technology Alliance, 2013, this short paper consists of
coherent excerpts from several prior papers.
[23] B. E. Aydemir, A. Bohannon, M. Fairbairn, J. N. Foster, B. C. Pierce,
P. Sewell, D. Vytiniotis, G. Washburn, S. Weirich, and S. Zdancewic,
“Mechanized metatheory for the masses: The poplmark challenge,” in
Proceedings of the 18th International Conference on Theorem Proving in Higher Order Logics, ser. TPHOLs’05. Berlin, Heidelberg:
Springer-Verlag, 2005, pp. 50–65.
[24] C. Klein, J. Clements, C. Dimoulas, C. Eastlund, M. Felleisen,
M. Flatt, J. A. McCarthy, J. Rafkind, S. Tobin-Hochstadt, and R. B.
Findler, “Run your research: On the effectiveness of lightweight
mechanization,” in Proceedings of the 39th Annual ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, ser.
POPL ’12. New York, NY, USA: ACM, 2012, pp. 285–296.
[25] X. Yang, Y. Chen, E. Eide, and J. Regehr, “Finding and understanding bugs in C compilers,” in Proceedings of ACM SIGPLAN 2011
Conference on Programming Language Design and Implementation,
2011.
[26] X. Leroy, “Formal verification of a realistic compiler,” Commun.
ACM, 2009.
[27] K. Bhargavan, C. Fournet, M. Kohlweiss, A. Pironti, and P.-Y. Strub,
“Implementing TLS with verified cryptographic security,” in IEEE
Symposium on Security & Privacy (Oakland), 2013, pp. 445–462.
[Online]. Available: http://www.ieee-security.org/TC/SP2013/papers/
4977a445.pdf
[28] “PolarSSL
verification
polarssl-verification-kit/, 2015.
kit,”
http://trust-in-soft.com/
[29] J. Yang and C. Hawblitzel, “Safe to the last instruction: Automated
verification of a type-safe operating system.” Association for Computing Machinery, Inc., June 2010.
[30] N. Swamy, C. Hriţcu, C. Keller, A. Rastogi, A. Delignat-Lavaud,
S. Forest, K. Bhargavan, C. Fournet, P.-Y. Strub, M. Kohlweiss, J.K. Zinzindohoue, and S. Z. Beguelin, “Dependent types and multimonadic effects in F*,” in POPL, 2016.
[31] A. Shamir, “How to share a secret,” Communications of the ACM,
vol. 22, no. 11, pp. 612–613, Nov. 1979.
[32] J. Launchbury, I. S. Diatchki, T. DuBuisson, and A. Adams-Moran,
“Efficient lookup-table protocol in secure multiparty computation,” in
ICFP, 2012.
[33] S. Zahur and D. Evans, “Obliv-c: A language for extensible
data-oblivious computation,” Unpublished, 2015, http://oblivc.org/
downloads/oblivc.pdf.
[34] J. B. Almeida, M. Barbosa, G. Barthe, G. Davy, F. Dupressoir,
B. Grgoire, and P.-Y. Strub, “Verified implementations for secure and
verifiable computation,” 2014.
[35] M. Backes, M. Maffei, and E. Mohammadi, “Computationally Sound
Abstraction and Verification of Secure Multi-Party Computations,” in
IARCS Annual Conference on Foundations of Software Technology
and Theoretical Computer Science (FSTTCS 2010), 2010.
[36] E. Meijer, B. Beckman, and G. Bierman, “Linq: Reconciling object,
relations and xml in the .net framework,” in Proceedings of the 2006
ACM SIGMOD International Conference on Management of Data,
ser. SIGMOD ’06. New York, NY, USA: ACM, 2006, pp. 706–706.
[Online]. Available: http://doi.acm.org/10.1145/1142473.1142552
[37] “Bedrock, a coq library for verified low-level programming,” http:
//plv.csail.mit.edu/bedrock/.
[38] The Coq development team, The Coq proof assistant. [Online].
Available: http://coq.inria.fr
[39] S. G. Choi, K.-W. Hwang, J. Katz, T. Malkin, and D. Rubenstein,
“Secure multi-party computation of boolean circuits with applications
to privacy in on-line marketplaces,” 2011, http://eprint.iacr.org/.
[40] Y. Huang, D. Evans, and J. Katz, “Private set intersection: Are garbled
circuits better than custom protocols?” in NDSS, 2012.
[41] R. Atkey, “Parameterised notions of computation,” Journal of
Functional Programming, vol. 19, pp. 335–376, 2009. [Online].
Available: http://journals.cambridge.org/article S095679680900728X
[42] A. Nanevski, J. G. Morrisett, and L. Birkedal, “Hoare type theory,
polymorphism and separation,” J. Funct. Program., vol. 18, no. 5-6,
pp. 865–911, 2008. [Online]. Available: http://ynot.cs.harvard.edu/
papers/jfpsep07.pdf
[43] N. Benton, “Simple relational correctness proofs for static analyses
and program transformations,” in Proceedings of the 31st ACM
SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, ser. POPL ’04. New York, NY, USA: ACM, 2004,
pp. 14–25. [Online]. Available: http://doi.acm.org/10.1145/964001.
964003
[44] M. R. Clarkson and F. B. Schneider, “Hyperproperties,” J. Comput.
Secur., vol. 18, no. 6, pp. 1157–1210, Sep. 2010. [Online]. Available:
https://www.cs.cornell.edu/fbs/publications/Hyperproperties.pdf
[45] A. Sabelfeld and A. C. Myers, “A model for delimited information release,” in Software Security - Theories and Systems, Second
Mext-NSF-JSPS International Symposium, ISSS 2003, Tokyo, Japan,
November 4-6, 2003, Revised Papers, 2003.
[46] G. Barthe, C. Fournet, B. Grégoire, P. Strub, N. Swamy, and S. Z.
Béguelin, “Probabilistic relational verification for cryptographic
implementations,” in The 41st Annual ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages, POPL ’14,
San Diego, CA, USA, January 20-21, 2014, 2014, pp. 193–206.
[Online]. Available: http://doi.acm.org/10.1145/2535838.2535847
[47] M. Felleisen and R. Hieb, “The revised report on the syntactic theories
of sequential control and state,” Theoretical computer science, vol.
103, no. 2, pp. 235–271, 1992.
[48] F. Henglein, “Dynamic typing: Syntax and proof theory,” Sci.
Comput. Program., vol. 22, no. 3, pp. 197–230, 1994. [Online].
Available: http://dx.doi.org/10.1016/0167-6423(94)00004-2
[49] A. Rastogi, P. Mardziel, M. Hammer, and M. Hicks, “Knowledge
inference for optimizing secure multi-party computation,” in PLAS,
2013.
[50] C. Fournet, M. Kohlweiss, and P.-Y. Strub, “Modular code-based
cryptographic verification,” in Proceedings of the 18th ACM Conference on Computer and Communications Security, 2011.
| 6cs.PL
|
Antifragility for Intelligent Autonomous Systems
arXiv:1802.09159v1 [cs.AI] 26 Feb 2018
Anusha Mujumdar, Swarup Kumar Mohalik, Ramamurthy Badrinath
Ericsson Research, Bangalore
anusha.pradeep.mujumdar, swarup.kumar.mohalik,[email protected]
Abstract
Antifragile systems grow measurably better in the
presence of hazards. This is in contrast to fragile
systems which break down in the presence of hazards, robust systems that tolerate hazards up to a
certain degree, and resilient systems that – like selfhealing systems – revert to their earlier expected
behavior after a period of convalescence. The notion of antifragility was introduced by Taleb for
economics systems, but its applicability has been
illustrated in biological and engineering domains
as well. In this paper, we propose an architecture
that imparts antifragility to intelligent autonomous
systems, specifically those that are goal-driven and
based on AI-planning. We argue that this architecture allows the system to self-improve by uncovering new capabilities obtained either through the
hazards themselves (opportunistic) or through deliberation (strategic). An AI planning-based case
study of an autonomous wheeled robot is presented.
We show that with the proposed architecture, the
robot develops antifragile behaviour with respect to
an oil spill hazard.
1
Introduction
How should an intelligent, autonomous system adapt to unforeseen situations? Answering this question has been the
focus of significant research effort over the last two decades
[Hayes-Roth, 1995]. However, the notion of adaptivity has
so far been limited to a system’s capability to cope with unknown situations and return, at best to its original performance. In this work, we argue that the presence of unforeseen
stressors, or hazards, offers a system an opportunity to selfimprove. Such self-improvement in the presence of hazards
is the central notion of antifragility.
The term antifragility was coined by Nicholas Taleb in his
book [Taleb, 2012], in which he argued that the opposites of
fragile systems are not robust or resilient systems, but systems that benefit from hazards and grow stronger as a result.
These concepts and their relationships are compelling and
have been illustrated in many domains such as economics, biology and engineering. However, the literature does not provide any suggestion towards formalization of the concepts,
due to which there are no concrete guidelines for designing an antifragile system. In this paper, we address this gap
by formalizing the notions of fragility, robustness, resilience
and antifragility in the context of intelligent systems, in particular, for a large subclass of systems that are based upon
knowledge-based reasoning and AI planning. We show that
antifragile intelligent systems can be designed using a refinement of MAPE-K (Monitor, Analyse, Planning and Execution - Knowledge), the autonomic architecture suggested by
IBM [IBM, 2005].
The paper is organized as follows. We briefly review some
background work on antifragility in section 2. Section 3 lays
out the scope of the current study, and describes our approach
towards introducing antifragility concepts into intelligent systems. Section 4 details the intelligent systems model followed, and formalizes concepts of fragility and antifragilility in such systems. Section 5 proposes design modules for
antifragility, to be augmented to an existing autonomic computing architecture of intelligent systems. In section 6, we
illustrate the developed concepts with the help of a robot path
planning example. We conclude in section 7 with some future
directions.
2
Related Work
Systems that degrade in behavior when acted upon by external hazards (or, stressors) are known as fragile systems. Robust systems display tolerance to hazards upto a certain predesigned level, and show no change in performance. If the
magnitude of the hazard exceeds this level, system performance rapidly degrades. An example of a robust system is
a building designed to withstand an earthquake of a particular magnitude. Resilient systems are affected by hazards, but
return to normal performance after a hiatus. Robust and resilient systems have been studied extensively as a means to
allow intelligent systems to cope with hazards [Russell et al.,
2015]; however, in such systems, the performance, at best,
retains its original behavior.
Taleb notes that certain systems, in fact, benefit from the
presence of hazards, and that these systems need to be distinguished from being merely robust or resilient. Such systems
abound in nature e.g., the immune system strengthens itself
upon encounter with an antigen and often retains the developed coping capability permanently [Monperrus, 2017]. Another example is muscular strength being developed by the
stressor of exercise [Jones, 2014]. The increase in strength so
developed improves the capability of the muscle to perform a
wide variety of tasks many of which may be unrelated to the
exercise that initially triggered it. In environmental ecosystems [Taleb, 2012], the presence of a hazard in the form of a
predator strengthens the surviving prey population, since they
learn from past experience.
Antifragile systems have been studied in the literature
in the context of communication systems [Lichtman et al.,
2016], infrastructure networks [Fang and Sansavini, 2017],
aerospace [Jones, 2014], health care [Clancy, 2015] and
cloud systems [Johnson and Gheorghe, 2013]. All of these
emphasize the following characteristics of antifragile systems:
• Such systems thrive in the presence of hazards, as opposed to robust systems that simply endure, and resilient
systems which merely survive and return to original performance at best.
• Antifragile systems learn from failure to handle hazards;
the failure may be their own or of other members in a
population-based system.
• When acted upon by hazards, the system responds in
a fashion that improves its performance over time (not
necessarily immediately); the improvement is through
acquisition of new capabilities caused by deliberate
strategies by the system or through the opportunistic exploitation of certain attributes of the hazard itself.
The literature, in general, does not provide formal definitions
of the concepts of fragility, robustness, resilience and antifragility. In [Lichtman et al., 2016], antifragility is demonstrated mathematically for a particular hazard in a communication system, but the definitions are specific to the problem
and are not generic enough to be applicable to other systems.
3
Our Scope and Approach
In this paper, we wish to formalize the concepts of antifragility in the context of intelligent systems. Specifically,
we consider the class of systems that are goal-driven and
based on AI planning. These systems have a fixed set of actions. Given goals and constraints, they attempt to achieve the
goals by deriving a sequence (called plan) from the available
actions and then executing the plan.
An intelligent system can be fragile because hazards encountered during execution may drive the system to states
from where it cannot find a plan to achieve its goals. If, on
the other hand, it is able to find an alternate plan to lead to
the goals no matter where the hazards drive it to, it is designated resilient. Note that if the system is able to find plans
that avoid the hazards altogether then occurrence of hazards
do not affect the plans - such a system is robust w.r.t. the hazards. This is possible in many cases since most hazards are
localized (e.g. traffic jam, corruption of memory block).
A crucial point to note is that robustness and resilience
uses plans over actions available with the system. However, antifragile systems, by definition, must become stronger
when a hazard occurs. In the intelligent systems, we correlate
strength with available plans and take a position that a system is stronger if there are more plans available. This can be
implemented by introducing new actions in the system. As
a consequence, a system that was fragile w.r.t. some hazards
may now become robust and/or resilient. This is consistent
with the observation in antifragility literature that new capabilities introduce redundancy, which brings beneficial effects
through resilience and robustness.
Modeling the above notion of antifragility in the context of
intelligent systems has a difficulty. In the antifragility literature, new capabilities are introduced through external entities (e.g. medicine) or through realization of already existing
capabilities (e.g. exercise). However, the intelligent systems
that we consider have fixed sets of actions and do not have access to any external source for new actions1 . We ameliorate
this situation by partitioning the set of actions into visible and
hidden subsets. At any point of time, the plan synthesis algorithms have access to only the visible subset. Actions can be
brought in from the hidden set to introduce new actions for
the planner. One can argue that such an approach is “artificial” because the system by default has access to all actions,
which results in the largest set of plans possible. However,
our approach is realistic and has a strong rationale: first, due
to the smaller visible set, planning is more efficient. More
importantly, there is a cost to support the planning and execution of actions (more sensors and actuators, more predicates,
larger state space). Therefore, it is advisable to consider only
a subset of the actions that are necessary in the current situation and enhance the subset only when in need (e.g. to handle
a hazard). This is accurately captured by the partition of the
actions into visible and hidden sets.
An antifragile intelligent system must therefore first decide
what actions are made visible, when and for what duration,
and then execute these decisions. We show that by suitable
refinement of the MAPE-K loop mentioned in the introduction, with refined action sets and appropriate modules, one
can obtain an architecture to design intelligent systems that
exhibit antifragility. In the rest of this section, we give some
details of the base MAPE-K loop.
Intelligent systems based on AI planning can be implemented using the MAPE-K loop (see Fig. 2). It comprises the
Monitor, Analyzer, Planner, Executor and Knowledge modules orchestrated in a loop by the Autonomic Manager. The
Monitor module is responsible for gathering relevant state information from the resources to detect the events of interest. This information is then used by the Analysis module
to determine if there is a need to change the current plan (sequence of actions) which is being executed. The change may
be necessitated due to a change in the goals of the system
or changes in the environment which force a course correction. The Planning module is then called upon to synthesize a
new plan for the changed goals/context. The Execution module uses the generated plan to resume its execution. During
the monitoring of the environment and execution, the system derives knowledge about the environment and possible
responses that it can store in a knowledge-base . This knowl1
We note that the generalization to “open” intelligent systems is
certainly interesting, and we leave this direction for future work.
edge is used to derive faster and better responses as the system
continues to interact with the environment.
4
System Model
We model an intelligent system A as a goal-driven, planningbased agent interacting with an environment E . The interaction is specified as a game where the environment supplies a goal and the agent tries to achieve the goal through
a plan (sequence of actions). Once a goal is achieved, the
environment supplies another goal and thus the game continues ad infinitum. During the execution of the plan, the environment may change the state of the system nondeterministically, capturing the notion of hazards. The type of response
from A determines whether it is fragile, robust, resilient or
antifragile.
Formally, we have a finite set of boolean predicates Q, and
the system states S are defined as all valuations V : Q →
{true, false}. The environment E is specified through a set of
goal states G ⊆ S, goal transitions TG ⊆ G × G (also called
missions) and a set of hazard transitions H ⊆ S × S. For a
hazard (s, t), s is referred to as the hazard source and t the
hazard consequence. There is a special subset W of G called
the waypoints.
The intelligent system A is defined over of a set
of actions Act where each action is specified by a
hprecondition, ef f ecti pair with precondition, effect ⊆ Q.
Act induces a deterministic labeled graph SAct with S as the
nodes and edges (s, a, t) when precondition(a) is consistent with s and t = s ← effect(a) denoting the values of the
predicates in s being overridden by effect(a). A sequence of
actions P , is a plan for a pair of states (c, g) if starting at the
initial state c and applying the actions in P consecutively, we
arrive at the goal state g. The corresponding path is denoted
as path(P, c, g). We assume that the system has a special
reset action which moves it to one of the waypoints i.e. for
every state s ∈ S, there is a waypoint w ∈ W such that
(s, reset, w) ∈ SAct .
The minimal notations above are sufficient to define the
following concepts succinctly.
Definition A plan P for (c, g) is fragile w.r.t H if there is
state s on path(P, c, g) and (s, t) ∈ H such that there is no
plan P 0 for (t, g).
A plan P for (c, g) is robust w.r.t a set of hazards H, if for
all (s, t) ∈ H, s is not on path(P, c, g).
A plan P for (c, g) is resilient w.r.t H, if for all (s, t) ∈ H,
s is on path(P, c, g) implies there is a plan P 0 for (t, g).
A fragile plan is one which goes through a state where a
hazard can occur and the from the hazard consequence state
one cannot achieve the intended goal state. A robust plan
completely avoids the states where hazards can occur. A resilient plan switches to another starting from the hazard consequence state when a hazard occurs. These plans are illustrated in Figure 1 (in (b) the dotted path on the left corresponds to a robust plan). One can extend these definitions to
the entire system.
Definition An intelligent system is fragile w.r.t H if there is
a mission (c, g) such that all plans for (c, g) are fragile w.r.t.
H.
Figure 1: Fragile, Robust and Resilient Plans
An intelligent system is robust w.r.t a set of hazards H if
for every mission (c, g), there is a robust plan w.r.t. H.
An intelligent system is resilient w.r.t. H if, for every mission (c, g), any plan P for (c, g) is resilient w.r.t. H.
Corollary 4.1 An intelligent system is not fragile w.r.t. H if
it is either robust or resilient w.r.t. H.
Robustness and resilience are independent concepts, i.e. a
system may be robust w.r.t. a hazard h because there is a
robust plan for every mission, but there may be other plans
which enable a hazard and are not resilient. Similarly, all the
plans for the missions may be resilient but there may not be a
single robust plan.
Fragility, robustness and resilience refers to a given system.
However, antifragility points to the capability of systems to
evolve from fragile to not fragile. This is how we define the
notion of “an antifragile system becomes stronger” due to the
occurrence of hazards.
Definition An intelligent system A is Antifragile w.r.t. a set
of hazards H if, A is fragile w.r.t H implies A eventually becomes either resilient or robust to H.
A corollary of the definition is that whereas there was no plan
for (t, g) for some hazard consequence t and goal g originally,
there needs to be such plan to have resiliency. Similarly, there
has to be new paths for some mission (s, g) to have robustness. This implies there is a need to introduce new actions for
the agent.
As one can see, the definition of antifragility leaves a number of dimensions unspecified. When there is a hazard, if
there is no alternate plan, it does not specify whether the improvement through new actions should happen immediately
or at a future time, whether the improvement should be for
a limited duration (say, till the current goal) or permanent,
whether the improvement should be for the specific hazard
only or a related larger set. We note that such decisions depend upon predictive capabilities of the system and cost considerations that must be left for the designer.
5
Implementation
In this section, we describe the refinements for the MAPE-K
loop and also suggest some design decisions for the unspecified dimensions mentioned in the previous section as a guide
for system designers.
({Now, Later}), D denoted duration of improvement ({Current, All}) and M denotes mode of handling({Robust, Resilient})
Note that the hW, D, M i’s are only recommendations. The
system will “improve” for any value of M. W denotes that
the urgency of improvement and D controls whether the new
actions are necessary for longer term.
5.4
Figure 2: Intelligent System with MAPE-K Architecture
5.1
Knowledge
The Knowledge module (KM) has a set of actions. As
observed in Section 3, the key idea for introducing new
actions is partitioning the actions Act into 3 classes:
Acte (empowering), Actv (visible) and Acth (hidden). All actions a have an associated visibility predicate visiblea . The
membership of action a in Actv or Acth is determined by the
boolean value of visiblea in the current state. Visibility predicates can be set to True only through empowering actions, or
by the hazards themselves, though they can be set to False by
all actions and hazards.
Apart from the actions, the knowledge base maintains a
history of hazards that have occurred and a set of visibility
predicates called internal goals. The latter are essentially the
predicates which can possibly be achieved through planning
and execution of empowering actions. Through this, the Manager can make some of the hidden actions visible in the future
in a deliberate and strategic way.
5.2
Monitor
The normal role of Monitor in MAPE-K is to access the current state of the system and detect anomalies if any. Here,
the Monitor detects a hazard through the inconsistency of the
current state (c) and the desired precondition of the action-tobe-executed. In case of a hazard, it records the hazard (e, c) in
the knowledge base, where e is the state that had been reached
after the execution of the previous action.
5.3
Analyzer
The design decisions we mentioned are mostly in the Analyzer module. The Analyzer refers to the history of the hazards, domain knowledge including cost etc (which could be
based upon analytics, for example, and taking into account
the impact of the hazard on the achievement of the current
goal) and outputs the following in the event of a hazard:
Planner
Given the current state, a goal state, set of actions, a set of
hazards and the recommendation from the analyzer, the Planner synthesizes a plan using only the visible actions. First,
it determines the current state space from the predicates used
in Actv . Note that since Actv is dynamic, the state space
also changes dynamically. It is expected that the Monitor can
query the environment and find the correct current state at
any level of detail (e.g. accessing the appropriate sensors and
processing modules outputting the predicates).
Plan synthesis may fail because of several reasons. All of
visible or hidden actions may not be sufficient to provide new
plans. This is a pathological case where external support is
mandatory. There may not be plans satisfying the recommendation in which case the best effort improvement is done. For
example, if a robust plan cannot be found, the Planner tries
to find a resilient one. Or, if robust plan cannot be found for
all missions, it finds one for the current mission and so on.
Specific behavior depends upon Planner policies.
5.5
Executor
The Executor module is pretty simple: given a plan which
is a labeled sequence of transitions in Sact , it executes the
actions associated with the labels in the sequence order taking
help of the Monitor to detect hazards. If there is a hazard,
the Executor halts the execution, signals the Manager that a
hazard has taken place and appends the hazard to the history
in the knowledge base. In the best case, the Executor finishes
executing the plan without countering any hazard and signals
the environment about plan completion. The Environment
then can issue a new goal which is handled by Autonomic
Manager.
When the Manager decides to issue a reset action, it is issued directly to the Executor. Just as it executes a plan, the
Executor executes the reset action which leads to a waypoint
(a special goal state). The reset completion signal to the Environment allows it to continue the game with a new goal.
5.6
Autonomic Manager
Autonomic manager is essentially an orchestrator for other
components in MAPE-K, but has a greater role in antifragile
intelligent systems since the life cycle is no longer just a loop.
The reactive behavior of the Manager is as follows:
1. It first determines a set of hazards (including the present
one) that needs to be handled, by predicting and/or correlating with the present hazard.
1. When E issues a new goal, Manager passes the goal
along with the Actv , corresponding current state, hazards and hW, D, M i recommendations to the Planner for
plan synthesis.
2. It outputs for each hazard a three-tuple hW, D, M i
where W denotes immediate or future improvement
2. When E introduces a hazard, the Manager gets the signal from the Executor. It invokes the Planner as in (1)
to first find if there is a resilient plan already in the current SActv . If no such plan exists, then it invokes the
Analyzer to get possibly a set of hazards H 0 with the
hW, D, M i recommendations. For space reasons, assume that there is a single hazard to be handled. The
case of multiple hazards is routine though tedious.
(a) Manager invokes Planner to use both visible and
hidden actions to come up with a plan P with minimum number of hidden actions if possible depending upon mode M of the hazard h. If it is possible
to synthesize a plan, then it records the visibility
predicates P redv (h) of the hidden actions in P .
(b) If W==Later, then Manager issues reset to the Executor, but in a separate, parallel thread triggers the
Planner to achieve the visibility predicates and execute the plan subsequently. This is so that during
the planning and execution for other missions, hidden actions are enabled gradually.
(c) If W==Now, then Manager triggers the planner to
first achieve P redv (h), which ensures new actions
in Actv and then invoking P for the current goal.
(d) If D == Current, then the visibility predicates are
toggled by the Manager as soon as the current goal
is achieved.
Thus, whenever there is a hazard, if there is a possibility of
having a robust or resilient system w.r.t. the hazard using
extra hidden actions, those actions are made visible. This ensures that whenever there is a new goal, the system is able
to find a plan (robust or resilient) for the goal. Note that for
completely different goals, the system may not find robust/resilient plans in SAct , hence a similar exercise is carried out to
make more actions visible.
In the antifragility literature, it is noted that certain systems may perform better under hazards (e.g. adrenaline rush
enabling great strength in stress conditions). In a way, this
attribute is captured in our setup by the fact that a hazard may
lead to a state where certain visibility predicates are set to
True and hence the agent has access to (possibly) more number of visible actions for planning. On the other hand, building up resources to enable more strength or immunity is a deliberate process and this is captured by the internal planning
with empowering actions.
6
Case Study
We illustrate the concepts discussed in this paper within
a path planning scenario. We consider a wheeled robot
A autonomously navigating in its world (e.g., a warehouse).
The mission is split into a sequence of goals, with the next
goal being planned for once the current goal is successfully
achieved. Each goal is nothing but a position the robot has to
reach on a sparse 5×6 grid of equally sized cells.
In a simplistic scenario, the robot has the following visible actions: MOVE and TURN using which it navigates the
sparse grid. Traditional path planning is used to compute the
optimal sequence of actions that take the robot from an initial
position to the given goal. The robot’s behavior also has two
other actions smallMOVE and smallTURN. These two are
initially not visible (hidden actions). These two require additional support from the environment during execution, for
example they need sensors to locate the robot on a finer grid.
Therefore we assume additional empowering actions are typically needed before one can use these hidden actions; the
preconditions for these actions account for that requirement
through appropriate predicates.
During execution, A follows the computed plan. An oil
spill, unknown at the time of planning, is encountered, which
throws the robot off the planned path and hence the original
plan is not applicable (shown by point (1) in Figure 3(b)).
Thus the robot discovers it has encountered a hazard.
The Autonomic Manager now uses external information
(camera images, image analytics etc.) to identify that the
hazard is an oil spill and uses domain knowledge to derive
the fact that MOVE and TURN actions cannot be used in the
oil spill. Therefore it makes these actions Hidden, and sets
the visibility predicates of smallMOVE and smallTURN as
an internal goal. The Planner uses the prerequisite empowering actions to prepare the current state and ultimately makes
the smallMOVE and smallTURN actions visible. The robot’s
domain view when these actions are enabled is shown in Figure 3(b). To support this world view, note that the environment needs to be able to provide the right kind of sensory
information (e.g., fine grid positions).
For the purposes of this illustration, we assume that
smallMOVE and smallTURN actions are deterministic and
have the intended effect2 . Various possibilities exist for replanning with these additional actions and it depends upon
the Planner. As shown in Figure 3(b), the robot can navigate
from within the oil spill, with smallMOVE and smallTURN
actions to a known state on the earlier plan (red path), it can
come out of the oil spill as early as possible and then navigate
the the known state (yellow path) and it could directly find
the a path to the current goal, abandoning the earlier plan. As
we have mentioned, getting out of the oil spill may be treated
as a hazard as a response to which MOVE and TURN actions
be made visible, and the robot now has all the four actions at
its disposal for planning the new paths.
Figure 4 shows a second goal sequence after the first goal
has been successfully reached. We note that the path avoids
known oil spills, due to the awareness of this new entity in
the environment. While following the path, a new oil spill
that did not exist at plan-time may be encountered. In this
case, the robot uses its newly visible actions to easily cope.
In Figure 4, the red curve (4) shows the path using these new
capabilities. We also note that these new actions may also be
used as planned actions to navigate through known oil spills
even at original plan time possibly resulting in a lower cost
path. Thus in reaching the second goal if the red oil spill
was known, it could still have chosen the path through the oil
spill, the blue line - (5), using the newly visible actions for
this stretch appropriately.
As mentioned in the Implementation, Analyzer with the
help of Knowledge module can generalize the specific oil
spill hazard to other hazards such as ice, slippery sand etc.
2
A reasonable assumption when referring to sufficiently small
steps on slippery surfaces.
(a) Originally planned path
(b) Hazard mode triggered by oil-spill hazard
Figure 3: Robot path planning example: first goal sequence
Figure 4: Robot path planning example: second goal sequence
by identifying that the hazard consequence state is similar
and also from the knowledge that these surfaces have similar
properties. With this, similar uncovering of hidden actions
can be done for all such hazards. This shows the dimension
of the antifragile architecture where occurrence of one hazard
builds antifragility for a class of other hazards.
This illustrative example has shown that it is possible for
an intelligent system to develop antifragility through the suggested refinements and suitable design decisions in the base
MAPE-K architecture.
7
Summary and Further Work
In this paper we have defined antifragility in the context of
intelligent systems which are goal-directed and based on AI
planning. We have shown that antifragile intelligent systems
can be built using the MAPE-K architecture though with several additional features. Primarily, the available action set
should be divided into empowering, visible, and hidden subsets. Depending on the strategy of the Autonomic Manager
module, the system is made robust or resilient to the hazard in different manners, correctly capturing the essence of
antifragility. The proposed architecture has been illustrated
through a robotics path planning example, where a robotic
agent A performs better after encountering a hazard.
One immediate problem arises that must be studied: that of
optimization of the selection process through which hidden
actions are made visible. Efficiency of algorithms to select
the minimal number of new actions that will allow robustness/resilience is critical from the view of real-time systems.
We observe that the improvement of the systems is triggered by hazards in antifragility literature. In an indirect way,
by defining hazards as any violation of cost metrics, one can
bring in improvements through new actions that will result
in more efficient plans even though the system does not encounter “unforeseen” hazards. The simian army in [Abid et
al., 2014] used to inject faults and create hazards to improve
the systems is similar in spirit.
We note that this work is an initial effort to study the concept of antifragility in the context of intelligent systems, and
therefore is rather preliminary. Undoubtedly, rigorous implementation and verification of the developed ideas is needed.
In addition, the notion of performance improvement from antifragility, albeit intuitively true, must be quantified in the current context. We believe that the real value of building antifragile systems will become evident as we develop complex
multi-agent systems. Therefore, distributed antifragility, built
into a population of systems, is an interesting future direction.
The potential applications of a system that improves from the
presence of hazards is vast. We hope that the promise of
antifragile intelligent systems will stimulate further research
within the intelligent systems community.
References
[Abid et al., 2014] Amal Abid, Mouna Torjmen Khemakhem, Soumaya Marzouk, Maher Ben Jemaa, Thierry
Monteil, and Khalil Drira. Toward antifragile cloud
computing infrastructures. Procedia Computer Science,
32(Antifragile):850–855, 2014.
[Clancy, 2015] Thomas R. Clancy. Complexity, Flow, and
Antifragile Healthcare Systems. JONA: The Journal of
Nursing Administration, 45(4):188–191, 2015.
[Fang and Sansavini, 2017] Yiping Fang and Giovanni
Sansavini. Emergence of antifragility by optimum postdisruption restoration planning of infrastructure networks.
Journal of Infrastructure Systems, 23(4):04017024, 2017.
[Hayes-Roth, 1995] Barbara Hayes-Roth. An architecture
for adaptive intelligent systems. Artificial Intelligence,
72(1-2):329–365, 1995.
[IBM, 2005] An architectural blueprint for autonomic computing. Technical report, IBM, 2005.
[Johnson and Gheorghe, 2013] John Johnson and Adrian V.
Gheorghe. Antifragility analysis and measurement framework for systems of systems. International Journal of Disaster Risk Science, 4(4):159–168, 2013.
[Jones, 2014] Kennie H. Jones. Engineering antifragile systems: A change in design philosophy. Procedia Computer
Science, 32(Antifragile):870–875, 2014.
[Lichtman et al., 2016] Marc Lichtman, Matthew T. Vondal,
T. Charles Clancy, and Jeffrey H. Reed. Antifragile Communications. IEEE Systems Journal, pages 1–12, 2016.
[Monperrus, 2017] Martin Monperrus. Principles of antifragile software. In Companion to the first International
Conference on the Art, Science and Engineering of Programming, page 32. ACM, 2017.
[Russell et al., 2015] Stuart Russell, Daniel Dewey, and Max
Tegmark. Research priorities for robust and beneficial artificial intelligence. AI Magazine, 36(4):105–114, 2015.
[Taleb, 2012] Nassim Nicholas Taleb. Antifragile: Things
that gain from disorder, volume 3. Random House Incorporated, 2012.
| 2cs.AI
|
A Polynomial Algorithm for Balanced Clustering via Graph
Partitioning∗
Luis Evaristo Caraballo†
José-Miguel Dı́az-Báñez‡
Nadine Kroher§
arXiv:1801.03347v2 [cs.CG] 25 Jan 2018
January 26, 2018
Abstract
The objective of clustering is to discover natural groups in datasets and to identify geometrical structures which might reside there, without assuming any prior knowledge on the
characteristics of the data. The problem can be seen as detecting the inherent separations
between groups of a given point set in a metric space governed by a similarity function. The
pairwise similarities between all data objects form a weighted graph adjacency matrix which
contains all necessary information for the clustering process, which can consequently be formulated as a graph partitioning problem. In this context, we propose a new cluster quality measure
which uses the maximum spanning tree and allows us to compute the optimal clustering under
the min-max principle in polynomial time. Our algorithm can be applied when a load-balanced
clustering is required.
1
Introduction
The objective of clustering is to divide a given dataset into groups of similar objects in an unsupervised manner. Clustering techniques find frequent application in various areas, including
computational biology, computer vision, data mining, gene expression analysis, text mining, social
network analysis, VLSI design, and web indexing, to name just a few. Commonly, a metric is used
to compute pair-wise similarities between all items and the clustering task is formulated as a graph
partitioning problem, where a complete graph is generated from the similarity matrix. In fact,
many graph-theoretical methods have been developed in the context of detecting and describing
inherent cluster structures in arbitrary point sets using a distance function[2].
Here, we propose a novel clustering algorithm based on a quality measure that uses the maximum
spanning tree of the underlying weighted graph and addresses a balanced grouping with the minmax principle. More specifically, we aim to detect clusters which are balanced with respect to
∗
This research has received funding from the projects COFLA2 (Junta de Andalucı́a, P12-TIC-1362) and GALGO
(Spanish Ministry of Economy and Competitiveness and MTM2016-76272-R AEI/FEDER,UE).
This work has also received funding from the European Union’s Horizon 2020 research and innovation
programme under the Marie Sklodowska-Curie grant agreement No 734922.
†
Department of Applied Mathematics II, University of Seville, Spain. L.E.C. is funded by the Spanish Government
under the FPU grant agreement FPU14/04705. Email: [email protected].
‡
Department of Applied Mathematics II, University of Seville, Spain. Email: [email protected].
§
Department of Applied Mathematics II, University of Seville, Spain. Email: [email protected].
1
Figure 1: Illustration of the desired cluster properties: The ratios of inner variance to distance to
other clusters is balanced among groups. Clusters 1 and 4 exhibit a higher variance but are also
further apart from the other clusters.
their ratio of intra-cluster variance to their distance to other data instances. In other words, we
allow clusters with a weaker inner edges to be formed, if they are located at large distance of other
clusters (Figure 1). We prove that an optimal clustering under this measure can be computed in
polynomial time using dynamic programming.
Such cluster properties are typically desired when grouping sensors in wireless sensor networks [1]: Each group communicates only with sensors in the same cluster and streams its information to a single command node located inside the cluster. The power consumption of sensors
heavily depends on their distance to the command node and ideally, a balanced consumption among
sensors is desirable. Consequently, a grouping should be balanced with respect to the ratio of interconnection (sharing information between clusters) and intra-connection (sharing information within
a cluster). A similar scenario occurs in the context of multi-robot task allocation in cooperative
robotics, where the goal is to allocate tasks to robots while minimizing costs. For example, in
monitoring missions, using a cooperative team of Unmanned Aerial Vehicles (UAVs), the goal is
to minimize the elapsed time between two consecutive observations of any point in the area. Techniques based in area partitioning achieve this by assigning a sub-area to each UAV according to its
capabilities. In this scenario, a load balanced clustering extends the life of the agents and allows
to perform the task in a distributed manner [8, 3]. Another possible application area arises from
the field of Music Information Retrieval[5], where several applications rely on the unsupervised
discovery of similar (but not identical) melodies or melodic fragments. In this context, clustering
methods can be used to explore large music collections with respect to melodic similarity, or to
detect repeated melodic patterns within a composition[7].
1.1
Related work
Graph clustering refers to the task of partitioning a given graph G(V, E) into a set of k clusters
C = {C1 , . . . , Ck } in such a way that vertices within a cluster are strongly connected whereas clusters are well separated. A number of exact and approximate algorithms have been proposed for
this task, targeting different types of graphs (directed vs. undirected, complete vs. incomplete,
etc.) and optimizing different cluster fitness values (i.e. maximizing densities, minimizing cuts).
For a complete overview of existing strategies and their taxonomy, we refer to [9]. The algorithm
proposed in this study operates on the maximum spanning tree M ST (G) of graph G. The idea
of using minimum or maximum spanning trees (when working with distances or similarities, respectively) for cluster analysis goes back as far as 1971, when Zahn demonstrated [12] various
properties which indicate that the minimum spanning tree serves as a suitable starting point for
graph clustering algorithms and proposes a segmentation algorithm based on a local edge weight
inconsistency criterion. This criterion was revisited and improved in [6]. Asano et al. [2] show
2
that both, the optimal partitioning minimizing the maximum intra-cluster distances and the partitioning maximizing the minimum inter-cluster distance, can be computed from the maximum and
minimum spanning trees. In the context of image processing, Xu et al. [11] propose a dynamic
programming algorithm for segmenting gray-level images which minimizes gray level variance in
the resulting subtrees. Felzenszwalb et al. [4] introduced a comparison predicate which serves as
evidence for a cluster boundary and provide a bottom-up clustering algorithm in O(n log n) time.
In the context of Gene Expression Data Clustering, Xu et al. [10] proposed three algorithms for
partitioning the minimum spanning tree optimizing different quality criteria.
2
Problem statement
Let V = {v1 , v2 , . . . , vn } be a set of points or nodes in a metric space and suppose that there exists a
function to estimate the similarity between two nodes. Let A be the matrix of similarity computed
for every pair of elements in V . The value A[i, j] is the similarity between the nodes vi and vj . If
A[i, j] > A[i, l], then the node vi is more similar to vj than vl . Our goal is to create groups, such
that similar nodes are in the same cluster and dissimilar nodes are in separate clusters.
Let G = (V, E, w) be a weighted and undirected graph induced by the similarity matrix A on the
set of nodes V . In the sequel, such graphs are simply referred to as “graph”. If E is the set of edges
and contains an edge for every unordered pair of nodes, and w is a weight function w : E −→ (0, 1)
such that w(e) is the similarity between the nodes connected by e (i.e. if e = {vi , vj }, then
w(e) = A[i, j]).
Let C ⊆ V be a cluster. The outgoing edges set of C, denoted by Out(C) , is the set of edges
connecting C with V \ C. Let M ST (C) be the maximum spanning tree of C. Let max(Out(C))
and min(M ST (C)) be the weights of the heaviest and lightest edges of Out(C) and M ST (C),
respectively.
We can use the following function as a quality measure of a cluster C:
0
if C = V,
max(Out(C)) if |C| = 1,
Φ(C) =
max(Out(C))
in other case
min(M ST (C))
Note that higher values of Φ(·) correspond to worse clusters. Also, note that, if |C| = 1 then
max(Out(C))
we can consider min(M ST (C)) = 1 and then Φ(C) = max(Out(C)) = min(M
ST (C)) . In addition, if
max(Out(C))
C = V then we can consider max(Out(C)) = 0 and then Φ(C) = 0 = min(M
ST (C)) .
Let C = {C1 . . . , Ck } be a clustering formed by k > 1 clusters of G. To evaluate the quality of
k
C we use the quality of the worst cluster of C, that is, Φ(C) = max{Φ(Ci )}.
i=1
Denoting the set of all possible k-clusterings (clustering scenarios formed by k clusters) on G
by P(k , G), we state the following optimization problem (MinMax Clustering Problem):
min Φ(C)
subject to: C ∈ P(k , G).
3
(1)
A0
A
v0
v
Figure 2: Representation of a graph’s bipartition. The crossing edges of this bipartition are indicated by dashed lines.
When the value of k is unknown, the problem can be stated as follows:
min Φ(C)
n
[
P(k , G).
subject to: C ∈
(2)
k=2
That is, to find the clustering C ∈ P(k , G) such that Φ(C) is minimum among all possible clusterings
with more than one cluster irrespective of the number of clusters contained in it.
3
Properties of the optimal clustering
Note that the problems stated above can be generalized to connected (not necessarily complete)
graphs, by simply setting P(k , G) as the set of all the possible partitions of G in k connected
components.
Lemma 1. Let G be a graph and let C ∗ be an optimal clustering of G for Problem 1 in P(k , G).
Then, Φ(C ∗ ) ≤ 1.
Proof. If k = 1, Φ(C ∗ ) ≤ 1. If k > 1, then take the maximum-spanning-tree of G and denote it
by T = M ST (G). Let C = {C1 , . . . , Ck } be the clustering induced by the k connected components
obtained by removing (or “cutting”) the k − 1 lightest edges from T . If |Ci | = 1 then Φ(Ci ) < 1
since w(e) ∈ (0, 1) for all edge e in the graph. If |Ci | > 1, then min(M ST (Ci )) ≥ max(Out(Ci ))
(according to properties of a maximum-spanning-tree), therefore Φ(C) ≤ 1. The result follows.
Before showing the next lemma, we recall the definition of the crossing edge. Let G = (V, E)
be a graph and let B = {A, A0 } be a bipartition of V . An edge {v, v 0 } ∈ E is a crossing edge of B
if v ∈ A and v 0 ∈ A0 (see Figure 2).
Lemma 2. Let G be a graph and let C ∗ be an optimal clustering of G for Problem 1. Let C be a
cluster in C ∗ . If |C| > 1, then every bipartition of C has a crossing edge in a maximum spanning
tree of G.
Proof. We prove by contradiction. Let C be a cluster in C ∗ with cardinality greater than 1 and let
{A, A0 } be a partition of C such that there is no edge from A to A0 in a maximum spanning tree of G.
4
Let e be the heaviest edge that crosses from A to A0 . Let M ST (G) be a spanning tree of G. Due to
our assumption, e ∈
/ M ST (G) and therefore, adding e to M ST (G) results in a cycle. All other edges
in this cycle have a weight equal or greater than e, given by the properties of the maximum spanning
tree. Let e0 be an edge in this cycle connecting a node in C with others in V \ C. If w(e) = w(e0 ),
then replacing e0 by e in M ST (G) we obtain another maximum spanning tree containing e and thus,
this is a contradiction. If w(e) < w(e0 ), then max(Out(C)) ≥ w(e0 ) > w(e) ≥ min(M ST (C)).This
is a contradiction by Lemma 1.
Theorem 3. Let G be a graph and let C ∗ ∈ P(k , G) be an optimal clustering of G for Problem 1.
For every cluster C ∈ C ∗ , the maximum spanning tree of C is a subtree of a maximum spanning
tree of G and the heaviest outgoing edge of C is in a maximum spanning tree of G.
Proof. Let C be a cluster of C ∗ . If |C| = 1 then, obviously, M ST (C) ⊂ M ST (G). If |C| > 1 then
M ST (C) ⊆ M ST (G) by Lemma 2. The second part of the theorem, claiming that the heaviest
edge in Out(C) is in M ST (G), is deduced from the properties of the maximum spanning tree of a
graph.
The following result is directly deduced from the theorem above.
S
Corollary 4. Let G be a graph and let C ∗ ∈ nk=1 P(k , G) be an optimal clustering of G for
Problem 2. For every cluster C ∈ C ∗ , the maximum spanning tree of C is a subtree of a maximum
spanning tree of G and the heaviest outgoing edge of C is in a maximum spanning tree of G.
v1
C1
v1
v2
v3
v2
v3
v7 C3
v4
v8
v7 C3
v4
v9
v5
C2
C1
C2
v6
(a)
v8
v9
v5
v6
(b)
Figure 3: (a) A graph G and a spanning tree T of G. The edges of T are bold. The clustering
C = {C1 , C2 , C3 } is represented by dotted strokes. (b) Obtaining an optimal clustering C ∗ =
{C1 , C2 , C3 } by cutting two edges in the maximum spanning tree.
Let G be a graph and let T be a spanning tree of G. Note that every possible clustering of T is a
valid clustering in G and therefore P(k , T ) ⊆ P(k , G) (see Figure 3a). However, a valid clustering
of G may be not feasible for T , for example, the clustering {{v1 , v3 , v5 }, {v4 , v6 }, {v2 , v7 , v8 , v9 }} is
valid for the graph G in Figure 3a, but is not feasible for T because the cluster {v1 , v3 , v5 } does
not constitute a connected component in T . Let C be a valid cluster for G and T . Consider
OutT (C) as the set of outgoing edges of C as described earlier, but restricted to the set of edges
forming T . For example, considering G and T in Figure 3a, the set OutT (C1 ) only contains the
edge {v3 , v4 }; however, the set Out(C1 ) contains {v3 , v4 }, {v1 , v5 }, {v1 , v6 }, . . . Analogously, we can
5
apply the same argument to the set of inner edges in the cluster C. We use the analogous notations
M STT (·) to denote the maximum spanning tree of a cluster using only the edges in T ; also, note that
M STT (C) is the subtree of T determined by the nodes of C. Consequently M STG (C) = M ST (C)
and OutG (C) = Out(C).
The previous explanation is needed to introduce the following notions: Let T be a spanning
tree of a graph G. Let C ∈ P(k , T ) be a clustering of T . The evaluation function ΦT (C) operates as
usual, but is restricted to the set of edges forming T . Therefore, the optimal solution for Problem 1
on T is C † ∈ P(k , T ) such that ΦT (C † ) ≤ ΦT (C) for every other clustering C ∈ P(k , T ).
Theorem 5. Let G be a graph and let T be a maximum spanning tree of G. If C ∗ ∈ P(k , G)
and C † ∈ P(k , T ) are the optimal clusterings (for Problem 1) on G and T , respectively; then
Φ(C ∗ ) = ΦT (C † ).
Proof. Let us first prove that Φ(C ∗ ) ≥ ΦT (C † ). Suppose that every cluster C ∈ C ∗ contains a single
connected component in T . Then, for every cluster C ∈ C ∗ , min(M STG (C)) = min(M STT (C))
follows from properties of a maximum spanning tree. Moreover max(OutG (C)) ≥ max(OutT (C))
(because T is a subgraph of G). Therefore, Φ(C) ≥ ΦT (C), which implies Φ(C ∗ ) ≥ ΦT (C ∗ ) ≥
ΦT (C † ).
Suppose now, that some cluster C ∈ C ∗ does not contain a single connected component in T . For
an illustration, see Figure 4 where the cluster C encompasses parts of two connected components
in T . Take an edge e ∈ M ST (C) connecting nodes in two of the different connected components
determined by C in T . Adding e to T will form a cycle in the edges of T . If there exists an edge e0
such that w(e) > w(e0 ) in this cycle, then T is not a maximum spanning tree. Consequently, for all
edges e0 6= e in this cycle w(e0 ) ≥ w(e). Note that this cycle includes edges in Out(C) (if all edges of
this cycle were inner edges of C, then e would connect nodes within the same connected component,
posing a contradiction), then max(Out(C)) ≥ w(e) ≥ min(M ST (C)). Thus 1 ≤ Φ(C) ≤ Φ(C ∗ ).
On the other hand, ΦT (C † ) ≤ 1 by using Lemma 1, so, ΦT (C † ) ≤ Φ(C ∗ ).
Finally, let us prove that Φ(C ∗ ) ≤ ΦT (C † ). For every cluster C ∈ C † , min(M STG (C)) =
min(M STT (C)) and max(OutG (C)) = max(OutT (G)) can be demonstrated using the properties of
the maximum spanning tree. Therefore Φ(C) = ΦT (C), which implies that Φ(C † ) = ΦT (C † ). Then
Φ(C ∗ ) ≤ Φ(C † )m because C ∗ is the optimal clustering on G. This completes the proof.
The following result is directly deduced from the theorem above.
S
Corollary
6. Let G be a graph and let T be a maximum spanning tree of G. If C ∗ ∈ nk=1 P(k , G)
S
and C † ∈ nk=1 P(k , T ) are the optimal clusterings (for Problem 2) on G and T , respectively; then
Φ(C ∗ ) = ΦT (C † ).
As a consequence of the above properties, the optimal k-clustering C ∗ for Problem 1 can be
obtained by “cutting” the k −1 appropriate edges in a maximum spanning tree of G (see Figure 3b).
These edges can be found combinatorially in O(nk−1 ) time. Thus, using a naive approach, the
solution of Problem 2 can be found in O(nn−1 ) time. In the next section, we show an algorithm
which solves both problems in polynomial time in n and k.
4
The algorithm
First, recall that Theorem 5 and Corollary 6 provide a nice property, which allows us to reduce
Problems 1 and 2 from a graph to its maximum spanning tree. Consequently, given a similarity
6
e
C
e0
e
C
e00
(a)
(b)
Figure 4: Representation of a graph G and a clustering C ∗ . Dotted black strokes mark clustering
boundaries. (a) The maximum spanning trees of the clusters are drawn in red. (b) A maximum
spanning tree T of G is drawn in blue. The two connected components partially covered by cluster
C in T are shaded gray.
graph, we can operate on its maximum spanning tree T = (V, E, w). From now on, we will use
E to denote the set of edges in the maximum spanning tree. Observe that every cluster C in T
determines only one subtree of T . Then, using M ST (C) to denote the maximum spanning tree in
C may be confusing or redundant. Therefore, instead of using M ST (C), we will use E(C) (set of
edges connecting nodes in C).
The following technical lemma is crucial for the correctness of our algorithm.
Lemma 7. Let C be a clustering of a tree T = (V, E, w). By removing an edge of T we induce two
clusterings, one for each generated subtree (see Figure 5). The evaluations of the induced clusterings
are at most Φ(C).
Proof. Let A and B denote the two induced clusterings and let e be the removed edge (see Figure 5).
For the sake of contradiction, suppose that one of the two induced clusterings has an evaluation
greater than C. W.l.o.g. assume that Φ(A) > Φ(C). Let A denote the cluster of A containing one
of the incident nodes of e (the other one is in a cluster of B).
If Φ(A) > Φ(A), then there is another cluster A0 ∈ A such that Φ(A0 ) = Φ(A). Note that A0 is
also in C and A0 is not affected when e is removed, so, Φ(C) ≥ Φ(A0 ). This is a contradiction since
we are assuming that Φ(A0 ) = Φ(A) > Φ(C).
If Φ(A) = Φ(A), then let mA = min(E(A)) (recall that we consider mA = 1 if |A| = 1) and
MA = max(Out(A)). Observe that the incident nodes of e may be both, in the same cluster of C
(see Figure 5a), or not (see Figure 5b). Suppose the incident nodes of e are in the same cluster
C ∈ C. Note that min(E(C)) ≤ mA and max(E(C)) ≥ MA , therefore
Φ(C) ≥ Φ(C) =
max(Out(C))
MA
≥
= Φ(A) = Φ(A).
min(E(C))
mA
This is another contradiction.
7
A B
A
C
e
(a)
A
A
A
B
e
(b)
Figure 5: Obtaining two clusterings, one per subtree, by removing an edge of a given clustering.
The removed edge in (a) is inside a cluster, and in (b) is a crossing edge.
Now, suppose that the incident nodes of e are in different clusters of C. In this case A ⊂ C
(C)
and then A is also in C. Let MA denote the weight of the heaviest outgoing edge of A in C, then
(C)
MA = max{w(e), MA } ≥ MA , and the evaluation of A in C is Φ(C) (A) =
(C)
MA
mA
, therefore
(C)
Φ(C) ≥ Φ(C) (A) =
MA
MA
≥
= Φ(A) = Φ(A).
mA
mA
This is a contradiction and completes the proof.
The proposed algorithm is based on dynamic programming. We show, that the stated problems
have an optimal substructure and construct the optimal solution in T from optimal solutions for
subtrees of T . From here on, we consider that the tree T is rooted at an arbitrary node r ∈ V . For
all v ∈ V , let c(v) be the set of children of v; and for all v ∈ V, v 6= r, let p(v) be the parent of v.
Recall that if c(v) is empty then we say that v is a leaf node.
Given a tree T , let S be a subtree of T and let v be the node with minimum depth in S. Then
we say that S is rooted at v. In the sequel, we only consider subtrees S rooted at v that contain
all the descendants of vertices v 0 ∈ S \ {v}. Figure 6a shows a subtree S 0 rooted at v. The leaves
hanging from v 0 and v 00 are not in S 0 , so S 0 is not considered as a subtree. Figure 6b shows an
example of a subtree to be considered. In addition, we say that S = Tv if S is rooted at v and
contains all the descendants of v.
The main idea of our algorithm is to work on (local) clusterings of a subtree and perform a
bottom-up dynamic programming strategy with two basic operations:
• UpToParent: knowing an optimal clustering of the subtree S = Tv 6= T , compute an optimal
clustering of the subtree formed by adding p(v) to S, see Figure 10a.
8
v
v
v
v 00
g→
v0
S0
S
S
(a)
(b)
(c)
Figure 6: (a) A subtree S 0 which is not considered. (b) A considered subtree S. (c) Representation
of a clustering C of S. The head cluster is above the curve g. The edges of OutS (h(C)) are the edges
in S crossed by g. The clusters of C that are below g constitute the headless clustering C \ {h(C)}.
• AddChildTree: knowing an optimal clustering of a subtree S rooted at p(v) and knowing
an optimal clustering of the subtree Q = Tv such that v 6∈ S, compute an optimal clustering
of the subtree formed by joining S and Q, see Figure 10b.
Now, we elaborate on a the (local) clustering C of a subtree S rooted at v, see Figure 6c.
A clustering of S is given by cutting some edges. We call a cluster containing the node v head
cluster of C, denoted h(C) (see Figure 6c). Note that if C ∈ C and C 6= h(C), then Out(C) is
entirely contained in S. However, Out(h(C)) is entirely contained in S only if S = T ; if v 6= r then
{v, p(v)} ∈ Out(h(C)) and {v, p(v)} is not in S; if v = r and S 6= T , then some node v 0 ∈ c(v) is not
in S and {v, v 0 } ∈ Out(h(C)) but {v, v 0 } is not in S. Thus, it is convenient to introduce OutS (C)
as the set of outgoing edges of C connecting nodes in S. In Figure 6c, OutS (h(C)) is formed by the
edges stabbed by the curve g.
Given a clustering C of a subtree S, let M be the weight of the heaviest edge in OutS (h(C)),
that is, M = max(OutS (h(C))). If h(C) contains all the nodes in S, then there are no descending
outgoing edges, and in these cases we set M = 0. On the other hand, let µ be the weight of
the lightest edge in E(h(C)), that is µ = min(E(h(C))). If h(C) is formed by single node, that is
h(C) = {v}, then E(h(C)) is empty, and in these cases we set µ = 1. For convenience, we introduce
the functions ΦS (·) and ΦS (·) as the restricted quality measures of a cluster and a clustering,
respectively. They work as usual but are restricted to the edges of the subtree S, thus:
ΦS (h(C)) =
M
.
µ
(3)
Note that if S = T , then ΦS (h(C)) = Φ(h(C)). If S = Tv 6= T , then:
Φ(h(C)) =
max{M, w({v, p(v)})}
.
µ
For every cluster C ∈ C, such that C is not the head cluster, the usual evaluation and the restricted
9
one have the same value, Φ(C) = ΦS (C). Consequently, the restricted evaluation of the “headless”
clustering C \ {h(C)} is:
ΦS (C \ {h(C)}) = Φ(C \ {h(C)}) = max { Φ(C) | C ∈ C \ {h(C)} } ,
(4)
therefore, the restricted evaluation of the clustering C is:
ΦS (C) = max {ΦS (h(C)), Φ(C \ {h(C)})} .
(5)
Let S be a subtree of T , and let H(l , S , µ) denote the set of l-clusterings of S in which µ is the
weight of the lightest edge in the head cluster. That is:
H(l , S , µ) = { C | C ∈ P(l , S )
and µ = min(E(h(C))) } .
We are now ready to state an encoding of a local solution and the invariant that allows us to
apply dynamic programming:
Notation 8. Suppose H(l , S , µ) is not empty, then a clustering C in H(l , S , µ) is encoded by the
ordered pair OS (l, µ) = (M, b), if the following properties are fulfilled:
1. M = max(OutS (h(C))) and b = ΦS (C).
2. ΦS (C) = min { ΦS (C 0 ) | C 0 ∈ H(l , S , µ)} .
3. max(OutS (h(C))) = min { max(OutS (h(C 0 ))) | C 0 ∈ H(l , S , µ) and ΦS (C 0 ) = ΦS (C) } .
If H(l , S , µ) is empty, then OS (l, µ) = (∞, ∞), where ∞ indicates the “infinity” value.
By Lemma 1, an optimal clustering C ∗ of T has an evaluation Φ(C ∗ ) ≤ 1 and according to
Lemma 7, if a clustering C of a subtree S is used to build C ∗ , then ΦS (C) ≤ 1 too. Therefore, we
set OS (l, µ) as (∞, ∞) if 1 < min { ΦS (C) | C ∈ H(l , S , µ) }. Then, given a subtree S, OS (·, ·) is a
function whose domain is N[1,k] × (w(E) ∪ {1}) and image {(∞, ∞)} ∪ (w(E) ∪ {0}) × R[0,1] where
w(E) = { w(e) | e ∈ E }.
Remark 9. Some times, it is more convenient to see OS (·, ·) as a table of k rows with labels
1, 2, . . . , k and n columns with labels w(e1 ), w(e2 ), . . . , w(en−1 ), 1 where e1 , e2 , . . . , en−1 is a labeling
of the edges in E from the lightest one to the heaviest one. In this way OS (l, µ) refers to the cell
with row-label l and column-label µ and it is value is the corresponding ordered pair (M, b).
If OS (l, µ) = (M, b) 6= (∞, ∞), then, by using equations (3), (4) and (5), we obtain that OS (l, µ)
encodes a clustering C (not necessarily unique) where:
M
ΦS (C) = b = max
, Φ(C \ {h(C)}) .
(6)
µ
For the sake of simplicity, we use the following notation for OS (l, µ) = (M, b) (not necessarily
distinct from (∞, ∞)): OS (l, µ)[1] = M and OS (l, µ)[2] = b.
If we have the function OT , then the evaluation of the optimal clusterings for Problems 1 and
2 are:
min { OT (k, µ)[2] | µ ∈ w(E) ∪ {1} } , and
min OT (l, µ)[2] | l ∈ N[2,n] and µ ∈ w(E) ∪ {1} , respectively.
The following lemma is a useful technical result:
10
vi+1
vi
vi+1
g0
(a)
{
g
g0
v = v0
}|
{
z
g
P
z
vi
v = v0
}|
P
}|
{
z
P
v = v0
vi
g
g0
vi+1
(b)
(c)
Figure 7: Let C and C 0 be two different clusterings in a subtree S. The head clusters h(C) and h(C 0 )
are formed by the nodes above g and g 0 , respectively. Let P = (v = v0 , v1 , . . . , vi ) be a path in S.
In the three pictures e = {vi , vi+1 } ∈ OutS (h(C)), note that e is stabbed by g. (a) e ∈ OutS (h(C 0 )),
note that e is stabbed by g 0 . (b) e ∈ h(C 0 ), note that vi and vi+1 are both above g 0 . (c) Some edge
in P is in OutS (h(C 0 )), note that P is stabbed by g 0 .
Lemma 10. Let S be a subtree rooted at v. Let C and C 0 be two different clusterings of S such
that min(E(h(C))) = min(E(h(C 0 ))) = µ. If ΦS (C) < ΦS (C 0 ) ≤ 1 then max(OutS (h(C))) ≤
max(OutS (h(C 0 ))).
Proof. For the sake of contradiction, suppose that max(OutS (h(C))) > max(OutS (h(C 0 ))). Observe that µ ≥ max(OutS (h(C))) > max(OutS (h(C 0 ))), since ΦS (C) < ΦS (C 0 ) ≤ 1. Note that
max(OutS (h(C))) > max(OutS (h(C 0 ))) implies that max(OutS (h(C))) > 0 and then OutS (h(C)) is
not empty. Let e be one of the heaviest edges of OutS (h(C)). Let P = (v = v0 , v1 , . . . , vi ) denote a
path from v to vi such that e = {vi , vi+1 } (see Figure 7).
If none of the edges of P are in OutS (h(C 0 )), then e ∈ OutS (h(C 0 )) or e ∈ E(h(C 0 )).
If e ∈ OutS (h(C 0 )) (Figure 7a) then there is a contradiction, because
w(e) = max(OutS (h(C))) > max(OutS (h(C 0 ))).
If e ∈ E(h(C 0 )) (Figure 7b), using min(E(h(C 0 ))) = µ yields µ ≤ w(e) and using µ ≥
max(OutS (h(C))) leads to µ ≥ w(e), so, µ = w(e). Therefore, there is a contradiction
because:
w(e)
1=
= ΦS (h(C)) ≤ ΦS (C) < ΦS (C 0 ) ≤ 1.
µ
Suppose that an edge e0 ∈ P is in OutS (h(C 0 )) (Figure 7c). Note that w(e0 ) ≤ max(OutS (h(C 0 ))).
Also, note that e0 is in E(h(C)), and consequently, µ ≤ w(e0 ). Therefore µ ≤ w(e0 ) ≤ max(OutS (h(C 0 )))
which is a contradiction since µ ≥ max(OutS (h(C))) > max(OutS (h(C 0 ))).
From the previous lemma, the following result is deduced directly:
Corollary 11. Let S be a subtree of T . For a given value OS (l, µ) = (M, b) 6= (∞, ∞), every
l-clustering C ∈ H(l , S , µ) fulfills that: ΦS (C) ≥ b, and max(OutS (h(C))) ≥ M .
11
The following lemma is the key of the proposed dynamic programming:
Lemma 12. Let S be a subtree rooted at v. Let OS (l, µ) = (M, b) 6= (∞, ∞) and let C be an
l-clustering of S encoded by OS (l, µ). Let Q be a subtree of S rooted at v 0 ∈ c(v). By removing the
edge e = {v 0 , v} from S an l0 -clustering A of Q is induced. Let µ0 = min(E(h(A))). By replacing
A with a clustering B encoded by OQ (l0 , µ0 ) and restoring the edge e, a new clustering C 0 of S is
obtained, which is also encoded as OS (l, µ) = (M, b) (see Figures 8 and 9).
Proof. By using Corollary 11, we have that
ΦQ (B) ≤ ΦQ (A), and max(OutQ (h(B))) ≤ max(OutQ (h(A))).
(7)
By using Lemma 7, we have that ΦQ (A) ≤ b, then:
b ≥ ΦQ (A) ≥ ΦQ (h(A)) =
max(OutQ (h(A)))
µ0
(8)
Moreover, using observations (7) and (8) yields to:
ΦQ (B) ≤ b.
(9)
Obviously, C and C 0 are both l-clusterings. We need to prove that:
min(E(h(C))) = min(E(h(C 0 ))) = µ,
ΦS (C) = ΦS (C 0 ) = b, and
max(OutS (h(C))) = max(OutS (h(C 0 ))) = M.
We divide the rest of the proof into two parts according to the two possible situations when e is
going to be removed, e ∈ OutS (h(C)) (see Figure 8) or e ∈ E(h(C)) (see Figure 9).
Let us start with the first case. Note that, by replacing A by B, the head cluster is not affected.
Consequently, max(OutS (h(C))) = max(OutS (h(C 0 ))) = M , and min(E(h(C))) = min(E(h(C 0 ))) =
µ. Let us prove that ΦS (C) = ΦS (C 0 ) = b. By Corollary 11, it is enough to prove that ΦS (C) ≤ b for
every cluster C ∈ C 0 . For every cluster C ∈ C 0 such that C 6∈ B, we have that C is also contained in C,
so, ΦS (C) ≤ b. For every cluster C ∈ B, such that C 6= h(B) we have ΦS (C) = ΦQ (C) ≤ ΦQ (B) ≤ b
by observation (9). From observation (7) we can deduce that ΦS (h(B)) ≤ ΦS (h(A)) ≤ b.
Let us analyze the second case. Let D denote the induced clustering of the (remaining) subtree
rooted at v (see Figure 9). Note that:
min(E(h(C 0 ))) = min {min(E(h(D))), w(e), min(E(h(B)))} , and
min(E(h(C))) = min {min(E(h(D))), w(e), min(E(h(A)))} .
Notice that min(E(h(A))) = min(E(h(B))) = µ0 by construction, therefore, min(E(h(C 0 ))) =
min(E(h(C))). By Corollary 11, it is enough to prove that max(OutS (h(C 0 ))) ≤ M and ΦS (C) ≤ b
for every cluster C ∈ C 0 . Note that:
max(OutS (h(C 0 ))) = max {max(OutS (h(D))), max(OutS (h(B)))} , and
max(OutS (h(C))) = max {max(OutS (h(D))), max(OutS (h(A)))} .
12
vC
v
A
v0
v
B
v0
v0
(b)
(a)
0
vC
v0
(c)
(d)
Figure 8: Removing the edge e = {v, v 0 } when e connects nodes in different clusters. (a) Initial
situation. (b) Induced clustering A when e is removed. (c) Replacing A with another clustering B.
(d) Restoring the edge e and obtaining a new clustering C 0 .
vC
v
D
D
B
v0
v0
(a)
v
A
0
vC
v0
(b)
(c)
v0
(d)
Figure 9: Removing the edge e = {v, v 0 } when e is inside a cluster. (a) Initial situation. (b) Induced
clustering A when e is removed. (c) Replacing A with another clustering B. (d) Restoring the edge
e and obtaining a new clustering C 0 .
Since max(OutS (h(B))) ≤ max(OutS (h(A))), therefore,
max(OutS (h(C 0 ))) ≤ max(OutS (h(C))) = M.
Finally, for every cluster C ∈ C 0 \ {h(C 0 )} such that C 6∈ B, C is also contained in C, so, ΦS (C) ≤ b.
For every cluster C ∈ (C 0 \ {h(C 0 )}) ∩ B we have ΦS (C) = ΦQ (C) ≤ ΦQ (B) ≤ b by observation (9).
From observation (7) we can deduce that ΦS (h(C 0 )) ≤ ΦS (h(C)) ≤ b.
In the next subsections we show how to perform the operations UpToParent and AddChildTree.
In order to simplify the formulas in the next subsections we introduce the following total order:
Let (M, b) and (M 0 , b0 ) be two ordered pairs. We say that (M, b) = (M 0 , b0 ) if M = M 0 and b = b0 .
We say that (M, b) < (M 0 , b0 ) if b < b0 , or if b = b0 and M < M 0 .
4.1
UpToParent: computing OS from OS
Let S = Tv 6= T be a subtree and let S denote the tree formed by the union of S and p(v). In this
section we will show how to compute OS , assuming that we already know OS . Let ω = w({v, p(v)}).
If we are computing OS (l, µ), then:
13
v
v
P
S
S
S
(a)
Q
(b)
Figure 10: (a) Subtree S formed by the subtree S = Tv and p(v). (b) Subtree P formed by joining
the subtree Q = Tv and a subtree S rooted at p(v).
ω
ω
ω
ω
v
v
v
v
S
S
S
S
S
(a)
S
(b)
Figure 11: Construction of a clustering of S based in one of S. (a) The edge {v, p(v)} is cut. (b)
The edge {v, p(v)} is not cut.
Claim 13. If µ = 1, then:
max{ω, OS (l − 1, µ0 )[1]}
0
max ω, OS (l − 1, µ )[2],
OS (l, µ) = ω, min
µ0
µ0
Proof. Let C be a clustering encoded as OS (l, µ) = (M, b). We have that h(C) = {p(v)} because
µ = 1. Then, the edge {v, p(v)} is cut and M = ω, see Figure 11a. Using (6) leads to b =
max{ω, Φ(C \{h(C)})}. Note that C \{h(C)} is an (l−1)-clustering C 0 of S. According to Lemma 12,
C 0 is encoded as OS (l − 1, µ0 ) for some µ0 , and then:
ΦS (C 0 ) = OS (l − 1, µ0 )[2] and Φ(h(C 0 )) =
max{ω, OS (l − 1, µ0 )[1]}
.
µ0
It is easy to see that:
Φ(C 0 ) = max Φ(h(C 0 )), Φ(C 0 \ {h(C 0 )}) ,
by using equation (4) we have:
Φ(C 0 ) = max Φ(h(C 0 )), ΦS (C 0 \ {h(C 0 )}) .
14
It is also easy to see that Φ(h(C 0 )) ≥ ΦS (h(C 0 )), therefore:
Φ(C 0 ) = max Φ(h(C 0 )), ΦS (h(C 0 )), ΦS (C 0 \ {h(C 0 )}) .
The previous equation can be reduced to Φ(C 0 ) = max {Φ(h(C 0 )), ΦS (C 0 )} .
Finally, Φ(C \ {h(C)}) = Φ(C 0 ) = max ΦS (C 0 ), Φ(h(C 0 ))
max{ω, OS (l − 1, µ0 )[1]}
0
.
= max OS (l − 1, µ )[2],
µ0
Claim 14. If 1 > µ > ω then OS (l, µ) = (∞, ∞).
Proof. It is impossible to build a clustering with this encoding. If we cut {v, p(v)}, then µ = 1,
and if we do not cut {v, p(v)}, then it is in the head cluster and then µ ≤ ω.
Claim 15. If µ = ω then:
OS (l, µ0 )[1]
0
0
OS (l, µ) = min
OS (l, µ )[1], max
, OS (l, µ )[2]
.
µ0 ≥ω
ω
Proof. Let C be a clustering encoded as OS (l, µ) = (M, b). In this case µ ≤ ω, so, {v, p(v)} is not
cut, see Figure 11b. According to Lemma 12, C is formed by adding p(v) to the head cluster of an
l-clustering C 0 of S, which is encoded as OS (l, µ0 ) for some µ0 ≥ ω. Note that max(OutS (h(C))) =
max(OutS (h(C 0 ))) and then M = OS (l, µ0 )[1]. It is easy to see that ΦS (C\{h(C)}) = ΦS (C 0 \{h(C 0 )})
and ΦS (h(C)) ≥ ΦS (h(C 0 )), so:
b = ΦS (C) = max ΦS (h(C)), ΦS (C \ {h(C)})
= max ΦS (h(C)), ΦS (h(C 0 )), ΦS (C 0 \ {h(C 0 )})
= max ΦS (h(C)), ΦS (C 0 ) .
The result follows.
Claim 16. If µ < ω, then:
OS (l, µ) = OS (l, µ).
Proof. Let C be a clustering encoded as OS (l, µ) = (M, b). In this case µ ≤ ω, and consequently
{v, p(v)} is not cut, see Figure 11b. According to Lemma 12, C is formed by adding p(v) to
the head cluster of an l-clustering C 0 of S which is encoded as OS (l, µ). It is easy to see that
max(OutS (h(C))) = max(OutS (h(C 0 ))) = M and ΦS (C) = ΦS (C 0 ) = b. The result follows.
Theorem 17. Let S = Tv such that Tv 6= T , and let S be the subtree formed by adding p(v) to S.
If we know the function OS , then the function OS can be computed in O(kn) time.
15
ω
ω
ω
ω
P
Q
S
S
P
Q
S
(a)
Q
S
Q
(b)
Figure 12: Construction of a clustering of S based in one of S. (a) The edge {v, p(v)} is cut. (b)
The edge {v, p(v)} is not cut.
Proof. Think in OS (·, ·) as table (see Remark 9), lets analyze the time to compute the values of
every cell in this table. By Claim 13, computing the values of form OS (l, 1) takes O(n) time per
cell and there are O(k) cells of this form, resulting in the total time of O(kn). By Claim 14, the
values of form OS (l, µ) with 1 > µ > ω take constant time per cell and given that there are O(kn)
cells of this form, the total time results to O(kn). By Claim 15, computing the values of the form
OS (l, ω) takes O(n) time per cell and there are O(k) cells of this form, yielding a total time of
O(kn). Finally, by Claim 16, computing the values of form OS (l, µ) with µ < ω takes constant
time per cell and there are O(kn) cells of this form, resulting in a total time of O(kn). The result
follows.
4.2
AddChildTree: computing OP from OS and OQ
Let Q = Tv 6= T and let S be a subtree rooted in p(v) such that S does not contain v. Let P denote
the subtree which results from joining on S and Q. We show how to compute OP from OS and OQ .
Let ω = w({v, p(v)}). Let Q be the subtree formed by adding p(v) to Q. Pre-compute OQ from
OQ using the claims of the previous subsection. If we are computing OP (l, µ), then we have that:
Claim 18. If µ = 1 then:
n
n
oo
OP (l, µ) = min
max {ω, OS (x, 1)[1]} , max OS (x, 1)[2], OQ (l − x + 1, 1)[2]
1≤x<l
Proof. Let C be a clustering encoded as OP (l, µ) = (M, b). We have that h(C) = {p(v)}, since
µ = 1. Then, the edge {v, p(v)} is cut, see Figure 12b. C is based in two clusterings C 0 and C 00 with
encodings OS (x, 1) and OQ (l − x + 1, 1) for some 1 ≤ x < l. It easy to see that:
M = max(OutP (h(C))) = max ω, max(OutS (h(C 0 )))
= max {ω, OS (x, 1)[1]} .
It is easy to see that:
b = ΦP (C) = max {ΦP (h(C)), ΦP (C \ {h(C)})} .
Note also that:
n
o
ΦP (h(C)) = max ΦS (h(C 0 )), ΦQ (h(C 00 )) , and
16
o
n
ΦP (C \ {h(C)}) = max ΦS (C 0 \ {h(C 0 )}), ΦQ (C 00 \ {h(C 00 )}) .
Then,
n
o
0
00
0
0
00
00
ΦP (C) = max ΦS (h(C )), ΦQ (h(C )), ΦS (C \ {h(C )}), ΦQ (C \ {h(C )}) .
And this can be rewritten as:
n
o
ΦP (C) = max ΦS (C 0 ), ΦQ (C 00 )
n
o
= max OS (x, 1)[2], OQ (l − x + 1, 1)[2] .
Claim 19. If 1 > µ > ω then:
M
M, max
OP (l, µ) = min
,
, OS (x, µ)[2], OQ (l − x + 1, 1)[2]
1≤x<l
µ
where M = max {ω, OS (x, µ)[1]}.
Proof. Let C be a clustering encoded as OP (l, µ) = (M, b). Note that {v, p(v)} is not included in
h(C) because µ > ω, and consequently this edge is cut, see Figure 12b. Furthermore, C is based in
two clusterings, C 0 and C 00 , with encodings OS (x, µ) and OQ (l − x + 1, 1) for some 1 ≤ x < l. It is
easy to see that M = max {ω, OS (x, µ)[1]}. Notice:
ΦP (h(C)) ≥ ΦS (h(C 0 )), and ΦP (h(C)) ≥ ΦQ (h(C 00 )).
Moreover,
n
o
ΦP (C \ {h(C)}) = max ΦS (C 0 \ {h(C 0 )}), ΦQ (C 00 \ {h(C 00 )}) .
Now, focus on the evaluation of C:
b = ΦP (C) = max {ΦP (h(C)), ΦP (C \ {h(C)})}
n
o
= max ΦP (h(C)), ΦS (h(C 0 )), ΦQ (h(C 0 )), ΦS (C 0 \ {h(C 0 )}), ΦQ (C 00 \ {h(C 00 )})
M
0
00
= max
, ΦS (C ), ΦS (C )
µ
M
= max
, OS (x, µ)[2], OQ (l − x + 1, 1)[2] .
µ
Claim 20. If µ = ω then:
M1
o1 = min
M1 , max
, OS (x, ω)[2], OQ (l − x + 1, 1)[2]
,
1≤x<l
ω
where M1 = max {ω, OS (x, ω)[1]} ;
M2
0
o2 =
min
M2 , max
, OS (x, µ )[2], OQ (l − x + 1, ω)[2]
,
1≤x<l, ω≤µ0
ω
n
o
where M2 = max OS (x, µ0 )[1], OQ (l − x + 1, ω)[1] ; and
OP (l, µ) = min{o1 , o2 }.
17
Proof. Let C be a clustering encoded as OP (l, µ) = (M, b). In this case, we have two options to build
C. The first one is using two clusterings, C 0 and C 00 , with encodings OS (x, ω) and OQ (l − x + 1, 1),
respectively. In this case, {v, p(v)} is cut. This case is analogous to the previous claim and
corresponds to o1 encoding. And the second one, using two clusterings C 0 and C 00 with encodings
OS (x, µ0 ) for some µ0 ≥ ω and OQ (l − x + 1, ω), respectively. In this case, {v, p(v)} is not cut. This
case corresponds to o2 encoding and we can prove it using ideas similar to the ones used in the
previous claims.
Claim 21. If µ < ω then:
M1
0
o1 =
min
M1 , max
, OS (x, µ)[2], OQ (l − x + 1, µ )[2]
,
1≤x<l, µ≤µ0
µ
n
o
where M1 = max OS (x, µ)[1], OQ (l − x + 1, µ0 )[1] ;
M2
0
o2 =
min
M2 , max
, OS (x, µ )[2], OQ (l − x + 1, µ)[2]
,
1≤x<l, µ≤µ0
µ
o
n
where M2 = max OS (x, µ0 )[1], OQ (l − x + 1, µ)[1] ;
and
OP (l, µ) = min{o1 , o2 }.
Proof. Let C be a clustering encoded as OP (l, µ) = (M, b). In this case µ < ω and then {v, p(v)} is
not cut (otherwise, the head cluster of C has an evaluation greater than 1). C is based in clusterings
C 0 and C 00 of S and Q, respectively. There are two possible ways to build C: the lightest edge into
h(C) is in h(C 0 ), o1 ; or the lightest edge into h(C) is in h(C 00 ), o2 . In both cases, the formulas can
be verified using the same ideas used in the previous claims.
Theorem 22. Let Q = Tv such that Tv 6= T , and let S be a subtree rooted at p(v) such that v is
not in S. Let P denote the subtree formed by joining S and Q. If we know the functions OS and
OQ , then the function OP can be computed in O(k 2 n2 ).
Proof. Analyzing the number of cells in each claim (18, 19, 20, 21) and the complexity to compute
the value of a cell in each case, we conclude that OP can be computed in O(k 2 n2 ).
4.3
Complexity of the algorithm
Given a tree T and a value k, we can calculate OT by computing OTv for every node v in T in a
bottom-up (from the leaves to the root) procedure using the mentioned operations. Note that, if v
is a leaf, then OTv (1, 1) = (0, 0) and OTv (l, µ) = (∞, ∞) if l > 1 or µ < 1. To compute the function
OTv of an inner node v, we proceed as follows: Let {v1 , . . . , vm } be the set of children of v. First,
considering S = Tv1 , compute OS from OS using the UpToParent operation. Subsequently, we
proceed with joining the subtrees Tvi one by one using the AddChildTree operation. When all
the children have been added, the resulting subtree corresponds to Tv . Note that we apply a single
operation per edge. Consequently, this algorithm takes O(k 2 n3 ) time. Note also that with this
algorithm, we obtain the evaluation of the optimal clustering; the clusters of an optimal solution
can be computed by “navigating backwards” through the computed functions.
18
Problem 2 can be solved using the same idea with a slightly more complex approach. We can
use a similar algorithm based on functions OS (µ), saving the parameter l, (which corresponds to
the number of clusters) and then the computing time is O n3 .
References
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: a
survey. Computer networks, 38(4):393–422, 2002.
[2] T. Asano, B. Bhattacharya, M. Keil, and F. Yao. Clustering algorithms based on minimum and
maximum spanning trees. In Proceedings of the fourth annual symposium on Computational
geometry, 1988.
[3] L. E. Caraballo, J. M. Dı́az-Báñez, I. Maza, and A. Ollero. The block-information-sharing
strategy for task allocation: A case study for structure assembly with aerial robots. European
Journal of Operational Research, 260(2):725–738, 2017.
[4] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient graph-based image segmentation. International Journal of Computer Vision, 59(2):167–181, 2004.
[5] M. Frühwirth and A. Rauber. Self-organizing maps for content-based music clustering. In
Neural Nets WIRN Vietri-01, pages 228–233. Springer, 2002.
[6] O. Grygorash, Y. Zhou, and Z. Jorgensen. Minimum spanning tree based clustering algorithms.
In Proceedings of the 18th IEEE International COnference on Tools with Artificial Intelligence
(ICTAI’06), 2006.
[7] N. Kroher, J.-M. Dı́az-Báñez, and A. Pikrakis. Discovery of repeated melodic phrases in folk
singing recordings. IEEE Transactions on Multimedia (Submitted, pending minor revision),
2017.
[8] A. Ollero and I. Maza. Multiple heterogeneous unmanned aerial vehicles. Springer Publishing
Company, Incorporated, 2007.
[9] S. E. Schaeffer. Graph clustering. Computer Science Rieveiw, 1:27–64, 2007.
[10] Y. Xu, V. Olman, and D. Xu. Minimum spanning trees for gene expression data clustering.
Genome Inform, 12:12:24, 2001.
[11] Y. Xu and E. C. Uberbacher. 2d image segmentation using minimum spanning tree. Image
and Vision Computing, 15:47–57, 1997.
[12] C. T. Zahn. Graph-theoretical methods for detecting and describing gestalt clusters. IEEE
Transactions on Computers, C-20(1):68–86, 1971.
19
| 8cs.DS
|
arXiv:1610.03625v2 [math.GR] 8 Apr 2017
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
MARC KEILBERG
Abstract. We investigate a possible connection between the F SZ properties
of a group and its Sylow subgroups. We show that the simple groups G2 (5)
and S6 (5), as well as all sporadic simple groups with order divisible by 56 are
not F SZ, and that neither are their Sylow 5-subgroups. The groups G2 (5)
and HN were previously established as non-F SZ by Peter Schauenburg; we
present alternative proofs. All other sporadic simple groups and their Sylow
subgroups are shown to be F SZ. We conclude by considering all perfect groups
available through GAP with order at most 106 , and show they are non-F SZ
if and only if their Sylow 5-subgroups are non-F SZ.
Introduction
The F SZ properties for groups, as introduced by Iovanov et al. [4], arise from
considerations of certain invariants of the representation categories of semisimple
Hopf algebras known as higher Frobenius-Schur indicators [5, 10, 11]. See [9] for
a detailed discussion of the many important uses and generalizations of these invariants. When applied to Drinfeld doubles of finite groups, these invariants are
described entirely in group theoretical terms, and are in particular invariants of
the group itself. The F SZ property is then concerned with whether or not these
invariants are always integers—which gives the Z in F SZ.
While the F SZ and non-F SZ group properties are well-behaved with respect to
direct products [4, Example 4.5], there is currently little reason to suspect a particularly strong connection to proper subgroups which are not direct factors. Indeed,
by [2, 4] the symmetric groups Sn are F SZ, while there exist non-F SZ groups of
order 56 . Therefore, Sn is F SZ but contains non-F SZ subgroups for all sufficiently
large n. On the other hand, non-F SZ groups can have every proper subquotient
be F SZ. Even the known connection to the one element centralizers—see the comment following Definition 1.1—is relatively weak. In this paper we will establish a
few simple improvements to this situation, and then proceed to establish a number
of examples of F SZ and non-F SZ groups that support a potential connection to
Sylow subgroups. We propose this connection as Conjecture 2.7.
We will make extensive use of GAP [3] and the AtlasRep[15] package. Most of
the calculations were designed to be completed with only 2GB of memory or (much)
less available—in particular, using only a 32-bit implementation of GAP—, though
in a few cases a larger workspace was necessary. In all cases the calculations can
2010 Mathematics Subject Classification. Primary: 20D08; Secondary: 20F99, 16T05, 18D10.
Key words and phrases. sporadic groups, simple groups, Monster group, Baby Monster group,
Harada-Norton group, Lyons group, projective symplectic group, higher Frobenius-Schur indicators, FSZ groups, Sylow subgroups.
This work is in part an outgrowth of an extended e-mail discussion between Geoff Mason,
Susan Montgomery, Peter Schauenburg, Miodrag Iovanov, and the author. The author thanks
everyone involved for their contributions, feedback, and encouragement.
1
2
MARC KEILBERG
be completed in workspaces with no more than 10GB of memory available. The
author ran the code on an Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz machine
with 12GB of memory. All statements about runtime are made with respect to this
computer. Most of the calculations dealing with a particular group were completed
in a matter of minutes or less, though calculations that involve checking large
numbers of groups can take several days or more across multiple processors.
The structure of the paper is as follows. We introduce the relevant notation,
definitions, and background information in Section 1. In Section 2 we present a
few simple results which offer some connections between the F SZ (or non-F SZ)
property of G and certain of its subgroups. This motivates the principle investigation of the rest of the paper: comparing the F SZ properties for certain groups and
their Sylow subgroups. In Section 3 we introduce the core functions we will need to
perform our calculations in GAP. We also show that all groups of order less than
2016 (except possibly those of order 1024) are F SZ. The remainder of the paper
will be dedicated to exhibiting a number of examples that support Conjecture 2.7.
In Section 4 we show that the simple groups G2 (5), HN , Ly, B, and M , as
well as their Sylow 5-subgroups, are all non-F SZ5 . In Section 5 we show that all
other sporadic simple groups (including the Tits group) and their Sylow subgroups
are F SZ. This is summarized in Theorem 5.4. The case of the simple projective
symplectic group S6 (5) is handled in Section 6, which establishes S6 (5) as the second
smallest non-F SZ simple group after G2 (5). It follows from the investigations
of Schauenburg [13] that HN is then the third smallest non-F SZ simple group.
S6 (5) was not susceptible to the methods of Schauenburg [13], and requires further
modifications to our own methods to complete in reasonable time. We finish our
examples in Section 7 by examining those perfect groups available through GAP,
and show that they are F SZ if and only if their Sylow subgroups are F SZ. Indeed,
they are non-F SZ if and only if their Sylow 5-subgroup is non-F SZ5 .
Of necessity, these results also establish that various centralizers and maximal
subgroups in the groups in question are also non-F SZ5 , which can be taken as
additional examples. If the reader is interested in F SZ properties for other simple
groups, we note that Schauenburg [13] has checked all simple groups of order at
most |HN | = 273,030,912,000,000 = 214 ·36 ·56 ·7·11·19, except for S6 (5) (which we
resolve here); and that several families of simple groups were established as F SZ
by Iovanov et al. [4].
We caution the reader that the constant recurrence of the number 5 and Sylow
5-subgroups of order 56 in this paper is currently more of a computationally convenient coincidence than anything else. The reasons for this will be mentioned during
the course of the paper.
1. Background and Notation
Let N be the set of positive integers. The study of F SZ groups is connected to
the following sets.
Definition 1.1. Let G be a group, u, g ∈ G, and m ∈ N. Then we define
Gm (u, g) = {a ∈ G : am = (au−1 )m = g}.
Note that Gm (u, g) = ∅ if u 6∈ CG (g), and that in all cases Gm (u, g) ⊆ CG (g).
In particular, letting H = CG (g), then when u ∈ H we have
Gm (u, g) = Hm (u, g).
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
3
The following will then serve as our definition of the F SZm property. It’s equivalence to other definitions follows easily from [4, Corollary 3.2] and applications of
the Chinese remainder theorem.
Definition 1.2. A group G is F SZm if and only if for all g ∈ G, u ∈ CG (g), and
n ∈ N coprime to the order of g, we have
|Gm (u, g)| = |Gm (u, g n )|.
We say a group is F SZ if it is F SZm for all m.
The following result is useful for reducing the investigation of the F SZ properties
to the level of conjugacy classes or even rational classes.
Lemma 1.3. For any group G and u, g, x ∈ G we have a bijection Gm (u, g) →
Gm (ux , g x ) given by a 7→ ax .
If n ∈ N is coprime to |G| and r ∈ N is such that rn ≡ 1 mod |G|, we also have
a bijection Gm (u, g n ) → Gm (ur , g).
Proof. The first part is [5, Proposition 7.2] in slightly different notation. The second
part is [14, Corollary 5.5].
All expressions of the form Gm (u, g n ) will implicitly assume that n is coprime to
the order of g. We are free to replace n with an equivalent value which is coprime
to |G| whenever necessary. Moreover, when computing cardinalities |Gm (u, g)| it
suffices to compute the cardinalities |Hm (u, g)| for H = CG (g), instead. This latter
fact is very useful when attempting to work with groups of large order, or groups
with centralizers that are easy to compute in, especially when the group is suspected
of being non-F SZ.
+
, the union of which yields
Remark 1.4. There are stronger conditions called F SZm
+
the F SZ + condition, which are also introduced by Iovanov et al. [4]. The F SZm
condition is equivalent to the centralizer of every non-identity element with order
not in {1, 2, 3, 4, 6} being F SZm , which is in turn equivalent to the sets Gm (u, g) and
Gm (u, g n ) being isomorphic permutation modules for the two element centralizer
CG (u, g) [4, Theorem 3.8], with u, g, n satisfying the same constraints as for the
F SZm property. Here the action is by conjugation. We note that while the F SZ
property is concerned with certain invariants being in Z, the F SZ + property is not
concerned with these invariants being non-negative integers. When the invariants
are guaranteed to be non-negative is another area of research, and will also not be
considered here.
Example 1.5. The author has shown that quaternion groups and certain semidirect products defined from cyclic groups are always F SZ [7, 8]. This includes
the dihedral groups, semidihedral groups, and quasidihedral groups, among many
others.
Example 1.6. Iovanov et al. [4] showed that several groups and families of groups
are F SZ, including:
• All regular p-groups.
• Zp ≀r Zp , the Sylow p-subgroup of Sp2 , which is an irregular F SZ p-group.
• P SL2 (q) for a prime power q.
4
MARC KEILBERG
• Any direct product of F SZ groups. Indeed, any direct product of F SZm
groups is also F SZm , as the cardinalities of the sets in Definition 1.1 split
over the direct product in an obvious fashion.
• The Mathieu groups M11 and M12 .
• Symmetric and alternating groups. See also [2].
Because of the first item, Susan Montgomery has proposed that we use the term
F S-regular instead of F SZ, and F S-irregular for non-F SZ. Similarly for F Sm regular and F Sm -irregular. These seem reasonable choices, but for this paper the
author will stick with the existing terminology.
Example 1.7. On the other hand, Iovanov et al. [4] also established that non-F SZ
groups exist by using GAP [3] to show that there are exactly 32 isomorphism classes
of groups of order 56 which are not F SZ5 .
Example 1.8. The author has constructed examples of non-F SZpj p-groups for
all primes p > 3 and j ∈ N in [6]. For j = 1 these groups have order pp+1 , which is
the minimum possible order for any non-F SZ p-group. Combined, [4, 6, 13] show,
among other things, that the minimum order of non-F SZ 2-groups is at least 210 ,
and the minimum order for non-F SZ 3-groups is at least 38 . It is unknown if any
non-F SZ 2-groups or 3-groups exist, however.
Example 1.9. Schauenburg [13] provides several equivalent formulations of the
F SZm properties, and uses them to construct GAP [3] functions which are useful
for testing the property. Using these functions, it was shown that the Chevalley
group G2 (5) and the sporadic simple group HN are not F SZ5 . These groups
were attacked directly, using advanced computing resources for HN , often with
an eye on computing the values of the indicators explicitly. We will later present
an alternative way of using GAP to prove that these groups, and their Sylow 5subgroups, are not F SZ5 . We will not attempt to compute the actual values of the
indicators, however.
One consequence of these examples is that the smallest known order for a nonF SZ group is 56 = 15,625. The groups with order divisible by pp+1 for p > 5
that are readily available through GAP are small in number, problematically large,
and frequently do not have convenient representations. Matrix groups have so far
proven too memory intensive for what we need to do, so we need permutation
or polycyclic presentations for accessible calculations. For these reasons, all of
the examples we pursue in the following sections will hone in on the non-F SZ5
property for groups with order divisible by 56 , and which admit known or reasonably
computable permutation representations. In most of the examples, 56 is the largest
power of 5 dividing the order, with the Monster group, the projective symplectic
group S6 (5), and the perfect groups of order 12 · 57 being the exceptions.
2. Obtaining the non-F SZ property from certain subgroups
Our first elementary result offers a starting point for investigating non-F SZm
groups of minimal order.
Lemma 2.1. Let G be a group with minimal order in the class of non-F SZm
groups. Then |Gm (u, g)| 6= |Gm (u, g n )| for some (n, |G|) = 1 implies g ∈ Z(G).
Proof. If not then CG (g) is a smaller non-F SZm group, a contradiction.
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
5
The result applies to non-F SZm groups in a class that is suitably closed under
the taking of centralizers. For example, we have the following version for p-groups.
Corollary 2.2. Let P be a p-group with minimal order in the class of non-F SZpj
p-groups. Then |Ppj (u, g)| 6= |Ppj (u, g n )| for some p ∤ n implies g ∈ Z(P ).
Example 2.3. From the examples in the previous section, we know the minimum
possible order for a non-F SZp p-group for p > 3 is pp+1 . It remains unknown if
the examples of non-F SZpj p-groups from [6] for j > 1 have minimal order among
non-F SZpj p-groups. We also know that to check if a group of order 210 or 38 is
F SZ it suffices to assume that g is central.
Next, we determine a condition for when the non-F SZ property for a normal
subgroup implies the non-F SZ property for the full group.
Lemma 2.4. Let G be a group and suppose H is a non-F SZm normal subgroup
with m coprime to [G : H]. Then G is non-F SZm .
Proof. Let u, g ∈ H and (n, |g|) = 1 be such that |Hm (u, g)| 6= |Hm (u, g n )|. By
the index assumption, for all x ∈ G we have xm ∈ H ⇔ x ∈ H, so by definitions
Gm (u, g) = Hm (u, g) and Gm (u, g n ) = Hm (u, g n ), which gives the desired result.
Corollary 2.5. Let G be a finite group and suppose P is a normal non-F SZpj
Sylow p-subgroup of G for some prime p. Then G is non-F SZpj .
Corollary 2.6. Let G be a finite group and P a non-F SZpj Sylow p-subgroup of
G. Then the normalizer NG (P ) is non-F SZpj .
Sadly, we will find no actual use for Corollary 2.5 in the examples we consider
in this paper. However, this result, [13, Lemma 8.7], and the examples we collect
in the remainder of this paper do suggest the following conjectural relation for the
F SZ property.
Conjecture 2.7. A group is F SZ if and only if all of its Sylow subgroups are F SZ.
Some remarks on why this conjecture may involve some deep results to establish
affirmatively seems in order.
Consider a group G and let u, g ∈ G and n ∈ N with (n, |G|) = 1. Suppose that
g has order a power of p, for some prime p. Then
[
Gpj (u, g) =
Gpj (u, g) ∩ P x ,
where the union runs over all distinct conjugates P x in CG (g) of a fixed Sylow
p-subgroup P of CG (g). Let Ppxj (u, g) = Gpj (u, g) ∩ P x . Then |Gpj (u, g)| =
S
S
|Gpj (u, g n )| if and only if there is a bijection P x (u, g) → P x (u, g n ). In the
special case u ∈ P , if P was F SZpj we would have a bijection P (u, g) → P (u, g n ),
x
x
but this does not obviously guarantee a bijection
g n ) for all conS x P (u, g)
S →xP (u,
n
jugates. Attempting to get a bijection P (u, g) → P (u, g ) amounts, via
the Inclusion-Exclusion Principle, to controlling the intersections of any number of
conjugates and how many elements those intersections contribute to Gpj (u, g) and
Gpj (u, g n ). There is no easy or known way to predict the intersections of a collection of Sylow p-subgroups for a completely arbitrary G, so any positive affirmation
of the conjecture will impose a certain constraint on these intersections.
6
MARC KEILBERG
Moreover, we have not considered the case of the sets Gm (u, g) where m has
more than one prime divisor, nor those cases where u, g do not order a power of
a fixed prime, so a positive affirmation of the conjecture is also expected to show
that the F SZm properties are all derived from the F SZpj properties for all prime
powers dividing m. On the other hand, a counterexample seems likely to involve
constructing a large group which exhibits a complex pattern of intersections in its
Sylow p-subgroups for some prime p, or otherwise exhibits the first example of a
group which is F SZpj for all prime powers but is nevertheless not F SZ.
Example 2.8. All currently known non-F SZ groups are either p-groups (for which
the conjecture is trivial), are nilpotent (so are just direct products of their Sylow
subgroups), or come from perfect groups (though the relevant centralizers need not
be perfect). The examples of both F SZ and non-F SZ groups we establish here
will also all come from perfect groups and p-groups. In the process we obtain,
via the centralizers and maximal subgroups considered, an example of a solvable,
non-nilpotent, non-F SZ group; as well as an example of a non-F SZ group which
is neither perfect nor solvable. All of these examples, of course, conform to the
conjecture.
3. GAP functions and groups of small order
The current gold standard for general purpose testing of the F SZ properties in
GAP [3] is the FSZtest function of Schauenburg [13]. In certain specific situations,
the function FSInd from [4] can also be useful for showing a group is non-F SZ.
However, with most of the groups we will consider in this paper both of these
functions are impractical to apply directly. The principle obstruction for FSZtest
is that this function needs to compute both conjugacy classes and character tables
of centralizers, and this can be a memory intensive if not wholly inaccessible task.
For FSInd the primary obstruction, beyond its specialized usage case, is that it
must completely enumerate, store, and sort the entire group (or centralizer). This,
too, can quickly run into issues with memory consumption.
We therefore need alternatives for testing (the failure of) the FSZ properties
which can sidestep such memory consumption issues. For Section 7 we will also
desire functions which can help us detect and eliminate the more ”obviously” FSZ
groups. We will further need to make various alterations to FSZtest to incorporate
these things, and to return a more useful value when the group is not F SZ.
The first function we need, FSZtestZ, is identical to FSZtest—and uses several of
the helper functions found in [13]—except that instead of calculating and iterating
over all rational classes of the group it iterates only over those of the center. It
needs only a single input, which is the group to be checked. If it finds that the
group is non-F SZ, rather than return false it returns the data that established
the non-F SZ property. Of particular importance are the values m and z. If the
group is not shown to be non-F SZ by this test, then it returns fail to indicate
that the test is typically inconclusive.
FSZtestZ := function (G)
l o c a l CT, zz , z , c l , div , d , c h i , m, b ;
c l := R a t i o n a l C l a s s e s ( Center (G) ) ;
c l := F i l t e r e d ( c l , c−>not Order ( R e p r e s e n t a t i v e ( c ) )
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
7
in [ 1 , 2 , 3 , 4 , 6 ] ) ;
for z z in c l do
z := R e p r e s e n t a t i v e ( z z ) ;
d i v := F i l t e r e d ( D i v i s o r s I n t ( Exponent (G) / Order ( z ) ) ,
m−>not Gcd (m, Order ( z ) ) in [ 1 , 2 , 3 , 4 , 6 ] ) ;
i f Length ( d i v ) < 1 then continue ; f i ;
CT := O r d i n a r y C h a r a c t er T a b le (G) ;
for c h i in I r r (CT) do
for m in d i v do
i f not I s R a t ( beta (CT, z , m, c h i ) )
then return [ z ,m, c h i ,CT ] ;
fi ;
od ;
od ;
od ;
#t h e t e s t i s i n c o n c l u s i v e i n g e n e r a l
return f a i l ;
end ;
This function is primarily useful for testing groups with minimal order in a class
closed under centralizers, such as in Lemma 2.1 and Corollary 2.2. Or for any group
with non-trivial center that is suspected of failing the F SZ property at a central
value.
We next desire a function which can quickly eliminate certain types of groups as
automatically being F SZ. For this, the following result on groups of small order is
helpful.
Theorem 3.1. Let G be a group with |G| < 2016 and |G| 6= 1024. Then G is F SZ.
Proof. By Lemma 2.1 it suffices to run FSZtestZ over all groups in the SmallGroups
library of GAP. This library includes all groups with |G| < 2016, except those of
order 210 = 1024. In practice, the author also used the function IMMtests introduced below, but where the check on the size of the group is constrained initially
to 100 by [4, Corollary 5.5], and can be increased whenever desired to eliminate all
groups of orders already completely tested. This boils down to quickly eliminating
p-groups and groups with relatively small exponent. By using the closure of the
F SZ properties with respect to direct products, one need only consider a certain
subset of the orders in question rather than every single one in turn, so as to avoid
essentially double-checking groups. We note that the groups of order 1536 take the
longest to check. The entire process takes several days over multiple processors,
but is otherwise straightforward.
8
MARC KEILBERG
We now define the function IMMtests. This function implements most of the
more easily checked conditions found in [4] that guarantee the F SZ property, and
calls FSZtestZ when it encounters a suitable p-group. The function returns true
if the test conclusively establishes that the group is F SZ; the return value of
FSZtestZ if it conclusively determines the group is non-F SZ; and fail otherwise.
Note that whenever this function calls FSZtestZ that test is conclusive by Corollary 2.2, so it must adjust a return value of fail to true.
IMMtests := function (G)
l o c a l sz , b , l , p2 , p3 , po ;
i f I s A b e l i a n (G)
then return true ;
fi ;
s z := S i z e (G) ;
i f ( s z < 2 0 1 6 ) and ( not s z =1024)
then return true ;
fi ;
i f IsPGroup (G) then
#R e g u l a r p−g r o u p s are a l w a y s FSZ .
l := C o l l e c t e d ( F a c t o r s I n t ( s z ) ) [ 1 ] ;
i f l [1] >= l [ 2 ] or Exponent (G) = l [ 1 ]
then return true ;
fi ;
s z := Length ( U p p e r C e n t r a l S e r i e s (G) ) ;
i f l [ 1 ] = 2 then
i f l [2 ] < 1 0 or s z < 3
or Exponent (G)<64
then return true ;
e l i f l [ 2 ] = 1 0 and s z >= 3
then
b := FSZtestZ (G) ;
i f I s L i s t ( b ) then return b ;
e l s e return true ;
fi ;
fi ;
e l i f l [ 1 ] = 3 then
i f l [2 ] < 8 or s z < 4
or Exponent (G)<27
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
9
then return true ;
e l i f l [ 2 ] = 8 and sz >=4
then
b := FSZtestZ (G) ;
i f I s L i s t ( b ) then return b ;
e l s e return true ;
fi ;
fi ;
e l i f sz < l [1]+1
then return true ;
e l i f s z = l [ 1 ] + 1 and s z=l [ 2 ]
then
b := FSZtestZ (G) ;
i f I s L i s t ( b ) then return b ;
e l s e return true ;
fi ;
fi ;
else
#c h e c k t h e e x p o n e n t f o r non−p−g r o u p s
l := F a c t o r s I n t ( Exponent (G) ) ;
p2 := Length ( P o s i t i o n s ( l , 2 ) ) ;
p3 := Length ( P o s i t i o n s ( l , 3 ) ) ;
po := F i l t e r e d ( l , x−>x > 3 );
i f F o r A l l ( C o l l e c t e d ( po ) , x−>x [ 2 ] < 2 ) and
( ( p2 < 4 and p3 < 4 )
or ( p2 < 6 and p3 < 2 ) )
then return true ;
fi ;
fi ;
#t e s t s were i n c o n c l u s i v e
return f a i l ;
end ;
We then incorporate these changes into a modified version of FSZtest, which we
give the same name. Note that this function also uses the function beta and its
corresponding helper functions from [13]. It has the same inputs and outputs as
FSZtestZ, except that the test is definitive, and so returns true when the group is
F SZ.
FSZtest := function (G)
l o c a l C, CT, zz , z , c l , div , d , c h i , m, b ;
b := IMMtests (G ) ; ;
10
MARC KEILBERG
i f not b= f a i l
then return b ;
fi ;
c l := R a t i o n a l C l a s s e s (G) ;
c l := F i l t e r e d ( c l , c−>not Order ( R e p r e s e n t a t i v e ( c ) )
in [ 1 , 2 , 3 , 4 , 6 ] ) ;
for z z in c l do
z := R e p r e s e n t a t i v e ( z z ) ;
C := C e n t r a l i z e r (G, z ) ;
d i v := F i l t e r e d ( D i v i s o r s I n t ( Exponent (C) / Order ( z ) ) ,
m−>not Gcd (m, Order ( z ) ) in [ 1 , 2 , 3 , 4 , 6 ] ) ;
i f Length ( d i v ) < 1 then continue ; f i ;
# Check f o r t h e e a s y c a s e s
b := IMMtests (C ) ;
i f b=true
then continue ;
e l i f I s L i s t ( b ) then
i f R a t i o n a l C l a s s (C, z)= R a t i o n a l C l a s s (C, b [ 2 ] )
then return b ;
fi ;
fi ;
CT := O r d i n a r y C h a r a c t er T a b le (C) ;
for c h i in I r r (CT) do
for m in d i v do
i f not I s R a t ( beta (CT, z , m, c h i ) )
then return [m, z , c h i ,CT ] ;
fi ;
od ;
od ;
od ;
return true ;
end ;
Our typical procedure will be as follows: given a group G, take its Sylow 5subgroup P and find u, g ∈ P such that |P5 (u, g)| 6= |P5 (u, g n )| for 5 ∤ n, and
then show that |G5 (u, g)| 6= |G5 (u, g n )|. The second entry in the list returned by
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
11
FSZtest gives precisely the g value we need. But it does not provide the u value
directly, nor the n. As it turns out, we can always take n = 2 when o(g) = 5, but
for other orders this need not necessarily hold.
In order to acquire these values we introduce the function FSIndPt below, which
is a variation on FSInd [4]. This function has the same essential limitation that
FSInd does, in that it needs to completely enumerate, store, and sort the elements
of the group. This could in principle be avoided, at the cost of increased run-time.
However our main use for the function is to apply it to Sylow 5-subgroups which
have small enough order that this issue does not pop up.
The inputs are a group G, m ∈ N and g ∈ G. It is best if one in fact passes in
CG (g) for G, but the function will compute the centralizer regardless. The function
looks for an element u ∈ CG (g) and an integer j coprime to the order of g such that
|Gm (u, g)| 6= |Gm (u, g j )|. The output is the two element list [u,j] if such data
exists, otherwise it returns fail to indicate that the test is normally inconclusive.
Note that by Lemma 1.3 and centrality of g in C = CG (g) we need only consider
the rational classes in C to find such a u.
FSIndPt:= function (G,m, g )
l o c a l GG, C, Cl , g u c o e f f , elG , Gm, a l i s t ,
a u l i s t , u m l i s t , npos , j , n , u , pr ;
C := C e n t r a l i z e r (G, g ) ;
GG := EnumeratorSorted (C ) ; ;
elG := S i z e (C ) ;
Gm := L i s t (GG, x−>P o s i t i o n (GG, xˆm) ) ;
pr := P r imeResidues ( Order ( g ) ) ;
for Cl in R a t i o n a l C l a s s e s (C) do
u := R e p r e s e n t a t i v e ( Cl ) ;
npos := [ ] ;
a l i s t := [ ] ;
a u l i s t := [ ] ;
u m l i s t := [ ] ;
g u c o e f f := [ ] ;
u m l i s t := L i s t (GG, a−>P o s i t i o n (GG,
( a∗ I n v e r s e ( u ) ) ˆm) ) ; ;
#The f o l l o w i n g computes t h e c a r d i n a l i t i e s
# o f G m( u , g ˆn ) .
for n in pr
12
MARC KEILBERG
do
npos := P o s i t i o n (GG, g ˆn ) ;
a l i s t := P o s i t i o n s (Gm, npos ) ;
a u l i s t := P o s i t i o n s ( u m l i s t , npos ) ;
g u c o e f f [ n ] := S i z e (
Intersection ( alist , aulist ));
#Check i f we ’ ve fou n d our u
i f not g u c o e f f [ n ] = g u c o e f f [ 1 ]
then return [ u , n ] ;
fi ;
od ;
od ;
#No u was fou n d f o r t h i s G,m, g
return f a i l ;
end ;
Lastly, we introduce the function FSZSetCards, which is the most naive and
straightforward way of computing both |Gm (u, g)| and |Gm (u, g n )|. The inputs are
a set C of group elements—normally this would be CG (g), but could be a conjugacy
class or some other subset or subgroup—; group elements u, g; and integers m, n
such that g 6= g n . The output is a two element list, which counts the number of
elements of C in Gm (u, g) in the first entry and the number of elements of C in
Gm (u, g n ) in the second entry. It is left to the user to check that the inputs satisfy
whatever relations are needed, and to then properly interpret the output.
FSZSetCards := function (C, u , g ,m, n )
l o c a l c o n t r i b s , apow , aupow , a ;
c o n t r i b s := [ 0 , 0 ] ;
for a in C do
apow := a ˆm;
aupow := ( a ∗ I n v e r s e ( u ) ) ˆm;
i f ( apow = g and aupow = g ) then
c o n t r i b s [ 1 ] := c o n t r i b s [ 1 ] + 1 ;
e l i f ( apow=g ˆn and aupow=gˆn ) then
c o n t r i b s [ 2 ] := c o n t r i b s [ 2 ] + 1 ;
fi ;
od ;
return ( c o n t r i b s ) ;
end ;
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
13
As long as C admits a reasonable iterator in GAP then this function can compute
these cardinalities with a very minimal consumption of memory. Any polycyclic or
permutation group satisfies this, as well as any conjugacy class therein. However,
for a matrix group GAP will attempt to convert to a permutation representation,
which is usually very costly.
The trade-off, as it often is, is in the speed of execution. For permutation groups
the run-time can be heavily impacted by the degree, such that it is almost always
worthwhile to apply SmallerDegreePermutationRepresentation whenever possible. If the reader wishes to use this function on some group that hasn’t been tested
before, the author would advise adding in code that would give you some ability
to gauge how far along the function is. By default there is nothing in the above
code, even if you interrupt the execution to check the local variables, to tell you if
the calculation is close to completion. Due to a variety of technical matters it is
difficult to precisely benchmark the function, but when checking a large group it
is advisable to acquire at least some sense of whether the calculation may require
substantial amounts of time.
Remark 3.2. Should the reader opt to run our code to see the results for themselves,
they may occasionally find that the outputs of FSZSetCards occur in the opposite
order we list here. This is due to certain isomorphisms and presentations for groups
calculated in GAP not always being guaranteed to be identical every single time you
run the code. As a result, the values for u or g may sometimes be a coprime power
(often the inverse) of what they are in other executions of the code. Nevertheless,
there are no issues with the function proving the non-F SZ property thanks to
Lemma 1.3, and there is sufficient predictability to make the order of the output
the only variation.
While very naive, FSZSetCards will suffice for most of our purposes, with all
uses of it completing in an hour or less. However, in Section 6 we will find an
example where the expected run-time for this function is measured in weeks, and
for which FSZtest requires immense amounts of memory—Schauenburg [13] says
that FSZtest for this group consumed 128 GB of memory without completing!
We therefore need a slightly less naive approach to achieve a more palatable runtime in this case. We leave this to Section 6, but note to the reader that the method
this section uses can also be applied to all of the other groups for which FSZSetCards
suffices. The reason we bother to introduce and use FSZSetCards is that the method
of Section 6 relies on being able to compute conjugacy classes, which can hit memory
consumption issues that FSZSetCards will not encounter. It is not our goal with
these functions to find the most efficient, general-purpose procedure. Instead we
seek to highlight some of the ways in which computationally problematic groups
may be rendered tractable by altering the approach one takes, and to show that the
non-F SZ property of these groups can be demonstrated in a (perhaps surprisingly)
short amount of time and with very little memory consumption.
4. The non-F SZ sporadic simple groups
The goal for this section is to show that the Chevalley group G2 (5), and all
sporadic simple groups with order divisible by 56 , as well as their Sylow 5-subgroups,
are non-F SZ5 . We begin with a discussion of the general idea for the approach.
Our first point of observation is that the only primes p such that pp+1 divides the
order of any of these groups have p ≤ 5. Indeed, a careful analysis of the non-F SZ
14
MARC KEILBERG
groups of order 56 found in [4] shows that several of them are non-split extensions
with a normal extra-special group of order 55 , which can be denoted in AtlasRep
notation as 51+4 .5. Consulting the known maximal subgroups for these groups we
can easily infer that the Sylow 5-subgroups of HN , G2 (5), B, and Ly have this same
form, and that the Monster has such a p-subgroup. Indeed, G2 (5) is a maximal
subgroup of HN , and B and Ly have maximal subgroups containing a copy of
HN , so these Sylow subgroups are all isomorphic. Furthermore, the Monster’s
Sylow 5-subgroup has the form 51+6 .52 , a non-split extension of the elementary
abelian group of order 25 by an extra special group of order 57 . Given this, we
suspect that these Sylow 5-subgroups are all non-F SZ5 , and that this will cause
the groups themselves to be non-F SZ5 .
We can then exploit the fact that non-trivial p-groups all have non-trivial centers
to obtain centralizers in the parent group that contain a Sylow 5-subgroup. In the
case of G = HN or G = G2 (5), we can quickly find u, g ∈ P , with P a Sylow
5-subgroup of G, such that |P5 (u, g)| 6= |P5 (u, g 2 )|, and show that for H = CG (g)
we have |H5 (u, g)| 6= |H5 (u, g 2 )|. Since necessarily |H5 (u, g)| = |G5 (u, g)| and
|H5 (u, g 2 )| = |G5 (u, g 2 )|, this will show that HN and G2 (5) are non-F SZ5 . Unfortunately, it turns out that P is not normal in H in either case, so the cardinalities of
these sets in H must be checked directly, rather than simply applying Corollary 2.5.
The remaining groups require a little more work, for various reasons.
In the case of the Monster, there is a unique non-identity conjugacy class yielding
a centralizer with order divisible by 59 . So we are free to pick any subgroup G
of M that contains a centralizer with this same order. Fortunately, not only is
such a (maximal) subgroup known, but Bray and Wilson [1] have also computed
a permutation representation for it. This is available in GAP via the AtlasRep
package. This makes all necessary calculations for the Monster accessible. The
Sylow 5-subgroup is fairly easily shown to be non-F SZ5 directly. However, the
centralizer we get in this way has large order, and its Sylow 5-subgroup is not
normal, making it impractical to work with on a personal computer. However,
further consultation of character tables shows that the Monster group has a unique
conjugacy class of an element of order 10 whose centralizer is divisible by 56 . So we
may again pick any convenient (maximal) subgroup with such a centralizer, and it
turns out the same maximal subgroup works. We construct the appropriate element
of order 10 by using suitable elements from Sylow subgroups of the larger centralizer,
and similarly to get the element u. Again it turns out that the Sylow 5-subgroup of
this smaller subgroup is not normal, so we must compute the set cardinalities over
the entire centralizer in question. However, this centralizer is about 1/8000-th the
size of the initial one, and we are subsequently able to calculate the appropriate
cardinalities in under an hour.
The Baby Monster can then be handled by using the fact that the Monster
contains the double cover of B as the centralizer of an involution to obtain the centralizer we need in B from a centralizer in M . The author thanks Robert Wilson
for reminding them of this fact. For the Lyons group, the idea is much the same
as for HN and G2 (5), with the additional complication that the AtlasRep package
does not currently contain any permutation representations for Ly. To resolve this,
we obtain a permutation representation for Ly, either computed directly in GAP
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
15
or downloaded [12]. This is then used to construct a suitable permutation representation of the maximal subgroup in question. Once this is done the calculations
proceed without difficulties.
These calculations all make extensive use of the functions given in Section 3.
4.1. Chevalley group G2 (5). We now show that G2 (5) and its Sylow 5-subgroups
are not F SZ5 . This was independently verified in [13]. Since G2 (5) is of relatively
small order, it can be attacked quickly and easily.
Theorem 4.1. The simple Chevalley group G2 (5) and its Sylow 5-subgroup are
non-F SZ5 .
Proof. The claims follow from running the following GAP code.
G := AtlasGroup ( ”G2 ( 5 ) ” ) ; ;
P := SylowSubgroup (G, 5 ) ; ;
# The f o l l o w i n g shows P i s n ot FSZ 5
g := FSZtestZ (P ) [ 2 ] ;
# Find u
u := FSIndPt (P , 5 , g ) [ 1 ] ; ;
C := C e n t r a l i z e r (G, g ) ; ;
#Check t h e c a r d i n a l i t i e s
FSZSetCards (C, u , g , 5 , 2 ) ;
The output is [0,625], so it follows that G and P are both non-F SZ5 as desired.
We note that P is not normal in C, and indeed C is a perfect group of order
375,000 = 23 · 3 · 56 .
The call to FSZSetCards above runs in approximately 11 seconds, which is approximately the amount of time necessary to run FSZtest on G2 (5) directly. In
this case, the use of FSZSetCards is not particularly efficient, as the groups in
question are of reasonably small sizes and permutation degree. Nevertheless, this
demonstrates the basic method we will employ for all subsequent groups.
4.2. The Harada-Norton group. For the group HN the idea proceeds similarly
as for G2 (5).
Theorem 4.2. The Harada-Norton simple group HN and its Sylow 5-subgroup are
not F SZ5 .
Proof. To establish the claims it suffices to run the following GAP code.
G := AtlasGroup ( ”HN” ) ; ;
P := SylowSubgroup (G, 5 ) ; ;
# G, t h u s P, has v e r y l a r g e d e g r e e .
# P o l y c y c l i c g r o u p s are e a s i e r t o work w i t h .
i s o P := IsomorphismPcGroup (P ) ; ;
P := Image ( i s o P ) ; ;
16
MARC KEILBERG
#Find u , g f o r P
g := FSZtestZ (P ) [ 2 ] ;
u := FSIndPt (P , 5 , g ) [ 1 ] ;
g := Image ( Inver seGener a lMa pping ( i s o P ) , g ) ; ;
u := Image ( Inver seGener a lMa pping ( i s o P ) , u ) ; ;
C := C e n t r a l i z e r (G, g ) ; ;
iso C := IsomorphismPcGroup (C ) ; ;
C := Image ( iso C ) ; ;
FSZSetCards (C, Image ( isoC , u ) , Image ( isoC , g ) , 5 , 2 ) ;
This code executes in approximately 42 minutes, with approximately 40 of that
spent finding P . The final output is [3125,0], so we conclude that both P and HN
are non-F SZ5 , as desired.
P is again not a normal subgroup of C, so we again must test the entire centralizer
rather than just P . We note that |C| = 25 56 = 500,000. Indeed, C is itself nonF SZ5 of necessity, and the fact that the call to IsomorphismPcGroup did not fail
means that C is solvable, and in particular not perfect and not a p-group.
4.3. The Monster group. We will now consider the Monster group M . The full
Monster group is famously difficult to compute in. But, as detailed in the beginning of the section, by consulting character tables of M and its known maximal
subgroups, we can find a maximal subgroup which contains a suitable centralizer
(indeed, two suitable centralizers) and also admits a known permutation representation [1].
Theorem 4.3. The Monster group M and its Sylow 5-subgroup are not F SZ5 .
Proof. The Sylow 5-subgroup of M has order 59 . Consulting the character table of
M , we see that M has a unique conjugacy class yielding a proper centralizer with
order divisible by 59 , and a unique conjugacy class of an element of order 10 whose
centralizer has order divisible by 56 ; moreover, the order of the latter centralizer is
precisely 12 million, and in particular is not divisible by 57 . It suffices to consider
any maximal subgroups containing such centralizers. The maximal subgroup of
shape 51+6
: 2.J2 .4, which is the normalizer associated to a 5B class, is one such
+
choice.
We first show that the Sylow 5-subgroup of M is not F SZ5 .
G :=
P :=
isoP
P :=
AtlasGroup ( ” 5 ˆ ( 1 + 6 ) : 2 . J2 . 4 ” ) ; ;
SylowSubgroup (G, 5 ) ; ;
:= IsomorphismPcGroup (P ) ; ;
Image ( i s o P ) ; ;
ex := FSZtestZ (P ) ;
The proper centralizer with order divisible by 59 is still impractical to work with.
So we will use the data for P to construct the element of order 10 mentioned above.
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
17
zp := ex [ 2 ] ; ;
zp := Image ( Inver seGener a lMa pping ( i s o P ) , zp ) ; ;
C := C e n t r a l i z e r (G, zp ) ; ;
Q := SylowSubgroup (C , 2 ) ; ;
zq := F i r s t ( Center (Q) , q−>Order ( q)>1 and
S i z e ( C e n t r a l i z e r (G, zp ∗q ) ) = 1 2 0 0 0 0 0 0 ) ; ;
#This g i v e s us t h e g and c e n t r a l i z e r we want .
g := zp ∗ zq ; ;
C := C e n t r a l i z e r (G, g ) ; ;
#Reducing t h e p e r m u t a t i o n d e g r e e w i l l
#s a v e a l o t o f compu t at ion t ime l a t e r .
iso C := S m a l l e r D e g r e e P e r m u t a t i o n R e p r e s e n t a t i o n (C ) ; ;
C := Image ( iso C ) ; ;
g := Image ( isoC , g ) ; ;
zp := Image ( isoC , zp ) ; ;
zq := Image ( isoC , zq ) ; ;
#Now
P :=
isoP
P :=
proceed to con st ru ct a choice of u .
SylowSubgroup (C , 5 ) ; ;
:= IsomorphismPcGroup (P ) ; ;
Image ( i s o P ) ; ;
ex := FSIndPt (P, 5 , Image ( iso P , zp ) ) ;
up := Image ( Inver seGener a lMa pping ( i s o P ) , ex [ 1 ] ) ; ;
#D e f i n e our c h o i c e o f u .
#In t h i s case , u has o r d e r 5 0 .
u := up∗ zq ; ;
#F i n a l l y , we compute t h e c a r d i n a l i t i e s
# of the relevant s e t s .
FSZSetCards (C, u , g , 5 , 7 ) ;
This final function yields [0,15000], which proves that M is not F SZ5 , as desired.
This final function call takes approximately 53 minutes to complete, while all
preceding operations can complete in about 5 minutes combined—though the conversion of C to a lower degree may take more than this, depending. The lower
degree C has degree 18, 125, but requires (slightly) more than 2 GB of memory to
acquire. This conversion can be skipped to keep the memory demands well under
2GB, but the execution time for FSZSetCards will inflate to approximately a day
and a half.
18
MARC KEILBERG
Remark 4.4. In the first definition of C above, containing the full Sylow 5-subgroup
of M , we have |C| = 9.45 × 1010 = 28 · 33 · 59 · 7. For the second definition of C,
corresponding to the centralizer of an element of order 10, we have |C| = 1.2×107 =
28 · 3 · 56 . The first centralizer is thus 7875 = 32 · 53 · 7 times larger than the second
one. Either one is many orders of magnitude smaller than |M | ≈ 8.1 × 1053 , but
the larger one was still too large to work with for practical purposes.
4.4. The Baby Monster. We can now consider the Baby Monster B.
Theorem 4.5. The Baby Monster B and its Sylow 5-subgroup are both non-F SZ5 .
Proof. The Baby Monster is well known to have a maximal subgroup of the form
HN.2, so it follows that B and HN have isomorphic Sylow 5-subgroups. By Theorem 4.2 HN has a non-F SZ5 Sylow 5-subgroup, so this immediately gives the
claim about the Sylow 5-subgroup of B.
From the character table of B we see that there is a unique non-identity conjugacy class whose centralizer has order divisible by 56 . This corresponds to an element of order 5 from the 5B class, and the centralizer has order 6,000,000 = 27 ·3·56 .
In the double cover 2.B of B, this centralizer is covered by the centralizer of an
element of order 10. This centralizer necessarily has order 12, 000, 000. Since M
contains 2.B as a maximal subgroup, and there is a unique centralizer of an element
of order 10 in M with order divisible by 12,000,000, these centralizers in 2.B and M
are isomorphic. We have already computed this centralizer in M in Theorem 4.3.
To obtain the centralizer in B, we need only quotient by an appropriate central
involution. In the notation of the proof of Theorem 4.3, this involution is precisely
zq.
GAP will automatically convert this quotient group D into a lower degree representation, yielding a permutation representation of degree 3125 for the centralizer.
This will require as much as 8GB of memory to complete. Moreover, the image
of zp from Theorem 4.3 in this quotient group yields the representative of the 5B
class we desire, denoted here by g. Using the image of up in the quotient for u,
we can then easily run FSZSetCards(C,u,g,5,2) to get a result of [15000,3125],
which shows that B is non-F SZ5 as desired. This final call completes in about 4
minutes.
Note that in M the final return values summed to 15,000, with one of the values
0, whereas in B they sum to 18,125 and neither is zero. This reflects how there
is no clear relationship between the F SZ properties of a group and its quotients,
even when the quotient is by a (cyclic) central subgroup. In particular, it does not
immediately follow that the quotient centralizer would yield the non-F SZ property
simply because the centralizer in M did, or vice versa.
Moreover, we also observe that the cardinalities computed in Theorem 4.2 implies that for a Sylow 5-subgroup P of B we have P5 (u, g 2 ) = ∅, so the 3125 ”extra”
elements obtained in B5 (u, g 2 ) come from non-trivial conjugates of P . This underscores the expected difficulties in a potential proof (or disproof) of Conjecture 2.7.
4.5. The Lyons group. There is exactly one other sporadic group with order
divisible by 56 (or pp+1 for p > 3): the Lyons group Ly.
Theorem 4.6. The maximal subgroup of Ly of the form 51+4 : 4.S6 has a faithful
permutation representation on 3,125 points, given by the action on the cosets of
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
19
4.S6 . Moreover, this maximal subgroup, Ly, and their Sylow 5-subgroups are all
non-F SZ5 .
Proof. It is well-known that Ly contains a copy of G2 (5) as a maximal subgroup,
and that the order of Ly is not divisible by 57 . Therefore Ly and G2 (5) have
isomorphic Sylow 5-subgroups, and by Theorem 4.1 this Sylow subgroup is not
F SZ5 .
Checking the character table for Ly as before, we find there is a unique nonidentity conjugacy class whose corresponding centralizer has order divisible by 56 .
In particular, the order of this centralizer is 2,250,000 = 24 ·32 ·56 , and it comes from
an element of order 5. So any maximal subgroup containing an element of order 5
whose centralizer has this order will suffice. The maximal subgroup 51+4 : 4.S6 is
the unique such choice.
The new difficulty here is that, by default, there are only matrix group representations available though the AtlasRep package for Ly and 51+4 : 4.S6 , which
are ill-suited for our purposes. However, faithful permutation representations for
Ly are known, and they can be constructed through GAP with sufficient memory
available provided one uses a well-chosen method. A detailed description of how to
acquire the permutation representation on 8,835,156 points, as well as downloads
for the generators (including MeatAxe versions courtesy of Thomas Breuer) can be
found on the web, courtesy Pfeiffer [12].
Using this, we can then obtain a permutation representation for the maximal subgroup 51+4 : 4.S6 on 8,835,156 points using the programs available on the online
ATLAS [16]. This in turn is fairly easily converted into a permutation representation on a much smaller number of points, provided one has up to 8 GB of memory
available, via SmallerDegreePermutationRepresentation. The author obtained
a permutation representation on 3125 points, corresponding to the action on the
cosets of 4.S6 . The exact description of the generators is fairly long, so we will not
reproduce them here. The author is happy to provide them upon request. One can
also proceed in a fashion similar to some of the cases handled in [1] to find such a
permutation representation.
Once this smaller degree representation is obtained, it is then easy to apply
the same methods as before to show the desired claims about the F SZ5 properties. We can directly compute the Sylow 5-subgroup, then find u, g through
FSZtestZ and FSIndPt irrespectively, set C to be the centralizer of g, then run
FSZSetCards(C,u,g,5,2). This returns [5000,625], which gives the desired nonF SZ5 claims.
Indeed, FSZtest can be applied to (both) the centralizer and the maximal subgroup once this permutation representation is obtained. This will complete quickly,
thanks to the relatively low orders and degrees involved. We also note that the centralizer C so obtained will not have a normal Sylow 5-subgroup, and is a perfect
group. The maximal subgroup in question is neither perfect nor solvable, and does
not have a normal Sylow 5-subgroup.
5. The F SZ sporadic simple groups
We can now show that all other sporadic simple groups and their Sylow subgroups
are F SZ.
20
MARC KEILBERG
Example 5.1. Any group which is necessarily F SZ (indeed, F SZ + ) by [4, Corollary 5.3] necessarily has all of its Sylow subgroups F SZ, and so satisfies the conjecture. This implies that all of the following sporadic groups, as well as their Sylow
p-subgroups, are F SZ (indeed, F SZ + ).
• The Mathieu groups M11 , M12 , M22 , M23 , M24 .
• The Janko groups J1 , J2 , J3 , J4 .
• The Higman-Simms group HS.
• The McLaughlin group M cL.
• The Held group He.
• The Rudvalis group Ru.
• The Suzuki group Suz.
• The O’Nan group O′ N .
• The Conway group Co3 .
• The Thompson group T h.
• The Tits group 2 F4 (2)′ .
Example 5.2. Continuing the last example, it follows that the following are the
only sporadic simple groups not immediately in compliance with the conjecture
thanks to [4, Corollary 5.3].
• The Conway groups Co1 , Co2 .
• The Fischer groups F i22 , F i23 , F i′24 .
• The Monster M .
• The Baby Monster B.
• The Lyons group Ly.
• The Harada-Norton group HN .
The previous section showed that the last four groups were all non-F SZ5 and have
non-F SZ5 Sylow 5-subgroups, and so conform to the conjecture. By exponent considerations the Sylow subgroups of the Conway and Fischer groups are all F SZ + .
The function FSZtest can be used to quickly show that Co1 , Co2 , F i22 , and F i23
are F SZ, and so conform to the conjecture.
This leaves just the largest Fischer group F i′24 .
Theorem 5.3. The sporadic simple group F i′24 and its Sylow subgroups are all
F SZ.
Proof. The exponent of F i′24 can be calculated from its character table and shown
to be
24,516,732,240 = 24 · 33 · 5 · 7 · 11 · 13 · 17 · 23 · 29.
As previously remarked, this automatically implies that the Sylow subgroups are all
F SZ (indeed, F SZ + ). By [4, Corollary 5.3] it suffices to show that every centralizer
of an element with order not in {1, 2, 3, 4, 6} in F i′24 that contains an element of
order 16 is F SZ. There is a unique conjugacy class in F i′24 for an element with
order (divisible by) 16. The centralizer of such an element has order 32, and is
isomorphic to Z16 × Z2 . So it suffices to consider the elements of order 8 in this
centralizer, and show that their centralizers (in F i′24 ) are F SZ. Every such element
has a centralizer of order 1536 = 29 · 3. So by Theorem 3.1 the result follows.
The following is GAP code verifying these claims.
G := AtlasGroup ( ” Fi2 4 ’ ” ) ; ;
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
21
GT := C h a r a c t e r T a b l e ( ” Fi2 4 ’ ” ) ; ;
P o s i t i o n s ( O r d e r s C l a s s R e p r e s e n t a t i v e s (GT) mod 1 6 , 0 ) ;
exp := Lcm( O r d e r s C l a s s R e p r e s e n t a t i v e s (GT) ) ;
C o l l e c t e d ( F a c t o r s I n t ( exp ) ) ;
SetExponent (G, exp ) ; ;
P := SylowSubgroup (G, 2 ) ; ;
#There are many ways t o g e t an e l e m e n t o f o r d e r 1 6 .
#Here ’ s a v e r y crude , i f non−d e t e r m i n i s t i c , one .
x := Random(P ) ; ;
while not Order ( x ) = 16 do x:=Random(P ) ; od ;
C := C e n t r a l i z e r (G, x ) ; ;
c e n t s := F i l t e r e d (C, y−>Order ( y ) = 8 ) ; ;
c e n t s := L i s t ( c e n t s , y−>C e n t r a l i z e r (G, y ) ) ; ;
L i s t ( cents , Size ) ;
The following then summarizes our results on sporadic simple groups.
Theorem 5.4. The following are equivalent for a sporadic simple group G.
(1) G is not F SZ.
(2) G is not F SZ5 .
(3) The order of G is divisible by 56 .
(4) G has a non-F SZ Sylow subgroup.
(5) The Sylow 5-subgroup of G is not F SZ5 .
Proof. Combine the results of this section and the previous one.
6. The symplectic group S6 (5)
In [13] it was mentioned that the symplectic group S6 (5) was likely to be the
second smallest non-F SZ simple group, after G2 (5). Computer calculations there
ran into issues when checking a particular centralizer, as the character table needed
excessive amounts of memory to compute. Our methods so far also place this group
at the extreme end of what’s reasonable. In principle the procedure and functions
we’ve introduced so far can decide that this group is non-F SZ in an estimated two
weeks of uninterrupted computations, and with nominal memory usage. However,
we can achieve a substantial improvement that completes the task in about 8 hours
(on two processes; 16 hours for a single process), while maintaining nominal memory
usage.
The simple yet critical observation comes from [13, Definition 3.3]. In particular,
if a ∈ Gm (u, g), then am = g implies that for all b ∈ classCG (g) (a) we have bm = g.
So while FSZSetCards acts as naively as possible and iterates over all elements of
22
MARC KEILBERG
C = CG (g), we in fact need to only iterate over the elements of those conjugacy
classes of C whose m-th power is g (or g n ). GAP can often compute the conjugacy
classes of a finite permutation or polycyclic group quickly and efficiently. So while
it is plausible that finding these conjugacy classes can be too memory intensive
for certain centralizers, there will nevertheless be centralizers for which all other
methods are too impractical for either time or memory reasons, but for which this
reduction to conjugacy classes makes both time and memory consumption a nonissue. The otherwise problematic centralizer of S6 (5) is precisely such a case, as we
will now see.
Theorem 6.1. The projective symplectic group S6 (5) and its Sylow 5-subgroup are
both non-F SZ5 .
Proof. As usual, our first task is to show that the Sylow 5-subgroup is non-F SZ5 ,
and then use the data obtained from that to attack S6 (5).
G :=
P :=
isoP
P :=
AtlasGroup ( ” S6 ( 5 ) ” ) ; ;
SylowSubgroup (G, 5 ) ; ;
:= IsomorphismPcGroup (P ) ; ;
Image ( i s o P ) ; ;
#Show P i s non−FSZ 5 , and
#g e t t h e g we need v i a FS Zt est Z
g := FSZtestZ (P ) [ 2 ] ;
#Get t h e u we need v i a FSIndPt
u := FSIndPt (P , 5 , g ) [ 1 ] ;
One can of course store the results of FSZtestZ and FSIndPt directly to see the
complete data returned, and then extract the specific data need.
We can then show that G = S6 (5) is itself non-F SZ5 by computing G5 (u, g) and
G5 (u, g 2 ) with the following code.
G :=
isoG
G :=
g :=
u :=
uinv
C e n t r a l i z e r (G, g ) ; ;
:= S m a l l e r D e g r e e P e r m u t a t i o n R e p r e s e n t a t i o n (G ) ; ;
Image ( isoG ) ; ;
Image ( isoG , g ) ; ;
Image ( isoG , u ) ; ;
:= I n v e r s e ( u ) ; ;
#Now we compute t h e c o n j u g a c y c l a s s e s
# of the c e n t r a l i z e r .
c l := C o n j u g a c y C l a s s e s (G ) ; ;
#We t h e n need o n l y c o n s i d e r t h o s e
# c l a s s e s w i t h a s u i t a b l e 5− t h power
cand1 := F i l t e r e d ( c l , x−>R e p r e s e n t a t i v e ( x)ˆ5=g ) ; ;
cand2 := F i l t e r e d ( c l , x−>R e p r e s e n t a t i v e ( x)ˆ5=g ˆ 2 ) ; ;
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
23
#There i s i n f a c t o n l y one c o n j u g a c y
# c l a s s in both cases .
Length ( cand1 ) ;
Length ( cand2 ) ;
cand1 := cand1 [ 1 ] ; ;
cand2 := cand2 [ 1 ] ; ;
#The f o l l o w i n g computes | G 5 ( u , g ) |
Number ( cand1 , x−>(x∗ uinv )ˆ5=g ) ;
#The f o l l o w i n g computes | G 5 ( u , g ˆ 2 ) |
Number ( cand2 , x−>(x∗ uinv )ˆ5=g ˆ 2 ) ;
This code shows that
|G5 (u, g)| = 1,875,000;
|G5 (u, g 2 )| = 375,000.
Therefore S6 (5) is non-F SZ5 , as desired.
The calculation of |G5 (u, g)| takes approximately 8.1 hours, and the calculation of |G5 (u, g 2 )| takes approximately 7.45 hours. The remaining calculations are
done in significantly less combined time. We note that the calculations of these
two cardinalities can be done independently, allowing each one to be calculated
simultaneously on separate GAP processes.
We also note that the centralizer in S6 (5) under consideration in the above
is itself a perfect group; is a permutation group of degree 3125 and order 29.25
billion; and has a non-normal Sylow 5-subgroup. Moreover, it can be shown that
the g we found yields the only rational class of P at which P fails to be F SZ. One
consequence of this, combined with the character table of S6 (5), is that, unlike in
the case of the Monster group, we are unable to switch to any other centralizer
with a smaller Sylow 5-subgroup to demonstrate the non-F SZ5 property.
Similarly as with the Baby Monster group, it is interesting to note that |P5 (u, g)| =
62,500 and |P5 (u, g 2 )| = 0 for P, u, g as in the proof. These cardinalities can be
quickly computed exactly as they were for S6 (5), simply restricted to P , or using the slower FSZSetCards, with the primary difference being that now there are
multiple conjugacy classes to check and sum over.
Before continuing on to the next section, where we consider small order perfect
groups available in GAP, we wish to note a curious dead-end, of sorts.
Lemma 6.2. Given u, g ∈ G with [u, g] = 1, let C = CG (g), D = CC (u), and
m ∈ N. Then a ∈ Gm (u, g) if and only if ad ∈ Gm (u, g) for some/any d ∈ D.
Proof. This is noted by Iovanov et al. [4] when introducing the concept of an F SZ +
group. It is an elementary consequence of the fact that D = CG (u, g) centralizes
both g and u by definition.
So suppose we have calculated those conjugacy classes in C whose m-th power is
g. As in the above code, we can iterate over all elements of these conjugacy classes
in order to compute |Gm (u, g)|. However, the preceding lemma shows that we could
instead partition each such conjugacy class into orbits under the D action. The
24
MARC KEILBERG
practical upshot then being that we need only consider a single element of each
orbit in order to compute |Gm (u, g)|.
In the specific case of the preceding theorem, we can show that the single conjugacy classes cand1 and cand2 both have precisely 234 million elements, and that
D is a non-abelian group of order 75,000, and is in fact the full centralizer of u in
S6 (5). Moreover the center of C is generated by g, and so has order 5. Thus in the
best-case scenario partitioning these conjugacy classes into D orbits can result in
orbits with |D/Z(C)| = 15,000 elements each. The cardinalities we computed can
also be observed to be multiples 15,000. That would constitute a reduction of more
than four orders of magnitude on the total number of elements we would need to
check. While this is a best-case scenario, since D also has index 390,000 in C it
seems very plausible that such a partition would produce a substantial reduction in
the number of elements to be checked. So provided that calculating these orbits can
be done reasonably quickly, we would expect a significant reduction in run-time.
There is a practical problem, however. The problem being that, as far as the
author can tell, there is no efficient way for GAP to actually compute this partition.
Doing so evidently requires that GAP fully enumerate and store the conjugacy class
in question. In our particular case, a conjugacy class of 234 million elements in a
permutation group of degree 3125 simply requires far too much memory—in excess
of 1.5 terabytes. As such, while the lemma sounds promising, it seems to be lacking
in significant practical use for computer calculations. It seems likely, in the author’s
mind, that any situation in which it is useful could have been handled in reasonable
time and memory by other methods. Nevertheless, the author cannot rule out the
idea as a useful tool.
7. Perfect groups of order less than 106
We now look for examples of additional non-F SZ perfect groups. The library
of perfect groups stored by GAP has most perfect groups of order less than 106 ,
with a few exceptions noted in the documentation. So we can iterate through the
available groups, of which there are 1097 at the time this paper was written. We
can use the function IMMtests from Section 3 to show that most of them are F SZ.
#Get a l l a v a i l a b l e s i z e s
G l i s t := F i l t e r e d ( S i z e s P e r f e c t G r o u p s ( ) ,
n−>N r P e r f e c t L i b r a r y G r o u p s ( n ) > 0 ) ; ;
#Get a l l a v a i l a b l e p e r f e c t g r o u p s
G l i s t := L i s t ( G l i s t ,
n−>L i s t ( [ 1 . . N r P e r f e c t L i b r a r y G r o u p s ( n ) ] ,
k−>P er fectGr o up ( IsPermGroup , n , k ) ) ) ; ;
G l i s t := F l a t ( G l i s t ) ; ;
#Remove t h e o b v i o u s l y FSZ on es
F l i s t := F i l t e r e d ( G l i s t , G−>not IMMtests (G)=true ) ; ;
This gives a list of 63 perfect groups which are not immediately dismissed as being
F SZ.
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
25
Theorem 7.1. Of the 1097 perfect groups of order less than 106 available through
the GAP perfect groups library, exactly 7 of them are not F SZ, all of which are
extensions of A5 . All seven of them are non-F SZ5 . Four of them have order
375,000 = 23 · 3 · 56 , and three of them have order 937,500 = 22 · 3 · 57 . Their perfect
group ids in the library are:
[375000, 2],
[375000, 8],
[375000, 9],
[937500, 3],
[937500, 4],
[937500, 5]
[375000, 11],
Proof. Continuing the preceding discussion, we can apply FSZtest to the 63 groups
in Flist to obtain the desired result. This calculation takes approximately two days
of total calculation time on the author’s computer, but can be easily split across
multiple GAP instances. Most of the time is spent on the F SZ groups of orders
375,000 and 937,500.
On the other hand, we can also consider the Sylow subgroups of all 1097 available
perfect groups, and test them for the F SZ property.
Theorem 7.2. If G is one of the 1097 perfect groups of order less than 106 available
through the GAP perfect groups library, then the following are equivalent.
(1)
(2)
(3)
(4)
G
G
G
G
is not F SZ.
has a non-F SZ Sylow subgroup.
has a non-F SZ5 Sylow 5-subgroup.
is not F SZ5 .
Proof. Most of the GAP calculations we need to perform now are quick, and the
problem is easily broken up into pieces, should it prove difficult to compute everything at once. The most memory intensive case requires about 1.7 GB to test.
With significantly more memory available than this, the cases can simply be tested
by FSZtest en masse, which will establish the result relatively quickly—a matter of
hours. We sketch the details here and leave it to the interested reader to construct
the relevant code. Recall that it is generally worthwhile to convert p-groups into
polycyclic groups in GAP via IsomorphismPcGroup.
Let Glist be constructed in GAP as before. Running over each perfect group,
we can easily construct their Sylow subgroups. We can then use IMMtests from
Section 3 to eliminate most cases. There are 256 Sylow subgroups, each from a
distinct perfect group, for which IMMtests is inconclusive; and there are exactly 4
cases where IMMtests definitively shows the non-F SZ property, which are precisely
the Sylow 5-subgroups of each of the non-F SZ perfect groups of order 375,000.
These 4 Sylow subgroups are all non-F SZ5 . We can also apply FSZtestZ to the
Sylow 5-subgroups of the non-F SZ perfect groups of order 937,500 to conclude
that they are all non-F SZ5 . All other Sylow subgroups remaining that come from
a perfect group of order less than 937,500 can be shown to be F SZ by applying
FSZtest without difficulty. Of the three remaining Sylow subgroups, one has a
direct factor of Z5 , and the other factor is easily tested and shown to be F SZ,
whence this Sylow subgroup is F SZ. This leaves two other cases, which are the
Sylow 5-subgroups of the perfect groups with ids [937500,7] and [937500,8]. The
second of these is easily shown to be FSZ by FSZtest. The first can also be
26
MARC KEILBERG
tested by FSZtest, but this is the case that requires the most memory and time—
approximately 15 minutes and the indicated 1.7 GB. In this case as well the Sylow
subgroups are F SZ. This completes the proof.
References
[1] John N. Bray and Robert A. Wilson. Explicit representations of maximal
subgroups of the monster. Journal of Algebra, 300(2):834 – 857, 2006.
ISSN 0021-8693. doi: http://dx.doi.org/10.1016/j.jalgebra.2005.12.017. URL
http://www.sciencedirect.com/science/article/pii/S0021869305007313.
[2] Pavel Etingof.
On some properties of quantum doubles of finite groups.
Journal of Algebra, 394:1 – 6, 2013.
ISSN 00218693.
doi:
http://dx.doi.org/10.1016/j.jalgebra.2013.07.004.
URL
http://www.sciencedirect.com/science/article/pii/S0021869313003529.
[3] GAP.
GAP – Groups, Algorithms, and Programming, Version 4.8.4.
http://www.gap-system.org, Jun 2016.
[4] M. Iovanov, G. Mason, and S. Montgomery. F SZ-groups and Frobenius-Schur
indicators of quantum doubles. Math. Res. Lett., 21(4):1–23, 2014.
[5] Yevgenia Kashina, Yorck Sommerhäuser, and Yongchang Zhu.
On
higher Frobenius-Schur indicators. Mem. Amer. Math. Soc., 181(855):
viii+65, 2006.
ISSN 0065-9266.
doi: 10.1090/memo/0855.
URL
http://dx.doi.org/10.1090/memo/0855.
[6] M. Keilberg. Examples of non-FSZ p-groups for primes greater than three.
ArXiv e-prints, September 2016. under review.
[7] Marc Keilberg.
Higher indicators for some groups and
their doubles.
J. Algebra Appl.,
11(2):1250030, 38,
2012.
ISSN 0219-4988.
doi:
10.1142/S0219498811005543.
URL
http://dx.doi.org/10.1142/S0219498811005543.
[8] Marc Keilberg.
Higher Indicators for the Doubles of Some Totally Orthogonal Groups.
Comm. Algebra, 42(7):2969–2998, 2014.
ISSN 0092-7872.
doi:
10.1080/00927872.2013.775651.
URL
http://dx.doi.org/10.1080/00927872.2013.775651.
[9] C. Negron and S.-H. Ng. Gauge invariants from the powers of antipodes. ArXiv
e-prints, September 2016.
[10] Siu-Hung Ng and Peter Schauenburg.
Higher Frobenius-Schur indicators for pivotal categories.
In Hopf algebras and generalizations, volume 441 of Contemp. Math., pages 63–90. Amer. Math.
Soc., Providence, RI, 2007.
doi: 10.1090/conm/441/08500.
URL
http://dx.doi.org/10.1090/conm/441/08500.
[11] Siu-Hung Ng and Peter Schauenburg. Central invariants and higher indicators for semisimple quasi-Hopf algebras. Trans. Amer. Math. Soc., 360(4):
1839–1860, 2008. ISSN 0002-9947. doi: 10.1090/S0002-9947-07-04276-6. URL
http://dx.doi.org/10.1090/S0002-9947-07-04276-6.
[12] Markus J. Pfeiffer.
Computing a (faithful) permutation representation of Lyons’ sporadic simple group,
2016.
URL
https://www.morphism.de/~markusp/posts/2016-06-20-computing-permutation-representation-ly
[13] P. Schauenburg. Higher frobenius-schur indicators for drinfeld doubles of finite groups through characters of centralizers. ArXiv e-prints, April 2016. in
preparation.
THE FSZ PROPERTIES OF SPORADIC SIMPLE GROUPS
27
[14] Peter Schauenburg. Some quasitensor autoequivalences of drinfeld doubles of
finite groups. Journal of Noncommutative Geometry, 11:51–70, 2017. doi:
10.4171/JNCG/11-1-2. URL http://dx.doi.org/10.4171/JNCG/11-1-2.
[15] R. A. Wilson, R. A. Parker, S. Nickerson, J. N. Bray, and T. Breuer. AtlasRep, a gap interface to the atlas of group representations, Version 1.5.1.
http://www.math.rwth-aachen.de/~Thomas.Breuer/atlasrep, Mar 2016.
Refereed GAP package.
[16] Robert Wilson, Peter Walsh, Jonathan Tripp, Ibrahim Suleiman, Richard
Parker, Simon Norton, Simon Nickerson, Steve Linton, John Bray, and
Rachel Abbott. ATLAS of Finite Group Representations - Version 3. URL
http://brauer.maths.qmul.ac.uk/Atlas/v3/. Accessed: October 3, 2016.
E-mail address: [email protected]
| 4math.GR
|
SceneFlowFields: Dense Interpolation of Sparse Scene Flow Correspondences
René Schuster1
Oliver Wasenmüller1
Georg Kuschk2
Christian Bailer1
Didier Stricker1
1
2
DFKI - German Research Center for Artificial Intelligence
BMW Group
arXiv:1710.10096v1 [cs.CV] 27 Oct 2017
firstname.lastname@{dfki,bmw}.de
Abstract
While most scene flow methods use either variational optimization or a strong rigid motion assumption, we show for
the first time that scene flow can also be estimated by dense
interpolation of sparse matches. To this end, we find sparse
matches across two stereo image pairs that are detected
without any prior regularization and perform dense interpolation preserving geometric and motion boundaries by
using edge information. A few iterations of variational energy minimization are performed to refine our results, which
are thoroughly evaluated on the KITTI benchmark and additionally compared to state-of-the-art on MPI Sintel. For
application in an automotive context, we further show that
an optional ego-motion model helps to boost performance
and blends smoothly into our approach to produce a segmentation of the scene into static and dynamic parts.
a)
b)
c)
1. Introduction
Scene flow describes the perceived 3D motion field with
respect to the observer. It can thereby be considered as an
extension to optical flow, which in comparison is the apparent motion field in 2D image space. Many applications
such as robot navigation, high level vision tasks, e.g. moving object detection [25], and driver assistance systems rely
on an accurate motion estimation of their surroundings. Especially the latter ones have great potential to make traffic
more comfortable and much safer. While scene flow is not
only a more detailed representation of the real world motion compared to optical flow, scene flow algorithms also
reconstruct the 3D geometry of the environment. Due to
the increased complexity when compared to depth or optical flow estimation, scene flow has only recently become
of bigger interest. At the same time, simply estimating
depth and optical flow separately to obtain scene flow (see
[19, 28, 29, 31]) is not exploiting the full potential of the
underlying data. The splitting produces incoherent results,
limits the exploitation of inherent redundancies, and in general yields a non-dense scene flow field. Approaches that
combine stereo depth estimation, e.g. [15], and 2D optical
d)
Figure 1. We present SceneFlowFields which uses stereo image
pairs (a), extracts scene flow boundaries (b), and computes a dense
scene flow field (c) that we compare to ground truth (d). The color
of the point clouds encodes the optical flow.
flow, e.g. [4, 30], have clearly been outperformed by a huge
margin in the KITTI benchmark [10].
Most state-of-the-art approaches are either designed for
indoor scenarios or describe outdoor scenes under mostly
stationary or rigid motion assumptions. Contrary to that,
our method is very versatile. In fact, we do not employ any
a-priori regularization. Our matching process inherently encodes a first order local smoothness assumption and is itself
solely data term based, while our interpolation scheme allows for very sharp discontinuities in the scene flow field.
An optional ego-motion extension that incorporates additional assumptions improves accuracy for challenging traf-
Figure 2. Overview of our SceneFlowFields. Blue color indicates
the optional ego-motion extension.
fic data sets like KITTI, but is no mandatory part of our
method. The basic version of the method and its optional
extension are illustrated in Figure 2. The single stages of
our approach are visualized in the supplementary video1 .
In more detail, we present a new scene flow approach –
called SceneFlowFields – that densely interpolates sparse
scene flow matches. The interpolation preserves boundaries of the 3D geometry and of moving objects by edgepreserving interpolation based on an improved edge detector to approximate scene flow boundaries. Matches are obtained by multi-scale propagation with random search to
compute a dense scene flow field that in turn is filtered to
remove outliers and leave sparse, robust correspondences
across all images. The combination of multi-scale matching
and edge-preserving interpolation sums to a novel method
which is in strong contrast with any existing method that
estimates scene flow. Matching can optionally be used to
estimate ego-motion and obtain sparse motion indicators
which can be interpolated to a dense motion segmentation.
Optionally using ego-motion, we can reconstruct the scene
flow for static parts of the scene directly. However, our
method does not rely on ego-motion estimation. In particular, our contribution consists of:
• A novel method to find scene flow matches.
• A new interpolation method for scene flow that preserves boundaries of geometry and motion.
• An improved edge detector to approximate scene flow
boundaries.
• An optional approach for straightforward integration
of ego-motion.
• Thorough evaluation of our method on KITTI and MPI
Sintel with comparison to state-of-the-art methods.
2. Related Work
Starting from Vedula et al. [32] who was among the first
to compute 3D scene flow, many variational approaches followed. First using pure color images as input [2, 17] and
later using RGB-D images [13, 18, 37]. While a varia1 https://youtu.be/7fsXq5EA0Rw
tional formulation is typically complex, [18] achieved realtime performance with a primal-dual framework. Yet, all
these approaches are sensitive to initialization and can not
cope with large displacements [36], which is why they use
a coarse-to-fine scheme. That in turn tends to miss finer
details. Furthermore, the RGB-D approaches rely on depth
sensors that either perform poorly in outdoor scenarios or
are accordingly very expensive [39].
Since it is hard to capture ground truth scene flow information, there exist only very few data sets to evaluate
scene flow algorithms on. Most of them use virtually rendered scenes to obtain the ground truth data [5, 9, 22]. To
the best of our knowledge the only realistic data set that
provides a benchmark for scene flow is the KITTI Vision
Benchmark [10] that combines various tasks for automotive
vision. Its introduction has played an important role in the
development of stereo and optical flow algorithms, and the
extension by [23] has also driven the progress in scene flow
estimation.
Due to the advent of a piece-wise rigid plane model [33],
scene flow has recently achieved a boost in performance.
The majority of top performing methods at the KITTI Vision Benchmark employ this model to enforce strong regularization [21, 23, 34]. In [34] the authors encode this model
by alternating assignment of each pixel to a plane segment
and each segment to a rigid motion, based on a discrete set
of planes and motions. In [23] the complexity of the model
is further lowered by the assumption that a scene consists
of only very few independently moving rigid objects. Thus
each plane segment only needs to be assigned to one object.
All segments assigned to the same object share the same
motion. By propagation of objects over multiple frames,
[24] achieves temporal consistency for [23]. The authors of
[21] solve the pixel-to-plane assignment and the plane-tomotion assignment in a continuous domain.
Another promising strategy builds on the decomposition
of a scene into static and moving parts [31]. While the motion of dynamic objects is estimated by solving a discrete labeling problem (as in [6]) using the Semi-Global-Matching
(SGM) [15] algorithm, the perceived motion of all static
parts is directly obtained from the 3D geometry of the scene
and the ego-motion of the camera. This approach is especially convenient for scenes, where only a small proportion
consists of moving objects, like it is typically the case in
traffic scenarios. However, any a-priori assumption limits
the versatility of a method. A rigid plane model performs
poorly when applied to deformable objects, and ego-motion
estimation for highly dynamic scenes is hard.
Our scene flow method differs from any of the mentioned
approaches. We find sparse scene flow matches that are interpolated to a dense scene flow field, recovering the geometry of the scene and the 3D motion. Our method has to be
distinguished from purely variational approaches. Although
we use variational optimization, it can be considered as a
post-processing step for refinement. During interpolation,
we assume that the geometry of a scene can be modeled
by small planar segments, but we do not initially presume
any segmentation. In fact, the size of our plane segments
only depends on the density of the matches, which leads
to smoothly curved shapes where matches are dense and to
planar patches where matches are sparse. The same holds
for our piece-wise affine motion model that is used to interpolate the 3D motion. These differences in the model and
additionally the difference in the optimization method draw
a clear boundary between our method and [21, 23, 34]. If
we apply our optional ego-motion model there is a conceptual overlap to [31] which also uses ego-motion to estimate
the 3D motion for the static parts of the scene. However, if
we do not apply the ego-motion model, both methods have
no noteworthy similarities. In any case, the way we estimate the ego-motion and the motion segmentation differs
essentially from [31].
Finally, one has to differentiate between dual-frame
[21, 23, 34] and multi-frame [24, 31, 34] approaches. Especially the images of KITTI have several characteristics
that make matching between two frame pairs much more
challenging than in a multi-view setting. First, considerably large stereo and flow displacements. Second, difficult
lighting conditions and many reflective surfaces. Third, fast
ego-motions combined with a low frame rate, which causes
large regions to move out of image bounds. This has to be
kept in mind when comparing results across these two categories.
3. SceneFlowFields
For scene flow computation we assume to have the typical stereo image information provided, i.e. two rectified
stereo image pairs (Il0 , Il1 , Ir0 , Ir1 ) at times t0 and t1 along
with the camera intrinsics. We further assume that the baseline B is known. For rectified images, the baseline describes
the relative pose between the left and right cameras as translation parallel to the image plane. We represent scene flow
as a 4D vector u = (u, v, d0 , d1 )T consisting of two optical flow components u, v and the disparity values d0 , d1
for both time steps. During matching, we jointly optimize
all four components to obtain coherent scene flow. Given
the mentioned information, we estimate a dense scene flow
field as follows: For k subscales we initialize the coarsest by
finding the best correspondences from kD-trees build with
feature vectors using Walsh-Hadamard-Transform (WHT)
[12]. For all k + 1 scales (the k subscales plus full resolution), we iteratively propagate scene flow vectors and adjust
them by random search. Afterwards, the dense scene flow
map on full resolution is filtered using an inverse scene flow
field and a region filter. The filtered scene flow map is further thinned out by only taking the best match in each non-
overlapping 3×3 block. Scene flow boundaries are detected
using a structured random forest. Geometry and 3D motion are separately interpolated based on a boundary-aware
neighborhood. Finally, we refine the 3D motion by variational optimization. An overview is outlined in Figure 2.
3.1. Sparse Correspondences
Matching Cost. The matching cost in our algorithm
solely depends on a data term. No additional smoothness
assumptions are made like e.g. in [13, 17, 21, 23, 33, 34].
Given a scene flow vector, we define its matching cost by
the sum of Euclidean distances between SIFTFlow features
[20] over small patches for three image correspondences.
These correspondences are the stereo image pair at time t0 ,
the temporal image pair for the left view point (standard optical flow correspondence) and a cross correspondence between the reference frame and the right frame at the next
time step. This leads to the following cost C for a scene
flow vector u at pixel p:
X
+
φ Il0 (p̃) − φ Il1 p̃ + (u, v)T
C=
p̃∈W (p)
+
φ Il0 (p̃) − φ Ir0 p̃ + (−d0 , 0)T
0
1
T
φ Il (p̃) − φ Ir p̃ + (u − d1 , v)
(1)
W (p) is a 7 × 7 patch window centered at pixel p and
φ (I (p)) returning the first three principal components of
a SIFT feature vector for image I and pixel p. The principal axes are computed for the combined SIFT features of all
four images. At image boundaries we replicate the boundary pixel to pad images.
Initialization. Initialization is based on kD-trees similar
to [11], but with three trees, using WHT features as in
[1, 35]. For each frame other than the reference frame, we
compute a feature vector per pixel and store them in a tree.
To initialize a pixel of the reference image, we query the
feature vector of the pixel to the pre-computed kD-trees.
Scene flow matches are then obtained by comparing all
combinations of the leafs for each queried node according
to the matching data term introduced before (Equation 1).
Since our stereo image pairs are rectified, for the images
observed from the right camera view, we create kD-trees
which regard the epipolar constraint, i.e. queries for such a
tree will only return elements, which lie on the same image row as the query pixel. This way, we can efficiently
lower the number of leaves per node for the epipolar trees,
which speeds up the initialization process without loss of
accuracy. For further acceleration, we use this initialization
on the coarsest resolution only, and let the propagation fill
the gaps when evolving to the next higher scale.
Multi-Scale Propagation. The initial matches get spread
by propagation and steadily refined by random search. This
is done over multiple scales which helps to distribute rare
correct initial matches over the whole image. For each
scale, we run several iterations of propagation in one out
of the four image quadrants so that each direction is used
equally. During propagation, a scene flow vector is replaced
if the propagated vector has a smaller matching cost. If this
is not the case, the propagation along this path continues
with the existing scene flow vector. After each iteration we
perform a random search. That means that for all pixels
we add a uniformly distributed random offset in the interval ]−1, 1[ in pixel units of the current scale to each of the
four scene flow components and check whether the matching cost decreases. Both propagation and random search
help to obtain a smoothly varying vector field and to find
correct matches even if the initialization is slightly flawed.
For the different scale spaces, we simulate the scaling by
smoothing the images and taking only every n-th pixel for
a subsampling factor of n = 2k so that the patches consist of the same number of pixels for all scales. This way,
we prevent (up)sampling errors because all operations are
performed on exact pixel locations on the full image resolution. Smoothing is done by area-based downsampling
followed by upsampling using Lanczos interpolation. Note
that this matching method has already been used in [1], but
while they use it for optical flow, we apply it to twice as
many dimensions in search space.
Consistency Check. The matching procedure yields a
dense map of scene flow correspondences across all images. However, many of the correspondences are wrong
because of occlusions, out-of-bounds motion or simply because of mismatching due to challenging image conditions.
To remove these outliers, we perform a two-step consistency check. First, we compute an inverse scene flow field
for which the reference image is the right image at time
t1 . Temporal order as well as points of view are swapped.
Everything else remains as explained above. During consistency check, optical flow and both disparity maps for each
pixel are compared to the corresponding values of the inverse scene flow field. If either difference exceeds a consistency threshold τc in image space, the scene flow vector
gets removed. Secondly, we form small regions of the remaining pixels as in [1], where a pixel is added to a region
if it has approximately the same scene flow vector. Afterwards, we check if we could add one of the already removed
outliers in the neighborhood following the same rule. If this
is possible and the region is smaller than sc pixels, we remove the whole region. This way we obtain the filtered
final scene flow correspondences of high accuracy and very
few outliers (cf. Table 2). Because most of the times the
joint filtering of the matches removes more disparity values than necessary, we fill up gaps with additional values.
These values are the result of a separate consistency check
for the disparity matches only. For the separate check we
a)
b)
c)
Figure 3. Sparse correspondences (left) and dense interpolation
(right). Optical flow (a) and disparities at t0 (b) and t1 (c).
compute a second disparity map with SGM [15] and use the
same threshold τc as before. The additional disparity values
that are retrieved this way are as accurate as the one from
our standard consistency check but much denser which is
shown in Figure 3 and Table 2.
3.2. Dense Interpolation
Sparsification. Before interpolating the filtered scene
flow field to recover full density, an additional sparsification
step is performed. This helps to extend the spatial support
of the neighborhood during interpolation and speeds up the
whole process [1]. For non-overlapping 3 × 3 blocks, we
select the match with the lowest consistency error during
filtering only. The remaining matches are called seeds with
respect to the interpolation.
Interpolation Boundaries. A crucial part of the interpolation is the estimation of scene flow boundaries. While
[1, 26] approximate motion boundaries for optical flow with
a texture-agnostic edge detector [8], our edge detector is
trained on semantic boundaries. We find that this models
geometric boundaries as well as motion boundaries much
better than image edges and is much more robust to lighting,
shadows, and coarse textures. To do so, we have gathered
about 400 images of the KITTI data set from [14, 27, 38]
that have been labeled with semantic class information.
Within these images we have merged semantic classes that
in general neither align with geometric nor motion discontinuities, e.g. lane markings and road, or pole and panel.
The boundaries between the remaining semantic labels are
used as binary edge maps to train our edge detector. To this
end, we utilize the framework of Structured Edge Detection
(SED) [8] and train a random forest with the same parameters as in their paper, except for the number of training
patches. We sample twice as many positive and negative
patches during training because we use a bigger data set
with images of higher resolution. The impact of the novel
boundary detector will be evaluated in Section 5.
Interpolation Models. For the interpolation of geometry
and motion, we use two different models. Both parts are
interpolated separately which leads to a more accurate re-
a)
b)
with the same closest seed share the same local neighborhood. And secondly, the distance between p̂ and its closest seed is a constant offset for all neighboring seeds which
can be neglected. It is therefore sufficient to find a labeling that assigns each pixel to its closest seed and to find the
local neighborhood for each seed. We use the graph-based
method of [26] for this. Where the distances between seeds
are geodesic distances that are directly based on the edge
maps from our boundary detector.
3.3. Variational Optimization
c)
Figure 4. Whereas SED [8] (b) detects all image boundaries, our
new boundary detector (c) suppresses lane markings and shadows.
construction of the scene. This is due to the fact that the
separate consistency check for disparity leaves more geometric matches where motion would leave image boundaries. Suppose a local, boundary-aware neighborhood of
seeds is given for each unknown scene flow vector û at
pixel p̂ for geometric and motion seeds respectively, Ngeo
and Nmotion . The depth of pixel p̂ is reconstructed by fitting a plane E(p̂) : a1 x + a2 y + a3 = d0 through all seeds
of the neighborhood Ngeo . This is done by solving a linear system of equations for all neighboring seed points pg
for which the disparity values are known, using weighted
least squares. The weights for each seed are obtained from
a Gaussian kernel g(D) = exp (−αD) on the distance
D(p̂, pg ) between target pixel and seed. The missing disparity value of p̂ is obtained by plugging the coordinates
of p̂ into the estimated plane equation. In a similar fashion, but using a neighborhood of motion seeds Nmotion ,
the missing 3D motion is obtained by fitting an affine 3D
transformation x1 = Ax0 + t using weighted least squares
on all motion seeds pm . Where xt = (xt , yt , zt )T are the
3D world coordinates of motion seed pm at time t0 and t1 ,
and [A|t] ∈ R3×4 is the affine 3D transformation of twelve
unknowns. The weights are computed by the same Gaussian kernel as for geometric interpolation, but using the distances D(p̂, pm ) between the target pixel and the motion
seeds. To summarize, for the full reconstruction of scene
flow û = (u, v, d0 , d1 )T at pixel p̂, we compute d0 using
the plane model E(p̂), reproject the point into 3D world
space, transform it according to its associated affine transformation [A|t], and project it back to image space to obtain
u, v and d1 .
Edge-Aware Neighborhood. To find the local neighborhoods, we follow the idea of [26] using both their approximations. That is first, the n closest seeds to a pixel p̂ are the
n − 1 closest seeds to the closest seed of p̂, thus all pixels
To further refine the 3D motion after interpolation, we
use variational energy minimization to optimize the objective
f low
cross
E(u, v, d0 ) = Edata
+ Edata
+ ϕ · Esmooth
(2)
Motion is represented in image space by optical flow and the
change in disparity d0 . The energy consists of three parts.
Two data terms, one temporal correspondence and one cross
correspondence, and an adaptively weighted smoothness
term for regularization. The data terms use the gradient
constancy assumption. Our experiments have shown, that
a term for the color constancy assumption can be neglected.
∗
Edata
(I1 , I2 , x, w) =
Z
2
β (x, w) · Ψ γ · |∇I2 (x + w) − ∇I1 (x)| dx
(3)
Ω
f low
∗
Edata
= Edata
Il0 , Il1 , x, (u, v)T
cross
Edata
=
∗
Edata
Il0 , Ir1 , x, (u
0
− d0 − d , v)
(4)
T
(5)
The data terms do not contribute to the energy if the function
(
T
1, if (x + w) ∈ Ω
β (x, w) =
(6)
0, otherwise
indicates that the scene flow is leaving the image domain.
The smoothness term
Z
2
2
2
Esmooth =
Ψ |∇u| + |∇v| + λ · |∇d0 | dx (7)
Ω
penalizes changes in the motion field and is weighted by
ϕ(x) = e−κB(x)
(8)
where B(x) is the edge value of our boundary detector
at
pixel x. All parts use the Charbonnier penalty Ψ x2 =
√
x2 + ε2 to achieve robustness. Since the smoothness term
rather enforces constancy if β for both data terms is zero, we
do not optimize the scene flow at pixels where the interpolated scene flow field leaves Ω. Our energy formulation is
inspired by [3, 17, 37]. We use linear approximations of the
Euler-Lagrange equations for the objective and apply the
framework of Brox et al. [3] without the coarse-to-fine steps
to find a solution by Successive Over-Relaxation (SOR).
a)
b)
c)
Figure 5. Example of our motion segmentation. Sparse motion
indicators as obtained during ego-motion computation (a), dense
segmentation by interpolation (b) and moving ground truth objects
as provided by KITTI [23] (c).
4. Ego-Motion Model
In Section 5 we will show that our approach as described
so far achieves results comparable to state-of-the-art. For
the special challenges of the KITTI data set, we make an
additional, optional assumption to further improve the performance of SceneFlowFields. Following [31], we argue
that most parts of a scene are static and thus that the 3D motion for these areas is fully determined by the ego-motion of
the observer. Given the ego-motion and a motion segmentation into static and dynamic areas, we apply the inverse egomotion to all static points in the scene. Using our matching
and interpolation scheme, both can be easily estimated with
almost no additional effort.
Ego-Motion Estimation. The filtered scene flow field before interpolation provides very accurate matches across all
images. We compute 3D-2D correspondences between the
reference frame and the temporally subsequent frame by triangulation with the stereo matches. We limit the depth of
these correspondences to 35 meters because disparity resolution for farther distances gets too inaccurate. This way, we
obtain a Perspective-n-Point problem, which we solve iteratively using Levenberg-Marquardt and RANSAC to find the
relative pose between the left cameras at time t0 and t1 by
minimizing the re-projection error of all correspondences.
For RANSAC, we consider a correspondence an outlier if
the re-projection error is above 1 pixel. After first estimation, we recompute the set of inliers with a relaxed threshold
of 3 pixels and re-estimate the pose P = [R|t] ∈ R3×4 . The
two stage process helps to avoid local optima and to find a
trade-off between diverse and robust correspondences.
Motion Segmentation. An initial sparse motion segmentation can directly be obtained as side product of the egomotion estimation. Outliers in the correspondences are
considered in motion, while points in conformity with the
estimated ego-motion are marked as static. We use our
boundary-aware interpolation to compute a dense segmentation (cf. Figure 5). Pixels labeled as moving are spread
up to the object boundaries within they are detected. Because the segmentation is only a binary labeling, no complex interpolation model is needed. An unknown pixel gets
assigned with the weighted mean of its local neighborhood.
The weights are again based on the geodesic distances between matches. This interpolation method is similar to
the Nadaraya-Watson estimator in [26]. The interpolated
motion field is then thresholded to obtain a dense, binary
motion segmentation. The quality of this segmentation is
evaluated in Section 5. Finally, the inverse estimated egomotion is applied to all points that are labeled as static.
5. Experiments and Results
We use the explicit values of the previous sections and
the following parameters for all our experiments, even
across different data sets. For k = 3 subscales and full
resolution we run 12 iterations of propagation and random
search. We use a consistency threshold of τc = 1 and a minimal region size of sc = 150 for the region filter. During interpolation we use geometry and motion neighborhoods of
160 and 80 seeds respectively and α = 2.2 for the Gaussian
kernel to weight the geodesic distances. For the variational
energy minimization we set κ = 5, γ = 0.77, λ = 10 and
ε = 0.001. We run two outer and one inner iteration in
our optimization framework with 30 iterations for the SOR
solver using a relaxation factor of ω = 1.9. We threshold
the interpolated motion field with τS = 0.4 to obtain a binary segmentation when applying the ego-motion model.
5.1. Boundary Detection
To test the impact of our motion boundary detector, we
evaluate different variants of our method twice. Once, using
standard edge detection as in [8], and a second time using
our structured random forest trained on semantic edges. The
results are compared in Table 2. The major improvements
are visualized in Figure 4. High image gradients at lane
markings or shadows (especially shadows of vehicles) are
effectively suppressed when using our boundary detector,
while at the same time it accurately detects all kinds of objects. This helps greatly to smoothly recover the street surface during interpolation, sharpen discontinuities in depth
and motion in general, and allows for accurate boundaries
when interpolating the motion segmentation.
5.2. KITTI Scene Flow Benchmark
Our main experiments are taken out on the KITTI Scene
Flow Benchmark [23]. The results of our public submission are presented in Table 1 where we compare to
state-of-the-art methods. At the time of writing, our method
Rank
Method
1
2
3
4
5
6
7
8
9
10
PRSM [34]
OSF+TC [24]
OSF [23]
FSF+MS [31]
CSF [21]
SceneFlowFields
PRSF [33]
SGM+SF [15, 16]
PCOF-LDOF [7]
PCOF+ACTF [7]
dual
x
x
x
x
x
x
x
D1-bg
D1-fg
D1-all
D2-bg
D2-fg
D2-all
Fl-bg
Fl-fg
Fl-all
SF-bg
SF-fg
SF-all
Runtime
3.02
4.11
4.54
5.72
4.57
5.12
4.74
5.15
6.31
6.31
10.52
9.64
12.03
11.84
13.04
13.83
13.74
15.29
19.24
19.24
4.27
5.03
5.79
6.74
5.98
6.57
6.24
6.84
8.46
8.46
5.13
5.18
5.45
7.57
7.92
8.47
11.14
14.10
19.09
19.15
15.11
15.12
19.41
21.28
20.76
21.83
20.47
23.13
30.54
36.27
6.79
6.84
7.77
9.85
10.06
10.69
12.69
15.60
20.99
22.00
5.33
5.76
5.62
8.48
10.40
10.58
11.73
20.91
14.34
14.89
13.40
13.31
18.92
25.43
25.78
24.41
24.33
25.50
38.32
60.15
6.68
7.02
7.83
11.30
12.96
12.88
13.83
21.67
18.33
22.43
6.61
7.08
7.01
11.17
12.21
12.48
13.49
23.09
25.26
25.77
20.79
20.03
26.34
33.91
33.21
32.28
31.22
34.46
49.39
67.75
8.97
9.23
10.23
14.96
15.71
15.78
16.44
24.98
29.27
32.76
300 s
3000 s
3000 s
2.7 s
80 s
65 s
150 s
2700 s
50 s
(0.08 s)
Variant
D1-bg
D1-fg
D1-all
D2-bg
D2-fg
D2-all
Fl-bg
Fl-fg
Fl-all
SF-bg
SF-fg
SF-all
Density
full+ego
full
no var
5.36
5.36
5.36
10.85
10.85
10.85
6.20
6.20
6.20
7.94
15.91
15.77
18.23
18.03
18.81
9.51
16.23
16.24
10.36
22.33
23.75
22.85
21.69
23.72
12.28
22.23
23.75
12.04
24.78
25.60
28.31
27.37
28.99
14.53
25.18
26.12
100.00 %
100.00 %
100.00 %
Ours
full+ego
full
no var
5.48
5.48
5.48
11.99
11.99
11.99
6.47
6.47
6.47
9.07
16.90
16.73
19.98
20.80
21.38
10.74
17.50
17.44
11.63
23.57
25.00
25.47
25.22
27.04
13.75
23.82
25.32
13.39
26.14
26.93
30.92
30.71
32.16
16.07
26.84
27.73
100.00 %
100.00 %
100.00 %
SED [8]
Table 1. Results on the KITTI Scene Flow Benchmark [23]. The column dual indicates whether only two frame pairs are used by this
method. Run times in parentheses are using a GPU. We achieve the third best result among all dual-frame methods. Our SceneFlowFields
yields especially good results at foreground regions (SF-fg).
Edges
matches
disparity
1.91
1.42
3.74
4.06
2.18
1.82
2.48
–
4.08
–
2.71
–
2.10
–
2.78
–
2.20
–
3.87
–
6.24
–
4.21
–
38.82 %
57.81 %
–
–
Table 2. Evaluation of the different parts of our method on KITTI [23] training data. Our new edge detector outperforms SED [8]. The egomotion model helps greatly to improve overall results. The bottom two rows show the amount of outliers for our sparse correspondences
before interpolation. The density is computed with respect to available KITTI ground truth.
was ranked 6th and achieved the 3rd best result out of all
dual-frame methods while at the same time being considerably faster than the top three performing methods (cf. Table 1). That our method generalizes better to other data sets
is shown in Section 5.3 where we often outperform the best
dual-frame method [23]. In Figure 6, we give an visual
example of our results and compare it to the two top performing methods in dual-frame [23] and multi-frame [34]
categories. It can be seen that our interpolation produces
very sharp edges. This in combination with our matching
method helps to obtain accurate scene flow, especially for
(moving) objects. Methods with comparable overall performance on KITTI ([21, 31]), perform worse on moving
foreground objects than our SceneFlowFields.
Apart from the official evaluation, we test the different
components of our method in Table 2 to evaluate the effect
of each part. We use all training images of the KITTI data
set, and evaluate our basic method without the variational
optimization (no var), the full basic approach (full) and our
method with the optional ego-motion extension (full+ego).
Additionally we compute the accuracy and densities with
respect to the KITTI ground truth of our sparse scene flow
matches (matches) and the separately filtered sparse stereo
correspondences (disparity). The variational optimization is
primarily useful for optical flow and foreground. All variants using our improved edge detector outperform their according variant using basic image edges.
Finally, we use the provided object maps of KITTI, to
test the performance of our motion segmentation (cf. Figure 5). To this end, we compute the precision and recall
for our binary segmentation. Precision is defined as the percentage of estimated pixels that are correctly labeled as in
motion. The recall is the relative amount of ground truth
pixels that are labeled as moving and covered by our estimation. Over all frames, we achieve a precision of about 28 %
and a recall of about 83 %. Most of the missed ground truth
foreground pixels belong to objects which are far away and
moving parallel to the direction of viewing. This way, the
re-projection error of the 3D-2D correspondences during
ego-motion estimation is below the threshold. Two remarks
have to be considered regarding the precision. First, KITTI
only annotates cars that are mostly visible, i.e. pedestrians,
cyclists, other vehicles, or partly occluded cars are not included in the ground truth, but will be marked as moving if
they are in motion. Secondly, since areas that are wrongly
classified as dynamic will be filled with our basic scene flow
estimation which is still of high quality, we did tune in favor
of a high recall.
5.3. MPI Sintel
We claim that our proposed method is very versatile and
not at all restricted to any setup. Therefore, we have additionally evaluated SceneFlowFields on MPI Sintel [5] without changing any parameters. The only difference to our
PRSM [34] – multi-frame
SceneFlowFields (ours) – dual-frame
OSF [23] – dual-frame
a)
b)
c)
d)
Figure 6. Exemplary visual comparison on KITTI Scene Flow Benchmark [23]. We show disparity (a) and optical flow (c) results along
with the corresponding error maps (b) and (d) for PRSM [34], OSF [23] and our SceneFlowFields. We accurately detect moving objects
and reconstruct sharp boundaries. More examples are visualized in the supplementary video and on the public homepage of KITTI.
evaluation on KITTI is, that we do not use our semantic
edge detector which was trained on KITTI imagery, but
instead SED to obtain edge maps. We test both, our basic approach and the ego-motion extension (+ego) for our
method. All training frames over all but two sequences for
which a subsequent frame exists are processed. The final
rendering passes for all images are used. We measure the
percentage of outliers according to the KITTI metric for
disparity and optical flow. The sequences cave 2 and sleeping 1 are left out because they have not been evaluated in
[31] – to which we want to compare – due to varying camera parameters. The relative amounts of outliers over all
evaluated sequences are given in Table 3 and are compared
to [23, 31, 34] using the results published by [31]. Our results can keep up with state-of-the-art scene flow methods,
although we have not tuned our method on MPI Sintel. For
sequences with close-up, non-rigid motion, e.g. ambush 7
or bandage 1, our depth estimation even beats the multiframe scene flow method that is ranked first on KITTI.
6. Conclusion
Our novel approach to interpolate sparse matches to a
dense scene flow achieves state-of-the-art performance on
different data sets. At time of submission SceneFlowFields
is ranked third on KITTI and achieves state-of-the-art performance on MPI Sintel. We have shown that a stochastic matching approach works for higher dimensional search
Sequence
PRSM
Disparity
OSF
FSF
Ours
PRSM
Optical Flow
OSF
FSF Ours
+ego
Average
15.99
19.84
15.35
18.15
13.70
28.16
18.32
29.24
22.20
alley 1
alley 2
ambush 2
ambush 4
ambush 5
ambush 6
ambush 7
bamboo 1
bamboo 2
bandage 1
bandage 2
cave 4
market 2
market 5
market 6
mountain 1
shaman 2
shaman 3
sleeping 2
temple 2
temple 3
7.43
0.79
41.77
24.09
17.72
29.41
35.07
7.34
17.06
21.22
22.44
4.27
5.27
15.38
8.99
0.42
25.49
33.92
1.74
4.92
11.04
5.28
1.31
55.13
24.05
19.54
26.18
71.58
9.71
18.08
19.37
23.53
5.86
6.61
13.67
10.29
0.78
28.27
52.22
2.97
5.54
16.62
5.92
2.08
36.93
23.30
18.54
30.33
23.47
9.67
19.27
20.93
22.69
6.22
6.81
13.25
10.63
0.23
24.77
27.09
3.52
5.96
10.65
8.81
1.73
51.72
37.78
25.52
37.13
16.34
14.53
19.89
16.42
21.77
6.20
6.71
26.66
14.53
0.15
21.13
35.37
3.07
6.98
8.61
1.58
1.08
51.33
41.99
25.23
41.98
3.35
2.41
3.58
3.30
4.06
16.32
4.77
28.38
10.72
3.71
0.46
1.75
0.00
9.51
32.10
7.33
1.44
87.37
49.16
44.70
54.75
22.47
4.04
4.86
18.40
13.12
33.94
10.08
29.58
16.39
88.60
1.67
11.45
0.01
10.52
81.39
2.11
1.20
72.68
45.23
24.82
44.05
27.87
4.11
3.65
4.00
4.76
14.62
5.17
26.31
13.13
17.05
0.56
1.31
0.02
9.66
62.34
5.94
2.85
90.92
60.03
46.92
57.06
13.66
6.11
5.84
3.82
10.72
15.63
7.11
40.77
28.92
90.60
8.85
15.91
0.61
29.58
72.28
3.95
0.87
83.84
42.65
29.86
47.65
7.35
4.15
3.97
4.03
9.06
12.95
6.09
28.87
16.69
89.57
4.31
8.51
0.03
12.57
49.18
Table 3. Results on MPI Sintel [5]. Average outliers show that
SceneFlowFields can keep up with state-of-the-art.
spaces and that the applied consistency filters produce very
robust correspondences. Boundary-aware interpolation has
turned out to be a powerful tool to fill the gaps in the scene
flow field due to filtering. To cope with missing correspondences across the images, we have applied an optional egomotion model that helps to overcome this issue. For future
work, we want to improve robustness of the ego-motion or
extend SceneFlowFields to use multiple frame pairs.
References
[1] C. Bailer, B. Taetz, and D. Stricker. Flow Fields: Dense
correspondence fields for highly accurate large displacement
optical flow estimation. In International Conference on
Computer Vision (ICCV), 2015. 3, 4
[2] T. Basha, Y. Moses, and N. Kiryati. Multi-view scene flow
estimation: A view centered variational approach. International Journal of Computer Vision (IJCV), 2013. 2
[3] T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping.
European Conference on Computer Vision (ECCV), 2004. 5
[4] T. Brox and J. Malik. Large displacement optical flow: Descriptor matching in variational motion estimation. Transactions on Pattern Analysis and Machine Intelligence (PAMI),
2011. 1
[5] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A
naturalistic open source movie for optical flow evaluation.
In European Conference on Computer Vision (ECCV), 2012.
2, 7, 8
[6] Q. Chen and V. Koltun. Full Flow: Optical flow estimation
by global optimization over regular grids. In Conference on
Computer Vision and Pattern Recognition (CVPR), 2016. 2
[7] M. Derome, A. Plyer, M. Sanfourche, and G. Le Besnerais.
A prediction-correction approach for real-time optical flow
computation using stereo. In German Conference on Pattern
Recognition (GCPR), 2016. 7
[8] P. Dollár and C. L. Zitnick. Structured forests for fast edge
detection. In International Conference on Computer Vision
(ICCV), 2013. 4, 5, 6, 7
[9] A. Gaidon, Q. Wang, Y. Cabon, and E. Vig. Virtual worlds
as proxy for multi-object tracking analysis. In Conference on
Computer Vision and Pattern Recognition (CVPR), 2016. 2
[10] A. Geiger, P. Lenz, and R. Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In
Conference on Computer Vision and Pattern Recognition
(CVPR), 2012. 1, 2
[11] K. He and J. Sun. Computing nearest-neighbor fields via
propagation-assisted kD-trees. In Conference on Computer
Vision and Pattern Recognition (CVPR), 2012. 3
[12] Y. Hel-Or and H. Hel-Or. Real-time pattern matching using projection kernels. Transactions on Pattern Analysis and
Machine Intelligence (PAMI), 2005. 3
[13] E. Herbst, X. Ren, and D. Fox. RGB-D Flow: Dense 3-D
motion estimation using color and depth. In International
Conference on Robotics and Automation (ICRA), 2013. 2, 3
[14] A. Hermans and G. Floros.
https://omnomnom.
vision.rwth-aachen.de/data/rwth_kitti_
semantics_dataset.zip. 4
[15] H. Hirschmuller. Stereo processing by semiglobal matching
and mutual information. Transactions on Pattern Analysis
and Machine Intelligence (PAMI), 2008. 1, 2, 4, 7
[16] M. Hornacek, A. Fitzgibbon, and C. Rother. SphereFlow: 6
DoF scene flow from RGB-D pairs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7
[17] F. Huguet and F. Devernay. A variational method for scene
flow estimation from stereo sequences. In International Conference on Computer Vision (ICCV), 2007. 2, 3, 5
[18] M. Jaimez, M. Souiai, J. Gonzalez-Jimenez, and D. Cremers.
A primal-dual framework for real-time dense RGB-D scene
flow. In International Conference on Robotics and Automation (ICRA), 2015. 2
[19] P. Lenz, J. Ziegler, A. Geiger, and M. Roser. Sparse scene
flow segmentation for moving object detection in urban environments. In Intelligent Vehicles Symposium (IV), 2011.
1
[20] C. Liu, J. Yuen, and A. Torralba. SIFT Flow: Dense correspondence across scenes and its applications. Transactions
on Pattern Analysis and Machine Intelligence (PAMI), 2011.
3
[21] Z. Lv, C. Beall, P. F. Alcantarilla, F. Li, Z. Kira, and F. Dellaert. A continuous optimization approach for efficient and
accurate scene flow. In European Conference on Computer
Vision (ECCV), 2016. 2, 3, 7
[22] N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers,
A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow
estimation. In Conference on Computer Vision and Pattern
Recognition (CVPR), 2016. 2
[23] M. Menze and A. Geiger. Object scene flow for autonomous
vehicles. In Conference on Computer Vision and Pattern
Recognition (CVPR), 2015. 2, 3, 6, 7, 8
[24] M. Neoral and J. Šochman. Object scene flow with temporal
consistency. In Computer Vision Winter Workshop (CVWW),
2017. 2, 3, 7
[25] C. Rabe, U. Franke, and S. Gehrig. Fast detection of moving
objects in complex scenarios. In Intelligent Vehicles Symposium (IV), 2007. 1
[26] J. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid.
EpicFlow: Edge-preserving interpolation of correspondences for optical flow. In Conference on Computer Vision
and Pattern Recognition (CVPR), 2015. 4, 5, 6
[27] G. Ros, S. Ramos, M. Granados, A. Bakhtiary, D. Vazquez,
and A. Lopez.
Vision-based offline-online perception
paradigm for autonomous driving. In Winter Conference on
Applications of Computer Vision (WACV), 2015. 4
[28] R. Schuster, C. Bailer, O. Wasenmüller, and D. Stricker.
Combining stereo disparity and optical flow for basic scene
flow. In Commercial Vehicle Technology Symposium (CVTS),
2018. 1
[29] R. Schuster, O. Wasenmüller, G. Kuschk, C. Bailer, and
D. Stricker. Towards flow estimation in automotive scenarios. In ACM Computer Science in Cars Symposium (CSCS),
2017. 1
[30] D. Sun, S. Roth, and M. J. Black. A quantitative analysis
of current practices in optical flow estimation and the principles behind them. International Journal of Computer Vision
(IJCV), 2014. 1
[31] T. Taniai, S. N. Sinha, and Y. Sato. Fast multi-frame stereo
scene flow with motion segmentation. In Conference on
Computer Vision and Pattern Recognition (CVPR), 2017. 1,
2, 3, 6, 7, 8
[32] S. Vedula, S. Baker, P. Rander, R. Collins, and T. Kanade.
Three-dimensional scene flow. In International Conference
on Computer Vision (ICCV), 1999. 2
[33] C. Vogel, K. Schindler, and S. Roth. Piecewise rigid scene
flow. In International Conference on Computer Vision
(ICCV), 2013. 2, 3, 7
[34] C. Vogel, K. Schindler, and S. Roth. 3D scene flow estimation with a piecewise rigid scene model. International
Journal of Computer Vision (IJCV), 2015. 2, 3, 7, 8
[35] A. S. Wannenwetsch, M. Keuper, and S. Roth. Probflow:
Joint optical flow and uncertainty estimation. In International Conference on Computer Vision (ICCV), 2017. 3
[36] O. Wasenmüller, B. Schenkenberger, and D. Stricker. Towards non-rigid reconstruction - How to adapt rigid RGB-D
reconstruction to non-rigid movements? In International
Conference on Computer Vision Theory and Applications
(VISAPP), 2017. 2
[37] A. Wedel, C. Rabe, T. Vaudrey, T. Brox, U. Franke, and
D. Cremers. Efficient dense scene flow from sparse or dense
stereo data. In European Conference on Computer Vision
(ECCV), 2008. 2, 5
[38] P. Xu, F. Davoine, J.-B. Bordes, H. Zhao, and T. Denœux.
Multimodal information fusion for urban scene understanding. Machine Vision and Applications (MVA), 2016. 4
[39] T. Yoshida, O. Wasenmüller, and D. Stricker. Time-of-flight
sensor depth enhancement for automotive exhaust gas. In
International Conference on Image Processing (ICIP), 2017.
2
| 1cs.CV
|
arXiv:1703.01913v1 [cs.DS] 6 Mar 2017
Near-Optimal Closeness Testing of Discrete Histogram
Distributions
Ilias Diakonikolas∗
University of Southern California
[email protected]
Daniel M. Kane†
University of California, San Diego
[email protected]
Vladimir Nikishkin‡
University of Edinburgh
[email protected]
March 7, 2017
Abstract
We investigate the problem of testing the equivalence between two discrete histograms. A
k-histogram over [n] is a probability distribution that is piecewise constant over some set of k
intervals over [n]. Histograms have been extensively studied in computer science and statistics.
Given a set of samples from two k-histogram distributions p, q over [n], we want to distinguish
(with high probability) between the cases that p = q and kp − qk1 ≥ ǫ. The main contribution
of this paper is a new algorithm for this testing problem and a nearly matching informationtheoretic lower bound. Specifically, the sample complexity of our algorithm matches our lower
bound up to a logarithmic factor, improving on previous work by polynomial factors in the
relevant parameters. Our algorithmic approach applies in a more general setting and yields
improved sample upper bounds for testing closeness of other structured distributions as well.
1
Introduction
In this work, we study the problem of testing equivalence (closeness) between two discrete structured
distributions. Let D be a family of univariate distributions over [n] (or Z). The problem of closeness
testing for D is the following: Given sample access to two unknown distribution p, q ∈ D, we want
to distinguish between the case that p = q versus kp − qk1 ≥ ǫ. (Here, kp − qk1 denotes the ℓ1 distance between the distributions p, q.) The sample complexity of this problem depends on the
underlying family D.
For example, if D is the class of all distributions over [n], then it is known [CDVV14] that
the optimal sample complexity is Θ(max{n2/3 /ǫ4/3 , n1/2 /ǫ2 }). This sample bound is best possible
only if the family D includes all possible distributions over [n], and we may be able to obtain
significantly better upper bounds for most natural settings. For example, if both p, q are promised
to be (approximately) log-concave over [n], there is an algorithm to test equivalence between them
using O(1/ǫ9/4 ) samples [DKN15a]. This sample bound is independent of the support size n, and
is dramatically better than the worst-case tight bound [CDVV14] when n is large.
∗
Supported by NSF Award CCF-1652862 (CAREER) and a Sloan Research Fellowship.
Supported by NSF Award CCF-1553288 (CAREER) and a Sloan Research Fellowship.
‡
Supported by a University of Edinburgh PCD Scholarship.
†
More generally, [DKN15a] described a framework to obtain sample-efficient equivalence testers
for various families of structured distributions over both continuous and discrete domains. While
the results of [DKN15a] are sample-optimal for some families of distributions (in particular, over
continuous domains), it was not known whether they can be improved for natural families of discrete
distributions. In this paper, we work in the framework of [DKN15a] and obtain new nearly-matching
algorithms and lower bounds.
Before we state our results in full generality, we describe in detail a concrete application of our
techniques to the case of histograms – a well-studied family of structured discrete distributions with
a plethora of applications.
Testing Closeness of Histograms. A k-histogram over [n] is a probability distribution p : [n] →
[0, 1] that is piecewise constant over some set of k intervals over [n]. The algorithmic difficulty in
testing properties of such distributions lies in the fact that the location and “size” of these intervals
is a priori unknown. Histograms have been extensively studied in statistics and computer science.
In the database community, histograms [JKM+ 98, CMN98, TGIK02, GGI+ 02, GKS06, ILR12,
ADH+ 15] constitute the most common tool for the succinct approximation of data. In statistics,
many methods have been proposed to estimate histogram distributions [Sco79, FD81, Sco92, LN96,
DL04, WN07, Kle09] in a variety of settings.
In recent years, histogram distributions have attracted renewed interested from the theoretical
computer science community in the context of learning [DDS12a, CDSS13, CDSS14a, CDSS14b,
DHS15, ADLS16, ADLS17, DKS16a] and testing [ILR12, DDS+ 13, DKN15b, Can16, CDGR16].
Here we study the following testing problem: Given sample access to two distributions p, q over [n]
that are promised to be (approximately) k-histograms, distinguish between the cases that p = q
versus kp − qk1 ≥ ǫ. As the main application of our techniques, we give a new testing algorithm
and a nearly-matching information-theoretic lower bound for this problem.
We now provide a summary of previous work on this problem followed by a description of our
new upper and lower bounds. We want to ǫ-test closeness in ℓ1 -distance between two k-histograms
over [n], where k ≤ n. Our goal is to understand the optimal sample complexity of this problem
as a function of k, n, 1/ǫ. Previous work is summarized as follows:
• In [DKN15a], the authors gave a closeness tester with sample complexity O(max{k 4/5 /ǫ6/5 , k1/2 /ǫ2 }).
• The best known sample lower bound is Ω(max{k2/3 /ǫ4/3 , k1/2 /ǫ2 }). This straightforwardly
follows from [CDVV14], since k-histograms can simulate any support k distribution.
Notably, none of the two bounds depends on the domain size n. Observe that the upper bound
of O(max{k4/5 /ǫ6/5 , k1/2 /ǫ2 }) cannot be tight for the entire range of parameters. For example, for
n = O(k), the algorithm of [CDVV14] for testing closeness between arbitrary support n distributions
has sample size O(max{k2/3 /ǫ4/3 , k1/2 /ǫ2 }), matching the above sample complexity lower bound,
up to a constant factor.
This simple example might suggest that the Ω(max{k2/3 /ǫ4/3 , k1/2 /ǫ2 }) lower bound is tight in
general. We prove that this is not the case. The main conceptual message of our new upper bound
and nearly-matching lower bound is the following:
The sample complexity of ǫ-testing closeness between two k-histograms over [n] depends
in a subtle way on the relation between the relevant parameters k, n and 1/ǫ.
We find this fact rather surprising because such a phenomenon does not occur for the sample
complexities of closely related problems. Specifically, testing the identity of a k-histogram over [n]
to a fixed distribution has sample complexity Θ(k1/2 /ǫ2 ) [DKN15b]; and learning a k-histogram
2
over [n] has sample complexity Θ(k/ǫ2 ) [CDSS14a]. Note that both these sample bounds are
independent of n and are known to be tight for the entire range of parameters k, n, 1/ǫ.
Our main positive result is a new closeness testing
algorithm for k-histograms over [n] with
4/3
2/3
4/3
sample complexity O k · log (2 + n/k) log(k)/ǫ
. Combined with the known upper bound of
[DKN15a], we obtain the sample upper bound of
O max min(k4/5 /ǫ6/5 , k2/3 log4/3 (2 + n/k) log(k)/ǫ4/3 ), k1/2 log2 (k) log log(k)/ǫ2 .
As our main negative result, we prove a lower bound of Ω(min(k2/3 log1/3 (2+ n/k)/ǫ4/3 , k4/5 /ǫ6/5 )).
The first term in this expression shows that the “log(2 + n/k)” factor that appears in the sample
complexity of our upper bound is in fact necessary, up to a constant power. In summary, these
bounds provide a nearly-tight characterization of the sample complexity of our histogram testing
problem for the entire range of parameters.
A few observations are in order to interpret the above bounds:
• When n goes to infinity, the O(k4/5 /ǫ6/5 ) upper bound of [DKN15a] is tight for k-histograms.
• When n = poly(k) and ǫ is not too small (so that the k1/2 /ǫ2 term does not kick in), then
the right answer for the sample complexity of our problem is (k2/3 /ǫ4/3 )polylog(k).
• The terms “k4/5 /ǫ6/5 ” and “k2/3 log4/3 (2 + n/k) log(k)/ǫ4/3 ” appearing in the sample complexity become equal when n is exponential in k. Therefore, our new algorithm has better
sample complexity than that of [DKN15a] for all n ≤ 2O(k) .
In the following subsection, we state our results in a general setting and explain how the
aforementioned applications are obtained from them.
1.1
Our Results and Comparison to Prior Work
For a given family D of discrete distributions over [n], we are interested in designing a closeness
tester for distributions in D. We work in the general framework introduced by [DKN15b, DKN15a].
Instead of designing a different tester for any given family D, the approach of [DKN15b, DKN15a]
proceeds by designing a generic equivalence tester under a different metric than the ℓ1 -distance.
This metric, termed Ak -distance [DL01], where k ≥ 2 is a positive integer, interpolates between
Kolmogorov distance (when k = 2) and the ℓ1 -distance (when k = n). It turns out that, for a range
of structured distribution families D, the Ak -distance can be used as a proxy for the ℓ1 -distance
for a value of k ≪ n [CDSS14a]. For example, if D is the family of k-histograms over [n], the
A2k distance between them is tantamount to their ℓ1 distance. We can thus obtain an ℓ1 closeness
tester for D by plugging in the right value of k in a general Ak closeness tester.
To formally state our results, we will need some terminology.
Notation. We will use p, q to denote the probability mass functions of our distributions. If p
is discrete over support [n] := {1, . . . , n}, we denote by pi the probability of element
Pn i in the
distribution. For two
discrete
distributions
p,
q,
their
ℓ
and
ℓ
distances
are
kp−qk
=
1
2
1
i=1 |pi −qi |
pPn
2
and kp − qk2 =
i=1 (pi − qi ) . Fix a partition of the domain I into disjoint intervals I :=
(Ii )ℓi=1 . For such a partition I, the reduced distribution pIr corresponding to p and I is the discrete
distribution over [ℓ] that assigns the i-th “point” the mass that p assigns to the interval Ii ; i.e., for
i ∈ [ℓ], pIr (i) = p(Ii ). Let Jk be the collection of all partitions of the domain I into k intervals.
def
For p, q : I → R+ and k ∈ Z+ , we define the Ak -distance between p and q by kp − qkAk =
P
maxI=(Ii )k ∈Jk ki=1 |p(Ii ) − q(Ii )| = maxI∈Jk kpIr − qrI k1 .
i=1
3
In this context, [DKN15a] gave a closeness testing algorithm under the Ak -distance using
O(max{k4/5 /ǫ6/5 , k1/2 /ǫ2 }) samples. It was also shown that this sample bound is information–
theoretically optimal (up to constant factors) for some adversarially constructed continuous distributions, or discrete distributions of support size n sufficiently large as a function of k. These
results raised two natural questions: (1) What is the optimal sample complexity of the Ak -closeness
testing problem as a function of n, k, 1/ǫ? (2) Can we obtain tight sample lower bounds for natural
families of structured distributions?
We resolve both these open questions. Our main algorithmic result is the following:
Theorem 1.1. Given sample access to distributions p and q on [n] and ǫ > 0 there exists an
algorithm that takes
O max min k4/5 /ǫ6/5 , k2/3 log4/3 (2 + n/k) log(2 + k)/ǫ4/3 , k1/2 log2 (k) log log(k)/ǫ2
samples from each of p and q and distinguishes with 2/3 probability between the cases that p = q
and kp − qkAk ≥ ǫ.
As explained in [DKN15b, DKN15a], using Theorem 1.1 one can obtain testing algorithms for
the ℓ1 closeness testing of various distribution families D, by using the Ak distance as a “proxy”
for the ℓ1 distance:
Fact 1.2. For a univariate distribution family D and ǫ > 0, let k = k(D, ǫ) be the smallest integer
such that for any f1 , f2 ∈ D it holds that kf1 − f2 k1 ≤ kf1 − f2 kAk + ǫ/2. Then there exists an ℓ1
closeness testing algorithm for D with the sample complexity of Theorem 1.1.
Applications. Our upper bound for ℓ1 -testing of k-histogram distributions follows from the above
by noting that for any k-histograms p, q we have kp − qk1 = kp − qkA2k . Also note that our upper
bound is robust: it applies even if p, q are O(ǫ)-close in ℓ1 -norm to being k-histograms.
Finally, we remark that our general Ak closeness tester yields improved upper bounds for various
other families of structured distributions. Consider for example the case that D consists of all kmixtures of some simple family (e.g., discrete Gaussians or log-concave), where the parameter k is
large. The algorithm of [DKN15a] leads to a tester whose sample complexity scales with O(k 4/5 ),
while Theorem 1.1 implies a Õ(k2/3 ) bound.
On the lower bound side, we show:
Theorem 1.3. Let p and q be distributions on [n] and let ǫ > 0 be less than a sufficiently small
constant. Any tester that distinguishes between p = q and kp − qkAk ≥ ǫ for some k ≤ n must use
Ω(m) samples for m = min(k2/3 log4/3 (2 + n/k)/ǫ4/3 , k4/5 /ǫ6/5 ).
Furthermore, for m = min(k2/3 log1/3 (2 + n/k)/ǫ4/3 , k4/5 /ǫ6/5 ), any tester that distinguishes
between p = q and kp − qkAk ≥ ǫ must use Ω(m) samples even if p and q are both guaranteed to be
piecewise constant distributions on O(k + m) pieces.
√
Note that a lower bound of Ω( k/ǫ2 ) straightforwardly applies even for p and q being khistograms. This dominates the above bounds for ǫ < k −3/8 .
We also note that our general lower bound with respect to the Ak distance is somewhat stronger,
matching the term “log4/3 (2 + n/k)” in our upper bound.
4
1.2
Related Work
During the past two decades, distribution property testing [BFR+ 00] – whose roots lie in statistical
hypothesis testing [NP33, LR05] – has received considerable attention by the computer science
community, see [Rub12, Can15] for two recent surveys. The majority of the early work in this field
has focused on characterizing the sample size needed to test properties of arbitrary distributions of
a given support size. After two decades of study, this “worst-case” regime is well-understood: for
many properties of interest there exist sample-optimal testers (matched by information-theoretic
lower bounds) [Pan08, CDVV14, VV14, DKN15b, DK16, DGPP16].
In many settings of interest, we know a priori that the underlying distributions have some “nice
structure” (exactly or approximately). The problem of learning a probability distribution under
such structural assumptions is a classical topic in statistics, see [BBBB72] for a classical book,
and [GJ14] for a recent book on the topic, that has recently attracted the interest of computer
scientists [DDS12a, DDS12b, CDSS13, DDO+ 13, CDSS14a, CDSS14b, ADH+ 15, DKS16d, DKS16e,
DKS16b, DDKT16, ADLS17, DKS16a, DKS16c].
On the other hand, the theory of distribution testing under structural assumptions is less fully
developed. More than a decade ago, Batu, Kumar, and Rubinfeld [BKR04] considered a specific
instantiation of this question – testing the equivalence between two unknown discrete monotone
distributions – and obtained a tester whose sample complexity is poly-logarithmic in the domain
size. A recent sequence of works [DDS+ 13, DKN15b, DKN15a] developed a framework to leverage
such structural assumptions and obtained more efficient testers for a number of natural settings.
However, for several natural properties of interest there is still a substantial gap between known
sample upper and lower bounds.
1.3
Overview of Techniques
To prove our upper bound, we use a technique of iteratively reducing the number of bins (domain
elements). In particular, we show that if we merge bins together in consecutive pairs, this does
not significantly affect the Ak distance between the distributions, unless a large fraction of the
discrepancy between our distributions is supported on O(k) bins near the boundaries in the optimal
partition. In order to take advantage of this, we provide a novel identity tester that requires few
samples to distinguish between the cases where p = q and the case where p and q have a large
ℓ1 distance supported on only k of the bins. We are able to take advantage of the small support
essentially because having a discrepancy supported on few bins implies that the ℓ2 distance between
the distributions must be reasonably large.
Our new lower bounds are somewhat more involved. We prove them by exhibiting explicit
families of pairs of distributions, where in one case p = q and in the other p and q have large Ak
distance, but so that it is information-theoretically impossible to distinguish between these two
families with a small number of samples. In both cases, p and q are explicit piecewise constant
distributions with a small number of pieces. In both cases, our domain is partitioned into a small
number of bins and the restrictions of the distributions to different bins are independent, making
our analysis easier. In some bins we will have p = q each with mass about 1/m (where m is the
number of samples). These bins will serve the purpose of adding “noise” making harder to read
the “signal” from the other bins. In the remaining bins, we will have either that p = q being
supported on some interval, or p and q will be supported on consecutive, non-overlapping intervals.
If three samples are obtained from any one of these intervals, the order of the samples and the
distributions that they come from will provide us with information about which family we came
from. Unfortunately, since triple collisions are relatively uncommon, this will not be useful unless
5
m ≫ max(k4/5 /ǫ6/5 , k1/2 /ǫ2 ). Bins from which we have one or zero samples will tell us nothing,
but bins from which we have exactly two samples may provide information.
For these bins, it can be seen that we learn nothing from the ordering of the samples, but we
may learn something from their spacing. In particular, in the case where p and q are supported on
disjoint intervals, we would suspect that two samples very close to each other are far more likely to
be taken from the same distribution rather than from opposite distributions. On the other hand,
in order to properly interpret this information, we will need to know something about the scale of
the distributions involved in order to know when two points should be considered to be “close”. To
overcome this difficulty, we will stretch each of our distributions by a random exponential amount.
This will effectively conceal any information about the scales involved so long as the total support
size of our distributions is exponentially large.
2
2.1
A Near-Optimal Closeness Tester over Discrete Domains
Warmup: A Simpler Algorithm
We start by giving a simpler algorithm establishing a basic version of Theorem 1.1 with slightly
worse parameters:
Proposition 2.1. Given sample access to distributions p and q on [n] and ǫ > 0 there exists an
algorithm that takes
√
O k2/3 log4/3 (3 + n/k) log log(3 + n/k)/ǫ4/3 + k log2 (3 + n/k) log log(3 + n/k)/ǫ2
samples from each of p and q and distinguishes with 2/3 probability between the cases that p = q
and kp − qkAk ≥ ǫ.
The basic idea of our algorithm is the following: From the distributions p and q construct new
distributions p′ and q ′ by merging pairs of consecutive buckets. Note that p′ and q ′ each have much
smaller
domains (of size about n/2). Furthermore, note that the Ak distance between p and q is
P
|p(I)
− q(I)| for some partition I into k intervals. By using essentially the same partition,
I∈I
we can show that kp′ − q ′ kAk should be almost as large as kp − qkAk . This will in fact hold unless
much of the error between p and q is supported at points near the endpoints of intervals in I. If
this is the case, it turns out there is an easy algorithm to detect this discrepancy. We require the
following definitions:
Definition 2.2. For a discrete distribution p on [n], the merged distribution obtained from p is
def
the distribution p′ on ⌈n/2⌉, so that p′ (i) = p(2i) + p(2i + 1). For a partition I of [n] , define the
divided partition I ′ of domain ⌈n/2⌉, so that Ii′ ∈ I ′ has the points obtained by point-wise gluing
together odd points and even points.
Note that one can simulate a sample from p′ given a sample from p by letting p′ = ⌈p/2⌉.
Definition 2.3. Let p and q be distributions on [n]. For integers k ≥ 1, let kp − qk1,k be the sum
of the largest k values of |p(i) − q(i)| over i ∈ [n].
We begin by showing that either kp′ − q ′ kAk is close to kp − qkAk or kp − qk1,k is large.
Lemma 2.4. For any two distributions p and q on [n], let p′ and q ′ be the merged distributions.
Then,
kp − qkAk ≤ kp′ − q ′ kAk + 2kp − qk1,k .
6
P
Proof. Let I be the partition of [n] into k intervals so that kp − qkAk = I∈I |p(I) − q(I)|. Let I ′
be obtained from I by rounding each upper endpoint of each interval except for the last down to
the nearest even integer, and rounding the lower endpoint of each interval up to the nearest odd
integer. Note that
X
X
|p(I) − q(I)| =
|p′ (I/2) − q ′ (I/2)| ≤ kp′ − q ′ kAk .
I∈I ′
I∈I ′
The partition I ′ is obtained from I by taking at most k points and moving them from one interval
to another. Therefore, the difference
X
I∈I
|p(I) − q(I)| −
X
I∈I ′
|p(I) − q(I)| ,
is at most twice the sum of |p(i) − q(i)| over these k points, and therefore at most 2kp − qk1,k .
Combing this with the above gives our result.
Next, we need to show that if two distributions have kp − qk1,k large that this can be detected
easily.
Lemma 2.5. Let p and q be distributions on [n]. √Let k > 0 be a positive integer, and ǫ > 0.
There exists an algorithm which takes O(k2/3 /ǫ4/3 + k/ǫ2 ) samples from each of p and q and, with
probability at least 2/3, distinguishes between the cases that p = q and kp − qk1,k > ǫ.
Note that if we needed to distinguish between p = q and kp − qk1 > ǫ, this would require
√
Ω(n2/3 /ǫ4/3 + n/ǫ2 ) samples. However, the optimal testers for this problem are morally ℓ2 √
testers. That is, roughly, they actually distinguish between p = q and kp − qk2 > ǫ/ n. From this
viewpoint, it is clear why it would
√ be easier to test for discrepancies in k − k1,k -distance, since if
kp − qk1,k > ǫ, then kp − qk2 > ǫ/ k, making it easier for our ℓ2 -type tester to detect the difference.
Our general approach will be by way of the techniques developed in [DK16]. We begin by giving
the definition of a split distribution coming from that paper:
Definition 2.6. Given a distribution p on [n] and a multiset S of elements of [n], define the split
distribution pS on [n + |S|] asPfollows: For 1 ≤ i ≤ n, let ai denote 1 plus the number of elements of
S that are equal to i. Thus, ni=1 ai = n + |S|. We can therefore associate the elements of [n + |S|]
to elements of the set B = {(i, j) : i ∈ [n], 1 ≤ j ≤ ai }. We now define a distribution pS with
support B, by letting a random sample from pS be given by (i, j), where i is drawn randomly from
p and j is drawn randomly from [ai ].
We now recall two basic facts about split distributions:
Fact 2.7 ([DK16]). Let p and q be probability distributions on [n], and S a given multiset of
[n]. Then: (i) We can simulate a sample from pS or qS by taking a single sample from p or q,
respectively. (ii) It holds kpS − qS k1 = kp − qk1 .
Lemma 2.8 ([DK16]). Let p be a distribution on [n]. Then: (i) For any multisets S ⊆ S ′ of [n],
kpS ′ k2 ≤ kpS k2 , and (ii) If S is obtained by taking m samples from p, then E[kpS k22 ] ≤ 1/m.
We also recall an optimal ℓ2 closeness tester under the promise that one of the distributions
has smal ℓ2 norm:
7
Lemma 2.9 ([CDVV14]). Let p and q be two unknown distributions
on [n]. There exists an
√
algorithm that on input n, b ≥ min{kpk2 , kqk2 } and 0 < ǫ < 2b, draws O(b/ǫ2 ) samples from
each of p and q and, with probability at least 2/3, distinguishes between the cases that p = q and
kp − qk2 > ǫ.
Proof of Lemma 2.5: We begin by presenting the algorithm:
Algorithm Small-Support-Discrepancy-Tester
Input: sample access to pdf’s p, q : [n] → [0, 1], k ∈ Z+ , and ǫ > 0.
Output: “YES” if q = p; “NO” if kq − pk1,k ≥ ǫ.
1. Let m = min(k2/3 /ǫ4/3 , k).
2. Let S be the multiset obtained by taking m independent samples from p.
3. Use the ℓ2 tester of Lemma 2.9 to distinguish between the cases that pS = qS and
kpS − qS k22 ≥ k−1 ǫ2 /2 and return the result.
The analysis is simple. By Lemma 2.8, with 90% probability kpS k2 = O(m−1/2 ), and therefore
the number of√ samples needed (using the ℓ2 tester from Lemma 2.9) is O(m + km−1/2 /ǫ2 ) =
O(k2/3 /ǫ4/3 + k/ǫ2 ). If p = q, then pS = qS and the algorithm will return “YES” with appropriate
probability. If kq − pk1,k ≥ ǫ, then kpS − qS k1,k+m ≥ ǫ. Since k + m elements contribute to total ℓ1
error at least ǫ, by Cauchy-Schwarz, we have that kpS − qS k22 ≥ ǫ2 /(k + m) ≥ k−1 ǫ2 /2. Therefore,
in this case, the algorithm returns “NO” with appropriate probability.
Proof of Proposition 2.1: The basic idea of our algorithm is the following: By Lemma 2.5, if kp −
qkAk is large, then so is either kp − qk1,k or kp′ − q ′ kAk . Our algorithm then tests whether kp − qk1,k
is large, and recursively tests whether kp′ − q ′ kAk is large. Since p′ , q ′ have half the support size, we
will only need to do this for log(n/k) rounds, losing only a poly-logarithmic factor in the sample
complexity. We present the algorithm here:
Algorithm Small-Domain-Ak-tester
Input: sample access to pdf’s p, q : [n] → [0, 1], k ∈ Z+ , and ǫ > 0.
Output: “YES” if q = p; “NO” if kq − pkAk ≥ ǫ.
def
1. For i := 0 to t = ⌈log2 (n/k)⌉, let p(i) , q (i) be distributions on [⌈2−i n⌉] defined by
p(i) = ⌈2−i p⌉ and q (i) = ⌈2−i q⌉.
2. Take Ck2/3 log4/3 (3 + n/k) log log(3 + n/k)/ǫ4/3 samples, for C sufficiently large, and
use these samples to distinguish between the cases p(i) = q (i) and kp(i) − q (i) k1,k >
ǫ/(4 log 2 (3 + n/k)) with probability of error at most 1/(10 log 2 (3 + n/k)) for each i
from 0 to t, using the same samples for each test.
3. If any test yields that p(i) 6= q (i) , return “NO”. Otherwise, return “YES”.
We now show correctness. In terms of sample complexity, we note that by taking a majority over
O(log log(3 + n/k)) independent runs of the tester from Lemma 2.5 we can run this algorithm with
the stated sample complexity. Taking a union bound, we can also assume that all tests performed
in Step 2 returned the correct answer. If p = q then p(i) = q (i) for all i and thus, our algorithm
8
returns “YES”. Otherwise, we have that kp − qkAk ≥ ǫ. By repeated application of Lemma 2.4, we
have that
kp − qkAk ≤
t−1
X
i=0
2kp
(i)
(i)
− q k1,k + kp
(t)
(t)
− q kAk ≤ 2
t
X
i=0
kp(i) − q (i) k1,k ,
where the last step was because p(t) and q (t) have a support of size at most k and so kp(t) −
q (t) kAk = kp(t) − q (t) k1 = kp(t) − q (t) k1,k . Therefore, if this is at least ǫ, it must be the case that
kp(i) − q (i) k1,k > ǫ/(4 log 2 (3 + n/k)) for some 0 ≤ i ≤ t, and thus our algorithm returns “NO”. This
completes our proof.
2.2
Full Algorithm
The improvement to Proposition 2.1 is somewhat technical. The key idea involves looking into the
analysis of Lemma 2.5. Generally speaking, choosing a larger value of m (up to the total sample
complexity), will decrease the ℓ2 norm of p, and thus the final complexity. Unfortunately, taking
m > k might lead to problems as it will subdivide the k original bins on which the error is supported
into ω(k) bins. This in turn could worsen the lower bounds on kp − qk2 . However, this will only
be the case if the total mass of these bins carrying the difference is large. Thus, we can obtain an
improvement to Lemma 2.5 when the mass of bins on which the error is supported is small. This
motivates the following definition:
Definition 2.10. For probability distributions p, q, an integer k and real number
P α > 0, dk,α (p, q)
is the maximum over sets T of size at most k so that p(i) ≤ α for all i ∈ T of i∈T |p(i) − q(i)|.
In other words, dk,α (p, q) is the biggest ℓ1 difference between p and q coming from at most k
bins of mass at most α. We have the following lemma:
Lemma 2.11. Let p and q be distributions on [n]. Let k > 0 be a positive integer, and ǫ, α > 0.
There exists an algorithm which takes O(k2/3 /ǫ4/3 (1 + mα)) samples from each of p and q and,
with probability at least 2/3, distinguishes between the cases that p = q and dk,α (p, q) > ǫ.
Proof. The algorithm and its analysis are nearly identical to that of Lemma 2.5. We include them
here for completeness:
Algorithm Small-Support-Discrepancy-Tester
Input: sample access to pdf’s p, q : [n] → [0, 1], k ∈ Z+ , and ǫ > 0 with kpk2 ≤ α.
Output: “YES” if q = p; “NO” if kq − pk1,k ≥ ǫ.
1. Let m = k2/3 /ǫ4/3 .
2. Let S be the multiset obtained by taking m independent samples from p.
3. Use the ℓ2 tester of Lemma
√ 2.9 to distinguish between the cases pS = qS and kpS −
2
−1
2
qS k2 ≥ k ǫ /(1 + O(αm/ k)) and return the result.
The analysis is quite simple. Firstly, we can assume that kpS k22 = O(1/m) as this happens
with 90% probability
over the choice of S. Next, let T be the set of size at most k such that
P
dk,α (p, q) = i∈T |p(i)−q(i)|. With 90% probability over the choice of S, we have that only O(mkα)
elements from S land in T . Assuming this is the case, it is sufficient to distinguish between
pS = q S
√
√
and kpS − qS k22 ≥ k−1 ǫ2 /(1 + O(αm)), which can be done in O(kǫ−2 (1 + O(αm/ k))/ m) =
O(k2/3 ǫ−4/3 (1 + O(αm))) samples. This completes the proof.
9
We are now prepared to prove Theorem 1.1. The basic idea behind the improvement is that we
want to avoid merging heavy bins. We do this by first taking a large set of elements and defining
the p(i) in a way that doesn’t involve merging elements of these sets.
Proof. We first note that given the algorithm from [DKN15a], it suffices to provide an algorithm
when ǫ > k −3/8 and n ≤ 2k .
Our algorithm is the following:
Algorithm Small-Domain-Ak-tester
Input: sample access to pdf’s p, q : [n] → [0, 1], k ∈ Z+ , and ǫ > 0.
Output: “YES” if q = p; “NO” if kq − pkAk ≥ ǫ.
1. Let m = k2/3 log4/3 (3 + n/k)/ǫ4/3 . Let C be a sufficiently large constant.
2. Let S be a set of Cm log(k) independent samples from p.
def
3. Let p(0) = p and q (0) = q. For i := 1 to t = ⌈log2 (n/k)⌉, define distributions p(i) , q (i)
inductively as follows:
(a) p(i) will be a flattening of p by merging all bins in certain dyadic intervals (i.e.,
intervals of the form [a · 2b + 1, (a + 1)2b ]).
(b) p(i+1) is obtained from p(i) by merging any pair of adjacent bins of p(i) that
correspond to intervals [a2i+1 + 1, a2i+1 + 2i ] and [a2i+1 + 2i + 1, (a + 1)2i ] where
neither of these subintervals contains a point of S.
(c) q (i+1) is obtained by merging bins in a similar way.
4. Take Cm log log(3 + n/k) samples, and use these samples to distinguish between the
cases p(i) = q (i) and dk,1/m (p(i) − q (i) ) > ǫ/(8 log 2 (3 + n/k)) with probability of error
at most 1/(10 log 2 (3 + n/k)) for each i from 0 to t, using the same samples for each
test.
5. If any test yields that p(i) 6= q (i) , return “NO”.
6. Otherwise, test if p(t) = q (t) of kp(t) − q (t) kAk > ǫ/2 using the algorithm from Proposition 2.1 and return the answer.
We now proceed with the analysis. Firstly, we note that the bins of p(t) corresponds to a dyadic
interval either containing an element of S or adjacent to such an element. Therefore, the domain
of p(t) is at most O(t|S|) = poly(k).
We also note that the sample complexity of
O(m log(k)) + O(k2/3 log4/3 (3 + n/k) log log(3 + n/k)/ǫ4/3 )
√
+ O((k2/3 log4/3 (k) log log(k)/ǫ4/3 + k log2 (k) log log(k)/ǫ2 )),
which is sufficient.
We now proceed to prove correctness. For completeness, if p = q, it is easy to see that p(i) = q (i)
for all i, and thus, by a union bound, we pass every test and our algorithm returns “YES” with
2/3 probability.
It remains to consider the soundness case, i.e., the case where kp − qkAk > ǫ. In this case, let
P
I = {Ii }1≤i≤k be a partition of [n] into intervals so that ki=1 |p(Ii ) − q(Ii )| > ǫ. We claim that
10
with high probability over the choice of S every dyadic interval that has mass (under p) at least
1/m and contains an endpoint of some Ii also contains an element of S. To prove this, we note that
the Ii contain only O(k) endpoints, and each endpoint is contained in a unique minimal dyadic
interval of mass at least 1/m. It suffices to show that each of these O(k) intervals of mass at least
1/m contains a point in S, but this follows easily by a union bound. Henceforth, we will assume
that the S we chose has this property.
Let I (i) be a partition of the bins for p(i) and q (i) defined inductively by I (0) = I and I (i+1) is
obtained from I (i) by flattening it and assigning new bins that partially overlap two of the intervals
in I (i) arbitrarily to one of the two corresponding intervals in I (i+1) .
We note that
X
X
|p(i) (I) − q (i) (I)| −
|p(i+1) (I) − q (i+1) (I)|
I∈I (i)
I∈I (i+1)
is at most twice a sum over k bins b, not containing an element of S of |p(i) (b) − q (i) (b)|. This in
turn is at most 2dk,1/m (p(i) , q (i) ). Inducting, we have that
kp − qkAk ≤ 2
t−1
X
i=0
dk,1/m (p(i) , q (i) ) + kp(t) − q (t) kAk .
Therefore, if kp − qkAk > ǫ, we have that either dk,1/m (p(i) − q (i) ) > ǫ/(8 log2 (3 + n/k)) for some i,
or kp(t) − q (t) kAk > ǫ/2. In either case, with probability at least 2/3, our algorithm will detect this
and reject. This completes the proof.
3
Nearly Matching Information-Theoretic Lower Bound
In this section, we prove a nearly matching sample lower bound. We first show a slightly easier
lower bound that holds even for distributions that are piecewise constant on a few pieces, and then
modify it to obtain the stronger general bound for testing closeness in Ak distance.
3.1
Lower Bound for k-Histograms
We begin with a lower bound for k-histograms (k-flat distributions). Before moving to the discrete
setting, we first establish a lower bound for continuous histogram distributions. Our bound on
discrete distributions will follow from taking the adversarial distribution from this example and
rounding its values to the nearest integer. In order for this to work, we will need ensure to that our
adversarial distribution does not have its Ak -distance decrease by too much when we apply this
operation. To satisfy this requirement, we will guarantee that our distributions will be piecewise
constant with all the pieces of length at least 1.
Proposition 3.1. Let k ∈ Z+ , ǫ > 0 sufficiently small, and W > 2 . Fix
m = min(k2/3 log1/3 (W )/ǫ4/3 , k4/5 /ǫ6/5 ) .
There exist distributions D, D ′ over pairs of distributions p and q on [0, 2(m + k)W ], where p and
q are O(m + k)-flat with pieces of length at least 1, so that: (a) when drawn from D, we have p = q
deterministically, (b) when drawn from D ′ , we have kp − qkAk > ǫ with 90% probability, and so that
o(m) samples are insufficient to distinguish whether or not the pair is drawn from D or D ′ with
better than 2/3 probability.
11
At a high-level, our lower bound construction proceeds as follows: We will divide our domain
into m+k bins so that no information about which distributions had samples drawn from a given bin
or the ordering of these samples will help to distinguish between the cases of p = q and otherwise,
unless at least three samples are taken from the bin in question. Approximately k of these bins
will each have mass ǫ/k and might convey this information if at least three samples are taken from
the bin. However, the other m bins will each have mass approximately 1/m and will be used to
add noise. In all, if we take s samples, we expect to see approximately s3 ǫ3 /k2 of the lighter bins
with at least three samples. However, we will see approximately s3 /m2 of our heavy bins with
three samples. In order for the signal to overwhelm the noise, we will need to ensure that we have
(s3 ǫ3 /k2 )2 > s3 /m2 .
The above intuitive sketch assumes that we cannot obtain information from the bins in which
only two samples are drawn. This naively should not be the case. If p = q, the distance between
two samples drawn from that bin will be independent of whether or not they are drawn from the
same distribution. However, if p and q are supported on disjoint intervals, one would expect that
points that are close to each other should be far more likely to be drawn from the same distribution
than from different distributions. In order to disguise this, we will scale the length of the intervals
by a random, exponential amount, essentially making it impossible to determine what is meant by
two points being close to each other. In effect, this will imply that two points drawn from the same
bin will only reveal O(1/ log(W )) bits of information about whether p = q or not. Thus, in order
for this information to be sufficient, we will need that (s2 ǫ2 /k)2 / log(W ) > (s2 /m). We proceed
with the formal proof below.
Proof of Proposition 3.1: We use ideas from [DK16] to obtain this lower bound using an information
theoretic argument.
We √
may assume that ǫ > k1/2 , because otherwise we may employ the standard lower bound
that Ω( k/ǫ2 ) samples are required to distinguish two distributions on a support of size k.
First, we note that it is sufficient to take D and D ′ be distributions over pairs of non-negative,
piecewise constant distributions with total mass Θ(1) with 90% probability so that running a
Poisson process with parameter o(m) is insufficient to distinguish a pair from D from a pair from
D ′ [DK16].
We construct these distributions as follows: We divide the domain into m+k bins of length 2W .
For each bin i, we independently generate a random ℓi , so that log(ℓi /2) is uniformly distributed
over [0, 2 log(W )/3]. We then produce an interval Ii within bin i of total length ℓi and with random
offset. In all cases, we will have p and q supported on the union of the Ii ’s.
For each i with probability m/(m + k), we have the restrictions of p and q to Ii both uniform
with p(Ii ) = q(Ii ) = 1/m. The other k/(m + k) of the time we have p(Ii ) = q(Ii ) = ǫ/k. In this
latter case, if p and q are being drawn from D, p and q are each constant on this interval. If they
are being drawn from D ′ , then p + q will be constant on the interval, with all of that mass coming
from p on a random half and coming from q on the other half.
Note that in all cases p and q are piecewise constant with O(m + k) pieces of length at least 1.
It is easy to show that with high probability the total mass of each of p and q is Θ(1), and that if
drawn from D ′ that kp − qkAk ≫ ǫ with at least 90% probability.
We will now show that if one is given m samples from each of p and q, taken randomly from
either D or D ′ , that the shared information between the samples and the source family will be
small. This implies that one is unable to consistently guess whether our pair was taken from D or
D′ .
Let X be a random variable that is uniformly at random either 0 or 1. Let A be obtained by
applying a Poisson process with parameter s = o(m) on the pair of distributions p, q drawn from
12
D if X = 0 or from D ′ if X = 1. We note that it suffices to show that the shared information
I(X : A) = o(1). In particular, by Fano’s inequality, we have:
Lemma 3.2. If X is a uniform random bit and A is a correlated random variable, then if f is any
function so that f (A) = X with at least 51% probability, then I(X : A) ≥ 2 · 10−4 .
Let Ai be the samples of A taken from the ith bin.
P Note that the Ai are conditionally independent on X. Therefore, we have that I(X : A) ≤ i I(X : Ai ) = (m + k)I(X : A1 ) . We will
proceed to bound I(X : A1 ).
We note that I(X : A1 ) is at most the integral over pairs of multisets a (representing a set of
samples from q and a set of samples from p), of
(Pr(A1 = a|X = 0) − Pr(A1 = a|X = 1))2
O
.
Pr(A1 = a)
Thus,
I(X : A1 ) =
∞ Z
X
h=0 |a|=h
O
(Pr(A1 = a|X = 0) − Pr(A1 = a|X = 1))2
Pr(A1 = a)
.
We will split this sum up based on the value of h.
For h = 0, we note that the distributions for p+q are the same for X = 0 and X = 1. Therefore,
the probability of selecting no samples is the same. Therefore, this contributes 0 to the sum.
For h = 1, we note that the distributions for p + q are the same in both cases, and conditioning
on I1 and (p + q)(I1 ) that E[p] and E[q] are the same in each of the cases X = 0 and X = 1.
Therefore, again in this case, we have no contribution.
For h ≥ 3, we note that I(X : A1 ) ≤ I(X : A1 , I1 ) ≤ I(X : A1 |I1 ) , since I1 is independent of
X. We note that Pr(A1 = a|X = 0, p(I1 ) = 1/m) = Pr(A1 = a|X = 1, p(I1 ) = 1/m). Therefore,
we have that
Pr(A1 = a|X = 0)−Pr(A1 = a|X = 1) = Pr(A1 = a|X = 0, p(I1 ) = ǫ/k)−Pr(A1 = a|X = 1, p(I1 ) = ǫ/k).
If p(I1 ) = ǫ/k, the probability that exactly h elements are selected in this bin is at most k/(m +
k)(2sǫ/k)h /h!, and if they are selected, they are uniformly distributed in I1 (although which of the
sets p and q they are taken from is non-uniform). However, the probability that h elements are
taken from I1 is at least Ω(m/(m + k)(sm)−h /h!) from the case where p(I1 ) = 1/m, and in this
case the elements are uniformly distributed in I1 and uniformly from each of p and q. Therefore, we
have that this contribution to our shared information is at most k2 /(m(m + k))O(sǫ2 m/k2 )h /h! .
We note that ǫ2 m/k2 < 1. Therefore, the sum of this over all h ≥ 3 is k2 /(m(m + k))O(sǫ2 m/k2 )3 .
Summing over all m + k bins, this is k−4 ǫ6 s3 m2 = o(1).
It remains to analyze the case where h = 2. Once again, we have that ignoring which of p and
q elements of A1 came from, A1 is identically distributed conditioned on p(I1 ) = 1/m and |A1 | = 2
as it is conditioned on p(I1 ) = ǫ/k and |A1 | = 2. Since once again, the distributions D and D ′ are
indistinguishable in the former case, we have that the contribution of the h = 2 terms to the shared
information is at most
(k/(k + m)(ǫs/k)2 )2
dTV ((A1 |X = 0, p(I1 )ǫ/k, |A1 | = 2), (A1 |X = 1, p(I1 ) = ǫ/k, |A1 | = 2))
O
m/(k + m)(s/m)2
or
O s2 mk−2 ǫ4 /(k + m) dTV ((A1 |X = 0, p(I1 ) = ǫ/k, |A1 | = 2), (A1 |X = 1, p(I1 ) = ǫ/k, |A1 | = 2)) .
13
It will suffice to show that conditioned upon p(I1 ) = ǫ/k and |A1 | = 2 that
dTV ((A1 |X = 0), (A1 |X = 1)) = O(1/ log(W )).
Let f be the order preserving linear function from [0, 2] to I1 . Notice that conditional on |A1 | = 2
and p(I1 ) = ǫ/k that we may sample from A1 as follows:
• Pick two points x > y uniformly at random from [0, 2].
• Assign the points to p and q as follows:
– If X = 0 uniformly randomly assign these points to either distribution p or q.
– If X = 1 randomly do either:
∗ Assign points in [0, 1] to q and other points to p.
∗ Assign points in [0, 1] to p and other points to q.
• Randomly pick I1 and apply f to x and y to get outputs z = f (x), w = f (y).
Notice that the four cases: (i) both points coming from p, (ii) both points coming from q, (iii) a
point from p preceding a point from q, (iv) a point from q preceding a point from p, are all equally
likely conditioned on either X = 0 or X = 1. However, we will note that this ordering is no longer
independent of the choice of x and y.
Therefore, we can sample from A1 subject to X = 0 and from A1 subject to X = 1 in such
a way that this ordering is the same deterministically. We consider running the above sampling
algorithm to select (x, y) while sampling from X = 0 and (x′ , y ′ ) when sampling from X = 1 so
that we are in the same one of the above four cases. We note that
dTV ((A1 |X = 0), (A1 |X = 1)) ≤ Ex,y,x′ ,y′ [dTV ((f (x), f (y)), (f (x′ ), f (y ′ )))] ,
where the variation distance is over the random choices of f .
To show that this is small, we note that |f (x) − f (y)| is distributed like ℓ1 (x − y). This means
that log(|f (x) − f (y)|) is uniform over [log(f (x) − f (y)), log(f (x) − f (y)) + 2 log(W )/3]. Similarly,
log(|f ′ (x′ )− f ′ (y ′ )|) is uniform over [log(f (x′ )− f (y ′ )), log(f (x′ )− f (y ′ ))+ 2 log(W )/3]. These differ
in total variation distance by
| log(f (x) − f (y))| + | log(f (x′ ) − f (y ′ ))|
O
.
log(W )
Taking the expectation over x, y, x′ , y ′ we get O(1/ log(W )). Therefore, we may further correlate
the choices made in selecting our two samples, so that z − w = z ′ − w′ except with probability
O(1/ log(W )). We note that after conditioning on this, z and z ′ are both uniformly distributed
over subintervals of [0, 2W ] of length at least 2(W − W 2/3 ). Therefore, the distributions on z
and z ′ differ by at most O(W −1/3 ). Hence, the total variation distance between A1 conditioned
on |A1 | = 2, p(I1 ) = ǫ/k, X = 0 and conditioned on |A1 | = 2, p(I1 ) = ǫ/k, X = 1 is at most
O(1/ log(W )) + O(W −1/3 ) = O(1/ log(W )). This completes our proof.
We can now turn this into a lower bound for testing Ak distance on discrete domains.
Proof of second half of Theorem 1.3: Assume for sake of contradiction that this is not the case, and
that there exists a tester taking o(m) samples. We use this tester to come up with a continuous
tester that violates Proposition 3.1.
14
We begin by proving a few technical bounds on the parameters involved. Firstly, note that we
already have a lower bound of Ω(k1/2 /ǫ2 ), so we may assume that this is much less than m. We
now claim that m = O(min(k2/3 log1/3 (3 + n/(m + k))/ǫ4/3 , k4/5 /ǫ6/5 ). If m ≤ k, there is nothing
to prove. Otherwise,
k2/3 log1/3 (3 + n/(m + k))/ǫ4/3 ≥ m(m/k)−1/3 log(3 + n/(m + k))1/3 .
Thus, there is nothing more to prove unless log(3 + n/(m + k)) ≫ m/k. But, in this case, log(3 +
n/(m + k)) ≫ log(m/k) and thus log(3 + n/(m + k)) = Θ(log(3 + n/k)), and we are done.
We now let W = n/(6(m + k)), and let D and D ′ be as specified in Proposition 3.1. We claim
that we have a tester to distinguish a p, q from D from ones taken from D ′ in o(m) samples. We
do this as follows: By rounding p and q down to the nearest third of an integer, we obtain p′ ,q ′
supported on set of size n. Since p and q were piecewise constant on pieces of size at least 1, it
is not hard to see that kp′ − q ′ kAk ≥ kp − qkAk /3. Therefore, a tester to distinguish p′ = q ′ from
kp′ − q ′ kAk ≥ ǫ can be used to distinguish p = q from kp − qkAk ≥ 3ǫ. This is a contradiction and
proves our lower bound.
3.2
The Stronger Lower Bound
In order to improve on the bound from the last section, we will need to modify our previous
construction in two ways both having to do with the contribution to the shared information coming
from the case where two samples are taken from the same bin. The first is that we will need
a different way of distinguishing between D and D ′ so that the variation distance between the
distributions obtained from taking a pair of samples from the same bin is O(1/ log 2 (W )) rather
than O(1/ log(W )). After that, we will also need a better method of disguising these errors. In
particular, in the current construction, most of the information coming from pairs of samples from
the same bin occurs when the two samples are very close to each other (as when this happens in
D ′ , the samples usually don’t come one from p and the other from q). This is poorly disguised by
noise coming from the heavier bins since these are not particularly likely to produce samples that
are close. We can improve our way of disguising this by having different heavy bins to better mask
this signal.
In order to solve the first of these problems, we will need the following construction:
Lemma 3.3. Let W be a sufficiently large integer. There exists a family E of pairs of distributions
p and q on [0, W ] so that the following holds:
Firstly, p and q are deterministically supported on disjoint intervals, and thus have A1 distance
2. Furthermore, let E0 be the family of pairs of distributions p and q on [W ] obtained by taking
(p′ , q ′ ) from E and letting p = q = (p′ + q ′ )/2. In other words, a sample from E0 can be thought of
as taking a sample from E and then re-randomizing the label. Consider the distribution obtained
by sampling (p, q) from E, and then taking two independent samples x and y from (p + q)/2. We
let E 2 be the induced distribution on x and y along with the labels of which of p and q each were
taken from. Define E02 similarly, and note that it is equivalent to taking a sample from E 2 and
re-randomizing the labels. Then dTV (E 2 , E02 ) = O(1/ log 2 (W )).
Proof. We note that it is enough to construct a family of continuous distributions p and q on [0, W ]
so that deterministically p and q are supported on intervals separated by distance 2, and so that
the second condition above holds. By then rounding the values of p and q to the nearest integer,
we obtain an appropriate discrete distribution.
The construction of E is straightforward. First, choose a uniformly from [W 2/3 , W − W 2/3 ], ℓ
uniformly from [0, log(W )/3], and b uniformly from {±1}. To sample from p, take an α uniformly
15
from [0, log(W )/3] and return a + beℓ eα . To sample from q, take an α uniformly from [0, log(W )/3]
and return a − beℓ eα .
It is clear that p and q are supported on disjoint intervals of distance at least 2. It remains to
prove the more complicated claim.
Let Es2 be the distribution obtained by picking a pair of distributions from E and then returning two independent samples from p. Let Ed2 be the distribution obtained by picking a pair
of distributions from E and then returning independent samples from p and q. We claim that
dTV (E 2 , E02 ) = O(dTV (Es2 , Ed2 )). This is because if a sample from E 2 has both points coming from p
or both from q, the points come from Es2 , whereas if one point comes from each, the points come
from Ed2 . On the other hand, in any of these cases, a pair of samples from E02 comes from (Es2 +Ed2 )/2.
Let (x, y) be a sample from Es2 and (w, z) a sample from Ed2 . We claim that dTV ((x, y), (w, z)) ≤
dTV (x − y, w − z) + O(W −1/3 ). This is because of the averaging over a in the definition of E. In
particular, consider the following mechanism for taking a sample from Es2 or Ed2 . First, randomly
select values of s and ℓ. Then select the α and α′ for the two sample points. Finally, sample the
defining value of a. Notice that the difference between the two final points does not depend on
the choice of a. In fact, after making all other choices, the final distribution is within O(W −1/3 )
of the uniform distribution over pairs of points in [0, W ] with this distance. Thus, (x, y) is close
distributionally to the distribution on pairs in [0, W ] with separation x−y. A similar statement holds
for (z, w) and points with separation z−w. Thus, dTV ((x, y), (w, z)) = dTV (x−y, w−z)+O(W −1/3 ),
as desired.
Next, we claim that dTV (x − y, z − w) = dTV (|x − y|, |z − w|). This is easily seen to be the case
by averaging over b. We have left to bound the latter distance. If x and y are chosen using αx and
αy , we have that |x − y| = eℓ |eαx − eαy |. Similarly, if z and w are chosen using αz and αw , we have
that |z − w| = eℓ |eαz + eαw |. Notice that if we fix αx , αy , αz and αw , the variation distance between
these two distributions (given the distributions over the values of ℓ) is
αx αy
|e −e |
log |e
αz +eαw |
.
O
log(W )
Therefore, the variation distance between |x − y| and |z − w| is O(1/ log(W )) times the earth
mover distance between log(|eαx − eαy |) and log(|eαz + eαw |). Correlating these variables so that
αx = αz = α and αy = αw = β, this is at most the expectation of | log(tanh((α − β)/2))|, which
can easily be seen to be O(1/ log(W )). This shows that dTV (E 2 , E02 ) = O(1/ log 2 (W )), completing
our proof.
We are now ready to prove the first part of Theorem 1.3.
Proof. The overall outline is very similar to the methods used in the last section. For sufficiently
large integers m, k, W and ǫ > 0 we are going to define families of pairs of pseudo-distributions D
and D ′ on [(k + 2m)W ] so that:
• With 90% probability a random sample from either D or D ′ consists of two pseudo-distributions
with total mass Θ(1).
• The distributions picked by a sample from D are always the same.
• The two distributions picked by a sample from D ′ have Ak distance Ω(ǫ) with 90% probability.
16
• Letting A be the outcome of a Poisson process with parameter m run on a random sample from
either D or D ′ , the family used cannot be reliably determined from A unless m ≫ k4/5 /ǫ6/5
or m ≫ k2/3 log4/3 (W )/ǫ4/3 .
Before we define D and D ′ , we will need to define one more family. Firstly, let E and E0 be the
families of distributions on [W ] from Lemma 3.3. Let E 2 and E02 be as described in that lemma.
We define another family, F of pairs of distributions on [W ] as follows. First select a point (x, y)
from the renormalized version of |E 2 − E02 |. Then return the pair of distributions p = q equals the
uniform distribution over {x, y}.
To define D and D ′ , we split [(k + 2m)W ] into k + 2m blocks of size W . A sample from D
assigns to each block independently the pseudo-distribution:
• E0 /m (i.e., a random sample from E0 scaled by a factor of 1/m) with probability m/(k + 2m)
• E0 ǫ/k with probability k/(k + 2m)
• F/m with probability m/(k + 2m).
A sample from D ′ assigns to each block independently the pseudo-distribution:
• E0 /m with probability m/(k + 2m)
• Eǫ/k with probability k/(k + 2m)
• F/m with probability m/(k + 2m).
It is easy to see that D and D ′ satisfy the first three of the properties listed above. To demonstrate the fourth, let X be a uniform Bernoulli random variable. Let A be obtained by applying a
Poisson process of parameter m to a sample from D if X = 0, and to a sample from D ′ if X = 1.
We will show that I(X : A) = o(1). Once again, letting A = (A1 , A2 , . . . , Ak+2m ), where Ai are
the samples taken from the ith block, we note that the Ai are conditionally independent on X and
therefore, I(X : A) ≤ (k + 2m)I(X : A1 ).
As before, no information is gained when |A1 | < 2, and the contribution when |A1 | ≥ 3 is
O((k/(k + m))2 (mǫ/k)6 /(m/(k + m))), which leads to a total contribution of o(1) when m =
o(k4/5 /ǫ6/5 ). It remains to consider the contribution from events where |A1 | = 2.
This is
X
(Pr(A1 = x|X = 0) − Pr(A1 = x|X = 1))2
.
O
Pr(A1 = x)
2
x∈([W ]×{p,q})
Note that the contribution to Pr(A1 = x|X = 0) − Pr(A1 = x|X = 1) from cases where D and D ′
on block 1 are E0 /m or F/m cancel out. Therefore, we have that
| Pr(A1 = x|X = 0) − Pr(A1 = x|X = 1)| = O((k/(k + m))(mǫ/k)2 |E 2 (x) − E02 (x)|)
= O((k/(k + m))(mǫ/k)2 F(x)/ log 2 (W )).
On the other hand, the Pr(A1 = x) is at least the probability that A1 = x when the restriction
to block 1 is F/m, which is Ω(m/(k + m)F(x)). Therefore, the contribution to I(X : A1 ) coming
17
from events where |A1 | = 2 is
X
x∈([W ]×{p,q})2
=
X
(Pr(A1 = x|X = 0) − Pr(A1 = x|X = 1))2
Pr(A1 = x)
((k/(k + m))(mǫ/k)2 F 2 (x)/ log2 (W ))2
O
m/(k + m)F 2 (x)
2
O
x∈([W ]×{p,q})
=
X
x∈([W ]×{p,q})2
F 2 (x)O(k2 (mǫ/k)4 log−4 (W )/(m(k + m)))
= O(m3 ǫ4 k−2 log−4 (W )/(m + k)).
Hence, the total contribution to I(X : A) from such terms is O(m3 ǫ4 k−2 log−4 (W )/(m + k)). This
is o(1) if m = o(k2/3 log4/3 (W )/ǫ4/3 ). This completes our proof.
References
[ADH+ 15] J. Acharya, I. Diakonikolas, C. Hegde, J. Li, and L. Schmidt. Fast and Near-Optimal
Algorithms for Approximating Distributions by Histograms. In 34th ACM SIGMODSIGACT-SIGAI Symposium on Principles of Database Systems, PODS 2015, pages
249–263, 2015.
[ADLS16] J. Acharya, I. Diakonikolas, J. Li, and L. Schmidt. Fast algorithms for segmented
regression. In Proceedings of the 33nd International Conference on Machine Learning,
ICML 2016, pages 2878–2886, 2016.
[ADLS17] J. Acharya, I. Diakonikolas, J. Li, and L. Schmidt. Sample-optimal density estimation in
nearly-linear time. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2017, pages 1278–1289, 2017. Full version available at
https://arxiv.org/abs/1506.00671.
[BBBB72] R.E. Barlow, D.J. Bartholomew, J.M. Bremner, and H.D. Brunk. Statistical Inference
under Order Restrictions. Wiley, New York, 1972.
[BFR+ 00] T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White. Testing that distributions are close. In IEEE Symposium on Foundations of Computer Science, pages
259–269, 2000.
[BKR04]
T. Batu, R. Kumar, and R. Rubinfeld. Sublinear algorithms for testing monotone and
unimodal distributions. In ACM Symposium on Theory of Computing, pages 381–390,
2004.
[Can15]
C. L. Canonne. A survey on distribution testing: Your data is big. but is it blue?
Electronic Colloquium on Computational Complexity (ECCC), 22:63, 2015.
[Can16]
C. L. Canonne. Are few bins enough: Testing histogram distributions. In Proceedings
of the 35th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, PODS 2016, pages 455–463, 2016.
[CDGR16] C. L. Canonne, I. Diakonikolas, T. Gouleakis, and R. Rubinfeld. Testing shape restrictions of discrete distributions. In 33rd Symposium on Theoretical Aspects of Computer
Science, STACS 2016, pages 25:1–25:14, 2016.
18
[CDSS13]
S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Learning mixtures of structured
distributions over discrete domains. In SODA, pages 1380–1394, 2013.
[CDSS14a] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Efficient density estimation via
piecewise polynomial approximation. In STOC, pages 604–613, 2014.
[CDSS14b] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Near-optimal density estimation in
near-linear time using variable-width histograms. In NIPS, pages 1844–1852, 2014.
[CDVV14] S. Chan, I. Diakonikolas, P. Valiant, and G. Valiant. Optimal algorithms for testing
closeness of discrete distributions. In SODA, pages 1193–1203, 2014.
[CMN98]
S. Chaudhuri, R. Motwani, and V. R. Narasayya. Random sampling for histogram
construction: How much is enough? In SIGMOD Conference, pages 436–447, 1998.
[DDKT16] C. Daskalakis, A. De, G. Kamath, and C. Tzamos. A size-free CLT for poisson multinomials and its applications. In Proceedings of the 48th Annual ACM Symposium on
the Theory of Computing, STOC ’16, New York, NY, USA, 2016. ACM.
[DDO+ 13] C. Daskalakis, I. Diakonikolas, R. O’Donnell, R.A. Servedio, and L. Tan. Learning
Sums of Independent Integer Random Variables. In FOCS, pages 217–226, 2013.
[DDS12a]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning k-modal distributions via
testing. In SODA, pages 1371–1385, 2012.
[DDS12b]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions. In STOC, pages 709–728, 2012.
[DDS+ 13] C. Daskalakis, I. Diakonikolas, R. Servedio, G. Valiant, and P. Valiant. Testing k-modal
distributions: Optimal algorithms via reductions. In SODA, pages 1833–1852, 2013.
[DGPP16] I. Diakonikolas, T. Gouleakis, J. Peebles, and E. Price. Collision-based testers are optimal for uniformity and closeness. Electronic Colloquium on Computational Complexity
(ECCC), 23:178, 2016.
[DHS15]
I. Diakonikolas, M. Hardt, and L. Schmidt. Differentially private learning of structured
discrete distributions. In NIPS, pages 2566–2574, 2015.
[DK16]
I. Diakonikolas and D. M. Kane. A new approach for testing properties of discrete
distributions. In FOCS, pages 685–694, 2016. Full version available at abs/1601.05557.
[DKN15a] I. Diakonikolas, D. M. Kane, and V. Nikishkin. Optimal algorithms and lower bounds
for testing closeness of structured distributions. In IEEE 56th Annual Symposium on
Foundations of Computer Science, FOCS 2015, pages 1183–1202, 2015.
[DKN15b] I. Diakonikolas, D. M. Kane, and V. Nikishkin. Testing identity of structured distributions. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2015, pages 1841–1854, 2015.
[DKS16a]
I. Diakonikolas, D. M. Kane, and A. Stewart. Efficient robust proper learning of logconcave distributions. CoRR, abs/1606.03077, 2016.
[DKS16b] I. Diakonikolas, D. M. Kane, and A. Stewart. The fourier transform of poisson multinomial distributions and its algorithmic applications. In Proceedings of STOC’16, 2016.
19
[DKS16c]
I. Diakonikolas, D. M. Kane, and A. Stewart. Learning multivariate log-concave distributions. CoRR, abs/1605.08188, 2016.
[DKS16d] I. Diakonikolas, D. M. Kane, and A. Stewart. Optimal Learning via the Fourier Transform for Sums of Independent Integer Random Variables. In COLT, volume 49, pages
831–849, 2016. Full version available at arXiv:1505.00662.
[DKS16e]
I. Diakonikolas, D. M. Kane, and A. Stewart. Properly learning poisson binomial distributions in almost polynomial time. In Proceedings of the 29th Conference on Learning
Theory, COLT 2016, pages 850–878, 2016. Full version available at arXiv:1511.04066.
[DL01]
L. Devroye and G. Lugosi. Combinatorial methods in density estimation. Springer
Series in Statistics, Springer, 2001.
[DL04]
L. Devroye and G. Lugosi. Bin width selection in multivariate histograms by the
combinatorial method. Test, 13(1):129–145, 2004.
[FD81]
D. Freedman and P. Diaconis. On the histogram as a density estimator:l2 theory.
Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete, 57(4):453–476, 1981.
[GGI+ 02]
A. C. Gilbert, S. Guha, P. Indyk, Y. Kotidis, S. Muthukrishnan, and M. Strauss.
Fast, small-space algorithms for approximate histogram maintenance. In STOC, pages
389–398, 2002.
[GJ14]
P. Groeneboom and G. Jongbloed. Nonparametric Estimation under Shape Constraints:
Estimators, Algorithms and Asymptotics. Cambridge University Press, 2014.
[GKS06]
S. Guha, N. Koudas, and K. Shim. Approximation and streaming algorithms for histogram construction problems. ACM Trans. Database Syst., 31(1):396–438, 2006.
[ILR12]
P. Indyk, R. Levi, and R. Rubinfeld. Approximating and Testing k-Histogram Distributions in Sub-linear Time. In PODS, pages 15–22, 2012.
[JKM+ 98] H. V. Jagadish, N. Koudas, S. Muthukrishnan, V. Poosala, K. C. Sevcik, and T. Suel.
Optimal histograms with quality guarantees. In VLDB, pages 275–286, 1998.
[Kle09]
J. Klemela. Multivariate histograms with data-dependent partitions. Statistica Sinica,
19(1):159–176, 2009.
[LN96]
G. Lugosi and A. Nobel. Consistency of data-driven histogram methods for density
estimation and classification. Ann. Statist., 24(2):687–706, 04 1996.
[LR05]
E. L. Lehmann and J. P. Romano. Testing statistical hypotheses. Springer Texts in
Statistics. Springer, 2005.
[NP33]
J. Neyman and E. S. Pearson. On the problem of the most efficient tests of statistical hypotheses. Philosophical Transactions of the Royal Society of London. Series
A, Containing Papers of a Mathematical or Physical Character, 231(694-706):289–337,
1933.
[Pan08]
L. Paninski. A coincidence-based test for uniformity given very sparsely-sampled discrete data. IEEE Transactions on Information Theory, 54:4750–4755, 2008.
[Rub12]
R. Rubinfeld. Taming big probability distributions. XRDS, 19(1):24–28, 2012.
20
[Sco79]
D. W. Scott. On optimal and data-based histograms. Biometrika, 66(3):605–610, 1979.
[Sco92]
D.W. Scott. Multivariate Density Estimation: Theory, Practice and Visualization.
Wiley, New York, 1992.
[TGIK02]
N. Thaper, S. Guha, P. Indyk, and N. Koudas. Dynamic multidimensional histograms.
In SIGMOD Conference, pages 428–439, 2002.
[VV14]
G. Valiant and P. Valiant. An automatic inequality prover and instance optimal identity
testing. In FOCS, 2014.
[WN07]
R. Willett and R. D. Nowak. Multiscale poisson intensity and density estimation. IEEE
Transactions on Information Theory, 53(9):3171–3187, 2007.
21
| 10math.ST
|
1
arXiv:cs/0509027v1 [cs.PL] 10 Sep 2005
Haskell’s overlooked object system
— 1 February 2008 —
Oleg Kiselyov
Fleet Numerical Meteorology and Oceanography Center, Monterey, CA
Ralf Lämmel
Microsoft Corp., Redmond, WA
Abstract
Haskell provides type-class-bounded and parametric polymorphism as opposed to subtype
polymorphism of object-oriented languages such as Java and OCaml. It is a contentious
question whether Haskell 98 without extensions, or with common extensions, or with
new extensions can fully support conventional object-oriented programming with encapsulation, mutable state, inheritance, overriding, statically checked implicit and explicit
subtyping, and so on.
In a first phase, we demonstrate how far we can get with object-oriented functional
programming, if we restrict ourselves to plain Haskell 98. In the second and major phase,
we systematically substantiate that Haskell 98, with some common extensions, supports
all the conventional OO features plus more advanced ones, including first-class lexically
scoped classes, implicitly polymorphic classes, flexible multiple inheritance, safe downcasts
and safe co-variant arguments. Haskell indeed can support width and depth, structural
and nominal subtyping. We address the particular challenge to preserve Haskell’s type
inference even for objects and object-operating functions. Advanced type inference is a
strength of Haskell that is worth preserving. Many of the features we get “for free”: the
type system of Haskell turns out to be a great help and a guide rather than a hindrance.
The OO features are introduced in Haskell as the OOHaskell library, non-trivially
based on the HList library of extensible polymorphic records with first-class labels and
subtyping. The library sample code, which is patterned after the examples found in OO
textbooks and programming language tutorials, including the OCaml object tutorial,
demonstrates that OO code translates into OOHaskell in an intuition-preserving way:
essentially expression-by-expression, without requiring global transformations.
OOHaskell lends itself as a sandbox for typed OO language design.
Keywords: Object-oriented functional programming, Object type inference, Typed objectoriented language design, Heterogeneous collections, ML-ART, Mutable objects, TypeClass-based programming, Haskell, Haskell 98, Structural subtyping, Duck typing, Nominal subtyping, Width subtyping, Deep subtyping, Co-variance
2
O. Kiselyov and R. Lämmel
Contents
1 Introduction
2 The folklore ‘shapes’ example
2.1 C++ reference encoding
2.2 OOHaskell encoding
2.3 Discussion of the example
2.3.1 Classes vs. interfaces
2.3.2 Extensibility and encapsulation
2.3.3 Subtyping technicalities
3 Alternative Haskell encodings
3.1 Map subtype hierarchy to an algebraic datatype
3.2 Map object data to tail-polymorphic record types
3.3 Functional objects, again with tail polymorphism
3.4 Mutable objects, again with tail polymorphism
3.5 Subtypes as composed record types with overloading
3.6 Variation — existential quantification
3.7 Variation — heterogeneous collections
4 Type-agnostic OOHaskell idioms
4.1 Objects as records
4.1.1 First-class labels
4.1.2 Mutable variables
4.1.3 HList records
4.1.4 OO test cases
4.2 Object generators
4.2.1 Constructor arguments
4.2.2 Construction-time computations
4.2.3 Implicitly polymorphic classes
4.2.4 Nested object generators
4.2.5 Open recursion
4.2.6 Instantiation checking
4.3 Reuse techniques
4.3.1 Single inheritance with extension
4.3.2 Class-polymorphic functionality
4.3.3 Single inheritance with override
4.3.4 Orphan methods
4.3.5 Flexible reuse schemes
4.4 Safe value recursion
5 Type-perceptive OOHaskell idioms
5.1 Semi-implicit upcasting
5.2 Narrow to a fixed type
5.3 Self-returning methods
5.4 Casts based on dynamics
5.5 Casts based on unions
5.6 Explicit type constraints
5.7 Nominal subtyping
5.8 Iso-recursive types
5.9 Width and depth subtyping
5.10 Co-variant method arguments
5.11 Anti-patterns for subtyping
6 Discussion
6.1 Usability issues
6.1.1 Usability of inferred types
6.1.2 Usability of type errors
6.1.3 Efficiency of object encoding
6.2 Related work
6.2.1 The ML-ART object encoding
6.2.2 Haskell language extensions
6.2.3 Object encodings for Haskell
6.3 More future work
7 Concluding remarks
References
3
5
5
8
10
10
11
11
11
12
14
17
20
23
27
28
29
30
30
31
32
33
34
34
34
35
36
37
38
39
39
40
40
41
41
43
46
46
48
49
50
52
54
55
58
59
62
64
66
66
66
68
70
71
71
72
73
74
74
76
Haskell’s overlooked object system
— 1 February 2008 —
3
1 Introduction
The topic of object-oriented programming in the functional language Haskell
is raised time and again on programming language mailing lists, on programming tutorial websites, and in verbal communication at programming
language conferences with remarkable intensity. Dedicated OO Haskell language extensions have been proposed; specific OO idioms have been encoded
in Haskell (Hughes & Sparud, 1995; Gaster & Jones, 1996; Finne et al., 1999;
Shields & Peyton Jones, 2001; Nordlander, 2002; Bayley, 2005). The interest in this
topic is not at all restricted to Haskell researchers and practitioners since there is a
fundamental and unsettled question — a question that is addressed in the present
paper:1
What is the relation between type-class-bounded and subtype polymorphism?
In this research context, we specifically (and emphatically) restrict ourselves to the
existing Haskell language (Haskell 98 and common extensions where necessary),
i.e., no new Haskell extensions are to be proposed. As we will substantiate, this
restriction is adequate, as it allows us to deliver a meaningful and momentous answer to the aforementioned question. At a more detailed level, we offer the following
motivation for research on OO programming in Haskell:
• In an intellectual sense, one may wonder whether Haskell’s advanced type
system is expressive enough to model object types, inheritance, subtyping,
virtual methods, etc. No general, conclusive result has been available so far.
• In a practical sense, one may wonder whether we can faithfully transport
imperative OO designs from, say, C#, C++, Eiffel, Java, VB to Haskell —
without totally rewriting the design and without foreign-language interfacing.
• From a language design perspective, Haskell has a strong record in prototyping semantics and encoding abstraction mechanisms, but one may wonder
whether Haskell can perhaps even serve as a sandbox for design of typed
object-oriented languages so that one can play with new ideas without the
immediate need to write or modify a compiler.
• In an educational sense, one may wonder whether more or less advanced
functional and object-oriented programmers can improve their understanding
of Haskell’s type system and OO concepts by looking into the pros and cons
of different OO encoding options in Haskell.
1
On a more anecdotal account, we have collected informative pointers to mailing list discussions, which document the unsettled understanding of OO programming in Haskell and the relation between OO classes and Haskell’s type classes:
http://www.cs.mu.oz.au/research/mercury/mailing-lists/mercury-users/mercury-users.0105/0051.html,
http://www.talkaboutprogramming.com/group/comp.lang.functional/messages/47728.html,
http://www.haskell.org/pipermail/haskell/2003-December/013238.html,
http://www.haskell.org/pipermail/haskell-cafe/2004-June/006207.html,
http://www.haskell.org//pipermail/haskell/2004-June/014164.html
4
O. Kiselyov and R. Lämmel
This paper delivers substantiated, positive answers to these questions. We describe OOHaskell — a Haskell-based library for (as of today: imperative) OO programming in Haskell. OOHaskell delivers Haskell’s “overlooked” object system.
The key to this result is a good deal of exploitation of Haskell’s advanced type system combined with a careful identification of a suitable object encoding. We instantiate and enhance existing encoding techniques (such as (Pierce & Turner, 1994;
Rémy, 1994; Abadi & Cardelli, 1996)) aiming at a practical object system that
blends well with the host language — Haskell. We take advantage of our previous
work on heterogeneous collections (Kiselyov et al., 2004) (the HList library). More
generally, we put type-class-based or type-level programming to work (Hallgren, 2001;
McBride, 2002; Neubauer et al., 2002; Neubauer et al., 2001).
The simplified story is the following:
- Classes are represented as functions that are in fact object generators.
- State is maintained through mutable variables allocated by object generators.
- Objects are represented as records of closures with a component for each method.
- Methods are monadic functions that can access state and self.
- We use HList’s record calculus (extensible records, up-casts, etc.).
- We use type-class-based functionality to program the object typing rules.
To deliver a faithful, convenient and comprehensive object system, several techniques had to be discovered and combined. Proper effort was needed to preserve
Haskell’s type inference for OO programming idioms (as opposed to explicit type
declarations or type constraints for classes, methods, and up-casts). The obtained
result, OOHaskell, delivers an amount of polymorphism and type inference that
is unprecedented. Proper effort was also needed in order to deploy value recursion
for closing object generators. Achieving safety of this approach was a known challenge (Rémy, 1994). In order to fully appreciate the object system of OOHaskell,
we also review less sophisticated, less favourable encoding alternatives.
Not only OOHaskell provides the conventional OO idioms; we have also
language-engineered several features that are either bleeding-edge or unattainable
in mainstream OO languages: for example, first-class classes and class closures; statically type-checked collection classes with bounded polymorphism of implicit collection arguments; multiple inheritance with user-controlled sharing; safe co-variant
argument subtyping. It is remarkable that these and more familiar object-oriented
features are not introduced by fiat — we get them for free. For example, the type
of a collection with bounded polymorphism of elements is inferred automatically
by the compiler. Also, abstract classes are uninstantiatable not because we say so
but because the program will not typecheck otherwise. Co- and contra-variant subtyping rules and the safety conditions for the co-variant method argument types
are checked automatically without any programming on our part. These facts suggest that (OO)Haskell lends itself as prime environment for typed object-oriented
language design.
Haskell’s overlooked object system
— 1 February 2008 —
5
Road-map of this paper
• Sec. 2: We encode a tutorial OO example both in C++ and OOHaskell.
• Sec. 3: We review alternative object encodings in Haskell 98 and beyond.
• Sec. 4 and Sec. 5: We describe all OOHaskell idioms. The first part focuses
on idioms where subtyping and object types do not surface the OO program
code. The second part covers all technical details of subtyping including casts
and variance properties.
• Sec. 6: We discuss usability issues, related work and future work.
• Sec. 7: We conclude the paper.
The main sections, Sec. 4 and Sec. 5, are written in tutorial style, as to ease digestion of all techniques, as to encourage OO programming and OO language design
experiments. There is an extended source distribution available.2
2 The folklore ‘shapes’ example
One of the main goals of this paper is to be able to represent the conventional
OO code, in as straightforward way as possible. The implementation of our system
may be not for the feeble at heart — however, the user of the system must be
able to write conventional OO code without understanding the complexity of the
implementation. Throughout the paper, we illustrate OOHaskell with a series of
practical examples as they are commonly found in OO textbooks and programming
language tutorials. In this section, we begin with the so-called ‘shapes’ example.
We face a type for ‘shapes’ and two subtypes for ‘rectangles’ and ‘circles’; see
Fig. 1. Shapes maintain coordinates as state. Shapes can be moved around and
drawn. The exercise shall be to place objects of different kinds of shapes in a
collection and to iterate over them as to draw the shapes. It turns out that this
example is a crisp OO benchmark.3
2.1 C++ reference encoding
The type of shapes can be defined as a C++ class as follows:
class Shape {
public:
// Constructor method
Shape(int newx, int newy) {
x = newx;
y = newy;
}
2
3
The source code can be downloaded at http://www.cwi.nl/~ralf/OOHaskell/ , and it is subject
to a very liberal license (MIT/X11 style). As of writing, the actual code commits to a few specific
extensions of the GHC implementation of Haskell — for reasons of convenience. In principle,
Haskell 98 + multi-parameter classes with functional dependencies is sufficient.
The ‘shapes’ problem has been designed by Jim Weirich and deeply explored by him and
Chris Rathman. See the multi-lingual collection ‘OO Example Code’ by Jim Weirich at
http://onestepback.org/articles/poly/ ; see also an even heavier collection ‘OO Shape Examples’ by Chris Rathman at http://www.angelfire.com/tx4/cus/shapes/.
6
O. Kiselyov and R. Lämmel
Fig. 1. Shapes with state and a subtype-specific draw method
// Accessors
int getX() { return
int getY() { return
void setX(int newx)
void setY(int newy)
x; }
y; }
{ x = newx; }
{ y = newy; }
// Move shape position
void moveTo(int newx, int newy) {
x = newx;
y = newy;
}
// Move shape relatively
void rMoveTo(int deltax, int deltay) {
moveTo(getX() + deltax, getY() + deltay);
}
// An abstract draw method
virtual void draw() = 0;
// Private data
private:
int x;
int y;
}
The x, y coordinates are private, but they can be accessed through getters and setters. The methods for accessing and moving shapes are inherited by the subclasses
of Shape. The draw method is virtual and even abstract; hence concrete subclasses
must implement draw.
The subclass Rectangle is derived as follows:
Haskell’s overlooked object system
— 1 February 2008 —
7
class Rectangle: public Shape {
public:
// Constructor method
Rectangle(int newx, int newy, int newwidth, int newheight)
: Shape(newx, newy) {
width = newwidth;
height = newheight;
}
// Accessors
int getWidth() { return width; }
int getHeight() { return height; }
void setWidth(int newwidth)
{ width = newwidth; }
void setHeight(int newheight) { height = newheight; }
// Implementation of the abstract draw method
void draw() {
cout << "Drawing a Rectangle at:("
<< getX() << "," << getY()
<< "), width " << getWidth()
<< ", height " << getHeight() << endl;
}
// Additional private data
private:
int width;
int height;
};
For brevity, we elide the similar derivation of the subclass Circle:
class Circle : public Shape {
Circle(int newx, int newy, int newradius)
: Shape(newx, newy) { ... }
...
}
The following code block constructs different shape objects and invokes their
methods. More precisely, we place two shapes of different kinds in an array, scribble,
and then loop over it to draw and move the shape objects:
Shape *scribble[2];
scribble[0] = new Rectangle(10, 20, 5, 6);
scribble[1] = new Circle(15, 25, 8);
for (int i = 0; i < 2; i++) {
scribble[i]->draw();
scribble[i]->rMoveTo(100, 100);
scribble[i]->draw();
}
The loop over scribble exercises subtyping polymorphism: the actually executed
implementation of the draw method differs per element in the array. The program
run produces the following output — due to the logging-like implementations of
the draw method:
8
O. Kiselyov and R. Lämmel
Drawing
Drawing
Drawing
Drawing
a
a
a
a
Rectangle at:(10,20), width 5, height 6
Rectangle at:(110,120), width 5, height 6
Circle at:(15,25), radius 8
Circle at:(115,125), radius 8
2.2 OOHaskell encoding
We now show an OOHaskell encoding, which happens to pleasantly mimic the
C++ encoding, while any remaining deviations are appreciated. Most notably, we
are going to leverage type inference: we will not define any type. The code shall be
fully statically typed nevertheless.
Here is the OOHaskell rendering of the shape class:
-- Object generator for shapes
shape newx newy self
= do
-- Create references for private state
x <- newIORef newx
y <- newIORef newy
-- Return object as record of methods
returnIO $ getX
.=. readIORef x
.*. getY
.=. readIORef y
.*. setX
.=. writeIORef x
.*. setY
.=. writeIORef y
.*. moveTo
.=. (\newx newy -> do
(self # setX) newx
(self # setY) newy )
.*. rMoveTo .=. (\deltax deltay ->
do
x <- self # getX
y <- self # getY
(self # moveTo) (x + deltax) (y + deltay) )
.*. emptyRecord
Classes become functions that take constructor arguments plus a self reference
and that return a computation whose result is the new object — a record of methods
including getters and setters. We can invoke methods of the same object through
self; cf. the method invocation self # getX and others. (The infix operator #
denotes method invocation.) Our objects are mutable, implemented via IORef.
(STRef also suffices.) Since most OO systems in practical use have mutable state,
OOHaskell does not (yet) offer functional objects, which are known to be challenging on their own. We defer functional objects to future work.
We use the extensible records of the HList library (Kiselyov et al., 2004), hence:
•
•
•
•
emptyRecord denotes what the name promises,
(.*.) stands for (right-associative) record extension,
(.=.) is record-component construction: label .=. value,
Labels are defined according to a trivial scheme, to be explained later.
Haskell’s overlooked object system
— 1 February 2008 —
9
The abstract draw method is not mentioned in the OOHaskell code because it
is not used in any other method, neither did we dare declaring its type. As a side
effect, the object generator shape is instantiatable whereas the explicit declaration
of the abstract draw method made the C++ class Shape uninstantiatable. We will
later show how to add similar declarations in OOHaskell.
We continue with the OOHaskell code for the shapes example.
-- Object generator for rectangles
rectangle newx newy width height self
= do
-- Invoke object generator of superclass
super <- shape newx newy self
-- Create references for extended state
w <- newIORef width
h <- newIORef height
-- Return object as
returnIO $
getWidth .=.
.*. getHeight .=.
.*. setWidth .=.
.*. setHeight .=.
record of methods
readIORef
readIORef
(\neww ->
(\newh ->
w
h
writeIORef w neww)
writeIORef h newh)
.*. draw .=.
do -- Implementation of the abstract draw method
putStr "Drawing a Rectangle at:(" <<
self # getX << ls "," << self # getY <<
ls "), width " << self # getWidth <<
ls ", height " << self # getHeight << ls "\n"
-- Rectangle records start from shape records
.*. super
This snippet illustrates the essence of inheritance in OOHaskell. Object generation for the supertype is made part of the monadic sequence that defines object
generation for the subtype; self is passed from the subtype to the supertype.
Subtype records are derived from supertype records through record extension (or
potentially also through record updates when overrides are to be modelled).
As in the C++ case, we elide the derivation of the object generators for circles:
circle newx newy newradius self
= do
super <- shape newx newy self
...
returnIO ... .*. super
Ultimately, here is the OOHaskell rendering of the ‘scribble loop’:
-- Object construction and invocation as a monadic sequence
myOOP = do
-- Construct objects
s1 <- mfix (rectangle (10::Int) (20::Int) 5 6)
s2 <- mfix (circle (15::Int) 25 8)
10
O. Kiselyov and R. Lämmel
-- Create a homogeneous list of different shapes
let scribble = consLub s1 (consLub s2 nilLub)
-- Loop over list with normal monadic map
mapM_ (\shape -> do
shape # draw
(shape # rMoveTo) 100 100
shape # draw)
scribble
The use of mfix (an analogue of the new in C++) reflects that object generators take ‘self’ and construct (part of) it. Open recursion enables inheritance. The
let scribble . . . binding is noteworthy. We cannot directly place rectangles and
circles in a normal Haskell list — the following cannot possibly type check:
let scribble = [s1,s2] -- s1 and s2 are of different types!
We have to homogenise the types of s1 and s2 when forming a Haskell list. To this
end, we use special list constructors nilLub and consLub as opposed to the normal
list constructors [] and (:). These new constructors coerce the list elements to the
least-upper bound type of all the element types. Incidentally, if the ‘intersection’ of
the types of the objects s1 and s2 does not include the methods that are invoked
later (i.e., draw and rMoveTo), we get a static type error which literally says so. As
a result, the original for-loop can be carried out in the native Haskell way: a normal
(monadic) list map over a normal Haskell list of shapes. Hence, we have exercised
a faithful model of subtype polymorphism, which also allows for (almost) implicit
subtyping. OOHaskell provides several subtyping models, as we will study later.
2.3 Discussion of the example
2.3.1 Classes vs. interfaces
The C++ code should not be misunderstood to suggest that class inheritance is the
only OO design option for the shapes hierarchy. In a Java-like language, one may
want to model Shape as an interface, say, IShape, with Rectangle and Circle as
classes implementing this interface. This design would not allow us to reuse the implementations of the accessors and the move methods. So one may want to combine
interface polymorphism and class inheritance. That is, the classes Rectangle and
Circle will be rooted by an additional implementation class for shapes, say Shape,
which hosts implementations shared among different shape classes — incidentally
a part of the IShape interface. The remainder of the IShape interface, namely the
draw method in our example, would be implemented in Rectangle and Circle.
More generally, OO designs that employ interface polymorphism alone are rare,
so we need to provide encodings for both OO interface polymorphism and OO class
inheritance in OOHaskell. One may say that the former mechanism is essentially
covered by Haskell’s type classes (modulo the fact that we would still need an object encoding). The latter mechanism is specifically covered by original HList and
OOHaskell contributions: structural subtyping polymorphism for object types,
based on polymorphic extensible records and programmable subtyping constraints.
(Sec. 5.7 discusses nominal object types in OOHaskell)˙
Haskell’s overlooked object system
— 1 February 2008 —
11
2.3.2 Extensibility and encapsulation
Both the C++ encoding and the OOHaskell encoding of the shapes example are
faithful to the encapsulation premise as well as the extensibility premise of the OO
paradigm. An object encapsulates both data (‘state’) and methods (‘behaviour’).
One may add new kinds of shapes without rewriting (or, perhaps, even re-compiling)
existing code.
Both premises are the subject of an unsettled debate in the programming language community, especially with regards to functional programming. The basic
OO paradigm has been criticised (Zenger & Odersky, 2004) for its over-emphasis
of extensibility in the subtyping dimension and for its neglect of other dimensions
such as the addition of new functions into a pre-existing subtyping hierarchy. While
we agree with this overall criticism, we avoid the debate in this paper. We simply
want OOHaskell to provide an object encoding that is compatible with the established OO paradigm. (Incidentally, some of the non-encapsulation-based encodings
in Sec. 3 show that Haskell supports extensibility in both the data and the functionality dimension.)
2.3.3 Subtyping technicalities
The “scribble loop” is by no means a contrived scenario. It is a faithful instance
of the ubiquitous composite design pattern (Gamma et al., 1994). In terms of expressiveness and typing challenges, this sort of loop over an array of shapes of
different kinds forces us to explore the tension between implicit and explicit subtyping. As we will discuss, it is relatively straightforward to use type-class-bounded
polymorphism to represent subtype constraints. It is however less straightforward
to accumulate entities of different subtypes in the same collection. With explicit
subtyping (e.g., by wrapping in a properly constrained existential envelope) the
burden would be on the side of the programmer. A key challenge for OOHaskell
was to make subtyping (almost) implicit in all the cases, where a OO programmer would expect it. This is a particular area in which OOHaskell goes beyond OCaml (Leroy et al. , 2004)— the de-facto leading strongly typed functional
object-oriented language. OOHaskell provides a range of subtyping notions, including one that even allows for safe downcasts for object types. This is again
something that has not been achieved in OCaml to date.
3 Alternative Haskell encodings
OOHaskell goes particularly far in providing an object system, when compared
to conservative Haskell programming knowledge. To this end, we put type-classbased or type-level programming to work. In this section, we will review more
conservative object encodings with their characteristics and limitations. All of them
require boilerplate code from the programmer.
Some of the ‘conservative’ encodings to come are nevertheless involved and enlightening. In fact, the full spectrum of encodings has not been documented before — certainly not in a Haskell context. So we reckon that their detailed analysis
12
O. Kiselyov and R. Lämmel
makes a useful contribution. Furthermore, several of the discussed techniques are
actually used in OOHaskell, where some of them are simply generalised through
the advanced use of Haskell’s type classes. Hence, the present section is an incremental preparation for the main sections Sec. 4 and Sec. 5.
For most of this section, we limit ourselves to Haskell 98. (By contrast, OOHaskell
requires several common Haskell 98 extensions.) Towards the end of the section, we
will investigate the value of dismissing this restriction.
3.1 Map subtype hierarchy to an algebraic datatype
We begin with a trivial and concise encoding. Its distinguishing characteristic is extreme simplicity.4 It uses only basic Haskell 98 idioms. The encoding is also seriously
limited, lacking extensibility with regard to new forms of shapes (cf. Sec. 2.3.2).
We define an algebraic datatype for shapes, where each kind of shape amounts to
a constructor declaration. For readability, we use labelled fields instead of unlabelled
constructor components.
data Shape =
Rectangle {
,
,
,
getX
getY
getWidth
getHeight
::
::
::
::
Int
Int
Int
Int }
|
Circle { getX
:: Int
, getY
:: Int
, getRadius :: Int }
Both constructor declarations involve labelled fields for the (x, y) position of a
shape. While this reusability dimension is not emphasised at the datatype level, we
can easily define reusable setters for the position. (There are some issues regarding
type safety, which we will address later.) For instance:
setX :: Int -> Shape -> Shape
setX i s = s { getX = i }
We can also define setters for Rectangle- and Circle-specific fields. For instance:
setWidth :: Int -> Shape -> Shape
setWidth i s = s { getWidth = i }
It is also straightforward to define functions for moving around shapes:
moveTo :: Int -> Int -> Shape -> Shape
moveTo x y = setY y . setX x
rMoveTo :: Int
rMoveTo deltax
where
x = getX s +
y = getY s +
4
-> Int -> Shape -> Shape
deltay s = moveTo x y s
deltax
deltay
Thanks to Lennart Augustsson for pointing out this line of encoding.
Cf. http://www.haskell.org/pipermail/haskell/2005-June/016061.html
Haskell’s overlooked object system
— 1 February 2008 —
13
The function for drawing shapes properly discriminates on the kind of shapes.
That is, there is one equation per kind of shape. Subtype polymorphism reduces to
pattern matching, so to say:
draw :: Shape -> IO ()
draw s@(Rectangle _ _ _ _)
= putStrLn ("Drawing a Rectangle at:("
++ (show (getX s))
++ ","
++ (show (getY s))
++ "), width " ++ (show (getWidth s))
++ ", height " ++ (show (getHeight s)))
draw s@(Circle _ _ _)
= putStrLn ("Drawing a Circle at:("
++ (show (getX s))
++ ","
++ (show (getY s))
++ "), radius "
++ (show (getRadius s)))
With this encoding, it is trivial to build a collection of shapes of different kinds and
to iterate over it such that each shape is drawn and moved (and drawn again):
main =
do
let scribble = [ Rectangle 10 20 5 6
, Circle 15 25 8
]
mapM_ ( \x ->
do
draw x
draw (rMoveTo 100 100 x))
scribble
Assessment of the encoding
• The encoding ignores the encapsulation premise of the OO paradigm.
• The foremost weakness of the encoding is the lack of extensibility. The addition of a new kind of shape would require re-compilation of all code; it would
also require amendments of existing definitions or declarations: the datatype
declaration Shape and the function definition draw.
• A related weakness is that the overall scheme does not suggest a way of
dealing with virtual methods: introduce the type of a method for a base
type potentially with an implementation; define or override the method for
a subtype. We would need a scheme that offers (explicit and implicit) open
recursion for datatypes and functions defined on them.
• The setters setX and setY happen to be total because all constructors end
up defining labelled fields getX and getY. The type system does not prevent
us from forgetting those labels for some constructor. It is relatively easy to
14
O. Kiselyov and R. Lämmel
resolve this issue to the slight disadvantage of conciseness. (For instance, we
may avoid labelling entirely, and use pattern matching instead. We may also
compose together rectangles and circles from common shape data and deltas.)
• The use of a single algebraic datatype Shape implies that Rectangle- and
Circle-specific functions cannot be defined as total functions. Such biased
functions, e.g., setWidth, are only defined for certain constructors. Once we go
beyond the simple-minded encoding model of this section, it will be possible to
increase type safety by making type distinctions for different kinds of shapes,
but then we will also encounter the challenge of subtype polymorphism.
3.2 Map object data to tail-polymorphic record types
There is a folklore technique for encoding extensible records (Burton, 1990) that
we can use to model the shapes hierarchy in Haskell 98. Simple type classes let us
implement virtual methods. We meet the remaining challenge of placing different
shapes into one list by making different subtypes homogeneous through embedding
shape subtypes into a union type (Haskell’s Either).
We begin with a datatype for extensible shapes; cf. shapeTail:
data Shape w =
Shape { getX :: Int
, getY :: Int
, shapeTail :: w }
For convenience, we also provide a constructor for shapes:
shape x y w = Shape { getX = x
, getY = y
, shapeTail = w }
We can define setters and movers once and for all for all possible extensions
of Shape by simply leaving the extension type parametric. The actual equations
are literally the same as in the previous section; so we only show the (different)
parametrically polymorphic types:
setX :: Int -> Shape w -> Shape w
setY :: Int -> Shape w -> Shape w
moveTo :: Int -> Int -> Shape w -> Shape w
rMoveTo :: Int -> Int -> Shape w -> Shape w
The presence of the type variable w expresses that the earlier definitions on Shape . . .
can clearly be instantiated to all subtypes of Shape. The draw function must be
placed in a dedicated type class, Draw, because we anticipate the need to provide
type-specific implementations of draw. (One may compare this style with C++
where one explicitly declares a method to be (pure) virtual.)
class Draw w
where
draw :: Shape w -> IO ()
Haskell’s overlooked object system
— 1 February 2008 —
15
Shape extensions for rectangles and circles are built according to a common
scheme. We only show the details for rectangles. We begin with the definition of
the “data delta” contributed by rectangles; each such delta is again polymorphic in
its tail.
data RectangleDelta w =
RectangleDelta { getWidth
:: Int
, getHeight
:: Int
, rectangleTail :: w }
We define the type of rectangles as an instance of Shape:
type Rectangle w = Shape (RectangleDelta w)
For convenience, we provide a constructor for rectangles. Here we fix the tail of
the rectangle delta to (). (We could still further instantiate Rectangle and define
new constructors later, if necessary.)
rectangle x y w h
= shape x y $ RectangleDelta {
getWidth
= w
, getHeight
= h
, rectangleTail = () }
The definition of rectangle-specific setters involves nested record manipulation:
setHeight :: Int -> Rectangle w -> Rectangle w
setHeight i s = s { shapeTail = (shapeTail s) { getHeight = i } }
setWidth :: Int -> Rectangle w -> Rectangle w
setWidth i s = s { shapeTail = (shapeTail s) { getWidth = i } }
The rectangle-specific draw function is defined through a Draw instance:
instance Draw (RectangleDelta w)
where
draw s
=
putStrLn ("Drawing a Rectangle at:("
++ (show (getX s))
++ ","
++ (show (getY s))
++ "), width "
++ (show (getWidth (shapeTail s)))
++ ", height "
++ (show (getHeight (shapeTail s))))
The difficult part is the ‘scribble loop’. We cannot easily form a collection of
shapes of different kinds. For instance, the following attempt will not type-check:
-- Wrong! There is no homogeneous element type.
let scribble = [ rectangle 10 20 5 6
, circle 15 25 8
]
16
O. Kiselyov and R. Lämmel
There is a relatively simple technique to make rectangles and circles homogeneous
within the scope of the scribble list and its clients. We have to establish a union
type for the different kinds of shapes.5 Using an appropriate helper, tagShape,
for embedding shapes into a union type (Haskell’s Either), we may construct a
homogeneous collection as follows:
let scribble = [ tagShape Left (rectangle 10 20 5 6)
, tagShape Right (circle 15 25 8)
]
The boilerplate operation for embedding is trivially defined as follows.
tagShape :: (w -> w’) -> Shape w -> Shape w’
tagShape f s = s { shapeTail = f (shapeTail s) }
Embedding (or tagging) clearly does not disturb the reusable definitions of functions
on Shape w. However, the loop over scribble refers to the draw operation, which
is defined for RectangleDelta and CircleDelta, but not for the union over these
two types. We have to provide a trivial boilerplate for generalising draw:
instance (Draw a, Draw b) => Draw (Either a b)
where
draw = eitherShape draw draw
(This instance actually suffices for arbitrarily nested unions of Shape subtypes.)
Here, eitherShape is variation on the normal fold operation for unions, i.e., either.
We discriminate on the Left vs. Right cases for the tail of a shape datum. This
boilerplate operation is independent of draw, but specific to Shape.
eitherShape :: (Shape w -> t) -> (Shape w’ -> t) -> Shape (Either w w’) -> t
eitherShape f g s
= case shapeTail s of
(Left s’) -> f (s { shapeTail = s’ })
(Right s’) -> g (s { shapeTail = s’ })
The Draw instance for Either makes it clear that we use the union type as an
intersection type. We may only invoke a method on the union only if we may
invoke the method on either branch of the union. The instance constraints make
that fact obvious.
Assessment of the encoding
• Again, the encoding ignores the encapsulation premise of the OO paradigm:
methods are not encapsulated along with the data.
• The encoding does not have the basic extensibility problem of the previous
section. We can introduce new kinds of shapes without rewriting and recompiling type-specific code.
5
Haskell 98 supports unions in the prelude: with the type name Either, and the two constructors
Left and Right for the branches of the union.
Haskell’s overlooked object system
— 1 February 2008 —
17
• Some patterns of subtype-polymorphic code may require revision, though. For
instance all program points that insert into a subtype-polymorphic collection
or that downcast must agree on the formation of the union type over specific
subtypes. If a new subtype must be covered, then the scattered applications
of embedding operations must be revised.
• We fail to put Haskell’s type inference to work as far as object types are
concerned. We end up defining explicit datatypes for all encoded classes. This
is acceptable from a mainstream OO point of view since nominal types (i.e.,
explicit types) dominate the OO paradigm. However, in Haskell, we would
like to do better by allowing for inference of structural class and interface
types. All subsequent encodings of this section will share this problem. (By
contrast, OOHaskell provides full structural type inference.)
• It is annoying enough that the formation of a subtype-polymorphic collection
requires explicit tagging of all elements; cf. Left and Right. What is worse,
the tagging is done on the delta position of Shape. This makes the scheme noncompositional: Each new base class requires its own functions like tagShape
and eitherShape.
• The encoding of final and virtual methods differs essentially. The former are
encoded as parametric polymorphic functions parameterised in the extension
type. Virtual methods are encoded as type-class-bounded polymorphic functions overloaded in the extension type. Changing a final method into virtual
or vice versa triggers code rewriting. This may be overcome by making all
methods virtual (and using default type class methods to reuse implementations). However, this bias will increase the amount of boilerplate code such
as the instances for Either.
• The subtyping hierarchy leaks into the encoding of subtype-specific accessors;
cf. setWidth. The derivation chain from a base type shows up as nesting
depth in the record access pattern. One may factor out these code patterns
into access helpers and overload them so that all accessors can be coded in a
uniform way. This will complicate the encoding, though.
• The approach is restricted to single inheritance.
3.3 Functional objects, again with tail polymorphism
So far we have defined all methods as separate functions that process “data records”.
Hence, we ignored the OO encapsulation premise: our data and methods were
divorced from each other. Thereby, we were able to circumvent problems of self
references that tend to occur in object encodings. Also, we avoided the classic
dichotomy ‘mutable vs. functional objects’. We will complement the picture by exploring a functional object encoding (this section) and a mutable object encoding
(next section). We continue to use tail-polymorphic records.
In a functional object encoding, object types are necessarily recursive because all
mutating methods are modelled as record components that return “self”. In fact,
18
O. Kiselyov and R. Lämmel
the type-theoretic technique is to use equi-recursive types (Pierce & Turner, 1994).
We must use iso-recursive types instead since Haskell lacks equi-recursive types.
Extensible shapes are modelled through the following recursive datatype:
data Shape w =
Shape { getX
, getY
, setX
, setY
, moveTo
, rMoveTo
, draw
, shapeTail
}
::
::
::
::
::
::
::
::
Int
Int
Int ->
Int ->
Int ->
Int ->
IO ()
w
Shape w
Shape w
Int -> Shape w
Int -> Shape w
This type reflects the complete interface of shapes, including getters, setters, and
more complex methods. The object constructor is likewise recursive. Recall that
recursion models functional mutation, i.e., the construction of a changed object:
shape x y d t
= Shape { getX
, getY
, setX
, setY
, moveTo
, rMoveTo
, draw
, shapeTail
}
=
=
=
=
=
=
=
=
x
y
\x’ -> shape x’ y d t
\y’ -> shape x y’ d t
\x’ y’ -> shape x’ y’ d t
\deltax deltay -> shape (x+deltax) (y+deltay) d t
d x y
t
As before, subtypes are modelled as instantiations of the base-type record. That
is, the Rectangle record type is an instance of the Shape record type, where instantiation fixes the type shapeTail somewhat:
type Rectangle w = Shape (RectangleDelta w)
data RectangleDelta w =
RectangleDelta { getWidth’
, getHeight’
, setWidth’
, setHeight’
, rectangleTail
}
::
::
::
::
::
Int
Int
Int -> Rectangle w
Int -> Rectangle w
w
We used primed labels because we wanted to save the unprimed names for the
actual programmer API. The following implementations of the unprimed functions
hide the fact that rectangle records are nested.
getWidth
getHeight
setWidth
setHeight
=
=
=
=
getWidth’
getHeight’
setWidth’
setHeight’
.
.
.
.
shapeTail
shapeTail
shapeTail
shapeTail
The constructor for rectangles elaborates the constructor for shapes as follows:
Haskell’s overlooked object system
— 1 February 2008 —
19
rectangle x y w h
= shape x y drawRectangle shapeTail
where
drawRectangle x y
= putStrLn ("Drawing a Rectangle at:("
++ (show x) ++ ","
++ (show y) ++ "), width "
++ (show w) ++ ", height "
++ (show h) )
shapeTail
= RectangleDelta { getWidth’
= w
, getHeight’
= h
, setWidth’
= \w’ -> rectangle x y w’ h
, setHeight’
= \h’ -> rectangle x y w h’
, rectangleTail = ()
}
The encoding of the subclass Circle can be derived likewise. (Omitted.)
This time, the scribble loop is set up as follows:
main =
do
let scribble = [ narrowToShape (rectangle 10 20 5 6)
, narrowToShape (circle 15 25 8)
]
mapM_ ( \x ->
do
draw x
draw (rMoveTo x 100 100) )
scribble
The interesting aspect of this encoding concerns the construction of the scribble
list. We cast or narrow the shapes of different kinds to a common type. This is a
general option, which we could have explored in the previous section (where we used
embedding into a union type instead). Narrowing takes a shape with an arbitrary
tail and returns a shape with tail ():
narrowToShape :: Shape w -> Shape
narrowToShape s = s { setX
=
, setY
=
, moveTo
=
, rMoveTo
=
, shapeTail =
}
()
narrowToShape . setX s
narrowToShape . setY s
\z -> narrowToShape . moveTo s z
\z -> narrowToShape . rMoveTo s z
()
Assessment of the encoding
• The encoding is faithful to the encapsulation premise of the OO paradigm.
• The specific extensibility problem of the ‘union type’ approach is resolved (cf.
assessment Sec. 3.2). Code that accesses a subtype-polymorphic collection
does not need to be revised when new subtypes are added elsewhere in the
program. The ‘narrowing’ approach frees the programmer from commitment
to a specific union type.
20
O. Kiselyov and R. Lämmel
• The narrowing approach (unlike the union-type one) does not permit downcasting.
• The implementation of the narrowing operation is base-type-specific just as
the earlier embedding helpers for union types. Boilerplate code of that kind
is, of course, not required from programmers in mainstream OO languages.
3.4 Mutable objects, again with tail polymorphism
We also review an object encoding for mutable objects, where we employ IORefs
of the IO monad to enable object state — as is the case for OOHaskell. The
functions (“methods”) in a record manipulate the state through IORef operations.
We continue to use tail-polymorphic records.
Extensible shapes are modelled through the following type:
data Shape w =
Shape { getX
, getY
, setX
, setY
, moveTo
, rMoveTo
, draw
, shapeTail
}
::
::
::
::
::
::
::
::
IO Int
IO Int
Int ->
Int ->
Int ->
Int ->
IO ()
w
IO ()
IO ()
Int -> IO ()
Int -> IO ()
The result type of all methods is wrapped in the IO monad so that all methods
may have side effects, if necessary. One may wonder whether this is really necessary
for getters. Even for a getter, we may want to add memoisation or logging, when we
override the method in a subclass, in which case a non-monadic result type would
be too restrictive.
The object generator (or constructor) for shapes is parameterised in the initial
shape position x and y, in a concrete implementation of the abstract method draw,
in the tail of the record to be contributed by the subtype, and in self so to enable
open recursion. The latter lets subtypes override method defined in shape. (We will
illustrate overriding shortly.)
shape x y concreteDraw tail self
= do
xRef <- newIORef x
yRef <- newIORef y
tail’ <- tail
returnIO Shape
{ getX
= readIORef xRef
, getY
= readIORef yRef
, setX
= \x’ -> writeIORef xRef x’
, setY
= \y’ -> writeIORef yRef y’
, moveTo
= \x’ y’ -> do { setX self x’; setY self y’ }
, rMoveTo
= \deltax deltay ->
do x <- getX self
y <- getY self
Haskell’s overlooked object system
— 1 February 2008 —
21
moveTo self (x+deltax) (y+deltay)
, draw
= concreteDraw self
, shapeTail = tail’ self
}
The type declarations for rectangles are the following:
type Rectangle w = Shape (RectangleDelta w)
data RectangleDelta w =
RectangleDelta { getWidth’
, getHeight’
, setWidth’
, setHeight’
, rectangleTail
}
::
::
::
::
::
IO Int
IO Int
Int -> IO ()
Int -> IO ()
w
Again, we define unprimed names to hide the nested status of the rectangle API:
getWidth
getHeight
setWidth
setHeight
=
=
=
=
getWidth’
getHeight’
setWidth’
setHeight’
.
.
.
.
shapeTail
shapeTail
shapeTail
shapeTail
We reveal the object generator for rectangles step by step.
rectangle x y w h
= shape x y drawRectangle shapeTail
where
-- to be cont’d
We invoke the generator shape, passing on the normal constructor arguments x
and y, a rectangle-specific draw method, and the tail for the rectangle API. We do
not yet fix the self reference, thereby allowing for further subtyping of rectangle.
We define the draw method as follows, resorting to C++-like syntax, <<, for daisy
chaining output:
drawRectangle self =
putStr "Drawing a Rectangle at:(" <<
getX self << ls "," << getY self <<
ls "), width " << getWidth self <<
ls ", height " << getHeight self <<
ls "\n"
Finally, the following is the rectangle part of the shape object:
shapeTail
= do
wRef <- newIORef w
hRef <- newIORef h
returnIO ( \self ->
RectangleDelta
{ getWidth’
, getHeight’
, setWidth’
, setHeight’
, rectangleTail
} )
=
=
=
=
=
readIORef wRef
readIORef hRef
\w’ -> writeIORef wRef w’
\h’ -> writeIORef hRef h’
()
22
O. Kiselyov and R. Lämmel
The overall subtype derivation scheme is at ease with overriding methods in subtypes. We illustrate this capability by temporarily assuming that the draw method
is not abstract. So we may revise the constructor for shapes as follows:
shape x y tail self
= do
xRef <- newIORef x
yRef <- newIORef y
tail’ <- tail
returnIO Shape
{ -- ... as before but we deviate for draw ...
, draw = putStrLn "Nothing to draw"
}
We override draw when constructing rectangles:
rectangle x y w h self
= do
super <- shape x y shapeTail self
returnIO super { draw = drawRectangle self }
As in the previous section, we use narrowToShape when building a list of different
shapes. Actual object construction ties the recursive knot for the self references with
mfix. Hence, mfix is our operator “new”.
main =
do
s1 <- mfix $ rectangle 10 20 5 6
s2 <- mfix $ circle 15 25 8
let scribble = [ narrowToShape s1
, narrowToShape s2
]
mapM_ ( \x ->
do
draw x
rMoveTo x 100 100
draw x )
scribble
The narrow operation is trivial this time:
narrowToShape :: Shape w -> Shape ()
narrowToShape s = s { shapeTail = () }
We just “chop off” the tail of shape objects. We may no longer use any rectangleor circle-specific methods. One may say that chopping off the tail makes the fields
in the tail and the corresponding methods private. The openly recursive methods,
in particular draw, had access to self that characterised the whole object, before
the chop off. The narrow operation becomes (potentially much) more involved or
infeasible once we consider self-returning methods, binary methods, co- and contravariance, and other advanced OO idioms.
Haskell’s overlooked object system
— 1 February 2008 —
23
Assessment of the encoding This encoding is actually very close to OOHaskell
except that the former uses explicitly declared, non-extensible record types. As
a result, the encoding requires substantial boilerplate code (to account for type
extension) and subtyping is explicit. Furthermore, OOHaskell leverages typelevel programming to lift restrictions like the limited narrowing capabilities.
3.5 Subtypes as composed record types with overloading
Many problems of tail-polymorphic record types prompt us to consider an alternative. We now compose record types for subtypes and use type classes to represent
the actual subtype relationships. Such use of type classes has first been presented
in (Shields & Peyton Jones, 2001) for encoding of OO interface polymorphism in
Haskell. We generalise this technique for class inheritance.
The compositional approach can be described as follows:
•
•
•
•
•
•
The data part of an OO class amounts to a record type.
Each such record type includes components for superclass data
The interface for each OO class amounts to a Haskell type class.
OO superclasses are mapped to Haskell superclass constraints.
Reusable OO method implementations are mapped to default methods.
A subtype is implemented as a type-class instance.
We begin with the record type for the data part of the Shape class:
data ShapeData =
ShapeData { valX :: Int
, valY :: Int }
For convenience, we also provide a constructor:
shape x y = ShapeData { valX = x
, valY = y }
We define a type class Shape that models the OO interface for shapes:
class Shape s
where
getX
:: s -> Int
setX
:: Int -> s -> s
getY
:: s -> Int
setY
:: Int -> s -> s
moveTo
:: Int -> Int -> s -> s
rMoveTo
:: Int -> Int -> s -> s
draw
:: s -> IO ()
-- to be cont’d
We would like to provide reusable definitions for most of these methods (except
for draw of course). In fact, we would like to define the accessors to shape data
once and for all. To this end, we need additional helper methods. While it is clear
how to define accessors on ShapeData, we must provide generic definitions that
are able to handle records that include ShapeData as one of their (immediate or
non-immediate) components. This leads to the following two helpers:
24
O. Kiselyov and R. Lämmel
class Shape s
where
-- cont’d from earlier
readShape :: (ShapeData -> t)
-> s -> t
writeShape :: (ShapeData -> ShapeData) -> s -> s
which let us define generic shape accessors:
class Shape s
where
-- cont’d from earlier
getX
= readShape valX
setX i
= writeShape (\s -> s { valX = i })
getY
= readShape valY
setY i
= writeShape (\s -> s { valY = i })
moveTo x y = setY y . setX x
rMoveTo deltax deltay s = moveTo x y s
where
x = getX s + deltax
y = getY s + deltay
We do not define an instance of the Shape class for ShapeData because the
original shape class was abstract due to the purely virtual draw method. As we
move to rectangles, we define their data part as follows:
data RectangleData =
RectangleData { valShape :: ShapeData
, valWidth :: Int
, valHeight :: Int
}
The rectangle constructor also invokes the shape constructor:
rectangle x y w h
= RectangleData { valShape = shape x y
, valWidth = w
, valHeight = h
}
“A rectangle is a shape.” We provide access to the shape part as follows:
instance Shape RectangleData
where
readShape f
= f . valShape
writeShape f s = s { valShape = readShape f s }
-- to be cont’d
We also implement the draw method.
-- instance Shape RectangleData cont’d
draw s
=
putStrLn ("Drawing a Rectangle at:("
++ (show (getX s)) ++ ","
++ (show (getY s)) ++ "), width "
++ (show (getWidth s)) ++ ", height "
++ (show (getHeight s)))
Haskell’s overlooked object system
— 1 February 2008 —
25
We also need to define a Haskell type class for the OO class of rectangles. OO
subclassing coincides in Haskell type-class subclassing.
class Shape s => Rectangle s
where
-- to be cont’d
The type class is derived from the corresponding OO class just as we explained for
the base class of shapes. The class defines the ‘normal’ interface of rectangles and
access helpers.
-- class Rectangle cont’d
getWidth
:: s -> Int
getWidth
= readRectangle valWidth
setWidth
:: Int -> s -> s
setWidth i
= writeRectangle (\s -> s
getHeight
:: s -> Int
getHeight
= readRectangle valHeight
setHeight
:: Int -> s -> s
setHeight i
= writeRectangle (\s -> s
{ valWidth = i })
{ valHeight = i })
readRectangle :: (RectangleData -> t)
-> s -> t
writeRectangle :: (RectangleData -> RectangleData) -> s -> s
“A rectangle is (nothing but) a rectangle.”
instance Rectangle RectangleData
where
readRectangle = id
writeRectangle = id
The subclass for circles can be encoded in the same way.
The scribble loop can be performed on tagged rectangles and circles:
main =
do
let scribble = [ Left (rectangle 10 20 5 6)
, Right (circle 15 25 8)
]
mapM_ ( \x ->
do
draw x
draw (rMoveTo 100 100 x)
)
scribble
We attach Left and Right tags at the top-level this time. Such simple tagging
was not possible with the tail-polymorphic encodings. We still need an instance for
Shape that covers tagged shapes:
instance (Shape a, Shape b) => Shape (Either a b)
where
readShape f = either (readShape f) (readShape f)
writeShape f = bimap (writeShape f) (writeShape f)
draw
= either draw draw
26
O. Kiselyov and R. Lämmel
The bi-functorial map, bimap, pushes writeShape into the tagged values. The
Either-specific fold operation, either, pushes readShape and draw into the tagged
values. For completeness, we recall the relevant facts about bi-functors and folds
on Either:
class BiFunctor f where
bimap :: (a -> b) -> (c -> d) -> f a c -> f b d
instance BiFunctor Either where
bimap f g (Left x) = Left (f x)
bimap f g (Right x’) = Right (g x’)
either :: (a -> c) -> (b -> c) -> Either a b -> c
either f g (Left x) = f x
either f g (Right y) = g y
We should mention a minor but useful variation, which avoids the explicit attachment of tags when inserting into a subtype-polymorphic collection. We use a
special cons operation, consEither, which replaces the normal list constructor (:):
-- ... so far ...
let scribble = [ Left (rectangle 10 20 5 6)
, Right (circle 15 25 8)
]
-- ... liberalised notation ...
let scribble = consEither
(rectangle 10 20 5 6)
[circle 15 25 8]
-- A union-constructing cons operation
consEither :: h -> [t] -> [Either h t]
consEither h t@(_:_) = Left h : map Right t
consEither _ _ = error "Cannot cons with empty tail!"
Assessment of the encoding
• This approach is highly systematic and general. For instance, multiple inheritance is immediately possible. One may argue that this approach does
not directly encode OO class inheritance. Rather, it mimics object composition. One might indeed convert native OO programs, prior to encoding, so
that they do not use class inheritance, but they use interface polymorphism
combined with (manually coded) object composition instead.
• A fair amount of boilerplate code is required (cf. readShape and writeShape).
Also, each transitive subtype relationship requires surprising boilerplate. For
example, let us assume an OO class FooBar that is a subclass of Rectangle.
The transcription to Haskell would involve three instances: one for the type
class that is dedicated to FooBar (“Ok”), one for Rectangle (still “Ok” except
the scattering of implementation), and one for Shape (“annoying”).
Haskell’s overlooked object system
— 1 February 2008 —
27
• The union-type technique improved compared to Sec. 3.2. The top-level tagging scheme eliminates the need for tagging helpers that are specific to the
object types. Also, the consEither operation relieves us from the chore of
explicitly writing sequences of tags. However, we must assume that we insert
into a non-empty list, and we also must accept that the union type increases
for each new element in the list — no matter how many different element
types are encountered. Also, if we want to downcast from the union type, we
still need to know its exact layout. To lift these restrictions, we have to engage
into proper type-class-based programming.
3.6 Variation — existential quantification
So far we have restricted ourselves to Haskell 98. We now turn to common extensions
of Haskell 98, in an attempt to improve on the problems that we have encountered.
In upshot, we cannot spot obvious ways for improvement.
Our first attempt is to leverage existential quantification for the implementation of subtype-polymorphic collections. Compared to the earlier Either-based approach, we homogenise shapes by making them opaque (Cardelli & Wegner, 1985)
as opposed to embedding them into the union type. This use of existentials could
be combined with various object encodings; we illustrate it here for the specific
encoding from the previous section.
We define an existential envelope for shape data.
data OpaqueShape = forall x. Shape x => HideShape x
“Opaque shapes are (still) shapes.” Hence, a Shape instance:
instance Shape
where
readShape f
writeShape f
draw
OpaqueShape
(HideShape x) = readShape f x
(HideShape x) = HideShape $ writeShape f x
(HideShape x) = draw x
When building the scribble list, we place shapes in the envelope.
let scribble = [ HideShape (rectangle 10 20 5 6)
, HideShape (circle 15 25 8)
]
Assessment of the encoding
• Compared to the ‘union type’ approach, programmers do not have to invent
union types each time they need to homogenise different subtypes. Instead, all
shapes are tagged by HideShape. The ‘narrowing’ approach was quite similar,
but it required boilerplate.
• We face a new problem. Existential quantification limits type inference.
We see that in the definition of OpaqueShape; viz. the explicit constraint Shape. It
is mandatory to constraint the quantifier by all subtypes whose methods may be
invoked. A reader may notice the similar problem for the ‘union type’ approach,
which required Shape constraints in the instance
28
O. Kiselyov and R. Lämmel
instance (Shape a, Shape b) => Shape (Either a b) where ...
That instance, however, was merely a convenience. We could have disposed of it
and used the fold operation either explicitly in the scribble loop:
main =
do
let scribble = [ Left (rectangle 10 20 5 6)
, Right (circle 15 25 8)
]
mapM_ (either scribbleBody scribbleBody) scribble
where
scribbleBody x = do
draw x
draw (rMoveTo 100 100 x)
By contrast, the explicit constraint for the existential envelope cannot be eliminated. Admittedly, the loss of type inference is a nuance in this specific example. In
general, however, this weakness of existentials is quite annoying. It is intellectually
dissatisfying since type inference is one of the added values of an (extended) Hindley/Milner type system, when compared to mainstream OO languages. Worse than
that, the kind of constraints in the example are not necessary in mainstream OO
languages (without type inference), because these constraints deal with subtyping,
which is normally implicit.
We do not use existentials in OOHaskell.
3.7 Variation — heterogeneous collections
We continue with our exploration of common extensions of Haskell 98. In fact,
we will offer another option for the difficult problem of subtype-polymorphic collections. We recall that all previously discussed techniques aimed at making it
possible to construct a normal homogeneous Haskell list in the end. This time,
we will engage into the construction of a heterogeneous collection in the first
place. To this end, we leverage techniques as described by us in the HList
paper (Kiselyov et al., 2004). Heterogeneous collections rely on multi-parameter
classes (Chen et al., 1992; Jones, 1992; Jones, 1995; Peyton Jones et al., 1997) with
functional dependencies (Jones, 2000; Duck et al., 2004).
Heterogeneous lists are constructed with dedicated constructors HCons and HNil —
analogues of (:) and []. One may think of a heterogeneous list type as a nested
binary product, where HCons corresponds to (,) and HNil to (). We use special
HList functions to process the heterogeneous lists; the example requires a map
operation. The scribble loop is now encoded as follows:
main =
do
let scribble =
HCons (rectangle 10 20 5 6)
(HCons (circle 15 25 8)
HNil)
hMapM_ (undefined::ScribbleBody) scribble
Haskell’s overlooked object system
— 1 February 2008 —
29
The operation hMapM_ is the heterogeneous variation on the normal monadic map
mapM_. The function argument for the map cannot be given inline; instead we pass a
proxy undefined::ScribbleBody. This detour is necessary due to technical reasons
that are related to the combination of rank-2 polymorphism and type-class-bounded
polymorphism.6
The type code for the body of the scribble loop is defined by a trivial datatype:
data ScribbleBody -- No constructors needed; non-Haskell 98
The heterogeneous map function is constrained by the Apply class, which models
interpretation of function codes like ScribbleBody. Here is the Apply class and the
instance dedicated to ScribbleBody:
class Apply f a r | f a -> r
where apply :: f -> a -> r
instance Shape s => Apply ScribbleBody s (IO ())
where
apply _ x =
do
draw x
draw (rMoveTo 100 100 x)
Assessment of the encoding
• This approach eliminates all effort for inserting elements into a collection.
• The approach comes with heavy surface encoding; cf. type code ScribbleBody.
• This encoding is at odds with type inference — just as in the case of existentials. That is, the Apply instance must be explicitly constrained by the
interfaces that are going to be relied upon in the body of the scribble loop.
Again, the amount of explicit typing is not yet disturbing in the example at
hand, but it is an intrinsic weakness of the encoding. The sort of required
explicit typing goes beyond standard OO programming practise.
4 Type-agnostic OOHaskell idioms
We will now systematically develop all important OOHaskell programming idioms. In this section, we will restrict ourselves to ‘type-agnostic’ idioms, as to clearly
substantiate that most OOHaskell programming does not require type declarations, type annotations, explicit casts for object types — thanks to Haskell’s type
inference and its strong support for polymorphism. The remaining, ‘type-perceptive’
6
A heterogeneous map function can encounter entities of different types. Hence, its argument function must be polymorphic on its own (which is different from the normal
map function for lists). The argument function typically uses type-class-bounded polymorphic functions to process the entities of different types. The trouble is that the map
function cannot possibly anticipate all the constraints required by the different uses of
the map function. The type-code technique moves the constraints from the type of the
heterogeneous map function to the interpretation site of the type codes.
30
O. Kiselyov and R. Lämmel
OOHaskell idioms (including a few advanced topics related to subtyping) are described in the subsequent section.
In both sections, we adopt the following style. We illustrate the OO idioms and
describe the technicalities of encoding. We highlight strengths of OOHaskell:
support for the traditional OO idioms as well as extra features due to the underlying
record calculus, and first-class status of labels, methods and classes. Finally, we
illustrate the overall programmability of a typed OO language design in Haskell.
As a matter of style, we somewhat align the presentation of OOHaskell with the
OCaml object tutorial. Among the many OO systems that are based on open records
(Perl, Python, Javascript, Lua, etc.), OCaml stands out because it is statically
typed (just as OOHaskell). Also, OCaml (to be precise, its predecessor ML-ART)
is close to OOHaskell in terms of motivation: both aim at the introduction of
objects as a library in a strongly-typed functional language with type inference. The
implementation of the libraries and the sets of features used or required are quite
different (cf. Sec. 6.2.1 for a related work discussion), which makes a comparison
even more interesting. Hence, we draw examples from the OCaml object tutorial,
to specifically contrast OCaml and OOHaskell code and to demonstrate the fact
that OCaml examples are expressible in OOHaskell, roughly in the same syntax,
based on direct, local translation. We also use the OCaml object tutorial because
it is clear, comprehensive and concise.
4.1 Objects as records
Quoting from (Leroy et al. , 2004)[§ 3.2]:7
“The class point below defines one instance variable varX and two methods getX and
moveX. The initial value of the instance variable is 0. The variable varX is declared mutable.
Hence, the method moveX can change its value.”
class point =
object
val mutable varX = 0
method getX
= varX
method moveX d
= varX <- varX + d
end;;
4.1.1 First-class labels
The transcription to OOHaskell starts with the declaration of all the labels that
occur in the OCaml code. The HList library readily offers 4 different models of
labels. In all cases, labels are Haskell values that are distinguished by their Haskell
type. We choose the following model:
• The value of a label is “⊥”.
7
While quoting portions of the OCaml tutorial, we take the liberty to rename some identifiers
and to massage some subminor details.
Haskell’s overlooked object system
— 1 February 2008 —
31
• The type of a label is a proxy for an empty type (empty except for “⊥”).8
data VarX; varX = proxy :: Proxy VarX
data GetX; getX = proxy :: Proxy GetX
data MoveX; moveX = proxy :: Proxy MoveX
where proxies are defined as
data Proxy e
proxy :: Proxy e
proxy = ⊥
-- A proxy type is an empty phantom type.
-- A proxy value is just ‘‘⊥’’.
Simple syntactic sugar can significantly reduce the length of the one-liners for label
declaration should this become an issue. For instance, we may think of the above
lines just as follows:
-- Syntax extension assumed; label is a new keyword.
label varX
label getX
label moveX
The explicit declaration of OOHaskell labels blends well with Haskell’s scoping
rules and its module concept. Labels can be private to a module, or they can be
exported, imported, and shared. All models of HList labels support labels as firstclass citizens. In particular, we can pass them to functions. The “labels as type
proxies” idea is the basis for defining record operations since we can thereby dispatch
on labels in type-level functionality. We will get back to the record operations
shortly.
4.1.2 Mutable variables
The OCaml point class is transcribed to OOHaskell as follows:
point =
do
x <- newIORef 0
returnIO
$ varX .=. x
.*. getX .=. readIORef x
.*. moveX .=. (\d -> do modifyIORef x (+d))
.*. emptyRecord
The OOHaskell code clearly mimics the OCaml code. While we use Haskell’s
IORefs to model mutable variables, we do not use any magic of the IO
monad. We could as well use the simpler ST monad, which is very well formalised (Launchbury & Peyton Jones, 1995). The source distribution for the paper
illustrates the ST option.
8
It is a specific GHC extension of Haskell 98 to allow for datatypes without any constructor
declarations. Clearly, this is a minor issue because one could always declare a dummy constructor
that is not used by the program.
32
O. Kiselyov and R. Lämmel
The Haskell representation of the point class stands revealed as a value binding
declaration of a monadic record type. The do sequence first creates an IORef for the
mutable variable, and then returns a record for the new point object. In general,
the OOHaskell records provide access to the public methods of an object and to
the IORefs for public mutable variables. We will often call all record components of
OOHaskell’s objects just ‘methods’. In the example, varX is public, just as in the
original OCaml code. In OOHaskell, a private variable would be encoded as an
IORef that is not made available through a record component. (Private variables
were explored in the shapes example.)
4.1.3 HList records
We may ask Haskell to tell us the inferred type of point:
ghci> :t point
point :: IO (Record (HCons (Proxy MutableX, IORef Integer)
(HCons (Proxy GetX, IO Integer)
(HCons (Proxy Move, Integer -> IO ())
HNil))))
The type reveals the use of HList’s extensible records (Kiselyov et al., 2004). We
explain some details about HList, as to make the present paper self-contained. The
inferred type shows that records are represented as heterogeneous label-value pairs,
which are promoted to a proper record type through the type-constructor Record.
-- HList constructors
data HNil
= HNil
-- empty heterogeneous list
data HCons e l = HCons e l -- non-empty heterogeneous list
-- Sugar for forming label-value pairs
infixr 4 .=.
l .=. v = (l,v)
-- Record type constructor
newtype Record r = Record r
The constructor Record is opaque for the library user. Instead, the library user
(and most of the library code itself) relies upon a constrained constructor:
-- Record value constructor
mkRecord :: HRLabelSet r => r -> Record r
mkRecord = Record
The constraint HRLabelSet r statically assures that all labels are pairwise distinct
as this is a necessary precondition for a list of label-value pairs to qualify as a record.
(We omit the routine specification of HRLabelSet r (Kiselyov et al., 2004).) We can
now implement emptyRecord, which was used in the definition of point:
emptyRecord = mkRecord HNil
The record extension operator, (.*.), is a constrained variation on the heterogeneous cons operation, HCons: we need to make sure that the newly added label-value
pair does not violate the uniqueness property for the labels. This is readily expressed
by wrapping the unconstrained cons term in the constrained record constructor:
Haskell’s overlooked object system
— 1 February 2008 —
33
infixr 2 .*.
(l,v) .*. (Record r) = mkRecord (HCons (l,v) r)
4.1.4 OO test cases
We want to instantiate the point class and invoke some methods. We begin with
an OCaml session, which shows some inputs and the responses from the OCaml
interpreter:9
# let p = new point;;
val p : point = <obj>
# p#getX;;
- : int = 0
# p#moveX 3;;
- : unit = ()
# p#getX;;
- : int = 3
In Haskell, we capture this program in a monadic do sequence because method
invocations can involve IO effects including the mutation of objects. We denote
method invocation by (#), just as in OCaml; this operation is a plain record lookup. Hence:
myFirstOOP =
do
p <- point -- no need for new!
p # getX >>= Prelude.print
p # moveX $ 3
p # getX >>= Prelude.print
OOHaskell and OCaml agree:
ghci> myFirstOOP
0
3
For completeness we outline the definition of “#”:
-- Sugar operator
infixr 9 #
obj # feature = hLookupByLabel feature obj
-- Type-level operation for look-up
class HasField l r v | l r -> v
where
hLookupByLabel:: l -> r -> v
This operation performs type-level (and value-level) traversal of the label-value
pairs, looking up the value component for a given label from the given record, while
9
OCaml’s prompt is indicated by a leading character “#”.
Method invocation is modelled by the infix operator “#”.
The lines with leading “val” or “–” are the responses from the interpreter.
34
O. Kiselyov and R. Lämmel
using the label type as a key. We recall that the term ‘field’ (cf. HasField) originates from record terminology. In OOHaskell, all ‘fields’ are ‘methods’. (We omit
the routine specification of HasField l r v (Kiselyov et al., 2004).) The class declaration reveals that HList (and thereby OOHaskell) relies on multi-parameter
classes (Chen et al., 1992; Jones, 1992; Jones, 1995; Peyton Jones et al., 1997) with
functional dependencies (Jones, 2000; Duck et al., 2004).
4.2 Object generators
In class-based, mainstream OO languages, the construction of new class instances
is normally regulated by so-called constructor methods. In OOHaskell, instances
are created by a function that serves as an object generator. The function can be
seen as the embodiment of the class itself. The point computation defined above
is a trivial example of an object generator.
4.2.1 Constructor arguments
Quoting from (Leroy et al. , 2004)[§ 3.1]:
“The class point can also be abstracted over the initial value of varX. The parameter
x_init is, of course, visible in the whole body of the definition, including methods. For
instance, the method getOffset in the class below returns the position of the object relative
to its initial position.”
class para_point x_init =
object
val mutable varX = x_init
method getX
= varX
method getOffset = varX - x_init
method moveX d
= varX <- varX + d
end;;
In OOHaskell, objects are created as the result of monadic computations producing records. We can parameterise these computations by turning them into
functions, object generators, which take construction parameters as arguments.
For instance, the parameter x_init of the OCaml class para_point ends up as a
plain function argument:
para_point x_init
= do
x <- newIORef x_init
returnIO
$ varX
.=. x
.*. getX
.=. readIORef x
.*. getOffset .=. queryIORef x (\v -> v - x_init)
.*. moveX
.=. (\d -> modifyIORef x (+d))
.*. emptyRecord
4.2.2 Construction-time computations
Quoting from (Leroy et al. , 2004)[§ 3.1]:
Haskell’s overlooked object system
— 1 February 2008 —
35
“Expressions can be evaluated and bound before defining the object body of the class.
This is useful to enforce invariants. For instance, points can be automatically adjusted to
the nearest point on a grid, as follows:”
class adjusted_point x_init =
let origin = (x_init / 10) * 10 in
object
val mutable varX = origin
method getX
= varX
method getOffset = varX - origin
method moveX d
= varX <- varX + d
end;;
In OOHaskell, we follow the suggestion from the OCaml tutorial: we use local
let bindings to carry out the constructor computations “prior” to returning the
constructed object:
adjusted_point x_init
= do
let origin = (x_init ‘div‘ 10) * 10
x <- newIORef origin
returnIO
$ varX
.=. x
.*. getX
.=. readIORef x
.*. getOffset .=. queryIORef x (\v -> v - origin)
.*. moveX
.=. (\d -> modifyIORef x (+d))
.*. emptyRecord
That “prior” is not meant in a temporal sense: OOHaskell remains a non-strict
language, in contrast to OCaml.
4.2.3 Implicitly polymorphic classes
A powerful feature of OOHaskell is implicit polymorphism for classes. For instance, the class para_point is polymorphic with regard to the point’s coordinate — without our contribution. This is a fine difference between the OCaml
model and our OOHaskell transcription. In OCaml’s definition of para_point,
the parameter x_init was of the type int — because the operation (+) in OCaml
can deal with integers only. The OOHaskell points are polymorphic — a point’s
coordinate can be any Num-ber, for example, an Int or a Double. Here is an example
to illustrate that:
myPolyOOP =
do
p <- para_point (1::Int)
p’ <- para_point (1::Double)
p # moveX $ 2
p’ # moveX $ 2.5
p # getX >>= Prelude.print
p’ # getX >>= Prelude.print
The OOHaskell points are actually bounded polymorphic. The point coordinate may be of any type that implements addition. Until very recently, one could
36
O. Kiselyov and R. Lämmel
not express this in Java and in C#. Expressing bounded polymorphism in C++ is
possible with significant contortions. In (OO)Haskell, we did not have to do anything at all. Bounded polymorphism (aka, generics) is available in Ada95, Eiffel and
a few other languages. However, in those languages, the polymorphic type and the
type bounds must be declared explicitly. (There are ongoing efforts to add some
specific bits of type inference to new versions of mainstream OO languages.) In
(OO)Haskell, the type system infers the (bounded) polymorphism on its own, in
full generality.
The implicit polymorphism of OOHaskell does not injure static typing. If we confuse Ints and Doubles in the above code, e.g., by attempting “p # moveX $ 2.5”,then
we get a type error saying that Int is not the same as Double. In contrast, the poor
men’s implementation of polymorphic collections, e.g., in Java < 1.5, which up-casts
an element to the most general Object type when inserting it into the collection,
requires runtime-checked downcasts when accessing elements.
4.2.4 Nested object generators
Quoting from (Leroy et al. , 2004)[§ 3.1]:
“The evaluation of the body of a class only takes place at object creation time. Therefore,
in the following example, the instance variable varX is initialised to different values for
two different objects.”
let x0 = ref 0;;
class incrementing_point :
object
val mutable varX = incr x0; !x0
method getX
= varX
method moveX d
= varX <- varX + d
end;;
We test this new class at the OCaml prompt:
#
#
-
new incrementing_point#getX;;
: int = 1
new incrementing_point#getX;;
: int = 2
The variable x0 can be viewed as a “class variable”, belonging to a class object.
Recall that classes are represented by object generators in OOHaskell. Hence to
build a class object we need a nested object generator:
incrementing_point =
do
x0 <- newIORef 0
returnIO (
do modifyIORef x0
x <- readIORef
returnIO
$ varX .=.
.*. getX .=.
(+1)
x0 >>= newIORef
x
readIORef x
Haskell’s overlooked object system
— 1 February 2008 —
37
.*. moveX .=. (\d -> modifyIORef x (+d))
.*. emptyRecord)
We can nest generators to any depth since we just use normal Haskell scopes. In the
example, the outer level does the computation for the point template (i.e., “class”);
the inner level constructs points themselves. Here is a more suggestive name for the
nested generator:
makeIncrementingPointClass = incrementing_point
This (trivial) example demonstrates that classes in OOHaskell are really firstclass citizens. We can pass classes as arguments to functions and return them as
results. In the following code fragment, we create a class in a scope, and bind it as
a value to a locally-scoped variable, which is then used to instantiate the created
class in that scope. The localClass is a closure over the mutable variable x0.
myNestedOOP =
do
localClass <- makeIncrementingPointClass
localClass >>= ( # getX ) >>= Prelude.print
localClass >>= ( # getX ) >>= Prelude.print
ghci> myNestedOOP
1
2
In contrast, such a class closure is not possible in Java, let alone C++. Java
supports anonymous objects, but not anonymous first-class classes. Nested classes
in Java must be linked to an object of the enclosing class. Named nested classes in
C# are free from that linking restriction. However, C# does not support anonymous
classes or class computations in a local scope (although anonymous delegates of C#
let us emulate computable classes). Nevertheless, classes, as such, are not first-class
citizens in any mainstream OO language.
4.2.5 Open recursion
The methods of an object may send messages to ‘self’. To support inheritance with
override that ‘self’ must be bound explicitly (Cook, 1989). Otherwise, inheritance
will not able to revise the messages to self that were coded in a superclass. Consequently, general object generators are to be given in the style of ‘open recursion’:
they take self and construct (some part of) self.
Quoting from (Leroy et al. , 2004)[§ 3.2]:
“A method or an initialiser can send messages to self (that is, the current object). For
that, self must be explicitly bound, here to the variable s (s could be any identifier, even
though we will often choose the name self.) ... Dynamically, the variable s is bound at
the invocation of a method. In particular, when the class printable_point is inherited,
the variable s will be correctly bound to the object of the subclass.”
class printable_point x_init =
object (s)
val mutable varX = x_init
38
O. Kiselyov and R. Lämmel
method getX
method moveX d
method print
end;;
= varX
= varX <- varX + d
= print_int s#getX
Again, this OCaml code is transcribed to OOHaskell very directly. The self argument, s, ends up as an ordinary argument of the monadic function for generating
printable point objects:
printable_point x_init s =
do
x <- newIORef x_init
returnIO
$ varX .=. x
.*. getX .=. readIORef x
.*. moveX .=. (\d -> modifyIORef x (+d))
.*. print .=. ((s # getX ) >>= Prelude.print)
.*. emptyRecord
In OCaml, we use printable_point as follows:
# let p = new printable_point 7;;
val p : printable_point = <obj>
# p#moveX 2;;
- : unit = ()
# p#print;;
9- : unit = ()
Although s does not appear on the line that constructs a point p with the new
construct, the recursive knot clearly is tied right there. In OOHaskell, we use the
(monadic) fixpoint function, mfix, rather than a special keyword new. This makes
the nature of openly recursive object generators manifest.
mySelfishOOP =
do
p <- mfix (printable_point 7)
p # moveX $ 2
p # print
ghci> mySelfishOOP
9
4.2.6 Instantiation checking
One potential issue with open recursion in OOHaskell is that some type errors
in messages to self will not be spotted until the first object construction is coded.
For instance, an OO library developer could, accidentally, provide object generators
that turn out to be uninstantiatable; the library user would notice this defect once
the generators are put to work. This issue is readily resolved as follows. When we
program object generators, we may use the concrete operation:
-- An assured printable point generator
concrete_printable_point x_init
= concrete $ printable_point x_init
Haskell’s overlooked object system
— 1 February 2008 —
39
-- The concrete operation
concrete generator self = generator self
where
_ = mfix generator
Operationally, concrete is the identity function. However, it constrains the type
of generator such that the application of mfix is typeable. This approach needs
to be slightly refined to cover abstract methods (aka pure virtual methods). To
this end, one would need to engage into local inheritance — adding vacuous (i.e.,
potentially undefined) methods for any needed virtual method. This generalised
concrete operation would take the virtual portion of a record, or preferably just a
proxy for it, so that the purpose of this argument is documented.
4.3 Reuse techniques
The first-class status of labels, methods and classes enables various, common and
advanced forms of reuse. Single inheritance boils down to (monadic) function composition of object generators. Multiple inheritance and object composition employ
more advanced operations of the record calculus.
4.3.1 Single inheritance with extension
Quoting from (Leroy et al. , 2004)[§ 3.7]:10
“We illustrate inheritance by defining a class of colored points that inherits from the
class of points. This class has all instance variables and all methods of class point, plus
a new instance variable color, and a new method getColor.”
class colored_point x (color : string) =
object
inherit point x
val color = color
method getColor = color
end;;
Here is the corresponding OCaml session:
# let p’ = new colored_point 5 "red";;
val p’ : colored_point = <obj>
# p’#getX, p’#getColor;;
- : int * string = (5, "red")
The following OOHaskell version does not employ a special inherit construct.
We compose computations instead. To construct a colored point we instantiate the
superclass while maintaining open recursion, and extend the intermediate record,
super, by the new method getColor.
10
We use British spelling consistently in this paper, except for some words that enter the text
through code samples: color, colored, ...
40
O. Kiselyov and R. Lämmel
colored_point x_init (color::String) self =
do
super <- printable_point x_init self
returnIO
$ getColor .=. (returnIO color)
.*. super
Here, super is just a variable rather than an extra construct.
myColoredOOP =
do
p’ <- mfix (colored_point 5 "red")
x <- p’ # getX
c <- p’ # getColor
Prelude.print (x,c)
OOHaskell and OCaml agree:
ghci> myColoredOOP
(5,"red")
4.3.2 Class-polymorphic functionality
We can parameterise computations with respect to classes.
myFirstClassOOP point_class =
do
p <- mfix (point_class 7)
p # moveX $ 35
p # print
ghci> myFirstClassOOP printable_point
42
The function myFirstClassOOP takes a class (i.e., an object generator) as an argument, instantiates the class, and moves and prints the resulting object. We can
pass myFirstClassOOP any object generator that creates an object with the slots
moveX and print. This constraint is statically verified. For instance, the colored
point class, which we derived from the printable point class in the previous section,
is suitable:
ghci> myFirstClassOOP $ flip colored_point "red"
42
4.3.3 Single inheritance with override
We can override methods and still refer to their superclass implementations (akin
to the super construct in OCaml and other languages). We illustrate overriding
with a subclass of colored_point whose print method is more informative:
colored_point’ x_init (color::String) self =
do
super <- colored_point x_init color self
return $ print .=. (
do putStr "so far - "; super # print
putStr "color - "; Prelude.print color )
.<. super
Haskell’s overlooked object system
— 1 February 2008 —
41
The first step in the monadic do sequence constructs an old-fashioned colored point,
and binds it to super for further reference. The second step in the monadic do
sequence returns super updated with the new print method. The HList operation
“.<.” denotes type-preserving record update as opposed to the familiar record
extension “.*.”. The operation “.<.” makes the overriding explicit (as it is in C#,
for example). We could also use a hybrid record operation, which does extension
in case the given label does not yet occur in the given record, falling back to typepreserving update. This hybrid operation would let us model the implicit overriding
in C++ and Java. Again, such a variation point demonstrates the programmability
of OOHaskell’s object system.
Here is a demo that shows overriding to properly affect the print method:
myOverridingOOP =
do
p <- mfix (colored_point’ 5 "red")
p # print
ghci> myOverridingOOP
so far - 5
color - "red"
4.3.4 Orphan methods
We can program methods outside of any hosting class. Such methods can be reused
across classes without any inheritance relationship. For instance, we may define a
method print_getX that can be shared by all objects that have at least the method
getX of the type Show a => IO a — regardless of any inheritance relationships:
print_getX self = ((self # getX ) >>= Prelude.print)
We can update the earlier code for printable_point as follows:
-- before: inlined definition of print
... .*. print
.=. ((s # getX ) >>= Prelude.print)
-- after: reusable orphan method
... .*. print
.=. print_getX s
4.3.5 Flexible reuse schemes
In addition to single class inheritance, there are several other established reuse
schemes in OO programming including object composition, different forms of mixins and different forms of multiple inheritance. Given the first-class status of all
OOHaskell entities and its foundation in a powerful record calculus, it should be
possible to reconstruct most if not all existing reuse schemes. We will use an (admittedly contrived) example to demonstrate a challenging combination of multiple
inheritance and object composition. To the best of our knowledge, this example
cannot be directly represented in any existing mainstream language.
We are going to work through a scenario of making a class heavy_point from
three different concrete subclasses of abstract_point. The first two concrete points
42
O. Kiselyov and R. Lämmel
Fig. 2. A complex reuse scenario
will be shared in the resulting heavy point, because we leave open the recursive
knot. The third concrete point does not participate in the open recursion and is
not shared. In C++ terminology, abstract_point is a virtual base class (with
respect to the first two concrete points) and a non-virtual base class at the same
time. See Fig. 2 for an overview.
The object template for heavy points starts as follows:
heavy_point x_init color self =
do super1 <- concrete_point1 x_init self
super2 <- concrete_point2 x_init self
super3 <- mfix (concrete_point3 x_init)
... -- to be continued
We bind all ancestor objects for subsequent reference. We pass self to the first two
points, which participate in open recursion, but we fix the third point in place. The
first two classes are thus reused in the sense of inheritance, while the third class is
reused in the sense of object composition. A heavy point carries print and moveX
methods delegating corresponding messages to all three points:
... -- continued from above
let myprint = do
putStr "super1: "; (super1
putStr "super2: "; (super2
putStr "super3: "; (super3
let mymove = ( \d -> do
super1 # moveX $ d
super2 # moveX $ d
super3 # moveX $ d
return
$
print .=. myprint
.*.
moveX .=. mymove
.*.
emptyRecord
... -- to be continued
# print)
# print)
# print)
)
Haskell’s overlooked object system
— 1 February 2008 —
43
The three points, with all their many fields and methods, contribute to the heavy
point by means of left-biased union on records, which is denoted by “.<++.” below:
... -.<++.
.<++.
.<++.
continued from above
super1
super2
super3
Here is a demo:
myDiamondOOP =
do
p <- mfix (heavy_point 42 "blue")
p # print -- All points still agree!
p # moveX $ 2
p # print -- The third point lacks behind!
ghci> myDiamondOOP
super1: 42
super2: 42
super3: 42
super1: 46
super2: 46
super3: 44
For comparison, in OCaml, multiple inheritance follows fixed rules. Only the
last definition of a method is kept: the redefinition in a subclass of a method that
was visible in the parent class overrides the definition in the parent class. Previous
definitions of a method can be reused by binding the related ancestor using a special
. . . as . . . notation. The bound name is said to be a ‘pseudo value identifier’ that can
only be used to invoke an ancestor method. Eiffel, C++, etc. have their own fixed
rules and notations for multiple inheritance. OOHaskell allows us to “program”
this aspect of the OO type system. Programmers (or language designers) may devise
their own inheritance and object composition rules.
4.4 Safe value recursion
The support for open recursion in an OO system has a subtle but fundamental
difficulty. Of the three ways to emulate open recursion – recursive types, existential
abstraction ((Rémy, 1994; Pierce & Turner, 1994)) and value recursion, the latter
is the simplest one. This is the one we have chosen for OOHaskell. Recall that
each object generator receives the self argument (representing the constructed
object), which lets the methods send the messages to the object itself. An object is
constructed by obtaining the fixpoint of the generator. Here is a variation on the
printable point example from Sec. 4.2.5 that illustrates the potential unsafety of
value recursion:
printable_point x_init self =
do
x <- newIORef x_init
self # print -- Unsafe!
44
O. Kiselyov and R. Lämmel
returnIO
$ varX .=. x
.*. getX .=. readIORef x
.*. moveX .=. (\d -> modifyIORef x (+d))
.*. print .=. ((self # getX ) >>= Prelude.print)
.*. emptyRecord
An object generator may be tempted to invoke methods on the received self argument, as self # print above. That code typechecks. However, the attempt to
construct an object by executing
mfix (printable_point 0)
reveals the problem: looping. Indeed, self represents the object that will be constructed. It is not proper to invoke any method on self when the object generation
is still taking place, because self as a whole does not yet exist.
In Haskell, accessing a not-yet-constructed object leads to “mere” looping. This,
so-called left-recursion problem (well-known in parsing (Aho et al., 1986)) is akin
to the divergence of the following trivial expression:
mfix (\self -> do { Prelude.print self; return "s" })
In a non-strict language like Haskell, determining the fixpoint of a value a->a
where a is not a function type is always safe: the worst can happen is the divergence,
but no undefined behaviour. In strict languages, the problem is far more serious:
accessing the field before it was filled in is accessing a dummy value (e.g., a null
pointer) that was placed into the field prior to the evaluation of the recursive
definition. Such an access results in undefined behaviour, and has to be prevented
with a run-time check. As noted in (Rémy, 1994), this problem has been widely
discussed but no satisfactory solution was found.
Although the problem is relatively benign in OOHaskell and never leads to
undefined behaviour, we would like to statically prevent it. To be precise, we impose the rule that the constructor may not execute any actions that involve notyet-constructed objects. With little changes, we statically enforce that restriction.
Object construction may be regarded as a sort of a staged computation; the problem of preventing the use of not-yet-constructed values is one of the key challenges
in multi-staged programming (Taha & Nielsen, 2003), where it has been recently
solved with environment classifiers. Our solution is related in principle (making the
stage of completion of an object a part of its type), but differs in technique (we
exploit compile-time tagging and monadic types rather than higher-ranked types).
We introduce a tag NotFixed to mark the objects that are not constructed yet:
newtype NotFixed a = NotFixed a -- data constructor opaque!
Because NotFixed is a newtype, this tag imposes no run-time overhead. We do not
export the data constructor NotFixed so the user may not arbitrarily introduce
or remove that tag. All operations on this tag are restricted to a new module
NotFixed that is part of the OOHaskell library. The module exports two new
operations: new and construct. The former is a variant of mfix for the IO monad.
The construct operation is a variation on returnIO.
Haskell’s overlooked object system
— 1 February 2008 —
45
new :: (NotFixed (Record a)
-> IO (NotFixed (Record a))) -- object generator
-> IO (Record a)
-- object computation
new f = mfix f >>= (\(NotFixed a) -> return a)
construct :: NotFixed (Record a)
-> (Record a -> Record b)
-> IO (NotFixed (Record b))
construct (NotFixed self) f = returnIO
-- self
-- constructor
-- object computation
$ NotFixed (f self)
Staged object construction proceeds as follows. The argument self passed to the
object generator is marked as NotFixed. After the fixpoint is computed, new removes the NotFixed tag. The function construct, while maintaining the NotFixed
tag, lifts the tag internally so that the methods being defined by the object generator could use the self reference. We can now write our example as follows:
printable_point x_init self =
do
x <- newIORef x_init
-- self # print
construct self (\self->
mutableX .=. x
.*. getX
.=. readIORef x
.*. moveX
.=. (\d -> modifyIORef x ((+) d))
.*. print
.=. ((self # getX ) >>= Prelude.print)
.*. emptyRecord)
test_pp
do
p
p
p
=
<- new (printable_point 7)
# moveX $ 2
# print
If we uncomment the statement self # print we will get the type error saying
that a NotFixed object does not have the method print. (There are no HasField
instances for the NotFixed type.) Within the body of construct, the reference to
self is available without the NotFixed tag; so one may be tempted to invoke methods on self and execute their actions. However, the second argument of construct
is a non-monadic function of the type Record a -> Record b. Because the result
type of the function does not include IO, it is not possible to read and write IORef
and do other IO actions within that function. In Haskell (in contrast to OCaml),
imperativeness of a function is manifest in its type.
The extension to the construction of inherited classes is straightforward. For
example, the colored_point example from Sec. 4.3.1 now reads:
colored_point x_init (color::String) self =
do
p <- printable_point x_init self
-- p # print -- would not typecheck.
construct p $ \p -> getColor .=. (returnIO color) .*. p
myColoredOOP =
do
46
O. Kiselyov and R. Lämmel
p’ <- new (colored_point 5 "red")
x <- p’ # getX
c <- p’ # getColor
Prelude.print (x,c)
The constructor colored_point receives the argument self marked as not-yetconstructed. We pass that argument to printable_point, which gives us a notyet-constructed object of the superclass. We cannot execute any methods on that
object (and indeed, uncommenting the statement p # print leads to a type error).
The execution of a superclass method may involve the invocation of a method
on self (as is the case for the method print), and self is not constructed yet.
The construct operation shown here is not fully general; the source distribution
illustrates safe object generation where methods refer both to self and super. The
technique readily generalises to multiple inheritance and object composition.
5 Type-perceptive OOHaskell idioms
So far we have avoided type declarations, type annotations and explicit coercions of
object types. We will now discuss those OOHaskell programming scenarios that
can benefit from additional type information, or even require it. We will pay special
attention to various subtyping and related cast and variance issues. In particular,
we will cover the technical details of subtype-polymorphic collections, which require
some amount of type perceptiveness, as we saw in the introductory shapes example
in Sec. 2.
5.1 Semi-implicit upcasting
There is an important difference between OOHaskell’s subtype polymorphism
(which we share to some extent with OCaml and ML-ART) and polymorphism in
C++ and other mainstream OO languages.11 In the latter languages, an object can
be implicitly coerced to an object of any of its superclasses (“upcast”). One may
even think that an object is polymorphic by itself, i.e., it has types of all of its
superclasses, simultaneously. Hence, there is no need for functions on objects (or
methods) to be polymorphic by themselves; they are monomorphic.
In OCaml and OOHaskell, it is the other way around: objects are monomorphic
(with regard to the record structure) and the language semantics does not offer
any implicit upcasting or narrowing.12 However, functions that take objects can be
polymorphic and can process objects of different types. To be precise, OOHaskell
exploits type-class-bounded polymorphism. A function that takes an object and
refers to its methods (i.e., record components) has in its inferred or explicit type
HasField constraints for these record components. The function therefore accepts
11
12
See, however, Sec. 5.7, where we emulate the mainstream nominal subtyping.
We prefer the term narrow over up-cast, as to emphasise the act of restricting the interface of
an object, as opposed to walking up an explicit (perhaps even nominal) subtyping hierarchy.
Haskell’s overlooked object system
— 1 February 2008 —
47
any object that has at least the components that satisfy the HasField constraints.13
Therefore, most of the time no (implicit or explicit) upcasting is needed; in fact, in
Sec. 4 we did not see any.
An explicit cast is usually understood as casting to an explicitly named target
type. We discuss such casts later in this section. Here we show that the established
explicit-vs.-implicit upcast dichotomy misses an intermediate option, which is admitted by OOHaskell. Namely, OOHaskell lets the programmer specify that
narrowing is to be performed — without giving an explicit target type, though. So
we continue to get by without specifying types, leaving it all to type inference — at
least for a while.
In OOHaskell (and OCaml), we must narrow an object if its expression context
has no type-class-bounded polymorphism left and requires an object of a different
type. The archetypal example is placing objects in a homogeneous collection, e.g., a
list. The original item objects may be of different types; therefore, we must establish
a common element type and narrow the items to it. This common element type does
not have to be specified explicitly, however. OOHaskell can compute that common
type as we add objects to the collection; the context will drive the narrowing. The
OOHaskell implementation of the shapes example in Sec. 2 involved this sort of
narrowing:
myOOP = do
s1 <- mfix (rectangle (10::Int) (20::Int) 5 6)
s2 <- mfix (circle (15::Int) 25 8)
let scribble = consLub s1 (consLub s2 nilLub)
... and so on ...
The designated list constructors nilLub and consLub incorporate narrowing into
their normal constructor behaviour. The specific element type of each new element
constraints the ultimate least-upper bound (LUB) element type for the final list.
Elements are continuously cast towards this LUB. The list constructors are defined
as follows:
-- A type-level code for the empty list
data NilLub
-- The empty list constructor
nilLub = ⊥ :: NilLub
-- Cons as a type-level function
class ConsLub h t l | h t -> l
where
consLub :: h -> t -> l
-- No coercion needed for a singleton list
instance ConsLub e NilLub [e]
where
consLub h _ = [h]
13
We oversimplify here by not talking about operations that add or remove fields. This is a
fair simplification though because we talk about normal OO functionality here, as opposed to
free-wheeling functionality for record manipulation.
48
O. Kiselyov and R. Lämmel
-- Narrow head and tail to their LUB type
instance LubNarrow e0 e1 e2 => ConsLub e0 [e1] [e2]
where
consLub h t = fst (head z) : map snd (tail z)
where
z = map (lubNarrow h) (⊥:t)
The important operation is lubNarrow:
class LubNarrow a b c | a b -> c
where lubNarrow :: a -> b -> (c,c)
Given two values of record types a and b, this operation returns a pair of narrowed
values, both of type c, where c is supposed to be the least-upper bound in the sense
of structural subtyping. The specification of lubNarrow once again illustrates the
capability of OOHaskell to ‘program’ OO type-system aspects. We exploit the
type-level reflection on HList records to define narrowing:
instance ( HZip la va a
, HZip lb vb b
, HTIntersect la lb lc
, H2ProjectByLabels lc a c aout
, H2ProjectByLabels lc b c bout
, HRLabelSet c
)
=> LubNarrow (Record a) (Record b) (Record c)
where
lubNarrow ra@(Record a) rb@(Record b) =
( hProjectByLabels lc ra
, hProjectByLabels lc rb
)
where
lc = hTIntersect la lb
(la,_) = hUnzip a
(lb,_) = hUnzip b
That is, given two records ra and rb, we compute the intersection lc of their labels
la and lb such that we can subsequently project both records to this shared label
set. It is possible to improve consLub so that we can construct lists in linear time.
We may also want to consider depth subtyping in addition to width subtyping, as
we will discuss in Sec. 5.9.
5.2 Narrow to a fixed type
The LUB narrowing is neither an explicit nor an implicit coercion. In the shapes
example, we explicitly apply special list constructors, which we know perform coercion, but the target type is left implicit. Such semi-implicit narrowing is a feature of
OOHaskell, not available in the otherwise similar systems OCaml and ML-ART.
In OCaml, building the scribble list in the shapes example requires fully explicit
narrowing (which OCaml calls upcast, “:>”):
Haskell’s overlooked object system
— 1 February 2008 —
49
let (scribble: shape list) = [
(new rectangle 10 20 5 6 :> shape);
(new circle 15 25 8 :> shape)] in ...
We can express such narrowing in OOHaskell as well:
s1 <- mfix (rectangle (10::Int) (20::Int) 5 6)
s2 <- mfix (circle (15::Int) 25 8)
let scribble :: [Shape Int]
scribble = [narrow s1, narrow s2]
The applications of narrow prepare the shape objects for insertion into the homogeneous Haskell list. We do not need to identify the target type per element:
specifying the desired type for the result list is enough. The operation narrow is
defined in a dedicated class:
class Narrow a b
where narrow :: Record a -> Record b
The operation narrow extracts those label-value pairs from a that are requested by
b. Its implementation uses the same kind of projection on records that we saw in
full in the previous section; cf. lubNarrow.
(Fully) explicit narrowing implies that we must declare appropriate types —
something that we managed to avoid so far. Here is the Shape type, which ‘drives’
narrowing in the example:
-- The Shape interface
type Shape a = Record (
:*:
:*:
:*:
:*:
:*:
:*:
:*:
GetX
GetY
SetX
SetY
MoveTo
RMoveTo
Draw
HNil )
:=:
:=:
:=:
:=:
:=:
:=:
:=:
IO
IO
(a
(a
(a
(a
IO
a
a
->
->
->
->
()
IO ())
IO ())
a -> IO ())
a -> IO ())
Two infix type synonyms add convenience to explicitly written types:
infixr 2 :*:
type e :*: l = HCons e l
infixr 4 :=:
type l :=: v = (l,v)
The Shape interface above explicitly includes the virtual operation draw because
the loop over scribble needs this method. We will see more applications of narrow
in subsequent sections.
5.3 Self-returning methods
A self-returning method is a method whose result type is the type of self or
is based on it. An example is a clone method. The typing of such methods (and of self) is known to be a difficult issue in typed object encodings;
50
O. Kiselyov and R. Lämmel
cf. (Cook et al., 1990; Abadi & Cardelli, 1996; Bruce et al., 2003) for some advanced treatments. In OOHaskell, we must not naively define a method me that
returns self, as is:
self_returning_point (x_init::a) self =
do
super <- printable_point x_init self
returnIO
$ me .=. self -- WRONG!
.*. super
If we wrote such code, then we get a type error, when we attempt to instantiate the
corresponding object (i.e., when we mfix the object generator). Haskell does not
permit (equi-)recursive types, which are needed to type self in the example. The
issue of recursive types and returning the full self is discussed in detail in Sec. 5.8.
Here, we point out a simpler solution: disallowing returning self and requiring the
programmer to narrow self to a specific desired interface. In the case of the clone
method, mainstream programming languages typically define its return type to be
the base class of all classes. The programmer is then supposed to use downcast to
the intended subtype.
We resolve the problem with the self-returning method as follows:
self_returning_point (x_init::a) self =
do
super <- printable_point x_init self
returnIO
$ me .=. (narrow self :: PPInterface a)
.*. super
type PPInterface a
= Record ( GetX :=: IO a
:*: MoveX :=: (a -> IO ())
:*: Print :=: IO ()
:*: HNil )
That is, me narrows self explicitly to the interface for printable points.
We should relate the explicit narrowing of the return type of me to the explicit
declaration of the return type of all methods in C++ and Java. The presented
narrowing approach does have a limitation however: all record components that
do not occur in the target interface are irreversibly eliminated from the result
record. We would prefer to make these components merely ‘private’ so they can
be recovered through a safe downcast. We offer two options for such downcastable
upcasts in the next two sections.
5.4 Casts based on dynamics
Turning again to the shapes benchmark, let us modify the loop over scribble,
a homogeneous list of shapes, so to single out circles for special treatment. This
requires downcast:
Haskell’s overlooked object system
— 1 February 2008 —
51
mapM_ (\shape -> maybe (putStrLn "Not a circle.")
(\circ -> do circ # setRadius $ 10;
circ # draw)
((downCast shape) ‘asTypeOf‘ (Just s2)))
scribble
In each iteration, we attempt to downcast the given shape object to the type of
s2 (which we recall is a circle object). A downcast may fail or succeed, hence the
Maybe type of the result.
Neither OOHaskell’s narrow operation nor OCaml’s upcast support such scenarios. OOHaskell’s narrow irrevocably removes record components. We can
define, however, other forms of upcast, which are reversible. We begin with a
technique that exploits dynamic typing (Abadi et al., 1989; Abadi et al., 1991;
Lämmel & Peyton Jones, 2003).
The new scribble list is built as follows:
let scribble :: [UpCast (Shape Int)]
scribble = [upCast s1, upCast s2]
where
data UpCast x = UpCast x Dynamic
The data constructor UpCast is opaque for the library user, who can only upcast through a dedicated upCast operation. The latter saves the original object by
embedding it into Dynamic. (We presume that record types readily instantiate the
type class Typeable.) Dually, downcast is then a projection from Dynamic to the
requested subtype:
upCast :: (Typeable (Record a), Narrow a b)
=> Record a -> UpCast (Record b)
upCast x = UpCast (narrow x) (toDyn x)
downCast :: (Typeable b, Narrow b a)
=> UpCast (Record a) -> Maybe (Record b)
downCast (UpCast _ d) = fromDynamic d
We want to treat ‘upcast objects’ as being objects too, and so we add a trivial HasField instance for looking up record components of upcast objects. This
instance delegates the look-up to the narrowed part of the UpCast value:
instance HasField l x v => HasField l (UpCast x) v
where
hLookupByLabel l (UpCast x _) = hLookupByLabel l x
This technique suffers from a few shortcomings. Although downcast is safe in a
sense that no ‘bad things can happen’ (cf. unsafe casts in C), this downcast does not
keep us from attempting so-called ‘stupid casts’, i.e., casts to types for which casting
cannot possibly succeed. In the following section, we describe a more elaborate upcast/downcast pair that statically prevents stupid downcasts. The dynamics-based
method also suffers from the full computational overhead of the narrow operation,
a value-level coercion that iterates over all record components.
52
O. Kiselyov and R. Lämmel
5.5 Casts based on unions
We turn to the subtyping technique from Sec. 3.4 (further refined in Sec. 3.5), which
used union types to represent the intersection of types. That techniques had several
problems: it could not easily deal with the empty list, could not minimise the union
type to the number of distinct element types, and could not downcast. We fully lift
these restrictions here by putting type-level programming to work.
We again make upcasts semi-implicit with dedicated list constructors:
myOOP = do
s1 <- mfix (rectangle (10::Int) (20::Int) 5 6)
s2 <- mfix (circle (15::Int) 25 8)
let scribble = consEither s1 (consEither s2 nilEither)
... and so on ...
The list constructors are almost identical to nilLub and consLub in Sec. 5.1. The
difference comes when we cons to a non-empty list; see the last instance below:
-- A type-level code for the empty list
data NilEither
-- The empty list constructor
nilEither = ⊥ :: NilEither
-- Cons as a trivial type-level function
class ConsEither h t l | h t -> l
where
consEither :: h -> t -> l
-- No coercion needed for a singleton list
instance ConsEither e NilEither [e]
where
consEither h _ = [h]
-- Construct union type for head and tail
instance ConsEither e1 [e2] [Either e1 e2]
where
consEither h t = Left h : map Right t
We extend the union type for the ultimate element type with one branch for each
new element, just as we did in the Haskell 98-based encoding of Sec. 3.5. However,
with type-level programming, we can, in principle, minimise the union type so that
each distinct element type occurs exactly once.14 This straightforward optimisation
is omitted here for brevity.15
Method look-up is generic, treating the union type as the intersection of record
fields of the union branches:
14
15
The same kind of constraints was covered in the HList paper (Kiselyov et al., 2004), cf. typeindexed heterogeneous collections.
In essence, we need to iterate over the existing union type and use type-level type equality to
detect if the type of the element to cons has already occurred in the union. If so, we also need
to determine the corresponding sequence of Left and Right tags.
Haskell’s overlooked object system
instance (HasField
=> HasField
where
hLookupByLabel l
hLookupByLabel l
— 1 February 2008 —
53
l x v, HasField l y v)
l (Either x y) v
(Left x) =
(Right y) =
hLookupByLabel l x
hLookupByLabel l y
Downcast is a type-driven search operation on the union type. We also want
downcast to fail statically if the target types does not appear among the branches.
Hence, we start downcast with a type-level Boolean, hFalse, to express that we
have not yet seen the type in question:
downCast = downCastSeen hFalse
Downcast returns Maybe because it can intrinsically fail at the value level:
class DownCastSeen seen u s
where downCastSeen :: seen -> u -> Maybe s
We process the union like a list. Hence, there are two cases: one for the nonsingleton union, and one for the final branch. Indeed, the details of the definition
reveal that we assume right-associative unions.
instance (DownCastEither seen b x y s, TypeEq x s b)
=> DownCastSeen seen (Either x y) s
where
downCastSeen seen = downCastEither seen (⊥::b)
instance (TypeCastSeen seen b x s, TypeEq x s b)
=> DownCastSeen seen x s
where
downCastSeen seen = typeCastSeen seen (⊥::b)
In both cases we test for the type equality between the target type and the (left)
branch type. We pass the computed type-level Boolean to type-level functions
DownCastEither (‘non-singleton union’) and TypeCastSeen (‘final branch’, a singleton union):
class TypeCastSeen seen b x y
where typeCastSeen :: seen -> b -> x -> Maybe y
instance TypeCast x y => TypeCastSeen seen HTrue x y
where
typeCastSeen _ _ = Just . typeCast
instance TypeCastSeen HTrue HFalse x y
where
typeCastSeen _ _ = const Nothing
The first instance applies when we have encountered the requested type at last.
In that case, we invoke normal, type-level type cast (cf. (Kiselyov et al., 2004)),
knowing that it must succeed given the earlier check for type equality. The second
instance applies when the final branch is not of the requested type. However, we
must have seen the target type among the branches, cf. HTrue. Thereby, we rule
out stupid casts.
The following type-level function handles ‘non-trivial’ unions:
class DownCastEither seen b x y s
where downCastEither :: seen -> b -> Either x y -> Maybe s
54
O. Kiselyov and R. Lämmel
instance (DownCastSeen HTrue y s, TypeCast x s)
=> DownCastEither seen HTrue x y s
where
downCastEither _ _ (Left x) = Just (typeCast x)
downCastEither _ _ (Right y) = downCastSeen hTrue y
instance DownCastSeen seen y s
=> DownCastEither seen HFalse x y s
where
downCastEither _ _
(Left x) = Nothing
downCastEither seen _ (Right y) = downCastSeen seen y
The first instances applies in case the left branch of the union type is of the target
type; cf. HTrue. It remains to check the value-level tag. If it is Left, we are done
after the type-level type cast. We continue the search otherwise, with seen set
to HTrue to record that the union type does indeed contain the target type. The
second instance applies in case the left branch of the union type is not of the target
type; cf. HFalse. In that case, downcast continues with the tail of union type, while
propagating the seen flag. Thereby, we rule out stupid casts.
5.6 Explicit type constraints
In some cases, it is useful to impose structural record type constraints on arguments
of an object generator, on arguments or the result type of a method. These constraints are akin to C++ concepts (Siek et al., 2005). The familiar narrow turns out
to be a convenient tool for imposition of such type constraints. This use of narrow
does no operations at run-time. A good example example of OO type constraints
is the treatment of virtual methods in OOHaskell.
Quoting from (Leroy et al. , 2004)[§ 3.4]:
“It is possible to declare a method without actually defining it, using the keyword virtual.
This method will be provided later in subclasses. A class containing virtual methods must be
flagged virtual, and cannot be instantiated (that is, no object of this class can be created).
It still defines type abbreviations (treating virtual methods as other methods.)
class virtual abstract_point x_init =
object (self)
val mutable varX = x_init
method print = print_int self#getX
method virtual getX : int
method virtual moveX : int -> unit
end;;
In C++, such methods are called pure virtual and the corresponding classes are
called abstract. In Java and C#, we can flag both methods and classes as being
abstract. In OOHaskell, it is enough to leave the method undefined. Indeed, in the
shapes example, we omitted any mentioning of the draw method when we defined
the object generator for shapes.
OCaml’s abstract point class may be transcribed to OOHaskell as follows:
Haskell’s overlooked object system
— 1 February 2008 —
55
abstract_point x_init self =
do
xRef <- newIORef x_init
returnIO $
varX
.=. xRef
.*. print .=. ( self # getX >>= Prelude.print )
.*. emptyRecord
This object generator cannot be instantiated with mfix because getX is used but not
defined. The Haskell type system effectively prevents us from instantiating classes
which use the methods neither they nor their parents have defined. There arises the
question of the explicit designation of a method as pure virtual, which would be of
particular value in case the pure virtual does not happen to be used in the object
generator itself.
OOHaskell allows for such explicit designation by means of adding type constraints to self. To designate getX and moveX as pure virtuals of abstract_point
we change the object generator as follows:
abstract_point (x_init::a) self =
do
... as before ...
where
_ = narrow self :: Record ( GetX :=: IO a
:*: MoveX :=: (a -> IO ())
:*: HNil )
We use the familiar narrow operation, this time to express a type constraint. We
must stress that we narrow here at the type level only. The result of narrowing is not
used (cf. “ ”), so operationally it is a no-op. It does however affect the typechecking
of the program: every instantiatable extension of abstract_point must define getX
and moveX.
One may think that the same effect can be achieved by adding regular type
annotations (e.g., on self). These annotations however must spell out the desired
object type entirely. Furthermore, a regular record type annotation rigidly and
unnecessarily restrains the order of the methods in the record as well as their types
(preventing deep subtyping, Sec. 5.9). One may also think object types can be
simply constrained by specifying HasField constraints. This is impractical in so far
that full object types would need to be specified then by the programmer; Haskell
does not directly support partial signatures. Our narrow-based approach solves
these problems.
5.7 Nominal subtyping
In OCaml and, by default, in OOHaskell, object types engage into structural
subtype polymorphism. Many other OO languages prefer nominal object types
with explicitly declared subtyping (inheritance) relationships. There is an enduring
debate about the superiority of either form of subtyping. The definite strength
of structural subtype polymorphism is that it naturally enables inference of object
56
O. Kiselyov and R. Lämmel
types. The downside is potentially accidental subtyping (Cardelli & Wegner, 1985):
a given object may be admitted as an actual argument of some function just because
its structural type fits. Nominal types allow us to restrict subtyping polymorphism
on the basis of explicitly declared subclass or inheritance relationships between
nominal (i.e., named) types.
Although OOHaskell is biased towards structural subtyping polymorphism,
OOHaskell, as a general sandbox for typed OO language design, does admit
nominal object types and nominal subtyping including multiple inheritance.
We revisit our familiar printable points and colored points, switching to nominal
types. First, we need to invent class names, or nominations:
data PP
data CP
= PP
= CP
-- Printable points
-- Colored points
As an act of discipline, we also register these types as nominations:
class Nomination f
instance Nomination PP
instance Nomination CP
We attach nomination to a regular, record-based OOHaskell object as a phantom type. To this end, we using the following newtype wrapper:
newtype N nom rec = N rec
The following two functions add and remove the nominations:
-- An operation to ‘nominate’ a record as nominal object
nominate :: Nomination nt => nt -> x -> N nt x
nominate nt x = N x
-- An operation to take away the type distinction
anonymize :: Nomination nt => N nt x -> x
anonymize (N x) = x
To be able to invoke methods on nominal objects, we need a HasField instance for
N, with the often seen delegation to the wrapped record:
instance (HasField l x v, Nomination f) => HasField l (N f x) v
where hLookupByLabel l o = hLookupByLabel l (anonymize o)
OO programming with nominal subtyping on PP and CP can now commence. The
object generator for printable points remains exactly the same as before except that
we nominate the returned object as an PP:
printable_point x_init s =
do
x <- newIORef x_init
returnIO $ nominate PP -- Nominal!
$ mutableX .=. x
.*. getX
.=. readIORef x
.*. moveX
.=. (\d -> modifyIORef x (+d))
.*. print
.=. ((s # getX ) >>= Prelude.print)
.*. emptyRecord
Haskell’s overlooked object system
— 1 February 2008 —
57
The nominal vs. structural distinction only becomes meaningful once we start
to annotate functions explicitly with the requested nominal argument type. We
will first consider request that insist on a specific nominal type, with no subtyping
involved. Here is a print function that only accepts nominal printable points.
printPP (aPP::N PP x) = aPP # print
To demonstrate nominal subtyping, we define colored points (‘CP’):
colored_point x_init (color::String) self =
do
super <- printable_point x_init self
returnIO $ nominate CP -- Nominal!
$ print .=. ( do putStr "so far - "; super # print
putStr "color - "; Prelude.print color )
.<. getColor .=. (returnIO color)
.*. anonymize super -- Access record!
We need to make CP a nominal subtype of PP. That designation is going to be
explicit. We introduce a type class Parents, which is an extensible type-level function from nominal types to the list of their immediate supertypes. A type may have
more than one parent: multiple inheritance. The following two instances designate
PP as the root of the hierarchy and CP as its immediate subtype:
class ( Nomination child, Nominations parents ) =>
Parents child parents | child -> parents
instance Parents PP HNil
instance Parents CP (HCons PP HNil)
-- PP has no parents
-- Colored points are printable points
The OOHaskell library also defines a general relation Ancestor, which is the
reflexive, transitive closure of Parents:
class ( Nomination f, Nomination anc ) =>
Ancestor f anc
We are now in the position to define an upcast operation, which is the basis for
nominal subtyping:
-- An up-cast operation
nUpCast :: Ancestor f g => N f x -> N g x
nUpCast = N . anonymize
We could also define some forms of downcast. Our nUpCast does no narrowing, so
operationally it is the identity function. This is consistent with the implementation of the nominal upcast in mainstream OO languages. The record type of an
OOHaskell object is still visible in its nominal type. Our nominal objects are
fully OOHaskell objects except that their subtyping is deliberately restricted.
We can define a subtype-polymorphic print function for printable points by ‘relaxing’ the non-polymorphic printPP function through upcast.16
16
We cannot define printPP’ in a point-free style because of Haskell’s monomorphism restriction.
58
O. Kiselyov and R. Lämmel
printPP (aPP::N PP x) = aPP # print -- accept PP only
printPP’ o = printPP (nUpCast o)
-- accept PP and nominal subtypes
The couple printPP and printPP’ clarifies that we can readily restrict argument
types of functions to either precise types or to all subtypes of a given base. This
granularity of type constraints it not provided by mainstream OO languages. Also,
the use of structural subtyping in the body of printPP hints at the fact that we
can blend nominal and structural subtyping with ease in OOHaskell. Again, this
is beyond state-of-the-art in mainstream OO programming.
5.8 Iso-recursive types
In the previous section, we have studied nominal types for the sake of nominal
subtyping. Nominal types are intrinsically necessary, when we need to model recursive object types in OOHaskell. In principle, a type system with equi-recursive
types would be convenient in this respect. However, adding such types to Haskell
was debated and then rejected because it will make type-error messages nearly useless (Hughes, 2002). Consequently, we encode recursive object types as iso-recursive
types; in fact, we use newtypes. (An alternative technique of existential quantification (Pierce & Turner, 1994) is discussed in Sec. 5.11.)
We illustrate iso-recursive types on uni-directionally linked dynamic lists. The
interface of such list objects has methods that also return list objects: a getter for
the tail and an insertion method.
-- The nominal object type
newtype ListObj a =
ListObj (ListInterface a)
-- The structural interface type
type ListInterface a =
Record (
IsEmpty :=: IO
:*: GetHead :=: IO
:*: GetTail :=: IO
:*: SetHead :=: (a
:*: InsHead :=: (a
:*: HNil )
Bool
a
(ListObj a)
-> IO ())
-> IO (ListObj a))
Recall that we had to define a HasField instance whenever we went beyond the
normal ‘objects as records’ approach. This is the case here, too. Each newtype for
iso-recursion has to be complemented by a trivial HasField instance:
instance HasField l (ListInterface a) v =>
HasField l (ListObj a) v
where
hLookupByLabel l (ListObj x) = hLookupByLabel l x
For clarity, we chose the implementation of ListInterface a with two OO
classes: for the empty and non-empty lists. A single OO list class would have sufficed
too. Empty-list objects fail for all getters. Here is the straightforward generator for
empty lists:
Haskell’s overlooked object system
— 1 February 2008 —
59
nilOO self :: IO (ListInterface a)
= returnIO
$ isEmpty .=. returnIO True
.*. getHead .=. failIO "No head!"
.*. getTail .=. failIO "No tail!"
.*. setHead .=. const (failIO "No head!")
.*. insHead .=. reusableInsHead self
.*. emptyRecord
The reusable insert operation constructs a new object of the consOO:
reusableInsHead list head
= do
newCons <- mfix (consOO head list)
returnIO (ListObj newCons)
Non-empty list objects hold a reference for the head, which is accessed by getHead
and setHead. Here is the object generator for non-empty lists:
consOO head tail self
= do
hRef <- newIORef
returnIO
$ isEmpty .=.
.*. getHead .=.
.*. getTail .=.
.*. setHead .=.
.*. insHead .=.
.*. emptyRecord
head
returnIO False
readIORef hRef
returnIO (ListObj tail)
writeIORef hRef
reusableInsHead self
OO programming on nominal objects commences without ado. They can be used
just like record-based OOHaskell objects before. As an example, the following
recursive function prints a given list. One can check that the various method invocations involve nominally typed objects.
printList aList
= do
empty <- aList # isEmpty
if empty
then putStrLn ""
else do
head <- aList # getHead
putStr $ show head
tail <- aList # getTail
putStr " "
printList tail
5.9 Width and depth subtyping
We have used the term subtyping in the informal sense of type-safe type substitutability. That is, we call the object type S to be a subtype of the object type T
if in any well-typed program P the typeability of method invocations is preserved
upon replacing objects of type T with objects of type S. This notion of subtyping is
60
O. Kiselyov and R. Lämmel
to be distinguished from behavioural subtyping, also known as Liskov Substitution
Principle (Liskov & Wing, 1994).
In OOHaskell, subtyping is enabled by the type of the method invocation
operator #. For instance, the function \o -> o # getX has the following inferred
type:
HasField (Proxy GetX) o v => o -> v
This type is polymorphic. The function will accept any object (i.e., record) o provided that it has the method labelled GetX whose type matches the function’s
desired return type v.
A basic form of subtyping or subsumption is width subtyping, whereupon an
object of type S is a subtype of T if the record type S has (at least) all the fields
of T with the exact same type. The HList library readily provides this subtyping
relation, Record.SubType. Corresponding constraints can be added to type signatures (although we recall that Sec. 5.6 devised a constraint technique that is more
convenient for OOHaskell). It is easy to see that if SubType S T holds for some
record types S and T , then substituting an object of type S for an object of type
T preserves the typing of every occurrence of # in a program. No method will be
missing and no method will be of a wrong type.
Width subtyping is only one form of subtyping. There are other subtyping relations, which too preserve the typing of each occurrence of # in a program — in
particular, depth subtyping. While width subtyping allows the subtype to have
more fields than the supertype, depth subtyping allows the fields of the subtype to
relate to the fields of the supertype by subtyping. Typed mainstream OO languages
like Java and C# do not support full depth subtyping.
We will now explore depth subtyping in OOHaskell. We define some new object
types and functions on the one-dimensional printable_point class from Sec. 4.2.5
and its extension colored_point from Sec. 4.3.3. We define a simple-minded onedimensional vector class, specified by two points for the beginning and the end,
which can be accessed by the methods getP1 and getP2:
vector (p1::p) (p2::p) self =
do
p1r <- newIORef p1
p2r <- newIORef p2
returnIO $
getP1
.=. readIORef
.*. getP2
.=. readIORef
.*. print
.=. do self #
self #
.*. emptyRecord
p1r
p2r
getP1 >>= ( # print )
getP2 >>= ( # print )
The local type annotations p1::p and p2::p enforce our intent that the two points
of the vector have the same type. It is clear that objects of type p must be able to
respond to the message print. Otherwise, the type of the points is not constrained.
Our object generator vector is parameterised over the class of points. In C++, the
close analogue is a class template. This example shows that Haskell’s normal forms
Haskell’s overlooked object system
— 1 February 2008 —
61
of polymorphism, combined with type inference, allow us to define parameterised
classes without ado.
We construct two vector objects, v and cv:
testVector = do
p1 <- mfix (printable_point 0)
p2 <- mfix (printable_point 5)
cp1 <- mfix (colored_point 10 "red")
cp2 <- mfix (colored_point 25 "red")
v <- mfix (vector p1 p2)
cv <- mfix (vector cp1 cp2)
-- ... to be continued ...
The former is the vector of two printable points; the latter is the vector of two
colored points. The types of v and cv are obviously different: the type checker will
remind us of this fact if we tried to put both vectors into the same homogeneous
list. The vectors v and cv are not related by width subtyping: indeed, both vectors
agree on method names, but the types of the methods getP1 and getP2 differ. In v,
the method getP1 has the type IO PrintablePoint whereas in cv the same method
has the type IO ColoredPoint. These different result types, PrintablePoint and
ColoredPoint, are related by width subtyping.
The type of cv is a deep subtype of v. In OOHaskell, we may readily use
functions (or methods) that exploit depth subtyping. For instance, we can define
the following function for computing the norm of a vector, and we can pass either
vector v or cv to the function.
norm v =
do
p1 <- v # getP1; p2 <- v # getP2
x1 <- p1 # getX; x2 <- p2 # getX
return (abs (x1 - x2))
The above test code continues thus:
-- ... continued ...
putStrLn "Length of v"
norm v >>= Prelude.print
putStrLn "Length of colored cv"
norm cv >>= Prelude.print
The method invocation operations within norm remain well-typed no matter
which vector, v or cv, we pass to that function. The typing of # is indeed compatible with both width and depth subtyping, and, in fact, their combination. Thus,
the object type S is a subtype of T if the record type S has all the fields of T
whose types are not necessarily the same but related by subtyping in turn. Here
we assume, for now, that subtyping on method types is defined in accordance to
conservative rules (Cardelli & Wegner, 1985; Abadi & Cardelli, 1996). (In the following formulation, without loss of generality, we assume that OOHaskell method
types are monadic function types.) If A1 → · · · → An → IO R is a method type
from T , then there must be a method type in S, with the same method name, and
with a type A′1 → · · · → A′n → IO R′ such that the following relationships hold:
62
O. Kiselyov and R. Lämmel
• A1 , . . . , An must be subtypes of A′1 , . . . , A′n .
• R′ must be a subtype of R.
(contra-variance)
(co-variance)
The above vector example exercises the co-variance of the result type for the getters
getP1 and getP2.
We never had to specifically assert that the types of two objects are related by
width or depth subtyping. This is because in each and every case, the compiler
checks the well-typedness of all method invocations directly, so no separate subtyping rules are needed. We contrast this with type systems like System F≤ , where the
subsumption rules are explicitly asserted. The only place where an OOHaskell
programmer has to make the choice of subtyping relationship explicit is in explicit
narrowing operations. The previously described operation narrow covers width subtyping; the OOHaskell library also includes an operation deep’narrow. For instance, we can place v and cv, into the same homogeneous list:
let vectors = [v, deep’narrow cv]
The operation deep’narrow descends into records, prefixes method arguments
by narrowing, and postfixes method results by narrowing. Deep narrowing is just
another record operation driven by the structure of method types. (We refer to
the source distribution for details.) Deep narrowing is not the only way of dealing
explicitly with depth subtyping in OOHaskell. We may also adopt the union-type
technique as of Sec. 5.5.
5.10 Co-variant method arguments
The variance of argument types is the subject of a significant controversy (Castagna, 1995; Surazhsky & Gil, 2004; Howard et al., 2003). The contravariant rule for method arguments entails type substitutability, i.e., it assures the
type safety of method invocation for all programs. However, argument type contravariance is known to be potentially too conservative. It is often argued that a
co-variant argument type rule is more suitable for modelling real-world problems.
If a method with co-variant argument types happens to receive objects of expected
types, then co-variance is safe — for that particular program. The proponents of
the co-variant argument type rule argue that because of the idiomatic advantages
of the rule we should admit it for those programs where it is safe. It is the job
of the compiler to warn the user when the co-variant rule is used unsafely. Alas,
in the case of Eiffel — the most established language with co-variance — the situation is the following: “No compiler currently available fully implements these
checks and behaviour in those cases ranges from run-time type errors to system
crashes.” (comp.lang.eiffel, 2004).
In this section we demonstrate the restrictiveness of contra-variance for methodargument types and show that OOHaskell’s subtyping naturally supports typesafe co-variance. The faithful implementation of the archetypal example from the
Eiffel FAQ (comp.lang.eiffel, 2004) is contained in the accompanying source code.
Continuing with the vector example from the previous section, we extend vector
Haskell’s overlooked object system
— 1 February 2008 —
63
with a method, moveO for moving the origin of the vector. The method receives the
new origin as a point object.
vector1 (p1::p) (p2::p) self =
do
super <- vector p1 p2 self
returnIO
$ moveO .=. (\pa -> do
p1 <- self # getP1
x <- pa # getX
p1 # moveX $ x)
.*. super
As in the previous section, we construct the vector1 of plain printable points
v1 and the vector1 of colored points cv1. If we intend cv1 to be substitutable
for v1 in all circumstances (by the virtue of depth subtyping), we must follow the
contra-variance rule, which requires the argument pa of moveO be either a plain
printable point (or an instance of its super-type). That requirement is responsible
for the longer-than-expected implementation of moveO. Furthermore, the supertyping requirement on pa precludes moveO’s changing the color of the origin point,
for the vector of colored points. That degrades the expressiveness.
To illustrate the subtyping of the vectors, we define the function that moves the
origin of its vector argument to 0:
move_origin_to_0 varg =
do
zero <- mfix (printable_point 0)
varg # moveO $ zero
We may indeed apply that function to either v1 or cv1. The function is polymorphic
and can take any vector1 of plain points and and its subtypes. The type of cv1 is
truly a deep subtype of the type of v1. (Again, OOHaskell does not require us
to assert the relevant subtype relationship in any way.)
We now turn to co-variant method argument types and so experiment with yet
another class of vectors. We also construct two instances of vector2.
vector2 (p1::p) (p2::p) self =
do
p1r <- newIORef p1
p2r <- newIORef p2
returnIO $
setO .=. writeIORef p1r
-- ... other fields as in vector ...
testVector = do
-- ... test case as before ...
v2 <- mfix (vectors p1 p2)
-- vector of printable points
cv2 <- mfix (vectors cp1 cp2) -- vector of colored points
Like vector1, vector2 provides for setting the origin point; cf. the method setO.
However, vector2 does that in a direct and simple way; also, only vector2 permits
changing the color of the origin point, in a vector of colored points. Although the
64
O. Kiselyov and R. Lämmel
method setO is more convenient and powerful than the method moveO, the method
setO has co-variant argument types — across printable-point and colored-point
vectors. For a vector of colored points, cv2, the argument type of setO must be a
colored point too, i.e., the same type as p1r — otherwise, the mutation writeIORef
cannot be typed.
Hence, the type of cv2 cannot be a subtype of the type of v2 (because setO
breaks the contra-variant argument type rule). An OO system that enforces the
contra-variant rule will not allow us to write functions that can take both v2 and
cv2. For example, we may want to devise the following function:
align_origins va vb =
do
pa <- va # getP1
vb # setO $ pa
It is always safe to apply align_origins to two vector2s of the same type.
OOHaskell does let us pass either two vector2s of printable points (such as
v2) to two vector2s of colored points (such as cv2), and so vector types can be
substitutable — despite a co-variant argument type of set0.
Substitutability is properly restricted for this function:
set_origin_to_0 varg =
do
zero <- mfix (printable_point 0)
varg # setO $ zero
We apply the function to v2, but if we try to apply it to cv2 we get the type error
message about a missing method getColor (which distinguishes colored points from
plain printable points). Likewise, we get an error if we attempt to place both v2
and cv2 in a homogeneous list like this:
let vectors = [v2, deep’narrow cv2]
In this case, we can narrow both vectors to the type of vector though, so that the
offending method setO will be projected out and becomes private.
OOHaskell typechecks actual operations on objects; therefore, OOHaskell
permits methods with co-variant argument types in situations where they are used
safely. The type checker will flag any unsafe use and force the programmer to remove
the offending method. Permitting safe uses of methods with co-variant argument
types required no programming on our part. We get this behaviour for free.
5.11 Anti-patterns for subtyping
We have seen several approaches to the construction of a subtype-polymorphic
collection, as needed for the ‘scribble’ loop in the running shapes example. In the
section on non-OOHaskell encodings, Sec. 3, we had discussed two additional
options:
• The use of HList’s heterogeneous lists.
• The use of “∃” to make the list element type opaque.
Haskell’s overlooked object system
— 1 February 2008 —
65
Albeit one might have expected these options to be of use, they turned out to
be problematic for OO programming with non-extensible Haskell records. In the
combination with OOHaskell (and its extensible records), these two options are
even less attractive.
In first approach, we construct the scribble list as:
let scribble = s1 ‘HCons‘ (s2 ‘HCons‘ HNil)
and use hMap_, Sec. 3.7, to iterate over the list:
hMapM_ (⊥::FunOnShape) scribble
where there must be an instance of type class Apply for FunOnShape, e.g.:
instance ( HasField (Proxy Draw) r (IO ())
, HasField (Proxy RMoveTo) r (Int -> Int -> IO ())
)
=> Apply FunOnShape r (IO ())
where
apply _ x = do
x # draw
(x # rMoveTo) 100 100
x # draw
Haskell’s type class system requires us to provide proper bounds for the instance,
hence the list of the method-access constraints (for “#”, i.e., HasField) above. The
form of these constraints strongly resembles the method types listed in the shape
interface type, Sec. 5.2. One may wonder whether we can somehow use the full type
synonym Shape, in order to constrain the instance. This is not possible in Haskell
because constraints are not first-class citizens in Haskell; we cannot compute them
from types or type proxies — unless we were willing to rely on heavy encoding or
advanced syntactic sugar. So we are doomed to manually infer and explicitly list
such method-access constraints for each such piece of polymorphic code.
The existential quantification approach falls short for essentially the same reason. Assuming a suitable existential envelope and following Sec. 3.6, we can build
scribble as
let scribble = [ HideShape s1, HideShape s2 ]
The declaration of the existential type depends on the function that we want to
apply to the opaque data. When iterating over the list, via mapM_, we only need to
unwrap the HideShape constructor prior to method invocations:
mapM_ ( \(WrapShape shape) -> do
shape # draw
(shape # rMoveTo) 100 100
shape # draw )
scribble
These operations have to be anticipated in the type bound for the envelope:
data OpaqueShape =
forall x. ( HasField (Proxy Draw) x (IO ())
, HasField (Proxy RMoveTo) x (Int -> Int -> IO ())
) => HideShape x
66
O. Kiselyov and R. Lämmel
This approach evidently matches the HList-based technique in terms of encoding
efforts. In both cases, we need to identify type class constraints that correspond
to the (potentially) polymorphic method invocations. This is impractical. Not even
mainstream OO languages with no advanced type inference, require this sort of
type information from the programmer.
Existential quantification can also be used for object encoding, e.g., for wrapping
up self. That lets us, for example, easily implement self-returning methods without
resorting to infinite types. Such use of existential quantification is not practical in
OOHaskell for the same reason: it requires us to exhaustively enumerate all type
classes an object and any of its types are or will be the instances of.
6 Discussion
We will first discuss usability issues of the current OOHaskell library, further constrained by current Haskell implementations. We will then summarise related work
on functional object-oriented programming in Haskell and elsewhere. Finally, we will
list topics for future work — other than just improving usability of OOHaskell.
6.1 Usability issues
6.1.1 Usability of inferred types
So far, we have not shown any type inferred by Haskell for our objects. One may
wonder how readable and comprehensible they are, if they can be used as means of
program understanding, and if a Haskell language extension is needed to improve
the presentation of the inferred types. In upshot, the inferred types are reasonable
for simple OO programming examples, but there is a fuzzy borderline beyond which
the volume and the idiosyncrasies of inferred types injure their usefulness. This
concern suggests an important topic for future work.
Let us see the inferred type of the colored point introduced in Sec. 4.3.1:
ghci6.4> :t mfix $ colored_point (1::Int) "red"
mfix $ colored_point (1::Int) "red" ::
IO (Record
(HCons (Proxy GetColor, IO String)
(HCons (Proxy VarX, IORef Int)
(HCons (Proxy GetX, IO Int)
(HCons (Proxy MoveX, Int -> IO ())
(HCons (Proxy Print, IO ())
HNil))))))
We think that this type is quite readable, even though it reveals the underlying
representation of records (as a heterogeneous list of label-value pairs), and gives
away the proxy-based model for labels. We may hope for a future Haskell implementation whose customisable ‘pretty printer’ for types would present the result of
type inference perhaps as follows:
ghci> :t mfix $ colored_point (1::Int) "red"
mfix $ colored_point (1::Int) "red" ::
Haskell’s overlooked object system
IO ( Record (
GetColor
:*: VarX
:*: GetX
:*: MoveX
:*: Print
:*: HNil ))
:=:
:=:
:=:
:=:
:=:
— 1 February 2008 —
67
IO String
IORef Int
IO Int
(Int -> IO ())
IO ()
The above example dealt with monomorphic objects. Let us also see the inferred
type of a polymorphic object generator, with ‘open recursion left open’. Here is the
(pretty-printed) type of the object generator for colored points:
ghci> :t colored_point
( Num a
, HasField (Proxy GetX) r (IO a1)
, Show a1
) => a
-> String
-> r
-> IO ( Record (
GetColor :=: IO String
:*: VarX
:=: IORef a
:*: GetX
:=: IO a
:*: MoveX
:=: (a -> IO ())
:*: Print
:=: IO ()
:*: HNil ))
The inferred type lists all the fields of an object, both new and inherited. Assumptions about self are expressed as constraints on the type variable r. The object
generator refers to getX (through self), which entails a constraint of the form
HasField (Proxy GetX) r (IO a1). The coordinate type for the point is polymorphic; cf. a for the initial value and a1 for the value retrieved by getX. Since
arithmetics is performed on the coordinate value, this implies bounded polymorphism: only Num-ber types are permitted. We cannot yet infer that a and a1 must
eventually be the same since ‘the open recursion is still open’.
We must admit that we have assumed a relatively eager instance selection in
the previous Haskell session. The Hugs implementation of Haskell is (more than)
eager enough. The recent versions of GHC have become quite lazy. In a session with
contemporary GHC (6.4), the inferred type would comprise the following additional
constraints, which all deal with the uniqueness of label sets as they are encountered
during record extension:
HRLabelSet (HCons
(HCons
likewise for MoveX,
likewise for MoveX,
likewise for MoveX,
(Proxy
(Proxy
Print,
Print,
Print,
MoveX, a -> IO ())
Print, IO ()) HNil)),
GetX
GetX, VarX
GetX, VarX, GetColor
Inspection of the HRLabelSet instances shows that these constraints are all satisfied,
no matter how the type variable a is instantiated. No ingenuity is required. A
simple form of strictness analysis were sufficient. Alas, GHC is consistently lazy in
68
O. Kiselyov and R. Lämmel
resolving even such constraints. Modulo HRLabelSet constraints, the inferred type
seems quite reasonable, explicitly listing all relevant labels and types of the record
components.
6.1.2 Usability of type errors
Due to OOHaskell’s extensive use of type-class-based programming, there is a
risk that type errors may become too complex. We will look at some examples. The
results clearly provide incentive for future work on the subject of type errors.
Let us first attempt to instantiate an abstract class, e.g., abstract_point from
Sec. 5.6. That object generator defined the print method, which invoked getX on
self. The latter is left to be defined in concrete subclasses. If we take the fixpoint
of such an ‘incomplete’ object generator, Haskell’s type checker (here: GHC 6.4)
gives the following error message:
ghci> let x = mfix (abstract_point 7)
No instance for (HasField (Proxy GetX) HNil (IO a))
arising from use of ‘abstract_point’ at <interactive>:1:14-27
Probable fix:
add an instance declaration for (HasField (Proxy GetX) HNil (IO a1))
In the first argument of ‘mfix’, namely ‘(abstract_point 7)’
In the definition of ‘x’: x = mfix (abstract_point 7)
We think that the error message is concise and to the point. The message succinctly lists just the missing field (The suggested ‘probable fix’ is not really helpful
here). In our next scenario, we use a version of abstract_point that comprises an
instantiation test by constraining self through narrow, as discussed in Sec. 5.6:
abstract_point (x_init::a) self =
do
... as before ...
where
_ = narrow self :: Record ( GetX :=: IO a
:*: MoveX :=: (a -> IO ())
:*: HNil )
When we now take the fixpoint again, we get a more complex error message:
ghci> let x = mfix (abstract_point 7)
No instance for (HExtract HNil (Proxy GetX) (IO a),
HExtract HNil (Proxy MoveX) (a -> IO ()),
HasField (Proxy GetX) HNil (IO a1))
arising from use of ‘abstract_point’ at <interactive>:1:14-27
Probable fix: ...
In the first argument of ‘mfix’, namely ‘(abstract_point 7)’
In the definition of ‘x’: x = mfix (abstract_point 7)
Compared to the earlier error message, there are two additional unsatisfied HExtract
constraints. Two out of the three constraints refer to GetX, and they complain about
the same problem: a missing method implementation for getX. The constraint regarding MoveX deals with a pure virtual method that is not used in the object
Haskell’s overlooked object system
— 1 February 2008 —
69
generator. The kinds and numbers of error messages for GetX and MoveX may lead
to confusion; internals of OOHaskell end up at the surface.
In order to improve on such problems, the Haskell type system and its errorhandling part would need to be opened up to allow for problem-specific error messages. We would like to refine Haskell’s type checker so that type error messages
directly refer to the involved OO concepts.
Let us consider yet another scenario. We turn to self-returning methods, as we
discussed them in Sec. 5.3. In the following flawed OOHaskell program, we attempt to return self right away:
self_returning_point (x_init::a) self =
do
super <- printable_point x_init self
returnIO
$ me .=. self -- assumes iso-recursive types
.*. super
The problem will go unnoticed until we try to mfix the generator, at which point
we get a type error:
Occurs check: cannot construct the infinite type:
a
=
Record (HCons (Proxy Me, a)
(HCons (Proxy MutableX, IORef a1)
(HCons (Proxy GetX, IO a1)
(HCons (Proxy MoveX, a1 -> IO ())
(HCons (Proxy Print, IO ()) HNil)))))
Expected type: a -> IO a
Inferred type: a -> IO (Record (HCons (Proxy Me, a)
(HCons (Proxy MutableX, IORef a1)
(HCons (Proxy GetX, IO a1)
(HCons (Proxy MoveX, a1 -> IO ())
(HCons (Proxy Print, IO ()) HNil))))))
In the application ‘self_returning_point 7’
In the first argument of ‘mfix’, namely ‘(self_returning_point 7)’
This error message is rather complex compared to the simple object types that
are involved. Although the actual problem is correctly described, the programmer
receives no help in locating the offending code, me .=. self. The volume of the
error message is the consequence of our use of structural types. One may think that
adding some type synonyms and using them in type signatures should radically
improve the situation. It is true that contemporary Haskell type checkers keep
track of type synonyms. However, an erroneous subexpression may just not be
sufficiently annotated or constrained by its context. Also, the mere coding of type
synonyms is very inconvenient. This situation suggests that a future Haskell type
checker could go two steps further. Our first proposal is to allow for the inference
of type synonyms; think of:
foo x y z = ... -- complex expression on structural object types
type Foo = typeOf foo
-- capture the type in an alias
70
O. Kiselyov and R. Lämmel
(Here, typeOf is an envisaged extension.) Our second proposal is to use type synonyms aggressively for the simplification of inferred types or type portions of error
messages. This is a challenging subject given Haskell’s forms of polymorphism.
The verbosity of OOHaskell error messages may occasionally compare to error
messages in C++ template instantiation, which can be immensely verbose, spanning several dozens of packed lines, and yet boost and similar C++ libraries, which
extensively use templates, are gaining momentum. In general, the clarity of error
messages is undoubtedly an area that needs more research, and such research is being carried out by Sulzmann and others (Stuckey et al., 2004), which OOHaskell
programmers and Haskell compiler writers may take advantage of.
The ultimate conclusion of our discussion of inferred types and type errors is
that such type information needs to be presented to the programmer in an abbreviated and OO-aware fashion. This proposal is based on the observation of OCaml’s
development. Although objects types shown by OCaml are quite concise, that has
not always been the case. In the ML-ART system, the predecessor of OCaml with
no syntactic sugar (Rémy, 1994), the printed inferred types were not unlike the
OOHaskell types we have seen in this section.
“Objects have anonymous, long, and often recursive types that describe all methods
that the object can receive. Thus, we usually do not show the inferred types of programs
in order to emphasise object and inheritance encoding rather than typechecking details.
This is quite in a spirit of ML where type information is optional and is mainly used
for documentation or in module interfaces. Except when trying top-level examples, or
debugging, the user does not often wish to see the inferred types of his programs in a
batch compiler.”
6.1.3 Efficiency of object encoding
Our representation of objects and their types is deliberately straightforward: polymorphic, extensible records of closures. This approach has strong similarities with
prototype-based systems (such as Self (Ungar & Smith, 1987)) in that mutable
fields and method ‘pointers’ are contained in one record. A more efficient representation based on separate method and field tables (as in C++ and Java) is
possible, in principle. Although our current encoding is certainly not optimal, it
is conceptually clearer. This encoding is used in such languages as Perl, Python,
Lua — and is often the first one chosen when adding OO to an existing language.
The efficiency of the current OOHaskell encoding is also problematic for reasons other than separation of fields and methods. For example, although record
extension is constant (run-)time, the field/method lookup is linear search. Clearly,
a more efficient encoding is possible: one representation of the labels in the HList
paper permits a total order among the labels types, which in turn, permits construction of efficient search trees. We may also impose an order on the components
per record type, complete with subtype-polymorphic record extension only to the
right, so that labels can be mapped to array indexes.
In the present paper, we chose conceptual clarity over such optimisations. Furthermore, a non-trivial case study is needed to drive optimisations. Mere improvements in object encoding may be insufficient however. The compilation time of
Haskell’s overlooked object system
— 1 February 2008 —
71
OOHaskell programs and their runtime efficiency is challenged by the number of
heavily nested dictionaries that are implied by our systematic type-class-based approach. It is quite likely that a scalable HList/OOHaskell style of programming
will require compiler optimisations that make type-class-based programming more
efficient — in general.
6.2 Related work
Throughout the paper we referenced related work whenever specific technical aspects suggested to do so. We will complete the picture by a broader discussion.
There are three overall dimensions of related work: foundations of object encoding (cf. Sec. 6.2.1), Haskell extensions for OO (cf. Sec. 6.2.2), and OO encoding in
Haskell (cf. Sec. 6.2.3).
The literature on object encoding is quite extensive. OOHaskell takes advantage of seminal work such as (Cardelli & Wegner, 1985; Abadi & Cardelli, 1996;
Ohori, 1995; Pierce & Turner, 1994; Bruce & Mitchell, 1992). Most often, typed
object encodings are based on polymorphic lambda calculi with subtyping, while
there are also object calculi that start, more directly, from objects or records. Due
to this overwhelming variety, we narrow down the discussion. We identify MLART (Rémy, 1994) by Rémy et al. (see also (Rémy & Vouillon, 1997)) as the closest to OOHaskell — in motivation and spirit, but not in the technical approach.
Hence, Sec. 6.2.1 is entirely focused on ML-ART, without further discussion of less
similar object encodings. The distinguishing characteristic of OOHaskell is the
use of type-class-bounded polymorphism.
6.2.1 The ML-ART object encoding
Both ML-ART and OOHaskell identify a small set of language features that make
functional object-oriented programming possible. In both projects, the aim was to
be able to implement objects — as a library feature. Therefore, several OO styles
can be implemented, for different classes of users and classes of problems. One does
not need to learn any new language and can discover OO programming progressively. Both ML-ART and OOHaskell base their object systems on polymorphic
extensible records. Both OOHaskell and ML-ART deal with mutable objects
(OOHaskell currently neglects functional objects since they are much less commonly used in practise). Both OOHaskell and ML-ART aim at preserving type
inference.
ML-ART adds several extensions to ML to implement objects: records with polymorphic access and extension, projective records, recursive types, implicit existential and universal types. As the ML-ART paper (Rémy, 1994) reports, none of the
extensions are new, but their combination is original and “provides just enough
power to program objects in a flexible and elegant way”.
We make the same claim for OOHaskell, but using a quite different set of
features. What fundamentally sets us apart from ML-ART is the different source
language: Haskell. In Haskell, we can implement polymorphic extensible records
72
O. Kiselyov and R. Lämmel
natively rather than via an extension. We use type-class-based programming to this
end.17 We avoid row variables and their related complexities. Our records permit
introspection and thus let us implement various type-safe cast operations appealing
to different subtyping relationships. For instance, unlike ML-ART, OOHaskell
can compute the most common type of two record types without requiring type
annotations. Quoting from the ML-ART paper:
“The same message print can be sent to points and colored points. However, both of
them have incompatible types and can never be stored in the same list. Some languages
with subtyping allow this set-up. They would take the common interface of all objects
that are mixed in the list as the interface of any single object of the list.”
Unlike ML-ART, we do not rely on existential or implicitly universal types, nor
recursive types. We use value recursion instead. That representation, a record of
recursive closures, abstracts the internal state of the object — its value as well
its type. Haskell helps us overcome what ML-ART calls “severe difficulties” with
value recursion. In ML, the difficulties are serious enough to abandon the value
recursion, despite its attractive features in supporting implicit subtyping, in favour
of more complex object encodings requiring extensions of the type system. The
subtle problem of value recursion is responsible for complicated and elaborate rules
of various mainstream OO languages that prescribe what an object constructor may
or may not do. The ML-ART paper mentions an unpublished attempt (by Pierce)
to take advantage of the facts that fixpoints in a call-by-name language are always
safe and that call-by-name can be emulated in a call-by-value language with the
help of extra abstraction (thunks). However, in that attempted implementation the
whole message table had to be rebuilt every time an object sends a message to self
and so that approach was not pursued further. Our simple scheme of Sec. 4.4 seems
to answer the ML-ART challenge — “to provide a clean and efficient solution that
permits restricted form of recursion on non-functional values.”
ML-ART uses a separate method table, whereas OOHaskell uses a single record
for both mutable fields and method ‘pointers’. The ML-ART encoding is more efficient than that of OOHaskell. All instances of an object (class) literally share
the same method table. ML-ART (and OCaml) is also more efficient simply because more elements of the object encoding are natively implemented. By contrast,
OOHaskell’s type system is programmed through type-class-based programming.
As a result, OOHaskellis definitely less fit for practical OO software development
than ML-ART (or rather OCaml).
6.2.2 Haskell language extensions
There were attempts to bring OO to Haskell by a language extension. An early
attempt is Haskell++ (Hughes & Sparud, 1995) by Hughes and Sparud. The authors motivated their extension by the perception that Haskell lacks the form of
17
The fact that such records are realisable in Haskell at all has been unknown, until the HList
paper, which we published in 2004. The assumed lack of extensible records in Haskell was
selected as prime topic for discussion at the Haskell 2003 workshop (H. Nilsson, 2003).
Haskell’s overlooked object system
— 1 February 2008 —
73
incremental reuse that is offered by inheritance in object-oriented languages. Our
approach uses common extensions of the Hindley/Milner type system to provide
the key OO notions. So in a way, Haskell’s fitness for OO programming just had to
be discovered, which is the contribution of this paper.
O‘Haskell (Nordlander, 1998; Nordlander, 2002) is a comprehensive OO variation
on Haskell designed by Nordlander. O‘Haskell extends Haskell with reactive objects
and subtyping. The subtyping part is a substantial extension. The reactive object
part combines stateful objects and concurrent execution, again a major extension.
Our development shows that no extension of Haskell is necessary for stateful objects,
and the details of the object system can be programmed in Haskell.
Another relevant Haskell variation is Mondrian. In the original paper on the
design and implementation of Mondrian (Meijer & Claessen, 1997), Meijer and
Claessen write: “The design of a type system that deals with subtyping, higherorder functions, and objects is a formidable challenge ...”. Rather than designing
a very complicated language, the overall principle underlying Mondrian was to obtain a simple Haskell dialect with an object-oriented flavour. To this end, algebraic
datatypes and type classes were combined into a simple object-oriented type system with no real subtyping, with completely co-variant type-checking. In Mondrian,
runtime errors of the kind “message not understood” are considered a problem akin
to partial functions with non-exhaustive case discriminations. OOHaskell raises
the bar by providing proper subtyping (“all message will be understood”) and other
OO concepts in Haskell without extending the Haskell type system.
6.2.3 Object encodings for Haskell
This paper may claim to provide the most authoritative analysis of possible object
encodings in Haskell; cf. Sec. 3. Previous published work on this subject has not
addressed general (functional) object-oriented programming, but it has focused instead on the import of foreign libraries or components into Haskell (Finne et al., 1999;
Shields & Peyton Jones, 2001; Pang & Chakravarty, 2004). The latter problem domain makes important simplifying assumptions:
•
•
•
•
•
•
Object state does not reside in Haskell data.
There are only (opaque) object ids referring to the foreign site.
State is solely accessed through methods (“properties”).
Haskell methods are (often generated) stubs for foreign code.
As a result, such OO styles just deal with interfaces.
No actual (sub)classes are written by the programmer.
In this restricted context, one approach is to use phantom types for recording
inheritance relationships (Finne et al., 1999). Each interface is represented by an
(empty) datatype with a type parameter for extension. After due consideration, it
turns out that this approach is a restricted version of what Burton called “type
extension through polymorphism”: even records can be made extensible through
the provision of a polymorphic dummy field (Burton, 1990). Once we do not maintain Haskell data for objects, there is no need to maintain a record type, but the
74
O. Kiselyov and R. Lämmel
extension point is a left over, and it becomes a phantom. We have “re-generalised”
the phantom approach in Sec. 3.2.
Another approach is to set up a Haskell type class to represent
the subtyping relationship among interfaces (Shields & Peyton Jones, 2001;
Pang & Chakravarty, 2004) where each interface is modelled as a dedicated (empty)
Haskell type. We have enhanced this approach by state in Sec. 3.5.
Based on our detailed analysis of both approaches, we submit that the second
approach seems to be slightly superior to the first one, while both approaches are
too cumbersome for actual functional OO programming.
Not in peer-referred publications, but in Haskell coding practise, some sorts
of OO-like encodings are occasionally found. For instance, it is relatively well
understood that Haskell’s type classes allow for interface polymorphism or for
abstract classes (type classes) vs. concrete classes (type class instances). As
of writing, the published Haskell reference solution for the shapes example,
http://www.angelfire.com/tx4/cus/shapes/, is a simple-to-understand encoding that does not attempt to maximise reuse among data declarations and accessors.
The encoding is specialised to the specific problem; the approach may fail to scale.
The encoding also uses existentials for handling subtype-polymorphic collections,
which is an inherently problematic choice, as we have shown in Sec. 5.11.
6.3 More future work
We have focused on mutable objects so far; studying functional objects appears to
be a natural continuation of this work, even though functional objects are of much
less practical relevance.
The notion of object construction as a multi-stage computation (cf. Sec. 4.4)
merits further exploration (as well as the clarification of the relationship with environment classifiers (Taha & Nielsen, 2003)).
OOHaskell should be elaborated to cover general forms of reflective programming and, on the top of that, general forms of aspect-oriented programming. A
simple form of reflection is already provided in terms of the type-level encoding
of records. We can iterate over records and their components in a generic fashion.
Further effort is needed to cover more advanced forms of reflection such as the
iteration over the object pool, or the modification of object generators.
Another promising elaboration of OOHaskell would be its use for the reusable
representation of design-pattern solutions.
7 Concluding remarks
The present paper addresses the intellectual challenge of seeing if the conventional
OO idioms can at all be implemented in Haskell (short of writing a compiler for
an OO language in Haskell). Peyton Jones and Wadler’s paper on imperative programming in Haskell (Peyton Jones & Wadler, 1993) epitomises such an intellectual tradition for the imperative paradigm. The same kind of intellectual challenge,
‘paradigm assimilation’, is addressed by FC++ (McNamara & Smaragdakis, 2004),
Haskell’s overlooked object system
— 1 February 2008 —
75
which implements in C++ the quintessential Haskell features: type inference, higherorder functions, non-strictness. The present paper, conversely, faithfully (i.e., in a
similar syntax and without global program transformation) realises a principal C++
trait — OO programming. According to Peyton Jones, Haskell is “the world’s finest
imperative programming language” (Peyton Jones, 2001). We submit that Haskell
is also a bleeding-edge OO programming language, while we readily restrict this
claim to mere OO language-design capability; much more work would be needed to
enable scalable OO software development with Haskell.
We have discovered an object system for Haskell that supports stateful objects,
inheritance and subtype polymorphism. We have implemented OO as a Haskell
library, OOHaskell, based on the polymorphic, extensible records with introspection and subtyping provided by the HList library (Kiselyov et al., 2004). Haskell
programmers can use OO idioms if it suits the problem at hand. We have demonstrated that OOHaskell programs are very close to the textbook OO code, which
is normally presented in mainstream OO languages. OOHaskell’s deviations are
appreciated. The OOHaskell library offers a comparatively rich combination of
OO idioms. Most notably, we have implemented parameterised classes, constructor methods, abstract classes, pure virtual methods, single inheritance, multiple
inheritance, object composition, structural types, and nominal types. The choice of
Haskell as a base language has allowed us to deliver extensive type inference, firstclass classes, implicit polymorphism of classes, and more generally: programmable
OO type systems. Starting from the existing OOHaskell library and the corresponding sample suite, one can explore OO language design, without the need to
write a compiler.
The present paper settles the question that hitherto has been open. The conventional OO idioms in their full generality are expressible in current Haskell without
any new extensions. It turns out, Haskell 98 plus multi-parameter type classes with
functional dependencies are sufficient. This combination is well-formalised and reasonably understood (Stuckey & Sulzmann, 2005). Even overlapping instances are
not essential (yet using them permits a more convenient representation of labels,
and a more concise implementation of some type-level functionality). The fact that
we found a quite unexpected (and unintended) use of the existing Haskell features is
reminiscent of the accidental discovery of C++ template meta-programming. The
latter is no longer considered an exotic accident or a type hack — rather, a real
feature of the language (Czarnecki et al., 2003), used in the Standard Template
Library and described in popular C++ books, e.g., (Alexandrescu, 2001).
Haskell has let us move beyond the mere curiosity of implementing OO idioms to
the point of making contributions to open and controversial OO problems. Haskell
has let us concisely specify and enforce the restrictions on the behaviour of object
constructors (preventing the constructor access not-yet-fully constructed objects).
The object encoding with recursive records can be made safe. Also, we were able to
effortlessly implement fine-grain notions of width and depth subtyping, with respect
to particular object operations, and thus safely permit methods with co-variant
argument subtyping. Not only OOHaskell is able to automatically compute the
least general interface of a heterogeneous collection of objects (through semi-implicit
76
O. Kiselyov and R. Lämmel
upcasts) and make the collection homogeneous, but it provides the means for safe
downcasts. Moreover, downcasts that cannot possibly succeed are flagged as type
errors. These are capabilities that go beyond state-of-the-art functional objectoriented programming with OCaml.
Just as C++ has become the laboratory for generative programming (Czarnecki et al., 2003) and lead to such applications as
FC++ (McNamara & Smaragdakis, 2004) and Boost (http://www.boost.org/),
we contend that (OO)Haskell would fit as the laboratory for advanced and typed
OO language design. All our experiments have shown that (OO)Haskell indeed
supports a good measure of experimentation — all without changing the type
system and the compiler.
Acknowledgements
We thank Keean Schupke for his major contributions to the HList and OOHaskell libraries. We thank Chung-chieh Shan for very helpful discussions. We also gratefully acknowledge feedback from Robin Green, Bryn Keller, Chris Rathman and several other
participants in mailing list or email discussions. The second author presented this work
at an earlier stage at the WG2.8 meeting (Functional Programming) in November 2004
at West Point. We are grateful for feedback received at this meeting.
References
Abadi, M., & Cardelli, L. (1996). A Theory of Objects. Monographs in Computer Science.
New York, NY: Springer-Verlag.
Abadi, M., Cardelli, L., Pierce, B., & Plotkin, G. 1989 (Jan.). Dynamic typing in a
statically-typed language. Pages 213–227 of: 16th ACM Conference on Principles of
Programming Languages.
Abadi, M., Cardelli, L., Pierce, B., & Plotkin, G. (1991). Dynamic typing in a staticallytyped language. TOPLAS, 13(2), 237–268.
Aho, A.V., Sethi, R., & Ullman, J.D. (1986). Compilers. principles, techniques and tools.
Addison-Wesley.
Alexandrescu, A. (2001). Modern C++ Design. Pearson Education.
Bayley, A. 2005 (June). Functional Programming vs. Object Oriented Programming.
Monad.Reader, http://www.haskell.org/tmrwiki/FpVsOo.
Bruce, K.B., & Mitchell, J.C. (1992). PER models of subtyping, recursive types and
higher-order polymorphism. Pages 316–327 of: POPL 1992: Proc. of the 19th ACM
SIGPLAN-SIGACT symposium on Principles of programming languages. ACM Press.
Bruce, K.B., Schuett, A., van Gent, R., & Fiech, A. (2003). PolyTOIL: A type-safe
polymorphic object-oriented language. TOPLAS, 25(2), 225–290.
Burton, F.W. (1990). Type extension through polymorphism. TOPLAS, 12(1), 135–138.
Cardelli, L., & Wegner, P. (1985). On Understanding Types, Data Abstraction, and
Polymorphism. ACM Computing Surveys, 17(4), 471–522.
Castagna, G. (1995). Covariance and contravariance: Conflict without a cause. TOPLAS,
17(3), 431–447.
Chen, K., Hudak, P., & Odersky, M. (1992). Parametric type classes. Pages 170–181 of:
Proceedings of the 1992 ACM Conference on LISP and Functional Programming. ACM
Press.
Haskell’s overlooked object system
— 1 February 2008 —
77
comp.lang.eiffel. 2004 (17 Apr.).
Frequently Asked Questions (FAQ).
http://www.faqs.org/faqs/eiffel-faq/.
Cook, W. R. (1989). A Denotational Semantics of Inheritance. Ph.D. thesis, Brown
University.
Cook, W.R., Hill, W., & Canning, P.S. (1990). Inheritance is not subtyping. Pages 125–
135 of: POPL ’90: Proceedings of the 17th ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages. New York, NY, USA: ACM Press.
Czarnecki, K., O’Donnell, J.T., Striegnitz, J., & Taha, W. (2003). DSL implementation
in MetaOCaml, Template Haskell, and C++. Pages 51–72 of: Lengauer, C., Batory,
D.S., Consel, C., & Odersky, M. (eds), Domain-specific program generation. LNCS, vol.
3016. Springer-Verlag.
Duck, G.J., Peyton Jones, S.L., Stuckey, P.J., & Sulzmann, M. (2004). Sound and Decidable Type Inference for Functional Dependencies. Pages 49–63 of: Schmidt, D.A.
(ed), Proceedings, 13th European Symposium on Programming, ESOP 2004, Barcelona,
Spain, March 29 - April 2, 2004. LNCS, vol. 2986. Springer-Verlag.
Finne, S., Leijen, D., Meijer, E., & Peyton Jones, S.L. (1999). Calling hell from heaven
and heaven from hell. Pages 114–125 of: ICFP ’99: Proceedings of the fourth ACM
SIGPLAN International Conference on Functional Programming. New York, NY, USA:
ACM Press.
Gamma, E., Helm, R., Johnson, R., & Vlissides, J. (1994). Design patterns: Elements of
reusable object-oriented software. Addison-Wesley.
Gaster, B.R., & Jones, M.P. 1996 (Nov.). A Polymorphic Type System for Extensible
Records and Variants. Technical report NOTTCS-TR-96-3. University of Nottingham,
Department of Computer Science.
H.
Nilsson.
(2003).
The
Future
of
Haskell
discussion
at
the
Haskell
Workshop.
Message
on
the
Haskell
mailing
list,
http://www.mail-archive.com/haskell\spacefactor\@mhaskell.org/msg13366.html.
Hallgren, T. (2001).
Fun with functional dependencies.
Joint Winter Meeting of the Departments of Science and Computer Engineering, Chalmers University of Technology and Goteborg University, Varberg, Sweden, Jan. 2001.
http://www.cs.chalmers.se/~hallgren/Papers/wm01.html.
Howard, M., Bezault, E., Meyer, B., Colnet, D., Stapf, E., Arnout, K., & Keller, M. 2003
(27 Apr.). Type-safe covariance: Competent compilers can catch all catcalls. Work done
as part of Eiffel language standardization by the TC39-TG4 committee of ECMA; Draft;
Available at http://www.inf.ethz.ch/~meyer/ongoing/covariance/recast.pdf.
Hughes,
J.
2002
(23
Aug.).
Re:
More
suitable
data
structure
needed.
Message
on
the
Haskell
mailing
list,
http://www.haskell.org/pipermail/haskell/2002-August/010335.html.
Hughes, J., & Sparud, J. (1995). Haskell++: An Object-Oriented Extension of Haskell.
Proc. of Haskell Workshop, La Jolla, California. YALE Research Report DCS/RR-1075.
Jones, M.P. (1992). A theory of qualified types. Pages 287–306 of: Symposium proceedings
on 4th European Symposium on Programming. Springer-Verlag.
Jones, M.P. (1995). Simplifying and improving qualified types. Pages 160–169 of: Proceedings of the 7th International Conference on Functional Programming Languages and
Computer Architecture. ACM Press.
Jones, M.P. (2000). Type Classes with Functional Dependencies. Pages 230–244 of:
Proceedings of the 9th European Symposium on Programming Languages and Systems.
Springer-Verlag.
Kiselyov, O., Lämmel, R., & Schupke, K. (2004).
Strongly typed heterogeneous collections.
ACM SIGPLAN Workshop on Haskell.
ACM Press.
See
78
O. Kiselyov and R. Lämmel
http://www.cwi.nl/~ralf/HList/ for an extended technical report and for the source
distribution.
Lämmel, R., & Peyton Jones, S.L. (2003). Scrap your boilerplate: a practical design
pattern for generic programming. ACM SIGPLAN Notices, 38(3), 26–37. Proc. of the
ACM SIGPLAN Workshop TLDI 2003.
Launchbury, J., & Peyton Jones, S.L. (1995). State in Haskell. Lisp and Symbolic Computation, 8(4), 293–342.
Leroy, Xavier, et al. . 2004 (July). The Objective Caml system, release 3.08, Documentation
and user’s manual. http://caml.inria.fr/ocaml/htmlman/index.html.
Liskov, B., & Wing, J.M. (1994). A Behavioral Notion of Subtyping. TOPLAS, 16(6),
1811–1841.
McBride, C. (2002). Faking It (Simulating Dependent Types in Haskell). Journal of
Functional Programming, 12(4–5), 375–392.
McNamara, B., & Smaragdakis, Y. (2004). Functional Programming with the FC++
Library. Journal of Functional Programming, 14(4), 429–472.
Meijer, E., & Claessen, K. (1997). The Design and Implementation of Mondrian. ACM
SIGPLAN Haskell Workshop. ACM Press.
Neubauer, M., Thiemann, P., Gasbichler, M., & Sperber, M. (2001). A Functional Notation
for Functional Dependencies. Pages 101–120 of: Proc. 2001 ACM SIGPLAN Haskell
Workshop, Firenze, Italy, September 2001.
Neubauer, M., Thiemann, P., Gasbichler, M., & Sperber, M. (2002). Functional logic
overloading. Pages 233–244 of: Proceedings of the 29th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages. ACM Press.
Nordlander, J. (1998). Pragmatic Subtyping in Polymorphic Languages. Pages 216–227 of:
Berman, M., & Berman, S. (eds), Proceedings of the third ACM SIGPLAN international
conference on functional programming (ICFP-98). ACM SIGPLAN Notices, vol. 34, 1.
New York: ACM Press.
Nordlander, J. (2002). Polymorphic subtyping in O’Haskell. Science of computer programming, 43(2–3), 93–127. Also in the Proceedings of the APPSEM Workshop on
Subtyping and Dependent Types in Programming, Ponte de Lima, Portugal, 2000.
Ohori, A. (1995). A polymorphic record calculus and its compilation. TOPLAS, 17(6),
844–895.
Pang, A.T.H., & Chakravarty, M.M.T. (2004). Interfacing Haskell with Object-Oriented
Languages. Pages 20–35 of: P.W. Trinder and G. Michaelson and R. Pena (ed), Implementation of Functional Languages, 15th International Workshop, IFL 2003, Edinburgh,
UK, September 8-11, 2003, Revised Papers. LNCS, vol. 3145. Springer-Verlag.
Peyton Jones, S.L. (2001). Tackling the Awkward Squad: monadic input/output, concurrency, exceptions, and foreign-language calls in Haskell. Pages 47–96 of: Hoare, C.A.R.,
Broy, M., & Steinbrueggen, R. (eds), Engineering theories of software construction,
Marktoberdorf Summer School 2000. NATO ASI Series. IOS Press.
Peyton Jones, S.L., & Wadler, P. 1993 (Jan.). Imperative functional programming. Pages
71–84 of: the symposium on principles of programming languages (POPL ’93). ACM.
Peyton Jones, S.L., Jones, M.P., & Meijer, E. (1997). Type classes: exploring the design
space. Launchbury, J. (ed), Haskell Workshop.
Pierce, B.C., & Turner, D.N. (1994). Simple Type-Theoretic Foundations for ObjectOriented Programming. Journal of functional programming, 4(2), 207–247.
Rémy, D. (1994). Programming Objects with ML-ART: An extension to ML with Abstract
and Record Types. Pages 321–346 of: Hagiya, M., & Mitchell, J.C. (eds), International
Haskell’s overlooked object system
— 1 February 2008 —
79
Symposium on Theoretical Aspects of Computer Software. LNCS, no. 789. Sendai,
Japan: Springer-Verlag.
Rémy, D., & Vouillon, J. (1997). Objective ML: a simple object-oriented extension of
ML. Pages 40–53 of: POPL ’97: Proceedings of the 24th ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages. New York, NY, USA: ACM
Press.
Shields, M., & Peyton Jones, S.L. (2001).
Object-Oriented Style Overloading for Haskell.
ENTCS, 59(1).
An extended TR is available from
http://research.microsoft.com/Users/simonpj/Papers/oo-haskell/.
Siek, J., Gregor, D., Garcia, R., Willcock, J., Järvi, J., & Lumsdaine, A. (2005). Concepts
for C++0x. Tech. rept. N1758=05-0018. ISO/IEC JTC 1, Information Technology,
Subcommittee SC 22, Programming Language C++.
Stuckey, P. J., & Sulzmann, M. (2005). A Theory of Overloading. TOPLAS. To appear.
Stuckey, P. J., Sulzmann, M., & Wazny, J. (2004). Improving type error diagnosis. Pages
80–91 of: Proc. of haskell’04. ACM Press.
Surazhsky, V., & Gil, J.Y. (2004). Type-safe covariance in C++. Pages 1496–1502 of:
SAC ’04: Proceedings of the 2004 ACM Symposium on Applied Somputing. ACM Press.
Taha, W., & Nielsen, M.F. (2003). Environment classifiers. Pages 26–37 of: POPL ’03:
Proceedings of the 30th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages. New York, NY, USA: ACM Press.
Ungar, D., & Smith, R. B. (1987). Self: The power of simplicity. Pages 227–242 of:
OOPSLA 1987: Conference Proceedings on Object-Oriented Programming Systems, Languages and Applications. ACM Press.
Zenger, M., & Odersky, M. 2004 (Mar.). Independently Extensible Solutions to the Expression Problem. Tech. rept. Ecole Polytechnique Fdérale de Lausanne. Technical Report
Nr. 200433.
| 2cs.AI
|
1
Minimax Rate-Optimal Estimation of Divergences
between Discrete Distributions
Yanjun Han, Student Member, IEEE, Jiantao Jiao, Student Member, IEEE, and Tsachy Weissman, Fellow, IEEE.
arXiv:1605.09124v2 [cs.IT] 24 Nov 2016
Abstract
We refine the general methodology in [1] for the construction and analysis of essentially minimax estimators for a wide class
of functionals of finite dimensional parameters, and elaborate on the case of discrete distributions with support size S comparable
with the number of observations n. Specifically, we determine the “smooth” and “non-smooth” regimes based on the confidence
set and the smoothness of the functional. In the “non-smooth” regime, we apply an unbiased estimator for a suitable polynomial
approximation of the functional. In the “smooth” regime, we construct a general version of the bias-corrected Maximum Likelihood
Estimator (MLE) based on Taylor expansion.
We apply the general methodology to the problem of estimating the KL divergence between two discrete probability measures
P and Q from empirical data in a non-asymptotic and possibly large alphabet setting. We construct minimax rate-optimal estimators
for D(P kQ) when the likelihood ratio is upper bounded by a constant which may depend on the support size, and show that
the performance of the optimal estimator with n samples is essentially that of the MLE with n ln n samples. Our estimator is
adaptive in the sense that it does not require the knowledge of the support size nor the upper bound on the likelihood ratio. We
show that the general methodology results in minimax rate-optimal estimators for other divergences as well, such as the Hellinger
distance and the χ2 -divergence. Our approach refines the Approximation methodology recently developed for the construction of
near minimax estimators of functionals of high-dimensional parameters, such as entropy, Rényi entropy, mutual information and
`1 distance in large alphabet settings, and shows that the effective sample size enlargement phenomenon holds significantly more
widely than previously established.
Index Terms
Divergence estimation, KL divergence, multivariate approximation theory, Taylor expansion, functional estimation, maximum
likelihood estimator, high dimensional statistics, minimax lower bound
I. I NTRODUCTION
Given jointly independent m samples from P = (p1 , · · · , pS ) and n samples from Q = (q1 , · · · , qS ) over some unknown
common alphabet of size S, consider the problem of estimating a functional of the distribution of the following form:
F (P, Q) =
S
X
f (pi , qi )
(1)
i=1
where f : A → R is a continuous function with some A ⊂ [0, 1]2 . Note that by allowing f to solely depend on p, this
problem generalizes the functional estimation problem considered in [1]. Among the most fundamental of such functionals is
the f -divergence [2]
Z
S
X
pi
dP
f
Df (P kQ) = f
dQ =
qi
(2)
dQ
qi
i=1
for some convex function f with f (1) = 0. The f -divergence serves as the fundamental information contained in binary
statistical models [3] and enjoys numerable applications in information theory [4] and statistics [5].
Among many f -divergences, we focus on the estimation problem of the Kullback–Leibler (KL) divergence (with f (t) =
t ln t − t + 1) in this paper, and the general approach naturally extends to the Hellinger distance and χ2 -divergence. The
KL divergence is an important measure of the discrepancy between two discrete distributions P = (p1 , · · · , pS ) and Q =
(q1 , · · · , qS ), defined as [6]
(PS
pi
if P Q,
i=1 pi ln qi
(3)
D(P kQ) =
+∞
otherwise,
where P Q denotes that the absolute continuity of P with respect to Q. Like the entropy and mutual information [7], the KL
divergence is a key information theoretic measure arising naturally in data compression [8], communications [9], probability
Yanjun Han, Jiantao Jiao, and Tsachy Weissman are with the Department of Electrical Engineering, Stanford University, CA, USA. Email: {yjhan, jiantao,
tsachy}@stanford.edu. The material in this paper was presented in part at the 2016 IEEE International Symposium on Information Theory and Its Applications
(ISITA), Monterey, USA.
2
theory [10], statistics [11], optimization [12], machine learning [13], [14], and many other disciplines. Throughout the paper
we use the squared error loss, i.e., the risk function for any estimator D̂ is defined as
L(D̂; P, Q) , E(P,Q) |D̂ − D(P kQ)|2 .
The maximum risk of an estimator D̂, and the minimax risk in estimating D(P kQ) are defined as
Rmaximum (D̂; U) ,
sup L(D̂; P, Q),
(4)
(5)
(P,Q)∈U
Rminimax (U) , inf
sup L(D̂; P, Q)
(6)
D̂ (P,Q)∈U
respectively, where U is a given collection of probability measures (P, Q), and the infimum is taken over all possible estimators
D̂. We aim to obtain the minimax risk Rminimax (U) for some properly chosen U.
Notations: for non-negative sequences aγ , bγ , we use the notation aγ . bγ to denote that there exists a universal constant
a
C such that supγ bγγ ≤ C, and aγ & bγ is equivalent to bγ . aγ . Notation aγ bγ is equivalent to aγ . bγ and bγ . aγ .
a
Notation aγ bγ means that lim inf γ bγγ = ∞, and aγ bγ is equivalent to bγ aγ . We write a ∧ b = min{a, b} and
d
a ∨ b = max{a, b}. Moreover, polyn denotes the set of all d-variate polynomials of degree no more than n, and En [f ; I]
denotes the distance of the function f to the space polydn in the uniform norm k · k∞,I on I ⊂ Rd . All logarithms are in the
natural base.
A. Background and main results
There have been several attempts to estimate the KL divergence for the continuous case, see [15]–[20] and references therein.
These approaches usually do not operate in the minimax framework, and focus on consistency but not rates of convergence,
unless strong smoothness conditions on the densities are imposed to achieve the parametric rate (i.e., Θ(n−1 ) in mean squared
error). In the discrete setting, [21] and [22] proved consistency of some specific estimators without arguing minimax optimality.
We note that in the discrete case, if the alphabet size S is fixed and the number of samples m, n go to infinity, the standard
Hájek–Le Cam theory of classical asymptotics shows that the plug-in approach is asymptotically efficient [23, Thm. 8.11,
Lemma 8.14]. The key challenge we face in the discrete setting is the regime where the support size S can be comparable to
or even larger than the number of observations m, n, which classical analyses do not address.
Now we consider the estimation of KL divergence between discrete distributions in a large-alphabet setting. For the choice
of U, it may appear natural to allow P to be any distribution which is absolutely continuous with respect to Q with alphabet
size S, i.e.,
US = {(P, Q) : P, Q ∈ MS , P Q}
(7)
where MS denotes the set of all probability measures with support size S. However, in this case, it turns out to be impossible
to estimate the KL divergence in the minimax sense, i.e., Rminimax (US ) = ∞ for any configuration (S, m, n) with S ≥ 2.
Intuitively, this is because that the observation from the Multinomial model depends continuously on (P, Q) while the KL
divergence does not at extremal points. A rigorous statement and proof of this result is given in Lemma 21 of the Appendix.
It seems natural then to consider an alternative uncertainty set with bounded likelihood ratio:
pi
US,u(S) = {(P, Q) : P, Q ∈ MS ,
≤ u(S), ∀i}
(8)
qi
where u(S) ≥ 1 is an upper bound on the likelihood ratio. Since u(S) = 1 results in the trivial case where D(P kQ) ≡ 0,
throughout we will assume that u(S) ≥ c for some constant c > 1.
The main result of this paper is as follows.
Theorem 1. For m & S/ ln S, n & Su(S)/ ln S, u(S) & (ln S)2 and ln S & ln(m ∨ n), we have
2
S
Su(S)
(ln u(S))2
u(S)
+
+
+
.
Rminimax (US,u(S) )
m ln m
n ln n
m
n
(9)
Furthermore, our estimator D̂A in Section III achieves this bound under the Poisson sampling model, and is adaptive in the
sense that it does not require knowledge of S or u(S).
The following corollary is a direct consequence of Theorem 1. Note that ln S & ln n and n & Su(S)/ ln S have already
implied that ln S & ln u(S), and thus ln(Su(S)) ln S.
Corollary 1. For our KL divergence estimator, the maximum mean squared error vanishes provided that m S/ ln S and
n Su(S)/ ln S. Moreover, if m . S/ ln S or n . Su(S)/ ln S, then the maximum risk of any estimator for KL divergence
is bounded away from zero.
3
Next we consider the plug-in approach in the context of minimax rate-optimality. Since it is possible that p̂i > 0 and q̂i = 0
for some i ∈ {1, · · · , S}, where Pm = (p̂1 , · · · , p̂S ), Qn = (q̂1 , · · · , q̂S ) are the respective empirical probability distributions,
the direct plug-in estimate D(Pm kQn ) may be infinity with positive probability. Hence, we use the following modification of
the direct plug-in approach: when we observe that p̂i > 0 and q̂i = 0, since naturally q̂i is an integral multiple of 1/n, we
manually change the value of q̂i to the closed lattice 1/n of zero. More precisely, we define
1
1
0
∨ q̂1 , · · · , ∨ q̂S
(10)
Qn =
n
n
and use the estimator D(Pm kQ0n ) to estimate the KL divergence. Note that Q0n may not be a probability distribution (in which
case D(Pm kQ0n ) is extended in the obvious way). The performance of this modified plug-in approach is summarized in the
following theorem.
Theorem 2. Under the Poisson sampling model, the modified plug-in estimator D(Pm kQ0n ) satisfies
2
(ln u(S))2
Su(S)
u(S)
S
+
Rmaximum (D(Pm kQ0n ); US,u(S) ) .
+
+
.
m
n
m
n
Moreover, for m ≥ 15S and n ≥ 4Su(S), we have
Rmaximum (D(Pm kQ0n ); US,u(S) )
&
Su(S)
S
+
m
n
2
+
(ln u(S))2
u(S)
+
.
m
n
(11)
(12)
The following corollary on the minimum sample complexity is immediate.
Corollary 2. The worst-case mean squared error of the modified plug-in estimator D(Pm kQ0n ) vanishes if and only if m
S ∨ (ln u(S))2 and n Su(S).
Hence, compared with the mean squared error or the minimum sample complexity of the modified plug-in approach, the
optimal estimator enjoys a logarithmic improvement. Note that (ln u(S))2 . (ln S)2 S is negligible under the condition in
Theorem 1, so there is no counterpart of m (ln u(S))2 in Corollary 1. Specifically, the performance of the optimal estimator
with (m, n) samples is essentially that of the plug-in approach with (m ln m, n ln n) samples, which is another manifestation
of the effective sample size enlargement phenomenon [1], [24]. Note that in the KL divergence example, the modified plug-in
estimator D(Pm kQ0n ) essentially exploits the plug-in idea.
After our submission of this work to arXiv, an independent study of the same problem was presented in ISIT 2016 [25]
without the construction of the optimal estimator, which was added to the full version [26] that appeared later on arXiv.
Specifically, the main result (i.e., Theorem 1) was also obtained in [26], while there are some differences. First, our estimator
is agnostic to both the support size S and the upper bound u(S) on the likelihood ratio, while the estimator in [26] requires
2
both. Second, as for Theorem 2, there is an unnecessary additional term (lnmS) in the upper bound (11) of the plug-in approach
in [26], though there is a minor difference between our choices of the plug-in estimator. Third, and most significant, [26] is
dedicated exclusively to the KL divergence case, while in our paper we propose a general approximation-based methodology
for the estimation of a wide class of functionals, with the estimation of KL divergence serving as the main example for concrete
illustration of the concepts. As additional examples, following the general recipe in the next subsection and the later analysis,
the result on estimating the L1 distance in [24] can be recovered, and for the Hellinger distance and the χ2 -divergence [2]
S
H 2 (P, Q) ,
χ2 (P, Q) ,
1X √
√ 2
( pi − qi )
2 i=1
(PS p2
i
i=1 qi − 1 if P Q,
+∞
otherwise
(13)
(14)
we can similarly obtain the following results on the optimal estimation rates in the large-alphabet setting.
Theorem 3. For m ∧ n & S/ ln S and ln S & ln(m ∨ n), for Hellinger distance we have
2
1
S
+
inf
sup
E(P,Q) T̂ − H 2 (P, Q)
(m ∧ n) ln(m ∧ n) m ∧ n
T̂ P ∈MS ,Q∈MS
(15)
and the estimator in Section V achieves this bound without the knowledge of S under the Poisson sampling model.
Theorem 4. For n &
S(u(S))2
, u(S)
ln S
& (ln S)2 and ln S & ln(m ∨ n), for χ2 -divergence we have
2 S(u(S))2 2 (u(S))2
(u(S))3
2
inf
sup
E(P,Q) T̂ − χ (P, Q)
+
+
n ln n
m
n
T̂ (P,Q)∈US,u(S)
(16)
4
and the estimator in Section V achieves this bound without the knowledge of either S or u(S) under the Poisson sampling
model.
The following corollaries on the minimum sample complexities follow directly from the previous theorems.
Corollary 3. For Hellinger distance over (P, Q) ∈ MS × MS , there exists an estimator with a vanishing maximum mean
squared error if and only if m ∧ n lnSS .
Corollary 4. For χ2 -divergence over (P, Q) ∈ US,u(S) , there exists an estimator with a vanishing maximum mean squared
2
error if and only if m (u(S))2 and n S(u(S))
∨ (u(S))3 .
ln S
B. Approximation: the general recipe
Estimation of KL divergence belongs to a large family of functional estimation problems: consider estimating the functional
G(θ) of a parameter θ ∈ Θ ⊂ Rp for an experiment {Pθ : θ ∈ Θ}. There
Pn has been a recent wave of study on functional
estimation of high dimensional parameters, e.g., the scaled `1 norm n1 i=1 |θi | in the Gaussian model [27], the Shannon
PS
PS
α
entropy
i=1 −pi ln pi [1], [28]–[30], the mutual information [1], the power sum function
i=1 pi [1], the Rényi entropy
PS
α
PS
ln i=1 pi
[31] and the `1 distance i=1 |pi − qi | [24] in Multinomial and Poisson models. Moreover, the effective sample size
1−α
enlargement phenomenon holds in all these examples: the performance of the minimax estimators with n samples is essentially
that of the plug-in approach with n ln n samples.
The optimal estimators in the previous examples all follow the general methodology of Approximation proposed in [1]:
suppose θ̂n is a consistent estimator for θ, where n is the number of observations. Suppose the functional G(θ) is analytic1
everywhere except at θ ∈ Θ0 . A natural estimator for G(θ) is G(θ̂n ), and we know from classical asymptotics [23, Lemma
8.14] that given the benign LAN (Local Asymptotic Normality) condition [23], G(θ̂n ) is asymptotically efficient for G(θ) for
θ∈
/ Θ0 if θ̂n is asymptotically efficient for θ. In the estimation of functionals of discrete distributions, Θ is the S-dimensional
probability simplex, and a natural candidate for θ̂n is the empirical distribution, which is unbiased for any θ ∈ Θ. Then the
following two-step procedure is conducted in estimating G(θ).
1) Classify the Regime: Compute θ̂n , and declare that we are in the “non-smooth” regime if θ̂n is “close” enough to Θ0 .
Otherwise declare we are in the “smooth” regime;
2) Estimate:
• If θ̂n falls in the “smooth” regime, use an estimator “similar” to G(θ̂n ) to estimate G(θ);
• If θ̂n falls in the “non-smooth” regime, replace the functional G(θ) in the “non-smooth” regime by an approximation
Gappr (θ) (another functional) which can be estimated without bias, then apply an unbiased estimator for the functional
Gappr (θ).
Simple as it may sound, this methodology has a few drawbacks and ambiguities. In our recent work [24], we applied this
general recipe to the estimation of `1 distance between two discrete distributions, where this recipe proves to be inadequate.
In the estimation of the `1 distance, a bivariate function f (x, y) = |x − y| which is non-analytic in a segment was considered,
which is completely different from the previous studies [1], [28]–[31] where a univariate function analytic everywhere except
a point is always taken into consideration. In particular, two more topics, i.e., multivariate approximation and localization via
confidence sets, were introduced and used.
Question 1. What if the domain of θ̂n is different from (usually larger than) Θ, the domain of θ?
Question 2. How to determine the “non-smooth” regime? What is its size?
Question 3. If θ̂n falls in the “non-smooth” regime, in which region should Gappr (θ) be a good approximation of G(θ) (e.g.,
the whole domain Θ, or a proper neighborhood of θ̂n )?
Question 4. If θ̂n falls in the “smooth” regime, how to construct an estimator “similar” to G(θ̂n )?
Other questions, such as what type/degree of approximation Gappr (θ) should be used, were answered in more detail in [1].
Among these questions, Question 1 is a relatively new one, where the estimation of KL divergence is the second example
so far for which it has arisen, where the first example on estimating the support size of a discrete distribution [32] did not
explicitly propose and answer this question. Question 2 and 3 were partially addressed in [1] and [24], but the answer to
Question 2 changes in view of Question 1, and further elaborations are also necessary for Question 3. As for Question 4,
the previous approaches can only handle order-one bias correction, while for some problems bias correction with an arbitrary
order is proved to be necessary [33]. Before answering these questions, we begin with a formal definition of confidence set in
statistical experiments, which is motivated by [24].
1A
function f is analytic at a point x0 if and only if its Taylor series about x0 converges to f in some neighborhood of x0 .
5
Definition 1 (Confidence set). Consider a statistical model (Pθ )θ∈Θ and an estimator θ̂ ∈ Θ̂ of θ, where Θ ⊂ Θ̂. For r ∈ [0, 1],
a confidence set of significance level r is a collection of sets {U (x)}x∈Θ̂ , where U (x) ⊂ Θ for any x ∈ Θ̂, and2
(17)
sup Pθ (θ ∈
/ U (θ̂)) ≤ r.
θ∈Θ
Moreover, every confidence set of significance level r can also induce a reverse confidence set {V (y)}y∈Θ of significance level
r, where V (y) , {x ∈ Θ̂ : y ∈ U (x)} ⊂ Θ̂ for any y ∈ Θ, and
(18)
sup Pθ (θ̂ ∈
/ V (θ)) ≤ r.
θ∈Θ
Intuitively, if {U (x)}x∈Θ̂ is a confidence set of significance level r, then after observing θ̂ we can conclude that θ ∈ U (θ̂)
with error probability at most r. More precisely, for any θ ∈ Θ, with probability at least 1 − r, we can get back to θ based on
U (·) after observing θ̂. Conversely, with probability at least 1 − r, we can also restrict θ̂ in the region V (θ). In other words,
the true parameter θ is localized at U (θ̂), and the observation θ̂ is localized at V (θ), from which the name localization via
confidence sets originates. Note that confidence set of any level exists for any statistical model (Pθ )θ∈Θ and estimator θ̂, since
U (x) ≡ Θ is always a feasible confidence set of level zero (and then V (y) ≡ Θ̂). In practice, we seek confidence sets which
5
are as small as possible. We also remark that, apart from the confidence set used in traditional hypothesis testing where r is
−A
usually chosen to be a fixed constant (e.g., 0.01), here we allow r to decay with n, e.g., rn n
with some constant A > 0.
−A
For
example,
in
the
Binomial
model
np̂
∼
B(n,
p)
with
Θ̂
=
[0,
1]
and
any
Θ
⊂
Θ̂,
for
r
n
,
by measure
concentration
−A
n
example, in the Binomial model np̂ ∼ B(n, p) with Θ̂ = [0, 1] and any Θ ⊂ Θ̂, for rn n , by measure
concentration
(cf.
(cf.
Lemma
28
in
Appendix
A)
the
collection
{U
(x)}
with
x∈[0,1]
Lemma 26 in Appendix A) the collection {U (x)}x∈[0,1]
with
ln n
ln n
[0, cc11ln
≤ cc11ln
n n ]]
n n ,,
[0, nq
ifif xx ≤
n
q
(19)
U(x)
(x) =
=Θ
Θ∩
∩
q
U
(17)
ln n
ln n
c ln n
[x − q cc11xxln
x+
+ cc11xxnln
< xx ≤
≤ 1.
1
n ] if c11 ln
n n ,,x
[x −
] if n n <
n
n
n
is aa (1
confidence
set of significance
level
assuming
the universal
c1 > 0 and
is large
enough, reverse
and theconfidence
induced reverse
is
− rn )-confidence
set assuming
thernuniversal
constant
c1 > 0 constant
is large enough,
the induced
set is
confidence
set
is
contained
in
contained in
4c1 ln n
2c1 ln n
[0, 4c
≤ 2c
ln n ,
ln n ]
1n
n
]
ifif yy ≤
,
[0, 1q
q
n
n
V
(y)
=
[0,
1]
∩
(20)
q
q
(18)
V (y) = [0, 1] ∩ [y − 2c1 y ln n , y + 2c1 y ln n ] if 2c1 ln n < y ≤ 1
[y − 2c1 ynln n , y + 2c1 ynln n ] if 2c1 nln n < y ≤ 1.
n
n
n
Θ̂
V (θ)
θ = (p1 , p2 )
q
l2 ∼ p1 nln n
w2 ∼
p̂2 ln n
n
U (θ̂)
p̂1 ln n
n
U (θ̂)
q
q
θ̂ = (p̂1 , p̂2 )
V (θ)
θ
w1 ∼
l1 ∼
Θ̂ = Θ
q
Θ
p2 ln n
n
which is
is of
of aa similar
similar structure.
structure. Figure
Figure 11 gives
gives aa pictorial
pictorial illustration
illustration of
of both
both the
the confidence
confidence set
set and
and the
the reverse
reverse confidence
confidence set
set
which
in
2D
Binomial
and
Gaussian
models,
respectively.
in 2D Binomial and Gaussian models, respectively.
θ̂
√
r2 ∼ σ ln n
√
r1 ∼ σ ln n
Fig.
Fig. 1:
1: Pictorial
Pictorial illustration
illustration of
of confidence
confidence set
set U
U ((θ̂)
θ̂) and
and reverse
reverse confidence
confidence set
set V
V (θ)
(θ) in
in 2D
2D Binomial
Binomial (left
(left panel)
panel) and
and Gaussian
Gaussian
(right
panel)
models.
In
the
Binomial
model,
n(p̂
,
p̂
)
∼
B(n,
p
)
×
B(n,
p
)
with
(p
,
p
)
∈
Θ,
(p̂
,
p̂
)
∈
Θ̂
and
(right panel) models. In the Binomial model, n(p̂11 , p̂22 ) ∼ B(n, p11 ) × B(n, p22 ) with (p11 , p22 ) ∈ Θ, (p̂11 , p̂22 ) ∈ Θ̂ and Θ
Θ⊂
⊂ Θ̂.
Θ̂. In
In
2
2 I2 ) with θ ∈ Θ, θ 0 ∈ Θ̂ and Θ = Θ̂.
the
Gaussian
model,
θ̂
∼
N
(θ,
σ
the Gaussian model, θ̂ ∼ N (θ, σ I2 ) with θ ∈ Θ, θ̂ ∈ Θ̂ and Θ = Θ̂.
Now
Now we
we provide
provide answers
answers to
to these
these questions
questions with
with the
the help
help of
of localization
localization via
via confidence
confidence sets.
sets.
1)
Question
1:
When
we
consider
the
non-analytic
region
of
G(·),
we
should
always
stick
1) Question 1: When we consider the non-analytic region of G(·), we should always stick to
to the
the domain
domain of
of θ̂θ̂nn instead
instead of
of
that
of
the
true
parameter
θ
(for
the
existence
of
G(
θ̂
),
here
we
assume
that
G(·)
is
well-defined
on
the
Θ̂
⊃
n
that of the true parameter θ (for the existence of G(θ̂n ), here we assume that G(·) is well-defined on the Θ̂ ⊃ Θ,
Θ, where
where
Θ̂
Θ̂ is
is the
the domain
domain of
of θ̂θ̂nn).
). In
In fact,
fact, we
we should
should distinguish
distinguish the
the “smooth”
“smooth” (resp.
(resp. “non-smooth”)
“non-smooth”) regime
regime of
of θθ and
and that
that of
of θ̂θ̂nn::
we
determine
the
corresponding
regimes
of
θ
first,
and
then
localize
θ
using
θ̂
since
θ
cannot
be
observed.
Hence,
n
we determine the corresponding regimes of θ first, and then localize θ using θ̂n since θ cannot be observed. Hence, in
in
the
the first
first step,
step, to
to make
make the
the plug-in
plug-in approach
approach G(
G(θ̂θ̂nn)) work
work for
for the
the estimation
estimation of
of G(θ),
G(θ), it
it must
must be
be ensured
ensured that
that with
with high
high
probability
probability θ̂θ̂nn does
does not
not fall
fall into
into the
the non-analytic
non-analytic region
region of
of G(·),
G(·), which
which is
is defined
defined over
over Θ̂
Θ̂ instead
instead of
of Θ.
Θ. As
As aa result,
result, the
the
non-analytic
region
of
G(·)
over
the
domain
of
θ̂
is
the
correct
region
to
consider.
n
non-analytic region of G(·) over the domain of θ̂n is the correct region
to
consider.
2) Question 2: We first determine the “smooth” regime Θs of θ. Let Θ00 ⊂ Θ̂ be the non-analytic region of G(·) over Θ̂.
2 In standard
By theterminology
previous answer
to testing,
Question
1, also
Θs the
should
be set
in statistical
this is
confidence
set to
of level 1 − r.
Θs , {θ ∈ Θ : Pθ (θ̂n ∈ Θ00 ) ≤ rn }
(19)
where the convergence rate rn (e.g., rn n−A for some constant A > 0) depends on the specific problem. Usually rn
can be of any negligible order compared to the minimax risk of the estimation problem. With the help of localization
6
2) Question 2: We first determine the “smooth” regime Θs of θ. Let Θ̂0 ⊂ Θ̂ be the non-analytic region of G(·) over Θ̂.
By the previous answer to Question 1, Θs should be set to
Θs , {θ ∈ Θ : Pθ (θ̂n ∈ Θ̂0 ) ≤ rn }
(21)
sup Pθ (θ̂n ∈ Θ̂0 ) ≤ sup Pθ (θ ∈
/ U (θ̂n )) ≤ sup Pθ (θ ∈
/ U (θ̂n )) ≤ rn
(22)
where the convergence rate rn (e.g., rn n−A for some constant A > 0) depends on the specific problem. Usually rn
can be of any negligible order compared to the minimax risk of the estimation problem. With the help of localization
via confidence sets, we can just set Θs = Θ − ∪x∈Θ̂0 U (x) for any confidence set {U (x)}x∈Θ̂ of significance level rn .
In fact, if θ ∈ Θs and θ̂n ∈ Θ̂0 , we have U (θ̂n ) ⊂ ∪x∈Θ̂0 U (x) = Θ − Θs and thus θ ∈
/ U (θ̂n ). As a result, by definition
of the confidence set we have
θ∈Θs
θ∈Θs
θ∈Θ
as desired. By taking complement we obtain the “non-smooth” regime Θns , Θ − Θs of θ.
Since we cannot observe θ, we need to determine the “smooth” regime Θ̂s based on θ̂n rather than θ. A natural choice
is given by confidence set: Θ̂s , {θ̂n ∈ Θ̂ : Θs ⊃ U (θ̂n )}, i.e., Θ̂s contains all observations whose confidence set for
the true parameter falls into the “smooth” regime Θs . Likewise, we can define Θ̂ns , {θ̂n ∈ Θ̂ : Θns ⊃ U (θ̂n )} for the
“non-smooth” regime based on θ̂n . Since Θs ∩ Θns = ∅, it can be easily seen that Θ̂s ∩ Θ̂ns = ∅ as well, but one problem
is that Θ̂ns ∪ Θ̂s $ Θ̂, i.e., some observation θ̂n is attributed to neither the “non-smooth” regime nor the “smooth” regime.
To solve this problem, we should expand Θs and Θns a little bit to ensure that Θ̂s and Θ̂ns form a partition of Θ̂.
In fact, this expansion can be done in many statistical models with satisfactory measure concentration properties (e.g.,
(2)
(1)
in Multinomial, Poisson and Gaussian models). Specifically, for some proper rn ≥ rn of order both negligible to
(2)
(1)
that of the minimax risk, there exists confidence sets {U1 (x)}x∈Θ̂ and {U2 (x)}x∈Θ̂ of significance level rn and rn ,
respectively, such that
Θ(1)
, Θ − ∪x∈Θ̂0 U1 (x)
s
Θ(2)
ns
, ∪x∈Θ̂0 U2 (x)
Θ̂s , {θ̂n ∈ Θ̂ :
Θ̂ns , {θ̂n ∈ Θ̂ :
Θ(1)
s
Θ(2)
ns
(23)
(24)
⊃ U1 (θ̂n )}
⊃ U2 (θ̂n )}
(25)
(26)
satisfy that Θ̂s ∪ Θ̂ns = Θ̂ (by passing through subsets it does not matter if Θ̂s ∩ Θ̂ns 6= ∅). Note that in this case we
(1)
(2)
must have Θs ∩ Θns 6= ∅, i.e., there exists some θ which belongs to both the “smooth” regime and the “non-smooth”
regime.
(1)
The interpretation of this approach is as follows. If the true parameter θ falls in the “smooth” regime Θs , then the plug(2)
in approach will work; conversely, if the true parameter θ falls in the “non-smooth” regime Θns , then the approximation
(1)
(2)
idea will work. Then Θs ∩ Θns 6= ∅ implies that there exists an intermediate regime such that both the plug-in approach
and the approximation approach work when θ falls into this regime. This intermediate regime is unnecessary when we
(1)
(2)
are given the partial information whether θ ∈ Θs or θ ∈ Θns , but it becomes important when we need to infer this
partial information based on θ̂n . Our target is as follows: if the true parameter θ does not fall in the “smooth” (resp.
“non-smooth”) regime, then with high probability we will also declare based on θ̂n that we are not in the “smooth” (resp.
(1)
(2)
“non-smooth”) regime. Mathematically, with high probability, θ ∈ Θ − Θs implies θ̂n ∈ Θ̂ns , and θ ∈ Θ − Θns implies
(2)
(1)
θ̂n ∈ Θ̂s . Note that if θ ∈ Θs ∩ Θns falls in the intermediate regime, either θ̂n ∈ Θ̂s or θ̂n ∈ Θ̂ns suffices for our
(1)
estimator to perform well. The key fact is that this target is fulfilled by the definition of confidence sets: if θ ∈ Θ − Θs
(1)
and θ̂n ∈
/ Θ̂ns , we have θ̂n ∈ Θ̂s , and by definition of Θ̂s we have U1 (θ̂n ) ⊂ Θs , which implies θ ∈
/ U1 (θ̂n ). As a
result,
sup
(1)
θ∈Θ−Θs
Pθ (θ̂n ∈
/ Θ̂ns ) ≤
sup
(1)
θ∈Θ−Θs
(2)
Pθ (θ ∈
/ U1 (θ̂n )) ≤ sup Pθ (θ ∈
/ U1 (θ̂n )) ≤ rn(1)
(27)
θ∈Θ
and similarly supθ∈Θ−Θ(2) Pθ (θ̂n ∈
/ Θ̂s ) ≤ rn . Hence, we successfully localize θ via confidence sets based on θ̂n such
ns
that the true parameter θ is very likely to belong to the declared regime based on θ̂n .
A pictorial illustration of this idea is shown in Figure 2.
3) Question 3: Given a confidence set {U (x)}x∈Θ̂ of a satisfactory significance level rn , after observing θ̂n ∈ Θ̂ns we
can always set the approximation region to be U (θ̂n ). Note that U (θ̂n ) ⊂ Θns by definition, and in fact U (θ̂n ) can be
considerably smaller than Θns , which makes it a desirable regime to approximate over rather than Θns and is proved
to be necessary in [24]. The reason why U (θ̂n ) is sufficient is as follows: by definition of confidence sets we have
supθ∈Θ Pθ (θ ∈ U (θ̂n )) ≤ rn , hence with probability at least 1 − rn , the approximation region U (θ̂n ) based on θ̂n covers
θ, which allows us to operate as if θ is conditioned to be inside U (θ̂). Note that in order to obtain a good approximation
7
Θ̂ = Θ
U2 (θ̂2 )
θ̂2
θ̂1
U1 (θ̂1 )
Θ̂ns
Θ̂0
Θ̂s
(1)
Θ − Θs
I
II
III
(2)
Θ − Θns
IV
Fig. 2: Pictorial explanation of the “smooth” and “non-smooth” regimes based on θ and θ̂n , respectively. In the above figure,
(1)
(2)
(1)
(2)
we have Θs = II ∪ III ∪ IV, Θns = I ∪ II ∪ III, Θ̂s = III ∪ IV, and Θ̂ns = I ∪ II. In particular, Θs ∩ Θns = II ∪ III is
the intermediate regime, where both the plug-in and the approximation approach are performing well. For θ̂1 ∈ Θ̂ns , we have
(2)
(1)
U1 (θ̂1 ) ⊂ Θns ; for θ̂2 ∈ Θ̂s , we have U2 (θ̂2 ) ⊂ Θs .
performance, we need to find a confidence set {U (x)}x∈Θ̂ as small as possible, which depends on the statistical model.
4) Question 4: there has been a long history of correcting the bias of the MLE based on Taylor expansion. For example, in
entropy estimation, one of the earliest investigations on reducing the bias of MLE in entropy estimation is due to Miller
[34]. Interestingly, it was already observed in 1969 by Carlton [35] that Miller’s bias correction formula should only be
applied when p 1/n, which is automatically satisfied when p belongs to the “smooth” regime [ lnnn , 1] defined in [1].
As a result, in the “smooth” regime, Miller’s idea was used in [1]. In our generalization of the “smooth” regime, by
definition of Θs , G(T ) (θ̂n ) remains bounded with high probability for any order T > 0 and θ ∈ Θs . Hence, it shows that
Miller’s bias-correction approach based on Taylor expansion can also be used in general. However, Miller’s approach
fails when high-order bias correction is desired, or equivalently, when T is large. To see why it is the case, we take a
look at the procedure considered in [1]. For Binomial random variable X ∼ B(n, p), denote the empirical frequency by
p̂ = X
n . Then it follows from Taylor’s theorem that
1
1
1 00
p(1 − p) 00
f (p)Varp (p̂) + O( 2 ) =
f (p) + O( 2 )
(28)
2
n
2n
n
where f 00 (p) is the second-order derivative of f at p. Hence, the bias-corrected estimator in [1] was proposed as follows:
Ef (p̂) − f (p) =
f 00 (p̂)p̂(1 − p̂)
.
(29)
2n
However, the plug-in approach was still used for the bias-correction term in the previous estimator, which should be
further corrected based on Taylor expansion again in order to achieve higher-order bias correction. Continuing this
approach, the further correction still suffers from the same problem and additional corrections need to be done, and so
on and so forth. As a result, the previous bias-correction fails to be generalized to high-order corrections, and a successful
bias-correction approach should avoid employing the plug-in approach for bias-correction terms.
One way to avoid the plug-in approach is as follows: instead of doing Taylor expansion of G(θ̂n ) near θ, we employ
Taylor expansion of G(θ) near θ̂n as
f c (p̂) = f (p̂) −
G(θ) ≈
T
X
G(k) (θ̂n )
k=0
k!
(θ − θ̂n )k .
(30)
The advantage is that, now G(k) (θ̂n ) is by definition an unbiased estimator of Eθ [G(k) (θ̂n )]. However, the unknown
θ in the RHS still prevents us from using this estimator explicitly. Fortunately, this difficulty can be overcome by the
(1)
(2)
standard sample splitting approach: we split samples to obtain independent θ̂n and θ̂n , both of which follow the same
class of distribution (with possibly different parameters) as θ̂n . We remark that sample splitting can be employed for
divisible distributions, including Multinomial, Poisson and Gaussian models [36], and it will be discussed in detail at
8
the beginning of Section III for Poisson models. Now our bias-corrected estimator is
k
T
(1)
X
G(k) (θ̂n ) X k
Sj (θ̂n(2) )(−θ̂n(1) )k−j
Ĝs (θ̂n ) =
k!
j
j=0
(31)
k=0
(2)
where Sj (θ̂n ) is an unbiased estimator of θj (which usually exists). Now it is straightforward to show that
" T
#
X G(k) (θ̂n(1) )
G(T +1) (ξ) (1)
(1) k
(θ − θ̂n ) − G(θ) = O Eθ
(θ̂n − θ)T +1
E[Ĝs (θ̂n )] − G(θ) = Eθ
k!
(T + 1)!
(32)
k=0
i.e., the estimator in (31) achieves the bias-correction of any desired order. Although in many scenarios (those in [1],
[24], [27]–[30]) no bias correction or only order-one bias correction is required in the “smooth” regime, bias correction
of an arbitrary order turns out to be crucial in our recent work on the estimation of nonparametric functionals [33]. We
also conjecture that this approach is crucial for the construction of the minimax rate-optimal estimator for the Rényi
entropy in the large alphabet setting, which [31] did not address completely.
The answers to these questions shed light on the detailed implementation of the general recipe and give rise to the important
concept of localization via confidence sets, which leads us to propose a refined two-step approach. As before, denote by Θ̂ ⊃ Θ
the set containing all possible values of the estimator θ̂n , and by Θ̂0 ⊂ Θ̂ the set on which G(·) is non-analytic. Let {U (x)}x∈Θ̂
be a satisfactory confidence set.
1) Classify the Regime:
• For the true parameter θ, declare that θ is in the “non-smooth” regime if θ is “close” enough to Θ̂0 in terms of
localization via confidence sets (cf. (24)). Otherwise declare θ is in the “smooth” regime (cf. (23));
• Compute θ̂n , and declare that we are in the “non-smooth” regime if the confidence set of θ̂n falls into the “nonsmooth” regime of θ (cf. (26)). Otherwise declare we are in the “smooth” regime (cf. (25));
2) Estimate:
• If θ̂n falls in the “smooth” regime, use an estimator “similar” to G(θ̂n ) to estimate G(θ);
• If θ̂n falls in the “non-smooth” regime, replace the functional G(θ) in the “non-smooth” regime by an approximation
Gappr (θ) (another functional which well approximates G(θ) on U (θ̂n )) which can be estimated without bias, then
apply an unbiased estimator for the functional Gappr (θ).
In this paper, we follow the refined recipe for the construction of our optimal estimator in estimating several divergences
between discrete distributions, including the KL divergence, Hellinger distance and χ2 -divergence, where only the KL divergence will be discussed in detail. Moreover, in the estimation of KL divergence, we will encounter a new phenomenon, i.e.,
multivariate approximation in polytopes, which is a highly non-trivial topic in approximation theory, and will also propose a
general tool to analyze the risk of the bias-corrected plug-in approach with the help of localization via confidence sets.
The rest of this paper is organized as follows. We first analyze the performance of the modified plug-in estimator and prove
Theorem 2 in Section II. In Section III, we first follow the general recipe to explicitly construct our estimator for the KL
divergence step by step, and show that it essentially achieves the bound in Theorem 1. Then we adopt and adapt some tricks to
construct another estimator which is rate-optimal, adaptive and easier to implement. The minimax lower bound for estimating
the KL divergence is proved in Section IV. For the Hellinger distance and the χ2 -divergence, we sketch the construction of
the respective minimax rate-optimal estimators in Section V. Conclusions are drawn in Section VI, and complete proofs of
the remaining theorems and lemmas are provided in the appendices. The Matlab code of estimating KL divergence has been
released on http://www.stanford.edu/∼tsachy/index hjw.
II. P ERFORMANCE OF THE MODIFIED PLUG - IN APPROACH
In this section, we give the upper bound and the lower bound of the worst-case mean squared error via the modified plug-in
approach, i.e., we prove Theorem 2. Throughout our analysis, we utilize the Poisson sampling model, i.e., each component
Xi (resp. Yi ) in the histogram X (resp. Y) has distribution Poi(mpi ) (resp. Poi(nqi )), and all coordinates of X (resp. Y) are
independent. In other words, instead of drawing fixed sample sizes m and n, there are i.i.d. samples from distributions P, Q
of sizes M ∼ Poi(m) and N ∼ Poi(n), respectively. Consequently, the observed number of occurrences of each symbol are
independent [37, Theorem 5.6]. We note that the Poisson sampling model is essentially the same as the Multinomial model,
and their minimax risks are related via Lemma 22 in Appendix A.
A. Proof of upper bounds
Recall that the empirical distribution Qn has been modified to
1
1
0
∨ q̂1 , · · · , ∨ q̂S
Qn =
n
n
(33)
9
the modified plug-in estimator D(Pm kQ0n ) is not the exact plug-in approach. However, it can be observed that this quantity
S
X
1
0
∨ q̂i
(34)
D(Pm kQn ) =
p̂i ln p̂i − p̂i ln
n
i=1
is close to the following natural plug-in estimator
D1 (Pm kQn ) =
where
gn (q) ,
S
X
i=1
[p̂i ln p̂i − p̂i gn (q̂i )] ,
(
−(1 + ln n) + nq
ln q
if 0 ≤ q <
if
1
n
1
n,
≤ q ≤ 1.
(35)
(36)
In view of this fact, we can apply the general approximation-based method in [38] to analyze the performance of the plug-in
approach.
By construction it is obvious that gn (q) is continuously differentiable on [0, 1], which coincides with g(q) = ln q on [ n1 , 1].
Moreover, since q̂i is a multiple of n1 , gn (q̂i ) only differs from ln( n1 ∨ q̂i ) at q̂i = 0 by | − (1 + ln n) − ln(1/n)| = 1. Hence, we
may consider the performance of the plug-in estimator p̂i (ln p̂i − gn (q̂i )) in estimating p(ln p − gn (q)), which is summarized
in the following lemma.
Lemma 1. Let np̂ ∼ Poi(mp) and nq̂ ∼ Poi(nq) be independent, and p ≤ u(S)q. Then we have
30u(S) 5 ln 2
+
,
n
m
2
4q 4u(S)
700u(S)
1
51
2
Var(p̂(ln p̂ − gn (q̂)) ≤ 2 +
p + p(ln u(S)) + 2 +
+
p+
.
m
m
e
en
n
m
|E[p̂(ln p̂ − gn (q̂))] − p(ln p − gn (q))| ≤
(37)
(38)
In particular,
1
u(S)
+ ,
n
m
pu(S) u(S) p(1 + ln u(S))2
q
1
+
+
+ .
Var(p̂(ln p̂ − gn (q̂))) . 2 +
m
n
mn
m
m
Hence, by Lemma 1, we conclude that
|E[p̂(ln p̂ − gn (q̂))] − p(ln p − gn (q))| .
|ED1 (Pm kQn ) − D1 (P kQ)| ≤
.
S
X
i=1
S
X
i=1
=
(39)
(40)
|E[pˆi (ln p̂i − gn (qˆi ))] − pi (ln pi − gn (qi ))|
(41)
1
u(S)
+
n
m
(42)
Su(S)
S
+
n
m
(43)
and
Var(D1 (Pm kQn )) =
.
S
X
i=1
S
X
i=1
Var(p̂i gn (q̂i ))
(44)
1
pi u(S) u(S) pi (1 + ln u(S))2
qi
+
+
+
+
2
m
n
mn
m
m
(45)
S
u(S) Su(S) (1 + ln u(S))2
+
+
+
m2
n
m
mn
S
u(S)
(Su(S))2
1
(ln u(S))2
. 2+
+
+
+
m
n
n2
m2
m
2
S
u(S) (Su(S))
(ln u(S))2
. 2+
+
+
.
m
n
n2
m
.
(46)
(47)
(48)
10
Combining these two inequalities yields, for any (P, Q) ∈ US,u(S) ,
2
2
E (D1 (Pm kQn ) − D1 (P kQ)) = |ED1 (Pm kQn ) − D1 (P kQ)| + Var(D1 (Pm kQn ))
2
u(S) (ln u(S))2
S
Su(S)
+
+
+
.
.
n
m
n
m
(49)
(50)
To prove Theorem 2, it remains to compute the difference between D and D1 . By the definition of gn (·), we have
E|D1 (Pm kQn ) −
D(Pm kQ0n )|2
S
X
=E
≤S
=S
2
p̂i 1(q̂i = 0)
i=1
S
X
(51)
E|p̂i 1(q̂i = 0)|2
i=1
S
X
p2i +
i=1
S
X
(52)
pi −nqi
e
m
(53)
u(S) −nqi
qi e
m
i=1
!
2
S
X
2
u(S)
≤S
(u(S))2
+
en
enm
i=1
2
Su(S)
S 2 u(S)
.
+
n
mn
2
S
Su(S)
+
.
n
m
≤S
where we have used the fact that
k −nx
sup x e
x∈[0,1]
=
(u(S))2 qi2 e−nqi +
k
en
k
(54)
(55)
(56)
(57)
(58)
.
Moreover, for any (P, Q) ∈ US,u(S) ,
|D1 (P kQ) − D(P kQ)| ≤
≤
S
X
i=1
S
X
i=1
pi | − (ln n + 1) + nqi − ln qi |1(qi <
pi (1 − ln(nqi ))1(qi <
≤ u(S) ·
S
X
i=1
1
)
n
1
)
n
qi (1 − ln(nqi ))1(qi <
(59)
(60)
1
)
n
(61)
Su(S)
(62)
n
where we have used that supq∈[0,1/n] q(1 − ln(nq)) = 1/n. Hence, by the triangle inequality, for any (P, Q) ∈ US,u(S) , we
have
.
2
E (D(Pm kQ0n ) − D(P kQ))
2
2
≤ 3 E (D1 (Pm kQn ) − D1 (P kQ)) + E (D1 (Pm kQn ) − D(Pm kQ0n )) + |D1 (P kQ) − D(P kQ)|2
2
Su(S)
S
u(S) (ln u(S))2
.
+
+
+
n
m
n
m
which completes the proof of the upper bound in Theorem 2.
(63)
(64)
11
B. Proof of lower bounds
By the bias-variance decomposition of the mean squared error, to prove that the squared term in Theorem 2 serves as a
lower bound, it suffices to find some (P, Q) ∈ US,u(S) such that
S
Su(S)
+ .
(65)
n
m
Note that here we prove this inequality based on the Multinomial model, and then obtain the result for the Poisson sampling
model via Lemma 22. The construction of (P, Q) is as follows: P = ( S1 , · · · , S1 ) is the uniform distribution, and Q =
1
1
S−1
( Su(S)
, · · · , Su(S)
, 1 − Su(S)
) is near-uniform. We first recall from [38] that, if m ≥ 15S, we have
|E(P,Q) D(Pm kQ0n ) − D(P kQ)| &
S
X
i=1
[E[p̂i ln p̂i ] − pi ln pi ] ≥
1
S2
S−1
−
.
−
2m
20m2
12m2
(66)
Next we give a lower bound for the term E(− ln(q̂i ∨ n1 )) − (− ln qi ) for qi ≥ n4 . We shall use the following lemma for
the approximation error of the Bernstein polynomial, which corresponds to the bias in the Multinomial model. Define the
Bernstein operator Bn as follows:
n
X
n i
i
n−i
,
f ∈ C[0, 1].
(67)
Bn [f ](x) =
x (1 − x)
·f
n
i
i=0
Lemma 2. [39] Let k ≥ 4 be an even integer. Suppose that the k-th derivative of f satisfies that f (k) ≤ 0 on (0, 1), and
Qk−1 is the Taylor polynomial of order k − 1 of f at some point x0 . Then for x ∈ [0, 1],
f (x) − Bn [f ](x) ≥ Qk−1 (x) − Bn [Qk−1 ](x).
(68)
Since our modification of ln(·) is not even differentiable, Lemma 2 cannot be applied directly. However, we can consider
the following function instead:
(
2
3
4
− ln n + n(x − n1 ) − n2 (x − n1 )2 + n3 (x − n1 )3 − n4 (x − n1 )4 if 0 ≤ x < n1 ,
(69)
hn (x) =
ln x
if n1 ≤ x ≤ 1.
By construction it is obvious that hn (x) ∈ C 4 [0, 1] which coincides with ln x on [ n1 , 1]. Moreover, hn (q̂) only differs from
(4)
ln( n1 ∨ q̂) at zero by |hn (0) + ln n| = 25
12 . Since hn (x) ≤ 0, Lemma 2 can be applied here to yield the following lemma.
Lemma 3. For
4
n
≤ x ≤ 1, we have
hn (x) − Bn [hn ](x) ≥
(1 − x)((n + 4)x − 2)
> 0.
n2 x2
Since our assumption n ≥ 4Su(S) ensures that for our choice of Q, qi ≥
of hn (·), we have
S
X
i=1
S−1
X
4
n
for any i. Hence, by Lemma 3 and the concavity
1
[ln qi − E[hn (q̂i )]]
S
i=1
S − 1 (Su(S))2
1
n+4
≥
·
1−
−2
S
n2
Su(S)
Su(S)
S − 1 (Su(S))2
1
n
≥
·
1−
·
S
n2
Su(S)
2Su(S)
(S − 1)u(S)
1
=
1−
.
2n
Su(S)
[pi ln qi − E[p̂i hn (q̂i )]] ≥
(70)
(71)
(72)
(73)
(74)
12
Now note that
S
X
i=1
S
|E[p̂i hn (q̂i )] − E[p̂i ln(
25 X
1
∨ q̂i )]| ≤
pi · P(q̂i = 0)
n
12 i=1
(75)
S
25 X
=
pi (1 − qi )n
12 i=1
(76)
≤
(77)
≤
where we have used the fact that
sup x(1 − x)n =
4
x∈[ n
,1]
4
n
1−
S
X
25
u(S)
qi (1 − qi )n
12
i=1
25Su(S)
3ne4
4
n
n
≤
4
.
ne4
A combination of these two inequalities yields
X
S
S
S
X
X
1
1
E[−p̂i ln( ∨ q̂i )] − (−pi ln qi ) ≥
[pi ln qi − E[p̂i hn (q̂i )]] −
|E[p̂i hn (q̂i )] − E[p̂i ln( ∨ q̂i )]|
n
n
i=1
i=1
i=1
1
25Su(S)
(S − 1)u(S)
1−
−
.
≥
2n
Su(S)
3ne4
Hence, when m ≥ 15S and n ≥ 4Su(S), combining (66) and (81) gives
|E(P,Q) D(Pm kQ0n ) − D(P kQ)| ≥ E(P,Q) D(Pm kQ0n ) − D(P kQ)
S
X
1
=
[E[p̂i ln p̂i ] − pi ln pi ] +
E[−p̂i ln( ∨ q̂i )] − (−pi ln qi )
n
i=1
i=1
S−1
S2
1
(S − 1)u(S)
1
25Su(S)
≥
−
−
+
1
−
−
2m
20m2
12m2
2n
Su(S)
3ne4
S
X
which gives (65), as desired.
For the remaining terms, we remark that
2
(ln u(S))2
u(S)
E(P,Q) D̂ − D(P kQ) &
+
m
n
(P,Q)∈US,u(S)
sup
(78)
(79)
(80)
(81)
(82)
(83)
(84)
(85)
holds for any estimator D̂ (and thus for the modified plug-in estimator D(Pm kQ0n )), and we postpone the proof to Section
IV. Now the proof of Theorem 2 is complete.
III. C ONSTRUCTION OF THE O PTIMAL E STIMATOR
We stay with the Poisson sampling model in this section. For simplicity of analysis, we conduct the classical “splitting”
operation [40] on the Poisson random vector X, and obtain three independent identically distributed random vectors Xj =
[X1,j , X2,j , . . . , XS,j ]T , j ∈ {1, 2, 3}, such that each component Xi,j in Xj has distribution Poi(mpi /3), and all coordinates in
Xi
i
Xj are independent. For each coordinate i, the splitting process generates a random sequence {Tik }X
k=1 such that {Tik }k=1 |X ∼
PXi
i
Multinomial(Xi ; (1/3, 1/3, 1/3)), and assign Xi,j = k=1 1(Tik = j) for j ∈ {1, 2, 3}. All the random variables {{Tik }X
k=1 :
1 ≤ i ≤ S} are conditionally independent given our observation X. The splitting operation is similarly conducted for the
Poisson random vector Y.
For simplicity, we re-define m/3 as m and n/3 as n, and denote
Xi,j
Yi,j
, q̂i,j =
,
i ∈ {1, 2, · · · , S}, j ∈ {1, 2, 3}.
(86)
m
n
We remark that the “splitting” operation is not necessary in implementation. We also note that for independent random variables
(X, Y ) such that nX ∼ Poi(mp), nY ∼ Poi(nq),
p̂i,j =
E
k−1
Y
r=0
r Y
s
Y −
= pk q l ,
m s=0
n
l−1
X−
for any k, l ∈ N. For a proof of this fact we refer to Withers [41, Example 2.8].
(87)
13
A. Estimator construction
Now we apply our general recipe to construct the estimator. Note that
D(P kQ) =
S
X
pi ln
i=1
S
S
S
X
X
X
pi
=
pi ln pi −
pi ln qi = −H(P ) −
pi ln qi
qi
i=1
i=1
i=1
(88)
PS
where H(P ) = i=1 −pi ln pi is the entropy function. Hence, the optimal estimator Ĥ for entropy [1], [28]–[30] can be used
PS
here and it remains to estimate the cross entropy i=1 pi ln qi , i.e., our target is the bivariate function f (p, q) = p ln q.
We first classify the regime. For the bivariate function f (p, q) = p ln q, the entire parameter set is Θ = {(p, q) ∈ [0, 1]2 : p ≤
u(S)q}, and the function is analytic everywhere except for the point Θ0 = {(0, 0)}. For all possible values of the estimator
(p̂, q̂), we have Θ̂ = [0, 1]2 , and non-analytic points are Θ̂0 = [0, 1] × {0}. For the confidence set of this two-dimensional
Poisson model (mp̂, nq̂) ∼ Poi(mp) × Poi(nq), we can set rn n−A for some universal constant A > 0 and use
m
n
n
m
[0, c ln
[0, c ln
if x ≤ c ln
if y ≤ c ln
m ]
m ,
n ]
n ,
q
q
q
q
×
U (x, y) = Θ ∩
[x − cx ln m , x + cx ln m ] if c ln m < x ≤ 1. [y − cy ln n , y + cy ln n ] if c ln n < y ≤ 1.
m
m
m
n
n
n
(89)
for some constant c > 0. Hence, by choosing c = c1 /2 and c = 2c1 respectively in (23) and (24) for some universal constant
c1 > 0 to be specified later, we get the “smooth” and “non-smooth” regimes for (p, q) as (for brevity we omit the superscripts
in (23) and (24))
c1 ln n
c1 ln n
, 1] = (p, q) ∈ [0, 1]2 :
≤ q ≤ 1, p ≤ u(S)q
(90)
Θs = Θ ∩ [0, 1] × [
2n
2n
2c1 ln n
2c1 ln n
] = (p, q) ∈ [0, 1]2 : 0 ≤ q ≤
, p ≤ u(S)q .
(91)
Θns = Θ ∩ [0, 1] × [0,
n
n
Further, by (25) and (26), the ultimate “smooth” and “non-smooth” regimes are given by
Θ̂s = [0, 1] × [
c1 ln n
, 1],
n
Θ̂ns = [0, 1] × [0,
c1 ln n
]
n
(92)
i.e., we are in the “non-smooth” regime if q̂ ≤ c1 nln n , and are in the “smooth” regime otherwise.
Next we construct the estimator in each regime. First, if we are in the “smooth” regime, our bias-corrected estimator (31)
of order T = 3 becomes
k
3
X
g (k) (q̂1 ) X k
Sj (q̂2 )(−q̂1 )k−j
(93)
T (3) (q̂1 , q̂2 ) =
j
k!
j=0
k=0
= ln q̂1 +
q̂2 − q̂1
(q̂2 − q̂1 )2
3q̂2
(q̂2 − q̂1 )3
q̂ 2
2q̂2
−
+
+
− 23 + 2 3
2
2
3
q̂1
2q̂1
2nq̂1
3q̂1
nq̂1
n q̂1
for estimating g(q) = ln q. Note that in the Poisson model nq̂2 ∼ Poi(nq), by (87) we have Sj (q̂2 ) =
for estimating f (p, q) = p ln q in the “smooth” regime, our estimator becomes
Ts (p̂1 , q̂1 ; p̂2 , q̂2 ) = p̂1 · T (3) (q̂1 , q̂2 ) = p̂1 (ln q̂1 +
(94)
Qj−1
k=0 (q̂2
− nk ). Then
q̂2 − q̂1
(q̂2 − q̂1 )2
3q̂2
(q̂2 − q̂1 )3
q̂22
2q̂2
+
+
−
+ 2 3 ).
−
q̂1
2q̂12
2nq̂12
3q̂13
nq̂13
n q̂1
(95)
To ensure that Ts is well-defined, it suffices to define an additional value of Ts (e.g., zero) when q̂1 = 0. Note that sample
splitting here is only used for the simplicity of analysis, and it is indeed not necessary in implementation. We can also replace
p̂2
p̂1 with p̂1 +
here to further reduce the variance.
2
Now consider the case where we are in the “non-smooth” regime, i.e., q̂ ≤ c1 nln n . By our general recipe, we should
approximate f (p, q) = p ln q in the approximation region given by the confidence set
ln m
[0, 2c1mln m ]
if p̂ ≤ c1 m
,
4c
ln
n
1
q
q
U (p̂, q̂) = Θ ∩
× [0,
] .
(96)
[p̂ − 1 c1 p̂ ln m , p̂ + 1 c1 p̂ ln m ] if c1 ln m < p̂ ≤ 1.
n
2
m
2
m
m
As a result, we further distinguish the “non-smooth” regime into two sub-regimes depending on p̂ ≤
ln m
localization via confidence sets is essentially equivalent to p ≤ c1 m
or not.
c1 ln m
m
or not, which by
14
If p̂ >
c1 ln m
m ,
the approximation region is given by
(
)
r
r
1
1
4c
ln
n
c
p̂
ln
m
c
p̂
ln
m
1
1
1
(p, q) ∈ [0, 1]2 : p̂ −
≤ p ≤ p̂ +
,0 ≤ q ≤
, p ≤ u(S)q
2
m
2
m
n
"
# "
!
#
r
r
r
1 c1 p̂ ln m
1 c1 p̂ ln m
1
1 c1 p̂ ln m
4c1 ln n
⊂ p̂ −
, p̂ +
×
p̂ −
,
2
m
2
m
u(S)
2
m
n
(97)
where the latter is a rectangle. Since q cannot hit zero in this approximation regime, and f (p, q) = p ln q is a product of p
and ln q, we can consider the best polynomial approximation of ln q in this regime. As a result, in this regime, we use the
approximation-based estimator
K
k−1
Y
c1 ln m X
j
Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) = 1(p̂2 ≥
)·
(98)
gK,k (p̂2 ) · p̂1
q̂1 −
3m
n
j=0
k=0
where
K
X
gK,k (p̂)z k = arg min 1
P ∈polyK
k=0
1
(p̂− 12
z∈[ u(S)
max
q
c1 p̂ ln m
4c ln n
), 1n
]
m
| ln z − P (z)|
(99)
is the best 1D order-K polynomial approximation of g(q) = ln q, where K = c2 ln n with universal constant c2 > 0 to be
ln m
ensures that the 1D approximation interval does not contain zero and is thus valid. We
specified later. Note that p̂2 ≥ c13m
call this regime as “non-smooth” regime I.
ln m
If p̂ ≤ c1 m
, the approximation region is given by
2c1 ln m
4c1 ln n
R = (p, q) ∈ [0, 1]2 : 0 ≤ p ≤
,0 ≤ q ≤
, p ≤ u(S)q .
(100)
m
n
Since q may be zero in R, the usual best 1D polynomial approximation of g(q) = ln q over this region does not work, and the
best 2D polynomial approximation of f (p, q) = p ln q should be employed here. Hence, in this regime the approximation-based
estimator is
l−1
k−1
X
Y
i Y
j
Tns,II (p̂1 , q̂1 ) =
hK,k,l
p̂1 −
q̂1 −
(101)
m j=0
n
i=0
k,l≥0,0<k+l≤K
where
X
hK,k,l wk z l = arg min 2
max |w ln z − P (w, z)|
P ∈polyK (w,z)∈R
k,l≥0,k+l≤K
(102)
is the best 2D order-K polynomial approximation of f (p, q) = p ln q in R, where K = c2 ln n. Note that the condition k +l > 0
in the summation ensures that the estimator is zero for unseen symbols.
We call this regime as “non-smooth” regime II.
PS
In summary, we have the following estimator construction for i=1 pi ln qi .
Estimator Construction 1. Conduct three-fold sample splitting to obtain i.i.d. samples (p̂i,1 , p̂i,2 , p̂i,3 ) and (q̂i,1 , q̂i,2 , q̂i,3 ).
PS
The estimator D̂0 for the cross entropy i=1 pi ln qi is constructed as follows:
S
X
c1 ln m
c1 ln n
c1 ln m
0
) + T̃ns,II (p̂i,1 , q̂i,1 )1(p̂i,3 ≤
) 1(q̂i,3 ≤
)
D̂ =
T̃ns,I (p̂i,1 , q̂i,1 ; p̂i,2 , q̂i,2 )1(p̂i,3 >
m
m
n
i=1
c1 ln n
+T̃s (p̂i,1 , q̂i,1 ; p̂i,2 , q̂i,2 )1(q̂i,3 >
) ,
(103)
n
where
T̃ns,I (x, y; x0 , y 0 ) , (Tns,I (x, y; x0 , y 0 ) ∧ 1) ∨ (−1)
T̃ns,II (x, y) , (Tns,II (x, y; x , y ) ∧ 1) ∨ (−1)
0
0
T̃s (x, y; x , y ) , Ts (x, y; x , y ) · 1(y 6= 0)
0
0
0
0
(104)
(105)
(106)
and Tns,I , Tns,II , Ts are given by (98), (101) and (95), respectively. A pictorial illustration of three regimes and our estimator
is displayed in Figure 3.
For the estimation of entropy, we essentially follow the estimator in [1]. Specifically, let LH (x) be the lower part function
15
p
1
(q̂1 , p̂1 )
p̂2 +
1
2
q
c1 ln m
m
· p̂2
“Smooth” regime:
(q̂2 , p̂2 )
plug-in approach with
“Non-smooth” regime I:
order-three bias correction
p̂2 × unbiased estimate of best
1D polynomial approximation of ln q
p̂2
p̂2 −
1
2
q
c1 ln m
m
· p̂2
c1 ln m
m
0
(q̂3 , p̂3 ) “Non-smooth” regime II:
unbiased estimate of best 2D
polynomial approximation of p ln q
q
c1 ln m
1
1
· p̂2
u(S) p̂2 − 2
m
c1 ln n
n
1
u(S)
q
PS
Fig. 3: Pictorial explanation of three regimes and our estimator for i=1 pi ln qi . The point (q̂1 , p̂1 ) falls in the “smooth”
regime, (q̂2 , p̂2 ) falls in the “non-smooth” regime I, and (q̂3 , p̂3 ) falls in the “non-smooth” regime II.
defined in [1], and UH (x) be defined as
1
(107)
2m
which gets rid of the interpolation function compared with the upper part function defined in [1]. Then the entropy estimator
is defined as
S
X
c1 ln m
c1 ln m
Ĥ =
UH (p̂i,1 )1(p̂i,3 >
) + LH (p̂i,1 )1(p̂i,3 ≤
) .
(108)
m
m
i=1
UH (x) , −x ln x +
Finally, the overall estimator D̂ for D(P kQ) is defined as
D̂ = −D̂0 − Ĥ
(109)
where c1 , c2 > 0 are suitably chosen universal constants.
B. Estimator analysis
In this subsection we will prove that the estimator constructed above achieves the minimax rate in Theorem 1. Recall that
the mean squared error of any estimator D̂ in estimating D(P kQ) can be decomposed into the squared bias and the variance
as follows:
E(P,Q) (D̂ − D(P kQ))2 = |Bias(D̂)|2 + Var(D̂)
(110)
where the bias and the variance are defined as
Bias(D̂) , E(P,Q) D̂ − D(P kQ)
(111)
Var(D̂) , E(P,Q) (D̂ − E(P,Q) D̂)
2
(112)
respectively. Hence, it suffices to analyze the bias and the variance in these three regimes.
ln n
.
1) “Smooth” regime: First we consider the “smooth” regime where the true parameter (p, q) belongs to Θs , i.e., q > c12n
In this regime, the estimator we employ is the plug-in approach whose bias is corrected by Taylor expansion, e.g., (94). Recall
16
that the bias of our bias-corrected plug-in estimator can be expressed as
|Eq T (r) (q̂1 , q̂2 ) − g(q)| = Eq
r
X
g (k) (q̂)
k!
k=0
g (r+1) (ξ)
(q − q̂)k+1
(r + 1)!
(q − q̂)k − g(q) ≤ Eq
(113)
where nq̂ ∼ Poi(nq), and ξ ∈ [q ∧ q̂, q ∨ q̂] depends on q̂. For smooth g(·) with |g (r+1) (ξ)| bounded everywhere, it suffices to
consider maxξ |g (r+1) (ξ)|, the upper bound of the (r + 1)-th derivative. However, the reason why the plug-in approach and
its bias-corrected version are both strictly suboptimal for the estimation of non-smooth functionals (e.g., the empirical entropy
[38]) is that the functional g(·) may have non-analytic points where the high-order derivatives may be unbounded. Hence, a
direct application of the Taylor expansion does not work for a general non-smooth g(·). However, now we are at the “smooth”
regime (i.e., (p, q) ∈ Θs ), by our general recipe we know that with high probability q̂ will not fall into the non-analytic region
Θ̂0 of g(·), thus g(·) is sufficiently smooth on the segment connecting q̂ and q, and maxq∧q̂≤ξ≤q∨q̂ |g (r+1) (ξ)| can be well
controlled. In other words, the bias can be upper bounded with the help of localization via confidence sets.
Motivated by the previous insights, we begin with the following general lemma.
(1)
Lemma 4. Assume that the estimator in (31) is well-defined, Eθ θ̂n = θ, and let {V (θ)}θ∈Θ be a reverse confidence set of
(1) (2)
level 1 − δ with δ ∈ (0, 1). Further suppose that for any k = 0, · · · , r, the function Hk (θ̂n , θ̂n ) coincides with
k
(1)
G(k) (θ̂n ) X k
Sj (θ̂n(2) )(−θ̂n(1) )k−j
(114)
Gk (θ̂n(1) , θ̂n(2) ) ,
k!
j
j=0
(2)
(1)
(1)
whenever θ̂n ∈ V (θ), and define H(θ̂n , θ̂n ) ,
(1)
Pr
(1)
k=0
(2)
Hk (θ̂n , θ̂n ). Then
Eθ |θ̂n − θ|r+1
|Eθ H(θ̂n(1) , θ̂n(2) ) − G(θ)| ≤
· sup |G(r+1) (θ̂)| + δ ·
(r + 1)!
θ̂∈V (θ)
and for any k = 0, 1, · · · , r,
Varθ (Hk (θ̂n(1) , θ̂n(2) )) ≤ Ar δ · sup |Hk (θ1 , θ2 )| +
+
k−1
X
j=0
θ1 ,θ2 ∈Θ̂
k
X
j=0
|G(θ)| + sup |H(θ1 , θ2 )|
θ1 ,θ2 ∈Θ̂
Eθ (θ̂n(1) − θ)2 · sup |G0k,j (θ̂)|2 + δ|Gk,j (θ)|2
θ̂∈V (θ)
!2
|Gk,j (θ)| + Eθ (θ̂n(1) − θ)2 · sup |G00k,j (θ̂)| + δ(|θ| + sup |θ̂|) · |G0k,j (θ)|
|Eθ Hk (θ̂n(1) , θ̂n(2) )| ≤
θ̂∈V (θ)
Eθ (θ̂n(1) − θ)2 · sup
Eθ |θ̂n(1) −θ|k
k!
!
θ̂∈Θ̂
!
2
· Eθ Sk−j
(θ̂n(2) )
· Varθ (Sk−j (θ̂n(2) ))
00
0
|G
(
θ̂)|
+
δ
·
sup
|H
(θ
,
θ
)|
+
|G
(θ)|(|θ|
+
sup
|
θ̂|)
1
1
2
θ1 ,θ2 ∈Θ̂
θ̂∈V (θ)
θ̂∈Θ̂
· supθ̂∈V (θ) |G(k) (θ̂)| + δ · supθ1 ,θ2 ∈Θ̂ |Hk (θ1 , θ2 )|
(115)
(116)
if k = 1,
if k ≥ 2.
(117)
where Gk,j (x) , xj G(k) (x), and Ar > 0 is a universal constant which depends on r only.
It can be seen from the previous lemma that the upper bounds of both the bias and the variance are very easy to compute,
for we only need to calculate the finite-order derivatives of G(·) and the moments of some usually well-behaved estimators
(2)
(i.e., Sj (θ̂n )). Moreover, with the help of localization via confidence sets, all bounds only depend on the local behavior of
(1) (2)
(1) (2)
(1)
function G(·) (so we only require that Hk (θ̂n , θ̂n ) coincide with Gk (θ̂n , θ̂n ) when θ̂n ∈ V (θ)) plus a negligible term
(1)
corresponding to the event that θ̂n ∈
/ V (θ). Note that this was the major difficult part in the analysis of the bias-corrected plugin estimator in [1], whose proof is quite lengthy (over four pages in the proof of Lemma 2) and requires the explicit construction
of the interpolation function in estimator construction. Note that in Lemma 4 we implicitly use the following “interpolation”
idea: we essentially condition on the event that θ̂n ∈ V (θ), which is similar as we “interpolate” the function G(θ̂n ) using
the rectangle window 1(θ̂n ∈ V (θ)) so as to prevent maxξ |G(r+1) (ξ)| becoming infinity. Note that this interpolation is done
only in the analysis but not in the construction of our estimator, and thus we remark that the explicit interpolation in [1] is
indeed unnecessary given the implicit interpolation by localization via confidence sets. Following this idea, although UH (·) in
(107) does not follow the same idea of our bias correction (31), the result in [1] can still be easily recovered without explicit
interpolation:
17
Lemma 5. Let p ≥
c1 ln m
2m ,
and mp̂ ∼ Poi(mp). If c1 ln m ≥ 4, the following inequalities hold:
6
+ 4m−c1 /24+2
c1 m ln m
p(2 − ln p)2
+ 4m−c1 /24
Var(UH (p̂)) ≤ A0
m
(118)
|EUH (p̂) + p ln p| ≤
(119)
where A0 is the universal constant appearing in Lemma 4.
Now we apply Lemma 4 to analyze the estimation performance of the bias-corrected plug-in estimator T (3) (q̂1 , q̂2 ) in (94).
ln n
, a natural reverse confidence set is given by
In the Poisson model nq̂ ∼ Poi(nq) with q ≥ c12n
#
"
r
r
c1 ln n
1 c1 q ln n
1 c1 q ln n
,
q≥
V (q) = q −
,q +
.
(120)
2
2n
2
2n
2n
By Lemma 28 we know that this reverse confidence set has level 1 − δ with
s
!2
1 1 c1 ln n
1
δ ≤ sup Pq (q̂ ∈
/ V (q)) ≤ exp −
· nq + exp −
3
2
2nq
2
c1 ln n
q≥
2n
1
2
s
c1 ln n
2nq
!2
· nq ≤ 2n−c1 /24
(121)
which can decay faster than any polynomial rate provided that c1 is large enough. In this special case we can simplify the
expressions in Lemma 4.
ln n
, (nq̂1 , nq̂2 ) ∼ Poi(nq) × Poi(nq) and c1 ln n ≥ 2, and g(·) be an (r + 1) times differentiable function
Lemma 6. Let q ≥ c12n
c1 ln n
on [ 4n , 1]. Suppose that for any k = 0, · · · , r, the function hk (q̂1 , q̂2 ) coincides with
k
g (k) (q̂1 ) X k
hk (q̂1 , q̂2 ) ,
Sj (q̂2 )(−q̂1 )k−j
(122)
j
k!
j=0
whenever q̂1 ∈ V (q), and define h(q̂1 , q̂2 ) ,
only such that
k=0
q r+1
2
|Eh(q̂1 , q̂2 ) − g(q)| ≤ Br
and for any k ≤ r,
Pr
n
·
hk (q̂1 , q̂2 ). Then there exists a universal constant Br > 0 depending on r
sup
ξ∈[q/2,2q]
Var(hk (q̂1 , q̂2 )) ≤ Br n−c1 /24 · sup |hk (q1 , q2 )| +
+
q1 ,q2
k−1
X
j=0
|Ehk (q̂1 , q̂2 )| ≤ Br
where gk,j (q) , q j g (k) (q).
q 2(k−j)−1
n
q k∨2
2
n
·
|g
(r+1)
k
X
j=0
(ξ)| + n
−c1 /24
!
· |g(q)| + sup |h(q1 , q2 )|
q
2(k−j)
q
sup |g 0 (ξ)|2 + n−c1 /24 |gk,j (q)|2
n ξ∈[q/2,2q] k,j
!
q
0
|gk,j (q)| +
sup |g 00 (ξ)| + n−c1 /24 |gk,j
(q)|
n ξ∈[q/2,2q] k,j
sup
ξ∈[q/2,2q]
(123)
q1 ,q2
2
,
!
|g (k∨2) (ξ)| + n−c1 /24 · (|g 0 (q)| + sup |hk (q1 , q2 )|) ,
q1 ,q2
!
k≥0
(124)
k≥1
(125)
Note that due to the nice property of the Poisson model, the previous lemma greatly simplifies the expression involving
Sj (q̂), the unbiased estimate of the monomial functions. Moreover, if g (s) (q) becomes a power function of q for some s,
all summands in Lemma 6 with k ≥ s will have the same order of magnitude and can thus be merged into one term. An
interesting observation is that, if we change r to r + 1 in this case, the order of the bias of our bias-corrected estimator is
ln n
1
multiplied by √1nq , which is at most of the order √ln
since q ≥ c12n
. Hence, by continuing this bias correction approach,
n
we can improve the bias of the plug-in approach by any desired logarithmic multiplicative factor.
ln n
Next we apply Lemma 6 to the bias-corrected estimator T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) in the smooth regime, where q ≥ c12n
and
0
g(q) = ln q satisfied the previous property (i.e., g (q) = 1/q is a power of q).
ln n
Lemma 7. Let p ∈ [0, 1], q ≥ c12n
, p ≤ u(S)q, and Poisson random variables m(p̂1 , p̂2 ) ∼ Poi(mp) × Poi(mp), n(q̂1 , q̂2 ) ∼
Poi(nq) × Poi(nq) be independent. Moreover, let c1 ln m ≥ 4 and c1 ln n ≥ 4.
18
If p ≤
2c1 ln m
m
is small, we have
192B3 u(S)
C
(2c1 ln m)4
+ 6B3 pn−c1 /24+3 +
+ 2−8c2 ln 2
c1 n ln n
m ln m
m
125
2(2c1 ln m)4
p
2
−c1 /24+3
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )) ≤
+ 1600B3 p +
+n
m2−8c2 ln 2
m
nq
2
p
B3 48 + 6n−c1 /24+3 − ln q
+
m
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )] + p ln(p/q)| ≤
where B3 , C are universal constants given in Lemma 6 and [1, Lemma 4], respectively. If p ≥
c1 ln m
2m
(126)
(127)
is large, we have
6
192B3 u(S)
+ 6B3 pn−c1 /24+3 +
+ 4m−c1 /24+2
(128)
c1 n ln n
c1 m ln m
8
204
4q 4u(S)
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )) ≤ 2 +
p + p(ln u(S))2 + 2 +
m
m
e
en
2800u(S)
1
p 125
2
−c1 /24+3
+
p+
+ 900B3 p +
+n
n
m
m
nq
2
9B32 p
48 + 3n−c1 /24+3 .
(129)
+ 4(1 + ln n)2 · n−c1 /2 +
m
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )] + p ln(p/q)| ≤
In particular, if c1 > 96 and 8c2 ln 2 < ∈ (0, 1), the previous bounds imply that for p ≤
2c1 ln m
,
m
u(S)
1
+
n ln n m ln m
p(1 + ln u(S))2
u(S) pu(S)
1
+
+
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )) . 2− +
m
m
mn
n
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )] + p ln(p/q)| .
and for p ≥
(130)
(131)
c1 ln m
2m ,
1
u(S)
+
(132)
n ln n m ln m
1
p(1 + ln u(S))2
q
u(S) pu(S)
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )) . 2 +
+
+
+
.
(133)
m
m
m
mn
n
Note that the variance bound given by Lemma 7 is a non-asymptotic result whose order coincides with that given by classical
asymptotics, where the asymptotic variance is the leading term and can be obtained easily via the delta method [23]. Now we
use Lemma 7 to analyze the property of the overall estimator
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )] + p ln(p/q)| .
T s (p̂, q̂) = T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )1(p̂3 ≤
in the “smooth” regime q ≥
c1 ln n
2n ,
c1 ln m
c1 ln m
) + UH (p̂1 )1(p̂3 >
)
m
m
(134)
where p̂, q̂ are the vector representations of p̂1 , p̂2 , p̂3 and q̂1 , q̂2 , q̂3 , respectively.
2c1 ln n
,
n
Lemma 8. Let p ∈ [0, 1], q ≥
p ≤ u(S)q, and mp̂ = m(p̂1 , p̂2 , p̂3 ) ∼ Poi(mp)3 , nq̂ = n(q̂1 , q̂2 , q̂3 ) ∼ Poi(nq)3 be
independent. Moreover, let c1 > 96 and 8c2 ln 2 < ∈ (0, 1), we have
1
u(S)
+
m ln m n ln n
2
1
p(1 + ln u(S))2
q
u(S) pu(S)
u(S)
Var(T s (p̂, q̂)) . 2− +
+
+
+
+
.
m
m
m
mn
n
n ln n
|ET s (p̂, q̂) + p ln(p/q)| .
(135)
(136)
ln m
2) “Non-smooth” regime I: Next we consider the “non-smooth” regime I where p ≥ c12m
, q ≤ 2c1nln n . By construction
c1 ln m
of the approximation-based estimator Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ), when p̂2 ≥ 3m and the approximation region
"
!
#
r
1
1 c1 p̂2 ln m
4c1 ln n
p̂2 −
,
(137)
u(S)
2
m
n
contains the true parameter q, the bias of this estimator is essentially the product of p and the best polynomial approximation
error of ln q in the previous approximation region. This approximation error can be easily obtained, for the 1D polynomial
approximation is well-understood (Lemma 23 gives
q an upper bound for the approximation error of ln x). Moreover, note that
ln m
the previous event occurs if c13m
≤ p̂2 ≤ p+ 12 c1 p̂2mln m , which holds with overwhelming probability by confidence sets. As
Qs−1
for the variance, it suffices to bound the variance of each term of the form l=0 (Y − nl ), where nY ∼ Poi(nq). Complicated
as it may seem, the present authors showed in [24] that the variance has an explicit expression in Poisson models, which is
19
the so-called Charlier polynomial [42].
Hence, we have good tools for the analysis of both the bias and the variance of T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ), which is presented in
the following lemma.
ln m
≤ p ≤ u(S)q, q ≤ 2c1nln n , and Poisson random variables m(p̂1 , p̂2 ) ∼ Poi(mp) × Poi(mp), n(q̂1 , q̂2 ) ∼
Lemma 9. Let c12m
Poi(nq) × Poi(nq) be independent. If c1 ≥ 2c2 and c2 ln n ≥ 1, we have
42c1 u(S)
|E[T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )] − p ln q| ≤ Cln pW
·
+ 2m−c1 /36 (1 − p ln q)
c22
pn ln n
1
16c1 u(S) ln n 2c1 u(S) ln n
2
2
+
42Cln
+ 1−11c2 ln 2
+ (c2 ln n)2 (Cln
+ (ln n)2 ) (138)
n
n
m
1
16c1 u(S) ln n 2c1 u(S) ln n
2
2
+
42Cln
+ (c2 ln n)2 (Cln
+ (ln n)2 ) + 2m−c1 /36
Var(T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )) ≤
n1−11c2 ln 2
n
m
(139)
where W (·) and Cln > 0 are given in Lemma 23. In particular, by Lemma 5, if c1 > 72 and 11c2 ln 2 < ∈ (0, 1), we have
1
u(S)
u(S)(ln n)5 u(S) ln n
1
|E[T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )] + p ln(p/q)| .
+ pW (
)+
+
(140)
m ln m
pn ln n
n1−
n
m
1
u(S)(ln n)5 u(S) ln n
1
p(1 + (ln p)2 )
+ 3+
+
(141)
Var(T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )) .
m
m
n1−
n
m
u(S)(ln n)3
p
1
u(S)(ln n)5 u(S) ln n
1
.
+
+ 3+
+
.
(142)
mn
m m
n1−
n
m
u(S)
Here pW ( pn
ln n ) corresponds to the polynomial approximation error, which will become the leading term in the bias.
3) “Non-smooth” regime II: Now we consider the “non-smooth” regime II where p ≤ 2c1mln m , q ≤ 2c1nln n and p ≤ u(S)q.
By the estimator construction, it is necessary to deal with the best 2D polynomial approximation of p ln q in
4c1 ln n
2c1 ln m
2
,0 ≤ q ≤
, p ≤ u(S)q .
(143)
R = (p, q) ∈ [0, 1] : 0 ≤ p ≤
m
n
We emphasize that the polynomial approximation in general multivariate case is extremely complicated. Rice [43] wrote:
“The theory of Chebyshev approximation (a.k.a. best approximation) for functions of one real variable has
been understood for some time and is quite elegant. For about fifty years attempts have been made to generalize
this theory to functions of several variables. These attempts have failed because of the lack of uniqueness of best
approximations to functions of more than one variable. ”
We also show in [24] that the non-uniqueness can cause serious trouble: some polynomial that achieves the best approximation
error cannot be used in our general methodology in functional estimation. What if we relax the requirement of computing the
best approximation in the multivariate case, and merely analyze the best approximation rate (i.e., the best approximation error
up to a multiplicative constant)? That turns out also to be extremely difficult. Ditzian and Totik [44, Chap. 12] obtained the
error rate estimate on simple polytopes3 , balls, and spheres, and it remained open until very recently Totik [45] generalized
the results to general polytopes. For results in balls and spheres, the readers are referred to Dai and Xu [46]. We still know
little about regimes beyond polytopes, balls, and spheres.
Complicated as the general multivariate case is, it is still possible to solve our problem since the approximation region R
in (143) is a convex polytope. Now we review the general theory of polynomial approximation in convex polytopes [45]. In
Rd we call a closed set K ⊂ Rd a convex polytope if it is the convex hull of finitely many points. Let x ∈ K and e ∈ Rd be
a direction (i.e., a unit vector). For continuous function f on K, we define the r-th symmetric difference in the direction e as
r
X
r
r
k r
(144)
∆he f (x) =
(−1)
f x + ( − k)he
k
2
k=0
with the understanding that
x with direction e intersects K at
= 0 if x + 2r he or
point Ae,x , Be,x , we
∆rhe f (x)
3A
x − 2r he does not belong to K. Moreover, letting the line le,x through
define the normalized distance as
q
d˜K (e, x) = kx − Ae,x k2 · kx − Be,x k2 .
(145)
simple polytope in Rd is a polytope such that each vertex has d edges.
20
Denoting by Sd−1 the set of all unit vectors in Rd , we define the Ditzian–Totik modulus of smoothness as follows:
r
ωK
(f, t) = sup sup sup |∆rhd˜
e∈S d−1 x∈K h≤t
K (e,x)e
f (x)|.
(146)
The significance of this quantity is presented in the following lemma.
Lemma 10. [45] Let K ⊂ Rd be a d-dimensional convex polytope and r = 1, 2, · · · . Then, for n ≥ rd, we have
r
En [f ; K] ≤ M ωK
(f,
1
)
n
n
M X
1
r
(k + 1)r−1 Ek [f ; K] ≥ ωK
(f, )
r
n
n
(147)
(148)
k=0
where the constant M > 0 only depends on r and d.
r
Hence, Lemma 10 shows that once we compute the Ditzian–Totik modulus of smoothness ωK
(f, 1/n) for f , we immediately
obtain an upper bound for the best polynomial approximation error. Moreover, Lemma 10 also shows that this is essentially
also the lower bound. In our case, K = R, and f (p, q) = p ln q. Choosing r = 2, Lemma 24 gives an upper bound for the
Ditzian–Totik modulus of smoothness. Moreover, tracing back to the proof for the simple polytope case in [44], it suffices to
take the supremum in (146) over all directions e which are parallel to some edge of the simple polytope K, which makes the
r
evaluation of ωK
(f, t) much simpler.
Now we are in position to bound the bias and the variance of T̃ns,II (p̂1 , q̂1 ), which are summarized in the following lemma.
Lemma 11. Let p ≤ 2c1mln m ∧u(S)q, q ≤ 2c1nln n , and Poisson random variables m(p̂1 , p̂2 ) ∼ Poi(mp)×Poi(mp), n(q̂1 , q̂2 ) ∼
Poi(nq) × Poi(nq) be independent. If c1 ≥ 2c2 and c2 ln n ≥ 1, we have
2c2 ln n
2c2 ln n !
213 c21 c42 (u(S))2 (ln n)4
ln n
n(ln m + ln n)
2M C0 u(S)
+
1+ 1+
+
|E[T̃ns,II (p̂1 , q̂1 )] − p ln q| ≤
c22 n ln n
n2−26c2 ln 2
ln m
mu(S) ln n
(149)
!
2c
ln
n
2c
ln
n
2
2
213 c21 c42 (u(S))2 (ln n)4
ln n
n(ln m + ln n)
Var(T̃ns,II (p̂1 , q̂1 )) ≤
1
+
1
+
+
.
(150)
2−26c
ln
2
2
n
ln m
mu(S) ln n
where the universal constant M > 0 is given by Lemma 10 (with r = d = 2), and the constant C0 which only depends on c1
is given by Lemma 24. In particular, by [1, Lemma 5], if lnnn . mu(S)
ln m , there exists some universal constant B > 0 such that
1
u(S)
(u(S))2
+
+ 2−c2 B
m ln m n ln n
n
1
(u(S))2
Var(T̃ns,II (p̂1 , q̂1 ) + LH (p̂1 )) . 2−c2 B + 2−c2 B .
m
n
|E[T̃ns,II (p̂1 , q̂1 ) + LH (p̂1 )] + p ln(p/q)| .
(151)
(152)
We remark that the condition lnnn . mu(S)
ln m can be removed later in the construction of the adaptive estimator. In fact, the
reason why we need this condition here is that we use an arbitrary best 2D polynomial approximation, which is not unique in
general. This point is very subtle: as was shown by the present authors in [24], not all polynomials which can achieve the best
uniform approximation error can be used to construct the rate-optimal estimator. Actually, we will show in the next subsection
that a special approximating polynomial can achieve the same rate without this condition. Moreover, a more careful design
of the approximating polynomial should require different degrees on p and q (instead of fixing a total degree K = c2 ln n),
but it has yet been unknown to approximation theorists that how to analyze the corresponding approximation error in general
polytopes.
4) Overall performance: Now we analyze the performance of the entire estimator D̂0 . For simplicity, we define
Tns (p̂i , q̂i ) , (T̃ns,I (p̂i,1 , q̂i,1 ; p̂i,2 , q̂i,2 ) + UH (p̂i,1 ))1(p̂i,3 >
ξ(p̂i , q̂i ) , Tns (p̂i , q̂i )1(q̂i,3 ≤
c1 ln n
) + T s (p̂i , q̂i )1(q̂i,3
n
c1 ln m
c1 ln m
) + (T̃ns,II (p̂i,1 , q̂i,1 ) + LH (p̂i,1 ))1(p̂i,3 ≤
) (153)
m
m
c1 ln n
>
)
(154)
n
where T s is given in (134), and p̂i = (p̂i,1 , p̂i,2 , p̂i,3 ) is the vector representation of the independent components, and similarly
for q̂i . Based on the current notations, we have
D̂ = −
S
X
i=1
ξ(p̂i , q̂i )
(155)
21
and by the independence between different symbols, we have
|Bias(D̂)| ≤
Var(D̂) =
S
X
i=1
S
X
|Bias(ξ(p̂i , q̂i ))|
(156)
Var(ξ(p̂i , q̂i )).
(157)
i=1
Hence, it suffices to analyze the bias and the variance of each ξ(p̂i , q̂i ) separately and then add them all. Based on Lemma 9
and Lemma 11, the next lemma first analyzes the bias and the variance of Tns (p̂i , q̂i ).
Lemma 12. Let 0 ≤ p ≤ 1 ∧ u(S)q, 0 ≤ q ≤ 2c1nln n , and mp̂ = m(p̂1 , p̂2 , p̂3 ) ∼ Poi(mp)3 and nq̂ = n(q̂1 , q̂2 , q̂3 ) ∼ Poi(nq)3
be independent. Moreover, we assume that n & u(S), lnnn . mu(S)
ln m , c1 > 72 and c2 (B ∨ 11 ln 2) < ∈ (0, 1) (where B is
given in Lemma 11). Then,
ln m
1) when p ≤ c12m
,
1
u(S)
(u(S))2
,
+
+
m ln m n ln n
n2−
1
(u(S))2
Var(Tns (p̂, q̂)) . 2− +
.
m
n2−
|Bias(Tns (p̂, q̂))| .
2) when
c1 ln m
2m
<p<
(158)
(159)
2c1 ln m
,
m
u(S)
u(S)
(u(S))2
1
u(S)
+ pW (
)+
+
+
,
m ln m
pn ln n
n ln n
n2−
mn1−
1
(u(S))2
u(S)
u(S) 2
Var(Tns (p̂, q̂)) . 2− + 2−2 +
+ [pW (
)] .
1−
m
n
mn
pn ln n
|Bias(Tns (p̂, q̂))| .
3) when p ≥
(160)
(161)
2c1 ln m
,
m
u(S)
(u(S))2
u(S)
1
+ pW (
)+
+
,
(162)
m ln m
pn ln n
n2−
mn1−
1
(u(S))2
u(S)
Var(Tns (p̂, q̂)) . 2− + 2−2 +
.
(163)
m
n
mn1−
Note that in Lemma 12, the condition n & u(S) is a natural requirement for the consistency of the optimal estimator in
view of Theorem 1, and lnnn . mu(S)
ln m is the additional condition in Lemma 11. Now based on Lemma 8 and Lemma 12, we
are about to analyze the bias and the variance of ξ(p̂, q̂).
|Bias(Tns (p̂, q̂))| .
Lemma 13. Let 0 ≤ p ≤ 1 ∧ u(S)q, 0 ≤ q ≤ 1, and mp̂ = m(p̂1 , p̂2 , p̂3 ) ∼ Poi(mp)3 and nq̂ = n(q̂1 , q̂2 , q̂3 ) ∼ Poi(nq)3 be
independent. Moreover, we assume that n & u(S), lnnn . mu(S)
ln m , c1 > 96 and c2 (B ∨ 11 ln 2) < ∈ (0, 1/2) (where B is
given in Lemma 11). Then,
ln n
1) when q ≤ c12n
,
1
u(S)
(u(S))2
ln m
p ≤ c12m
,
m ln m + n ln n + n2−
2
u(S)
u(S)
(u(S))
u(S)
2c1 ln m
c
ln
m
1
1
(164)
|Bias(ξ(p̂, q̂))| . m ln m + pW ( pn ln n ) + n ln n + n2− + mn1−
,
2m < p <
m
2
1
u(S)
(u(S))
u(S)
p ≥ 2c1mln m .
m ln m + pW ( pn ln n ) + n2− + mn1−
2
ln m
1
+ (u(S))
p ≤ c12m
,
n2−
m2−
2
(u(S))
u(S)
u(S)
c
ln
m
2c1 ln m
1
2
1
Var(ξ(p̂, q̂)) . m2− + n2−2 + mn1− + [pW ( pn ln n )]
(165)
,
2m < p <
m
2
1
(u(S))
u(S)
p ≥ 2c1mln m .
m2− + n2− + mn1−
2) when
c1 ln n
2n
<q<
|Bias(ξ(p̂, q̂))| .
Var(ξ(p̂, q̂)) .
2c1 ln n
,
n
1
m ln m
+
1
m ln m
+
(u(S))2
n2−
u(S)
pW ( pn ln n ) + nu(S)
ln n
u(S)
n ln n
+
+
(u(S))2
n2−
+
u(S)
mn1−
p≤
c1 ln m
2m ,
ln m
p ≥ c12m
.
u(S)
1
p(1 + ln u(S))2
q
pu(S) (u(S))2
mn
+
+
+
+ 2−2 + (u(S))2
2
2−2 + [pW ( u(S) )]2
m
m
m
n
n
n
pn ln n
(166)
p≤
p>
c1 ln m
2m ,
c1 ln m
2m .
(167)
22
3) when q ≥
2c1 ln n
,
n
1
u(S)
+
,
m ln m n ln n
p(1 + ln u(S))2
1
q
u(S) pu(S)
u(S) 2
Var(ξ(p̂, q̂)) . 2− +
+
+
+
+(
) .
m
m
m
mn
n
n ln n
|Bias(ξ(p̂, q̂))| .
(168)
(169)
Based on Lemma 13, we can analyze the total bias and variance of our estimator D̂. By differentiation, we have
pW (
and the maximum is attained at p =
|Bias(D̂)| ≤
Var(D̂) =
S
X
i=1
S
X
i=1
u(S)
en ln n .
u(S)
u(S)
)≤
pn ln n
en ln n
(170)
Hence,
|Bias(ξ(p̂i , q̂i ))| .
Var(ξ(p̂i , q̂i )) .
S
Su(S)
Su(S) S(u(S))2
+
+
+
2−2
m ln m
n ln n
n
mn1−
S(u(S))2
Su(S)
(1 + ln u(S))2
S
u(S)
+
+
+
+
.
m2−
n2−2
mn1−2
m
n
(171)
(172)
If we further require that ln S & ln(m ∨ n) and n & Su(S)/ ln S, the previous results can be further upper bounded as (let
→ 0)
Su(S)
Su(S) 2
S
S(u(S))2
S
+
+(
) +
+
(173)
|Bias(D̂)| .
m ln m
n ln n
n ln n
m2
n2−2
S
Su(S)
Su(S) 2
S
Su(S) 2
.
+
+(
) +
+(
)
(174)
2
m ln m
n ln n
n ln n
m
n ln n
Su(S)
S
+
(175)
.
m ln m
n ln n
S(u(S))2
S
S
Su(S) 2
u(S)
(1 + ln u(S))2
Var(D̂) . (
)2 + (
) +
+
+
(176)
+
m ln m
n ln n
m2
n2−4
m
n
S
Su(S) 2 (ln u(S))2
u(S)
.(
)2 + (
) +
+
.
(177)
m ln m
n ln n
m
n
Hence we come to the following theorem.
Theorem 5. Let m & lnSS , n & Su(S)
ln S , ln S & ln(m ∨ n) and
from the general recipe, we have
2
sup
E(P,Q) D̂ − D(P kQ) .
(P,Q)∈US,u(S)
n
ln n
.
mu(S)
ln m .
Then for our estimator D̂ in (103) constructed
S
Su(S)
+
m ln m
n ln n
2
+
(ln u(S))2
u(S)
+
.
m
n
(178)
Moreover, D̂ does not require the knowledge of the support size S.
C. An adaptive estimator
So far we have obtained an essentially minimax estimator via our general recipe. However, since this estimator is purely
obtained from the general method, it is not surprising that it is also subject to some disadvantages. Firstly, in the estimator
we do not specify the explicit form of the best 2D polynomial approximation in the “non-smooth” regime II. Although the
best 1D polynomial approximation is unique and can be efficiently obtained via the Remez algorithm [47], which has been
efficiently implemented in Matlab [48], the best 2D polynomial approximation is not unique and hard to compute. Moreover,
as what we have remarked before, the non-uniqueness forces us to add an unnecessary condition lnnn . mu(S)
ln m in Lemma 11
and thus in Theorem 5. Secondly, although the estimator does not require the knowledge of the support size S (we remove the
constant term in the polynomial approximation), but it requires the upper bound on the likelihood ratio u(S) (in the design of
“non-smooth” regime I). In practice, we wish to obtain an adaptive estimator which achieves the minimax rate and is agnostic to
both S and u(S). Thirdly, for the estimator construction in the “non-smooth” regime I, the approximating polynomial depends
on the empirical probabilities (i.e., we cannot store the polynomials in advance), which incurs large computational complexity.
To resolve these issues, we need to apply some tricks to explicitly construct an approximating polynomial for f (p, q) = p ln q
in the “non-smooth” regime, i.e., q ≤ 2c1nln n . We first suppose that there exists a 1D polynomial T (q) in q with degree . ln n
such that pT (q) has the desired approximation property for f (p, q) = p ln q in the entire “non-smooth” regime, i.e., we need
not to distinguish “non-smooth” regimes I and II. We remark that either the 1D approximation or only one approximation on
the entire “non-smooth” regime is not always doable in general. For example, for estimating the `1 distance, it has been shown
23
in [24] that not only any single approximation in the entire “non-smooth” regime will always fail to give the correct order of
the approximation error, but any 1D polynomial approximation of |p − q| in p − q will also not work when both p and q are
small. Nevertheless, this ambitious target can be achieved in our special example. Motivated by Lemma 9 and Lemma 11, the
correct order of the approximation error is nu(S)
ln n , i.e., T (q) should satisfy that
sup
0≤q≤
2c1 ln n
n
Since p ≤ u(S)q, it suffices to have
|T (q) − ln q| .
|pT (q) − p ln q| .
1
,
qn ln n
u(S)
.
n ln n
∀q ∈ (0,
2c1 ln n
]
n
(179)
(180)
i.e., to find a 1D polynomial approximation which satisfies the desired pointwise bound. However, it is easy to show that there
exists some polynomial T0 (q) on [0, 2c1nln n ] such that deg(T0 ) . ln n and
sup
0≤q≤
2c1 ln n
n
|T0 (q) − q ln q| .
1
.
n ln n
(181)
Hence, if we remove the constant term of T0 (q) and define T (q) = T0 (q)/q, then T (q) will have the desired property.
Motivated by the previous observations, we can construct an explicit estimator as follows:
Estimator Construction 2. Conduct three-fold sample splitting to obtain i.i.d. samples (p̂i,1 , p̂i,2 , p̂i,3 ) and (q̂i,1 , q̂i,2 , q̂i,3 ).
The adaptive estimator D̂A for the KL divergence D(P kQ) is
S
X
c1 ln m
c1 ln m
LH (p̂i,1 )1(p̂i,3 ≤
D̂A = −
) + UH (p̂i,1 )1(p̂i,3 >
)
m
m
i=1
c1 ln n
c1 ln n
+T̃ns (p̂i,1 , q̂i,1 )1(q̂i,3 ≤
) + T̃s (p̂i,1 , q̂i,1 ; p̂i,2 , q̂i,2 )1(q̂i,3 >
)
(182)
n
n
where T̃s is given by (106), LH (x), UH (x) are given by [1] and (108), respectively, and
T̃ns (x, y) = (Tns (x, y) ∧ 1) ∨ (−1)
−k
K
k−1
X
Y
2c1 ln n
l
Tns (x, y) =
gK,k+1
·x
y−
n
n
k=0
(184)
l=0
where the coefficients {gK,k }K+1
k=1 are given by the best polynomial approximation of x ln x as follows:
2c1 ln n
gK,1 = rK,1 + ln
,
gK,k = rK,k , 2 ≤ k ≤ K + 1
n
K+1
X
rK,k xk = arg min1
sup |P (x) − x ln x|.
k=0
(183)
P ∈polyK+1 x∈[0,1]
(185)
(186)
The parameters c1 , c2 > 0 are suitably chosen universal constants. A pictorial illustration of D̂A is displayed in Fig. 4.
Recall that the entropy estimator in [1] does not require the knowledge of S, we conclude that D̂A always sets zero to
unseen symbols and does not depend on u(S). In other words, the estimator D̂ for D(P kQ) is agnostic to both S and u(S),
and is thus adaptive. Moreover, the estimator D̂A is easy to implement in practice with near-linear computational complexity,
and the coefficients {gK,k }K+1
k=1 can be obtained offline via the Remez algorithm before observing any samples.
Now we analyze the performance of T̃ns (p̂, q̂) when q ≤ 2c1nln n .
Lemma 14. Let 0 ≤ p ≤ u(S)q, 0 ≤ q ≤
c1 ≥ 2c2 and c2 ln n ≥ 1, we have
2c1 ln n
,
n
and mp̂ ∼ Poi(mp), nq̂ ∼ Poi(nq) be independent random variables. If
Cu(S)
16c2
2u(S)
4(u(S))2
+ 11c2 2ln 2 (C + 2c1 (ln n)3 )2 ·
+
n ln n
n
c1 mn ln n
n2
2u(S)
16c2
4(u(S))2
Var(T̃ns (p̂, q̂)) ≤ 11c2 2ln 2 (C + 2c1 (ln n)3 )2 ·
+
.
n
c1 mn ln n
n2
|E[T̃ns (p̂, q̂)] − p ln q| ≤
(187)
(188)
24
p
1
“Non-smooth” regime:
p̂× unbiased estimate of a special
1D polynomial approximation of ln q
0
“Smooth” regime:
plug-in approach with
order-three bias correction
q
c1 ln n
n
Fig. 4: Pictorial explanation of our adaptive estimator for
PS
i=1
pi ln qi .
where C > 0 is a constant which only depends on c1 and c2 . In particular, if 11c2 ln 2 < ∈ (0, 1), by [1, Lemma 5] we have
(u(S))2
u(S)
u(S)
+
+
2−
n ln n
n
mn1−
2
(u(S))
u(S)
Var(T̃ns (p̂, q̂)) .
+
.
n2−
mn1−
|ET̃ns (p̂, q̂) − p ln q| .
(189)
(190)
Note that in Lemma 14 we have removed the condition lnnn . mu(S)
ln m in Lemma 11. Moreover, since the upper bounds of the
bias and variance of T̃ns presented in Lemma 14 are no worse than those in Lemma 9 and Lemma 11, by the same argument
in Lemma 12 and Lemma 13 we conclude that the adaptive estimator D̂A is rate-optimal, and thereby satisfies Theorem 1.
IV. M INIMAX L OWER B OUND
In this section, we prove the minimax lower bounds presented in Theorem 1. There are two main lemmas that we employ
towards the proof of the minimax lower bound. The first is the Le Cam two-point method, which helps to prove the minimax
lower bound corresponding to the variance, or equivalently, the classical asymptotics. Suppose we observe a random vector
Z ∈ (Z, A) which has distribution Pθ where θ ∈ Θ. Let θ0 and θ1 be two elements of Θ. Let T̂ = T̂ (Z) be an arbitrary
estimator of a function T (θ) based on Z. Le Cam’s two-point method gives the following general minimax lower bound.
Lemma 15. [5, Sec. 2.4.2] The following inequality holds:
|T (θ1 ) − T (θ0 )|
1
inf sup Pθ |T̂ − T (θ)| ≥
≥ exp (−D (Pθ1 kPθ0 )) .
2
4
T̂ θ∈Θ
(191)
The second lemma is the so-called method of two fuzzy hypotheses presented in Tsybakov [5]. Suppose we observe a
random vector Z ∈ (Z, A) which has distribution Pθ where θ ∈ Θ. Let σ0 and σ1 be two prior distributions supported on Θ.
Write Fi for the marginal distribution of Z when the prior is σi for i = 0, 1. Let T̂ = T̂ (Z) be an arbitrary estimator of a
function T (θ) based on Z. We have the following general minimax lower bound.
Lemma 16. [5, Thm. 2.15] Given the setting above, suppose there exist ζ ∈ R, s > 0, 0 ≤ β0 , β1 < 1 such that
σ0 (θ : T (θ) ≤ ζ − s) ≥ 1 − β0
σ1 (θ : T (θ) ≥ ζ + s) ≥ 1 − β1 .
(192)
(193)
25
If V (F1 , F0 ) ≤ η < 1, then
1−η−β −β
0
1
,
inf sup Pθ |T̂ − T (θ)| ≥ s ≥
2
T̂ θ∈Θ
(194)
where Fi , i = 0, 1 are the marginal distributions of Z when the priors are σi , i = 0, 1, respectively.
Here V (P, Q) is the total variation distance between two probability measures P, Q on the measurable space (Z, A).
Concretely, we have
Z
1
|p − q|dν,
(195)
V (P, Q) , sup |P (A) − Q(A)| =
2
A∈A
dQ
where p = dP
dν , q = dν , and ν is a dominating measure so that P ν, Q ν.
Su(S) 2
S
By the proof of the achievability results in previous sections, we observe that ( m ln
m + n ln n ) corresponds to the squared
(ln S)2
u(S)
bias term, and m + n corresponds to the variance term. In the sequel we will also prove the minimax lower bound for
the squared bias term and the variance term separately.
A. Minimax lower bound for the “variance”
First we prove that when n & u(S), we have
inf
D̂
2
u(S)
.
E(P,Q) D̂ − D(P kQ) &
n
(P,Q)∈US,u(S)
sup
(196)
1
1
, · · · , 2(S−1)
, 12 ). Applying Lemma 15 to our Poisson sampling model nq̂i ∼ Poi(nqi ), 1 ≤ i ≤ S, we know
Fix P = ( 2(S−1)
that for feasible θ1 = P ×Q1 = (p1 , · · · , pS )×(q1 , · · · , qS ), θ0 = P ×Q0 = (p1 , · · · , pS )×(q10 , · · · , qS0 ) with θ0 , θ1 ∈ US,u(S) ,
D (Pθ1 kPθ0 ) =
=
=
=
S
X
i=1
S
X
(197)
D(Poi(nqi )kPoi(nqi0 ))
(198)
i=1
∞
S X
X
i=1 k=0
S
X
qi
P (Poi(nqi ) = k) · k ln 0 − n(qi − qi0 )
qi
npi ln
i=1
then Markov’s inequality yields
D (Poi(mpi ) × Poi(nqi )kPoi(mpi ) × Poi(nqi0 ))
S
X
qi
(qi − qi0 )
−
n
qi0
i=1
= nD(Q1 kQ0 ),
2
|D(θ1 ) − D(θ0 )|2
E(P,Q) D̂ − D(P kQ) ≥
×
4
D̂ (P,Q)∈US,u(S)
|D(θ1 ) − D(θ0 )|
inf sup P |D̂ − D(P kQ)| ≥
2
D̂ P ∈MS
|D(P kQ1 ) − D(P kQ0 )|2
≥
exp (−nD(Q1 kQ0 ))
16
where we are operating under the Poisson sampling model.
Fix ∈ (0, 1/2) to be specified later. Letting
1
1
1
Q1 =
,··· ,
,1 −
,
(S − 1)u(S)
(S − 1)u(S)
u(S)
1+
1−
1+
1−
1
Q0 =
,
,··· ,
,
,1 −
,
(S − 1)u(S) (S − 1)u(S)
(S − 1)u(S) (S − 1)u(S)
u(S)
inf
(199)
(200)
(201)
sup
(202)
(203)
(204)
(205)
26
where without loss of generality we assume that S is an odd integer. Direct computation yields
1
1
ln(1 + ) −
ln(1 − )
2u(S)
2u(S)
1
=−
ln(1 − 2 )
2u(S)
2
,
≤
u(S)
D(Q1 kQ0 ) = −
(206)
(207)
(208)
and
Hence, by choosing =
q
1
1
|D(P kQ1 ) − D(P kQ0 )| ≥ − ln(1 + ) − ln(1 − )
4
4
1
2
= − ln(1 − )
4
2
≥ .
4
u(S)
n ,
(209)
(210)
(211)
we know that
inf
D̂
2
u(S)
E(P,Q) D̂ − D(P kQ) ≥
256en
(P,Q)∈US,u(S)
sup
under the Poisson sampling model. The result for the multinomial case can be obtained via Lemma 22.
Next we apply Lemma 15 to show that
2
(ln u(S))2
sup
E(P,Q) D̂ − D(P kQ) &
inf
.
m
D̂ (P,Q)∈US,u(S)
1
1
, · · · , 2(S−1)u(S)
,1 −
Now fix Q = ( 2(S−1)u(S)
1
2u(S) ),
(213)
and consider
1
,···
2(S − 1)
1−
,···
P0 =
2(S − 1)
P1 =
(212)
1
1
,
2(S − 1) 2
1− 1+
,
,
2(S − 1) 2
,
(214)
(215)
with ∈ (0, 21 ) to be specified later. By the same argument, for θ1 = P1 × Q ∈ US,u(S) , θ0 = P0 × Q ∈ US,u(S) , we have
2
|D(P1 kQ) − D(P0 kQ)|2
sup
E(P,Q) D̂ − D(P kQ) ≥
exp (−mD(P1 kP0 ))
(216)
inf
16
D̂ (P,Q)∈US,u(S)
under the Poisson sampling model. It is straightforward to compute that
D(P1 kP0 ) =
and
1
1
1
1
1
2
ln
+ ln
= − ln(1 − 2 ) ≤
2 1− 2 1+
2
4
1
u(S)
1−
1 + (1 + )u(S)
1
ln u(S) + ln
−
(ln u(S) + ln(1 − )) −
ln
2
2 2u(S) − 1
2
2
2u(S) − 1
1+
1−
u(S)
=
ln u(S) −
ln(1 + ) −
ln(1 − ) − ln
2
2
2
2 2u(S) − 1
ln u(S)
≥
.
2
|D(P1 kQ) − D(P0 kQ)| =
(217)
(218)
(219)
(220)
1
Combining these inequalities and setting = m− 2 completes the proof of (213).
B. Minimax lower bound for the “squared bias”
We employ Lemma 16 to prove the minimax lower bounds corresponding to the squared bias terms. First we show that
when m & lnSS and u(S) & (ln S)2 , ln S & ln m, we have
2 S 2
inf
sup
E(P,Q) D̂ − D(P kQ) &
.
(221)
m ln m
D̂ (P,Q)∈US,u(S)
27
In fact, by choosing Q to be the uniform distribution, the estimation of KL divergence reduces to the estimation of entropy of
for all i = 1, · · · , S. Since in the proof of the
discrete distribution P , subject to an additional constraint pi ≤ u(S)qi = u(S)
S
minimax lower bound in [30], all pi satisfy that pi . lnmm , and our assumption implies
(ln S)2
u(S)
ln m
.
.
(222)
m
S
S
i.e., the additional condition is automatically satisfied. Hence, we can operate as if we do not have the additional condition,
and [30] gives
2 S 2 S 2
E(P,Q) D̂ − D(P kQ) &
.
(223)
inf
sup
m ln S
m ln m
D̂ (P,Q)∈US,u(S)
pi .
where we have used the condition m & lnSS and ln S & ln m to give ln S ln m here.
2
Now we are about the prove that when m & lnSS , n & Su(S)
ln S and u(S) & (ln S) , ln S & ln n, we have
2 Su(S) 2
inf
E(P,Q) D̂ − D(P kQ) &
.
sup
n ln n
D̂ (P,Q)∈US,u(S)
(224)
We begin with a lemma to construct two measures with matching moments and large difference on the functional value,
which corresponds to the duality between function space and measure space.
Lemma 17. [1, Lemma 10, Lemma 12] For any bounded interval I ⊂ R, positive integer K > 0 and continuous function f
on I, there exist two probability measures ν0 and ν1 supported on I such that
R
R
1) R tl ν1 (dt) = tRl ν0 (dt), for all l = 0, 1, 2, · · · , L;
2) f (t)ν1 (dt) − f (t)ν0 (dt) = 2EK [f ; I].
Recall that EK [f ; I] is the distance in the uniform norm on I from the function f (x) to the space spanned by {1, x, · · · , xK }.
d1 ln n
1
Based on Lemma 17, we choose I = [ n ln
n,
n ], K = d2 ln n with universal constants d1 , d2 > 0 to be specified later,
and f (x) = ln x. The following lemma presents a lower bound of the approximation error EK [f ; I].
d1 ln n
1
Lemma 18. [30] For K = d2 ln n, I = [ n ln
n,
n ] and f (x) = ln x, we have
EK [f ; I] ≥ c0
(225)
where the constant c0 > 0 only depends on d1 , d2 .
R
R
Define µ , tν1 (dt) = tν0 (dt), by construction we have µ ≤ d1 nln n . Now the two fuzzy hypotheses σ0 , σ1 in Lemma
16 are constructed as follows: for each i = 0, 1, σi fixes
u(S)
u(S)
(S − 1)u(S)
P =
,··· ,
,1 −
(226)
n ln n
n ln n
n ln n
and assigns νiS−1 to the vector (q1 , · · · , qS−1 ), and fixes qS = 1 − (S − 1)µ. Note that by assumption,
Sµ .
S ln n
(ln S)2
.
.1
n
u(S)
(227)
thus qS takes positive value and is thus valid under proper parameter configurations. Moreover, it is straightforward to verify
that (P, Q) ∈ US,u(S) with probability one under σi . Since under σi , Q may not form a probability measure, we consider the
set of approximate probability vectors
(
)
S
X
US,u(S) () , (P, Q) : P ∈ MS , |
qi − 1| < , pi ≤ u(S)qi , ∀i
(228)
i=1
with parameter ∈ (0, 1) to be specified later. Further define the minimax under the Poisson sampling model for estimating
D(P kQ) with (P, Q) ∈ US,u(S) () as
2
RP (S, m, n, u(S), ) , inf
sup
E(P,Q) D̂ − D(P kQ) .
(229)
D̂ (P,Q)∈US,u(S) ()
The equivalence between RP (S, m, n, u(S), ) and R(S, m, n, u(S)) defined in (273) is established in the following lemma.
Lemma 19. For any S, m, n ∈ N and ∈ (0, 1/2), we have
R(S, m, n, u(S)) ≥
1
u(S)
m
n
RP (S, 2m, 2n,
, ) − (ln u(S))2 exp(− ) + exp(− ) − 22 .
2
1+
4
4
(230)
28
Then condition σi on the event
S
EK [f ; I]
1 X
(ln qj − Eσi ln qj ) ≤
,
Ei , US,u(S) () ∩ (P, Q) :
S j=1
2
i = 0, 1
(231)
and define the conditional probability distribution as
πi (·) ,
σi (· ∩ Ei )
,
σi (Ei )
(232)
i = 0, 1.
By setting
d1 = 1,
we have
σi (US,u(S) ()c ) = σi
S
X
j=1
q j − 1 ≥ ≤
d2 = 10e,
=
S
n ln n
(233)
2
S−1
S−1
1 X
d2 (ln n)4
1 X
S d1 ln n
2
= 1
Var
(q
)
≤
E
(q
)
≤
→ 0 (234)
νi j
νi j
2
2
2
j=1
j=1
n
S
and by Lemma 18,
S
S−1
X
X
(ln(n ln n))2
(ln S)2
E
[f
;
I]
4
1
K
≤ 4
(ln qj − Eσi ln qj ) ≥
Var
·
.
→ 0.
σi
(ln
q
)
≤
ν
j
i
S j=1
2
(c0 S)2 j=1
(c0 )2
S
S
Hence, by the union bound,
(235)
S
X
EK [f ; I]
1
(ln qj − Eσi ln qj ) ≥
σi (Eic ) ≤ σi (US,u(S) ()c ) + σi
→ 0.
S j=1
2
(236)
V (F0 , F1 ) ≤ V (F0 , G0 ) + V (G0 , G1 ) + V (G1 , F1 )
(237)
Denote by Fi , Gi the marginal probability under prior πi and σi , respectively, for each i = 0, 1. Now by the triangle inequality
and [1, Lemma 11], we have
≤
σ0 (E0c )
+ V (G0 , G1 ) +
S
≤ σ0 (E0c ) + 6 + σ1 (E1c )
n
→ 0.
σ1 (E1c )
(239)
(240)
Moreover, by the definition of πi , the first two conditions of Lemma 16 hold with β0 = β1 = 0 for
(S − 1)u(S)
(S − 1)u(S) Eν0 (ln q) + Eν1 (ln q)
·
− 1−
ln(1 − (S − 1)µ)
ζ = H(P ) −
n ln n
2
n ln n
(S − 1)u(S) Eν1 (ln q) − Eν0 (ln q) − 2 ·
·
n ln n
2
Hence, by Lemma 16, we conclude that
s=
EK [f ;I]
2
RP (S, m, n, u(S), ) & s &
2
(238)
=
(S − 1)u(S) EK [f ; I]
Su(S)
·
&
.
n ln n
2
n ln n
Su(S)
n ln n
2
and the desired bound (224) follows from Lemma 19.
Hence, the combination of (196), (213), (221) and (224) yields that when u(S) & (ln S)2 , n & Su(S)
ln S , m &
ln S & ln(m ∨ n), we have
2 Su(S) 2 S 2 (ln u(S))2
u(S)
inf
sup
E(P,Q) D̂ − D(P kQ) &
+
+
+
n ln n
m ln m
m
n
D̂ (P,Q)∈US,u(S)
(241)
(242)
(243)
S
ln S
and
(244)
and the proof of Theorem 1 is complete.
V. O PTIMAL E STIMATORS FOR H ELLINGER D ISTANCE AND χ2 -D IVERGENCE
Having analyzed the minimax rate-optimal estimator for the KL divergence thoroughly, in this section we apply the general
recipe to other divergence functions, i.e., the Hellinger distance and the χ2 -divergence. Specifically, we explicitly construct the
minimax rate-optimal estimators for the Hellinger distance and the χ2 -divergence, and sketch the proof of the achievability
29
part in Theorem 3 and 4. For brevity, we omit the complete proof and remark that it can be obtained in a similar fashion as
in the analysis of the KL divergence.
A. Optimal estimator for the Hellinger distance
√
√
For the Hellinger distance, the bivariate function of interest is f (p, q) = ( p − q)2 . We first classify the regime. In this
case, Θ = Θ̂ = [0, 1]2 , and the non-analytic regime is
Θ̂0 = {(p, q) ∈ [0, 1]2 : p = 0 or q = 0}.
(245)
Based on the confidence sets in Poisson models, the “smooth” and “non-smooth” regimes for (p, q) can obtained via (23) and
(24) as
c1 ln n
c1 ln m
, 1]) ∪ ([
, 1] × [0, 1])
(246)
2n
2m
2c1 ln m
2c1 ln n
]) ∪ ([0,
] × [0, 1])
(247)
Θns = ([0, 1] × [0,
n
m
where c1 > 0 is some universal constant. Further, by (25) and (26), we obtain the “smooth” and “non-smooth” regimes based
on observations (p̂, q̂) as
Θs = ([0, 1] × [
c1 ln n
c1 ln m
, 1]) ∪ ([
, 1] × [0, 1])
n
m
c1 ln m
c1 ln n
]) ∪ ([0,
] × [0, 1]).
Θ̂ns = ([0, 1] × [0,
n
m
√
√
Next we estimate the quantity f (p, q) = ( p − q)2 in each regime. In the “smooth” regime where p ≥
q ≥ c1 nln n , we simply employ the plug-in approach with no bias correction:
p
p
Ts (p̂, q̂) = ( p̂ − q̂)2 .
Θ̂s = ([0, 1] × [
(248)
(249)
c1 ln m
m
and
(250)
ln m
In the “non-smooth” regime, by symmetry it suffices to consider the case where p ≤ c1 m
. Now we need to find a proper
√
√ 2
√
polynomial P (p, q) to approximate f (p, q) = ( p − q) = p + q − 2 pq. Recall that the degree of the approximating
polynomial P (p, q) is determined by the bias-variance tradeoff, and we will have the following result after careful analysis:
c ln n
• when q̂ ≥ 1 n , the resulting polynomial P (p, q) should be of degree c2 ln m on p and of degree 0 on q;
c1 ln n
• when q̂ <
n , the resulting polynomial P (p, q) should be of degree c2 ln m on p and of degree c2 ln n on q.
Now
we
explicitly
give the expression of P (p, q) in both cases in terms of the following best approximating polynomial of
√
x:
QK (x) =
K
X
k=0
aK,k xk , arg min 1 max |Q(z) −
Q∈polyK
z∈[0,1]
√
z|.
(251)
Recall that we use the sample splitting technique to determine the approximation region and approximate the functional,
respectively, i.e., the approximation region is based on (p̂2 , q̂2 ), while
q the polynomial
q is evaluated using (p̂1 , q̂1 ). Hence,
when q̂ ≥ c1 nln n , the approximation region is [0, 2c1mln m ]×[q̂2 − c1 q̂2nln n , q̂2 + c1 q̂2nln n ], and by the degree requirements
√
√
of P (p, q), a natural choice is P (p, q) = ∆m QKm ( ∆pm ) · q̂2 with Km = c2 ln m, ∆m = 2c1mln m ;
c ln n
2c ln m
• when q̂ < 1 n , the approximation region is [0, 1m
] × [0, 2c1nln n ], and by the degree requirements of P (p, q), a
√
√
p
q
natural choice is P (p, q) = ∆m QKm ( ∆m )· ∆n QKn ( ∆n ) with Km = c2 ln m, ∆m = 2c1mln m and Kn = c2 ln n, ∆n =
2c1 ln n
.
n
In summary, the estimator is constructed as follows.
•
Estimator Construction 3. Conduct three-fold sample splitting to obtain i.i.d. samples (p̂i,1 , p̂i,2 , p̂i,3 ) and (q̂i,1 , q̂i,2 , q̂i,3 ).
The estimator T̂ for the Hellinger distance H 2 (P, Q) is given by
S
X
p
p
c1 ln m
c1 ln m
c1 ln n
c1 ln n
T̂ = 1 −
p̂i,2 1(p̂i,3 ≥
) + R̃m (p̂i,1 )1(p̂i,3 <
)
q̂i,2 1(q̂i,3 ≥
) + R̃n (q̂i,1 )1(q̂i,3 <
)
m
m
n
n
i=1
(252)
where for l = m, n
R̃l (x) , (Rl (x) ∧ 1) ∨ (−1)
−k+ 12 k−1
cX
2 ln l
Y
2c1 ln l
j
Rl (x) ,
ac2 ln l,k
(x − )
l
l
j=0
k=1
(253)
(254)
30
coefficients {aK,k } are given by (251), and c1 , c2 > 0 are some suitably chosen universal constants.
The previous estimator does not require the knowledge of S since it assigns zero to unseen
(we remove the constant
PS symbols
√
term in the expression of Rl (x)). Note that since the Hellinger distance H 2 (P, Q) = 1− i=1 pi qi enjoys a natural separation
in its variables (pi , qi ), the pair (pi , qi ) is also separated in our resulting estimator. Moreover, in our estimator construction we
can also merge (p̂i,1 , p̂i,2 ) and (q̂i,1 , q̂i,2 ) to result in a two-fold sample splitting.
Now we analyze the bias of the previous estimator. When p < 2c1mln m is small, by [1, Lemma 17], we know that
√
p
p
∆m
1
√
√
|ERm (p̂) − p| = | ∆m QKm (
.√
.
(255)
) − p| .
2
∆m
Km
m ln m
When p ≥
2c1 ln m
m
is large, by Lemma 6 with r = 0 we know that
√
p
p
d2 ξ
1
1
√
. √ .√
.
|E p̂ − p| .
· sup
2
m ξ∈[p/2,2p] dξ
m p
m ln m
(256)
Hence, the total bias of T̂ can be upper bounded as
s
r
r
√
√
S
X
qi
pi
S
S
S
√
|Bias(T̂ )| .
+√
≤
+
m ln m
n ln n
(m ∧ n) ln(m ∧ n)
m ln m
n ln n
i=1
(257)
as shown in Theorem 3. The variance can also be obtained in a similar fashion, and we omit the details.
B. Optimal estimator for the χ2 -divergence
2
For the χ2 -divergence χ2 (P, Q) over US,u(S) , the bivariate function of interest is f (p, q) = pq . We first classify the regime.
Since this function shares very similar analytic properties as the function p ln q used in the KL divergence case, and our
parameter set US,u(S) remains the same, here the “smooth” and “non-smooth” regimes are also given by
c1 ln n
, 1]
n
c1 ln n
Θ̂ns = [0, 1] × [0,
]
n
Θ̂s = [0, 1] × [
(258)
(259)
with some universal constant c1 2> 0.
Next we estimate f (p, q) = pq in each regime. In the “smooth” regime where q ≥ c1 nln n , we seek to correct the bias of
the plug-in estimator 1/q̂ in estimating 1/q. Based on our general bias correction technique (31), we simply use the following
order-three bias correction:
T̂ (3) (q̂1 , q̂2 ) =
1
(q̂2 − q̂1 )2
4q̂2
(q̂2 − q̂1 )3
3q̂22
6q̂2
q̂2 − q̂1
+
−
−
+
− 2 4.
−
q̂1
q̂12
q̂13
nq̂13
q̂14
nq̂14
n q̂1
(260)
1
Since p2 admits an unbiased estimate p̂(p̂ − m
) in the Poisson model mp̂ ∼ Poi(mp), the overall estimator in the “smooth”
regime is given by
1
T̂s (p̂1 , q̂1 ; p̂2 , q̂2 ) = p̂1 p̂1 −
· T̂ (3) (q̂1 , q̂2 ) · 1(q̂1 6= 0)
(261)
m
1
1
q̂2 − q̂1
(q̂2 − q̂1 )2
4q̂2
(q̂2 − q̂1 )3
3q̂ 2
6q̂2
= p̂1 p̂1 −
·
−
+
− 3−
+ 24 − 2 4 · 1(q̂1 6= 0). (262)
2
3
4
m
q̂1
q̂1
q̂1
nq̂1
q̂1
nq̂1
n q̂1
In the “non-smooth” regime, as is in the KL divergence case, we can further distinguish into “non-smooth” regime I and
“non-smooth” regime II and employ best polynomial approximation in these regimes, respectively. However, motivated by the
adaptive estimator D̂A for the KL divergence where a single polynomial approximation is enough, we wonder whether or not
it is also the case in the estimation of χ2 -divergence. Specifically, we seek a polynomial QK (q) of degree K on [0, ∆n ] such
that the following quantity
sup
q∈(0,∆n ]
q2
1
− QK (q)
q
(263)
is as small as possible, where ∆n = 2c1nln n . In other words, we seek to approximate the linear function q using q 2 , · · · , q K+2
on [0, ∆n ]. Fortunately, this task can be done with the help of the Chebyshev polynomial, and we summarize the result in the
following lemma.
31
Lemma 20. Let TK (x) = cos(K arccos x) be the degree-K Chebyshev polynomial on [−1, 1]. Then
p
K+2
− 2(K + 2)2 x/∆n
K T2(K+2) ( x/∆n ) − (−1)
QK (x) = (−1)
2
2(K + 2) (x/∆n )2
(264)
is a degree-K polynomial such that
sup
x2
x∈(0,∆n ]
1
∆n
.
− QK (x) =
x
(K + 2)2
(265)
On the other hand, using the Ditzian–Totik modulus of smoothness and Lemma 10 (or by Chebyshev’s alternating theorem
since {x2 , x3 , · · · } satisfies the Haar condition [49]), it is not hard to prove that no other degree-K polynomial can achieve
n
an approximation error of order o( ∆
K 2 ). Hence, the polynomial QK defined in Lemma 20 achieves the rate-optimal uniform
approximation error.
Motivated by Lemma 20, define Q̂K (x) be another polynomial such that EQ̂K (q̂) = QK (q) for nq̂ ∼ Poi(nq). Now in the
“non-smooth” regime, we choose K = c2 ln n, and use the following estimator:
1
T̂ns (p̂1 , q̂1 ) = p̂1 p̂1 −
· Q̂K (q̂1 ).
(266)
m
In summary, we have arrived at the following estimator construction.
Estimator Construction 4. Conduct three-fold sample splitting to obtain i.i.d. samples (p̂i,1 , p̂i,2 , p̂i,3 ) and (q̂i,1 , q̂i,2 , q̂i,3 ).
The estimator T̂ for χ2 -divergence χ2 (P, Q) is given by
S
X
c1 ln n
c1 ln n
) + T̃ns (p̂i,1 , q̂i,1 )1(q̂i,3 <
) −1
(267)
T̂s (p̂i,1 , q̂i,1 ; p̂i,2 , q̂i,2 )1(q̂i,3 ≥
T̂ =
n
n
i=1
where
T̃ns (x, y) , (T̂ns (x, y) ∧ 1) ∨ (−1)
(268)
and T̂s (x, y; x0 , y 0 ), T̂ns (x, y) are given by (262) and (266), respectively, and c1 , c2 > 0 are some suitably chosen universal
constants.
By construction, the previous estimator does not require the knowledge of S nor u(S), and is thus adaptive. For the analysis
of its performance, when q < c1 nln n is small, with the help of Lemma 20 we know that
ET̂ns (p̂, q̂) −
Moreover, for q ≥
c1 ln n
n
ET̂s (p̂1 , q̂1 ; p̂2 , q̂2 ) −
p2
1
(u(S))2
p2
= p2 Q(q) −
≤ (u(S))2 q 2 Q(q) −
.
.
q
q
q
n ln n
(269)
is large, applying Lemma 6 with r = 3 yields
p2
1
q
d4 (ξ −1 )
p2
(u(S))2
(u(S))2
= p2 ET̂ (3) (q̂1 , q̂2 ) −
. p2 · ( )2 sup
. 2 3 .
.
. (270)
4
2
q
q
n ξ∈[q/2,2q]
dξ
n q
n q
n ln n
Hence, the total bias of the previous estimator can be upper bounded as
|Bias(T̂ )| .
S
X
(u(S))2
i=1
n ln n
=
S(u(S))2
n ln n
(271)
which coincides with the term in Theorem 4. The variance can be dealt with analogously, and we omit the lengthy proofs.
VI. C ONCLUSIONS AND F UTURE W ORK
We proposed a general and detailed methodology for the construction of minimax rate-optimal estimators for low-dimensional
functionals of high-dimensional parameters, especially when the functional of interest is non-smooth in some part of its domain.
We elaborate on the insights of [1] which shows that the bias is the dominating term in the estimation of functionals and
approximation is the key for an efficient bias reduction, and find an interesting interplay between the functional itself and
the statistical model. Specifically, we show that the “smooth” and “non-smooth” regimes are determined by both the nonanalytic region of the underlying functional which is only related to the smoothness of the functional, and the confidence sets
given by concentration of measures which solely depend on the statistical model. Moreover, in the “non-smooth” regime, the
approximation region is determined by the confidence sets, while the approximation error is determined by the smoothness of
the functional in this region. Our general recipe is based on the interplay between these two factors, and successfully yields
the minimax rate-optimal estimators for various divergences including KL divergence, Hellinger distance and χ2 -divergence.
32
We have also explored the ideas behind the polynomial approximation and the plug-in approach in bias reduction. For
polynomial approximation, the uniform approximation error corresponds to the bias of the resulting estimator, and thus the
best approximating polynomials are usually used. We remark that it is a highly non-trivial task and remains open in general
to obtain and analyze the best polynomial approximation error for multivariate functionals, while for some special cases (e.g.,
general polytopes, balls and spheres) there are powerful tools from approximation theory. The plug-in approach corrects the
bias with the help of high-order Taylor expansions, which only works for the region where the functional is analytic. For bias
correction of the plug-in approach, in this paper we propose a general unbiased estimator of the Taylor series up to an arbitrary
order.
Following [1], this paper presents another second step towards a general theory of functional estimation. Despite our progress,
the interplay between the smoothness of the functional and the statistical model has yet to be completely revealed, and the
choice of the approximating polynomial in the “non-smooth” regime has thus far required functional-specific “tricks”. An
ambitious but worthy goal is to establish a general explanation of the effective sample size enlargement phenomenon in the
parametric case, and to find the counterpart in the estimation of non-smooth nonparametric functionals beyond the insights
provided by [33], [50].
A PPENDIX A
AUXILIARY L EMMAS
We first prove that the worst-case mean squared error of any estimator is infinity if we allow to choose any P which is
absolutely continuous with respect to Q.
Lemma 21. Let US = {(P, Q) : P, Q ∈ MS , P Q}. Then for any configuration (S, m, n) with S ≥ 2, we have
Rminimax (US ) = ∞.
(272)
The next lemma relates the minimax risk under the Poisson sampling model and that under the Multinomial model. We define
the minimax risk for Multinomial model with (m, n) observations with (P, Q) ∈ US,u(S) for estimating the KL divergence
D(P kQ) as
R(S, m, n, u(S)) , inf
sup
D̂ (P,Q)∈US,u(S)
EMultinomial D̂ − D(P kQ)
2
,
(273)
and the counterpart for the Poisson sampling model as
RP (S, m, n, u(S)) , inf
sup
D̂ (P,Q)∈US,u(S)
2
EPoisson D̂ − D(P kQ) .
(274)
Lemma 22. The minimax risks under the Poisson sampling model and the Multinomial model are related via the following
inequalities:
n
m n
m
(275)
RP (S, 2m, 2n, u(S)) − (ln u(S))2 exp(− ) + exp(− ) ≤ R(S, m, n, u(S)) ≤ 4RP (S, , , u(S)).
4
4
2 2
The next lemma gives the approximation properties of ln x.
Lemma 23. There exists a universal constant Cln > 0 such that for any 0 < a < b,
( b
b ≤ ean2 ,
2
b
En [ln x; [a, b]] ≤ Cln W ( 2 ) ≡ Cln · ean
an
ln( b 2 ) b > ean2 .
(276)
an
Lemma 24. For f (p, q) = p ln q and the region R given in (143), there exists a universal constant C0 only depending on c1
such that
u(S) ln n
1
2
.
(277)
ωR
(f, ) ≤ C0 ·
K
K 2n
The following lemma gives an upper bound for the second moment of the unbiased estimate of (p − q)j in Poisson model.
Lemma 25. [51] Suppose nX ∼ Poi(np), p ≥ 0, q ≥ 0. Then, the estimator
j
k−1
X
Y
j
h
j−k
gj,q (X) ,
(−q)
X−
k
n
k=0
h=0
(278)
33
is the unique unbiased estimator for (p − q)j , j ∈ N, and its second moment is given by
j 2
X
j
pk k!
2
E[gj,q (X) ] =
(p − q)2(j−k) k
n
k
k=0
p j
n(p − q)2
= j!
Lj −
assuming p > 0,
n
p
where Lm (x) stands for the Laguerre polynomial with order m, which is defined as:
m
X
m (−x)k
Lm (x) =
k!
k
(279)
(280)
(281)
k=0
If M ≥
n(p−q)2
p
∨ j, we have
E[gj,q (X)2 ] ≤
2M p
n
j
.
(282)
In order to bound the coefficients of best polynomial approximations, we need the following result by Qazi and Rahman
[52, Thm. E] on the maximal coefficients of polynomials on a finite interval.
Pn
Lemma 26. Let pn (x) = ν=0 aν xν be a polynomial of degree at most n such that |pn (x)| ≤ 1 for x ∈ [−1, 1]. Then,
|an−2µ | is bounded above by the modulus of the corresponding coefficient of Tn for µ = 0, 1, . . . , bn/2c, and |an−1−2µ | is
bounded above by the modulus of the corresponding coefficient of Tn−1 for µ = 0, 1, . . . , b(n − 1)/2c. Here Tn (x) is the n-th
Chebyshev polynomials of the first kind.
Moreover, it is shown in Cai and Low [27, Lemma 2] that all of the coefficients of Chebyshev polynomial T2m (x), m ∈ Z+
are upper bounded by 23m . Hence, we can obtain the following result when the approximation interval is not centered at zero.
Pn
Lemma 27. Let pn (x) = ν=0 aν xν be a polynomial of degree at most n such that |pn (x)| ≤ A for x ∈ [a, b], where
a + b 6= 0. Then
−ν
n
a+b
b+a
7n/2
|aν | ≤ 2
A
+1 ,
ν = 0, · · · , n.
(283)
2
b−a
The following lemma gives some tail bounds for Poisson and Binomial random variables.
Lemma 28. [53, Exercise 4.7] If X ∼ Poi(λ) or X ∼ B(n, nλ ), then for any δ > 0, we have
λ
2
eδ
P(X ≥ (1 + δ)λ) ≤
≤ e−δ λ/3 ∨ e−δλ/3
1+δ
(1 + δ)
λ
2
e−δ
≤ e−δ λ/2 .
P(X ≤ (1 − δ)λ) ≤
(1 − δ)1−δ
(284)
(285)
The following lemmas deal with the upper bound of the variance in different scenarios.
Lemma 29. For independent random variables X, Y with finite second moment, we have
Var(XY ) = (EY )2 Var(X) + (EX)2 Var(Y ) + Var(X)Var(Y ).
(286)
Lemma 30. [27, Lemma 4] Suppose 1(A) is an indicator random variable independent of X and Y , then
Var(X 1(A) + Y 1(Ac )) = Var(X)P(A) + Var(Y )P(Ac ) + (EX − EY )2 P(A)P(Ac ).
(287)
Lemma 31. [27, Lemma 5] For any two random variables X and Y ,
Var(X ∧ Y ) ≤ Var(X) + Var(Y ).
(288)
In particular, for any random variable X and any constant C,
Var(X ∧ C) ≤ Var(X).
(289)
34
A PPENDIX B
P ROOF OF M AIN L EMMAS
A. Proof of Lemma 1
First we give an upper bound of Var(gn (q̂)) for nq̂ ∼ Poi(nq). Note that gn (q) is continuously differentiable on [0, 1], we
have
Var(gn (q̂)) ≤ E(gn (q̂) − gn (q))2
(290)
q
q
= E(gn (q̂) − gn (q)) 1(q̂ ≥ ) + E(gn (q̂) − gn (q))2 1(q̂ < )
2
2
q
q
= E[gn0 (ξ1 )]2 (q̂ − q)2 1(q̂ ≥ ) + E[gn0 (ξ2 )]2 (q̂ − q)2 1(q̂ < )
2
2
q
≤ sup |gn0 (ξ1 )|2 · E(q̂ − q)2 + sup |gn0 (ξ2 )| · q 2 P(q̂ < )
2
ξ1 ∈[q/2,1]
ξ2 ∈[0,1]
4 q
≤ 2 · + n2 · q 2 e−nq/8
q n
3
4
n2
24
≤
+
·
nq
q
en
700
≤
nq
2
(291)
(292)
(293)
(294)
(295)
(296)
where in the previous steps we have used Lemma 28 and the fact
k
k
q k e−cnq ≤
ecn
(297)
for any q ∈ [0, 1].
Now we are ready to bound the bias. By independence and the triangle inequality, we have
(298)
|E[p̂gn (q̂)] − pgn (q)| = p|E[gn (q̂) − gn (q)]|
(299)
≤ u(S)q|E[gn (q̂) − gn (q)]|
(300)
≤ u(S) (|E[q̂gn (q̂)] − qgn (q)| + |E[(q̂ − q)gn (q̂)]|) .
We bound these two terms separately. For the first term, it can be obtained similar to [38] (via the second-order Ditzian–Totik
modulus of smoothness defined in [44]) that
5 ln 2
n
for any q ∈ [0, 1]. For the second term, first note that E[q̂] = q, we have
(301)
|E[q̂gn (q̂)] − qgn (q)| ≤
|E[(q̂ − q)gn (q̂)]| = |E[(q̂ − q)(gn (q̂) − Egn (q̂))]|.
(302)
|E[(q̂ − q)gn (q̂)]|2 ≤ E|q̂ − q|2 · E|gn (q̂) − Egn (q̂)|2
(303)
Hence, by the Cauchy-Schwartz inequality and the previous bound on Var(gn (q̂)), we have
= E|q̂ − q|2 · Var(gn (q̂))
q 700
≤ ·
n nq
700
= 2.
n
A combination of these two inequalities yields the bias bound
!
√
5 ln 2
700
30u(S)
|E[p̂gn (q̂)] − pgn (q)| ≤ u(S)
+
≤
n
n
n
(304)
(305)
(306)
(307)
ln 2
which together with |E[p̂ ln p̂] − p ln p| ≤ 5 m
in [38] yields the desired bias bound.
Next we bound the variance as follows:
Var(p̂(ln p̂ − gn (q̂))) ≤ E[p̂(ln p̂ − gn (q̂)) − p(ln p − gn (q))]2
2
2
2
≤ 3 E[p̂(ln p̂ − ln p)] + E[p̂(gn (q̂) − gn (q))] + E[(p̂ − p)(ln p − gn (q))]
≡ 3(A1 + A2 + A3 )
(308)
(309)
(310)
35
We bound A1 , A2 , A3 separately. To bound A1 , we further decompose A1 as
A1 = E[p̂(ln p̂ − ln p)]2 1(p̂ ≤ p) + E[p̂(ln p̂ − ln p)]2 1(p̂ > p)1(p ≥
where
1
1
) + E[p̂(ln p̂ − ln p)]2 1(p̂ > p)1(p < )
m
m
≡ B1 + B2 + B3
"
#
p̂ 2
p
B1 ≤ E sup | | · |p̂ − p|2 1(p̂ ≤ p) ≤ E|p̂ − p|2 =
ξ
m
ξ≥p̂
#
"
1
1
E[p̂2 (p̂ − p)2 ]
p̂ 2
2
1(p ≥ )
B2 ≤ E sup | | · |p̂ − p| 1(p̂ > p) 1(p ≥ ) ≤
2
m
p
m
ξ≥p ξ
p
p
1
5
1
6
+
=
+
1(p ≥ ) ≤ 2 + .
m3 p m2
m
m
m
m
(311)
(312)
(313)
(314)
(315)
Upper bounding B3 requires more delicate analysis. First note that by differentiation with respect to p, for any k ≥ 1 we have
sup
1
p≤ m
k 2
k 2
(mp)k k 2
(mp)k k 2
· 2 (ln
) ≤ sup
· 2 (2 + ln
)
k!
m
mp
k!
m
mp
p≤ 1
(316)
m
≤
(2 + ln k)2 k 2
· 2.
k!
m
(317)
Hence, expanding the expectation of B3 yields
B3 =
∞
X
e−mp
k=1
where the infinite sum converges to
∞
(mp)k k 2
k 2
1
1 X k 2 (2 + ln k)2
45
· 2 (ln
) 1(p < ) ≤ 2
< 2
k!
m
mp
m
m
k!
m
∞
X
k 2 (2 + ln k)2
k=1
Hence, A1 can be upper bounded as
(318)
k=1
k!
(319)
≈ 44.17 < 45.
A1 = B1 + B2 + B3 ≤
51
2p
+ .
m2
m
(320)
As for A2 , since we have proved that E(gn (q̂) − gn (q))2 ≤ 700
nq , by independence we have
p
700
700u(S)
1
700
= p2 +
·
≤
p+
.
A2 ≤ E(p̂2 ) ·
nq
m
nq
n
m
For A3 , it is clear that
p
2p
A3 = (ln p − gn (q))2 ≤
m
m
p
(ln )2 + (gn (q) − ln q)2
q
where we have used the fact that for p ≤ u(S)q,
2
≤
m
4q 4u(S)
p(ln u(S)) + 2 +
e
en
2
p
4q
4q
p(ln )2 ≤ p(ln u(S))2 ∨ 2 ≤ p(ln u(S))2 + 2
q
e
e
1
1
4u(S)
p(gn (q) − ln q)2 ≤ p(1 − ln(nq))2 1(q < ) ≤ u(S)q(1 − ln(nq))2 1(q < ) ≤
.
n
n
en
A combination of the upper bounds of A1 , A2 , A3 yields
51
2
4q 4u(S)
700u(S)
1
2
Var(p̂(ln p̂ − gn (q̂))) ≤ 2 +
p + p(ln u(S)) + 2 +
+
p+
.
m
m
e
en
n
m
The proof is complete.
(321)
(322)
(323)
(324)
(325)
36
B. Proof of Lemma 3
Braess and Sauer [39, Prop. 4] showed the following equalities for the Bernstein polynomials:
x0 (1 − x0 )
n
x0 (1 − x0 )(1 − 2x0 )
3
Bn [(x − x0 ) ](x0 ) =
.
n2
Bn [(x − x0 )2 ](x0 ) =
(326)
(327)
Hence, choosing x0 = x, we have
1 x(1 − x)
(1 − x)((n + 4)x − 2)
2 x(1 − x)(1 − 2x)
·
=
.
− 3·
x2
n
x
n2
n2 x2
Then the desired inequality is a direct result of Lemma 2.
(328)
Q3 (x) − Bn [Q3 ](x) =
C. Proof of Lemma 4
For the first statement, define the remainder term of the Taylor expansion as
R(θ̂n(1) ) ,
r
(1)
X
G(k) (θ̂n )
k=0
(1)
k!
(θ − θ̂n(1) )k − G(θ)
(329)
and denote by E the event that θ̂n ∈ V (θ). By the definition of reverse confidence sets, Pθ (E c ) ≤ δ, and
|Eθ H(θ̂n(1) , θ̂n(2) ) − G(θ)| ≤ |Eθ (H(θ̂n(1) , θ̂n(2) ) − G(θ))1(E)| + |Eθ (H(θ̂n(1) , θ̂n(2) ) − G(θ))1(E c )|
≤ |Eθ (Ĝ(r) (θ̂n(1) , θ̂n(2) ) − G(θ))1(E)| + δ ·
=
θ1 ,θ2 ∈Θ̂
!
θ1 ,θ2 ∈Θ̂
Eθ [|θ̂n − θ|r+1 1(E)]
≤
· sup |G(r+1) (θ̂)| + δ ·
(r + 1)!
θ̂∈V (θ)
(1)
(1)
(332)
!
|G(θ)| + sup |H(θ1 , θ2 )|
θ1 ,θ2 ∈Θ̂
(1)
Eθ |θ̂n − θ|r+1
≤
· sup |G(r+1) (θ̂)| + δ ·
(r + 1)!
θ̂∈V (θ)
(331)
|G(θ)| + sup |H(θ1 , θ2 )|
1(E)| + δ · |G(θ)| + sup |H(θ1 , θ2 )|
|Eθ R(θ̂n(1) )
(330)
!
!
|G(θ)| + sup |H(θ1 , θ2 )| .
θ1 ,θ2 ∈Θ̂
(333)
(334)
(2)
As for the variance of Hk (θ̂n , θ̂n ) with k ≥ 0, we first note by triangle inequality that
Varθ (Hk (θ̂n(1) , θ̂n(2) )) ≤ 2Varθ (Hk (θ̂n(1) , θ̂n(2) )1(E)) + 2Varθ (Hk (θ̂n(1) , θ̂n(2) )1(E c ))
≤
1(E)) + 2δ · sup |Hk (θ1 , θ2 )| .
2Varθ (Gk (θ̂n(1) , θ̂n(2) )
2
(335)
(336)
θ1 ,θ2 ∈Θ̂
Hence, it suffices to upper bound Varθ (Gk (θ̂n , θ̂n )1(E)). Note that Gk (θ̂n , θ̂n ) is a linear combination of terms of the
(1)
(1)
(2)
form G(k) (θ̂n )(θ̂n )k−j Sj (θ̂n ) with 0 ≤ j ≤ k, we employ the triangle inequality again to reduce the problem of bounding
the total variance to bounding the variance of individual terms. By independence and Lemma 29, it further suffices to upper
(1)
(1)
(1)
(1)
bound |Eθ G(k) (θ̂n )(θ̂n )k−j 1(E)| and Varθ (G(k) (θ̂n )(θ̂n )k−j 1(E)), respectively. In fact, defining Gk,j (θ) = θj G(k) (θ),
by Taylor expansion again we have
(1)
(2)
(1)
(2)
|Eθ Gk,j (θ̂n(1) )1(E)| ≤ |Eθ (Gk,j (θ̂n(1) ) − Gk,j (θ))1(E)| + |Gk,j (θ)|
1
= |Eθ (G0k,j (θ)(θ̂n(1) − θ) + G00k,j (ξ)(θ̂n(1) − θ)2 )1(E)| + |Gk,j (θ)|
2
!
1
00
(1)
2
c
0
≤ |Eθ Gk,j (ξ)(θ̂n − θ) 1(E)| + |Gk,j (θ)| + P(E ) · |Gk,j (θ)|(|θ| + sup |θ̂|)
2
θ̂∈Θ̂
!
(1)
2
Eθ (θ̂n − θ)
00
0
≤
· sup |Gk,j (θ̂)| + |Gk,j (θ)| + δ · |Gk,j (θ)|(|θ| + sup |θ̂|)
2
θ̂∈V (θ)
θ̂∈Θ̂
(337)
(338)
(339)
(340)
37
and
Varθ (Gk,j (θ̂n(1) )1(E)) ≤ Eθ (Gk,j (θ̂n(1) )1(E) − Gk,j (θ))2
≤
≤
≤
(341)
1
2Eθ ((Gk,j (θ̂n(1) ) − Gk,j (θ)) (E))2 + 2|Gk,j (θ)|2 · P(E c )
2Eθ (G0k,j (ξ)(θ̂n(1) − θ) (E))2 + 2δ · |Gk,j (θ)|2
2Eθ (θ̂n(1) − θ)2 · sup |G0k,j (θ̂)|2 + 2δ · |Gk,j (θ)|2
θ̂∈V (θ)
(342)
1
(343)
(344)
which establishes the desired variance bound.
(1) (2)
Finally it remains to bound the quantity |Eθ Hk (θ̂n , θ̂n )| for k ≥ 1. If k ≥ 2, as above, by triangle inequality, we conclude
that
|Eθ Hk (θ̂n(1) , θ̂n(2) )| ≤ |Eθ Hk (θ̂n(1) , θ̂n(2) )1(E)| + |Eθ Hk (θ̂n(1) , θ̂n(2) )1(E c )|
(345)
1(E)| + δ · sup |Hk (θ1 , θ2 )|
≤
|Eθ Gk (θ̂n(1) , θ̂n(2) )
=
1
|Eθ G(k) (θ̂n(1) )(θ − θ̂n(1) )k 1(E)| + δ · sup |Hk (θ1 , θ2 )|
k!
θ1 ,θ2 ∈Θ̂
(346)
θ1 ,θ2 ∈Θ̂
(347)
(1)
≤
Eθ |θ̂n − θ|k
· sup |G(k) (θ̂)| + δ · sup |Hk (θ1 , θ2 )|.
k!
θ1 ,θ2 ∈Θ̂
θ̂∈V (θ)
(348)
(1)
For k = 1, we further note that Eθ θ̂n = θ, and conduct order-one Taylor expansion to yield
|Eθ H1 (θ̂n(1) , θ̂n(2) )| ≤ |Eθ G0 (θ̂n(1) )(θ − θ̂n(1) )1(E)| + δ · sup |H1 (θ1 , θ2 )|
(349)
θ1 ,θ2 ∈Θ̂
= |Eθ (G0 (θ) + (θ̂n(1) − θ)G00 (ξ))(θ − θ̂n(1) )1(E)| + δ · sup |H1 (θ1 , θ2 )|
(350)
θ1 ,θ2 ∈Θ̂
≤
Eθ (θ̂n(1)
00
2
− θ) · sup |G (θ̂)| + δ ·
θ̂∈V (θ)
0
sup |H1 (θ1 , θ2 )| + |G (θ)|(|θ| + sup |θ̂|)
θ1 ,θ2 ∈Θ̂
θ̂∈Θ̂
!
(351)
as desired.
D. Proof of Lemma 5
Replacing n by m, we adopt the notations of V (p) and δ in (121). Denote by E the event p̂ ∈ V (p), we have
1
|EUH (p̂) + p ln p| = E −p̂ ln p̂ +
+ p ln p
2m
(p̂ − p)2
(p̂ − p)3
|E(p̂ − p)3 |
≤ E −p̂ ln p̂ − (1 + ln p)(p̂ − p) +
−
+
+
p
ln
p
2p
6p2
6p2
2
3
(p̂ − p)
(p̂ − p)
≤ E −p̂ ln p̂ − (1 + ln p)(p̂ − p) +
−
+ p ln p 1(E) +
2p
6p2
(p̂ − p)2
(p̂ − p)3
|E(p̂ − p)3 |
c
P(E ) · sup −p̂ ln p̂ − (1 + ln p)(p̂ − p) +
−
+
p
ln
p
+
2p
6p2
6p2
p̂∈[0,1]
4
1
1
1
1
E|p̂ − p|
sup |(−x ln x)(4) | + 2m−c1 /24 ·
− p ln p + 1 − ln p +
+ 2 +
≤
24
e
2p 6p
6pm2
x∈V (p)
≤
p + 3mp2
2
1
sup
+ 4m−c1 /24+2 +
24m3 x∈V (p) x3
6pm2
where we have used mp ≥ c1 ln m/2 ≥ 2. Since any x ∈ V (p) satisfies
r
1 c1 p ln m
p
x≥p−
≥
2
2m
2
(352)
(353)
(354)
(355)
(356)
(357)
38
by the previous inequality we have
2
1
p + 3mp2
·
+ 4m−c1 /24+2 +
24m3
(p/2)3
6pm2
2
1
4mp2
·
+ 4m−c1 /24+2 +
≤
24m3 (p/2)3
6pm2
3
≤ 2 + 4m−c1 /24+2
m p
6
≤
+ 4m−c1 /24+2
c1 m ln m
(358)
|EUH (p̂) + p ln p| ≤
(359)
(360)
(361)
as desired.
As for the variance, since the constant bias correcting term does not affect variance, applying Lemma 4 with k = 0 yields
p
(362)
Var(UH (p̂)) ≤ A0 2m−c1 /24 · 1 + (1 − ln(p/2))2 + 2m−c1 /24 · (p ln p)2
m
2
p(2 − ln p)
≤ A0
+ 4m−c1 /24 .
(363)
m
E. Proof of Lemma 6
The only non-trivial part in deducing the third inequality from Lemma 4 is to prove that for any k ∈ N and q ≥
q k2
Eq |q̂1 − q|k .
.
n
1
n,
(364)
In fact, since Eq exp(sq̂1 ) = exp(nq(es/n − 1)), we have Eq exp(s(q̂1 − q)) = exp(nq(es/n − 1 − s/n)). Hence, by comparing
the coefficient of sk at both sides of
k
∞
∞
∞
X
Eq (q̂1 − q)k k X (nq)k X 1 s j
s =
( )
(365)
k!
k!
j! n
j=2
k=0
k=0
yields that for even k, Eq (q̂1 − q) can be expressed as
k
k
k
Eq (q̂1 − q) =
2
X
ak,j
j=0
qj
nk−j
(366)
k/2
for some coefficients {ak,j }j=0 . Now for even k, the desired inequality follows from the assumption q ≥ n−1 . For odd k,
Cauchy-Schwartz inequality yields
k+1 1
1
q k2
q k−1
1
1
2 ·2+ 2 ·2
Eq |q̂1 − q|k ≤ Eq |q̂1 − q|k−1 2 Eq |q̂1 − q|k+1 2 .
=
.
(367)
n
n
Hence, the third inequality follows. The first inequality also follows from this fact, Lemma 4, δ ≤ 2n−c1 /24 and V (q) ⊂ [ 2q , 2q].
Now it remains to deduce the second inequality from Lemma 4. By (87), we know that Sj (q̂2 ) is a linear combination of
q̂2j−i
ni with constant coefficients and i = 0, 1, · · · , j. By the triangle inequality for the variance, it suffices to upper bound the
variance of each individual term
any k ≥ 0 and q ≥ n−1 ,
q̂2j−i
ni .
Using the same approach based on moment generating function, we conclude that for
Varq (q̂2k )
=
Eq (q̂22k )
As a result, for 0 ≤ i ≤ j,
Varq (
and thus Varq (Sj (q̂2 )) .
q 2j−1
n .
−
(Eq (q̂2k ))2
.
2k−1
X
j=0
qj
n2k−j
.
q 2k−1
.
n
q̂2j−i
q 2j−2i−1
q 2j−1
)
.
≤
ni
n2i+1
n
Finally it suffices to note that for q ≥
c1 ln n
2n
(369)
≥ n−1 ,
Eq [Sj2 (q̂2 )] = (Eq Sj (q̂2 ))2 + Varq (Sj (q̂2 )) . q 2j +
which completes the proof of the second inequality.
(368)
q 2j−1
. q 2j
n
(370)
39
F. Proof of Lemma 7
To invoke Lemma 6, we remark that g(q) = ln q, r = 3, and
h0 (q̂1 , q̂2 ) = ln q̂1 · 1(q̂1 6= 0)
q̂2 − q̂1
h1 (q̂1 , q̂2 ) =
· 1(q̂1 6= 0)
q̂
1
q̂2
(q̂2 − q̂1 )2
+
· 1(q̂1 6= 0)
h2 (q̂1 , q̂2 ) = −
2q̂12
2nq̂12
q̂2
q̂22
2q̂2
(q̂2 − q̂1 )3
+ 2 − 3 + 2 3 · 1(q̂1 6= 0).
h3 (q̂1 , q̂2 ) =
3q̂13
nq̂1
nq̂1
n q̂1
(371)
(372)
(373)
(374)
ln n
Noting that supξ∈[q/2,2q] |g (k) (ξ)| = |g (k) (q/2)| for any k ≥ 1 and q ≥ c12n
, we have
2
q
q −4
−c1 /24
2
3
|ET̂ (3) (q̂1 , q̂2 ) · 1(q̂1 6= 0) − ln q| ≤ B3
·
6(
)
+
n
·
−
ln
q
+
ln
n
+
n
+
n
+
2n
n2
2
96
−c1 /24+3
+
6n
≤ B3
n2 q 2
192
−c1 /24+3
≤ B3
+ 6n
c1 nq ln n
(375)
(376)
(377)
and thus by independence,
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 )] − p ln q| = p|ET̂ (3) (q̂1 , q̂2 ) · 1(q̂1 6= 0) − ln q|
192
−c1 /24+3
+ 6n
≤ pB3
c1 nq ln n
192B3 u(S)
≤
+ 6B3 pn−c1 /24+3 .
c1 n ln n
Moreover, for any 0 ≤ k ≤ 3,
(378)
(379)
(380)
q
24
· ( k−j+1 )2 + n−c1 /24 · q 2(j−k) (− ln q)
n q
j=0
k−1
2
X q 2(k−j)−1
q
q
+
q j−k + · 20( )j−k−2 + 4n−c1 /24 q j−k−1
n
n
2
j=0
!
2
576(k + 1) 2k
80
−c1 /24
3
2
≤ B3
+
1+
+n
n + (k + 1) ln n + 32kn
nq
nq
nq
12500
≤ B3
+ 100n−c1 /24+3
nq
Var(hk (q̂1 , q̂2 )) ≤ B3 n−c1 /24 · n3 +
k
X
q 2(k−j)
where in the last step we have used the fact that nq ≥ 2. For 1 ≤ k ≤ 3, we also have
q k∨2
q
|Ehk (q̂1 , q̂2 )| ≤ B3 ( ) 2 · 6( )−(k∨2) + n−c1 /24 · (q −1 + 2n3 )
2
n
≤ B3 48 + 3n−c1 /24+3 .
Now we are about to bound the bias and the variance for small p and large p, respectively. If p ≤
LH (x) = SK,H (x) ∧ 1 with SK,H (x) defined in [1]. It was shown in [1, Lemma 4] that
|ESK,H (p̂1 ) + p ln p| ≤
C
m ln m
(381)
(382)
(383)
(384)
(385)
2c1 ln m
,
m
first note that
(386)
4
(2c1 ln m)
(387)
m2
where we note that the constant c1 in [1] corresponds to the constant c1 /2 in our paper. Then applying Lemma 31, we have
2
ESK,H
(p̂1 ) ≤ m8c2 ln 2
2
Var(LH (p̂1 )) ≤ Var(SK,H (p̂1 )) ≤ ESK,H
(p̂1 ) ≤
(2c1 ln m)4
m2−8c2 ln 2
(388)
40
and thus
|ELH (p̂1 ) + p ln p| ≤ |ESK,H (p̂1 ) + p ln p| + E|SK,H (p̂1 )|1(SK,H (p̂1 ) ≥ 1)
(389)
2
(390)
≤ |ESK,H (p̂1 ) + p ln p| + E|SK,H (p̂1 )|
(2c1 ln m)4
C
+ 2−8c2 ln 2 .
≤
m ln m
m
Hence, the total bias can be upper bounded as
(391)
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )] + p ln(p/q)| ≤ |E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 )] − p ln q| + |ELH (p̂1 ) + p ln p|
192B3 u(S)
C
(2c1 ln m)4
≤
+ 6B3 pn−c1 /24+3 +
+ 2−8c2 ln 2 .
c1 n ln n
m ln m
m
(392)
(393)
As for the total variance, Lemma 29 can be used here to obtain
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 )) = E[p̂21 ] · Var(
3
X
k=0
hk (q̂1 , q̂2 )) + Var(p̂1 ) · (ET̂ (3) (q̂1 , q̂2 ) · 1(q̂1 6= 0))2
p
· 16B3
≤ p +
m
2
≤ 1600B3
12500
+ 100n−c1 /24+3
nq
p
+
m
B3
192
+ 6n−c1 /24+3
c1 nq ln n
(394)
2
p
p 125
−c1 /24+3
2
+n
+
B3 48 + 6n−c1 /24+3 − ln q .
p +
m
nq
m
− ln q
2
(395)
(396)
Now the desired variance bound follows from the triangle inequality Var(X + Y ) ≤ 2(Var(X) + Var(Y )).
ln m
If p ≥ c12m
, by Lemma 5 and the triangle inequality we have
|E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )] + p ln(p/q)| ≤ |E[T̃s (p̂1 , q̂1 ; p̂2 , q̂2 )] − p ln q| + |EUH (p̂1 ) + p ln p|
6
192B3 u(S)
≤
+ 6B3 pn−c1 /24+3 +
+ 4m−c1 /24+2
c1 n ln n
c1 m ln m
(397)
(398)
which is the desired bias bound. As for the variance, we have
3
T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 ) = −p̂1 ln
and the triangle inequality gives
X
p̂1
1
· 1(q̂1 6= 0) + p̂1 ·
hk (q̂1 , q̂2 ) +
q̂1
2m
(399)
k=1
!
3
X
p̂1
Var(T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )) ≤ 2Var p̂1 ln
hk (q̂1 , q̂2 )
· 1(q̂1 6= 0) + 2Var p̂1 ·
q̂1
(400)
k=1
≡ 2(B1 + B2 ).
Now we bound these two terms separately. For B1 , recall that Lemma 1 gives
51
2
4q 4u(S)
700u(S)
1
Var(p̂1 (ln p̂1 − gn (q̂1 ))) ≤ 2 +
p + p(ln u(S))2 + 2 +
+
p+
m
m
e
en
n
m
(401)
(402)
and the difference between these two quantities is upper bounded by
E[p̂1 (gn (q̂1 ) − ln q̂1 · 1(q̂1 6= 0))]2 ≤ (1 + ln n)2 · P(q̂1 = 0)
2
Hence, by triangle inequality again, we have
≤ (1 + ln n) · n
−c1 /2
.
B1 ≤ 2Var(p̂1 (ln p̂1 − gn (q̂1 ))) + 2E[p̂1 (gn (q̂1 ) − ln q̂1 · 1(q̂1 6= 0))]2
102
4
4q 4u(S)
1400u(S)
1
≤ 2 +
p + p(ln u(S))2 + 2 +
+
p+
+ 2(1 + ln n)2 · n−c1 /2 .
m
m
e
en
n
m
(403)
(404)
(405)
(406)
41
As for B2 , Lemma 29 is employed to obtain
B2 = E[p̂21 ] · Var(
3
X
k=1
hk (q̂1 , q̂2 )) + Var(p̂1 ) · (E
3
X
hk (q̂1 , q̂2 ))2
(407)
k=1
2
12500
p
p
· 9B3
3B3 48 + 3n−c1 /24+3
+ 100n−c1 /24+3 +
≤ p2 +
m
nq
m
2
125
2
p
9B
p
3
= 900B3 p2 +
48 + 3n−c1 /24+3 .
+ n−c1 /24+3 +
m
nq
m
(408)
(409)
The desired variance bound then follows from the upper bounds of B1 and B2 .
For the rest of the results, the only non-trivial observation is that when p ≤ 2c1mln m , we have
p(ln q)2 ≤ p(− ln p + ln u(S))2
2
(410)
2
(411)
≤ 2p(ln p) + 2p(ln u(S))
2
m
4c1 ln m
ln
+ 2p(ln u(S))2
≤
m
2c1 ln m
4c1 (ln m)3
+ 2p(ln u(S))2
≤
m
(412)
(413)
since p ≤ u(S)q and 2c1 ln m ≥ 8.
G. Proof of Lemma 8
For simplicity, we define
T s,I (p̂, q̂) , T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + UH (p̂1 )
(414)
T s,II (p̂, q̂) , T̃s (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )
(415)
then
T s (p̂, q̂) = T s,I (p̂, q̂)1(p̂3 >
c1 ln m
c1 ln m
) + T s,II (p̂, q̂)1(p̂3 ≤
).
m
m
(416)
By Lemma 7, the bias can be upper bounded as
(417)
|T s (p̂, q̂) + ln(p/q)| ≤ |ET s,I (p̂, q̂) + p ln(p/q)| + |ET s,II (p̂, q̂) + p ln(p/q)|
u(S)
1
+
.
.
m ln m n ln n
As for the variance, by Lemma 30 we have
Var(T s (p̂, q̂)) ≤ Var(T s,I (p̂, q̂)) + Var(T s,II (p̂, q̂)) + (ET s,I (p̂, q̂) − ET s,II (p̂, q̂))2
2
as desired.
≤ Var(T s,I (p̂, q̂)) + Var(T s,II (p̂, q̂)) + 2|ET s,I (p̂, q̂) + p ln(p/q)| + 2|ET s,II (p̂, q̂) + p ln(p/q)|
2
1
p(1 + ln u(S))2
q
u(S) pu(S)
1
u(S)
. 2− +
+
+
+
+
+
m
m
m
mn
n
m ln m n ln n
2
2
2
p(1 + ln u(S))
q
u(S) pu(S)
1
u(S)
1
. 2− +
+
+
+
+
+
m
m
m
mn
n
m ln m
n ln n
2
2
1
p(1 + ln u(S))
q
u(S) pu(S)
u(S)
. 2− +
+
+
+
+
m
m
m
mn
n
n ln n
(418)
(419)
2
(420)
(421)
(422)
(423)
42
H. Proof of Lemma 9
Denote by E the event that
2p
3
q
c1 p̂2 ln m
,
m
then by Lemma 28 we have
r
1 c1 p̂2 ln m
2p
c
) + P(p̂2 > p +
)
P(E ) ≤ P(p̂2 <
3 r
2
m
r
c1 p ln m
1 c1 p ln m
≤ P(p̂2 < p −
) + P(p̂2 > p +
)
18m
2
m
s
s
!2
!2
1
1
c
ln
m
c
ln
m
1
1
· mp + exp −
· mp
≤ exp −
2
18mp
3
4mp
≤ p̂2 ≤ p +
1
2
(424)
(425)
(426)
≤ 2e−c1 m/36 .
(427)
ln m
Note that conditioning on the event E, we have p̂2 ≥ c13m
, and
!
!
r
√
3
2p
1
1 c1 p̂2 ln m
p
2c1 ln n
1
0<
≤
p̂2 −
p̂2 ≤
p̂2 −
≤
≤q≤
21u(S)
u(S)
2
u(S)
2
m
u(S)
n
(428)
i.e., the approximation region contains q.
We first analyze the variance:
Var(T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )) ≤ E[T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )2 ]
(429)
1(E)] + E[T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) 1(E )]
≤ E[Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) 1(E)] + 1 · P(E c ).
= E[T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )
2
2
c
2
Note that conditioning on E, we have
Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) = p̂1 ·
PK
By construction, k=0 gK,k (p̂2 )xk is the best polynomial
"
4c1 ln n
1
2p
,
⊃R=
p̂2 −
21u(S)
n
u(S)
K
X
k=0
gK,k (p̂2 )
k−1
Y
l=0
(q̂2 −
l
).
n
approximation of ln q on R, where
!
#
r
1 c1 p̂2 ln m
4c1 ln n
2c1 ln n 4c1 ln n
,
⊃
,
2
m
n
n
n
(430)
(431)
(432)
(433)
where we have used (428) here. Since W (·) in Lemma 23 is an increasing function, conditioning on E the approximation
error can be upper bounded as
K
X
4c1 ln n/n
42c1 u(S)
k
gK,k (p̂2 )x − ln x ≤ Cln W
sup
= Cln W
·
.
(434)
(c2 ln n)2 · (2p/21u(S))
c22
pn ln n
x∈R
k=0
Note that W (x) ≤ 1 ∨ ln x, we conclude that for any x ∈ [ 2c1nln n , 4c1nln n ] ⊂ R, we have
K
X
k=0
K
X
n
2c1 ln n
k=0
42c1 u(S)
≤ Cln ln
·
∨ 1 + ln n ≡ A.
c22
pn ln n
gK,k (p̂2 )xk ≤
gK,k (p̂2 )xk − ln x + ln
Now we are about to apply Lemma 27 to bound each coefficient |gK,k (p̂2 )|. Lemma 27 yields
−k
−k
3c1 ln n
3c1 ln n
|gK,k (p̂2 )| ≤ 27K/2 A ·
(3K + 1) ≤ 211K/2 A ·
n
n
(435)
(436)
(437)
43
for any k = 0, 1, · · · , K = c2 ln n conditioning on E. Hence, by the triangle inequality,
!2
K
k−1
X
Y
l
E[Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )2 1(E)] = E(p̂21 ) · E
gK,k (p̂2 )
(q̂2 − ) 1(E)
n
k=0
l=0
#
−k !2 "k−1
K
X
Y
3c
ln
n
l
1
≤ E(p̂21 ) · (K + 1)
211K/2 A ·
E
(q̂2 − )2
n
n
k=0
l=0
"
#
−2k
K
k−1
Y
p X 3c1 ln n
l 2
11K
2
2
·
E
=2
(K + 1)A p +
(q̂2 − ) .
m
n
n
k=0
(438)
(439)
(440)
l=0
To evaluate the expectation, Lemma 25 with q = 0 is applied here to yield
"k−1
#
k
k
2k
Y
l 2
2q(k ∨ nq)
4c1 q ln n
3c1 ln n
E
(q̂2 − ) ≤
≤
≤
n
n
n
n
(441)
l=0
thus
p
p
≤ 211K+2 K 2 A2 p2 +
.
E[Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )2 1(E)] ≤ 211K (K + 1)2 A2 p2 +
m
m
By differentiation it is easy to show
2
42c1 u(S)
84c1 u(S)
84c1 u(S)
p ln
·
∨
1
≤ 2 ·
∨p≤ 2 ·
+p
c22
pn ln n
c2
n ln n
c2
n ln n
and note that p ≤ u(S)q ≤
(442)
(443)
2c1 u(S) ln n
,
n
we have
2
1
84c1 Cln
u(S)
2
2
E[Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )2 1(E)] ≤ 211K+3 K 2 p +
·
+
C
p
+
p(ln
n)
(444)
ln
m
c22
n ln n
2
2c1 u(S) ln n
1
84c1 Cln
2c1 u(S) ln n
u(S)
2
≤ 211K+3 K 2
+
+ (Cln
+ (ln n)2 )
(445)
·
2
n
m
c
n ln n
n
2
16c1 u(S) ln n 2c1 u(S) ln n
1
2
2
=
+
42Cln
+ (c2 ln n)2 (Cln
+ (ln n)2 )
(446)
n1−11c2 ln 2
n
m
which together with (431) is the variance bound.
Now we start to analyze the bias of T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ). By triangle inequality,
(447)
ET̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − p ln q ≤ E T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − p ln q 1(E) + E T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − p ln q 1(E c )
≤ |E (Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − p ln q) 1(E)| + E T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) 1(E)
+ 2m−c1 /36 (1 − p ln q)
(448)
+ 2m−c1 /36 (1 − p ln q)
(449)
≤ |E (Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − p ln q) 1(E)| + E |Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )1(|Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )| ≥ 1)1(E)|
−c1 /36
≡ A1 + A2 + 2m
(1 − p ln q).
(450)
Now we bound A1 and A2 separately. For A1 , since conditioning on E, the approximation region contains q, by (434) we
get
!
K
X
42c1 u(S)
k
A1 ≤ pE
gK,k (p̂2 )q − ln q 1(E) ≤ Cln pW
·
.
(451)
c22
pn ln n
k=0
As for A2 , since for any random variable X with finite second moment, we have
by (446) we get
E[|X|1(|X| ≥ 1)] ≤ E[|X|2 1(|X| ≥ 1)] ≤ E[X 2 ],
(452)
A2 ≤ E[Tns,I (p̂1 , q̂1 ; p̂2 , q̂2 )2 1(E)]
1
16c1 u(S) ln n 2c1 u(S) ln n
2
2
≤
+
42Cln
+ (c2 ln n)2 (Cln
+ (ln n)2 ) .
n1−11c2 ln 2
n
m
(453)
Now combining A1 and A2 completes the proof.
(454)
44
I. Proof of Lemma 11
First we bound the variance of Tns,II (p̂1 , q̂1 ). Recall that
X
Tns,II (p̂1 , q̂1 ) =
hK,k,l
k−1
Y
i=0
k,l≥0,k+l≤K
i
p̂1 −
m
We first bound the coefficients |hK,k,l |. It is straightforward to see that
sup
X
(x,y)∈R k,l≥0,k+l≤K
hK,k,l xk y l ≤ 2 sup |x ln y| ≤
(x,y)∈R
8c1 u(S) ln n
ln
n
l−1
Y
j
.
q̂1 −
n
j=0
n
4c1 ln n
≤
(455)
8c1 u(S)(ln n)2
≡ A.
n
(456)
We distinguish into two cases.
ln m
1) Case I: If mu(S)
≤ lnnn , we have
Hence, for any (x, y) ∈ R1 , we have
2c1 ln m
2c1 ln n 4c1 ln n
R ⊃ 0,
×
,
, R1 .
m
n
n
K
X
K−k
X
k=0
hK,k,l y l
l=0
!
(457)
xk ≤ A.
(458)
By Lemma 27, we conclude that for any y ∈ [ 2c1nln n , 4c1nln n ],
−k
K−k
X
c1 ln m
hK,k,l y l ≤ 27K/2+1 A ·
,
m
(459)
k = 0, 1, · · · , K.
l=0
Using Lemma 27 again, we have
−k
−l
3c1 ln n
c1 ln m
·
(3K−k + 1)
m
n
−k
−l
c1 ln m
3c1 ln n
≤ 29K+1 A
,
∀k, l ≥ 0, k + l ≤ K.
m
n
|hK,k,l | ≤ 27(K−k)/2 · 27K/2+1 A
2) Case II: If
ln m
mu(S)
>
ln n
n ,
define t , p/q, we have
2c1 ln n
R ⊃ (t, q) : 0 ≤ t ≤ u(S), 0 ≤ q ≤
, R2
n
and for any (t, q) ∈ R2 , we have
K
X
k=0
L−k
X
hK,k,l q k+l
l=0
!
X
tk =
k,l≥0,k+l≤K
By Lemma 27, we conclude that for any q ∈ [0, c1 nln n ],
L−k
X
hK,k,l q
k+l
l=0
By Lemma 27 again, we have
7K/2+1
≤2
A·
−k
(463)
(464)
.
−k
−(k+l)
u(S)
c1 ln n
|hK,k,l | ≤ 2
·2
A
·
2
n
−k
−l
c1 u(S) ln n
c1 ln n
= 27K+2 A
,
∀k, l ≥ 0, k + l ≤ n.
2n
n
7K/2+1
7K/2+1
Hence, combining these two cases yields
9K+1
|hK,k,l | ≤ 2
A
"
c1 u(S) ln n
2n
−k
+
c1 ln m
m
−k #
(461)
(462)
hK,k,l (qt)k q l ≤ A.
u(S)
2
(460)
c1 ln n
n
−l
.
(465)
(466)
(467)
45
Moreover, by Lemma 25, we have
#
"k−1
k
k
2k
Y
2p(k ∨ mp)
4c1 p(ln n + ln m)
3c1 (ln m + ln n)
i 2
(p̂1 − ) ≤
≤
≤
E
m
m
m
m
i=0
l
l
2l
l−1
Y
2q(l ∨ nq)
8c1 q ln n
6c1 ln n
j 2
E
≤
≤
.
(q̂1 − ) ≤
n
n
n
n
j=0
Now by the triangle inequality and previous inequalities, we have
"k−1
2 #
2
l−1
X
Y
Y
j
i
E[Tns,II (p̂1 , q̂1 )2 ] ≤ (K + 1)2
|hK,k,l |2 E
p̂1 −
E
q̂1 −
m
n
i=0
j=0
(468)
(469)
(470)
k,l≥0,0<k+l≤K
2k
2k !
n(ln m + ln n)
ln n
2k
+ 12
≤2
(K + 1) A
6
3
1+
ln m
mu(S) ln n
k,l≥0,0<k+l≤K
2K
2K !
ln n
n(ln m + ln n)
≤ 218K+3 122K (K + 1)4 A2 1 + 1 +
+
ln m
mu(S) ln n
2K
2K !
ln n
n(ln m + ln n)
26K+7 4 2
≤2
K A 1+ 1+
+
.
ln m
mu(S) ln n
18K+3
2
X
2
2l
2k
(471)
(472)
(473)
Hence, by Lemma 31 we get
Var(T̃ns,II (p̂1 , q̂1 )) ≤ Var(Tns,II (p̂1 , q̂1 )) ≤ E[Tns,II (p̂1 , q̂1 )2 ]
2K
2K !
n(ln m + ln n)
ln n
26K+7 4 2
+
≤2
K A 1+ 1+
ln m
mu(S) ln n
(474)
(475)
which is the desired variance bound.
As for the bias, Lemma 10 and Lemma 24 give
|ETns,II (p̂1 , q̂1 ) − p ln q| ≤ EK [p ln q; R] + |hK,0,0 |
2
≤ 2EK [p ln q; R] ≤ 2M ωR
(p ln q,
(476)
1
2M C0 u(S)
)≤
·
K
c22
n ln n
(477)
where |hK,0,0 | ≤ EK [p ln q; R] is obtained by setting (p, q) = (0, 0) ∈ R in
sup
X
(x,y)∈R k,l≥0,k+l≤K
hK,k,l xk y l − x ln y ≤ EK [p ln q; R].
(478)
Hence, by triangle inequality and (452), we get
|E[T̃ns,II (p̂1 , q̂1 )] − p ln q| ≤ |ETns,II (p̂1 , q̂1 ) − p ln q| + E|Tns,II (p̂1 , q̂1 )1(|Tns,II (p̂1 , q̂1 )| ≥ 1)|
2M C0 u(S)
≤
·
+ E[Tns,II (p̂1 , q̂1 )2 ]
c22
n ln n
2K
2K !
ln
n
n(ln
m
+
ln
n)
2M C0 u(S)
·
+ 226K+7 K 4 A2 1 + 1 +
+
≤
c22
n ln n
ln m
mu(S) ln n
(479)
(480)
(481)
as desired.
J. Proof of Lemma 12
We distinguish into three cases based on different values of p. For simplicity, we define
T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) , T̃ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + LH (p̂1 )
(482)
T ns,II (p̂1 , q̂1 ) , T̃ns,II (p̂1 , q̂1 ) + LH (p̂1 ).
(483)
46
1) Case I: We first consider the case where p ≤
c1 ln m
2m .
By the triangle inequality, the bias can be decomposed into
|Bias(Tns (p̂, q̂))| ≤ |ET ns,II (p̂1 , q̂1 ) + p ln(p/q)| + (E|T ns,II (p̂1 , q̂1 )| + E|T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )|)P(p̂3 ≥
c1 ln m
)
m
(484)
c1 ln m
2
≤ |ET ns,II (p̂1 , q̂1 ) + p ln(p/q)| + 4 · (e/4)
(485)
2
c1
1
u(S)
(u(S))
.
+
+ 2−c2 B + m− 2 ln(4/e)
(486)
m ln m n ln n
n
2
u(S)
(u(S))
1
(487)
+
+
.
m ln m n ln n
n2−
where we have used Lemma 11 and Lemma 28 here. Similarly, by Lemma 30, the variance can be upper bounded as
Var(Tns (p̂, q̂)) ≤ Var(T ns,II (p̂1 , q̂1 )) + Var(T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ))P(p̂3 ≥
c1 ln m
≤ Var(T ns,II (p̂1 , q̂1 )) + 22 · (e/4) 2 + 42 · (e/4)
c1
(u(S))2
1
. 2−c2 B + 2−c2 B + m− 2 ln(4/e)
m
n
(u(S))2
1
.
. 2− +
m
n2−
2) Case II: Next we consider the case where
bounded as
c1 ln m
2m
<p<
c1 ln m
c1 ln m
) + |ET ns,II − ET ns,I |2 P(p̂3 ≥
) (488)
m
m
c1 ln m
2
2c1 ln m
.
m
(489)
(490)
(491)
By Lemma 9 and Lemma 11, the bias can be upper
|Bias(Tns (p̂, q̂))| ≤ |ET ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + p ln(p/q)| + |ET ns,II (p̂1 , q̂1 ) + p ln(p/q)|
u(S)
u(S)(ln n)5 u(S) ln n
1
u(S)
(u(S))2
1
1
+ pW (
) + 1−11c2 ln 2
+
+
+ 2−c2 B
.
+
m ln m
pn ln n
n
n
m
m ln m n ln n
n
u(S)
u(S)
(u(S))2
u(S)
1
+ pW (
)+
+
+
.
.
m ln m
pn ln n
n ln n
n2−
mn1−
(492)
(493)
(494)
The variance is obtained by Lemma 30 as follows:
Var(Tns (p̂, q̂)) ≤ Var(T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )) + Var(T ns,II (p̂1 , q̂1 )) + (ET ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) − ET ns,II (p̂1 , q̂1 ))2
u(S)
u(S)(ln n)3
p
1
u(S)(ln n)5 u(S) ln n
1
(ln m)4
.
+
+ 3 + 1−11c2 ln 2
+
+ 2−c2 B + 2−c2 B
mn
m m
n
n
m
m
n
2
1
u(S)
u(S)
(u(S))2
u(S)
+
+ pW (
)+
+
+
m ln m
pn ln n
n ln n
n2−
mn1−
4
2
(u(S))
(u(S))
u(S)
u(S) 2
1
+ [pW (
)]
. 2− + 4−2 + 2−2 +
m
n
n
mn1−
pn ln n
(u(S))2
u(S)
u(S) 2
1
)]
+ [pW (
. 2− + 2−2 +
m
n
mn1−
pn ln n
where in the last step we have used that n & u(S).
3) Case III: Finally we consider the case where p ≥
2c1 ln m
.
m
(495)
(496)
(497)
(498)
By Lemma 9 and Lemma 28,
|Bias(Tns (p̂, q̂))| ≤ |ET ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + p ln(p/q)| + (E|T ns,II (p̂1 , q̂1 )| + E|T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 )|)P(p̂3 ≤
c ln m
− 12
≤ |ET ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + p ln(p/q)| + 4 · e
1
u(S)
u(S)(ln n)5 u(S) ln n
1
.
+ pW (
) + 1−11c2 ln 2
+
+ m−c1 /2
m ln m
pn ln n
n
n
m
1
u(S)
(u(S))2
u(S)
.
+ pW (
)+
+
m ln m
pn ln n
n2−
mn1−
c1 ln m
)
m
(499)
(500)
(501)
(502)
47
and the variance is given by Lemma 30 that
c1 ln m
c1 ln m
) + |ET ns,II − ET ns,I |2 P(p̂3 ≤
) (503)
m
m
c1 ln m
c1 ln m
(504)
≤ Var(T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + 22 · e− 2 + 42 · e− 2
u(S)(ln n)3
u(S)(ln n)5 u(S) ln n
p
1
1
.
+ 1−11c2 ln 2
+
+
+
+ m−c1 /2
(505)
mn
m m3
n
n
m
(u(S))2
u(S)
1
+
.
(506)
. 2− +
2−
m
n
mn1−
A combination of these three cases completes the proof.
Var(Tns (p̂, q̂)) ≤ Var(T ns,I (p̂1 , q̂1 ; p̂2 , q̂2 ) + Var(T ns,II (p̂1 , q̂1 ))P(p̂3 ≤
K. Proof of Lemma 13
As in the proof of Lemma 12, we also distinguish into three cases.
ln n
1) Case I: We first consider the case where q ≤ c12n
. By Lemma 12 and Lemma 28,
|Bias(ξ(p̂, q̂))| ≤ |ETns (p̂, q̂) + p ln(p/q)| + (E|Tns (p̂, q̂)| + E|T s (p̂, q̂)|)P(q̂3 ≥
c1 ln n
)
n
1
≤ |ETns (p̂, q̂) + p ln(p/q)| + (2 + ln n + 5n3 + 1 +
+ 1) · (e/4)
2n
1
2
ln m
+ u(S) + (u(S))
p ≤ c12m
,
n2−
m ln m n ln n
2
u(S)
u(S)
(u(S))
u(S)
c
ln
m
1
1
. m ln m + pW ( pn ln n ) + n ln n + n2− + mn1−
2m < p <
2
1
u(S)
(u(S))
u(S)
p ≥ 2c1mln m .
m ln m + pW ( pn ln n ) + n2− + mn1−
(507)
c1 ln n
2
(508)
2c1 ln m
,
m
(509)
where we have used the fact that
|T s (p̂, q̂)| ≤ |T̃s (p̂1 , q̂1 ; p̂2 , q̂2 )| + |UH (p̂1 )| + |LH (p̂1 )| ≤ 5n3 + ln n + 1 +
1
+1
2n
(510)
here. Similarly, by Lemma 30, the variance can be upper bounded as
c1 ln n
)
Var(ξ(p̂, q̂)) ≤ Var(Tns (p̂, q̂)) + Var(T s (p̂, q̂)) + (ETns (p̂, q̂) − ET s (p̂, q̂))2 P(q̂3 ≥
n
p(1 + ln u(S))2
q
u(S) pu(S)
u(S) 2
1
+
+
+
+
+(
)
. Var(Tns (p̂, q̂)) +
2−
m
m
m
mn
n
n ln n
c1 ln n
1
+22 + (5n3 + ln n + 1 +
+ 1)2 · (e/4) 2
2n
(u(S))2
ln m
1
p ≤ c12m
,
m2− + n2−
2
(u(S))
u(S)
u(S)
c
ln
m
2c1 ln m
1
2
1
. m2− + n2−2 + mn1− + [pW ( pn ln n )]
,
2m < p <
m
2
1
(u(S))
u(S)
p ≥ 2c1mln m .
m2− + n2− + mn1−
2) Case II: Next we consider the case where
c1 ln n
2n
<q<
2c1 ln n
.
n
(511)
(512)
(513)
By Lemma 8 and Lemma 12,
|Bias(ξ(p̂, q̂))| ≤ |ET s (p̂, q̂) + p ln(p/q)| + |ETns (p̂, q̂) + p ln(p/q)|
1
2
+ nu(S)
+ (u(S))
2−
m
ln
m
ln
n
n
1
u(S)
u(S)
u(S)
(u(S))2
u(S)
1
.
+
+ m ln
m + pW ( pn ln n ) + n ln n + n2− + mn1−
m ln m n ln n
1
u(S)
(u(S))2
u(S)
m ln m + pW ( pn ln n ) + n2− + mn1−
2
ln m
1 + u(S) + (u(S))
p ≤ c12m
,
m ln m
n ln n
n2−
.
2
u(S)
ln m
1 + pW ( u(S) ) + u(S) + (u(S))
p ≥ c12m
.
m ln m
pn ln n
n ln n
n2− + mn1−
(514)
p≤
c1 ln m
2m ,
c1 ln m
2m
p≥
<p<
2c1 ln m
,
m
(515)
2c1 ln m
.
m
(516)
48
As for the variance, Lemma 30 is used here to yield
Var(ξ(p̂, q̂)) ≤ Var(Tns (p̂, q̂)) + Var(T s (p̂, q̂)) + |ETns (p̂, q̂) − ET s (p̂, q̂)|2
p(1 + ln u(S))2
1
q
u(S) pu(S)
u(S) 2
. Var(Tns (p̂, q̂)) + 2− +
+
+
+
+(
)
m
m
m
mn
n
n ln n
2
2
+ |ETns (p̂, q̂) + p ln(p/q)| + |ET s (p̂1 , q̂1 ; p̂2 , q̂2 ) + p ln(p/q)|
( u(S)
1
p(1 + ln u(S))2
q
pu(S) (u(S))2
mn
. 2+
+
+
+ 2−2 +
u(S)
u(S)
m
m
m
n
n
)]2
1−2 + [pW (
mn
where we have used the fact that n & u(S) here.
3) Case III: Finally we come to the case where q ≥
2c1 ln n
.
n
pn ln n
(517)
(518)
p≤
p>
c1 ln m
2m ,
c1 ln m
2m .
By Lemma 8 and Lemma 28,
|Bias(ξ(p̂, q̂))| ≤ |ET s (p̂, q̂) + p ln(p/q)| + (E|Tns (p̂, q̂)| + E|T s (p̂, q̂)|)P(q̂3 ≤
c1 ln n
)
n
c1 ln n
1
u(S)
1
+
+ (2 + 5n3 + ln n + 1 +
+ 1)e− 2n
m ln m n ln n
2n
1
u(S)
.
+
.
m ln m n ln n
The variance bound is obtained in a similar fashion via Lemma 30:
c1 ln n
)
Var(ξ(p̂, q̂)) ≤ Var(T s (p̂, q̂)) + Var(Tns (p̂, q̂)) + (ETns (p̂, q̂) − ET s (p̂, q̂))2 P(q̂3 ≤
n
2
1
p(1 + ln u(S))
q
u(S) pu(S)
u(S) 2
. 2− +
+
+
+
+(
)
m
m
m
mn
n
n ln n
"
2 #
c1 ln n
1
+ 22 + 22 + 5n3 + ln n + 1 +
+1
· e− 2n
2n
.
p(1 + ln u(S))2
q
u(S) pu(S)
u(S) 2
+
+
+
+(
) .
m
m
mn
n
n ln n
Combining these three cases yields the desired result.
.
1
m2−
(519)
+
(520)
(521)
(522)
(523)
(524)
(525)
L. Proof of Lemma 14
As before, we first analyze the variance. By [1, Lemma 20], we know that there exists some constant C > 0 such that for
any x ∈ [0, 2c1nln n ],
1−k
K+1
X
C
2c1 ln n
gK,k
xk − x ln x ≤
.
(526)
n
n ln n
k=1
2c1 ln n
],
n
By triangle inequality, for any x ∈ [0,
1−k
1−k
K+1
K+1
X
X
2c1 ln n
2c1 ln n
k
gK,k
x ≤
gK,k
xk − x ln x + |x ln x|
n
n
k=1
(527)
k=1
2c1 ln n
n
C
+
· ln
n ln n
n
2c1 ln n
C
2c1 (ln n)2
≤
+
≡ A.
n ln n
n
As a result, by Lemma 27, for any k = 1, · · · , K + 1, we have
1−k
−k
2c1 ln n
c1 ln n
|gK,k |
≤ 27K/2+1 A ·
.
n
n
≤
(528)
(529)
(530)
49
Hence, by triangle inequality again and Lemma 25, we have
Var(Tns (p̂, q̂)) ≤ E[Tns (p̂, q̂)2 ]
and by Lemma 31, we have
(531)
"k−1
#
K
Y
X
l
2c
ln
n
1
· E[p̂2 ]E
(q̂ − )2
≤ (K + 1)
|gK,k+1 |2
n
n
l=0
k=0
−2(k+1)
K
4qc ln n k
X
c1 ln n
p
1
≤ 27K+2 (K + 1)A2
·
+ p2
n
m
n
k=0
K
−2
X
p
c1 ln n
·
≤ 211K+2 (K + 1)A2
+ p2
n
m
k=0
(C + 2c1 (ln n)3 )2
2u(S)
4(u(S))2
≤ 211K+2 (K + 1)2
·
+
(ln n)2
c1 mn ln n
n2
4(u(S))2
2u(S)
≤ 211K+4 c22 (C + 2c1 (ln n)3 )2 ·
+
c1 mn ln n
n2
Var(T̃ns (p̂, q̂)) ≤ Var(Tns (p̂, q̂)) ≤
−2k
211K+4 c22 (C
3 2
+ 2c1 (ln n) ) ·
As for the bias, by construction we have
2u(S)
4(u(S))2
+
c1 mn ln n
n2
(532)
(533)
(534)
(535)
(536)
.
(537)
−k
2c1 ln n
q k − ln q
n
k=0
1−k
K+1
X
2c1 ln n
gK,k
≤ u(S)q
q k−1 − ln q
n
k=1
1−k
K+1
X
2c1 ln n
gK,k
= u(S)
q k − q ln q
n
(539)
≤
(541)
|ETns (p̂, q̂) − p ln q| = p
K
X
gK,k+1
(538)
(540)
k=1
Cu(S)
.
n ln n
Hence, by triangle inequality and (452), we get
as desired.
|Bias(T̃ns (p̂, q̂))| ≤ |ETns (p̂, q̂) − p ln q| + E|Tns (p̂, q̂)1(|Tns (p̂, q̂)| ≥ 1)|
Cu(S)
+ E[Tns (p̂1 , q̂1 )2 ]
≤
n ln n
Cu(S)
2u(S)
4(u(S))2
≤
+ 211K+4 c22 (C + 2c1 (ln n)3 )2 ·
+
n ln n
c1 mn ln n
n2
(542)
(543)
(544)
M. Proof of Lemma 19
Fix δ > 0. Let D̂(X, Y) be a near-minimax estimator of D(P kQ) under the multinomial model with an upper bound
(1 + )u(S) on the likelihood ratio. Note that the estimator D̂ obtains the sample sizes m, n from observations. By definition,
we have
2
sup
EMultinomial D̂ − D(P kQ) ≤ R(S, m, n, (1 + )u(S)) + δ.
(545)
(P,Q)∈US,u(S)
Now given (P, Q) ∈ US,u(S) (), let X = [X1 , · · · , XS ]T and Y = [Y1 , · · · , YS ]T with (Xi , Yi ) ∼ Poi(mpi ) × Poi(nqi ). Write
PS
PS
m0 = i=1 Xi and n0 = i=1 Yi , we use the estimator D̂(X, Y) to estimate D(P kQ).
Note that conditioned on m0 = M , we have X ∼ Multinomial(M, PSP p ), and similarly for Y. Moreover, (P, Q) ∈
i=1
i
50
US,u(S) () implies that (P, PSQ
i=1
qi
) ∈ US,(1+)u(S) by construction. By the triangle inequality we have
2
1
1
RP (S, m, n, u(S), ) ≤ E(P,Q) D̂ − D(P kQ)
2
2
Q
≤ E(P,Q)
≤
≤
=
∞
X
k,l=0
∞
X
k,l=0
(546)
D̂ − D P k PS
i=1 qi
E(P,Q)
!!2
Q
D̂ − D P k PS
i=1 qi
+
!
Q
D P k PS
0
0
!
i=1 qi
!2
m = k, n = l
− D(P kQ)
!2
0
0
P(m = k)P(n = l) + (ln
(547)
S
X
qi )2 (548)
i=1
(R(S, k, l, (1 + )u(S)) + δ) P(m0 = k)P(n0 = l) + 22
X
k≥ m
2 ,l≥
(549)
R(S, k, l, (1 + )u(S))P(m0 = k)P(n0 = l)
(1−)n
2
m
(1 − )n
) + P(n0 <
)) + δ + 22
(550)
2
2
(1 − )n
m (1 − )n
m
, (1 + )u(S)) + (ln[(1 + )u(S)])2 exp(− ) + exp(−
) + δ + 22
≤ R(S, ,
2
2
8
8
(551)
+ (ln[(1 + )u(S)])2 (P(m0 <
where we have used Lemma 28. Then the result follows from the arbitrariness of δ.
N. Proof of Lemma 20
The properties of Chebyshev polynomials were well studied in [54]. In particular, the Chebyshev polynomial T2(K+2) (x)
is an even function and takes the form
T2(K+2) (x) = S2K (x)x4 − 2(−1)K (K + 2)2 x2 + (−1)K
(552)
for some even polynomial S2K (x) of degree 2K. Since |T2(K+2) (x)| ≤ 1 for any x ∈ [−1, 1], by triangle inequality
(−1)K
|T2(K+2) (x)|
1
1
S2K (x)x4
− x2 ≤
+
≤
.
2(K + 2)2
2(K + 2)2
2(K + 2)2
(K + 2)2
(553)
Now the desired result follows from the variable substitution y = ∆n x2 ∈ [0, ∆n ].
A PPENDIX C
P ROOF OF AUXILIARY L EMMAS
A. Proof of Lemma 21
Fix m, n and an arbitrary (possibly randomized) estimator D̂. Denote by µ the (possibly randomized) decision made by D̂
conditioning on the event E where m first symbols and no other symbols from P , and n second symbols and no other symbols
from Q are observed. Note that µ is a probability measure on R. Choose P = (1, 0, · · · , 0) and Q = (δ, 1 − δ, 0, · · · , 0) with
δ ∈ (0, 1) to be specified in the sequel, then P Q. Hence, with probability at least (1 − δ)n , the event E holds, and thus
Z
2
n
sup E(P,Q) D̂ − D(P kQ) ≥ (1 − δ) · (a − ln(1/δ))2 µ(da).
(554)
(P,Q)∈US
R
As a result, denote by a1/2 = inf{a ∈ R : µ((−∞, a]) ≥ 12 } a median of µ, choosing δ = 12 ∧ exp(−a1/2 − M ) for any
M > 0 yields
Z
2
1
M2
sup E(P,Q) D̂ − D(P kQ) ≥ (1 − )n ·
(a − ln(1/δ))2 µ(da) ≥ n+1 .
(555)
2
2
(P,Q)∈US
(−∞,a1/2 ]
Letting M → ∞ yields the desired result.
B. Proof of Lemma 22
Similar to the proof of [1, Lemma 16], we can show that
RP (S, m, n, u(S), π) =
∞
X
k,l=0
R(S, k, l, u(S), π)P(Poi(m) = k)P(Poi(n) = l)
(556)
51
where R(·, ·, ·, ·, π) and RP (·, ·, ·, ·, π) represent the Bayes error given prior π under the Multinomial model and the Poisson
sampling model, respectively. On one hand, we have
X
RP (S, m, n, u(S), π) ≥
R(S, k, l, u(S), π)P(Poi(m) = k)P(Poi(n) = l)
(557)
0≤k≤2m,0≤l≤2n
≥ R(S, 2m, 2n, u(S), π)P(Poi(m) ≤ 2m)P(Poi(n) ≤ 2n)
1
≥ R(S, 2m, 2n, u(S), π)
4
(558)
(559)
where we have used the Markov inequality to get P(Poi(m) ≤ 2m) ≥ 21 and P(Poi(n) ≤ 2n) ≥ 12 . On the other hand, note
that D(P kQ) ≤ ln u(S) whenever (P, Q) ∈ US,u(S) , by the Poisson tail bound in Lemma 28 we also have
X
RP (S, m, n, u(S), π) ≤
R(S, k, l, u(S), π)P(Poi(m) = k)P(Poi(n) = l)
k>m/2,l>n/2
≤
m
n
+ (ln u(S))2 P(Poi(m) ≤ ) + P(Poi(n) ≤ )
(560)
2
2
X
m n
m
n
R(S, , , u(S), π)P(Poi(m) = k)P(Poi(n) = l) + (ln u(S))2 exp(− ) + exp(− )
2 2
8
8
k>m/2,l>n/2
m
n
m n
≤ R(S, , , u(S), π) + (ln u(S))2 exp(− ) + exp(− ) .
2 2
8
8
By the minimax theorem [55], taking supremum over all priors π yields the desired result.
(561)
(562)
C. Proof of Lemma 23
We apply the general approximation theory on convex polytopes to our one-dimensional case where [a, b] is an interval.
Note that by polynomial scaling,
a
En [ln x; [a, b]] = En [ln ((b − a)x + a) ; [0, 1]] = En [ln(x +
); [0, 1]]
(563)
b−a
a
it suffices to consider the function h(x) = ln(x + ∆) defined on [0, 1], where ∆ = b−a
> 0. In this case, the second-order
Ditzian–Totik modulus of smoothness in (146) is reduced to [44]
u+v
u+v
2
2
: u, v ∈ [0, 1], |u − v| ≤ 2tϕ(
)
(564)
ω[0,1] (f, t) = ωϕ (f, t) , sup f (u) + f (v) − 2f
2
2
p
where ϕ(x) = x(1 − x). For the evaluation of ωϕ2 (h, t) for t ∈ [0, 1], we write u = r + s, v = r − s with u, v ∈ [0, 1] and
0 ≤ s ≤ tϕ(r) in the definition. Then for ∆ > t2 , by Taylor expansion we have
k
∞
∞
∞
X
X
|h(2k) (r)| 2k X 1
s2k
1 t2 r(1 − r)
|h(r + s) + h(r − s) − 2h(r)| ≤ 2
s =
·
≤
.
(565)
(2k)!
k (r + ∆)2k
k (r + ∆)2
k=1
By differentiation, it is easy to show that the maximum of
1
is 4∆(1+∆)
. Hence, by (565) we have
|h(r + s) + h(r − s) − 2h(r)| ≤
2
∞
X
k=1
t2
4∆(∆ + 1)
k=1
r(1−r)
(r+∆)2
k
k=1
is attained at r =
∞
X
t2
≤
4∆(∆ + 1)
k=1
∆
2∆+1 ,
and the corresponding maximum
k−1
1
t2
t2
=
≤
4
3∆(∆ + 1)
3∆
(566)
t
i.e., we conclude that ωϕ2 (h, t) ≤ 3∆
when ∆ > t2 .
2
For ∆ ≤ t , the concavity of ln(·) yields that the maximum-achieving pair (u, v) must satisfy one of the following: (1)
u = 1; (2) v = 0; (3) s = tϕ(r).
t2
1
We start with the case where s = tϕ(r), then 1+t
2 ≤ r ≤ 1+t2 . In this case (565) still holds, but now the maximum of
2
r(1−r)
t
(r+∆)2 is attained at r = 1+t2 , and the corresponding inequality becomes
k X
k
∞
∞
X
1
t4
1
t4
≤
k (t2 + (1 + t2 )∆)2
k (t2 + ∆)2
k=1
k=1
2
t4
(t2 + ∆)2
2t + ∆
= − ln 1 − 2
= ln
≤ ln
.
(t + ∆)2
∆(∆ + 2t2 )
∆
|h(r + s) + h(r − s) − 2h(r)| ≤
Note that this inequality only requires r ≥
t2
1+t2 ,
hence it also holds for the case where u = 1.
(567)
(568)
52
Now we are only left with the case where v = 0, then s = r and r ≤
t2
1+t2 .
As a result,
|h(r + s) + h(r − s) − 2h(r)| = 2 ln(r + ∆) − ln(2r + ∆) − ln(∆)
t2
r2
r
≤ ln 1 +
= ln 1 +
≤ ln 1 +
.
∆(2r + ∆)
2∆
2∆
In summary, for ∆ ≤ t2 we can obtain that
ωϕ2 (h, t)
2t2
≤ ln 1 +
∆
(569)
(570)
(571)
and Lemma 23 follows from the previous upper bounds on ωϕ2 (h, n−1 ) and Lemma 10.
D. Proof of Lemma 24
It suffices to prove the claim for ωT2 (f, 1/K), where T is the following triangle containing R:
2c1 ln n
T , (x, y) : 0 ≤ y ≤
, 0 ≤ x ≤ u(S)y .
n
(572)
Denote by E1 , E2 , E3 three edges of this triangle (excluding endpoints):
2c1 ln n
2c1 u(S) ln n
,y =
}
(573)
n
n
2c1 ln n
E2 , {(x, y) : x = 0, 0 < y <
}
(574)
n
2c1 ln n
E3 , {(x, y) : x = u(S)y, 0 < y <
}.
(575)
n
Since |∆2hd˜ (e,x,y)e f (x, y)| is a continuous function with respect to (x, y, e, h) ∈ T × S 1 × [0, 1], which is a compact set, we
T
can assume that (x0 , y0 , e0 , h0 ) achieves the supremum. Let A, B be the intersection of the line ` passing through (x0 , y0 )
with direction e0 with the triangle T .
If either A or B belongs to E1 , say, B ∈ E1 , then for sufficiently small > 0, the line connecting A and (x0 + , y0 )
(resp. (x0 − , y0 )) intersects E1 with direction e1 (resp. e2 ). Hence, by the similarity relation in geometry, the y-coordinates
of (x0 , y0 ) + h0 d˜T (e0 , x0 , y0 ) and (x0 + , y0 ) + h0 d˜T (e1 , x0 + , y0 ) are equal, and similarly for others. Hence, by linearity
of f (x, y) = x ln y in x,
E1 , {(x, y) : 0 < x <
2|∆2h
˜
0 dT (e0 ,x0 ,y0 )e0
f (x0 , y0 )| = |∆2h
˜
f (x0 + , y0 ) + ∆2h
˜
f (x0 + , y0 )| + |∆2h
0 dT (e1 ,x0 +,y0 )e1
≤ |∆2h
0 dT (e1 ,x0 +,y0 )e1
˜
0 dT (e2 ,x0 −,y0 )e2
f (x0 − , y0 )|
˜
0 dT (e2 ,x0 −,y0 )e2
f (x0 − , y0 )|
(576)
(577)
i.e., we can always perturb (x0 , y0 , e0 ) such that ` does not intersect E1 .
Now we assume that A = (0, y1 ), B = (u(S)y2 , y2 ) with y1 , y2 ∈ [0, 2c1nln n ]. If y2 = y1 , then f (x, y) is linear on `, and
2
ωT (f, 1/K) is zero. If y2 6= y1 , the function on ` becomes
h(y) =
y − y1
y2 u(S) ln y,
y2 − y1
y ∈ [u, v]
(578)
where u , y1 ∧ y2 and v , y1 ∨ y2 . Hence, by the sub-additivity of the Ditzian–Totik modulus of smoothness, for t ∈ [0, 1]
we have
y2 u(S)
y1 y2 u(S)
2
2
ωT2 (f, t) = ω[u,v]
(h, t) ≤
· ω 2 (y ln y, t) +
· ω[u,v]
(ln y, t)
(579)
|y2 − y1 | [u,v]
|y2 − y1 |
" 2
#
ω[u,v] (y ln y, t)
u
≤ u(S)v
+
· ω 2 (ln y, t)
(580)
v−u
v − u [u,v]
By the proof of Lemma 23 (where we have used ln(1 + x) ≤ x when ∆ ≤ t2 ), we have
3t2 (v − u)
.
u
2
As for ω[u,v]
(y ln y, t), we distinguish into two cases. If v ≤ 2u, by Taylor expansion we have
2
ω[u,v]
(ln y, t) ≤
2
ω[u,v]
(y ln y, t) ≤ sup |(y ln y)00 | · (v − u)2 t2 =
y∈[u,v]
(v − u)2 t2
≤ (v − u)t2 .
u
(581)
(582)
53
2
Otherwise, if v > 2u, since it has been shown in [38] that ω[0,1]
(y ln y, t) =
2t2 ln 2
1+t2 ,
by scaling [u, v] 7→ [u/v, 1] we have
2
2
2
(y ln y, t) ≤ v · ω[u/v,1]
(y ln y, t) ≤ v · ω[0,1]
(y ln y, t) ≤ 2 ln 2 · vt2 ≤ 4 ln 2 · (v − u)t2 .
ω[u,v]
A combination of the previous three inequalities yields
4 ln 2 · (v − u)t2
u
3t2 (v − u)
ωT2 (f, t) ≤ u(S)v
+
·
v−u
v−u
u
2
= (3 + 4 ln 2)u(S)vt
(6 + 8 ln 2)c1 u(S) ln n 2
·t
≤
n
where we have used v = y1 ∨ y2 ≤
(6 + 8 ln 2)c1 .
2c1 ln n
.
n
(583)
(584)
(585)
(586)
Now the desired result follows directly by choosing t = 1/K 2 and C0 =
E. Proof of Lemma 27
First we assume that b = t + 1, a = t − 1, t 6= 0, and write
pn (x) =
n
X
aν xν =
ν=0
n
X
bν (x − t)ν .
ν=0
(587)
By Lemma 26 and the related discussions, we know that |bν | ≤ 23n/2 for any ν = 0, · · · , n. Comparing coefficients yields
n
X
µ
|aν | =
(−t)µ−ν bµ
(588)
ν
µ=ν
n
X
µ
3n/2 −ν
≤2
|t|
|t|µ
(589)
ν
µ=ν
≤ 23n/2 |t|−ν 2n (n + 1)(|t|n + 1)
7n/2
≤2
|t|
−ν
(590)
n
(591)
(|t| + 1).
In the general case, the desired result is obtained by scaling.
F. Proof of Lemma 29
It is straightforward to show
Var(XY ) = E[(XY )2 ] − (E[XY ])2
2
2
2
= E[X ]E[Y ] − (EX) (EY )
(592)
2
2
(593)
2
2
2
= (Var(X) + (EX) )(Var(Y ) + (EY ) ) − (EX) (EY )
= Var(X)(EY )2 + Var(Y )(EX)2 + Var(X)Var(Y )
(594)
(595)
as desired.
R EFERENCES
[1] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of functionals of discrete distributions,” Information Theory, IEEE Transactions on,
vol. 61, no. 5, pp. 2835–2885, 2015.
[2] I. Csisz et al., “Information-type measures of difference of probability distributions and indirect observations,” Studia Sci. Math. Hungar., vol. 2, pp.
299–318, 1967.
[3] F. Liese and K.-J. Miescke, “Statistical decision theory,” in Statistical Decision Theory. Springer, 2007, pp. 1–52.
[4] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. New York: Wiley, 2006.
[5] A. Tsybakov, Introduction to Nonparametric Estimation. Springer-Verlag, 2008.
[6] S. Kullback and R. A. Leibler, “On information and sufficiency,” The annals of mathematical statistics, vol. 22, no. 1, pp. 79–86, 1951.
[7] C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, pp. 379–423, 623–656, 1948.
[8] O. Catoni and J. Picard, Statistical learning theory and stochastic optimization: Ecole d’Eté de Probabilités de Saint-Flour, XXXI-2001. Springer
Science & Business Media, 2004, vol. 31.
[9] I. Csiszar and J. Körner, Information theory: coding theorems for discrete memoryless systems. Cambridge University Press, 2011.
[10] I. Sanov, On the probability of large deviations of random variables. United States Air Force, Office of Scientific Research, 1958.
[11] S. Kullback, Information theory and statistics. Courier Corporation, 1997.
[12] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via the em algorithm,” Journal of the royal statistical society.
Series B (methodological), pp. 1–38, 1977.
[13] C. M. Bishop, “Pattern recognition,” Machine Learning, vol. 128, 2006.
[14] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013.
54
[15] Q. Wang, S. R. Kulkarni, and S. Verdú, “Divergence estimation of continuous distributions based on data-dependent partitions,” Information Theory,
IEEE Transactions on, vol. 51, no. 9, pp. 3064–3074, 2005.
[16] Y. K. Lee and B. U. Park, “Estimation of Kullback–Leibler divergence by local likelihood,” Annals of the Institute of Statistical Mathematics, vol. 58,
no. 2, pp. 327–340, 2006.
[17] A. Gretton, K. M. Borgwardt, M. Rasch, B. Schölkopf, and A. J. Smola, “A kernel method for the two-sample-problem,” in Advances in neural information
processing systems, 2006, pp. 513–520.
[18] F. Pérez-Cruz, “Kullback-Leibler divergence estimation of continuous distributions,” in Information Theory, 2008. ISIT 2008. IEEE International
Symposium on. IEEE, 2008, pp. 1666–1670.
[19] Q. Wang, S. R. Kulkarni, and S. Verdú, “Divergence estimation for multidimensional densities via-nearest-neighbor distances,” Information Theory, IEEE
Transactions on, vol. 55, no. 5, pp. 2392–2405, 2009.
[20] X. Nguyen, M. J. Wainwright, and M. I. Jordan, “Estimating divergence functionals and the likelihood ratio by convex risk minimization,” Information
Theory, IEEE Transactions on, vol. 56, no. 11, pp. 5847–5861, 2010.
[21] H. Cai, S. R. Kulkarni, and S. Verdú, “Universal divergence estimation for finite-alphabet sources,” Information Theory, IEEE Transactions on, vol. 52,
no. 8, pp. 3456–3475, 2006.
[22] Z. Zhang and M. Grabchak, “Nonparametric estimation of Küllback-Leibler divergence,” Neural computation, vol. 26, no. 11, pp. 2570–2593, 2014.
[23] A. W. Van der Vaart, Asymptotic statistics. Cambridge university press, 2000, vol. 3.
[24] J. Jiao, Y. Han, and T. Weissman, “Minimax estimation of the l1 distance,” in 2016 IEEE International Symposium on Information Theory (ISIT). IEEE,
2016, pp. 750–754.
[25] Y. Bu, S. Zou, Y. Liang, and V. V. Veeravalli, “Estimation of KL divergence between large-alphabet distributions,” in 2016 IEEE International Symposium
on Information Theory (ISIT). IEEE, 2016, pp. 1118–1122.
[26] ——, “Estimation of KL divergence: Optimal minimax rate,” arXiv preprint arXiv:1607.02653, 2016.
[27] T. T. Cai and M. G. Low, “Testing composite hypotheses, Hermite polynomials and optimal estimation of a nonsmooth functional,” The Annals of
Statistics, vol. 39, no. 2, pp. 1012–1041, 2011.
[28] G. Valiant and P. Valiant, “The power of linear estimators,” in Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium on.
IEEE, 2011, pp. 403–412.
[29] P. Valiant and G. Valiant, “Estimating the unseen: improved estimators for entropy and other properties,” in Advances in Neural Information Processing
Systems, 2013, pp. 2157–2165.
[30] Y. Wu and P. Yang, “Minimax rates of entropy estimation on large alphabets via best polynomial approximation,” arXiv preprint arXiv:1407.0381, 2014.
[31] J. Acharya, A. Orlitsky, A. T. Suresh, and H. Tyagi, “The complexity of estimating Rényi entropy.” SODA, 2015.
[32] Y. Wu and P. Yang, “Chebyshev polynomials, moment matching, and optimal estimation of the unseen,” arXiv preprint arXiv:1504.01227, 2015.
[33] Y. Han, J. Jiao, R. Mukherjee, and T. Weissman, “Optimal estimation of Lr norm of functions in Gaussian white noise,” in preparation, 2016.
[34] G. A. Miller, “Note on the bias of information estimates,” Information Theory in Psychology: Problems and Methods, vol. 2, pp. 95–100, 1955.
[35] A. Carlton, “On the bias of information estimates.” Psychological Bulletin, vol. 71, no. 2, p. 108, 1969.
[36] A. Nemirovski, “Topics in non-parametric,” Ecole dEté de Probabilités de Saint-Flour, vol. 28, p. 85, 2000.
[37] M. Mitzenmacher and E. Upfal, Probability and computing: Randomized algorithms and probabilistic analysis. Cambridge University Press, 2005.
[38] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Non-asymptotic theory for the plug-in rule in functional estimation,” arXiv preprint arXiv:1406.6959,
2014.
[39] D. Braess and T. Sauer, “Bernstein polynomials and learning theory,” Journal of Approximation Theory, vol. 128, no. 2, pp. 187–206, 2004.
[40] A. Tsybakov, “Aggregation and high-dimensional statistics,” Lecture notes for the course given at the École dété de Probabilités in Saint-Flour, URL
http://www. crest. fr/ckfinder/userfiles/files/Pageperso/ATsybakov/Lecture notes SFlour. pdf, vol. 16, p. 20, 2013.
[41] C. S. Withers, “Bias reduction by Taylor series,” Communications in Statistics-Theory and Methods, vol. 16, no. 8, pp. 2369–2383, 1987.
[42] G. Peccati and M. S. Taqqu, “Some facts about Charlier polynomials,” in Wiener Chaos: Moments, Cumulants and Diagrams. Springer, 2011, pp.
171–175.
[43] J. R. Rice, “Tchebycheff approximation in several variables,” Transactions of the American Mathematical Society, pp. 444–466, 1963.
[44] Z. Ditzian and V. Totik, Moduli of smoothness. Springer, 1987.
[45] V. Totik, Polynomial Approximation on Polytopes. Memoirs of the American Mathematical Society, 2013, vol. 232.
[46] F. Dai and Y. Xu, Approximation theory and harmonic analysis on spheres and balls. Springer, 2013.
[47] E. Y. Remez, “Sur la détermination des polynômes dapproximation de degré donnée,” Comm. Soc. Math. Kharkov, vol. 10, pp. 41–63, 1934.
[48] Chebfun Version 5, The Chebfun Development Team, 2014, http://www.chebfun.org/.
[49] A. Haar, “Die minkowskische geometrie und die annäherung an stetige funktionen,” Mathematische Annalen, vol. 78, no. 1, pp. 294–311, 1917.
[50] O. Lepski, A. Nemirovski, and V. Spokoiny, “On estimation of the Lr norm of a regression function,” Probability theory and related fields, vol. 113,
no. 2, pp. 221–253, 1999.
[51] J. Jiao, Y. Han, and T. Weissman, “Minimax estimation of divergence functions,” in preparation.
[52] M. Qazi and Q. Rahman, “Some coefficient estimates for polynomials on the unit interval,” Serdica Math. J, vol. 33, pp. 449–474, 2007.
[53] M. Mitzenmacher and E. Upfal, Probability and computing: Randomized algorithms and probabilistic analysis. Cambridge University Press, 2005.
[54] J. C. Mason and D. C. Handscomb, Chebyshev polynomials. CRC Press, 2002.
[55] A. Wald, Statistical decision functions. Wiley, 1950.
| 10math.ST
|
arXiv:1605.00716v1 [cs.LG] 3 May 2016
Radio Transformer Networks: Attention Models for
Learning to Synchronize in Wireless Systems
Timothy J. O’Shea
Latha Pemula
Dhruv Batra
T. Charles Clancy
Virginia Tech
Arlington, VA
[email protected]
Virginia Tech
Blacksburg, VA
[email protected]
Virginia Tech
Blacksburg, VA
[email protected]
Virginia Tech
Arlington, VA
[email protected]
Abstract—We introduce learned attention models into the radio machine learning domain for the task of modulation recognition by leveraging spatial transformer networks and introducing
new radio domain appropriate transformations. This attention
model allows the network to learn a localization network capable
of synchronizing and normalizing a radio signal blindly with zero
knowledge of the signal’s structure based on optimization of the
network for classification accuracy, sparse representation, and
regularization. Using this architecture we are able to outperform
our prior results in accuracy vs signal to noise ratio against an
identical system without attention, however we believe such an
attention model has implication far beyond the task of modulation
recognition.
Attention models have recently been gaining widespread
adoption in the computer vision community for a number
of important reasons. They introduce a learned model for
attention capable of removing numerous variances and parametric search spaces in the input data and focuses on the
task of extracting a canonical form attention patch with these
variations removed to make downstream tasks easier and of
lower complexity. These were first introduced as recurrent
networks [4] which were quite expensive, but have made
significant progress since then.
Keywords—Radio Transformer Networks, Radio communications, Software Radio, Cognitive Radio, Deep Learning, Convolutional Autoencoders, Neural Networks, Machine Learning,
Attention Models, Spatial Transformer Networks, Synchronization,
RadioML, Signal Processing
I.
I NTRODUCTION
Cognitive radio and signal processing in general has long
relied on a relatively well defined set of expert systems and
expert knowledge to operate. Unfortunately in the realization
of cognitive radio, this has greatly limited the ability of systems
to generalize and perform real learning and adaptation to
new and unknown signals and tasks. By approaching signal
recognition, synchronization, and reasoning from a ground-up
feature learning angle, we seek to be able to build cognitive
radio systems which truly generalize and adapt without running
into barriers of expert knowledge such as many of the current
day solutions which address more narrowly scoped problems.
In our prior work, we looked at the application of deep
convolutional neural networks to the task of modulation recognition [9] through blind feature learning on time domain radio
signals. We were able to achieve excellent classification performance at both low and high SNR by learning time domain
features directly from a dataset with harsh channel impairments
(oscillator drift, clock drift, fading, noise). However we had
no notion of attention in this work and instead forced the
discriminative network to learn features invariant to each of
these channel effects. In communications receivers (and many
iterative expert modulation classification algorithms), we typically perform synchronization on the signal before performing
additional signal processing steps. This synchronization can
be thought of as a form of attention which estimates a time,
frequency, phase, and sample timing offset in order to create
a normalized version of the signal.
Figure 1.
Generalized Transformer Network Architecture
Spatial transformer networks (STNs) [8] were recently
introduced to provide an end to end feed-forward model of
attention which can be trained directly from loss on each
training example and compactly evaluated on new samples.
They consist of a trained Localization network which performs
parameter regression, a fixed parametric transform operation,
and a trained discriminative classifier to select a class estimate.
In the image domain, where these have so far been applied, a
2D Affine transform is used to extract an attention patch which
is shifted, scaled, and rotated in 2-space from the original
image according to a 2x3 parameter vector θ, generalized in
figure 1.
In this work, we propose a Radio Transformer Network
(RTN), which leverages the generalization of the STN architecture, but introduces radio-domain specific parametric
transforms. This attention model can be used to learn directly
how to synchronize in wireless systems, and enables our
modulation recognition system to outperform the attentionless version of itself by assisting in the normalization of the
received signal prior to classification. By constructing this
normalized received signal with an attention model we greatly
simplify the task of the discriminative network by relaxing
the requirements on various variations of the received signal
it must recognize, reducing the complexity and increasing the
performance necessary in a discriminative network.
This is an important result in modulation recognition but
also more widely in radio communications and signal processing, as it demonstrates that we can learn to synchronize
rather than relying on expert systems and estimators derived
through a costly analytic process. We believe attention models
will play an important and wide-spread role in forthcoming
machine learning based signal processing systems.
II.
L EARNING TO C LASSIFY S IGNALS
Figure 3.
Original ConvNet Performance vs Expert Statistics
believe the correction of fading through equalization has the
potential to also be addressed as an attention model, either
jointly or subsequently to the transformations addressed here.
A. Timing and Symbol-Rate Recovery
Figure 2.
Original ConvNet Architecture without Attention
In prior work [9] we compare supervised learning using
a deep convolutional network with no expert features, to a
handful of widely used machine learning techniques on expert
signal amplitude, phase, and envelope moments. We used the
architecture shown in figure 2 using a convolutional frontend,
a dense backend, and a softmax with categorical cross-entropy
training using Adam/SGD against a synthetic dataset.
In figure 3 we summarize the results of that experiment
without an attention model, demonstrating a significant improvement over the moment based features and conventional
classifiers.
This is an exciting result as it demonstrates that feature
learning on raw RF data does work, and in this case it is
working better than conventional widely used expert features!
III.
U SING ATTENTION TO S YNCHRONIZE
To effectively synchronize to a wireless signal, we must
develop a transform which, with the right parameters is able to
correct for channel induced variation. Within the scope of this
paper, we consider channel variation due to time offset, time
dilation, frequency offset, and phase offset. These are effects
which exist in any real system containing transmitters and
receivers whose oscillators and clocks are not locked together.
For now we do not address the problem of fading, but we
Timing and symbol-rate recovery are relatively straightforward processes involving the re-sampling of the input signal at
the correct starting offset and sampling increment. This is very
much akin to the extraction of visual pixels at the correct offset
of a 1D Affine transformation, and so we treat it as such by
directly leveraging the Affine transformation used in the image
domain. We represent our data as a 2D image, with a two rows
containing I/Q and an N columns containing samples in time.
A full 2D Affine transformation allows for translation,
rotation, and scaling in 2-dimensions given by a 2x6 element
parameter vector. To restrict this to 1-Dimensional translation
and scaling in the time dimension, we can simply introduce
the following mask 1 and then readily use 2D Affine transform
implementations from the image domain.
θ0
0
0
θ1
θ2
0
(1)
B. Phase and Frequency Offset Recovery
Phase and frequency offset recovery is a task which doesn’t
have an immediate analogue in the vision domain. However
this transform in signal processing is relatively straightforward.
We simply mix our signal with a complex sinusoid with the
proper initial phase and frequency as defined by two new
unknown parameters.
yn = xn ∗ exp(nθ3 + θ4 )
(2)
We directly implement this transform as a new layer in
Keras (on top of Theano and Tensorflow), and cascade it
before the Affine transform in the Transformer module of our
network.
C. Parameter Estimation
The task of synchronization now becomes the task of
parameter estimation of θi values passed into our transformer
module. We experimentally try a number of different neural
network architectures for performing this parameter regression
task, and ultimately introduce two new domain appropriate
layers into Keras to help assist in their estimation.
1) Complex Convolution 1D Layer: Complex neural networks are not widely used and still faced some theoretical
issues especially in automated differentiation, so we represent
our signal as a two row 2D matrix with the real component in
the first row and imaginary component in the second row. In
theory real valued convolutions in a neural network can learn
the relationship between these components to some extent, but
by introducing a complex convolution operation, we simplify
the learning task and ensure that we learn a filter with the
properties we are used to working with.
For a complex valued input vector X of size 2xN, we define
a weight vector W of M complex filters each 2xK in length.
We may then compute the output for each of the k output
values as.
conv(X0,: , Wk,0,: ) − conv(X1,: , Wk,1,: )
conv(X0,: , Wk,1,: ) + conv(X1,: , Wk,0,: )
(3)
This allows us to leverage existing, highly optimized real
convolution operations and obtain a differentiable operation
which can be trained with back-propagation.
4) Training Details: We train this network using Keras [7]
on top of Theano [2] and TensorFlow [6] using an NVidia
Geforce Titan X inside a Digits Devbox. We use dropout [5]
of 0.5 between each layer for regularization, and the Adam [3]
method of stochastic gradient descent to fit network parameters
on our training set.
We train with a batch size of 1024 and an initial learning
rate of 0.001. We train for roughly 350 epochs, reducing our
learning rate by half each validation loss stops decreasing.
Training takes about 3 hours on a Titan X GPU, but feedforward evaluation or signal classification takes less than 10ms.
IV.
DATA - SET AND M ETHODOLOGY
For our classifier performance evaluation, in this work and
in prior work, we leverage the RadioML.com 2016.04C open
source dataset and perform at 60/40 split between train and test
sets. This consists of 11 modulations (8 digital and 3 analog) at
varying SNR levels, with random walk simulations on center
frequency, sample clock rate, sample clock offset, and initial
phase, as well as limited multi-path fading. We believe it is
critically important to test with real channel effects early to
ensure realistic assumptions early in our models. This dataset
is labeled with SNR and Modulation-Type, when performing
supervised training we use only the modulation-type labels
of the training set, and evaluate the classification accuracy
performance at each SNR label step for the test set.
Training and validation loss along with learning rate are
shown throughout the trainin in figure 5.
2) Complex to Power and Phase: Creating a differentiable
Cartesian to Polar operation which makes it easier for the
network to operate directly on input phase and magnitude is
slightly more involved. We compute magnitude squared simply
as mn = pow(x0,n , 2) + pow(x1,n , 2), but for phase computation we use a simplified and differentiable approximation of
atan2 without conditionals implemented in Keras on top of
Theano and Tensorflow.
z = Xq/K.clip(K.abs(Xi),1e-3,1e6)
z = z*K.sign(xi)
zd1 = 1+0.28*K.pow(z,2)
z1 = z/zd1 + PI*K.sign(Xq)
zd2 = K.pow(z,2) + 0.28
z2 = z/zd2 + (K.sign(Xq)-1)*0.5*PI
zc = K.abs(z) - 1
atan2 = (K.sign(zc)-1)*(-0.5)*z1 +
(K.sign(zc)+1)*0.5*z2
3) Network Architecture: We evaluate dozens of localization network architectures with slight variation between dense
connections, convolutional layers, and complex-convolutional
layers, using various activation functions, and achieve our best
performance using the composite network shown in figure
4. This uses both the complex convolutional layer and the
complex to polar layers within the localization network and
an identical discriminative network to the one we used without
attention in front of it for comparison.
Figure 5.
Training Loss and Learning Rate
V.
C LASSIFICATION P ERFORMANCE
Evaluating the performance of the RTN on the test set, we
obtain slightly increased performance over the model without
attention. Similar accuracies are obtained at slightly lower SNR
values ( 1dB) and high SNR performance is slightly improved
and more stable as shown in figure 6.
We suspect the complexity of the discriminative network
could now be reduced due to lower complexity of the normalized signal but we do not investigate this work further here for
fair comparison of the same discriminative network.
Figure 4.
Radio Transformer Network Architecture
Figure 6.
Radio Transformer Network Performance
Performance of the convolutional neural network without
attention is also improved from our prior work by increasing
dropout and better learning rate policy to match that used in
our RTN training. This is reflected in figure 6.
VI.
Figure 7.
Performance of the RTN at High SNR (18dB)
Figure 8.
Density Plots of Pre- and Post-Transformed Input Constellations
ATTENTION L EARNING P ERFORMANCE
We have shown that classification performance is improved
while using a radio localization network to extract normalized
patches on the dataset. However, we can not only look at
classification performance, it is also interesting to look at the
radio sample data before and after transformation to observe
any normalization that has occurred. It is difficult to visualize
exactly what has occured here, looking at time domain data
does not yield clean, obvious performance improvement on the
data. Upon attempting to plot an eye-diagram of the QPSKclass signals, it is clear that synchronizaiton is at this point
still horribly noisy and partial. However, if we plot the the
constellation density for 50 test examples over a range of 20
time samples each, shown in figure 8, we can start to see a bit
more density forming around the constellation points vs what
we started with, which is a good sign to start with.
[5]
Clearly much work needs to be done to improve and
quantify synchronization performance, in-fact we have no real
reason to expect perfect synchronization from a classification
task, just enough normalization to make things easier on the
discriminative network. We will continue to investigate this
area, and additional tasks other than modulation recognition
which may improve the synchronization properties for demodulation beyond what has been achieved here.
[6]
VII.
C ONCLUSIONS
By developing a feed forward model for radio attention, we
have demonstrated that we can effectively learn to synchronize
using deep convolutional neural networks with domain specific
transforms and layer configurations. Normalizing out time,
time-dilation, frequency and phase offsets using learned estimators does effectively improve our modulation classification
performance and requires no expert knowledge about the
signals of interest to train.
While the training complexity of such a network is high,
the feed-forward execution of it is actually quite compact
and viable for real-world use and deployment. Platforms such
as the TX1 with highly parallel, low clock rate GPGPU
architectures further enable the low-SWaP deployment of these
algorithms for which they are exceptionally well suited.
ACKNOWLEDGMENTS
The authors would like to thank the Bradley Department
of Electrical and Computer Engineering at the Virginia Polytechnic Institute and State University,the Machine Learning &
Perception Group, the Hume Center, and DARPA all for their
generous support in this work.
This research was developed with funding from the Defense Advanced Research Projects Agency’s (DARPA) MTO
Office under grant HR0011-16-1-0002. The views, opinions,
and/or findings expressed are those of the author and should
not be interpreted as representing the official views or policies
of the Department of Defense or the U.S. Government.
R EFERENCES
[1]
C. Clancy, J. Hecker, E. Stuntebeck, and T. O’Shea,
“Applications of machine learning to cognitive radio
networks”, Wireless Communications, IEEE, vol. 14, no.
4, pp. 47–52, 2007.
[2] J. Bergstra, O. Breuleux, F. Bastien, P. Lamblin, R.
Pascanu, G. Desjardins, J. Turian, D. Warde-Farley, and
Y. Bengio, “Theano: a CPU and GPU math expression compiler”, in Proceedings of the Python for Scientific Computing Conference (SciPy), Oral Presentation,
Austin, TX, Jun. 2010.
[3] D. Kingma and J. Ba, “Adam: a method for stochastic
optimization”, arXiv preprint arXiv:1412.6980, 2014.
[4] V. Mnih, N. Heess, A. Graves, et al., “Recurrent models
of visual attention”, in Advances in Neural Information
Processing Systems, 2014, pp. 2204–2212.
[7]
[8]
[9]
N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting”, The Journal of Machine
Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.
M. Abadi, A. Agarwal, et al., TensorFlow: large-scale
machine learning on heterogeneous systems, Software
available from tensorflow.org, 2015. [Online]. Available:
http://tensorflow.org/.
F. Chollet, Keras, https://github.com/fchollet/keras, 2015.
M. Jaderberg, K. Simonyan, A. Zisserman, et al., “Spatial
transformer networks”, in Advances in Neural Information Processing Systems, 2015, pp. 2008–2016.
T. J. O’Shea and J. Corgan, “Convolutional radio modulation recognition networks”, CoRR, vol. abs/1602.04105,
2016. [Online]. Available: http : / / arxiv. org / abs / 1602 .
04105.
| 3cs.SY
|
Analysis of planar ornament patterns via motif asymmetry assumption and
local connections
V. Adanova, S. Tari∗
arXiv:1710.04623v1 [cs.CV] 12 Oct 2017
Department of Computer Engineering, Middle East Technical University, 06800 Ankara
Abstract
Planar ornaments, a.k.a. wallpapers, are regular repetitive patterns which exhibit translational symmetry
in two independent directions. There are exactly 17 distinct planar symmetry groups. We present a fully
automatic method for complete analysis of planar ornaments in 13 of these groups, specifically, the groups
called p6m, p6, p4g, p4m, p4, p31m, p3m, p3, cmm, pgg, pg, p2 and p1. Given the image of an ornament
fragment, we present a method to simultaneously classify the input into one of the 13 groups and extract
the so called fundamental domain (FD), the minimum region that is sufficient to reconstruct the entire
ornament. A nice feature of our method is that even when the given ornament image is a small portion
such that it does not contain multiple translational units, the symmetry group as well as the fundamental
domain can still be defined. This is because, in contrast to common approach, we do not attempt to first
identify a global translational repetition lattice. Though the presented constructions work for quite a wide
range of ornament patterns, a key assumption we make is that the perceivable motifs (shapes that repeat)
alone do not provide clues for the underlying symmetries of the ornament. In this sense, our main target is
the planar arrangements of asymmetric interlocking shapes, as in the symmetry art of Escher.
Keywords: ornaments, wallpaper groups, mosaics, regular patterns, Escher style planar patterns
1. Introduction
Planar ornaments, a.k.a. wallpapers, are repetitive patterns which exhibit translational symmetry
in two independent directions. They form a tiling
of the plane. They are created by repeating base
unit in a predictable manner, using four primitive
planar geometric operations: translation, rotation,
reflection and glide reflection (Fig. 1). Using combinations of these primitive operations applied on
a base unit, different patterns can be generated.
An interesting fact is that the four primitive operations can be combined in exactly seventeen different
ways to tile a plane, forming the so called 17 P lane
Symmetry Groups [1].
We present an illustrative example in Fig. 2.
Firstly, observe that the pattern Fig. 2 (a) can be
generated by replicating the equilateral triangular
∗ Corresponding
author
Email addresses: [email protected]
(V. Adanova), [email protected] (S. Tari )
Preprint submitted to Elsevier
(a)
(b)
(c)
(d)
Figure 1: 4 primitive operations.(a) Translation, (b) reflection, (c) glide reflection, (d) 4 types of rotational symmetry.
fragment depicted in Fig. 2(b) by 60◦ rotations in a
systematic manner. This equilateral triangle fragment is the smallest fragment of the pattern that is
sufficient to construct the entire pattern in Fig. 2(a)
using four primitive operations. As such, it is reOctober 13, 2017
letter c which stands for centered cell. The primitive cell is a unit cell with the centers of highest
order of rotation at the vertices. The centered cell
is encountered only in two cases (cm and cmm symmetry groups), and is chosen so that the reflection
axis is normal to one or both sides of the cell. The
digit that follows the letters p or c indicates the
highest order of rotation, whereas the letter characters m and g respectively stand for mirror and glide
reflections. When there are two positions containing either of m and g, it is understood that the first
reflection is normal to x axis and the second is at an
angle α. The digit denoting the highest order of rotation in a symmetry group can take only values 1,
2, 3, 4, and 6. This restriction is introduced by the
crystallographic restriction theorem, which states
that the patterns repeating in two dimension can
only exhibit 180◦ , 120◦ , 90◦ , and 60◦ rotations.
In the present paper, given an image fragment
from an ornament belonging to either of the mentioned 13 groups, we present a robust method to
extract the fundamental domain along with the underlying symmetry operations. That is, we provide a complete analysis of an ornamental pattern
in 13 groups based on an image. An important feature of our computational scheme is even when the
given input image is only a small portion that a
full unit cell does not fit to image (e.g., Fig. 4(a)),
the symmetry group as well as the fundamental domain can still be defined. Of course, once the symmetry group and fundamental domain are defined,
it becomes trivial to deduce the translational unit
cell. We remark that most of the existing methods
rely on first discovering the underlying lattice via
a translational repetition structure. This requires
global calculations, e.g., autocorrelations. In contrast, we do not search for a translational repetition
or lattice. We directly look for local connections
among motifs (protiles) from which we deduce symmetry clues that are later integrated via a decision
tree.
Using a decision tree is indeed the classical
method for grouping tiles into symmetry groups
based on individual clues [2]; the classical decision
tree is depicted in Fig. 5; observe that the check
for mirror reflections dominate the yes/no question
set. The rational behind the given question sequence could be that it is easy for humans to spot
mirror reflection and rotational symmetries while
much harder to spot glides. While this is true, the
robustness in the computational case might not be
the same.
Figure 2: An example. (a) The ornament image, (b) fundamental domain (FD), (c) unit cell (UC), (d) symmetries
and the relation between FD and UC, (e) symmetries, UC
and FD superimposed on the ornament image.
ferred as the fundamental domain (FD). Secondly,
by rotating the FD 120◦ twice around its top corner
and then rotating 180◦ around the middle point of
its base yields the rhombus depicted in Fig. 2(c).
This is the smallest translational unit; the pattern
in (a) can be generated by simply translating it
along two independent directions; hence, it is referred as the unit cell (UC). The abstracted unit
cell of the tile, the relation between the fundamental domain and the unit cell along with the symmetries, is shown in Fig. 2 (d). The blue hexagons
on the four corners of the rhombus indicate six-fold
rotation centers, and the red triangles and the pink
diamonds located on the centers and the side midpoints of the two triangles forming the rhombus respectively indicate three-fold and two-fold rotation
centers. In the final illustration (Fig. 2(e)) a sample cell is shown as superimposed on the original
ornament image.
Through a common naming provided by Crystallographic notation, the presented example happens
to belong to a group called as p6, indicating six fold
rotations. In the Crystallographic notation, the remaining 16 groups are named as p1, pm, pg, cm,
p2, pmm, pmg, pgg, cmm, p3, p3m1, p31m, p4,
p4m, p4g and p6m. In Fig. 3, we depict the cell
structures of each of the 13 symmetry groups that
we are interested in.
In the group name, each character position defines a group property: The first position is either
the letter p which stands for primitive cell or the
2
Figure 3: Unit cell structures for 13 Wallpaper Groups. Darker regions indicate fundamental domains.
(a)
(b)
the groups, pm, pmm, cm, pmg and gives us the 13
groups listed above. For the 13 groups, our goal is
to obtain the fundamental domain robustly up to a
mirror reflection.
In accumulating symmetry clues, we only resort to
local connections among motifs (protiles). Many of
the classical ornaments, such as those found in Islamic art, are constructed from symmetric protiles
such as stars that provide a clue to the symmetry
group of the tile. In some works, the symmetries
of the protiles themselves are used as a clue to the
symmetry. However, we believe that inferring symmetries of the motifs are not robust due to possible
noise in motif extraction. Moreover, the ornament
artist may be using nearly symmetric motifs that
do not reflect the symmetries of the ornament itself (several such examples exist in Escher’s art).
Hence, we neither attempt to recover individual motifs correctly nor check their symmetries. Indeed,
we even expect the patterns we analyze contain
sufficient number of asymmetric motifs or at least
the most significant symmetries are not completely
swallowed by motifs having the same symmetry.
Motif asymmetry assumption is not as restrictive as it may strike at first. Consider two valid
examples shown in Fig. 6. The first one is of group
p6 and contains exactly three all symmetric protiles – brown, beige and gray– respectively having
six-fold, three-fold and two-fold rotational symmetries. That is, the ornament does not contain any
asymmetric motif and all three types of rotation
centers coincide with the motif centers. Nevertheless, the local relations among the two-fold rotational symmetric gray motifs reveal both the sixfold and three-fold rotational centers, providing suf-
(c)
Figure 4: An example case for a pattern with few repetitions. (a) Original ornament, (b) fundamental domain, (c)
unit cell (black quadrilateral). Observe that the ornament
fragment does not contain the full unit cell.
We propose an alternative decision tree, of which
details are given in § 3. Initially, we accumulate
indirect clues to mirror reflection as opposed to
searching for mirror reflection axis. We postpone
mirror reflection control till the last stage, and then
at the last stage use mirror reflection check about
the predicted axes only to eliminate possible false
alarms resulting from indirect clues. That is, we
use mirror reflection check only to eliminate false
alarms, not to catch missed ones. This means that
we are willing to sacrifice mirror reflections in order
not to wrongly assume that the tile has a mirror reflection. Our rational is as follows: If we only miss
the mirror reflection, say classify a p6m tile as p6,
we still get the correct unit cell and a redundant
(twice the size) fundamental domain. Hence, the
pattern can be correctly generated. If, however, we
falsely classify a p6 tile as p6m, then the extracted
fundamental domain is not sufficient, as we falsely
assume the existence of a mirror reflection. A consequence of not searching for mirror reflection is that
we can only recognize groups that contain sufficient
symmetries other than mirror. This leaves out 4 of
3
Figure 5: Classical decision tree.
e.g., [3], with a goal of revealing social structures
and their interaction via dominant symmetries used
in the ornament designs of individual cultures or
geographical regions. In mathematics, the ornament patterns are studied in terms of the groups
formed by the symmetry operations. Few examples
include [4, 5, 6]. The Dutch artist, Escher took a
particular interest in patterns formed by repeating
asymmetric shapes and discovered a local structure
leading to the same wallpaper patterns; his work
on symmetry is examined in [7]. Regular repetitive patterns such as Wallpaper and Frieze groups
are even utilized in quite practical problems; for
example, to analyze human gait [8] or to achieve
automatic fabric defect detection in 2D patterned
textures [9, 10, 11].
ficient clues for p6. The second ornament, at a first
glance, might give the impression that it contains
only a single form that has all the symmetries of the
ornament itself, i.e., three-fold rotational symmetry
and mirror reflection. Nevertheless, due to texture,
there are also other detectable motifs in the form of
circles or circle fragments, of which local relations
provide clues for various three-fold rotation centers.
In general, our constructions can not handle ornaments of single motif such that the motif contains
all the symmetries of the symmetry group; for example, a checkerboard pattern or a uniform pack
of triangles. We remark, however, that such ornaments are typically the most obvious ones to recognize. Furthermore, even for those class of ornaments, it is possible to identify the translation grid
though full analysis revealing FD can not be performed.
The rest of the paper is organized as follows. § 2
is on related work. The Method details are given in
§ 3 and the results on 100 tile set is in § 4. Finally,
§ 5 is the Summary and Conclusion.
In the general pool of works in computational
symmetry, the main focus has been finding symmetry axes in single objects. Since a single object can
exhibit only mirror reflections and rotational symmetries, the efforts are heavily focused on reflections and rotations [12, 13, 14, 15, 16, 17, 18, 19]
or finding local symmetries a.k.a shape skeletons.
To our knowledge, [20, 21] are the only works that
address finding a glide reflection axis in an image,
though the goal is to study one dimensional arrangements of symmetry, e.g., leaves. The works
targeting shape symmetry, whether directly from
an image or from a segmented region, fall out of
our focus. Our focus is on the symmetries of planar
2. Related Work
Ornament patterns have always been a source of
curiosity and interest, not only in arts and crafts
but also other fields including mathematics, computation, cultural studies etc. Early researchers
mostly examined ornaments in cultural contexts,
4
termined. Since the possible lattice types that
can be associated with are restricted by the symmetries the ornament pattern exhibits, detecting
the lattice types reduces the number of symmetry groups. Though the lattice detection is commonly performed using peak heights in autocorrelation [30], in [31] an alternative peak detection algorithm based on so called regions of dominance is
used to detect patterns translational lattice. The
region of dominance is defined as the largest circle
centered on the candidate peak such that no higher
peaks are contained in the circle. The authors argue that the region of dominance is more important
than the height of the peak. A Hough transform is
used to detect two shortest translation vectors that
best explains the majority of the point data. In
order to test whether a pattern has certain symmetry, the conjectured symmetry is applied to entire
pattern and then the similarity between the original image and the transformed one is computed.
The representative motif is chosen to be the most
symmetrical figure.
patterns formed by regular repetition of shapes via
four primitive geometric transformations.
Figure 6: Possible valid inputs.
In analyzing planar periodic patterns, translational symmetry, the most primitive repetition operation, is encountered in several works on recurring
structure discovery [22, 23, 24, 25, 26]. The general
flow of such works is to detect visual words and
cluster them based on their appearance and spatial layout. Among these works, [23, 24, 25] further
perform image retrieval based on the discovered recurring structures. In [23, 24] instead of directly using the recurring structures for image matching, the
authors first detect a translational repetition lattice
of an image. There can be multiple lattices for an
image. Thus, given a query image with various detected lattices, they search a database for images
with equivalent lattices. In each search, the matching score between two lattices is a product of two
measurements: the similarity of the grayscale mean
of the representative unit cell and the similarity of
the color histograms. In [27, 28] detection of deformed lattice in a given pattern is proposed. They
first propose a seed lattice from detected interest
points. Using those interest points a commonly occurring lattice vectors are extracted. Subsequently,
the seed lattice is refined, and grown outward until
it covers whole pattern.
In [29], for translational symmetry, model based
lattice estimation is performed where the model
comparison for hypotheses generated via peaks of
the autocorrelation is implemented using approximate marginal likelihood.
Some works move further beyond computing a
translational lattice to address classifying repeated
patterns according to the 17 plane symmetry groups
(wallpapers) [30, 31, 32, 33, 34]. As a first step,
lattice detection is performed. After the lattice
detection step, sequence of yes/no questions are
answered, until the final symmetry group is de-
Recently, [35], combined lattice extraction and
point symmetry groups of individual motifs to analyse Islamic patterns in mosaics. This method
specifically targets Islamic ornaments in which motifs such as n-stars typically provide clues to the
underlying plane symmetry group. As such, it is
not readily applicable if motifs can not be robustly
extracted or motifs do not reflect the symmetries.
In, [36], rotation groups are detected to analyze Islamic rosette patterns.
It is also possible to perform a continuous characterization of the ornament by comparing ornament images. This is for example encountered in
[37], where ornament images are classified according to a symmetry feature vector calculated based
on a prior lattice extraction and yes/no questions;
for lattice detection, they used method in [38]. In
[39], ornament images are directly compared in a
transformed domain after applying a global transformation.
Note that among the works addressing planar
patterns, there are also several interesting works
on pattern synthesis, including how to generate an
ornament in a certain symmetry group, how to use
a given motif to tile the plane in a certain style, or
how to map a given wallpaper pattern to a curved
surface [40, 41].
5
number of clusters to be specificied. (This is unlike k-means). However, it requires a bandwidth
parameter which indirectly influence the number of
detected clusters. We automate the clustering process by iteratively using mean shift clustering increasing the bandwidth at each iteration as follows.
3. The Method
Our symmetry detection system has three modules:
image processing module, local connectivity analysis, and final symmetry detection. Each of them are
separately explained below.
At the initial step, the bandwidth parameter for
mean shift algorithm is set to binit and the number
of clusters is observed. Then at each iteration, the
bandwidth is increased by bstep . For all our images
binit = Nc and bstep = binit /2. The iterations are
stopped whenever the number of clusters k drop
below Nc . The resulting k clusters are used to
define k binary images. The resulting bandwidth is
taken as the image dependent bandwidth estimate
b∗ .
3.1. Image Processing
The input to the image processing module is an arbitrary ornament I which may be a noisy scanned
image or screen shot of a part of an ornament drawn
using a computer tool. That is, input ornament images are acquired in arbitrary imaging conditions.
The processing proceeds in three stages: gamma
correction, initial clustering, and refinement. At
the end of the refinement step, k binary images for
each ornament are obtained. These binary images
will be called masks for that ornament image.
The number of masks is a result of adaptive clustering. Quite often, the number of masks coincides
with the number of colors in that ornament image.
This, however, is not always so because the image
processing module is not given the number of colors.
Refinement. The computation of initial clusters
as outlined above is performed in the color space.
Hence, spatial proximity of the pixels are not taken
into account. In the next stage, a sequential combination of median filtering in the pixel space and
mean shift clustering in the color space are applied
iteratively for the fixed bandwidth b∗ . The median
filtering is realized as follows: If a pixel of class ci
is surrounded by a pixel of class cj , it is assigned to
class cj , and L ∗ a ∗ b∗ channels of that pixel to the
cluster center of the cluster cj . This sequential application of median filtering followed by mean shift
with fixed bandwith b∗ is performed only few times.
For all of our images, five iterations seemed sufficient. More iterations may cause the components
of different colors to join.
Gamma Correction. The first step of the image processing module is gamma-correction [42] to
brighten the black lines and shadows. It is performed on the Y component of xyY color space
using the following formula:
Yout = Ymin + (Ymax − Ymin ) ∗ (
Y − Ymin γ
)
Ymax − Ymin
Resulting image depends on the γ parameter. If
one chooses γ < 1 then the lightness of an image
is higher than in original image and darker colors
have more contrast. When γ = 1 no effect on
original image is observed, while γ > 1 makes
colors darker than in original image.
Final clusters in the pixel space may have small
holes. These holes may result either from an insufficient application of the iterative and sequential filtering step outlined above or simply from a
small feature such as an eye of a bird. To remove
holes, all background (foreground) connected components with radius smaller than a given threshold
R are converted to foreground (background) pixels.
Performing elimination based on component radius
rather than component area is more reliable, because it might be the case that all (or some) of the
components join giving large areas causing the necessary but separate components to be eliminated.
Initial Clustering. The next step is an iterative
application of a clustering algorithm till the number of initial clusters drop below a pre-defined value
Nc . Here, we assume that the number of distinct
colors is less than Nc ; hence, if the number of colors is more, then some color groups are merged to
form bigger motif groups. For clustering, we use
fast and robust mean shift algorithm [43]. It produces clusters based on a given feature space. In
our case, the features are L ∗ a ∗ b∗ channels of the
gamma corrected image. The appealing feature of
this algorithm for us is that it does not require the
A sample result of the image processing module
is depicted in Fig.7.
6
fed to the mean shift. A small and fixed bandwidth,
bc∗ is used. The number of connection groups as
discovered by the mean shift clustering could be
different than the initial n.
In Fig.9, a sample connectivity graph and individual connection groups, discovered by clustering
using the connection length, is depicted.
Figure 7: A sample result for image processing module.
The input ornament image (left) and three masks.
3.2. Local connectivity analysis I: from masks to
connections
The process of local connectivity analysis starts
with consistent keypoint detection on connected
components. An obvious means is to detect centroid of each foreground component. However, this
will be very sensitive to the output of the image
processing module. Separate repeating motifs may
touch one another, for example. Furthermore, this
joining of the motifs may be inconsistent throughout the ornament plane. What is important for the
purpose of further analysis is that the detected keypoints are consistent throughout the pattern. It is,
however, not critical whether they really coincide
with the true motif centers. Hence, we call these
keypoints as nodes.
Towards robustly locating nodes, a continuous labelling on each binary image, namely mask, is performed, yielding a continuous image label∗ , whose
values are in (−1, 1). The labelling stage is detailed at the end of this subsection as a separate
paragraph. The nodes are calculated as the centroids of the positive valued connected components
of the label∗ image. Some sample label images are
depicted in Fig.8. A node is merely a robustly computed keypoint.
Once the nodes are detected, a graph called connectivity graph is constructed via an iterative extraction of local node relations. In the first iteration, the minimal pairwise distance is found, and
then all connections with similar distances are extracted (using a fixed tolerance tol). In the next
iteration, excluding the extracted connections, next
minimal distance is computed to extract new connections. After n iterations, connections of various
sizes are obtained. Note that the connections with
large distances are not much meaningful; at best
they provide redundant information. Hence, it is
better to keep n small. Neither the choice of the
parameter n nor the tolerance tol is critical, because we later re-classify the connections using the
mean shift algorithm.
Given a set of connections stored at the connectivity graph, the connection length as the feature is
Figure 8: Sample label images. Observe that the motif
centers obtain the highest values.
Continuous labelling on the binary masks.
For each mask of each ornament, each pixel p on
each foreground connected component is assigned
an initial label, label(0, p), reflecting whether or not
the probability of the pixel belonging to a centroid
is higher than that of the pixel belonging to the
component boundary. For this purpose, distance
from each foreground pixel to the nearest background pixel is computed, and then those which
have bigger distance than the half of the maximum
of all distances are assigned a positive constant a
where the others −a.
After this label initialization, a relaxation is performed. In the relaxation step, the value is increased (decreased) depending on whether the current label at a pixel p is less (more) than the average of neighboring labels. Formula-wise, relaxation
is expressed as follows:
label(i + 1, p) = label(i, p) + r ∗ relaxation
where r is the relaxation constant which we take as
positive, and the relaxation is
relaxation = Avg (label(i, p))−label(i, p)+label(0, p)
where the Avg (label(i, p)) is computed as the average of 4-neighboors. At convergence or after sufficient number iterations, whichever comes first, the
continuous labelling of binary masks is achieved as
label∗ (p) ← label(K, p)
where K is last performed iteration.
7
a
b
Figure 9: Connectivity graph. (a) Extracted connections for each mask of the ornament, (b) connection groups.
3.3. Local connectivity analysis II: from connections to symmetry elements
The connection groups extracted as described above
are further analyzed in order to detect node relations. The way the nodes are related gives a hint on
various symmetry elements. Recall that the connections with small sizes are favored, since larger connections repeat the same node relations at a larger
scale. Thus, the analyses begin with the connection
group of the smallest size and continue in ascending
order. The analyses are done as following.
Each individual connection group is divided into
connected graphs. Afterwards, each connected
graph is further analyzed independent of the other
connected graphs in the connection group. Given a
connected graph, the following decisions are made.
tation. Since two-node connections might
occur by accident, the two-fold centers are
accepted only if the number of nodes involved in such connection in a connection
group is more than 60% of entire node
number in a mask.
2b If the graph contains more than three
nodes, the polynomial of order one is fit
to the given nodes. For a graph of zigzag
structure the line should pass through the
centers of the edges between two adjacent
nodes. In this case, the distances between
the centers of the edges of two adjacent
nodes and a line are computed. The
probability of the graph being of zigzag
structure is computed by taking the product of the above distance ratios. If the
probability is more than 0.5, then there
is a glide reflection axis passing through
edge centers. On the contrary, if the
nodes themselves (not the edge centers)
lie on the fit polynomial, then this is a
line, representing translational symmetry.
1 If a graph is a cycle graph, i.e., the nodes are
connected in a closed chain, then the probability of the graph being either equilateral triangle, square or regular hexagon is computed
(Fig.10(top row)). The probability is the product of ratios of polygon edges. Thus, if a graph
is constructed from the connection of three
(four, six) nodes, and the probability computed as a product of its edge ratios is higher
than by chance ( i.e., 0.5), the center of the
triangle (square, hexagon) is taken as a center
of three-fold (four-fold, six-fold) rotation.
Detecting implicit node relations. After defining symmetry centers, elimination of repeating
symmetry centers is performed. For example, if
from early connection group a point on a mask has
been marked as a center of four-fold rotation, and
the same point is again marked as four-fold rotation center from subsequent connection group, the
later one is discarded. This means that a symmetry
center of one type is obtained only from one connection group. However, the symmetry center of
one type may coincide with the symmetry centers
of other types. For example, the symmetry center
2 If the graph is acyclic then the nodes might
be related either by two-fold rotation (a graph
containing only two nodes) or glide reflection (a graph containing nodes connected in
a zigzag form) (Fig.10(bottom row)).
2a If the graph contains only two nodes, their
center is taken as the center of two-fold ro8
Figure 10: Examplar connections. Top row illustrates
cyclic graphs, while acyclic ones are shown at the bottom row. Equilateral triangles indicate three-fold rotations,
squares indicate four-fold rotations, and regular hexagons
indicate six-fold rotations. Two-point connections represent
two-fold rotations and zigzag structures indicate glide reflections.
(a)
(b)
(c)
(d)
Figure 11: Implicit node relations. (a) Double two-fold rotation centers, (b) triple two-fold rotation centers, (c) double
three-fold rotation centers, (d) double glide reflections.
marked as four-fold rotation center from some connection group, may also be the symmetry center
of two-fold rotation derived from some other connection group. A point in a mask may represent
multiple symmetry centers of the same type, if and
only if they are derived from the same connection
group. Thus, two three-fold rotation centers (double triangle) at the same point might occur only
when those are detected from the same connection
group. Such repetitions are used to detect implicit
node relations. If there is a double triangle at a
point, this indicates that there is a six-fold rotation
around that point. Similarly, if there is double twofold (triple two-fold) rotation center at the point
this is an indicator of four-fold (six-fold) rotation
around that point if the angle between two-point
lines is 90◦ (60◦ ). Double glide reflection axes represent reflection symmetry which is perpendicular
to the glide reflection axis. Samples of such connections are shown in Fig.11.
Another case is paired two-point connections (see
Fig.12). They too indicate reflection symmetry.
Actual two-fold centers lie at the centers of those
paired lines. Moreover, they are the indicators of reflection axis passing through two-fold rotation centers. The algorithm detects them by computing the
minimal distance between two-fold rotation centers
and connecting nodes that are in this minimal distance from each other. If a graph of maximal degree
four is obtained then these are actual two-fold rotation centers. However, if two-point connections
occur then these are paired two-fold rotation centers and are handled accordingly.
Figure 12: Samples for paired two-fold connections. The
actual two-fold rotational centers lie at the centers of paired
lines.
Refinement. Notice that up to now the symmetries for each mask are detected. After these steps,
the algorithm yields a mask structure which contains fields for various symmetry centers and their
classes. For example, if a mask contains three-fold
rotation centers, the structure contains their centers and classes. The class of symmetry element is
the number of connection group from which it was
extracted. Thus, all three-fold rotational centers
extracted from the same connection group are of
the same class. The next step is to collect all symmetries detected on individual masks. This is a necessary step, since each mask may contain only one
class of some symmetry type, while a whole pattern
contains more classes of that symmetry type. Initially, all symmetries of the same type are collected
without considering their classes. Say, the threefold rotation centers extracted from all masks are
marked on an ornament. Then for each center, the
algorithm defines which classes fall into that particular center. Thus, if three-fold rotation centers of
9
the first mask are of class 2, 5 for mask two and 3 for
mask three, the centers contain classes 2 and 5 (25),
2 and 3 (23), and, 5 and 3 (53). These numbers define the classes of the symmetry centers and they
are classified accordingly into three groups. This
example is illustrated in Fig.13. It might be the
case that the symmetry center is detected only for
one mask, and that center has not been detected for
another masks. Then this center forms a group on
its own, leading to fourth symmetry class. To eliminate such groups, the algorithm counts the number
of centers of each class, sorts them in ascending
order, computes the minimal distance between the
nodes in one class, and beginning with the largest
class propagates its group to the nodes that are of
similar distance. This is done iteratively, until no
changes occur.
Once all the symmetry types are combined and
(a)
(b)
(c)
(d)
roneously, two classes were detected, the class intersecting with the unit cell edges is left, the other
type is eliminated. If the maximal order of rotation
is six and there are more than one classes of threefold rotations, then the one intersecting with the
unit cell centers is left, and the others eliminated.
3.4. Final symmetry group detection
Individual cues in the form of symmetry elements as
described in the previous subsection are integrated
to yield the final symmetry group decision via a decision tree which we propose for the reduced set of
13 groups, The tree is shown in Fig. 14. Comparing
our tree to the classical decision tree in Fig. 5, observe that mirror reflection checks are postponed to
be performed at the last stage and their number is
significantly reduced. Furthermore, with the exception of two-fold rotations case, a mirror reflection
check is performed only when other cues indirectly
imply it. For example, if the maximal order of rotation is six and glide reflections or paired two-fold rotations are detected, then a mirror reflection check
is in order. If the probability of mirror reflection
is calculated to be more than 0.5 then the ornament is classified as belonging to p6m group and
else to p6 group. The reason has been explained
before in the Introduction section: An erroneously
detected mirror reflection is less desirable than a
missed one. In case of a missed mirror reflection,
the fundamental domain will just be twice as big as
it really should be, the half of it being the mirror
of the other half; hence, the whole pattern can be
correctly generated.
There is one exception: The case when two twofold classes are detected. In this case, there are two
possibilities. The first possibility is that the ornament is of group pgg. The second possibility is that
the ornament is a cmm pattern formed by all mirror symmetric protiles. In the latter case, the third
two-fold center for cmm is missed because it happens to be on the glide reflection axis and cannot
be captured from the connection graphs. Hence, a
cmm tile is classified as pgg due to a missed third
two-fold center. This is not tolerable because unlike p6m versus p6 or p4m versus p4, the mirror
reflection is not the only distinction between the
two groups of cmm and pgg, hence, the fundamental domain is not correct up to a mirror. For this
reason, in the case of two two-fold centers, the mirror reflection check is performed even if it is not
implied via indirect cues.
Figure 13: (a-c) An example of symmetries detected for
three masks of an ornament. For each mask one class of
three-fold rotation centers are detected. (d) when the symmetries of all three masks are collected, three classes of threefold rotation centers are obtained.
their classes are determined the maximal order of
rotation is defined. If no rotations are observed
then the maximal order of rotation is taken as 1
and it is checked whether the pattern contains glide
reflections. If no glide reflection observed then the
tile is of p1 group, i.e., contains only translational
symmetry. After defining the maximal order of rotation, further elimination is done using the symmetry group information. Thus, if the maximal
order of rotation is four or six then there can be
only one class of two-fold rotation centers. If, er10
Figure 14: Proposed decision tree.
Mirror Reflection. Mirror reflection checks are
performed on a single unit cell. Note that, in case
of rotational groups of three, four or six fold orders,
the unit cells of the regular and reflectional groups
are identical, the difference being in the fundamental domain. In the case of two-fold rotations, the
unit cell of the pgg group is employed. At any rate,
given a unit cell, all the objects lying in it are extracted using the masks of the ornament pattern.
Then the unit cell is divided into two along the expected reflection axis (based on group information).
For each object lying on both sides, the area (A),
the perimeter (P ), the distance of the object center
to the reflection axis (dRef ), the distance of the
farthest point on the object to the reflection axis
(fdRef) and the point on the reflection axis which
is the closest to the object center (pRef ) are computed. A pair of objects, one from the first part of
the unit cell (obj1) and the other from the second
(obj2), is picked. The probability of the first object
being a reflection symmetry of the second object
is estimated via the product of the corresponding
feature ratios:
prob =
Aobj1
Aobj2
∗
Pobj1
Pobj2
∗ exp(
∗
dRefobj1
dRefobj2
∗
mirror object on the other side. Then, the mean
of all the highest scores in the first part is taken
to indicate with what probability the unit cell is
symmetrical along the given reflection axis. Note
that there might be more than one reflection axes
for a given symmetry group. The final probability
is the mean value computed for all reflection axes.
Unit Cell. Recall that, in all rotational groups,
the corners of the units cells are the points of
maximal order of rotation. With the exception of
p4g and pgg, the unit cell is readily constructed by
connecting the nearest maximal rotation centers of
the same class. For the two groups, p4g and pgg,
however, the maximal rotational symmetry center
appearing on the unit cell corners also appears
on the unit cell center. Therefore, for the latter
two groups, a point pt of certain class is chosen,
then four closest points to pt of similar class are
selected. Among these four points pt1 and pt4 are
selected such that the length of (pt, pt1) equals
to the length of (pt, pt4) and are in opposing
directions. Then the rest two points pt3 and pt2
should also be at equal distance to the point pt
and be in opposing directions. If all points are
exhausted, the algorithm reports a failure.
f dRefobj1
f dRefobj2
−d(pRefobj1 ,pRefobj2 )
)
σ
In all experiments, σ = 10. For each object in
the first part, the above probability is computed
for all its pairs. The highest score is picked. It
indicates with what probability the object has a
If there are no rotations but glide reflections,
then the lattice nodes are centers of two-point
connections of the zigzag structure with similar
directions. To detect a unit cell, a point (pt1)
11
three-fold center (pt2) that is closest to the six-fold
center. A point pt3 is selected so that the length
(pt3, pt2) is equal to the length of (pt1, pt2) and
the angle between these two lines is 120◦ . If no
three-fold centers are found then two-fold centers
are searched. Then the two-fold center (p) closest
to the six-fold center is used to extract the rest of
the fundamental domain. In this case pt2 is a point
which is equal to d(pt1, p) and in opposing direction. This new point is actually a six-fold center.
Then pt3 is found using the first case where two
six-fold centers are known. If no other symmetries
detected except for one six-fold center then algorithm fails to detect the fundamental domain. Note
that it might be the case that two six-fold centers
are found in two different locations so that in between six-fold centers are missed by the algorithm.
In such cases the first case gives large fundamental
domain. If more than one type of symmetries detected so that more than one cases described above
hold, then the fundamental domain is computed for
all of the cases. The final fundamental domain is
the one with the smallest region.
This approach is extended to other symmetry
groups using their own properties.
of particular class is chosen as a first node for
unit cell. Then two closest points (pt2 and pt3)
to pt1 are chosen so that the angle between lines
(pt1, pt2) and (pt1, pt3) is less than or equal to 90◦
and bigger than or equal to 60◦ . The last point pt4
is chosen so that the angle between lines (pt4, pt2)
and (pt4, pt3) is equal to the angle between lines
(pt1, pt2) and (pt1, pt3), and the length of lines
(pt4, pt2) and (pt4, pt3) are equal to lines (pt1, pt3)
and (pt1, pt2), respectively. Recall that all lattice
nodes are of the same class. If no such points
detected, then another point is chosen as pt1 and
the same operations performed according to this
point. If all points have been exhausted, then the
algorithm reports a failure. If an ornament pattern
has neither rotations nor glide reflections then the
unit cell nodes are merely the motif centers.
Fundamental Domain. If the unit cell is successfully constructed, then the fundamental domain
extraction is straightforward, recall Fig. 3. For example, the fundamental domain for a p6 ornament
is 1/6th of its unit cell whereas it is 1/12th for a
p6m ornament. If, however, the algorithm reports a
failure during unit cell construction, then algorithm
returns to the stage before the attempt to construct
a unit cell to make use of all the previously collected
information on individual symmetries. (Note that
failure during unit cell construction may arise due
to lack of sufficient translational repetition.) Let us
explain the fundamental domain extraction step in
case of failure via an example.
Suppose only two centers of six-fold rotation (pt1,
pt2) are found, hence, the unit cell cannot be constructed. Assuming that this two centers are both
nodes of the same unit cell the distance rd between
these two points is computed. Since it is known
that the unit cell for ornaments with six-fold rotations is composed of two equilateral triangles, there
should be a three-fold center (pt3) which is in a
√
3rd distance from the both points.
In order to
√
detect pt3, two circles of radius 3rd centering at
pt1 and pt2 are drawn. There will be two points at
which this two circles intersect. Any one of them
can be selected as pt3. In this manner, three points
indicating nodes of a fundamental domain are computed. For the cases when more than two six-fold
rotations exist, two closest nodes are selected as pt1
and pt2. If, on the other hand, only one six-fold
rotation center (pt1) is found, then three-fold centers are searched. Then the fundamental domain is
extracted using the relation of six-fold center and
4. Experiments
4.1. Data
We formed a labelled ornament data set produced
under different imaging conditions by a variety of
ornament artists including the authors using iOrnament tool [44]. This set is enriched with 14 representative ornament fragments that are equivalent to
p1, p2,pg, pgg, p4, p4g, p3, p3m1, p31m if color permutations are ignored. The total number of ornaments in the set add up to 100. In forming the data
set, we paid attention to cover a variety of styles in
terms of brush, color, tone, and motif choice. We
further paid attention that half of the ornaments
are mimicking Escher’s style with asymmetric interlocking forms and the set contains enough representative elements in each of the 13 groups.
4.2. Parameters
For all 100 ornaments, identical parameter values
are used as detailed below.
Gamma correction. γ is set to 0.5.
12
Clustering. Nc is set to 10. For the refinement
stage, R is set to R = 0.3 ∗ maxR, where maxR is
the maximum radius of the connected components.
For the last two ornaments, respectively in pgg
and pg groups, the collected cues are not sufficient.
In each case, only one of the glide axes are detected.
Hence, these two cases are inconclusive. The reason
for the failure is that it is harder to detect glides in
the absence of sufficient repetition. As we demonstrate in later examples, detecting pg or pgg is possible when there are slightly more samples.
The rest of the results for the remaining 85 tiles
are organized as follows. (Note that the result for
one p6 ornament had been shown previously in the
Introduction.) To save space, 45 samples are placed
in the Appendix. They are split into three figures
each showing the results for 15 patterns. To be
illustrated in this section, 25 illustrative samples
from 8 groups of higher order rotations are selected.
The remaining 15 are selected from the remaining
5 groups. The patterns with higher order rotations
are further organized into two groups. The first one
contains the groups with triangular lattice structures (p6m, p6, p31m, p3m, p3); for these groups
the unit cell consists of two equilateral triangles.
The other contains the groups with square lattice
(p4m, p4g, p4). The results are respectively shown
in Figs. 16 and 17. Original ornaments are shown
in their full sizes whereas the ones depicting the results are cropped in order to make the symmetries,
unit cells and fundamental domains visible.
In all samples of the higher order rotation groups,
the symmetry groups are correctly identified, except up to mirror reflection in some samples. Because a p31m ornament has two three-fold rotation
centers, it is automatically classified correctly without a need for mirror reflection check. In our p6m
examples, the mirror reflections were implied as a
result of detected glides, so the flow in the decision
tree proceeded to mirror reflection check; hence, the
fundamental domains are correctly identified. The
fundamental domain for a p6m ornament is half the
size of a p6 one, and that of a p3m ornament is half
the size of a p3 one. This is because the other halves
can be obtained by mirror reflection. In our p3m
examples, the fundamental domains are double the
size they should be. This is because the samples
are classified as belonging to p3 as a consequence of
missed mirror reflections. Nevertheless, using the
generation the rules for the p3 group, instead of for
the p3m, the original tiles can still be recreated.
As for the two samples of p4m (Fig. 17), the first
one successfully passed the reflection double check
(implied via glides and then checked). But for the
second sample, the mirror reflection test is not per-
Connection extraction. The connectivity graph
is constructed using n = 5 iterations with tol = 5.
The bandwidth fed to mean shift for connection
group clustering is set to bc∗ = 2.
4.3. Results
We first present our result on 14 representative ornaments taken from Escher’s collection (Fig. 15).
For the first 12 out of 14 ornaments, depicted in
Fig. 15 (a) through (l), our method works successfully. In each group, (a) through (l), the first rows
show the original input. The second rows show the
detected symmetries, the unit cells and the fundamental domains superimposed on the input. The
third rows show the fundamental domains cut out
automatically from the original patterns. Unit cells
(in the forms as previously described in Fig. 3) are
shown by red quadrilaterals.
For illustration purposes, the region belonging to
the fundamental domain is made lighter while the
rest of the pattern is made slightly darker. In case
an ornament image is too small to fit a whole unit
cell in it, only the fundamental domain is shown.
Half of the cases (6 out of 12) are like this.
Notice that in some cases, (d)-(f), the letter m
of the group names are in red color. This is to indicate that our method in these three cases missed
the mirror reflection. For none of these three examples, mirror reflection checks at the last stage
are performed because indirect clues did not imply
mirror reflection. As a result, the fundamental domains are twice as big; though, the unit cells are
correct. Recall that in general mirror symmetry
can not cause any problems, since the centers of
the maximal order of rotations do not reside on the
centers of the respective protiles.
For the famous mariposas pattern (Fig. 15 (a))
only three three-fold rotation centers all from different classes are detected. For the second p4
tile (Fig. 15 (h)) only two four-fold rotation centers both from different classes are detected. Yet,
for both patterns we obtain enough information to
identify their symmetry groups and detect fundamental domains. The results in Fig. 15 (h)-(i) show
two-fold rotation centers that are detected in wrong
places. Nevertheless, since the symmetry group depends on the maximal rotation order, incorrect twofold centers do not influence final decision.
13
(a) p3
(b) p3
(c) p31m
(g) p4
(h) p4
(i) p4
(m)
(d) p3m
(j) p4g
(e) p3m
(k) p2
(f) p3m
(l) p1
(n)
Figure 15: Symmetries detected for ornaments painted by Escher. Observe that although most of the images are small and
have few repetitions of the symmetries, we are able to find the fundamental domains. Having insufficient repetition seems to
pose a problem only to two glide group ornaments shown in the last row. The letter m in the group names under the ornaments
are shown in red if the mirror reflection is missed. In those cases, the fundamental domains are double the sizes they should
be.
14
p6
p6
p6m
p6m
p6m
p3
p3
p3
p3
p3
p3m
p3m
p31m
p31m
p31m
Figure 16: Results for the ornaments with six- and three-fold rotations.
formed because indirect clues did not indicate its
existence. Hence, for this particular ornament of
the p4m group, the fundamental domain is double
the size that it should be. Similar to the previously
discussed cases of p3m ornaments, using the generation the rules for the p4 group, instead of for the
p4m, the original p4m pattern can still be recreated.
As for the p4g ornaments, their fundamental do-
mains are accurately identified. In this group of
ornaments, we observed that glide reflections are
easily detected, while for a human it is hard to perceive this type of symmetry. For an ornament to
be classified to p4g group it is enough to detect one
class of four-fold rotation centers. Hence, mirror
reflection check is unnecessary.
25 more results on tiles with higher order rota15
p4
p4
p4g
p4g
p4
p4g
p4m
p4m
p4g
p4g
Figure 17: Symmetries detected for ornaments with four-fold rotations.
p6m
p6m
p6m
p31m
p31m
p31m
p3m
p4m
p4m
p4g
p4g
p4g
p4g
p3m
p31m
p31m
p4m
p4m
Figure 18: Fundamental domains for ornaments with mirror reflections.
16
p3m
p4g
p4g
p2
p2
pgg
pgg
pg
pg
p2
pgg
pg
p2
p2
cmm
cmm
p1
p1
Figure 19: Sample results for ornaments in p2, pgg, cmm, pg, p1 groups.
tions are given in Appendix Figs. A1 and A2. There
are in total 50 ornaments out of 100 with higher
order rotations (excluding Escher ornaments), and
21 of them contain mirror reflections. The detected fundamental domains for these 21 ornaments
are given in Fig. 18. The fundamental domains
of the ornaments for which mirror reflections have
not been detected are framed in black boxes. Ob-
serve that for 5 ornaments, the mirror reflections
are missed. Hence, they are classified as belonging to the corresponding reflection-less groups, and
their fundamental domains are double the sizes they
should be.
Fig. 19 shows results for the samples of the remaining five symmetry groups with lower order rotations: p2, pgg, cmm, pg, p1. The first five orna17
ments belong to p2 group. This group of ornaments
contain four distinct classes of two-fold symmetry
centers. The second and the third groups respectively contain ornaments with two and three distinct classes of two-fold centers. Both of the groups
(pgg and cmm) have two distinct classes of glide
reflection axes perpendicular to each other. The
second sample of cmm group (yellow and purple
ornament) shows a case when the third class of the
two-fold centers is not detected. This is because we
employ two-point connections to indicate any binary connections, i.e., both the two-fold rotations
and the glide fragments forming zigzag structures.
In this particular cmm example, our algorithm detects zigzag structures and defines them as glide
reflection axes. During this process, however, our
algorithm loses track of the two-fold rotations that
are also indicated by the same two-point connections.
In general, we observe that when two-fold rotation centers lie on the glide axes and all the protiles are mirror symmetric (as in the case of the
present cmm sample), our algorithm loses track of
the one of the two-fold rotation centers. When one
of the two-fold rotation centers is lost, the symmetries of the cmm pattern becomes similar to the
symmetries of pgg one except for the mirror reflection symmetry: while the pgg group does not have
mirror reflections, the cmm group does. Thus, if
exactly two two-fold rotation centers detected, we
always need to perform mirror reflection check. For
the present cmm sample, checking for the mirror reflection identifies the correct group. We performed
further tests with 6 additional samples as presented
in the Appendix Fig. A3. With the help of mirror
reflection checks, our algorithm achieves pgg and
cmm separation.
Finally, the last row of Fig. 19 illustrates the results for ornaments without rotational symmetries.
If two distinct classes of glides which are all parallel to each other are detected, then the ornament
is classified as pg. For the p1 group, the group of
pure translation symmetry, the algorithm detects
all grids or lines indicating translations.
tiling, the pattern formed by repeating shapes. As
long as an ornament contains sufficient number of
motifs (protiles) that are either rotationally asymmetric, or strongly concave, or at least less symmetric than the higher order rotational symmetry of the
group, our method works. The asymmetry assumption is not a serious restriction. Even in cases where
all motifs are symmetric, as long as they do not all
centered at the corners of the translational unit,
the method still works. If all motifs are symmetric and centered at the corners of the translational
unit, our method can not determine the symmetry
group. Nevertheless, it is still possible to extract a
translational repetition lattice.
As a proof of concept, to show a range of ornaments for which our method works, we have compiled an ornament database of 100 images. In the
set, 14 of the ornaments are images painted by Escher. All of them are classics, such as the famous
Mariposas, Angles and Damons and Lizards. The
remaining 86 ornament images are constructed via
iOrnament software either by the authors or by several iOrnament artists.
Because we do not explicitly check for the existence of mirror symmetry, unless indirectly implied
by other clues, we sometimes miss mirror reflections. This causes for 3 groups (p6m, p4m, p3m)
to be classified as belonging to respective reflectionless groups of lower symmetry (p6, p4, p3). Nevertheless, since their fundamental domains (double the sizes they should be) contain mirror reflected copies, recreation of the original patterns using the generation rule of the respective reflectionless groups is possible. Indeed, this forms our motivation for postponing the hard mirror reflection
checks till they are implied by indirect clues, such
as glide reflections.
Acknowledgements
The work is funded by TUBITAK Grant 114E204.
References
[1] G. Polya, Über die analogie der kristallsymmetrie in
der ebene, Zeitschrift für Kristallographie 60 (1) (1924)
278–282.
[2] D. Schattschneider, The plane symmetry groups: their
recognition and notation, The American Mathematical
Monthly 85 (6) (1987) 439–450.
[3] D. K. Washburn, D. W. Crowe, Symmetries of culture:
theory and practice of plane pattern analysis, University of Washington Press, 1988.
5. Summary and Conclusion
We have presented a fully automated method to detect the symmetry group and extract fundamental
domains of ornaments belonging to 13 symmetry
groups. We have focused on ornaments where motifs do not hint the symmetries of the underling
18
[4] B. Grünbaum, Z. Grünbaum, G. C. Shepard, Symmetry
in Moorish and other ornaments, Computers and Mathematics with Applications 12 (3-4) (1986) 641–653.
[5] M. Senechal, Color symmetry, Computers and Mathematics with Applications 16 (5-8) (1988) 545–553.
[6] H. S. M. Coxeter, Coloured symmetry, in: M. C. Escher:
Art and Science, North-Holland, 1986, pp. 15–33.
[7] D. Schattschneider, M.C. Escher: Visions of Symmetry,
Thames & Hudson, 2004.
[8] Y. Liu, R. T. Collins, Y. Tsin, Gait sequence analysis using frieze patterns, in: European Conference on
Computer Vision, 2002, pp. 657–671.
[9] H. Y. T. Ngan, G. K. H. Pang, N. H. C. Yung, Motifbased defect detection for patterned fabric, Pattern
Recognition 41 (6) (2008) 1878–1894.
[10] V. Asha, P. Nagabhushan, N. U. Bhajantri, Automatic
extraction of texture-periodicity using superposition of
distance matching functions and their forward differences, Pattern Recognition Letters 33 (5) (2012) 629–
640.
[11] H. Y. T. Ngan, G. K. H. Pang, N. H. C. Yung, Ellipsoidal decision regions for motif-based patterned fabric defect detection, Pattern Recognition 43 (6) (2010)
2132–2144.
[12] C. Sun, J. Sherrah, 3D symmetry detection using the
extended Gaussian image, IEEE Transactions on Pattern Analysis and Machine Intelligence 19 (2) (1997)
164–169.
[13] C. Sun, D. Si, Fast reflectional symmetry detection using orientation histograms, Real-Time Imaging 5 (1)
(1999) 63–74.
[14] Y. Keller, Y. Shkolnisky, A signal processing approach
to symmetry detection, Transactions on Image Processing 15 (8) (2006) 2198–2207.
[15] G. Loy, J. Eklundh, Detecting symmetry and symmetric
constellations of features, in: European Conference on
Computer Vision, 2006, pp. 508–521.
[16] B. Prasad, V. S. N. amd Yegnanarayana, Finding axes
of symmetry from potential fields, IEEE Transactions
on Image Processing 13 (12) (2004) 1559–1566.
[17] V. S. N. Prasad, L. S. Davis, Detecting rotational symmetries, in: IEEE International Conference on Computer Vision, 2005, pp. 954–961.
[18] S. Lee, R. Collins, Y. Liu, Rotation symmetry group
detection via frequency analysis of frieze-expansions,
in: IEEE Conference on Computer Vision and Pattern
Recognition, 2008, pp. 1–8.
[19] S. Lee, Y. Liu, Skewed rotation symmetry group detection, IEEE Transactions on Pattern Analysis and Machine Intelligence 32 (9) (2010) 1659–1672.
[20] J. Liu, Y. Liu, Curved reflection symmetry detection
with self-validation, in: Asian Conference on Computer
Vision, 2010, pp. 102–114.
[21] S. Lee, Y. Liu, Curved glide-reflection symmetry detection, IEEE Transactions on Pattern Analylsis and
Machine Intelligence 34 (2) (2012) 266–278.
[22] J. Liu, Y. Liu, GRASP recurring patterns from a single
view, in: Computer Vision and Pattern Recognition,
2013, pp. 2003–2010.
[23] P. Doubek, J. Matas, Detection of 2D lattice patterns
of repetitive elements and their use for image retrieval,
Tech. Rep. CTU-CMP-2009-16, Department of Cybernetics, Czech Technical University (2009).
[24] P. Doubek, J. Matas, M. Perdoch, O. Chum, Image
matching and retrieval by repetitive patterns, in: Inter-
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
19
national Conference on Pattern Recognition, 2010, pp.
3195–3198.
A. Torii, J. Sivic, T. Pajdla, M. Okutomi, Visual place
recognition with repetitive structures, in: Computer Vision and Pattern Recognition, 2013, pp. 883–890.
J. Gao, Y. Hu, J. Liu, R. Yang, Unsupervised learning of higher-order structural semantics from images,
in: International Conference on Computer Vision, 2009,
pp. 2122–2129.
M. Park, R. T. Collins, Y. Liu, Deformed lattice discovery via efficient mean-shift belief propagation, in:
European Conference on Computer Vision, 2008, pp.
474–485.
M. Park, K. Brocklehurst, R. T. Collins, Y. Liu,
Deformed lattice detection in real-world images using mean-shift belief propagation, IEEE Transactions
on Pattern Analysis and Machine Intelligence 31 (10)
(2009) 1804–1816.
J. Han, S. J. Mckenna, Lattice estimation from images
of patterns that exhibit translational symmetry, Image
and Vision Computing 32 (1) (2014) 64–73.
Y. Liu, R. Collins, Frieze and wallpaper symmetry
groups classification under affine and perspective distortion, Tech. rep., Robotics Institute (1998).
Y. Liu, R. T. Collins, A computational model for repeated pattern perception using frieze and wallpaper
groups, in: Computer Vision and Pattern Recognition,
2000, pp. 537–544.
Y. Liu, R. Collins, Periodic pattern analysis under affine
distortions using wallpaper groups, in: International
Workshop on Algebraic Frames for the Perception Action Cycle, Springer-Verlag Belin, 2000, pp. 241 – 250.
Y. Liu, R. T. Collins, Y. Tsin, A computational model
for periodic pattern perception based on frieze and wallpaper groups, IEEE Transactions on Pattern Analysis
and Machine Intelligence 26 (3) (2004) 354–371.
Y. Liu, R. T. Collins, Skewed symmetry groups, in:
Computer Society Conference on Computer Vision and
Pattern Recognition, 2001, pp. 872–879.
F. Albert, J. M. Gomis, J. Blasco, J. M. Valiente,
N. Aleixos, A new method to analyse mosaics based on
Symmetry Group theory applied to Islamic Geometric
Patterns, Computer Vision and Image Understanding
130 (2015) 54–70.
A. El-Ouaazizi, A. Nasri, R. Benslimane, A rotation
symmetry group detection technique for the characterization of Islamic Rosette Patterns, Pattern Recognition
Letters 68 (P1) (2015) 111–117.
M. Agusti-Melchor, A. Rodas-Jorda, J. M. ValienteGonzalez, Computational framework for symmetry
classification of repetitive patterns, in: International
Conference on Computer Vision Theory and Applications, 2013, pp. 257–270.
M. Agusti-Melchor, J. M. Valiente-Gonzalez, A. RodasJorda, Lattice extraction based on symmetry analysis,
in: International Conference on Computer Vision Theory and Applications, 2008, pp. 396–402.
V. Adanova, S. Tari, Beyond symmetry groups: a
grouping study on Escher’s Euclidean ornaments,
Graphical Models 83 (C) (2016) 15–27.
C. S. Kaplan, D. H. Salesin, Escherization, in: Computer Graphics and Interactive Techniques, 2000, pp.
499–510.
M. von Gagern, J. Richter-Gebert, Hyperbolization of
Euclidean ornaments, The Electronic Journal of Com-
binatorics 16 (2) (2009) 1–29.
[42] J. Dijk, P. W. Verbeek, Lightness filtering in color images with respect to the gamut, in: European Conference on Color in Graphics, Imaging, and Vision, 2006,
pp. 330–335.
[43] K. Fukunaga, L. Hostetler, The estimation of the gradient of a density function, with applications in pattern
recognition, IEEE Transactions on Information Theory
(1975) 32–40.
[44] J. Richter-Gebert, Science-to-touch, http://www.
science-to-touch.com/en/index.html (2012).
Appendix
The Appendix contains the results for the 45 ornaments that are not included in the main sections of
the paper. They are organized according to their
symmetry groups and placed in three figures.
20
p6
p3
p3
p6
p6
p3
p3
p3
p3
p3
p3
p3
p3m
p31m
p31m
Figure A1: Remaining ornaments - 1
21
p4
p4
p4
p4
p4
p4
p4
p4m
p4m
p4g
p2
p2
p2
p2
p2
Figure A2: Remaining ornaments - 2
22
p2
pgg
pgg
pgg
pgg
cmm
cmm
pg
pg
pg
pg
p1
p1
p1
p1
Figure A3: Remaining ornaments - 3
23
| 1cs.CV
|
1
Distributed second order methods with variable
number of working nodes
arXiv:1709.01307v2 [cs.IT] 15 Dec 2017
Nataša Krklec Jerinkić, Dušan Jakovetić, Nataša Krejić, Dragana Bajović
Abstract—Recently, an idling mechanism has been introduced
in the context of distributed first order methods for minimization
of a sum of nodes’ local convex costs over a generic, connected
network. With the idling mechanism, each node i, at each
iteration k, is active – updates its solution estimate and exchanges
messages with its network neighborhood – with probability pk ,
and it stays idle with probability 1 − pk , while the activations
are independent both across nodes and across iterations. In
this paper, we demonstrate that the idling mechanism can be
successfully incorporated in distributed second order methods
also. Specifically, we apply the idling mechanism to the recently
proposed Distributed Quasi Newton method (DQN). We first
show theoretically that, when pk grows to one across iterations
in a controlled manner, DQN with idling exhibits very similar
theoretical convergence and convergence rates properties as
the standard DQN method, thus achieving the same order of
convergence rate (R-linear) as the standard DQN, but with
significantly cheaper updates. Simulation examples confirm the
benefits of incorporating the idling mechanism, demonstrate the
method’s flexibility with respect to the choice of the pk ’s, and
compare the proposed idling method with related algorithms
from the literature.
Index Terms—Distributed optimization, Variable sample
schemes, Second order methods, Newton-like methods, Linear
convergence.
I. I NTRODUCTION
Context and motivation. The problem of distributed minimization of a sum of nodes’ local costs across a (connected)
network has received a significant and growing interest in
the past decade, e.g., [1], [2], [3], [4]. Such problem arises
in various application domains, including wireless sensor
networks, e.g., [5], smart grid, e.g., [6], distributed control
applications, e.g., [7], etc.
In the recent paper [8], a class of novel distributed first
order methods has been proposed, motivated by the so-termed
“hybrid”
Pn methods in [9] for (centralized) minimization of a
sum
i=1 fi (x) of convex component functions. The main
idea underlying a hybrid method in [9] is that it is designed
as a combination of 1) an incremental or stochastic gradient
method (or, more generally, an incremental or stochastic
Newton-like method) and 2) a full (standard) gradient method
Part of results of this paper was presented at the IEEE Global Conference
on Signal and Information Processing (GlobalSIP), Washington, DC, VA,
USA, Dec. 2016. The work of first three authors is supported by the Serbian
Ministry of Education, Science, and Technological Development, Grant no.
174030. The first three authors are with the Department of Mathematics
and Informatics, Faculty of Sciences, University of Novi Sad, Trg Dositeja
Obradovića 4, 21000 Novi Sad, Serbia. The fourth author is with the
Department of Power, Electronics and Communication Engineering, Faculty
of Technical Sciences, University of Novi Sad, Trg Dositeja Obradovića
6, 21000 Novi Sad, Serbia. Authors’ emails: [email protected],
[email protected], [email protected], [email protected].
(or a standard Newton-like method); it behaves as the stochastic method at the initial algorithm stage, and as the standard
method at a later stage. An advantage of the hybrid method
is that it potentially inherits some favorable properties of
both incremental/stochastic and standard methods, while eliminating their important drawbacks. For example, the hybrid
exhibits fast convergence at initial iterations k while having
inexpensive updates, just like incremental/stochastic methods.
On the other hand, it eliminates the oscillatory behavior of
incremental methods around the solution for large k’s (because
for large k’s it behaves as a full/standard method). Hybrid
methods calculate a search direction at iteration k based on a
subset (sample) of the fi ’s, where the sample size is small
at the initial iterations (mimicking a stochastic/incremental
method), while it approaches the full sample n for large k’s
(essentially matching a full, standard method).
With distributed first order methods in [8], the sample size at
iteration k translates into the number of nodes that participate
in the distributed algorithm at k. More precisely, therein we
introduce an idling mechanism where each node in the network
at iteration k is active with probability pk and stays idle
with probability 1 − pk , where pk is nominally increasing
to one with k, while the activations are independent both
across nodes and through iterations. Reference [8] analyzes
convergence rates for a distributed gradient method with the
idling mechanism and demonstrates by simulation that idling
brings significant communication and computational savings.
Contributions. The purpose of this paper is to demonstrate
that the idling mechanism can be incorporated in distributed
second order, i.e., Newton-like methods also, by both 1)
establishing the corresponding convergence rate analytical
results, and 2) showing through simulation examples that
idling continues to bring significant efficiency improvements.
Specifically, we incorporate here the idling mechanism in the
Distributed Quasi Newton method (DQN) [10]. (The DQN
method has been proposed and analyzed in [10], only for the
scenario when all nodes are active at all times.) DQN and its
extension PMM-DQN are representative distributed second order methods that exhibit competitive performance with respect
to the current distributed second order alternatives, e.g., [4],
[11], [16].
Our main results are as follows. We first carry out a
theoretical analysis of the idling-DQN assuming that the fi ’s
are twice continuously differentiable with bounded Hessians.
We show that, as long as pk converges to one at least as fast
as 1 − 1/k 1+ζ (ζ > 0 arbitrarily small), the DQN method with
idling converges in the mean square sense and almost surely to
the same point as the standard DQN method that activates all
2
nodes at all times. Furthermore, when pk converges to one at a
geometric rate, then the DQN algorithm with idling converges
to its limit at a R-linear rate in the mean square sense.
Simulation examples demonstrate that idling can bring to DQN
significant improvements in computational and communication
efficiencies.
We further demonstrate by simulation significant flexibility
of the proposed idling mechanism in terms of tuning of the
activation sequence pk . The simulations show that the idlingDQN method is effective for scenarios when pk increases but
does not eventually converge to one (stays bounded away
from one), even when pk is kept constant across iterations.
The latter two cases are relevant in practice when, due to
node/link failures or asynchrony, the networked nodes do
not have a full control over designing the pk ’s, or when an
increasing sequence of the pk ’s may be difficult to implement.
We also compare by simulation the proposed idling-DQN
method with a very recent distributed second order method
with randomized nodes’ activations in [17]. With constant
activation probabilities, idling-DQN performs very similarly
to [17], while when the activation probabilities are tuned as
proposed here, idling-DQN performs favorably over [17].
From the technical side, extending the analysis of either
distributed gradient methods with idling [8] or DQN without
idling [10] to the scenario considered here is highly nontrivial.
With respect to the standard DQN (without idling), here we
need to cope with inexact variants of the DQN-type second
order search directions. With respect to gradient methods
with idling, showing boundedness of the sequence of iterates
and consequently bounding the “inexactness amounts” of
the search directions are considerably more challenging and
require a different approach.
Brief literature review. There has been a significant
progress in the development of distributed second order methods in past few years. Reference [4] proposes a method based
on a penalty-like interpretation [18] of the problem of interest,
by applying Taylor expansions of the Hessian of the involved
penalty function. In [11], the authors develop a distributed
version of the Newton-Raphson algorithm based on consensus
and time-scale separation principles. References [16] and [19]
propose distributed second order methods based on the alternating direction method of multipliers and the proximal
method of multipliers, respectively. References [20], [30]
develop distributed second order methods for the problem formulations that are related but different than what we consider
in this paper, namely they study network utility maximization
type problems. All the above works [4], [11], [10], [16], [19],
[20], [30] assume that all nodes are active across all iterations,
i.e., they are not concerned with designing nor analyzing
methods with randomized nodes’ activations.
More related with our work are papers that study distributed
first and second order methods with randomized nodes’ or
links’ activations. References [1], [21] consider distributed
first order methods with deterministically or randomly varying communication topologies. Reference [29] proposes a
distributed gossip-based first order method where only two
randomly picked nodes are active at each iteration. The authors
of [22], [23] carry out a comprehensive analysis of (first order)
diffusion methods, e.g., [2], under a very general model of
asynchrony in nodes’ local computations and communications.
Reference [24] proposes a proximal distributed first order
method that provably converges to the exact solution under
a very general model of asynchronous communication and
asynchronous computation. Other relevant works on first order
or alternating direction methods include, e.g., [26], [25].
The authors of [17] propose an asynchronous version of
the (second order) Network Newton method in [4], wherein
a randomly selected node becomes active at a time and
performs a Network Newton-type second order update. The
paper [12], see also [13], [14], proposes and analyzes an
asynchronous distributed quasi-Newton method that is based
on an asynchronous implementation of the BroydenFletcherGoldfarbShanno (BFGS) matrix update.
Among the works discussed above, perhaps the closest to
our randomized activation model are the models studied in,
e.g., [22], [23], [24], [12]. However, these works are still very
different from ours. While they are primarily concerned with
establishing convergence guarantees under various asynchrony
effects that are not in control of the networked nodes, our
aim here is to demonstrate that a carefully designed, nodecontrolled “sparsification” of the workload across network –
as inspired by work [9] from centralized optimization – can
yield significant savings in communication and computation.
Importantly, and as demonstrated in simulations here, significant savings can be achieved even when nodes have only a
partial control over the network-wide workload orchestration,
due to, e.g., asynchrony, link failures, etc.
Finally, in a companion paper [15], we presented a brief
preliminary version of the current paper, wherein a subset of
the results here are presented without proofs. Specifically, [15]
considers convergence of DQN with idling only when the
activation probabilities geometrically converge to one, while
here we also consider the scenarios where the activation
probability converges to one sub-linearly; we also include here
extensions where this parameter stays bounded away from one
or is kept constant (and less than one) across iterations.
Paper organization. Section 2 describes the model that
we assume and gives the necessary preliminaries. Section 3
presents the DQN algorithm with idling, while Section 4
analyzes its convergence and convergence rate. Section 5
considers DQN with idling in the presence of persisting idling,
i.e., it considers the extension when pk does not necessarily
converge to one. Section 6 provides numerical examples.
Finally, we conclude in Section 7.
II. M ODEL AND PRELIMINARIES
Subsection 2.1 gives preliminaries and explains the network
and optimization models that we assume. Subsection 2.2
briefly reviews the DQN algorithm proposed in [10].
A. Optimization–network model
We consider distributed optimization where n nodes in a
connected network solve the following unconstrained problem:
minx∈Rp f (x) =
n
X
i=1
fi (x).
(1)
3
Here, fi : Rp → R is a convex function known only by node i.
We impose the following assumptions on the fi ’s.
Assumption A1. Each function fi : Rp → R, i = 1, . . . , n
is twice continuously differentiable, and there exist constants
0 < µ ≤ L < ∞ such that for every x ∈ Rp
µI ∇2 fi (x) LI.
Here, I denotes the p × p identity matrix, and P Q
(where P and Q are symmetric matrices) means that Q − P is
positive semidefinite. Assumption A1 implies that each fi is
strongly convex with strong convexity parameter µ, and it also
has Lipschitz continuous gradient with Lipschitz constant L,
i.e., for all i = 1, ..., n, for every x, y ∈ Rp , there holds:
µ
fi (y) ≥ fi (x) + ∇fi (x)T (y − x) + kx − yk2
2
k∇fi (x) − ∇fi (y)k ≤ L kx − yk.
Here, notation k · k stands for the Euclidean norm of its vector
argument and the spectral norm of its matrix argument. Under
Assumption A1, problem (1) is solvable and has the unique
solution x∗ ∈ Rp .
Nodes i = 1, ..., n constitute an undirected network G =
(V, E), where V is the set of nodes and E is the set of
edges. Denote by m the total number of (undirected) edges (the
cardinality of E). The presence of edge {i, j} ∈ E means that
the nodes i and j can directly exchange messages through a
communication link. Further, let Oi be the set of all S
neighbors
of a node i (excluding i), and define also Ōi = Oi {i}.
Assumption A2. The network G = (V, E) is connected,
undirected and simple (no self-loops nor multiple links).
We associate with network G a n × n weight matrix that
has the following properties.
Assumption A3. Matrix W = W T ∈ Rn×n is stochastic,
with elements wij such that
wij > 0 if {i, j} ∈ E, wij = 0 if {i, j} ∈
/ E, i 6= j,
X
and wii = 1 −
wij ;
j∈Oi
further, there exist constants wmin and wmax such that for
i = 1, . . . , n, there holds:
0 < wmin ≤ wii ≤ wmax < 1.
np
p
Let x
∈ R , with xi ∈ R , and denote
by Z ∈ R
the Kronecker product of W and the identity
I ∈ Rp×p , Z = W ⊗ I.1 We will make use of the following
penalty reformulation of (1) [18]:
n
X
Wd
Z
Zd
= diag(W ), Wu = W − Wd
(3)
= Z(W ) = Zd + Zu , where
= Wd ⊗ I = diag(Z), and Zu = Wu ⊗ I.
(Here, diag(P ) is the diagonal matrix with the diagonal
elements equal to those of matrix P .)
Further, decompose the Hessian of Φ as:
∇2 Φ(x) = A(x) − G,
with A : R
np
→R
(np)×(np)
(4)
given by:
A(x) = α∇2 F (x) + (1 + θ)(I − Zd ),
(5)
G = G(Z, θ) = Zu + θ(I − Zd ),
(6)
and
for some θ ≥ 0.
We close this subsection with the following result that
will be needed in subsequent analysis. For claim (a), see
Lemma 3.1 in [28]; for claim (b), see, e.g., Lemma 4.2 in
[27].
Lemma II.1 Consider a deterministic sequence {ak } converging to zero, with ak > 0, k = 0, 1, ..., and let ν ∈ (0, 1).
(a) Then, there holds:
Denote by λ1 ≥ . . . ≥ λn the eigenvalues of W. Then,
we have that λ1 = 1, all the remaining eigenvalues of W
are strictly less than one in modulus, and the eigenvector that
corresponds to the unit eigenvalue is e := √1n (1, . . . , 1)T .
= (xT1 , . . . , xTn )T
np×np
Here, α > 0 is a constant that, as we will see ahead, plays
the role of a step size with the distributed algorithms that we
consider. The rationale behind introducing problem (2) and
function Φ is that they enable one to interpret the distributed
first order method in [1] to solve (1) as the (ordinary) gradient
method applied on Φ, which in turn facilitates the development
of second order methods;
[4], for details. Denote by
Tsee, e.g.,
x∗ = (x∗1 )T , ..., (x∗n )T
∈ Rnp the solution to (2), where
x∗i ∈ Rp , for all i = 1, ..., n. It can be shown that, for all
i = 1, ..., n, kx∗i − x∗ k = O(α), i.e., the distance from the
desired solution x∗ of (1) and x∗i is of the order
step size α,
Pof
n
e.g., [4]. Define also F : Rnp → R, F (x) = i=1 fi (xi ).
We study distributed second order algorithms to minimize
function Φ (and hence, find a near optimal solution of (1)).
Therein, the Hessian of function Φ and its splitting into a
diagonal and an off-diagonal part will play an important role.
Specifically, first consider the splitting:
1
minx∈Rnp Φ(x) := α
fi (xi ) + xT (I − Z)x.
2
i=1
(2)
1 Throughout, we shall use “blackboard bold” upper-case letters for matrices
of size (np) × (np) (e.g., Z), and standard upper-case letters for matrices of
size n × n or p × p (e.g., W ).
k
X
ν k−t at−1 → 0 as k → ∞.
(7)
t=1
(b) If, moreover, {ak } converges to zero R-linearly, then the
sum in (7) also converges to zero R-linearly.
B. Algorithm DQN
We will incorporate the idling mechanism in the algorithm
DQN proposed in [10]. The main idea behind DQN is to
approximate the Newton direction with respect to function Φ
in (2) in such a way that distributed implementation is possible,
while the error in approximating the Newton direction is not
large. For completeness, we now briefly review DQN. All
nodes are assumed to be synchronized according to a global
4
clock and perform in parallel iterations k = 0, 1, ... The algorithm maintains iterates xk = ( (xk1 )T , ..., (xkn )T )T ∈ Rnp ,
over iterations k = 0, 1, ..., where xki ∈ Rp plays the role of
the solution estimate of node i, i = 1, ..., n. DQN is presented
in Algorithm 1 below. Therein, Ak := A(xk ), where A(x)
is given in (5); also, notation k · k stands for the Euclidean
norm of its vector argument and spectral norm for its matrix
argument.
Algorithm 1: DQN in vector format
Given x0 ∈ Rnp , ε, ρ, θ, α > 0. Set k = 0.
(1) Chose a diagonal matrix Lk ∈ Rnp×np such that
Algorithm 2: DQN – distributed implementation
At each node i, require x0i ∈ Rp , ε, ρ, θ, α > 0.
(1) Initialization: Each node i sets k = 0 and x0i ∈ Rp .
(2) Each node i transmits xki to all its neighbors j ∈ Oi and
receives xkj from all j ∈ Oi .
(3) Each node i calculates
X
−1
α ∇fi (xki ) +
wij xki − xkj .
dki = Aki
j∈Oi
kLk k ≤ ρ.
(8)
k
sk = −(I − Lk G)A−1
k ∇Φ(x ).
(9)
(2) Set
(3) Set
xk+1 = xk + εsk , k = k + 1.
Lk and Ak . Similarly, Gij is the p × p block at the (i, j)-th
position of G.
(10)
We now briefly comment on the algorithm and the involved
parameters. DQN takes the step in the direction sk scaled
with a positive step size ε > 0; direction sk in (9) is an
approximation of the Newton direction
−1
skN = − ∇2 Φ(xk )
∇Φ(xk ).
Unlike Newton direction skN , direction sk admits an efficient
distributed implementation. Inequality (8) corresponds to a
safeguarding step that is needed to ensure that sk is a descent
direction with respect to function Φ; the nonnegative parameter
ρ controls the safeguarding (see [10] for details on how to set
ρ for a given problem). The diagonal matrix Lk controls the
off-diagonal part of the Hessian inverse approximation [10].
Various choices for Lk that are easy to implement and do
not induce large extra computational and communication costs
have been introduced in [10]. Possible and easy-to-implement
choices are Lk = 0 and Lk = −ρ I.
As it is usually the case with second order methods, step
size ε should be in general strictly smaller than one to ensure
global convergence. However, extensive numerical simulations
on quadratic and logistic losses demonstrate that both DQN
and DQN with idling converge globally with the full step size
ε = 1 and choices Lk = 0, and Lk = −I.
The remaining algorithm parameters are as follows. Quantity θ ≥ 0 controls the splitting (5); reference [10] shows
by simulation that it is usually beneficial to adopt a small
positive value of θ or θ = 0. Finally, α > 0 defines the penalty
function Φ in (2) and results in the following tradeoff in the
performance of DQN: a smaller value of α leads to a better
asymptotic accuracy of the algorithm, while it also slows down
the algorithm’s convergence rate. By asymptotic accuracy, we
assume here the distance between the point of convergence
of DQN x∗ (which is actually equal to the solution of (2) –
see [10]) and the solution x∗ of (1). (As noted before, the
corresponding distance is O(α) [4].)
In Algorithm 2, we present DQN from the perspective of
distributed implementation. Therein, we denote by Λki and Aki ,
respectively, the p×p block on the (i, i)-th position of matrices
(4) Each node i transmits dki to all its neighbors j ∈ Oi and
receives dkj from all j ∈ Oi .
(5) Each node i chooses a diagonal p × p matrix Λki , such
that kΛki k ≤ ρ.
(6) Each node i calculates:
X
ski = −dki + Λki
Gij dkj .
j∈Ōi
(7) Each node i updates its solution estimate as:
xk+1
= xki + ε ski .
i
(8) Set k = k + 1 and go to step 2.
Note that, when Λki ≡ 0, for all i, k, steps 4-6 are skipped,
and the algorithm involves a single communication round per
iteration, i.e., a single transmission of a p-dimensional vector
(step 2) by each node; when Λki ≡ −ρk I, ρk 6= 0, then two
communication rounds (steps 2 and 4) per k are involved.
III. A LGORITHM DQN WITH IDLING
Subsection 3.1 explains the idling mechanism, while Subsection 3.2 incorporates this mechanism in the DQN method.
A. Idling mechanism
We incorporate in DQN the following idling mechanism.
Each node i, at each iteration k, is active with probability pk ,
and it is inactive with probability 1−pk .2 Active nodes perform
updates of their solution estimates (xki )’s and participate in
each communication round of an iteration, while inactive
nodes do not perform any computations nor communications,
i.e., their solution estimates (xki )’s remain unchanged. Denote
by ξik the Bernoulli random variable that governs the activity
of node i at iteration k. Then,
we have that probability
P ξik = 1 = 1 − P ξik = 0 = pk , for all i. We furthermore
assume that ξik and ξj` are mutually independent over all i 6= j
and k 6= `. Throughout the paper, we impose the following
Assumption on sequence {pk }.
Assumption A4. Consider the sequence of activation probabilities {pk }. We assume that pk ≥ pmin , for all k, for some
pmin > 0. Further, {pk } is a non-decreasing sequence with
limk→∞ pk = 1. Moreover, we assume that
0 ≤ uk ≤
Cu
,
(k + 1)1+ζ
(11)
2 We continue to assume that all nodes are synchronized according to a
global iteration counter k = 0, 1, ...
5
where uk = 1 − pk , Cu is a positive constant and ζ > 0 is
arbitrarily small.
Assumption A4 means that, on average, an increasing number of nodes becomes involved in the optimization process;
that is, intuitively, in a sense the precision of the optimization
process increases with the increase of the iteration counter k.
(Extensions to the scenarios when pk does not necessarily
converge to one is provided in Section 5.) We also assume
that pk converges to one sufficiently fast, where sublinear
convergence 1 − 1/k 1+ζ is sufficient.
For future reference, we also define the diagonal (np)×(np)
(random) matrix Yk = diag(ξ1k , . . . , ξnk ) ⊗ I, where I is the
p × p identity matrix. Also, we define the p × p random matrix
k
k k
k
k
k
W
ij ] by wij = wij ξi ξj for i 6= j, and wii = 1 −
P = [w
k
k
k
i6=j wij . Further, we let Z := W ⊗ I, and, analogously to
(3) and (6), we let:
Wdk
(12)
Zk
= Z(W k ) = Zkd + Zku , where
Zkd
= Wdk ⊗ I = diag(Zk ), and Zku = Wuk ⊗ I.
Gk
= G(Zk , θ) = Zku + θ(I − Zkd ).
(13)
P
P
k
k k
Notice that wii = 1 − i6=j wij ξi ξj ≥ 1 − i6=j wij =
wii ≥ wmin . Further, recall A : Rnp → R(np)×(np) in (5).
Using results from [10], we obtain the following important
bounds:3
kA
(x)k
≤
(αµ + (1 + θ)(1 − wmax ))
:= CA , x ∈ R
k
kG k
Gk A−1 (x)
≤
Algorithm 3: DQN with idling – distributed implementation
At each node i, require x0i ∈ Rp , ε, ρ, θ, α > 0, {pk }.
(1) Initialization: Node i sets k = 0 and x0i ∈ Rp .
(2) Each node i generates ξik ; if ξik = 0, node i is idle, and
goes to step 9; else, if ξik = 1, node i is active and goes
to step 3; all active nodes do steps 3-8 below in parallel.
(3) (Active) node i transmits xki to all its active neighbors
j ∈ Oi and receives xkj from all active j ∈ Oi .
(4) Node i calculates
X
α
−1
k
dki = Aki
∇fi (xki ) +
wij
xki − xkj .
pk
j∈Oi
= diag(W k ), Wuk = W k − Wdk
−1
Aki is the p × p block at the (i, i)-th position in Ak , while Gkij
is the p × p block of Gk at the (i, j)-th position.
−1
(14)
np
(1 + θ)(1 − wmin )
:= CG , k = 0, 1, ...
(1 + θ)(1 − wmin )
≤
,
αµ + (1 + θ)(1 − wmin )
x ∈ Rnp , k = 0, 1, ...
(15)
(16)
More generally, for A(x) there holds:
(αµ + (1 + θ)(1 − wmax ))I A(x)
(αL + (1 + θ)(1 − wmin ))I, x ∈ Rnp .
(17)
Also, notice that kYk k ≤ 1, and Zk I, for every k.
B. DQN with idling
We now incorporate the idling mechanism in the DQN
method. To avoid notational
T clutter, we continue to denote by
the algorithm iterates, k = 0, 1, ...,
xk = (xk1 )T , ..., (xkn )T
where xki is node i’s estimate of the solution to (1) at
iteration k. DQN with idling operates as follows. If the
activation variable ξik = 1, node i performs an update; else, if
ξik = 0, node i stays idle and lets xk+1
= xki . The algorithm
i
is presented in Algorithm 3 below. Therein, Ak := A(xk ), and
3 Throughout subsequent analysis, we shall state several relations (equalities
and inequalities) that involve random variables. These relations hold either
surely (for every outcome), or in expectation. E.g., relation (15) holds surely.
It is clear from notation which of the two cases is in force. Also, auxiliary
constants that arise from the analysis will be frequently denoted by the capital
calligraphic letter C with a subscript that indicates a quantity related with the
constant in question; e.g., see CA in (14).
dki
(5) Node i transmits
to all its active neighbors j ∈ Oi
and receives dkj from all the active j ∈ Oi .
(6) Node i chooses a diagonal p × p matrix Λki , such that
kΛki k ≤ ρ.
(7) Node i calculates:
X
ski = −dki + Λki
Gkij dkj .
j∈Ōi
(8) Node i updates its solution estimate as:
xk+1
= xki + ε ski .
i
(9) Set k = k + 1 and go to step 2.
We make a few remarks on Algorithm 3. First, note that,
unlike Algorithm 2, the iterates xki with Algorithm 3 are
random variables. (The initial iterates x0i , i = 1, ..., n, in
Algorithm 3 are assumed deterministic.) Next, note that we
implicitly assume that all nodes have agreed beforehand on
scalar parameters ε, ρ, θ, α; this can actually be achieved in a
distributed way with a low communication and computational
overhead (see Subsection 4.2 in [10]). Nodes also agree
beforehand on the sequence of activation probabilities {pk }.
In other words, sequence {pk } is assumed to be available at all
nodes. For example, as discussed in more detail in Sections 4
and 6, we can let pk = 1−σ k+1 , k = 0, 1, ..., where σ ∈ (0, 1)
is a scalar parameter known by all nodes. As each node is
aware of the global iteration counter k, each node is then able
to implement the latter formula for pk . The nodes’ beforehand
agreement on σ can be achieved similarly to the agreement
on other parameters ε, ρ, θ, α [10]. Tuning of parameter σ is
discussed in Remark 3 further ahead.
Parameters , ρ, and α, and the diagonal matrices Λki play
the same role as in DQN. An important difference with respect
to standard DQN appears in step 4, where the local active
node i’s gradient contribution is pαk ∇fi (xki ) = pαk ξik ∇fi (xki ),
while with standard DQN this contribution equals α ∇fi (xki ).
Note that the division by pk for DQN with idling makes the
terms of the two algorithms balanced on average, because
E[ξik ] = pk .
Using notation (12), we represent DQN with idling in Algo
rithm 4 in a vector format. (Therein, Lk = diag Λk1 , ..., Λkn .)
6
controlled by the activation probability pk . In order to simplify
notation in the analysis, we introduce the following quantities:
Algorithm 4: DQN with idling in vector format
Given x0 ∈ Rnp , ε, ρ, θ, α > 0, {pk }. Set k = 0.
(1) Chose a diagonal matrix Lk ∈ Rnp×np such that
Hk
bk
H
kLk k ≤ ρ.
g
(2) Set
sk = −(I−Lk Gk ) A−1
k
α
Yk ∇F (xk ) + (I − Zk ) xk .
pk
(3) Set
xk+1 = xk + εsk , k = k + 1.
IV. C ONVERGENCE ANALYSIS
In this Section, we carry out convergence and convergence
rate analysis of DQN with idling. We have two main results,
Theorems IV.4 and IV.5. The former result states that, under
Assumptions A1-A4, DQN with idling converges to the solution x∗ of (2) in the mean square sense and almost surely.
We then show that, when activation probability pk converges
to one at a geometric rate, the mean square convergence
towards x∗ occurs at a R-linear rate. Therefore, the order of
convergence (R-linear rate) of the DQN method is preserved
despite the idling. We note that the above result does not
explicitly establish computational and communication savings
with respect to standard DQN. An explicit quantification of
these savings is very challenging even for distributed first
order methods [8], and even more so here. However, the
theoretical results in the current section are complemented
in Section 6 with numerical examples; they demonstrate that
communication and computational savings usually occur in
practice.
The analysis is organized as follows. In Subsection 4.1, we
relate the search direction of DQN with idling and the search
direction of DQN, where the former is viewed as an inexact
version of the latter. Subsection 4.2 establishes the mean
square boundedness of the iterates of DQN with idling and its
implications on the “inexactness” of search directions. Finally,
Subsection 4.3 makes use of the results in Subsections 4.1
and 4.2 to prove the main results on the convergence and
convergence rate of DQN with idling.
A. Quantifying inexactness of search directions
We now analyze how “inexact” is the search direction of the
DQN with idling with respect to the search direction of the
standard DQN. For xk the iterate of DQN with idling, denote
by ŝk the search direction as with the standard DQN evaluated
at xk , i.e.:
k
k
ŝk = −(I − Lk G)A−1
k (α∇F (x ) + (I − Z)x ).
(18)
Then, the search direction sk of DQN with idling can be
viewed as an approximation, i.e., an inexact version, of ŝk .
We will show ahead that the error of this approximation is
k
ĝ k
= I − Lk Gk
= I − Lk G
α
= A−1
Yk ∇F (xk ) + (I − Zk )xk )
k (
pk
k
k
= A−1
k (α∇F (x ) + (I − Z)x ).
b k ĝ k and sk = −Hk g k . Notice that
Therefore, ŝk = −H
k
b k.
kH k ≤ 1 + ρCG := CH and that the same is true for H
We have the following result on the error in approximating ŝk
with sk . In the following, we denote by either ai or [a]i the
i-th p × 1 block of a (np)-dimensional vector a; for example,
we write gik for the i-th p-dimensional block of g k .
In the following Theorem, claims (20) and (21) are from
Theorem 3.2 in [10]; claim (19) is a straightforward generalization of (20), mimicking the proof steps of Theorem 3.2
in [10]; hence, proof details are omitted.
Theorem IV.1 [10] Let assumptions A1-A3 hold. Further, let,
ρ ∈ [0, ρDQN ] where
αµ + (1 + θ)(1 − wmax )
ρDQN =
(1 − wmin )(1 + θ)
1
×
−δ ,
αL + (1 + θ)(1 − wmin )
for some constant δ ∈ (0, 1/(αL + (1 + θ)(1 − wmin ))).
b
b k −1
Consider random matrices Rk := Hk A−1
k and Rk := H Ak .
Then, there holds:
Rk + RTk
λmin
≥ δ
(19)
2
!
bk + R
bT
R
k
λmin
≥ δ,
(20)
2
where λmin (·) denotes the minimal eigenvalue. Moreover,
quantity ŝk in (18) satisfies the following bounds:
∇T Φ(xk )ŝk ≤ −δk∇Φ(xk )k2 and kŝk k ≤ βk∇Φ(xk )k,
(21)
1+ρ(1+θ)(1−wmin )
.
where constant β = αµ+(1+θ)(1−w
max )
Moreover, it was shown in [10] that δ < β, which we use
in the proof of the subsequent result. Furthermore, using the
Mean value theorem and Lipschitz continuity of ∇Φ we obtain
(see the proof of Theorem 3.3 in [10] for instance):
Φ xk + ε sk − Φ(x∗ ) ≤ Φ(xk )
(22)
1
−Φ(x∗ ) + ε2 LΦ ksk k2 + ε∇T Φ(xk )sk , (23)
2
∗
where we recall that x is the (unique) solution of (2) and
LΦ is a Lipschitz gradient continuity parameter of Φ, which
equals LΦ = αL + 2(1 − wmin ). Next, we prove that DQN
with idling exhibits a kind of nonmonotone behavior where the
“nonmonotonicity” term depends on the difference between
the search directions sk and ŝk .
7
Theorem IV.2 Let assumptions A1-A3 hold, ρ ∈ [0, ρDQN ]
and ε ≤ εDQN , where
εDQN =
δ−q
,
2LΦ (β 2 + q 2 )
q ∈ (0, δ).
(24)
comes from the fact that ϕ̂(ε) ≤ ϕ(ε) for every ε > 0.
Finally, taking into account both cases and the fact that ek is
nonnegative, we conclude that the following inequality holds
with ν(ε) ∈ (0, 1) for all ε ∈ (0, εDQN ]
Φ(xk+1 ) − Φ(x∗ ) ≤ (Φ(xk ) − Φ(x∗ ))ν(ε) + ek .
Then
2
Φ(xk+1 ) − Φ(x∗ ) ≤ (Φ(xk ) − Φ(x∗ ))ν(ε) + ek ,
where ν(ε) ∈ (0, 1) is a constant, and ek = (ε2 LΦ +ε/q)ksk −
ŝk k2 .
Proof. We start by considering sk in (22) and using the bounds
from Theorem IV.1, i.e.:
Φ(xk+1 ) − Φ(x∗ ) ≤ Φ(xk ) − Φ(x∗ )
1 2
+
ε LΦ ksk ± ŝk k2 + ε∇T Φ(xk )(sk ± ŝk )
2
≤ Φ(xk ) − Φ(x∗ ) + ε2 LΦ ksk − ŝk k2 + ε2 LΦ kŝk k2 +
+
ε∇T Φ(xk )(sk − ŝk ) + ε∇T Φ(xk )ŝk
≤
Φ(xk ) − Φ(x∗ ) + ε2 LΦ ksk − ŝk k2
+
ε2 LΦ β 2 k∇Φ(xk )k2 +
+
εk∇Φ(xk )kk(sk − ŝk )k − εδk∇Φ(xk )k2
(25)
We distinguish two cases. First, assume that ksk − ŝk k ≤
qk∇Φ(xk )k. In this case, (25) implies that
Φ(xk+1 ) − Φ(x∗ ) ≤ Φ(xk ) − Φ(x∗ ) + ϕ(ε)k∇Φ(xk )k2
where ϕ(ε) = ε2 LΦ (β 2 + q 2 ) − ε(δ − q). Notice that ϕ is
convex, ϕ(0) = 0 and it attains its minimum at εDQN given in
(24) with ϕ(εDQN ) = −((δ − q)2 )/(4LΦ (β 2 + q 2 )). This also
implies that ϕ(ε) is negative for all ε ∈ (0, εDQN ]. Moreover,
Φ is strongly convex and there holds Φ(xk ) − Φ(x∗ ) ≤
1
k 2
where µΦ = αµ is the strong convexity
µΦ k∇Φ(x )k
parameter. Putting all together we obtain
Φ(xk+1 ) − Φ(x∗ ) ≤ (Φ(xk ) − Φ(x∗ ))ν(ε)
(26)
where ν(ε) = 1 + µΦ ϕ(ε). Notice that ν(ε) < 1 for all ε ∈
(0, εDQN ]. Moreover,
ν(ε) ≥ 1 + µΦ ϕ(εDQN ) = 1 −
µΦ (δ − q)2
.
4LΦ (β 2 + q 2 )
2
2
B. Mean square boundedness of the iterates and search directions
We next show that the iterates xk of DQN with idling are
uniformly bounded in the mean square sense. Below, E(·)
denotes the expectation operator.
Lemma IV.1 Let the sequence of random variables {xk } be
generated by Algorithm 4, and let assumptions A1-A4 hold.
Then, there exist positive constants ρ and ε depending on
α, µ, L, θ, wmin , wmax and pmin , such
that, for all ρ ∈ [0, ρ]
and ε ∈ (0, ε], there holds: E kxk k2 ≤ Cx , k = 0, 1, ..., for
some positive constant Cx .
Proof. It suffices to prove that E Φ(xk ) is uniformly
bounded for all k = 0, 1, ..., since Φ is strongly convex and
therefore it holds that Φ(x) ≥ Φ(x∗ )+ µ2Φ kx−x∗ k2 , x ∈ Rnp .
Further, for the sake of proving boundedness, without loss of
generality we can assume that fi (x) ≥ 0, for all x ∈ Rp ,
for every i = 1, ..., n.4 For k = 0, 1, ..., define function
Φk : Rnp → R, by
n
α
α X
1
F (x) + xT (I − Z)x =
fi (xi )
pk
2
pk i=1
X
wij kxi − xj k2 .
Φk (x) =
+
1
2
{i,j}∈E,i<j
Notice that Φk (x) = Φ(x), x ∈ Rnp , if pk = 1. Also note
that, for every k = 0, 1, ..., we have:
Φ(x) ≤ Φk+1 (x) ≤ Φk (x) ≤ Φ0 (x),
2
As µΦ = αµ < αL < LΦ and (δ − q) < δ < β <
β 2 + q 2 we conclude that ν(ε) is positive. Therefore, for all
ε ∈ (0, εDQN ], (26) holds with ν(ε) ∈ (0, 1).
Now, assume that ksk − ŝk k > qk∇Φ(xk )k, i.e.
k∇Φ(xk )k < ksk − ŝk k/q. Together with (25), this implies
Φ(xk+1 ) − Φ(x∗ ) ≤ Φ(xk ) − Φ(x∗ ) + ϕ̂(ε)k∇Φ(xk )k2 + ek
(27)
where ϕ̂(ε) = ε2 LΦ β 2 −εδ and ek = (ε2 LΦ +ε/q)ksk −ŝk k2 .
The function ϕ̂ has similar characteristics as ϕ and it retains
its minimum at ε0 = δ/(2LΦ β 2 ) with ϕ̂(ε0 ) = −δ 2 /(4LΦ β 2 ).
Since εDQN ≤ ε0 , ϕ̂(ε) < 0 holds for all ε ∈ (0, εDQN ].
Again, strong convexity of Φ and (27) imply that for all ε ∈
(0, εDQN ]
since pk is assumed to be non-decreasing.
The core of the
proof isto upper bound Φk+1 xk+1 with a quantity involving
Φk xk (see ahead (34)), and after that to “unwind” the
resulting recursion.
To start, notice that Φk is strongly convex and has Lipschitz
continuous gradient for every k. More precisely, for every k =
0, 1, ..., and for every x ∈ Rnp , we have that
µ̄I ∇2 Φk (x) L̄I,
where µ̄ = αµ and L̄ = αL/pmin + 1. Denote by
α
yk = ∇Φk (xk ) =
∇F (xk ) + (I − Z) xk .
pk
Φ(xk+1 ) − Φ(x∗ ) ≤ (Φ(xk )
−Φ(x∗ ))(1 + µΦ ϕ̂(ε)) + ek ≤ (Φ(xk ) − Φ(x∗ ))ν(ε) + ek ,
where we recall µΦ = αµ, and where the last inequality
4 Otherwise,
since each of the fi ’s is lower bounded, we can re-define
each fi as fbi (x) = fi (x) + c, where c is a constant larger than or equal
maxi=1,...,n |inf x∈Rp fi (x)|, and work with the fbi ’s throughout the proof.
8
Further, let us define two more auxiliary maps, as follows. Let
b k : Rnp × {0, 1}m → R, and Φ
e k : Rnp × {0, 1}m → R, be
Φ
given by:
b k (x; ξ)
Φ
e k (x; ξ)
Φ
=
α
pk
+
1
2
n
X
1
2
k
b k is strongly convex (with respect to x)
If ξmax
= 1 then Φ
with the same parameters as Φk , i.e.,
b k (x) L̄I.
µ̄I ∇2 Φ
X
b k with respect
wij ξi ξj kxi − xj k2Now, denote by x̂∗k = x̂∗k (ξ k ) the minimizer of Φ
b k is nonnegative, and that it has
i=1
{i,j}∈E,i<j
to x; using the fact that Φ
n
a
Lipschitz
continuous
gradient
and is strongly convex, we
X
b k (x) = α
= Φk (x) − Φ
(1 − ξi )fi (xi )
obtain
p
ξi fi (xi ) +
k i=1
X
2
wij (1 − ξi ξj )kxi − xj k ,
{i,j}∈E,i<j
where {0, 1}m denotes the set of all m-dimensional vectors
with the entries from set {0, 1}. Introduce also the short-hand
notation
n
X
b k (x) = Φ
b k (x; ξ k ) = α
Φ
ξ k fi (xi )
pk i=1 i
X
1
wij ξik ξjk kxi − xj k2
+
2
kb
yk k2 ≤ L̄2 kxk −x̂∗k k2 ≤
2
2L̄2 b
b k (x̂∗ )) ≤ 2L̄ Φ
b k (xk ).
(Φk (xk )−Φ
k
µ̄
µ̄
b k (x) ≤ Φk (x), for all x, the previous
Using the fact that Φ
inequality yields
q
p
kb
yk k ≤ L̄ 2/µ̄ Φk (xk ).
(28)
k
On the other hand, if ξmax
= 0 then Yk = 0 and Zk = I
which implies ybk = 0 and the previous inequality obviously
holds.
{i,j}∈E,i<j
e k (x)
Φ
e k (x; ξ k ) = Φk (x) − Φ
b k (x)
= Φ
n
α X
=
(1 − ξik )fi (xi )
pk i=1
X
1
+
wij (1 − ξik ξjk )kxi − xj k2 ,
2
{i,j}∈E,i<j
where we recall that ξ k = (ξ1k , ..., ξnk )T is the node activation
vector at iteration k. Hence, note that, for any fixed x ∈ Rnp ,
b k (x) is a random variable, measurable with respect to the
Φ
σ-algebra generated by ξ k . On the other hand, for any fixed
b k (x) is a (deterministic)
value that variable ξ k takes, x 7→ Φ
function, mapping Rnp to R; analogous observations hold for
b k (xk ), Φ
e k (xk )
Φ̂k as well. We will be interested in quantities Φ
k+1
k+1
b k (x
e k (x
Φ
), and Φ
). Note that they are all random
variables, measurable with respect to the σ-algebra generated
bk
by {ξ s }s=0,1,...,k . We will also work with the gradients of Φ
e k with respect to x, evaluated at xk , that we denote by
and Φ
b k (xk ) =
ybk = ∇Φ
α
Yk ∇F (xk ) + (I − Zk )xk
pk
and
e k (xk ).
ỹk = yk − ybk = ∇Φ
These quantities are also valid random variables, measurable
with respect to the σ-algebra generated by {ξ s }s=0,1,...,k .
Now, recall that, for each fixed i, ξik , k = 0, 1, ..., are
independent identically distributed (i.i.d.) Bernoulli random
variables. The same is true for the minimum and the maximum
k
ξmin
= min ξik ,
i=1,...,n
k
ξmax
= max ξik .
i=1,...,n
k
Also, notice that E(ξmin
) = pnk and
q
k
k
k
1 − ξmin
= 1 − ξmin
, 1 − ξik ≤ 1 − ξmin
,
e k and ỹk . First,
We perform a similar analysis considering Φ
k
e
notice that Φk is also non-negative. Furthermore, if ξmin
=0
e
bk
then Φk is strongly convex with the same parameters as Φ
and the following holds
q
p
e k (xk ).
kỹk k ≤ L̄ 2/µ̄ Φ
Moreover, notice that
e k (xk ) ≤ (1 − ξ k )Φk (xk )
Φ
min
and thus
q
p
k
)Φk (xk )
kỹk k ≤ L̄ 2/µ̄ (1 − ξmin
q
p
k
= (1 − ξmin
)L̄ 2/µ̄ Φk (xk ).
(29)
(30)
Let us return to Φk . This function also satisfies (22), i.e.,
1
Φk xk + ε sk ≤ Φk (xk ) + ε2 L̄ksk k2 + εykT sk . (31)
2
For the search direction sk in step 2 of Algorithm 4, and
recalling Rk = (I−Lk Gk ) A−1
k from Theorem IV.1, we obtain
that
sk = −Rk ybk .
Using the bounds (14) and (15) we conclude that kRk k ≤
(1 + ρCG )CA := CR and therefore
ksk k ≤ CR kb
yk k.
(32)
Also, by Theorem IV.1, the following holds for ε ≤ εDQN , and
ρ ≤ ρDQN (where ρDQN and εDQN are given in Theorems
IV.1 and IV.2, respectively):
Rk + RTk
T
T
ybk Rk ybk = ybk
ybk ≥ δkb
yk k 2 .
(33)
2
k
1 − ξik ξjk ≤ 1 − ξmin
.
b k and ybk , regarded as functions of x ∈ Rnp .
Consider now Φ
From now on, we assume that ε ≤ εDQN , and ρ ≤ ρDQN .
9
the following inequality holds for every k and some positive
constant Cs :
1
Φk (xk+1 ) ≤ Φk (xk ) + ε2 L̄R2 kb
yk k2 − ε(b
yk + ỹk )T Rk ybk
2
E(ksk − ŝk k2 ) ≤ (1 − pk )Cs .
(39)
1
k
2
2
2
≤ Φk (x ) + ε L̄R kb
yk k
(34)
2
Proof.
2
− εδkb
yk k + εCR kb
yk kkỹk k
We first split the error as follows:
1 2 2
k
= Φk (x ) + ( ε L̄CR
E ksk − ŝk k2
2
b k ĝ k + H
b k gk − H
b k g k k2
− εδ)kb
yk k2 + εCR kb
yk kkỹk k
(35)
= E kHk g k − H
2δ
2
− εδ) ≤ 0 for ε ≤ L̄C
Since ( 12 ε2 L̄CR
2 , we conclude that
b k )g k k2
b k k2 k(g k − ĝ k )k2 + E k(Hk − H
≤ 2 E kH
R
Φk (xk+1 ) ≤ Φk (xk ) + εCR kb
yk kkỹk k
≤ 2 (CH )2 E k(ĝ k − g k )k2
2
2εL̄CR
k
b k )g k k2 .
(40)
+E k(Hk − H
≤ Φk (xk ) +
(1 − ξmin
)Φk (xk ),
µ̄
We will estimate the expectations above separately. Start by
for
2δ
observing that:
(36)
, εDQN , ρ ≤ ρDQN ,
ε ≤ min
2
L̄CR
X
wij (xki − xkj )
ĝik − gik = −(Aki )−1 α∇fi (xki ) +
where εDQN and ρDQN are given in Theorems IV.1 and IV.2.
Substituting (32) and (33) into (31) we obtain
2εL̄C 2
j∈Oi
R
Denoting B =
µ̄ , and using the fact that Φk+1 (x) ≤
np
Φk (x), for all x ∈ R , we obtain
k
Φk+1 (xk+1 ) ≤ (1 + B(1 − ξmin
))Φk (xk ).
X
α
+ ξik (Aki )−1 ∇fi (xki ) +
wij ξjk (xki − xkj )
pk
(37)
j∈Oi
ξk
= ( i − 1)α(Aki )−1 ∇fi (xki )
pk
X
+ (Aki )−1
wij (ξik ξjk − 1)(xki − xkj ).
Applying expectation we obtain
E(Φk+1 (xk+1 )) ≤ (1 + B(1 − pnk ))E(Φk (xk )),
k
where we use independence between ξmin
and xk . Furthermore, recall that uk = 1 − pk and notice that 1 − pnk ≤ nuk .
Moreover, 1 + t ≤ et for t > 0 and thus
ξk
Next, by unwinding the recursion, we obtain
Pk−1
j=0
uj
ξk
Notice that (14) implies k( pik − 1)α(Aki )−1 ∇fi (xki )k2 ≤
2( pik −1)2 α2 (CA )2 k∇fi (xki )k2 , with CA defined in (14). Also,
the assumptions on the wij ’s (Assumption A3) and convexity
of the scalar quadratic function V(µ) = µ2 yield:
E(Φk+1 (xk+1 )) ≤ enBuk E(Φk (xk )).
E(Φk (xk )) ≤ enB
j∈Oi
Φ0 (x0 ) := CΦ,2 .
2
X
By assumption, {uk } is summable, and since Φ(x) ≤ Φk (x),
for all x, we conclude that
j∈Oi
2
E(Φ(xk )) ≤ CΦ,2 .
≤
Finally, since Φ is strongly convex, the desired result holds.
2
Notice that an immediate consequence of Lemma IV.1 is
that the gradients are uniformly bounded in the mean square
sense. Indeed,
E k∇F (xk )k2
= E k∇F (xk ) − ∇F (x̃∗ )k2
≤ E L2 kxk − x̃∗ k2
≤ 2L2 (E kxk k2 + kx̃∗ k2 )
≤ 2L2 (Cx + kx̃∗ k2 ) := CF ,
wij (ξik ξjk − 1)(xki − xkj )
(38)
where we recall Cx in Lemma IV.1.
Next, we show that the “inexactness” of the search directions of DQN with idling are “controlled” by the activation
probabilities pk ’s.
Theorem IV.3 Let assumptions A1-A4 hold, and consider ρ
and ε as in Lemma IV.1. Then, for all ρ ∈ [0, ρ] and ε ∈ (0, ε],
X
wij (1 − ξik ξjk )kxki − xkj k
j∈Oi
≤
X
wij (1 − ξik ξjk )2 kxki − xkj k2 .
j∈Oi
Therefore,
ξk
kĝik − gik k2 ≤ 2( i − 1)2 α2 (CA )2 k∇fi (xki )k2
pk
X
2
+2(CA )
wij (1 − ξik ξjk )2 kxki − xkj k2 .
j∈Oi
Applying the expectation and using the fact that ξik ’s are
independent, identically distributed (i.i.d.) across i and across
iterations, the fact that
E((
ξik
− 1)2 ) = (1 − pk )/pk ≤ (1 − pk )/pmin ,
pk
and
E((1 − ξik ξjk )2 ) = 1 − p2k = (1 − pk )(1 + pk ) ≤ 2(1 − pk )
10
we obtain the following inequality
Moreover,
2
α
E(kĝik − gik k2 ) ≤ 2(CA )2 (1 − pk )(
E(k∇fi (xki )k2 )
pmin
X
wij E(kxki − xkj k2 )).
+2
E(kg k k2 )
j∈Oi
Pn
k
k 2
Further, note that E(kĝ k − g k k2 ) =
i=1 E(kĝi − gi k ).
Moreover,
as
a
consequence
of
Lemma
IV.1
we
have
Pn
k 2
k 2
E(k∇f
(x
)k
)
=
E(k∇F
(x
)k
)
≤
C
,
and
i i
F
i=1
n X
X
i=1 j∈Oi
n X
X
≤
and thus
wij E(kxki
−
xkj k2 )
b k )g k k2 ) ≤ (1 − pk )ρ2 4(1 + θ2 )Cg,2 .
E(k(Hk − H
Finally, returning to (40), the previous inequality and (42)
imply
E ksk − ŝk k2 ≤ (1−pk )2((CH )2 Cg +ρ2 Cg,2 ) := (1−pk )Cs .
wij 2 E(kxki k2 + kxkj k2 )
i=1 j∈Oi
n
X
=
α
Yk ∇F (xk ) + (I − Zk )xk )k2 )
pk
α2
≤ (CA )2 2( 2 E(kYk k2 )E(k∇F (xk )k2 )
pmin
+ E(kI − Zk k2 )E(kxk k2 )) )
α2
(44)
≤ (CA )2 2( 2 CF + Cx ) := Cg,2
pmin
= E(kA−1
k (
2(E(
kxki k2
i=1
k 2
X
wij ) + E(
X
kxkj k2
j∈Oi
j∈Oi
k 2
n
X
wij ))
i=1
2
≤ 2(E(kx k + E(kx k ))
≤ 4Cx ,
(41)
C. Main results
The next result – first main result – follows from the
theorems stated above. Let us define constants ρidl =
min{ρ, ρDQN } and εidl = min{ε, εDQN } where ρ and ε are
2
k
k 2
2 α
E(kĝ −g k ) ≤ (1−pk )2(CA ) (
CF +8Cx ) := (1−pk )Cg . as in Theorem IV.3 and ρDQN and εDQN are as in Theorem
pmin
(42) IV.2.
Now, we will estimate the second expectation term in (40).
Again, consider an arbitrary block i of the vector under Theorem IV.4 Let {xk } be the sequence of random variables
expectation. We obtain the following:
generated by Algorithm 4. Further, let assumptions A1-A4
hold. In addition, let ρ ∈ [0, ρidl ] and ε ∈ (0, εidl ]. Then:
k
k k
k k
b
[(H − H )g ]i = [Lk (G − G )g ]i
E Φ(xk+1 ) − Φ(x∗ ) ≤ E Φ(xk ) − Φ(x∗ ) ν + CΦ (1 − pk ),
= [Lk (Zu − Zku + θ(Zkd − Zd ))g k ]i
X
(45)
k
k
= Λki
(wij − wij
)gjk + Λki θ(wii
− wii )gik
where ν ∈ (0, 1) is a constant, and CΦ is a positive constant.
j∈Oi
Moreover, the iterate sequence {xk } converges to the soluX
X
k
k k k
k
k
k
= Λi
wij (1 − ξi ξj )gj + Λi θ
(wij − wij )gi
tion x∗ of (2) in the mean square sense and almost surely.
where Cx is given in Lemma IV.1. Combining the bounds
above, we conclude
j∈Oi
=
Λki
X
j∈Oi
wij (1 −
ξik ξjk )(gjk
+
θgik ).
j∈Oi
Moreover, applying the norm and the convexity argument like
above we get
X
b k )g k ]i k2 ≤ kΛk k2 k
k[(Hk − H
wij (1 − ξik ξjk )(gjk + θgik )k2
i
j∈Oi
≤
ρ
2
X
wij (1 −
ξik ξjk )2 kgjk
+ θgik k2 .
(43)
Proof. Claim (45) follows by taking expectation in Theorem IV.3. The remaining two claims follow similarly to
the proof of Theorem 2 in [8]. We briefly demonstrate the
main arguments for completeness. Namely, unwinding the
recursion (45), we obtain for k = 1, 2, ...:
E(Φ(xk ) − Φ(x∗ )) ≤ Φ(x0 ) − Φ(x∗ ) ν k
(46)
+ CΦ
j∈Oi
k
X
ν k−t (1 − pt−1 ).
t=1
k
b k )g k ]i k2 ) ≤ ρ2 P
Therefore, E(k[(H − H
j∈Oi wij 2(1 −
pk )Ekgjk + θgik k2 ) and using the steps similar to the ones
in (41) we obtain the inequality
b k )g k k2 ) ≤ (1 − pk )ρ2 4(1 + θ2 )E(kg k k2 ).
E(k(Hk − H
Now, we apply Lemma IV.1. From this result and (46), it
follows directly that E Φ(xk ) − Φ(x∗ ) → 0 as k → ∞,
because it is assumed that pk → 1. Furthermore, using
inequality Φ(xk ) − Φ(x∗ ) ≥ µ2Φ kxk − x∗ k2 , the mean square
convergence of xk towards x∗ follows. It remains to show
that xk → x∗ almost surely, as well. Using condition (11),
inequality (46) implies that:
E Φ(xk ) − Φ(x∗ ) ≤ Φ(x0 ) − Φ(x∗ ) ν k (47)
+
CΦ Cu
k
X
ν k−t
t=1
t1+ζ
.
11
It can be shown
that (47) implies (see, e.g., [8]):
E Φ(xk ) − Φ(x∗ ) = O 1/k 1+ζ , which further implies:
E kxk − x∗ k2 = O 1/k 1+ζ .
(48)
Applying the Markov inequality for the random variable kxk −
x∗ k2 , we obtain, for any κ > 0:
1
P kxk − x∗ k2 > κ ≤ E kxk − x∗ k2 = O 1/k 1+ζ .
κ
(49)
Inequality (49) implies that:
∞
X
P kxk − x∗ k2 > κ < ∞,
k=0
and so, by the first Borel-Cantelli
lemma, we get
P kxk − x∗ k2 > κ, infinitely often = 0, which implies
that xk → x∗ , almost surely. 2
Next, we state and prove our second main result.
Theorem IV.5 Let {xk } be the sequence of random variables
generated by Algorithm 4. Further, let assumptions of Theorem IV.4 hold, and let pk = 1 − σ k+1 , with σ ∈ (0, 1). Then,
{xk } converges to the solution x∗ of problem (2) in the mean
square sense at an R-linear rate.
Proof. Denote zk = CΦ (1 − pk ). For the specific choice of
pk we obtain zk = CΦ σ k+1 which obviously converges to zero
R-linearly. Furthermore, repeatedly applying the relation (45)
we obtain
E Φ(xk ) − Φ(x∗ ) ≤ (Φ(x0 ) − Φ(x∗ ))ν k + ak
Pk
where ak = j=1 ν j−1 zk−j . Moreover, it can be shown (see
Lemma II.1) that ak also converges to zero R-linearly which
implies the R-linear convergence of E Φ(xk ) − Φ(x∗ ) . Now,
using the strong convexity of Φ (with strong convexity constant µΦ = αµ), we get: kxk −x∗ k2 ≤ (Φ(xk )−Φ(x∗ ))2/µΦ ,
which in turn implies:
E kxk − x∗ k2 ≤ E Φ(xk ) − Φ(x∗ ) 2/µΦ .
The last inequality means that E kxk − x∗ k2 also converges
to zero R-linearly. 2
Theorem IV.5 shows that the DQN method with idling
converges at an R-linear rate when pk = 1−σ k+1 . Parameter σ
plays an important role in the practical performance of the
method. We recommend the tuning σ = 1 − c α µ, with
c ∈ [30, 80], for α < 1/(80L). 5 The rationale for the
tuning above comes from distributed first order methods with
idling [8], where we showed analytically that it is optimal (in
an appropriate sense) to set σ • = (1 − αµ)2 ≈ 1 − 2αµ. The
value σ • is set to balance 1) the linear convergence factor of
the method without idling and 2) the convergence factor of
the convergence of pk to one. As DQN has a better (smaller)
convergence factor than the distributed first order method (due
to incorporation of the second order information), one has to
adjust the rule 1 − 2αµ, replacing 2 with a larger constant c;
5 Condition α < 1/(80L) is not restrictive, as we observed experimentally
that usually one needs to take an α smaller than 1/(80L) in order to achieve
a satisfactory limiting accuracy.
experimental studies suggest values for c on the order 30-80. In
addition, for very small values of αµ, in order to prevent very
small pk ’s at initial iterations on the one hand and the σ’s very
close to one on the other hand,
we can utilize a “safeguarding”
and modify pk to pk = max p, 1 − (min{σ, σ})k+1 , where
p can be taken, e.g., as 0.2, and σ as 0.9999.
V. E XTENSIONS : DQN UNDER PERSISTING IDLING
This section investigates DQN with idling when activation
probability pk does not converge to one asymptotically, i.e.,
the algorithm is subject to persisting idling. This scenario
is of interest when activation probability pk is not in full
control of the algorithm designer (and the networked nodes
during execution). For example, in applications like wireless
sensor networks, inter-node messages may be lost, due to, e.g.,
random packet dropouts. In addition, an active node may fail
to perform its solution estimate update at a certain iteration,
because the actual calculation may take longer than the time
slot allocated for one iteration, or simply due to unavailability
of sufficient computational resources.
Henceforth, consider the scenario when pk may not converge to one. In other words, regarding Assumption A4, we
only keep the requirement that the sequence {pk } is uniformly
bounded from below. We make here an additional assumption
that the iterates
are bounded in the mean square sense, i.e.,
E kxk k2 is uniformly bounded from above by a positive
constant. Now, consider relation (45); it continues to hold
under the assumptions of the Theorem IV.4, i.e., we have
E Φ(xk ) − Φ(x∗ ) ≤ (Φ(x0 ) − Φ(x∗ ))ν k
+
k
X
ν j−1 CΦ (1 − pk−j ).
j=1
Therefore, using the previous inequality we obtain
CΦ (1 − pmin )
.
E Φ(xk ) − Φ(x∗ ) ≤ (Φ(x0 ) − Φ(x∗ ))ν k +
1−ν
Using strong convexity of Φ (with strong convexity constant
µΦ ) and letting k go to infinity we obtain
2CΦ (1 − pmin )
lim sup E kxk − x∗ k2 ≤
:= E.
µΦ (1 − ν)
k→∞
Therefore, the proposed algorithm converges (in the mean
square sense) to a neighborhood of the solution x∗ of (2).
Hence, an additional limiting error (in addition to the error
due to the difference between the solutions of (1) and (2))
is introduced with respect to the case pk → 1. In order to
analyze further quantity E, we unfold CΦ to get
E = (1 − pmin )h(ε)l(ρ),
where
h(ε) =
4
ε2 LΦ + ε/q
, l(ρ) =
((1 + ρCG )2 Cg + ρ2 Cg,2 );
1 − ν(ε)
µΦ
and ν(ε) is as in the proof of Theorem IV.2. (Recall CG , Cg ,
and Cg,2 in (15), (42), and (44), respectively.) It can be shown
that h(ε) is an increasing function of ε so taking smaller
step size ε brings us closer to the solution. However, the
12
10
0
all nodes working at all iter.
variable num. of working nodes
rel. error
convergence factor ν(ε) is also increasing with ε; thus, there
is a tradeoff between the precision and the convergence rate.
Furthermore, considering l(ρ), we can see that it is also an
increasing function. However, as it is expected, l(0) is strictly
positive. In other words, the error remains positive when the
safeguarding parameter ρ = 0. Finally, the size of the error is
proportional to (1−pmin ) – the closer pmin to one, the smaller
the error. Simulation examples in Section 6 demonstrate that
the error is only moderately increased (with respect to the case
pk → 1), even in the presence of very strong persisting idling.
10
-1
VI. N UMERICAL RESULTS
n
1 X kxi − x∗ k ∗
, x 6= 0,
n i=1 kx∗ k
evolves with the elapsed total number of activations per node.
Note that the number of activations relates directly to both the
communication and computational costs of the algorithm.
The parameters for both algorithms are set in the same way,
and the only difference is in the activation schedule. For the
method with idling, we set pk = 1 − σ k+1 , k = 0, 1, ... We
set σ = 1 − c α µ, with c = 40. (Clearly, for the method
without idling, pk ≡ 1, for all k.) The remaining algorithm
parameters are as follows. We set α = 1/(100 L), where L
is the Lipschitz constant of the gradients of the fi ’s that we
take as maxi=1,...,n kAi k. Further, we let θ = 0, and = 1
(full step size). We consider two choices for Λki in step 6
of Algorithm 3: Λki = 0, for all i, k; and Λki = −I, for all
i, k. (We apply no safeguarding on the above choices of Λki ,
0
100
200
300
400
total cost (# activations per node)
(a)
10
0
variable num. of working nodes
all nodes working at all iter.
rel. error
This section demonstrates by simulation significant computational and communication savings incurred through the
idling mechanism within DQN. It also shows that persisting
idling (pk not converging to one) induces only a moderate
additional limiting error, i.e., the method continues to converge
to a solution neighborhood even under persisting idling.
We consider the problem with strongly convex local
quadratic costs; that is, for each i = 1, ..., n, we let fi : Rp →
R, fi (x) = 12 (x − bi )T Ai (x − bi ), p = 10, where bi ∈ Rp
and Ai ∈ Rp×p is a symmetric positive definite matrix.
The data pairs Ai , bi are generated at random, independently
across nodes, as follows. Each bi ’s entry is generated mutually
independently from the uniform distribution on [1, 31]. Each
Bi is generated as Bi = Qi Di QTi ; here, Qi is the matrix of
bi + B
b T ), and B
bi is a matrix
orthonormal eigenvectors of 21 (B
i
with independent, identically distributed (i.i.d.) standard Gaussian entries; and Di is a diagonal matrix with the diagonal
entries drawn in an i.i.d. fashion from the uniform distribution
on [1, 31].
The network is a n = 100-node instance of the random
geometric
graph model with the communication radius r =
q
ln(n)
n , and it is connected. The weight matrix W is set as
1
follows: for {i, j} ∈ E, i 6= j, wij = 2(1+max{d
, where
i ,dj })
di is the node
i’s
degree;
for
{i,
j}
∈
/
E,
i
=
6
j,
w
=
0; and
ij
P
wii = 1 − j6=i wij , for all i = 1, ..., n.
We compare the standard DQN method and the DQN
method with incorporated idling mechanism. Specifically, we
study how the relative error (averaged across nodes):
10
-1
0
50
100
150
200
250
300
total cost (# activations per node)
(b)
Fig. 1. Relative error versus total cost (number of activations per node) for
quadratic costs and n = 100-node network; Figure (a): Λki = 0; Figure (b):
Λki = −I. The solid lines correspond to all nodes working at all iterations;
the dashed lines correspond to the method with idling (increasing number of
working nodes).
i.e., we let ρ = 1.) Note that the former choice corresponds
to the algorithms with a single communication round per
iteration k, while the latter corresponds to the algorithms with
two communication rounds per k.
Figure 1 (a) plots the relative error versus total cost (equal to
total number of activations per node up to the current iteration)
for one sample path realization, for Λki = 0. We can see
that incorporating the idling mechanism significantly improves
the efficiency of the algorithm: for the method to achieve the
limiting accuracy of approximately 0.025, the method without
idling takes about 410 activations per node, while the method
with idling takes about 310 activations. Hence, the idling
mechanism reduces total cost by approximately 24%. Figure 1
(b) repeats the plots for Λki = −I, still showing clear gains of
idling, though smaller than with Λki = 0.
To account for randomness of the DQN method with idling
that arises due to the random nodes’ activation schedule,
13
14
10
0
pk = 1-k+1
12
all nodes working at all iterations
variable num. of working nodes
pk = 0.7 ( 1-k+1 )
rel. error
10
8
6
all nodes working at all iter.
pk = 0.7 = const
10
-1
4
2
0
180
10
200
220
240
260
total cost to reach rel. error 0.04
-2
280
0
(a)
10
0
pk = 1 - k+1
pk = 1 = const
variable num. of working nodes
pk = 0.7 ( 1 - k+1 )
rel. error
10
8
6
1000
(a)
14
12
200
400
600
800
total cost (# activations per node)
all nodes working all iter.
pk = 0.5 ( 1 - k+1 )
10
-1
pk = 0.9 ( 1 - k+1 )
4
2
0
140
10
160
180
total cost to reach rel. error 0.04
200
-2
0
200
400
600
800
total cost (# activations per node)
1000
(b)
(b)
Fig. 2. Total cost (number of activations per node) to reach relative error 0.04
for quadratic costs and n = 100-node network; Figure (a): Λki = 0;
Figure (b): Λki = −I. The histograms corresponds to the DQN algorithm
with idling; the arrow indicates the total cost needed by standard DQN.
Fig. 3. Relative error versus total cost (number of activations per node) for
strongly convex quadratic costs, n = 40-node network, and Λki = 0. The
Figures compare the following scenarios: 1)
pk ≡ 1 (standard DQN); 2)
pk = 1 − σ k+1 ; 3) pk = pmax 1 − σ k+1 ; and 4) pk = pmax , for all k.
we include histograms of the total cost needed to achieve a
fixed level of relative error. Specifically, in Figure 2 we plot
histograms of the total cost (corresponding to 50 generated
sample paths – 50 different realizations of the ξik ’s along
iterations) needed to reach the relative error equal 0.04;
Figure 2 (a) corresponds to Λki = 0, while Figure 2 (b)
corresponds to Λki = −I. The Figures also indicate with
arrows the total cost needed by standard DQN to achieve the
same accuracy. The results confirm the gains of idling. Also,
the variability of total cost across different sample paths is
small relative to the gain with respect to standard DQN.
Figures 3 and 4 investigate the scenarios when the activation
probability pk may not asymptotically converge to one. The
network is a (connected)
q random geometric graph instance
with n = 40 and r = ln(n)
n ; step size α = 1/(200L); the
remaining system and algorithmic parameters are the same
as with the previous simulation example. We consider the
following choices for pk : 1) pk ≡ 1 (standard DQN); 2)
pk = 1 − σ k+1 ; 3) pk = pmax 1 − σ k+1 ; and 4) pk = pmax ,
for all k. With the third and fourth choices, the presence of
pmax models external effects on the pk (out of control of the
networked nodes), e.g., due to link failures and unavailability
of computing resources at certain iterations; it is varied within
the set {0.5, 0.7, 0.9}. Figure 3 (a) compares the methods for
one sample path realization with the four choices of pk above,
with pmax = 0.7, for Λki = 0. Figure 3 (b) compares for the
same experiment the standard DQN (pk ≡ 1), pk = 1 − σ k+1 ,
and pk = pmax 1 − σ k+1 , with pmax ∈ {0.5, 0.7, 0.9}.
Several important observations stand out from the experiments.
First, we can see that the limiting error increases when pk
does not converge to one with respect to the case when it
converges to one. However, this increase (deterioration) is
moderate, and the algorithm still manages to converge to a
good solution neighborhood despite the persisting idling. In
14
10
0
pk = 1-k+1
pk = 1 = const
rel. error
pk = 0.7 ( 1-k+1 )
10
10
pk = 0.7 = const
-1
-2
0
200
400
600
total cost (# activations per node)
(a)
(a)
10
0
pk = 1 - k+1
pk = 1= const
rel. error
pk = 0.9 ( 1 - k+1 )
10
10
pk = 0.7 ( 1 - k+1 )
-1
pk = 0.5 ( 1 - k+1 )
-2
0
200
400
600
total cost (# activations per node)
(b)
(b)
Fig. 4. Relative error versus total cost (number of activations per node) for
strongly convex quadratic costs, n = 40-node network, and Λki = −I. The
Figures compare the following scenarios: 1)
pk ≡ 1 (standard DQN); 2)
pk = 1 − σ k+1 ; 3) pk = pmax 1 − σ k+1 ; and 4) pk = pmax , for all k.
Fig. 5. Relative error versus total cost (number of activations per node)
for strongly convex quadratic costs and n = 30-node network. Figure (a):
α = 1/(100 L); Figure (b): α = 1/(200 L). The red, dotted line corresponds
to the proposed DQN with idling and pk = 1−σ k+1 ; blue, solid line to [17];
green, dashed line to the proposed DQN with idling and pk = 0.8 = const;
and black, dashed line to the method in [29].
particular, from Figure 3 (b), we can see that the limiting
relative error increases from about 10−2 (with pk → 1) to
about 3.5 · 10−2 with pmax = 0.5–a case with a strong
persisting idling. This corroborates that DQN with idling is
an effective method even when activation probability pk is
not in full control of the algorithm designer. Second, the
limiting error decreases when pmax increases, as it is expected
– see Figure 3 (b). Finally, from Figure 3 (a),
we can
see that the method with pk = pmax 1 − σ k+1 performs
significantly better than the method with pk = pmax , for all
k. In particular, the methods have the same limiting error
(approximately 2 · 10−2 ), while the former approaches this
error much faster. This confirms that the proposed judicious
design of increasing pk ’s, as opposed to just keeping them
constant, significantly improves the algorithm performance.
Figure 4 repeats the same experiment for Λki = −I. We can
see that the analogous conclusions can be drawn.
In the next experiment, we compare the proposed DQN
method with idling with other existing methods that utilize
randomized activations of nodes. Specifically, we consider
the very recent asynchronous (second order) network Newton
method proposed in [17] that is an asynchronous version of the
method in [4]. We refer to this method here as asynchronous
Network Newton (NN). We also consider the first order gossipbased method in [29]. The method in [17] and the DQN with
idling and matrix L = −I both utilize two p-dimensional
communications per node activation (they have equal communication cost per node activation). The two methods also have
a similar computational cost per activation. The method in [29]
has a twice cheaper communication cost per activation, and it
has in general a lower computational cost per activation (due
to incorporating only the first order information into updates).
The comparison is carried out on a n = 30-node (con-
15
nected) random geometric graph instance with 91 links, for the
variable dimension p = 5 and strongly convex quadratic fi ’s
generated analogously to the previous experiments. With the
proposed idling-DQN, we consider two choices of activation
probabilities: 1) pk = 1 − σ k+1 , with σ = 1 − c α µ, c = 40,
as in the previous experiment; and 2) pk = 0.8 = const.
The weight matrices of the proposed method and that in [17]
are set in the same way, as in prior experiments. With both
idling-DQN and the method in [17], we set step size = 1. We
consider two different choices of parameter α: α = 1/(100 L),
and α = 1/(200 L). This is how step-sizes are set for the
method in [17] and for the idling-DQN with pk = 1 − σ k+1 .
Quantity α for the method in [29] (the algorithm’s step size)
and for the idling-DQN with pk = 0.8 are then adjusted
(decreased) for a fair comparison, so that the four different
methods achieve the same asymptotic relative error; we then
look how many per-node activations each method takes to
reach the saturating relative error.
Figure 5 plots the relative error versus total number of
activations for the four methods; Figure (a) corresponds to
α = 1/(100 L) and Figure (b) is for α = 1/(200 L). We
can see from Figure 5 (a) that the proposed idling-DQN with
pk = 1 − σ k+1 outperforms the other methods: it takes about
400 per-node activations to reach the relative error 0.025;
the method in [17] and the idling-DQN with pk = 0.8
take about 600; and the method in [29] needs at least 1600
activations for the same accuracy. Note that, even when we
half the number of activations for [29] to account for its
twice cheaper communication cost, the proposed idling-DQN
is still significantly faster – it compares versus [29] as 400
versus 800 “normalized activations.” Interestingly, the idling
DQN with constant pk and the method in [17] practically
match in performance. Figure 5 (b) repeats the comparison for
α = 1/(200 L); we can see that similar conclusions can be
drawn from this experiment. In summary, the idling-DQN reduces communication cost with respect to the method in [29],
which is expected as it utilizes more (second order) computations at each activation. The two second order methods
with randomized activations, the idling-DQN and the method
in [17], exhibit very similar performance when the idlingDQN uses constant pk policy. With the increasing pk policy,
the idling-DQN performs better than [17]. This, together with
previous experiments, demonstrates that a carefully designed
workload orchestration with idling-DQN leads to performance
improvements both with respect to a “pure” random activation
policy and with respect to the all-nodes-work-all-time policy.
VII. C ONCLUSION
We incorporated an idling mechanism, recently proposed
in the context of distributed first order methods [8], into
distributed second order methods. Specifically, we study the
DQN algorithm [10] with idling. We showed that, as long
as pk converges to one at least as fast as 1/k 1+ζ , ζ > 0
arbitrarily small, the DQN algorithm with idling converges in
the mean square sense and almost surely to the same point
as the standard DQN method that activates all nodes at all
iterations. Furthermore, when pk grows to one at a geometric
rate, DQN with idling converges at a R-linear rate in the mean
square sense. Therefore, DQN with idling achieves the same
order of convergence (R-linear) as standard DQN, but with
significantly cheaper iterations. Simulation examples corroborate communication and computational savings incurred by
incorporating the idling mechanism and show the method’s
flexibility with respect to the choice of activation probabilities.
The proposed idling-DQN method, and also other existing
distributed second order methods (involving local Hessian’s
computations) with randomized nodes’ activations, e.g., [17],
[12], are not exact in the sense that they converge to a solution
neighborhood. An interesting future research direction is to
develop and analyze an idling-based second order method with
exact convergence.
R EFERENCES
[1] A. Nedić, A. Ozdaglar, “Distributed subgradient methods for multi-agent
optimization,” IEEE Transactions on Automatic Control, vol. 54, no. 1,
pp. 48–61, 2009.
[2] F. Cattivelli, A. H. Sayed, “Diffusion LMS strategies for distributed
estimation,” IEEE Transactions on Signal Processing, vol. 58, no. 3,
pp. 1035–1048, 2010.
[3] W. Shi, Q. Ling, G. Wu, W. Yin, “EXTRA: an Exact First-Order
Algorithm for Decentralized Consensus Optimization,” SIAM Journal
on Optimization, vol. 2, no. 25, pp. 944-966, 2015.
[4] A. Mokhtari, Q. Ling, A. Ribeiro, “Network Newton Distributed Optimization Methods,” IEEE Trans. Signal Processing, vol. 65, no. 1, pp.
146-161, 2017.
[5] L. Xiao, S. Boyd, S. Lall, “A scheme for robust distributed sensor fusion
based on average consensus,” in IPSN ’05, Information Processing in
Sensor Networks, 63–70, Los Angeles, California, 2005.
[6] G. Hug, S. Kar, C. Wu, “Consensus + Innovations Approach for
Distributed Multiagent Coordination in a Microgrid,” IEEE Trans. Smart
Grid, vol. 6, no. 4, pp. 1893-1903, 2015.
[7] F. Bullo, J. Cortes, S. Martnez, “Distributed Control of Robotic Networks: A Mathematical Approach to Motion Coordination Algorithms,”
Princeton University Press, 2009.
[8] D. Bajović, D. Jakovetić, N. Krejić, N. Krklec Jerinkić, “Distributed
Gradient Methods with Variable Number of Working Nodes,” IEEE
Trans. Signal Processing, vol. 64, no. 15, pp. 4080-4095, 2016.
[9] M. P. Friedlander, M. Schmidt, “Hybrid deterministic-stochastic methods for data fitting,” SIAM Journal on Scientific Computing, vol. 34,
pp. 1380-1405, 2012.
[10] D. Bajović, D. Jakovetić, N. Krejić, N. Krklec Jerinkić, “Newton-like
method with diagonal correction for distributed optimization,” SIAM J.
Opt., vol. 27, no. 2, pp. 11711203. 2017.
[11] D. Varagnolo, F. Zanella, A. Cenedese, G. Pillonetto, and L. Schenato,
“Newton-Raphson Consensus for Distributed Convex Optimization,”
IEEE Trans. Aut. Contr., vol. 61, no. 4, 2016.
[12] M. Eisen, A. Mokhtari, A. Ribeiro, “Decentralized quasi-Newton methods,” IEEE Transactions on Signal Processing, vol. 65, no. 10, pp. 2613–
2628, May 2017.
[13] M. Eisen, A. Mokhtari, A. Ribeiro, “A Decentralized Quasi-Newton
Method for Dual Formulations of Consensus Optimization,” IEEE 55th
Conference on Decision and Control (CDC), Las Vegas, NV, Dec. 2016.
[14] M. Eisen, A. Mokhtari, A. Ribeiro, “An asynchronous quasi-Newton
method for consensus optimization,” IEEE Global Conference on Signal
and Information Processing (GlobalSIP), Washington, DC, VA, USA,
Dec. 2016.
[15] D. Bajović, D. Jakovetić, N. Krejić, N. Krklec Jerinkić, “Distributed
first and second order methods with variable number of working
nodes,” IEEE Global Conference on Signal and Information Processing,
Washington DC, VA, USA, Dec. 2016
[16] A. Mokhtari, W. Shi, Q. Ling, A. Ribeiro, “DQM: Decentralized
Quadratically Approximated Alternating Direction Method of Multipliers,” IEEE Trans. Signal Processing, vol. 64, no. 19, pp. 5158-5173,
2016.
[17] F. Mansoori, E. Wei, “Superlinearly Convergent Asynchronous
Distributed Network Newton Method,” 2017, available at
https://arxiv.org/abs/1705.03952
16
[18] D. Jakovetić, J. M. F. Moura, J. Xavier, “Distributed Nesterov-like
gradient algorithms”, CDC’12, 51st IEEE Conference on Decision and
Control, Maui, Hawaii, December 2012, pp. 5459–5464.
[19] A. Mokhtari, W. Shi, Q. Ling, A. Ribeiro, “A Decentralized Second
Order Method with Exact Linear Convergence Rate for Consensus
Optimization,” IEEE Trans. Signal and Information Processing over
Networks, vol. 2, no. 4. pp. 507-522, 2016.
[20] M. Zargham, A. Ribeiro, A. Jadbabaie, “Accelerated dual descent for
constrained convex network flow optimization,” Decision and Control
(CDC), 2013 IEEE 52nd Annual Conference on, Firenze, Italy, 2013.
pp. 1037-1042.
[21] I. Lobel, A. Ozdaglar, D. Feijer, “Distributed Multi-agent Optimization
with State-Dependent Communication,” Mathematical Programming,
vol. 129, no. 2, pp. 255-284, 2014.
[22] X. Zhao, A. H. Sayed, “Asynchronous Adaptation and Learning Over
Networks–Part I: Modeling and Stability Analysis,” IEEE Transactions
on Signal Processing, vol. 63, no. 4, Feb. 2015.
[23] X. Zhao, A. H. Sayed, “Asynchronous Adaptation and Learning Over
Networks–Part II: Performance Analysis,” IEEE Transactions on Signal
Processing, vol. 63, no. 4, Feb. 2015.
[24] T. Wu, K. Yuan, Q. Ling, W. Yin, A. H, Sayed, “Decentralized Consensus Optimization with Asynchrony and Delays,” IEEE Transactions
on Signal and Information Processing over Networks, to appear, 2017,
DOI: 10.1109/TSIPN.2017.2695121
[25] J. Liu, S. Wright, “Asynchronous stochastic coordinate descent: Parallelism and convergence properties,” SIAM J. Opt., vol. 25, no. 1, 2015.
[26] T. H. Chang, L. Wei-Cheng, M. Hong, X. Wang, “Distributed ADMM
for large-scale optimization part II: Linear convergence analysis and
numerical performance,” IEEE Trans. Sig. Proc., vol. 64, no. 12, 2016.
[27] N. Krejić, N. Krklec Jerinkić, “Nonmonotone line search methods with
variable sample size,” Numerical Algorithms, vol. 68, pp. 711–739, 2015.
[28] S. Ram, A. Nedic, V. Veeravalli, “Distributed stochastic subgradient
projection algorithms for convex optimization,” J. Optim. Theory Appl.,
vol. 147, no. 3, pp. 516-545, 2011.
[29] S. S. Ram, A. Nedić, V. Veeravalli, “Asynchronous gossip algorithms for
stochastic optimization,” CDC ’09, 48th IEEE International Conference
on Decision and Control, Shanghai, China, December 2009, pp. 3581 –
3586.
[30] E. Wei, A. Ozdaglar, A. Jadbabaie, “A distributed Newton method for
network utility maximization–I: Algorithm,” IEEE Transactions on
Automatic Control, vol. 58, no. 9, pp. 2162- 2175, 2013.
| 7cs.IT
|
Leveraging Diversity and Sparsity in Blind Deconvolution
Ali Ahmed and Laurent Demanet∗
arXiv:1610.06098v2 [cs.IT] 15 Dec 2017
December 18, 2017
Abstract
This paper considers recovering L-dimensional vectors w, and x1 , x2 , . . . , xN from their
circular convolutions yn = w ∗ xn , n = 1, 2, 3, . . . , N . The vector w is assumed to be S-sparse
in a known basis that is spread out in the Fourier domain, and each input xn is a member of a
known K-dimensional random subspace.
We prove that whenever K + S log2 S . L/ log4 (LN ), the problem can be solved effectively
by using only the nuclear-norm minimization as the convex relaxation, as long as the inputs are
sufficiently diverse and obey N & log2 (LN ). By “diverse inputs”, we mean that the xn ’s belong
to different, generic subspaces. To our knowledge, this is the first theoretical result on blind
deconvolution where the subspace to which w belongs is not fixed, but needs to be determined.
We discuss the result in the context of multipath channel estimation in wireless communications. Both the fading coefficients, and the delays in the channel impulse response w are
unknown. The encoder codes the K-dimensional message vectors randomly and then transmits
coded messages xn ’s over a fixed channel one after the other. The decoder then discovers all of
the messages and the channel response when the number of samples taken for each received message are roughly greater than (K + S log2 S) log4 (LN ), and the number of messages is roughly
at least log2 (LN ).
1
Introduction
This paper addresses the problem of recovering a vector w from its circular convolutions individually
with a series of unknown vectors {xn }n := x1 , x2 , . . . , xN . Consider a linear, time-invariant (LTI)
system, characterized by some unknown impulse response w. The system is driven by a series of
inputs {xn }n and one wants to identify the system by observing only the outputs, which in this case
are the convolutions of the inputs with the system impulse response. This problem is referred to as
the blind system identification: jointly discover the inputs, and the system impulse response from
the outputs, and is one of the core problems in the field of system theory and signal processing.
∗
Ali Ahmed is currently with the Information Technology University, Lahore, Pakistan. He was associated until
recently with the Department of Mathematics, MIT, Cambridge, MA. Laurent Demanet is with the Department
of Mathematics, MIT, Cambridge, MA. Email for the corresponding author: [email protected]. Both authors are
sponsored by AFOSR grants FA9550-12-1-0328 and FA9550-15-1-0078. LD is also funded by NSF, ONR, and Total
SA. We thank Augustin Cosse for interesting discussions. Some preliminary results in this direction were presented
in an earlier conference publication, namely, A convex approach to blind deconvolution with diverse inputs, in Proc.
IEEE CAMSAP, Cancun, December 2015. c This work has been submitted to the IEEE Transactions on Information
Theory for possible publication. Copyright may be transferred without notice, after which this version may no longer
be accessible.
1
DRAFT – December 18, 2017 – 2:26
When w is expected to be sparse, the problem can be recast in a now standard fashion as the
recovery of a simultaneously sparse and rank-1 matrix. We relax this formulation by dropping the
sparsity contraint and using nuclear-norm minimization.
We then leverage results in the well understood area of low-rank recovery from underdetermined
systems of equations to give the conditions on the unknown impulse response, and inputs under
which they can be deconvolved exactly.
Roughly, the results say the input vectors x1 , x2 , . . . , xN , each of which lives in some known
“generic” K-dimensional subspace of RL , and a vector w ∈ RL that is “incoherent” in the Fourier
domain, and is only assumed to be S-sparse is some known basis, are separable with high probability provided K + S ∼ L, up to log factors, and with appropriate coherences appearing in the
constants.
More precisely, we state the problem as follows. Assume that each of the input xn lives in a known
K-dimensional subspace of RL , i.e.,
xn = Cn mn ,
n = 1, 2, 3, . . . , N
(1)
for some L×K basis matrix Cn with K ≤ L, whose columns span the subspace in which xn resides.
Moreover, the vector w is only assumed to be S-sparse in an L × L basis matrix B, i.e.,
w = Bh, such that khk0 ≤ S.
(2)
(It can be convenient to think of B as the identity upon first reading.) Given the basis matrices
B and Cn , all we need to know are the expansion coefficients mn , and h to discover the inputs
xn for each n, and w. The structural assumptions on w are much weaker than on xn , in that we
only need w to be sparse in some known basis, whereas each xn resides in a generic and known
subspace.
We observe the circular convolutions:
yn = w ∗ xn , n = 1, 2, 3, . . . , N.
(3)
An entry in the length-L observation vector yn , for each n ∈ {1, 2, 3, . . . , N } is
yn [`] =
L
X
`0 =1
w[`0 ]xn [` − `0 + 1 mod L], (`, n) ∈ {1, 2, 3, . . . , L} × {1, 2, 3, . . . , N },
where modulo L is what makes the convolution circular. Given no information about the inputs xn
and the impulse response w, it is clear that both of these quantities cannot be uniquely identified
from the observations (3).
We want to put this result in perspective from the outset by comparing it with a related result
in [2], where a single (N = 1) input blind deconvolution problem is analyzed. Mathematically, the
main result in [2] shows that in a single input deconvolution problem, the vectors x, and w can
be recovered from the circular convolution y = w ∗ x when x lives in a known generic subspace as
above, however, unlike above the incoherent vector w also lives in a known subspace. In this paper,
we do not have a known subspace assumption on w, which makes it a significant improvement over
the results in [2] and has concrete implications in important applications as will be explained in
Section 1.4
2
DRAFT – December 18, 2017 – 2:26
1.1
Notations
We use upper, and lower case bold letters for matrices and vectors, respectively. Scalars are
represented by upper, and lower case, non-bold letters. The notation x∗ (xT ) denotes a row vector
formed by taking the transpose with (without) conjugation of a column vector x. By x̄, we mean a
column vector obtained from x by conjugating each entry. Linear operators are represented using
script letters. We repeatedly use the notation k ∼K n to indicate that the index k takes value in
the range {(n − 1)K + 1, . . . , nK} for some scalar K. We use [N ] to denote the set {1, 2, 3, . . . , N }.
The notation IK denotes K × K identity matrix for a scalar K. For a set Ω ⊂ [L], IL×Ω denotes an
L × |Ω| submatrix of an L × L identity obtained by selecting columns indexed by the set Ω. Also
we use Dn to represent a KN × KN matrix IK ⊗ en e∗n with ones along the diagonal at locations
k ∼K n and zeros elsewhere, where {en }n denote standard N -dimensional basis vectors, and ⊗
is the conventional Kronecker product. We write vec(A) for the vector formed by stacking the
columns of a matrix A. Given two matrices A, and B, we denote by A B, the rank-1 matrix:
[vec(A)][vec(B)]∗ . Similarly, P : CL×M → CL×M for some L, and M takes an L × M matrix X
∗
to IL×Ω IL×Ω
X. We will use L ∼ a → b to show that a variable L varies between scalar a, and b.
Lastly, the operator E refers to the expectation operator, and P represents the probability measure.
1.2
Lifting and convex relaxation
In this section, we recast the blind system identification from diverse inputs as a simultaneously
sparse, and rank-1 matrix recovery problem, and set up a semidefinite program (SDP) to solve it.
Begin with defining F , the L × L discrete Fourier transform (DFT) matrix,
1
F [ω, `] = √ e−j2π(ω−1)(`−1)/L ,
L
(ω, `) ∈ [L] × [L],
(4)
and let f`∗ denote the `th row of F . In the Fourier domain, the convolutions in (3) are
√
√
ŷn = Lŵ x̂n , or ŷn [`] = Lhf` , wihf` , xn i, (`, n) ∈ [L] × [N ],
where ŵ = F w, x̂n = F xn , and
denotes the Hadamard product. Using the fact that xn =
Cn mn , and w = Bh, we obtain
√
ŷn [`] = LhB ∗ f` , hihCn∗ f` , mn i = hb` , hihmn , c`,n i, (`, n) ∈ [L] × [N ],
where the last equality follows by substituting b` = B ∗ f` , c̄`,n =
hx, yi = hy, xi∗ . This can be equivalently expressed as
ŷn [`] = hb` c∗`,n , hm∗n i = hb` φ∗`,n , hm∗ i,
√
LCn∗ f` , and using the fact that
(`, n) ∈ [L] × [N ],
(5)
where with matrices as its arguments, the notation h·, ·i denotes the usual trace inner product,
m = [m∗1 , m∗2 , . . . , m∗N ]∗ , and φ`,n denotes a length KN vector of zeros except the c`,n in the
position indexed by k ∼K n, i.e.,
φ`,n = c`,n ⊗ en ,
(`, n) ∈ [L] × [N ]
(6)
with en denoting the standard N -dimensional basis vectors. It is clear that the measurements
are non-linear in h ∈ RL and m ∈ RKN but are linear in their outer product X0 = hm∗ .
3
DRAFT – December 18, 2017 – 2:26
Since the expansion coefficients h are S-sparse, this shows that the inverse problem in (3) can
be thought of as the question of recovering hm∗ ; a rank-1 matrix with S-sparse columns, from
its linear measurements obtained by trace inner products against known measurement matrices
A`,n = b` φ∗`,n .
Define a linear map A : RL×KN → CLN as
A(X) : = {hb` φ∗`,n , Xi | (`, n) ∈ [L] × [N ]}
(7)
The number of unknowns in X are LKN and there are only LN linear measurements available. This
means that the linear map A is severely underdetermined except in the trivial case when K = 1. In
all other cases when K > 1, infinitely many candidate solutions satisfy the measurements constraint
owing to the null space of A.
Of course, we can take advantage of the fact that the unknown matrix will always be simultaneously
sparse, and rank-1 and hence the inherent dimension is much smaller. Information theoretically
speaking, the number of unknowns is only ∼ S log L + KN , and if we can effectively solve for the
simultaneously, sparse and rank-1 matrices then inverting the system of equations for xn ’s, and w
might be possible for a suitable linear map A when LN & S log L + KN . If it were possible, a
single unknown input (N = 1) under certain structural assumptions would suffice to identify the
system completely.
However, it is only known how to individually relax the low-rank and sparse structures [7, 9, 10,
14, 24], namely using nuclear and `1 norms, but it remains an open question to efficiently relax
those structures simultaneously.1 Instead, if we ignore the sparsity altogether and only cater to the
rank-1 structure, the problem remains in principle solvable because the inherent number L+KN of
unknowns in this case become smaller than the number LN of observations as soon as the number
L
N of inputs exceeds L−K
> 1. Therefore, the main idea of this paper is to use multiple inputs,
which allow us to forego the use of a sparsity penalty in the relaxed program.
Before we formulate the optimization program, it is worth mentioning that the recovery of the
rank-1 matrix X0 only guarantees the recovery of h and m to within a global scaling factor α, i.e.,
we can only recover h̃ = αh, and m̃ = α−1 m, which is not of much concern in practice.
The inverse problem in (3) can be cast into a rank-1 matrix recovery problem from linear measurements as follows:
find
X
subject to ŷn [`] = hb` φ∗`,n , Xi,
rank(X) = 1.
(`, n) ∈ [L] × [N ]
The optimization program is non convex and in general NP hard due to the combinatorial rank
constraint. Owing to the vast literature [7,13,15,24] on solving optimization programs of the above
1
The most natural choice of combining the nuclear and `1 norms to constitute a convex penalty for simultaneously
sparse and low-rank is known to be suboptimal [22]. In fact, for the special case of rank-1, and sparse matrices, no
effective convex relaxation exists [1]. Thus, even if the low-rank and sparse structures can be individually handled
with effective convex relaxations, no obvious convex penalty is known for the simultaneously spase, and low-rank
structure.
4
DRAFT – December 18, 2017 – 2:26
form, it is well known that a good convex relaxation is
X̂ := argmin
X
kXk∗
(8)
subject to ŷn [`] = hb` φ∗`,n , Xi,
(`, n) ∈ [L] × [N ],
where the nuclear norm kXk∗ is the sum of the singular values of X. The system identification
problem is successfully solved if we can guarantee that the minimizer to the above convex program
equals hm∗ . Low-rank recovery from under determined linear map has been of interest lately in
several areas of science and engineering, and a growing literature [7, 15, 23, 24] has been concerned
with finding the properties of the linear map A under which we can expect to obtain the true
solution after solving the above optimization program.
1.3
Main results
In this section, we state the main result claiming that the optimization program in (8) can recover
the sparse, and rank-1 matrix hm∗ almost always when the inputs xn ’s reside in relatively dense
“generic” K dimensional subspaces of RL , and that w ∈ RL satisfies the nominal conditions of
S-sparsity in some known basis, and “incoherence” in the Fourier domain. Before stating our main
theorem, we define the terms “generic” and “incoherence” concretely below. Recall that w = Bh.
The incoherence of the basis B̂ = F B introduced in (2) is quantified using a coherence parameter
µ2max ,
µ2max := L · kB̂k2∞ ,
(9)
where k · k∞ is the entrywise uniform norm. Using the fact that B̂ is an L × L orthonormal matrix,
it is easy to see that 1 ≤ µ2max ≤ L. A simple example of a matrix that achieves minimum µ2max
would be the DFT matrix.
The incoherence of w in the Fourier domain is measured by µ20 ,
(
)
00 k2
0 k2
2
k
B̂h
k
B̂h
0 ∞
k
B̂hk
n,p
∞
n,n
∞
µ20 := L · max
,
,
,
khk22
kh0n,p k22
kh00n,n0 k22
where each ratio is a measure of diffusion in the Fourier domain. The spirit of the definition is
2
∞
mainly captured by the first term, L · kB̂hk
— scaled peak value of w in the Fourier domain.
khk2
2
The other terms involve quantities h0n,p and h00n,n0 that are defined in the sequel (they are random
perturbations of h), and are only present for technical reasons. Notice that the first term is small,
O(1), when w is diffuse in the frequency domain, and can otherwise be as large as L.
To keep our results as general as possible, we introduce an extra incoherence parameter ρ20 that
quantifies the distribution of energy among the inputs {mn }n , and is defined as
ρ20 := N · max
n
kmn k22
,
kmk22
(10)
which is bounded as 1 ≤ ρ20 ≤ N . The coherence ρ20 achieves the lower bound when the energy is
equally distributed among the inputs, and the upper bound is attained when all of the energy is
localized in one of the inputs, and the rest of them are all zero.
5
DRAFT – December 18, 2017 – 2:26
As mentioned earlier, we want each of the inputs to reside in some “generic” K-dimensional subspace, which we realize by choosing Cn ’s to be iid Gaussian matrices, i.e.,
Cn [`, k] ∼ Normal 0, L1 ∀(`, n, k) ∈ [L] × [N ] × [K].
(11)
A “generic” K-dimensional subspace refers to most of the K-dimensional subspaces in the entire
continuum of K-dimensional subspaces of RL , however, one must also be mindful that such generic
subspaces may not arise naturally in applications, and may have to be introduced by design as will
be demonstrated in a stylized channel-estimation application in Section 1.4.
√
Ultimately, we are working with the rows c`,n ’s of the matrix LF Cn as defined in (5). As the
columns of Cn are real and F is an orthonormal matrix, the columns of F Cn are also Gaussian
vectors with a conjugate symmetry. Hence, the rows c`,n are distributed as2
(
Normal(0, I)
` = 1, (L/2) + 1, n ∈ [N ]
c`,n =
(12)
−1/2
−1/2
Normal(0, 2
I) + jNormal(0, 2
I) ` = 2, . . . , (L/2), n ∈ [N ]
c`,n = c̄L−`+2,n , ` = (L/2) + 2, . . . , L, n ∈ [N ].
Note that the vectors c`,n ’s are independently instantiated for every n ∈ [N ]. On the other hand,
the vectors c`,n are no longer independent for every ` ∈ [L], rather the independence is retained
only for ` ∈ {1, 2, . . . , L/2 + 1}. However, the c`,n ’s are still uncorrelated for ∀` ∈ [L]; a fact which
is crucial in the analysis to follow later. We are now ready to state the main result.
2
2
Theorem 1. Suppose the bases {Cn }N
n=1 are constructed as in (12) , and the coherences µmax , µ0 ,
2
N
and ρ0 of the basis matrix B, and the expansion coefficients h, and {mn }n=1 are as defined above.
Furthermore, to ease the notation, set
α1 = log(K log(LN )),
and
α2 = log(S log(LN )).
Then for a fixed β ≥ 4, there exists a constant Cβ0 = O(β), such that if
max(µ20 α1 K, µ2max Sα2 log2 S) ≤
L
Cβ0 α1 log2 (LN )
and N ≥ Cβ0 ρ20 α1 log(LN ),
then X0 = h(m∗1 , m∗2 , . . . , m∗N ) is the unique solution to (8) with probability at least 1−O((LN )4−β ),
and we can recover N inputs {xn }N
n=1 and w (within a scalar multiple) from N convolutions
N
{yn = w ∗ xn }n=1 .
The result above crudely says that in an L dimensional space, an incoherent vector, S-sparse
in some known basis, can be separated successfully almost always from N vectors (with equal
energy distribution) lying in known random subspaces of dimension K whenever K + S log2 S .
L/ log4 (LN ), and N & log2 (LN ).
1.4
Application: Blind channel estimation using random codes
A stylized application of the blind system identification directly arises in multipath channel estimation in wireless communications. The problem is illustrated in Figure 1. A sequence of length-K
2
The construction in (12) is explicitly for even L but can be easily adapted to the case when L is odd.
6
DRAFT – December 18, 2017 – 2:26
messages m1 , m2 , . . . , mN are coded using taller L × K coding matrices C1 , C2 , . . . , CN , respectively. The coded messages xn = Cn mn , n ∈ 1, 2, 3, . . . , N are then transmitted one after the
other over an unknown multipath channel, characterized by a sparse impulse response w ∈ RL .
The transmitted message xn arrives at the receiver through multiple paths. Each path introduces
its own delay and fading. All the delayed and scaled copies of xn overlap in the free space communication medium. The received signal is modeled as the convolution of xn with w. This action
is repeated with same delay and fading coefficients for every xn . In other words, we are assuming
here that the channel’s impulse response is more or less fixed over the duration of the transmission
of these N coded messages, which justifies the use of a fixed impulse response w in each of the
convolutions. The task of the decoder is to discover both the impulse response and the messages
by observing their convolutions yn = w ∗ xn , n = 1, 2, 3, . . . , N , and using the knowledge of the
coding matrices.
Our main result in Theorem 1 took w as a vector that is sparse in some incoherent basis B. In the
application discussed in the last paragraph, we can simply take the basis B to be the standard basis;
perfectly incoherent. The location of each non-zero entry in w depicts the delay in the arrival time
of a copy of coded message at the receiver from a certain path and the value of the entry known
as the fading coefficient incorporates the attenuation and the phase change encountered in that
path. The coherence parameter µ20 is roughly just the peak value of the normalized frequency in
the spectrum of the channel response. For this particular application, we can assume that ρ20 ≈ 1
as the transmitter energy is equally distributed among the message signals. Our results prove
that if each of the message is coded using a random coding matrix, and the channel response has
approximately a flat spectrum, then we can recover the messages and the channel response jointly
almost always by solving (8), whenever the length K of the messages, the sparsity S of the channel
impulse response w, and the codeword length L obey K + S log2 S . L/ log4 (LN ), and the number
N of messages that convolve with the same instantiation of the channel roughly exceed log2 (LN ).
Our results here can be thought of as an extension to the blind deconvolution result that appeared
in [2], where we only have a one-time look at the unknown channel — we observe only a single
convolution of the impulse response with a randomly coded message. Consequently, only fading
coefficients could be resolved in [2] and not the delays in the impulse response w of the channel. In
other words, one needs to know the subspace or support of w in advance. In general, both fading
coefficient, and delays are equally important pieces of information to decipher the received message
in wireless communications. In this paper, we take advantage of several looks at the same channel
as it remains fixed during the transmission of N messages. This enables us to estimate both the
fading coefficients and the unknown delays at the same time. In general, we do not assume that
the vector w lives in a known subspace as was the case in [2].
1.5
Related work
In a nutshell, and to our knowledge, this paper is the first in the literature to theoretically deal
with an impulse response that belongs to a low-dimensional subspace that is not fixed ahead of
time but needs to be discovered.
The lifting strategy to linearize the bilinear blind deconvolution problem was proposed in [2],
and it was rigorously shown that two convolved vectors in RL can be separated blindly if their
K, and S dimensional subspaces are known and one of the subspace is generic and the other is
incoherent in the Fourier domain. It is further shown using the dual certificate approach in low7
DRAFT – December 18, 2017 – 2:26
x1
x2
..
.
..
.
C N mN
yN
.
.
y2
..
..
.
..
.
..
.
xN
=
y1
C 1 m1
C 2 m2
..
=
{y n }n
channel
.
x1
x2
{xn }n
..
encode
{mn }n
xN
decode
{m̃n }n
h̃
⇤
h
Figure 1: Blind channel estimation. Each meassage mn in the block {mn }n = {m1 , m2 , . . . , mN } of
messages is coded with a corresponding tall coding matrix Cn and the block of coded messages {xn }n is
sequentially transmitted over an arbitrary unknown channel. This results in convolution of each of the coded
messages {xn }n with an unknown impulse response h. The decoder receives the convolutions {yn }n and it
discovers both the messages {mn }n and the unknown channel h (within a global scalar).
rank matrix recovery literature [7, 15, 23] that both the vectors can be deconvolved exactly when
(K + S) . L/ log3 L. This paper extends the single input blind deconvolution result to multiple
diverse inputs, where we observe the convolutions of N vectors with known subspaces with a fixed
vector only known to be sparse in some known basis.
A natural question that arises is whether multiple (N > 1) inputs xn ’s are necessary in our problem
to identify w in (2). The answer is no in this specific case as even in the single input case N = 1,
under the same random subspace assumption on x1 , and replacing the nuclear norm in (8) with the
standard `1 norm (sum of absolute entries) will separate x1 , and w, however, the sample complexity
L will be suboptimal, and of the order of SK to within log factors. In the general single input case;
under no random subspace assumption on x1 , it is shown in [12] that w, and x1 are not identifiable
from y = w ∗ x1 .
A related question, in a sense dual to that presented in the previous section, is multichannel
blind deconvolution. See Figure 2. In discrete time this problem can be modeled as follows. An
unknown noise source w ∈ RL feeds N unknown multipath channels characterized by K-sparse
impulse responses xn ∈ RL , n = 1, 2, 3, . . . , N . The receiver at each channel observes several
delayed copies of w overlapped with each other, which amounts to observing the convolutions
yn = w ∗ xn , n = 1, 2, 3, . . . , N . The noise w can be modeled as a Gaussian vector, and is well
dispersed in the frequency domain, i.e., the vector w is incoherent according to the definition
(26). The fading coefficients of the multipath channels are unknown, however, we assume that
the delays are known. This amounts to knowing the subspace of the channels and the unknown
impulse responses can be expressed as xn = Cn mn for every n ∈ [N ], where the columns of the
known L × K coding matrices are now the trivial basis vectors and mn contain K unknown fading
coefficients in each channel. The indices (delays) of the non-zeros of every impulse response xn can
be modeled as random, in which case the coding matrices are composed of the random subset of
the columns of the identity matrix. With the coding matrix known and random, the multichannel
blind deconvolution problem is in spirit the dual of the blind system identification from diverse
inputs presented in this paper, where the roles of the channel and the source signal are reversed.
However, the results in Theorem 1 are explicitly derived for dense Gaussian coding matrices and
8
DRAFT – December 18, 2017 – 2:26
not for random sparse matrices. It is worth mentioning here that in many practical situations
the non zeros in the channel impulse response are concentrated in the top few indices making the
assumption of known subspaces (delays) plausible.
After [2], a series of results on blind deconvolution appeared under different sets of assumptions on
the inputs. For example, the result in [3] considers an image debluring problem, where the receiver
observes the N subsampled circular convolutions of an L-dimensional image x, modulated with
random binary waveforms, with an L-dimensional bandpass blur kernel h that lives in a known
K-dimensional subspace. Then it is possible to recover both the image and a incoherent blur kernel
using lifting and nuclear norm minimization, whenever N & log3 L log log K, where N/L is also
the number of subsampling factor of each convolution. The result shows that it is possible to
deconvolve two unknown vectors by observing multiple convolutions—each time one of the vectors
is randomly modulated and is convolved with the other vector living in a known subspace. We are
also observing multiple convolutions but one of the vectors in the convolved pair is changing every
time and the subspace of the other is also unknown, this makes our result much broader.
Another relevant result is blind deconvolution plus demixing [20], where one observes sum of N
different convolved pairs of L-dimensional vectors lying in K, and S dimensional known subspaces;
one of which is generic and the other is incoherent in the Fourier domain. Each generic basis is
chosen independently of others. The blind deconvolution plus the demixing problem is again cast
as a rank-N matrix recovery problem. The algorithm is successful when N 2 (K + S) . L/ log4 L.
An important recent article from the same group settles the recovery guarantee for a regularized
gradient descent algorithm for blind deconvolution, in the single-input case and with the scaling
K + S . L/ log2 L [19]. This result, however, makes the assumption of a fixed subspace for the
sparse impulse response. Note that gradient descent algorithms are expected to have much more
favorable runtimes than semidefinite programming, when their basin of attraction can be established
to be wide enough, as in [19].
The multichannel blind deconvolution was first modeled as a rank-1 recovery problem in [25] and the
experimental results show the successful joint recovery of Gaussian channel responses with known
support that are fed with a single Gaussian noise source. Other interesting works include [16, 33],
where a least squares method is proposed. The approach is deterministic in the sense that the input
statistics are not assumed to be known though the channel subspaces are known. Some of the results
with various assumptions on input statistics can be found in [29]. Owing to the importance of the
blind deconvolution problem, an expansive literature is available and the discussion here cannot
possibly cover all the related material, however, an interested reader might start with the some
nice survey articles [18, 21, 28] and the references therein.
It is also worth mentioning here a related line of research in the phase recovery problem from
phaseless measurements [6, 11], which happen to be quadratic in the unknowns. As in bilinear
problems, it is also possible to lift the quadratic phase recovery problem to a higher dimensional
space, and solve for a positive-definite matrix with minimal rank that satisfies the measurement
constraints.
9
DRAFT – December 18, 2017 – 2:26
x1
y 1 = w ⇤ x1
x2
..
.
w
y 2 = w ⇤ x2
xN
y N = w ⇤ xN
Figure 2: Blind multichannel estimation. An unknown noise source w feeds N unknown multipath channels
characterized by sparse impulse responses {x1 , x2 , . . . , xN }. We observe the convolutions at the receivers
and the task is to recover the channel responses together with the noise signal. The problem can be thought
of as the dual to the blind channel estimation problem where the roles of channels and the source signals
are reversed: a fixed incoherent vector is now fed into all the channels. The channel impulse responses can
be reliably modeled with Bernoulli Gaussian distribution.
2
Numerical Simulations
As an alternative to the computationally expensive semidefinite program in (8), we rely on a
heuristic non-linear program:
Ĥ, M̂ := argmin
H,M
kHk2F + kM k2F
subject to ŷn [`] = hb` φ∗`,n , HM ∗ i,
(13)
(`, n) ∈ [L] × [N ],
which solves for matrices H ∈ CL×R , and M ∈ CKN ×R . The semidefinite constraint in (8)
is always satisfied under the substitution X = HM ∗ . The non-linear program was proposed
in [4], and the results therein showed that all the local minima of (13) are the global minima of
(8) when R > rank(X̂), where X̂ is the optimal solution of (8). Since in our case the optimal
solution hm∗ is rank-1, we solve (13) with R = 2, and declare recovery when Ĥ, and M̂ are rank
deficient. The best rank-1 approximation of Ĥ M̂ ∗ constitutes the solution of (8). The non-linear
program considerably speeds up the simulations as instead of operating in the lifted space like (8)
with LKN variables involved, it operates almost in the natural parameter space with much fewer
number 2(L + KN ) of variables. We use an implementation of LBFGS available in [27] to solve
(13). An additional advantage of (13) is that no suitable initialization is required. Comparatively,
the recently proposed gradient descent scheme [19] for bilinear problems not only requires to solve
a separate optimization program to initialize well but also the gradient updates involve additional
unnatural regularizer to control the incoherence.
We present phase transitions that validate the sample complexity results in Theorem 1. The shade
in the phase transitions represents the probability of failure determined by counting the frequency of
failures in twenty five experiments for each pixel in the phase transitions. We classify the recovered
solution X̂ as a failure if kX̂ − hm∗ kF > 10−1 .
10
DRAFT – December 18, 2017 – 2:26
In all of the phase transitions, we take w, and m to be Gaussian vectors. Observe that w is
constructed to be a dense vector with no sparse model. Recall that Theorem 1 restricts w to be
a sparse vector, however, our simulation results show successful recovery in a more general case of
dense w. This observation is in conformation with our belief that the sparsity assumption on w is
a result of merely a technical requirement due to the proof method. Very similar phase transitions
can be obtained under restrictive sparse model on w.
We will present two sets of phase transitions. Each set contains three phase transition diagrams; in
each diagram, we fix one of the variables L, K, and N , and vary the other two in small increments
and compute the probability of failure every time as outlined earlier in this section.
In the first set, we mimic the channel estimation problem discussed in Section 1.4, and shown in
Figure 1. We take Cn ’s to be Gaussian matrices as in (11). Figure 3(a) shows that for a fixed
N = 40, we are able to recover all of the inputs xn ’s, and w as soon as L ≥ 10K. The phase
diagrams in Figure 3(b) and 3(c) mainly show that the performance of the algorithm become
roughly oblivious to the number N of inputs as soon as N ≥ 10 for the particular range of K, and
L considered in the phase transition diagrams.
In the second set shown in Figure 4, we simulate the blind channel estimation problem discussed in
Section 1.5, and shown in Figure 2. This set contains similar phase diagrams as in first set under
the same assumptions, the only difference is that the matrices Cn ’s are now the random subsets
of the columns of identity. In other words, we take the support of xn ’s to be random and known.
The results are almost exactly the same as in the first set.
3
Proof of Theorem 1
We are observing the linear measurements as in (5) of an unknown X0 = hm∗ such that h is an
S-sparse vector. Define Ω := supp(h), where |Ω| = S. To show that the solution to the SDP in (8)
equals X0 with high probability, we establish the existence of a valid dual certificate [7, 15]. The
proof of low-rank recovery using dual certificate method is a standard approach now and has been
employed in the literature [2, 15, 23] many times before. Our construction of the dual certificate
uses the golfing scheme [15], but is unusually technical in that there is a probabilistic dependence
between the iterates, which in turn precludes the use of matrix concentration inequalities.
Let u ∈ RL , and v ∈ RKN be arbitrary vectors, and h, and m be as defined earlier. Let T1 be the
linear space of matrices with rank at most two defined as
T1 := {X|X = αhv ∗ + βum∗ , and α, β ∈ R},
and T2 be the space of matrices with rows supported on index set Ω, and is defined as
T2 := {X | X ∈ RL×KN , X[`, j] = 0 for ` ∈ Ω⊥ },
where Ω⊥ = [L]\Ω. Then we have
T1 ∩ T2 := {X|X = αhv ∗ + βum∗ , uΩ⊥ = 0, and α, β ∈ R}.
Note that the matrix of interest X0 is a member of the space T1 ∩ T2 . Let us now define the related
projection operators. We start by defining P that takes a matrix or a vector with L rows and sets
11
DRAFT – December 18, 2017 – 2:26
L = 800
N = 40
1
5
10
1
50
0.8
90
0.6
30
0.8
K
K
55
130
0.4
170
0.2
200
0.6
80
0.4
105
500
800
1100
1400
1700
2000
0.2
130
0
5
30
55
80
105
L
130
155
180
0
N
(a)
(b)
K = 40
1
100
0.8
200
300
L
0.6
400
500
0.4
600
0.2
700
800
10
30
50
70
90
110 130 150 170 190
0
N
(c)
Figure 3: Empirical success rate for the deconvolution of w and x1 , x2 , . . . , xN . Recall that xn = Cn mn
for every n = 1, 2, 3, . . . , N . In these experiments, the vectors w, mn are Gaussian, and L × K matrices
Cn , n = 1, 2, 3, . . . , N are also independent, and Gaussian. (a) Fix N = 40, and vary L ∼ 200 → 2000,
K ∼ 10 → 190. Successful reconstruction is obtained with probability one when L ≥ 10K. (b) Fix L = 800,
and vary K ∼ 5 → 150, N ∼ 5 → 200. Successful reconstruction is obtained with probability one when
K ≤ L/10, and N ≥ 10. (c) Fix K = 40, and vary L ∼ 50 → 800, N ∼ 10 → 200. Successful reconstruction
occurs with probability one when L ≥ 10K, and N ≥ 10.
all the rows that are not indexed by the index set Ω ⊂ [L] to zero. Mathematically, we can define
the projection on the index set Ω as
∗
X
P(X) := IL×Ω IL×Ω
(14)
where IL×Ω denotes the submatrix of the L × L identity matrix with columns indexed by set Ω.
The orthogonal projector R onto T1 ∩ T2 is then defined as
R(Z) := hh∗ Z + P(Zmm∗ ) − hh∗ Zmm∗ ,
(15)
and the projector R⊥ onto the orthogonal complement T1⊥ ∪ T2⊥ of T1 ∩ T2 is then simply R⊥ (Z) =
Z − R(Z). Note that in the definition of the projection above, we assume without loss of generality
12
DRAFT – December 18, 2017 – 2:26
L = 800
N = 40
5
1
50
0.8
90
0.6
130
0.4
170
0.2
30
1
0.8
55
0.6
K
K
10
80
0.4
105
0.2
130
0
200
500
800 1100 1400 1700 2000
5
L
30
55
80
105 130 155 180
0
N
(a)
(b)
K = 40
1
100
0.8
200
300
0.6
L
400
500
0.4
600
0.2
700
800
10
50
90
130
170
0
N
(c)
Figure 4: Empirical success rate for the deconvolution of w and x1 , x2 , . . . , xN . In these experiments, the
vector w is Gaussian, and every xn is a sparse vector with random support and its K non-zero entries are
Gaussian. (a) Fix N = 40, and vary L ∼ 200 → 2000, K ∼ 10 → 190. Successful reconstruction is obtained
with probability one when L ≥ 10K. (b) Fix L = 800, and vary K ∼ 5 → 150, N ∼ 5 → 200. Successful
reconstruction is obtained with probability one when K ≤ L/10, and N ≥ 10. (c) Fix K = 40, and vary
L ∼ 50 → 800, N ∼ 10 → 200. Successful reconstruction occurs with probability one when L ≥ 10K, and
N ≥ 10.
that kmk2 = khk2 = 1 as the optimality conditions, presented in Lemma 1 below, for the success
of nuclear norm minimization 8 only involve normalized h, and m.
The following lemma gives sufficient conditions on a dual certificate Y ∈ Range(A∗ ) to guarantee
that nuclear-norm minimization program in (8) produces X0 = hm∗ as the solution. The proof
of the lemma is almost exactly the same as in [7], the only difference is that now instead of just
working with a space of rank-2 matrices T1 , we are dealing with the space T1 ∩ T2 of rank-2 and
row-sparse matrices. We repeat the proof here to show that all the details in [7] also work out for
the space T1 ∩ T2 .
13
DRAFT – December 18, 2017 – 2:26
Lemma 1 (Optimality Conditions). Let A be as defined in (7), and γ > 0 be a positive number
such that kAk ≤ γ, and
√
2kAR(Z)kF ≥ kR(Z)kF
(16)
for all Z ∈ Null(A). Then the matrix X0 = hm∗ is the unique minimizer of (8) if there exists a
Y ∈ Range(A∗ ) such that
khm∗ − R(Y )kF ≤
1
,
4γ
1
kR⊥ (Y )k ≤ .
2
(17)
Proof. Let X̂ denote the solution to the optimization program in (8). This implies that kX̂k∗ ≤
kX0 k∗ . Given this, it is enough to show
kX0 + Zk∗ > kX0 k∗ , ∀Z ∈ Null(A), Z 6= 0,
where Z = X̂ − X0 , to establish exact recovery, since the two conflicting requirements on X̂ above
would directly mean that X̂ = X0 , or Z = 0.
The sub-differential of nuclear norm at point X0 is (see [32] for details)
∂kX0 k∗ := {hm∗ + W : R⊥ (W ) = W , and kW k ≤ 1}.
Since by the definition of sub-differentials
kX0 + Zk∗ − kX0 k∗ ≥ h∆, Zi, for every ∆ ∈ ∂kX0 k∗ ,
we obtain
kX0 + Zk∗ − kX0 k∗ ≥ hhm∗ , Zi + hR⊥ (W ), Zi
= hhm∗ − Y , Zi + hW , R⊥ (Z)i, for every Y ∈ Range(A∗ ).
Using the fact that hm∗ ∈ T1 ∩ T2 , and also maximizing the inner product hW , R⊥ (Z)i with
respect to kW k ≤ 1 gives us
kX0 + Zk∗ − kX0 k∗ ≥ hhm∗ − R(Y ), R(Z)i − hR⊥ (Y ), R⊥ (Z)i + kR⊥ (Z)k∗
≥ −khm∗ − R(Y )kF kR(Z)kF − kR⊥ (Y )kkR⊥ (Z)k∗ + kR⊥ (Z)k∗ .
Now the inequality kAR(Z)kF ≥ 2−1/2 kR(Z)kF for any Z ∈ Null(A) implies that
0 = kA(Z)kF ≥ kAR(Z)kF − kAR⊥ (Z)kF
≥ 2−1/2 kR(Z)kF − γkR⊥ (Z)kF .
2 ≤ (2γ 2 +1)kR (Z)k2 , which in turn means R (Z) 6=
The above inequality implies firstly that kZk√
⊥
⊥
F
F
0 whenever Z 6= 0; secondly, kR(Z)kF ≤ 2γkR⊥ (Z)k∗ . Using these results, the bound in the
earlier inequality gives
√
kX0 + Zk∗ − kX0 k∗ ≥ −khm∗ − R(Y )kF 2γ − kR⊥ (Y )k + 1 kR⊥ (Z)k∗ .
Now under the conditions in (17), the right hand side above is strictly positive, which means
kX0 + Zk∗ > kX0 k∗ ; enough to exhibit the uniqueness.
14
DRAFT – December 18, 2017 – 2:26
The next lemma provides an upper bound γ on the operator norm of the linear map A.
Lemma 2 (Operator norm of A). Let A be as defined in (7). For any β ≥ 1,
p
kAk ≤ βK log(LN )
with probability at least 1 − (LN )−β+2 .
Proof. The operator norm can be calculated by using the fact that hA`1 ,n1 , A`2 ,n2 i = 0, ∀`1 6=
`2 ∈ [L], or ∀n1 P
6= n2 ∈ [N ]. This implies that kAk = max`,n kA`,n kF . We can write kA`,n k2F =
2
2
kb` k2 kφ`,n k2 = k∼K n |c`,n [k]|2 , where |c`,n [k]|2 are chi-squared random variables with degree 2
when 2 ≤ ` ≤ L/2, and degree 1 when ` = 1, or ` = L/2 + 1. In both cases E |c`,n [k]|2 = 1, and
P{|c`,n [k]|2 > u} ≤ e−u .
The maximum is taken over (L/2 + 1) · 2 · KN unique c`,n [k] and
P{max |c`,n [k]|2 ≥ u} ≤ (L + 2)KN · e−u ⇒ P{max kA`,n k2F ≥ Ku} ≤ (LN )2 e−u ,
`,n,k
`,n
where the last
Choose u = β log(LN ), which gives
p inequality follows from the fact that K ≤ L. −β+2
kAk ≤ γ := βK log(LN ) with probability at least 1 − (LN )
.
In the following section, we focus on constructing a dual certificating using golfing scheme, which
is then shown to satisfy the uniqueness conditions in the lemma above.
3.1
Linear operators on golfing partition
To prove the uniqueness conditions in (17), we use a dual certificate Y constructed using a variation
of the golfing scheme [15]. To this end, we partition the index set [L] × [N ] into P disjoint sets
Γ1 , Γ2 , . . . , ΓP defined as
Γp := (`, n) | (`, n) ∈ {(∆n,p , n)}N
(18)
n=1 ,
where ∆n,p is a subset of [L], chosen uniformly at random
S for every n and p, such that |∆n,p | = Q =
0
0
L/P , and for every n, ∆n,p ∩ ∆n,p = ∅ for p 6= p , and p ∆n,p = [L]. The parameter Q is adjusted
through the proof, and we assume here without loss of generality that it is an integer3 . In words,
the partition of the set {1, 2, 3, . . . , L} × {1, 2, 3, . . . , N } is obtained by dividing {1, 2, 3, . . . , L} into
P randomly chosen disjoint sets for a given n, and then we repeat this process independently for
every n to obtain a total of P N sets ∆1,1 , . . . , ∆N,1 , ∆1,2 , . . . , ∆N,P . We then define Γp as in (18).
For every n, the disjointness among the sets ∆n,1 , ∆n,2 , . . . , ∆n,P is of critical importance as it
ensures that no dependence arises due to reuse of the same φ`,n in different partitioned sets. We
define a linear map Ap : RL×KN → CQN that returns the measurements indexed by (`, n) ∈ Γp :
X
Ap (Z) := {hb` φ∗`,n , Zi : (`, n) ∈ Γp },
A∗p Ap (Z) :=
hb` φ∗`,n , Zib` φ∗`,n .
(19)
(`,n)∈Γp
3
We are assuming here that P is a factor of L; this can be achieved in the worst case by increasing the number
L of measurements in each convolution by no more than a factor 2, which only affects the measurements bounds in
Theorem 1 by a multiplicative constant.
15
DRAFT – December 18, 2017 – 2:26
In tandem with this partitioning of the measurements, we also require to partition the L rows of
B̂ into P sets of Q × L submatrices that behave roughly as an isometry on the sparse vectors.
Quantitatively, we want the rows b∗` of matrix B̂ in each of the sets ∆n,p to obey
(1 − δ)kzk22 ≤
L X ∗ 2
|b` z| ≤ (1 + δ)kzk22
Q
(20)
`∈∆n,p
for all vectors z supported on the set Ω such that |Ω| ≤ S. A reader familiar with compressive
sensing will readily recognize this as the restricted isometry property (RIP) [5] on the submatrices
of B̂. A result [26] from compressive sensing says that for each submatrix of rows b` with coherence
µ2max , defined in (9) if the index set ∆n,p is chosen uniformly at random, then there exists a constant
C such that for any 0 < δ < 1, and 0 < < 1,
S log L
S log L
−1 2
Q ≥ Cδ µmax
log
log2 S
2
2
2
implies that the RIP in (20) holds with probability exceeding 1 − e−c(δ/) . Given the partition
{∆n,p }n,p is chosen uniformly at random, the above result with δ = 1/4, and −2 = β log(LN )
means that if
Q ≥ Cβµ2max S log2 (LN ) log(βS log(LN )) log2 S,
(21)
then
sup
X
|Ω|≤S `∈∆
n,p
|b∗` z|2 ≤
5Q
kzk22 ,
4L
with probability at least 1 − O((LN )−β ), for all z supported on Ω for a given ∆n,p . A simple union
bound over all P N number of sets {∆n,p }n,p shows that
X
X
5Q
Q
Q
|b∗` z|2 ≤
max sup
kzk22 =⇒ max
Pb` b∗` P − P ≤
(22)
n,p
n,p
4L
L
4L
|Ω|≤S
`∈∆n,p
`∈∆n,p
holds with probability at least 1 − (P N )O((LN )−β ) ≥ 1 − O((LN )1−β ). In the rest of the article,
we take the results in (22) as given.
Before constructing a dual certificate, we define some nomenclature. Let
X
−1 ∗
‡
∗
Sn,p :=
(Pb` )(Pb` )∗ , and Sn,p
:= IL×Ω IL×Ω
Sn,p IL×Ω
IL×Ω ,
(23)
`∈∆n,p
where IL×Ω is an L × |Ω| matrix containing the columns of an L × L identity matrix indexed by
set Ω. A direct conclusion of (22) gives
max kSn,p k ≤
n,p
5Q
,
4L
‡
max kSn,p
k≤
n,p
4L
.
3Q
(24)
In addition, set Dn = IK ⊗ en e∗n . A linear operator Sp‡ is then defined by its action on an L × KN
matrix X as follows
X
‡
Sp‡ (X) =
Sn,p
XDn .
(25)
n
16
DRAFT – December 18, 2017 – 2:26
3.2
Coherence
With all the development above, we are now in position to precisely define the coherence parameter
µ20 that was first introduced in Section 1.3. The diffusion of the impulse response w is quantified
using the following definition
(
)
‡
‡
kB̂Sn,2
Sn0 ,1 hk2∞
kB̂Sn,p
hk2∞
kB̂hk2∞ Q2
2
· max
µ0 := L · max
,
, max0
.
(26)
n,n
khk22 L2 n,p
khk22
khk22
The quantities h00n,n0 and h00n,n0 from Section 1.3 can be easily read off from this expression, and
from `2 norm equivalences resulting from (24)4 . The following lemma presents upper and lower
bounds on µ20 .
Lemma 3 (Range of µ20 ). Let µ20 , and µ2max be as defined in (26), and (9), respectively. Assume
that (24) holds. Then
20
5
≤ µ20 ≤ µ2max S.
3
9
Proof. We assume without loss of generality that khk22 = 1. Since B̂ is an orthonormal matrix, it
‡
is easy to see that 1/L ≤ kB̂hk2∞ ≤ 1. As far as kB̂Sn,p
hk2∞ is concerned, an upper bound on it is
‡
‡
‡
h|2 ≤ max kb` k2∞ max kSn,p
hk21
max kB̂Sn,p
hk2∞ = max max |b∗` Sn,p
n,p
`
n,p
`
n,p
1
16
L
‡
≤ µ2max SkSn,p
hk22 ≤ µ2max S 2 ,
L
9
Q
where the first inequality is Hölder’s result; the second one follows from the definition of coherence
‡
µ2max in (9), the equivalence of `1 , and `2 -norms, and the fact that the vector Sn,p
h is S-sparse;
and the last one is the result of (24). A lower bound can be obtained by summing over ` ∈ [L] as
follows:
X
16 L2
‡
‡
‡
h|2 = max kSn,p
hk22 ≥
h|2 ≥ max
|b∗` Sn,p
L · max max |b∗` Sn,p
,
n,p
n,p
n,p
`
25 Q2
`
where the equality is due to the fact that B̂ is an orthonormal matrix, and last inequality follows
‡
from (24). In a similar manner, we can compute the upper and lower bounds on kB̂Sn,2
Sn0 ,1 hk2∞ ,
and the result is
5
20
‡
≤ max0 kB̂Sn,2
Sn0 ,1 hk2∞ ≤ µ2max S.
n,n
3
9
Combining all these results, the claim in the lemma follows.
The spirit of the incoherence is captured by the first term in the maximum; namely, kB̂hk2∞ , which
is small when w is diffuse in the frequency domain, and large otherwise. The other two terms are
mainly due to technical reasons in the proof presented later. It is a hard question to characterize
them exactly, but they are qualitatively expected to be of the same order as the first term because
‡
the matrices, Sn,p , and Sn,p
are random by construction and are not expected to make the vectors
‡
‡
Sn,p
h, and Sn,2
Sn0 ,1 h more aligned with the rows of B̂ than h was.
4
So that µ20 as defined here, and its illustration listed in the introduction, are equivalent to within an inconsequential
multiplicative factor.
17
DRAFT – December 18, 2017 – 2:26
3.3
Construction of a dual certificate via golfing
We now iteratively build a dual certificate Y ∈ Range(A∗ ) in P iterations with initial value Y0 = 0
as follows
Y1 =
L ∗
A A1 (hm∗ )
Q 1
and
Yp = Yp−1 + A∗p Ap Sp‡ (hm∗ − R(Yp−1 ))
for p ≥ 2.
(27)
Note that in (27), Yp ∈ Range(A∗ ) for every p. This approach to build the dual certificate was
first developed in [15]. Projecting both sides on T1 ∩ T2 results in
L
RA∗1 A1 (hm∗ )
Q
R(Yp ) = R(Yp−1 ) − RA∗p Ap Sp‡ (R(Yp−1 ) − hm∗ ) ,
R(Y1 ) =
p ≥ 2.
Denoting
W0 = −hm∗ ,
and Wp : = R(Yp ) − hm∗ for p ≥ 1
results in a recursion
L
∗
RA1 A1 R − R W0 ,
W1 =
Q
Wp = (RA∗p Ap Sp‡ R − R)Wp−1 for p ≥ 2,
(28)
(29)
which in turn implies
Wp0
p0
Y
(RA∗p Ap Sp‡ R
p=2
= −
L
∗
RA1 A1 R − R W0 .
− R)
Q
(30)
Running the iteration (27) till p = P gives us our candidate for dual certificate Y := YP . To
establish that X0 is the unique solution to (8), we need only show that kWP kF ≤ 1/4γ and
kR⊥ (YP )k < 1/2 in light of (17).
The Frobenius norm of WP is upper bounded by
P
Y
kWP kF ≤
kRA∗p Ap Sp‡ R − Rk
p=2
L
RA∗1 A1 R − R
Q
kW0 kF .
The difference in the construction of iterates for p = 1, and p ≥ 2 is mainly to avoid technical
difficulties that arise in the proofs later owing to the dependencies between Yp−1 and Ap for p ≥ 2.
The dependencies stem from the fact that although for every p, the random set ∆n,p is independent
of ∆n0 ,p , where n 6= n0 . However, for every n, the sets ∆n,p , and ∆n,p0 are dependent by construction.
This directly implies that the sets Γ1 , . . . , ΓP are dependent; therefore, Yp−1 , and hence Wp−1 are
dependent on Ap for p ≥ 2. As is shown in detail in the proofs to follow that the introduction of
Sp‡ for p ≥ 2 is to avoid this dependence problem and it ensures that
E RA∗p Ap Sp‡ (Wp−1 ) = Wp−1 ,
which is of critical importance in controlling some of the random quantities in the proofs to follow.
For p = 1, however, we do not introduce S1‡ as unlike Wp for p ≥ 1, the matrix W0 = hm∗ is fixed,
and there is no bias between (L/Q)RA∗1 A1 (hm∗ ), and hm∗ .
18
DRAFT – December 18, 2017 – 2:26
A bound on the operator norm for the term with p = 1 above can be achieved by using a simple
triangle inequality followed by an application of Lemma 5, and 6, in Section 3.5 below, to obtain
L
L
L
RA∗1 A1 R − R ≤ kRA∗1 A1 R − E RA∗1 A1 Rk +
E RA∗1 A1 R − R
Q
Q
Q
r
1 Q
≤
4 L
(31)
with probability at least 1 − 2(LN )−β+1 (the constant 2 upfront comes from the union bound), for
a parameter β > 2 that controls the choice of Q, and N . Note that the expectation E is only w.r.t.
the random construction of c`,n in (12), and not w.r.t. to the randomness due to the sets ∆n,p .
The operator norm of the remaining terms (p ≥ 2) in the expression (30) can all be bounded using
Lemma 4 in Section 3.5 below to conclude that
r
Q
−p−1
kWp kF ≤ 2
, for every p ∈ {1, . . . , P },
(32)
L
holds with probability
at least 1 − O(L−β+1 ). This means using the crude union bound that
p
−P
−1
Q/L holds with probability at least 1 − O(P L−β+1 ) ≥ 1 − O(L−β+2 ). Now
kWP kF ≤ 2
choosing P = (L/Q) = 0.5 log2 (4βK log(LN )) is more than sufficient to imply that kWP kF ≤
(4γ)−1 , where the value of γ is dictated by Lemma 2. This proves the first half of (17).
To prove the second half of (17), use the construction in (27) to write
P
X
L
YP = − A∗1 A1 (hm∗ ) +
A∗p Ap Sp‡ (Wp−1 ) .
Q
p=2
Since Wp ∈ T1 ∩ T2 , for every p ∈ {1, 2, 3, . . . , P }, this means
R⊥ (YP ) = −R⊥
L ∗
A A1 (hm∗ ) − hm∗
Q 1
−
P
X
p=2
R⊥ A∗p Ap Sp‡ (Wp−1 ) − Wp−1 .
Taking the operator norm and a triangle inequality, followed by an application of the fact that
kR⊥ k ≤ 1, shows that
P
X
L ∗
kR⊥ (YP )k ≤
A A1 (hm∗ ) − hm∗ +
A∗p Ap Sp‡ (Wp−1 ) − Wp−1 .
Q 1
p=2
Note that by adding and subtracting (L/Q) E A∗1 A1 (hm∗ ) in the first term and similarly adding
and subtracting E A∗p Ap Sp‡ (Wp−1 ) from every term in the summation above, and subsequently using
the triangle inequality, we obtain
L
L
kA∗1 A1 (hm∗ ) − E A∗1 A1 (hm∗ )k +
E A∗1 A1 (hm∗ ) − hm∗ +
Q
Q
P h
i
X
A∗p Ap Sp‡ (Wp−1 ) − E A∗p Ap Sp‡ (Wp−1 ) + E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 .
kR⊥ (YP )k ≤
p=2
Note that the expectation E is only w.r.t. the random construction of c`,n in (12), and not w.r.t.
to the randomness due to the sets ∆n,p . Now each of the term on the right hand side above can
19
DRAFT – December 18, 2017 – 2:26
be controlled using Corollary 2; and Lemma 12, 10, and 11 in Section 3.5 below, respectively that
hold under the choices of Q, and N that conform to Theorem 1 to give us the upper bound
P
kR⊥ (YP )k ≤
1 1 1 X −p 1
+ +
2 ≤ ,
8 8 2
2
p=2
which, using the union bound, holds with probability at least 1 − O((LN )−β+1 ).
Remark 1. Lemma 11 to control E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 relies on the uniform result in
(20) for sparse vectors to overcome the statistical dependence between Ap , and Wp−1 . Being able
to control this term is one of the main reasons of working with T1 ∩ T2 instead of T1 . Since
Wp−1 ∈ T1 ∩ T2 , the columns of Wp−1 are always S-sparse enabling us to employ the uniform
result. This technical requirement restricts our results to only S-sparse w. However, we think that
the proof technique may be improved to work for a completely dense vector w as is suggested by the
numerical experiments in Section 2.
We also need to show that (16) holds. To that end, note that by Corollary 1 in Section 3.5 below,
the linear map A is well-conditioned on T1 ∩ T2 , and hence, for Z ∈ Null(A), we have
kAR(Z)k2F = |hZ, RA∗ AR(Z)i|
≥ |hZ, (RA∗ AR − R)Zi + hZ, R(Z)i|
≥ kR(Z)k2F − kRA∗ AR − RkkR(Z)k2F .
Corollary 1 shows that kRA∗ AR − Rk ≤ 1/8. Using this fact in the inequality above proves (16).
Finally, the choice of upper bounds on L and N in the statement of the theorem is the tightest
upper bound that conforms to all the lemmas and corollaries, and uses the fact that L/Q = P =
0.5 log2 (4βK log(LN )) derived above. The nuclear norm minimization recovers the true solution
when all of the above conclusion hold true. The failure probability of each of the lemmas and
corollaries is less than or equal to (LN )−β+4 , and hence, using the union bound, the probability
that none of the above items fail is 1 − O((LN )−β+4 ).
This completes the proof of Theorem 1.
3.4
Concentration Inequalities
Most of the lemmas below require an application of either the uniform version, or the Orlicz-norm
version of the matrix Bernstein inequality to control the operator norm of the sum of independent
random matrices. Before stating them, we give an overview of the Orlicz-norm results that are
used later in the exposition.
We begin by giving basic facts about subgaussian and subexponential random variables that are
used throughout the proofs. The proofs of these facts can be found in any standard source; see, for
example, [31].
The Orlicz norms of a scalar random X are defined as
|X|α
kXkψα := inf u > 0 : E exp
≤ 2 , α ≥ 1.
uα
20
DRAFT – December 18, 2017 – 2:26
The Orlicz-norm of a vector kzkψα , and a matrix kZkψα are then defined by setting X = kzk2 ,
and X = kZk, respectively, in the above definition. Therefore, we restrict our discussion below to
scalar random variables, and it can trivially extended to vectors and matrices using above mentioned
equivalence.
Some of the key facts relating the Orlicz norms of subgaussian and subexponential random variables
are as follows. A subgaussian random variable X can be characterized by the fact that its Orlicz-2
norm is always finite, i.e., kXkψ2 < ∞. Similarly, for a subexponential r.v., we have kXkψ1 < ∞.
A random variable X is subgaussian iff X 2 is subexponential. Furthermore,
kXk2ψ2 ≤ kX 2 kψ1 ≤ 2kXk2ψ2 .
(33)
At some points in the proof, we are interested in bounding kX − E Xkψα . A coarse bound is
obtained by using the triangle inequality,
kX − E Xkψα ≤ kXkψα + k E Xkψα
followed by Jensen’s inequality, k E Xkψα ≤ E kXkψα = kXkψα , which further implies that
kX − E Xkψα ≤ 2kXkψα .
(34)
We also find it handy to have a generalized version of the above fact; namely, the product of two
subgaussian random variables X1 , and X2 is subexponential, and
kX1 X2 kψ1 ≤ CkX1 kψ2 kX2 kψ2 ,
(35)
for some C > 0.
As we are working with Gaussian random variables mostly in the proofs, some of the useful identities
for a Gaussian vector g ∼ Normal(0, IM ) are: For a fixed vector z, the random variable hg, zi is
also Gaussian, and hence, |hg, zi|2 must have a subexponential tail behavior, and it can be easily
verified that
2
P{|hg, zi|2 > λ} ≤ e−λ/kzk2
for every scalar λ ≥ 0. Moreover, the `2 -norm of g is strongly concentrated about its mean, and
there exists C > 0 such that
P{kgk22 > λM } ≤ Ce−λ .
This tail behavior of a random variable completely defines its Orlicz-norm. Specifically, for a
subexponential random variable X,
P{X > u} ≤ αe−βu =⇒ kXkψ1 ≤ (1 + α)/β.
(36)
This completes the required overview.
We now state the matrix Bernstein inequalities, which is heavily used in the proofs below.
Proposition 1 (Uniform Version [30], [17]). Let Z1 , Z2 , . . . , ZQ be iid random matrices with dimensions M × N that satisfy E(Zq ) = 0. Suppose that kZq k < U almost surely for some constant
U , and all q = 1, 2, 3, . . . , Q. Define the variance as
Q
Q
X
X
2
σZ
= max
(E Zq Zq∗ ) ,
(E Zq∗ Zq )
.
(37)
q=1
q=1
21
DRAFT – December 18, 2017 – 2:26
Then, there exists C > 0 such that , for all t > 0, with probability at least 1 − e−t ,
n p
o
kZ1 + Z2 + · · · + ZQ k . max σZ t + log(M + N ), U (t + log(M + N )) .
(38)
The version of Bernstein listed below depends on the Orlicz norms kZkψα , α ≥ 1 of a matrix Z,
defined as
kZkα
≤ 2}, α ≥ 1.
(39)
kZkψα = inf{u > 0 : E exp
uα
Proposition 2 (Orlicz-norm Version [17]). Let Z1 , Z2 , . . . , ZQ be iid random matrices with dimensions M × N that satisfy E(Zq ) = 0. Suppose that kZq kψα ≤ Uα for some constant Uα > 0, and
2 as in (37). Then, there exists C > 0 such that, for all
q = 1, 2, 3, . . . , Q. Define the variance σZ
−t
t > 0, with probability at least 1 − e ,
p
1
QUα2
α
kZ1 + Z2 + · · · + ZQ k . max σZ t + log(M + N ), Uα log
(t + log(M + N )) . (40)
2
σZ
3.5
Key lemmas
This section provides the important lemmas that constitute the main ingredients to establish the
uniqueness conditions (16), and (17) for our construction of the dual certificate Y in the previous
section.
Conditioning on T1 ∩ T2
The results in this section concern the conditioning of the linear maps A, and Ap when restricted
to the space T1 ∩ T2 .
Lemma 4. Let the coherences µ2max , µ20 , and ρ20 be as defined in (9), (26), and (10), respectively.
Fix β ≥ 2. Choose the subsets Γp := {(∆n,p , n)}n constructed as in Section 3.1, so that, for
p ∈ {2, 3, . . . , P },
|∆n,p | = Q ≥ Cβ(µ20 K + µ2max S) log2 (LN ),
for some sufficiently large C > 0. Then the linear operators Ap , and Sp‡ defined in (19) and (25),
obey
max
2≤p≤P
RA∗p Ap Sp‡ R − R ≤
1
2
(41)
with probability at least 1 − (P − 1)(LN )−β ≥ 1 − (LN )−β+1 .
Lemma 5. Let the coherences µ2max , µ20 , ρ20 be as in Lemma 4. Fix β ≥ 1. Choose ∆n,1 such that
|∆n,1 | = Q ≥ Cβ(µ20 K + µ2max S)(L/Q)1/2 log2 (LN ),
for some sufficiently large C > 0. Then the linear operator A1 defined in (19) obeys
r
L
1 Q
∗
∗
kRA1 A1 R − E RA1 A1 Rk ≤
Q
8 L
with probability at least 1 − (LN )−β .
22
DRAFT – December 18, 2017 – 2:26
(42)
(43)
Corollary 1 (Corollary of Lemma 5). Let the coherences µ2max , µ20 , and ρ20 be as in Lemma 4. Fix
β ≥ 1. Assume that
L ≥ Cβ(µ20 K + µ2max S) log2 (LN ),
for some sufficiently large C > 0. Then the linear operator A defined in (7) obeys
1
8
kRA∗ AR − Rk ≤
with probability at least 1 − (LN )−β .
Lemma 6. Let µ2max be as in Lemma 4. Fix β ≥ 2. Then there exists a constant C such that
|∆n,1 | = Q ≥ Cβµ2max S(L/Q) log(LN )
implies that the linear operator A1 defined in (19) obeys
L
1
E RA∗1 A1 R − R ≤
Q
8
r
Q
L
with probability at least 1 − (LN )−β+1 .
Coherences of iterates
In this section, we define the coherences of the iterates Wp in (29), and show that these coherences
can be bounded in terms of µ0 , and ρ0 in (26), and (10), respectively. The coherences are defined,
‡
among other variables, in terms of Sn,p
in (23), and Dn := IK ⊗ en e∗n . The partition {(∆n,p , n)}n,p
is as defined in Section 3.1, and we assume the implications of restricted isometry property in (22)
as given. Moreover, we also take the results of Lemma 4, 5, and 6 as true. The following results
are in order then.
Lemma 7. Define5
ρ2p :=
Q
N max
L 1≤n0 ≤N
b∗`0 Sn‡ 0 ,p+1 Wp Dn0
X
`0 ∈∆n0 ,p+1
Then
r
−p
ρp ≤ 2
2
2
for every p ∈ {1, 2, 3, . . . , P }.
Q
ρ0 for every p ∈ {1, 2, 3, . . . , P }.
L
(44)
(45)
Lemma 8. Define
Q2
νp2 :=
N max
L 1≤n0 ≤N
"
max
`0 ∈∆n0 ,p+1
b∗`0 Sn‡ 0 ,p+1 Wp Dn0
2
2
#
, for every p ∈ {1, 2, 3, . . . , P }.
(46)
Fix β ≥ 1. Then there exists a constant C such that
Q ≥ Cβ(µ20 K + µ2max S) log2 (LN )
implies
νp ≤ 2−p+3 µ0 ρ0 for every p ∈ {1, 2, 3, . . . , P }
with probability at least 1 −
(47)
(LN )−β .
5
We use the index variables `0 , and n0 as ` and n are already reserved to index the set Γp in the proofs of these
lemmas.
23
DRAFT – December 18, 2017 – 2:26
Lemma 9. Define
µ2p
N
Q2 X
:=
L 0
n =1
"
max
`0 ∈∆n0 ,p
b∗`0 Sn‡ 0 ,p+1 Wp Dn0
2
2
#
for every p ∈ {1, 2, 3, . . . , P }.
(48)
Let Q, and N be the same as in Lemma 8 for sufficiently large C. Then
µp ≤ 2−p+2 µ0 for every p ∈ {1, 2, 3, . . . , P }
(49)
with probability at least 1 − (LN )−β .
Range of A∗
Finally, the results in this section help us establish that the dual certificate Y mostly lies in T1 ∩ T2 ,
that is, kR⊥ (Y )k ≤ 1/2; one of the uniqueness conditions in (17). Let Ap , Sp‡ and Wp be as in
(19), (25), and (29), respectively. In addition, let the coherences µp , ρp , and νp be as defined in
(48), (44), and (46), respectively. We shall take (49), (45), and (47) as given. The following results
are in order then.
Lemma 10. Fix β ≥ 4. Then there exists a constant C such that
Q ≥ Cβµ20 K max{(L/Q), log2 (LN )} log(LN ), and N ≥ Cβρ20 (L/Q) log(LN )
implies that
A∗p Ap Sp‡ (Wp−1 ) − E A∗p Ap Sp‡ (Wp−1 ) ≤ 2−p−1 for all p ∈ {2, 3, . . . , P }
holds with probability at least 1 − (LN )−β+4
Corollary 2 (Corollary of Lemma 10). Let A1 be as in (19), and coherences µ20 , and ρ20 be as in
(26), and (10), respectively. Fix β ≥ 1. Then there exist a constant C such that
Q ≥ Cβµ20 K max{(L/Q), log2 (LN )} log(LN ), and N ≥ Cβρ20 (L/Q) log(LN )
implies that
L
1
kA∗1 A1 (hm∗ ) − E A∗1 A1 (hm∗ )k ≤
Q
8
with probability exceeding 1 − (LN )−β .
Lemma 11. Assume further that the restricted isometry property in (22) holds. Then
k E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 k ≤ 2−p−1 for all p ∈ {2, 3, . . . , P }
with probability at least 1 − (LN )−β+1
Lemma 12. Let A1 , µ2max , and ρ20 be as in (19), (9), and (10), respectively. Fix β ≥ 1. Then
there exists a constant C such that
L ≥ Cβµ2max S(L/Q) log(LN ), and N ≥ Cβρ20 (L/Q) log(LN )
implies that
L
1
E A∗1 A1 (hm∗ ) − hm∗ ≤
Q
8
with probability at least 1 − (LN )−β .
24
DRAFT – December 18, 2017 – 2:26
4
Proofs of the Key Lemmas
This section provides the proofs of all the key lemmas laid out in Section 1.3. All of the main
lemmas involve bounding the operator norm of a sum of independent random matrices with high
probability. The matrix Bernstein inequality is used repeatedly to compute such probability tail
bounds.
In the proof of the lemma below, the following calculations come in handy. Using the definition of
the projection R in (15), we can see that
R(b` φ∗`,n ) = (hh∗ b` )φ∗`,n + Pb` (mm∗ φ`,n )∗ − (hh∗ b` )(mm∗ φ`,n )∗ .
It also follows from the definition (15) that PR = RP = R. Another quantity of interest is
kR(b` φ∗`,n )k2F = hR(b` φ∗`,n ), b` φ∗`,n i, which can be expanded as
R(b` φ∗`,n )
2
F
= hhh∗ b` φ∗`,n , b` φ∗`,n i + hPb` φ∗`,n mm∗ , b` φ∗`,n i − hhh∗ b` φ∗`,n mm∗ , b` φ∗`,n i
= kφ`,n k22 |b∗` h|2 + kPb` k22 |m∗ φ`,n |2 − |b∗` h|2 |m∗ φ`,n |22
≤ kφ`,n k22 |b∗` h|2 + kPb` k22 |m∗ φ`,n |2 .
(50)
Moreover,
‡
R(Sn,p
b` φ∗`,n )
2
F
‡
‡
‡
‡
b` φ∗`,n i
b` φ∗`,n mm∗ , Sn,p
b` φ∗`,n i + hPSn,p
= hhh∗ Sn,p
b` φ∗`,n , Sn,p
‡
‡
b` φ∗`,n i
− hhh∗ Sn,p
b` φ∗`,n mm∗ , Sn,p
‡
‡
‡
h|2 |m∗ φ`,n |22
h|2 + kSn,p
b` k22 |m∗ φ`,n |2 − |b∗` Sn,p
= kφ`,n k22 |b∗` Sn,p
‡
‡
h|2 + kSn,p
b` k22 |m∗ φ`,n |2 .
≤ kφ`,n k22 |b∗` Sn,p
(51)
We are now ready to move on to the proof of Lemma 4 given below.
4.1
Proof of Lemma 4
Proof. The lemma concerns bounding the quantity kRA∗p Ap Sp‡ R−Rk for p ≥ 2. Using the definition
of Ap , and Sp‡ in (19), and (25), respectively, we expand the quantity RA∗p Ap Sp‡ R and evaluate its
expectation with the sets Γp fixed as follows
h
i
X
‡
E RA∗p Ap Sp‡ R =
R b` b∗` Sn,p
⊗ E φ`,n φ∗`,n R
(`,n)∈Γp
=
X
(`,n)∈Γp
h
i
‡
R b` b∗` Sn,p
⊗ Dn R.
From the construction of the sets Γp in Section 3.1, it is clear that (`, n) ∈ Γp means that for every
n, the index ` traverses the set ∆n,p . This means we can split the summation over Γp into an outer
‡
‡
sum over n and an inner sum over ` ∈ ∆n,p . Moreover, by definition R = RP, and PSn,p
= Sn,p
.
The following equality is now in order.
h
i
X X
‡
E RA∗p Ap Sp‡ R =
R Pb` b∗` PSn,p
⊗ IKN R.
n `∈∆n,p
25
DRAFT – December 18, 2017 – 2:26
‡
It is now easy to see by using the definition of Sn,p
in Section 1.3 that for p ≥ 2,
X
‡
Pb` b∗` PSn,p
= P,
`∈∆n,p
which implies that E RA∗p Ap Sp‡ R = R. Given this, we only need to control the term kRA∗p Ap Sp‡ R−
R(E A∗p Ap )Sp‡ Rk. By definitions of Ap , Sp‡ , and R in Section 3.1, we can write
h
i
X
‡
R b` b∗` Sn,p
⊗ φ`,n φ∗`,n R.
RA∗p Ap Sp‡ R =
(`,n)∈Γp
‡
Note that the action of the linear map Z`,n := R(b` b∗` Sn,p
⊗ φ`,n φ∗`,n )R on an L × KN matrix
h
i
‡
‡
X is Z`,n (X) = R b` b∗` Sn,p
⊗ φ`,n φ∗`,n R(X) = hX, R(Sn,p
b` φ∗`,n )iR(b` φ∗`,n ). It is clear from
this definition that Z`,n are rank-1 operators. Thus, we are asking the question of bounding the
operator norm of sum of independent operators Z`,n . Subtracting the expectation, we get
h
X
i
‡
R(A∗p Ap − E A∗p Ap )Sp‡ R =
⊗ φ`,n φ∗`,n − E φ`,n φ∗`,n R
R b` b∗` Sn,p
(`,n)∈Γp
=
X
(`,n)∈Γp
Z`,n − E Z`,n .
2 ; the
The operator norm of the sum can be controlled using Bernstein’s inequality. The variance σZ
main ingredient to compute the Bernstein bound, is in this case
X
X
2
∗
∗
σZ
:= max
− (E Z`,n )(E Z`,n )∗ ,
Z`,n − (E Z`,n )∗ (E Z`,n )
E Z`,n Z`,n
E Z`,n
(`,n)∈Γp
(`,n)∈Γp
X
X
∗
∗
≤ max
E Z`,n Z`,n
,
Z`,n
,
(52)
E Z`,n
(`,n)∈Γp
(`,n)∈Γp
where the last inequality follows from the fact that for two positive semidefinite (PSD) matrices
A, and B, kA − Bk ≤ kAk, whenever A − B is a PSD matrix. The first term in the maximum
in the variance expression is simplified below. As mentioned earlier, the linear operator Z`,n can
‡
be visualized as a rank-1 matrix R(b` φ∗`,n ) R(Sn,p
b` φ∗`,n ), and Z`,n (X) is just the product of the
rank-1 matrix above with the vectorized X, i.e.,
h
i
‡
Z`,n (X) = R(b` φ∗`,n ) R(Sn,p
b` φ∗`,n ) vec(X).
It is then easy to see that
∗
‡
Z`,n Z`,n
= kR(Sn,p
b` φ∗`,n )k2F · R(b` φ∗`,n ) R(b` φ∗`,n ) ,
where
R(b` φ∗`,n ) R(b` φ∗`,n ) = R b` b∗` ⊗ φ`,n φ∗`,n R.
Thus,
X
(`,n)∈Γp
∗
E Z`,n Z`,n
=
X
(`,n)∈Γp
‡
E kR(Sn,p
b` φ∗`,n )k2F R b` b∗` ⊗ φ`,n φ∗`,n R .
26
DRAFT – December 18, 2017 – 2:26
We remind the reader that the expectation is only over φ`,n , not over the randomness in the
construction of the partition Γp . Using the expansion in (50), the above quantity is upper bounded
by
X
(`,n)∈Γp
‡
‡
E kφ`,n k22 |b∗` Sn,p
h|2 + kSn,p
b` k22 |m∗ φ`,n |2 R b` b∗` ⊗ φ`,n φ∗`,n R
‡
≤ max kB̂Sn,p
hk2∞
X
n
(`,n)∈Γp
E kφ`,n k22 R b` b∗` ⊗ φ`,n φ∗`,n R
‡
+ max kSn,p
b` k22
`,n
‡
= max kB̂Sn,p
hk2∞
X
n
+
(`,n)∈Γp
X
(`,n)∈Γp
E |m∗ φ`,n |2 R b` b∗` ⊗ φ`,n φ∗`,n R
R b` b∗` ⊗ E(kφ`,n k22 φ`,n φ∗`,n ) R
16 2 SL
µ
9 max Q2
X
(`,n)∈Γp
R b` b∗` ⊗ E |m∗ φ`,n |2 φ`,n φ∗`,n R ,
(53)
where in the last equality, we made use of the linearity of Kronecker operator, and the bound
‡
‡
max kSn,p
b` k22 ≤ max kPb` k22 kSn,p
k22 ≤
`,n
`,n
16 2 SL
µ
,
9 max Q2
where we have in turn used the Cauchy Schwartz inequality; the fact that the non-zero rows and
‡
columns of Sn,p
are supported on the index set Ω; and the last inequality is the result of (9),
together with (24). It is then easy to verify that
E kφ`,n k22 φ`,n φ∗`,n = (K + 2)Dn ,
and
E |m∗ φ`,n |2 φ`,n φ∗`,n = (kmn k22 + 2mn m∗n )Dn .
Applying this and the facts that R = RP, and kRk ≤ 1, the above expression simplifies to
X
‡
∗
E Z`,n Z`,n
≤ (K + 2) max kB̂Sn,p
hk2∞
X
n
(`,n)∈Γp
+
(`,n)∈Γp
16 2 SL
µmax 2 max kmn k22
n
3
Q
Pb` b∗` P ⊗ Dn
X
(`,n)∈Γp
Using the definition of coherences in (26), and (10) and the fact that
that for an arbitrary matrix A, kA ⊗ Ik = kAk,
X X
n `∈∆n,p
P
(K + 2)L
2
2 SL
µ20
+
µ
ρ
· max
max 0 2
n
Q2
Q N
5 2 K + 2 20 2
2 S
≤
µ
+ µmax ρ0
,
4 0 Q
3
QN
∗
E Z`,n Z`,n
≤
27
DRAFT – December 18, 2017 – 2:26
Pb` b∗` P ⊗ Dn .
n IK
⊗ en e∗n = IKN , and
X
`∈∆n,p
Pb` b∗` P
where the last inequality follows from (22). The computation for the second term in the maximum
of variance expression (52) follows a very similar route. In short,
X
(`,n)∈Γp
∗
E Z`,n
Z`,n =
X
(`,n)∈Γp
‡
E kR(b` φ∗`,n )k2F R(b` b∗` Sn,p
⊗ φ`,n φ∗`,n )R
2
2 S
2K
· max
≤ 3 µ0 + µmax ρ0
n
L
LN
K
S
,
≤ 3 µ20 + µ2max ρ20
L
LN
X
`∈∆n,p
‡
Pb` b∗` PSn,p
(54)
‡
the last line is the result of the definition of Sn,p
in (23). This completes the calculation of the
variance term.
The second ingredient in the Bernstein bound is the calculation of the Orlicz norms of the summands. From (34), it follows that
kZ`,n − E Z`,n kψ1 ≤ 2kZ`,n kψ1 .
(55)
Since Z`,n is rank-1, its operator norm simplifies to
‡
kZ`,n k = kR(b` φ∗`,n )kF kR(Sn,p
b` φ∗`,n )kF ,
and it is also easy to show that k E Z`,n k = 1. Let us begin by showing that the random variable
kR(b` φ∗`,n )kF is subgaussian. For this it is enough to prove that kkR(b` φ∗`,n )kF k2ψ2 < ∞. Now
using (33), we have
kkR(b` φ∗`,n )kF k2ψ2 ≤ kkR(b` φ∗`,n )k2F kψ1
S
2
2
2
2
≤ khk∞ · kkφ`,n k2 kψ1 + µmax k|hm, φ`,n i| kψ1
L
K
S
≤ C µ20 + µ2max ρ20
.
L
LN
(56)
In a very similar manner, one shows that
‡
‡
b` φ∗`,n )kF k2ψ2 ≤ kkR(Sn,p
b` φ∗`,n )k2F kψ1
kkR(Sn,p
S
‡
2
2
2
2
≤ C kBSn,p hk∞ · kkφ`,n k2 kψ1 + µmax k|hm, φ`,n i| kψ1
L
KL
SL
≤ C µ20 2 + µ2max ρ20 2
.
Q
Q N
Using (35), we now obtain
‡
kZ`,n kψ1 ≤ CkkR(Sn,p
b` φ∗`,n )kF kψ2 · kkR(b` φ∗`,n )kF kψ2
L
2K
2
2 S
≤C
µ0 + µmax ρ0
.
Q
L
LN
Thus kZ`,n − E Z`,n k is a sub-exponential random variable; hence, α = 1 in (40), which gives
(QN )U12
log
≤ C log M,
(57)
2
σZ
28
DRAFT – December 18, 2017 – 2:26
where M = µ20 KN + µ2max ρ20 S ≤ C(LN ), the last inequality being the result of µ20 ≤ L, µ2max ≤ L,
and ρ20 ≤ N , and that L ≥ S, K.
Plugging the upper bound on the variance in (54), and (57) in (40), we have by using t = β log(LN )
that
kRA∗p Ap Sp‡ R − Rk ≤
(s
)
K
K
S
S
C max
log(LN ), β µ20 + µ2max ρ20
log2 (LN )
β µ20 + µ2max ρ20
Q
QN
Q
QN
(58)
with probability at least 1 − (LN )−β . The result in the statement of the lemma follows by choosing
Q, and N as in the statement of the lemma for a sufficiently large constant C.
4.2
Proof of Lemma 5
Proof. Lemma 4 considers bounding the quantity kRA∗p Ap Sp‡ R − Rk for p ≥ 2. The proof of this
corollary is very similar to the proof of Lemma 4, and basically follows by replacing Ap with A1 ,
and Sp‡ with (L/Q)P. Therefore, we lay out the steps very briefly. Start by expressing the quantity
of interest as a sum of independent random linear maps
L
L X
R(A∗1 A1 − E A∗1 A1 )R =
R b` b∗` ⊗ φ`,n φ∗`,n − E φ`,n φ∗`,n R,
Q
Q
(`,n)∈Γ1
Define Z`,n := (L/Q)R b` b∗` ⊗ φ`,n φ∗`,n R. Since the linear maps Z`,n can be thought of as
symmetric matrices
Z`,n = R(b` φ∗`,n ) R(b` φ∗`,n ),
∗
∗ Z
this means Z`,n
`,n = Z`,n Z`,n . Therefore, for the variance it suffices to compute
X
(`.n)∈Γ1
∗
Z`,n =
Z`,n
L2
Q2
X
(`.n)∈Γ1
kR(b` φ∗`,n )k2F R b` b∗` ⊗ φ`,n φ∗`,n R .
Using (50), and a calculation that goes along the same lines as before, and employing (52), we can
compute the upper bound on the variance
2
σZ
L2
S
2
2
2
≤ 2 3KkB̂hk∞ + 3Cµmax
max kmn k2 · max
n
n
Q
L
µ2
S 1
≤ 3 K 0 + µ2max ρ20
.
Q
Q N
X
`∈∆n,1
Pb` b∗` P
The Orlicz norm turns out to be kZ`,n − E Z`,n kψ1 ≤ 2kZ`,n kψ1 ≤ C(L/Q)kkR(b` φ∗`,n )k2F kψ1 , which
from (56) is
2K
2
2 S
kZ`,n − E Z`,n kψ1 ≤ C µ0 + µmax ρ0
.
Q
QN
This shows that Z`,n are sub-exponential, and thus using α = 1 in the definition of U1 in Proposition
2, one obtains
QN U12
log
≤ C log µ20 KN + µ2max ρ20 S ≤ C log(LN ).
2
σZ
29
DRAFT – December 18, 2017 – 2:26
With this, we have all the ingredients to compute the deviation bound. Apply Bernstein inequality
in Proposition 2, and choose t = (β − 1) log(LN ) to obtain
L
kRA∗1 A1 R − E RA∗1 A1 Rk
Q
(s
)
K
K
S
S
2
2
2
2
≤ C max
β µ20 + µ2max ρ20
log(LN ), β µ0 + µmax ρ0
log (LN )
Q
QN
Q
QN
(59)
with probability at least 1−(LN )−β . The choice of Q, and N in (42)
q for a sufficiently large constant
C in there guarantees that the right hand side is smaller than 18 Q
L.
4.3
Proof of Corollary 1
Proof. The proof is exactly same as in Lemma 5 except now instead of using A1 that is defined
over a the subset Γ1 ⊂ [L] × [N ], we instead use a linear operator A defined over the entire set
Γ := {(`, n) ∈ [L] × [N ]} of measurements. This means, we only need to replace A1 with A, defined
in (7), and Q = |∆n,1 | with L = |Γ|. In addition, note that E A∗ A = I, where I is an identity
operator, i.e., I(X) = X for a matrix X. Making these changes in the proof of Lemma 5, one
obtains the result claimed in the statement of the corollary.
4.4
Proof of Lemma 6
Proof. Note that by using the definition of A1 in (19), we have
X
R E A∗1 A1 R =
R b` b∗` ⊗ E φ`,n φ∗`,n R
(`,n)∈Γ1
=
X
n
R
X
`∈∆n,1
6 R.
b` b∗` ⊗ Dn R =
The operator norm of the quantity of interest can then be simplified using the definition of Sn,1 in
Section 3.1, and the facts that R = PR = RP, and RPR = R as follows
X
X
L
L
R E A∗1 A1 R − R =
R
b` b∗` ⊗ Dn R − R
Q
Q
n
`∈∆n,1
XL
X
=
R
Pb` b∗` P ⊗ Dn R − R
Q
n
`∈∆n,1
≤
X X
L
Q
L
Pb` b∗` P ⊗ Dn − P kRk ≤
kRk
· max
n
Q
L
Q
n
`∈∆n,1
X
`∈∆n,1
Pb` b∗` P −
Q
P . (60)
L
We can apply the uniform result in (22) to bound the last quantity above, but really a weaker
nonuniform result from the compressive sensing literature [8] suffices, and also results in an overall
30
DRAFT – December 18, 2017 – 2:26
tighter bound on Q than a simple application of the uniform result in (21) would give. The nonuniform result says that there exists a constant C, such that for any 0 < < 1, and 0 < δ < 1,
whenever
µ2 S
max{log L, log(1/δ)},
Q ≥ C max
2
then
X
Q
Q
Pb` b∗` P − P ≤
L
L
`∈∆n,1
p
with probability at least 1 − δ. Specifically, taking = 1/8 Q/L, δ = (LN )−β says that taking
Q ≥ Cβµ2max S(L/Q) log(LN ),
is enough to guarantee that
X
`∈∆n,1
Pb` b∗` P
Q
1
− P ≤
L
8
Q
L
3/2
with probability at least 1 − (LN )−β . Now a union bound over N sets in the partition {∆n,1 }n
means that
X
1 Q 3/2
Q
∗
max
Pb` b` P − P ≤
n
L
8 L
`∈∆n,1
with probability at least 1 − N (LN )−β ≥ 1 − (LN )1−β .
Plugging this result in (60) proves the lemma.
4.5
Proof of Lemma 7
Proof. As a direct implication of the restricted isometry property in (22), and later using (24), one
obtains
X
5Q ‡
20 L
kb∗`0 Sn‡ 0 ,p Wp Dn0 k22 ≤
kSn0 ,p Wp Dn0 k22 ≤
kWp Dn0 k2F .
4L
9
Q
0
` ∈∆n0 ,p
From the definition of Wp in (29), and Cauchy Schwartz inequality, we have
" p
#"
#
2
Y
L
kWp Dn0 k2F ≤
kRA∗p Ap Sp‡ R − Rk2
max
kW0 Dn0 k2F .
RA∗1 A1 R − R
n0
Q
k=2
Using (31), and (41), we can verify that
ρ2p ≤ 4−p
Q
N max
kW0 Dn0 k2F , for each p ∈ {1, 2, 3, . . . , P }.
n0
L
Finally, from the definition of ρ20 , and that W0 = −hm∗ , we can conclude that
kW0 Dn0 k2F = khk22 km∗ Dn0 k22 = kmn0 k22 ≤ ρ20 /N,
which, if plugged back in the above bound on ρ2p , completes the proof of the lemma.
31
DRAFT – December 18, 2017 – 2:26
4.6
Proof of Lemma 8
Proof. Our exposition is different for p ∈ {2, 3, . . . , P }, and p = 1 owing to the difference in the
iterative construction of the dual certificate for these choice of p. We start by considering νp2 for
p ∈ {2, 3, . . . , P }. The following lemma provides an upper bound on kb∗`0 Sn‡ 0 ,p+1 Wp Dn0 k22 that is in
turn used to bound νp2 .
Lemma 13. Let
Π2p
Q
:= C K|b∗`0 Sn‡ 0 ,p+1 h|2
L
"
#
max
`∈∆n0 ,p
kb∗` Sn‡ 0 ,p Wp−1 Dn0 k22
+
S
Cµ2max
Q
km
n0
k22
max max
n
`∈∆n,p
‡
kb∗` Sn,p
Wp−1 Dn k22
,
(61)
and
Λ2p := C
µ20 K ∗ ‡
S2
|b`0 Sn0 ,p+1 h|2 + µ4max ρ20 2 kmn0 k22
L
Q N
"
max
`∈∆n0 ,p
b∗` Sn‡ 0 ,p Wp−1 Dn0
2
2
#
.
(62)
‡
Let Wp , Sn,p
be as defined in (28), and (23); and assume that the restricted isometry property in
(22) holds. Fix β ≥ 1. Then there exists a constant C such that
kb∗`0 Sn‡ 0 ,p+1 Wp Dn0 k22 ≤ C max βΠ2p log(LN ), β 2 Λ2p log4 (LN ) , for every p ∈ {2, 3, . . . , P }. (63)
with probability at least 1 − (LN )−β .
Lemma 8 is established in Section 5. Using (63), it is clear that the upper bound on νp2 in (46)
can be obtained by evaluating the maximum of the quantities Πp , and Λp over `0 ∈ ∆n0 ,p+1 , and
n0 ∈ [N ].
Putting this together with (26), (10), and (46) directly implies that
"
#
L
2
2K
2
2 S
2
max
max Πp ≤ C µ0 + µmax ρ0
νp−1
,
n0
`0 ∈∆n0 ,p+1
Q
QN
Q2 N
and
"
max
0
n
#
max
`0 ∈∆n0 ,p+1
Λ2p
S2
L
4 K
4 4
2
≤ C µ0 2 + ρ0 µmax 2 2 νp−1
.
Q
Q N
Q2 N
Choosing Q, and N as in Lemma 8 results in
1 2
νp2 ≤ νp−1
for each p ∈ {2, 3, . . . , P }
4
(64)
with probability at least 1 − (LN )−β .
For the remaining case of p = 1,
Q2
ν12 =
N max
n0
L
"
max
`0 ∈∆n0 ,2
b∗`0 Sn‡ 0 ,2 W1 Dn0
32
DRAFT – December 18, 2017 – 2:26
2
2
#
,
and using the definition of W1 in (29)
#
"
2
2
L
Q
ν12 =
N max
max b∗`0 Sn‡ 0 ,2
RA∗1 A1 (hm∗ ) − hm∗ Dn0
n0
`0 ∈∆n0 ,2
L
Q
2
"
#
2
2
L
Q
L
∗
∗
∗
∗
∗ ‡
≤ 2 N max
max b`0 Sn0 ,2
RA1 A1 (hm ) − E RA1 A1 (hm ) Dn0
+
n0
`0 ∈∆n0 ,2
L
Q
Q
2
#
"
2
L
Q2
∗ ‡
∗
∗
∗
.
max b`0 Sn0 ,2
E RA1 A1 (hm ) − hm Dn0
2 N max
n0
`0 ∈∆n0 ,2
L
Q
2
(65)
The following corollary provides a bound on the first term in the sum above.
Corollary 3 (Corollary of Lemma 13). Let
Q
∗L
2
∗ ‡
2
2
Π1 := C K|b`0 Sn0 ,2 h| max max kb` PW0 Dn0 k2
n `∈∆n,1
L
Q
S
2
2
2
∗L
+ Cµmax kmn0 k2 max max kb` PW0 Dn k2 ,
n `∈∆n,1
Q
Q
(66)
and
Λ21 := C
µ20 K ∗ ‡
S2
|b`0 Sn0 ,2 h|2 + µ4max ρ20 2 kmn0 k22
L
Q N
"
max max
n
`∈∆n,1
L
b∗` PW0 Dn0
Q
2
#
.
(67)
2
Let Q be as in Lemma 13. Let
s∗`0 ,n0 := b∗`0 Sn‡ 0 ,2
L
L
RA∗1 A1 (hm∗ ) − E RA∗1 A1 (hm∗ ) Dn0 ,
Q
Q
(68)
Assume that the restricted isometry property in (22) holds. Fix β ≥ 1. Then there exists a constant
C such that
ks`0 ,n0 k22 ≤ C max βΠ21 log(LN ), β 2 Λ21 log4 (LN )
with probability at least 1 − (LN )−β .
The proof of this corollary is provided in Section 5.
It is easy to see using the definitions (26), and (10) that
L
2
2K
2
2 S
max Π1 ≤ C µ0 + µmax ρ0
max
µ20 ρ20 2 ,
0
0
n ` ∈∆n0 ,2
Q
QN
Q N
and
max
max
0
0
n
` ∈∆n0 ,2
Λ21
S
L
4 K
4 4
≤ C µ0 2 + ρ0 µmax
µ20 ρ20 2 .
Q
QN
Q N
Using these calculations, the first term in (65) can be bounded by applying Corollary 3, and choosing
Q as in Lemma 47 for a large enough constant C to achieve
"
#
Q2
1
2
(69)
N max
max ks`0 ,n0 k2 ≤ µ20 ρ20 .
n0
`0 ∈∆n0 ,2
L
4
As for the second term in the sum in (65), the lemma below provides an upper bound.
33
DRAFT – December 18, 2017 – 2:26
Lemma 14. Define
r`∗0 ,n0
:=
b∗`0 Sn‡ 0 ,2
L
∗
∗
∗
E RA1 A1 (hm ) − hm Dn0 ,
Q
and assume that (24) holds. Then
9 ∗ ‡
L2
2
2
∗ ‡
2
kr`0 ,n0 k2 ≤
kb 0 S 0 hk + 2 2 max kb`0 Sn0 ,2 Sn,1 hk2 kmn0 k22 .
2 ` n ,2 2
Q n
(70)
(71)
Proof. The proof of this lemma is provided in Section 5.
As for the second term in (65), we appeal to Lemma 14, and directly obtain after evaluating the
maximum over `0 ∈ ∆n0 ,2 , and n0 ∈ [N ] in (71), and using the definition of coherences µ20 , and ρ20 ,
that
"
#
Q2
2
N max
max kr`0 ,n0 k2 ≤ 6.5µ20 ρ20 .
(72)
n0
`0 ∈∆n0 ,2
L
Plugging (69), and (72) in (65) shows that ν1 ≤ 4µ0 ρ0 . Combining this fact with (64) completes
the proof of the lemma.
4.7
Proof of Lemma 9
Proof. We first consider the case for p ∈ {2, 3, . . . , P }. Using Lemma 13, we have a bound on the
‡
Wp Dn0 k22 for p ≥ 2 in terms of Λ2p , and Π2p in (62), and (61), respectively. A
quantity kb∗`0 Sn,p
corresponding bound on µ2p above is then
µ2p
(
"
#
"
#
)
X
X
Q2
2
2
2
4
≤C
max
max Πp β log(LN ),
max Λp β log (LN ) .
`0 ∈∆n0 ,p
`0 ∈∆n0 ,p
L
0
0
n
(73)
n
Using the definitions of the coherences νp2 , and µ2p of P
the iterates Wp in (46), and (48); the definition
2
of the coherence µ0 of h in (26); and the fact that n0 kmn0 k22 = 1, one can verify from (62), and
(61) that
"
#
X
L
16 2
SL
2
2 2
2
max Πp ≤ C µ0 µp−1 K 3 + µmax νp−1 3
,
`0 ∈∆n0 ,p
Q
9
Q N
0
n
and
"
X
n0
#
max
`0 ∈∆n0 ,p
Λ2p
≤C
µ40
2
K
4
4 S
+
µ
ρ
max 0 2 2
Q2
Q N
µ2p−1
L
.
Q2
Plugging back in (73), and choosing Q and N as in Lemma 8 for a large enough C is sufficient to
guarantee that
1 2
2
2
(74)
µp ≤ δ µp−1 + νp−1 , for each p ∈ {2, 3, . . . , P }.
N
34
DRAFT – December 18, 2017 – 2:26
for arbitrarily small number δ that lies between zero and one. Bounding µ2p iteratively using the
2 , p ∈ {2, 3, . . . , P }, and now using (47), and the fact
above relation gives µ2p ≤ δ p−1 µ21 + (2δ/N )νp−1
2
that ρ0 ≤ N , we have
µ2p ≤ δ p−1 µ21 + 2δ4−p+3 µ20 , for each p ∈ {2, 3, . . . , P }.
All that remains now is to bound µ21 , which from (48) is
"
2 X
Q
µ21 :=
max b∗0 S ‡ 0 W1 Dn0
L 0 `0 ∈∆n0 ,2 ` n ,2
n
2
2
(75)
#
.
Using the definition of W1 in (28) followed by an application of a simple identity (a+b)2 ≤ 2(a2 +b2 ),
we have
2
L
L
∗ ‡
∗ ‡
∗
∗
∗
∗
b`0 Sn0 ,2 W1 Dn0 ≤ 2 b`0 Sn0 ,2
RA1 A1 (hm ) − E RA1 A1 (hm ) Dn0 +
Q
Q
2
2
2
L
E RA∗1 A1 (hm∗ ) − hm∗ Dn0 .
2 b∗`0 Sn‡ 0 ,2
Q
2
An application of Corollary 3 gives
2
Q2 X
L
L
∗
∗
∗
∗
∗ ‡
0
max
S
RA
A
(hm
)
−
E
RA
A
(hm
)
D
b
0
0
n
1 1
1 1
L 0 `0 ∈∆n0 ,2 ` n ,2 Q
Q
2
n
(
)
N
Q2 X
max β 0 max Π21 log(LN ), β 2 0 max Λ21 log4 (LN ) .
≤C
` ∈∆n0 ,2
` ∈∆n0 ,2
L 0
(76)
n =1
Using the fact that W0 = hm∗ , it is easy to see that
"
#
!
N
Q2 X
SL
2
∗ ‡
2
2
2
∗ 2
max Π ≤ C KQ max
max |b`0 Sn0 ,2 h| + µmax
max kmn k2
max max |b` h| ,
n `∈∆n,1
n0 `0 ∈∆n0 ,2
L 0 `0 ∈∆n0 ,2 1
Q n
n =1
where we have used the fact that the fact that
have the upper bound
P
n0
kmn0 k22 = 1. By definitions (26), and (10), we
N
Q2 X
2
4K
2
2 2 S
max Π ≤ C µ0 + µmax µ0 ρ0
.
L 0 `0 ∈∆n0 ,2 1
Q
QN
n =1
In a similar manner, one can show that
N
Q2 X
S2 4 1 2
2
4 K
4
max Λ ≤ C µ0 2 + µmax 2 ρ0 2 µ0 .
L 0 `0 ∈∆n0 ,2 1
Q
Q
N
n =1
Now with the choice of Q, and N in the statement of lemma for some suitably large constant, one
can show that
2
Q2 X
1
L
L
∗ ‡
∗
∗
∗
∗
0
max
b
S
RA
A
(hm
)
−
E
RA
A
(hm
)
D
≤ µ20 ,
0
0
n
1 1
1 1
L 0 `0 ∈∆n0 ,2 ` n ,2 Q
Q
2
2
n
35
DRAFT – December 18, 2017 – 2:26
where the last inequality follows from ρ20 ≤ N . For the remaining term, using Lemma 14, and taking
summation over n0 followed by maximum over `0 ∈ ∆n0 ,2 on both sides, and using the definition
(26), we obtain
"
#
2
Q2 X
9 2
L
∗
∗
∗
2
∗ ‡
max b 0 S 0
E RA1 A1 (hm ) − hm Dn0
µ0 + 2µ0 = 6.5µ20 .
≤
L 0 `0 ∈∆n0 ,2 ` n ,2 Q
2
2
n
Plugging back returns µ21 ≤ 14µ20 . Combining this with (75) for a small enough δ means that we
can bound
µ2p ≤ 4−p+2 µ20 , for each p ∈ {1, 2, 3, . . . , P }.
This completes the proof.
4.8
Proof of Lemma 10
Proof. The proof concerns bounding deviation of A∗p Ap Sp‡ (Wp−1 ) from its mean, and we resort to
the matrix Bernstein inequality to control it. By definition of Ap , and Sp‡ in Section 1.3, we have
X
X
‡
‡
Wp−1 φ`,n φ∗`,n ,
b` b∗` Sn,p
⊗ φ`,n φ∗`,n )(Wp−1 ) =
A∗p Ap Sp‡ (Wp−1 ) =
(b` b∗` Sn,p
(`,n)∈Γp
(`,n)∈Γp
and its expected value is
E A∗p Ap Sp‡ (Wp−1 ) =
X
X
‡
Wp−1 E φ`,n φ∗`,n =
b` b∗` Sn,p
‡
Wp−1 Dn .
b` b∗` Sn,p
(77)
(`,n)∈Γp
(`,n)∈Γp
Given this, the quantity of interest can now be expressed as the sum
X
‡
‡
A∗p Ap Sp‡ (Wp−1 ) − E A∗p Ap Sp‡ (Wp−1 ) =
b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n − b` b∗` Sn,p
Wp−1 Dn
(`,n)∈Γp
of mean zero, independent random matrices
‡
‡
Wp−1 Dn .
Z`,n := b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n − b` b∗` Sn,p
An application of the matrix Bernstein inequality in Proposition 2 requires us to compute a bound
2 . To this end
on the variance σZ
X
(`,n)∈Γp
∗
E Z`,n
Z`,n ≤
=
X
∗
‡
‡
E b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n
b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n
(`,n)∈Γp
X
(`,n)∈Γp
‡
E |b∗` Sn,p
Wp−1 φ`,n |2 φ`,n φ∗`,n ,
where again the first inequality is the result of the fact that kA − Bk ≤ kAk when A, B, and
A − B are PSD. An application of Lemma 15 shows that
‡
‡
E |b∗` Sn,p
Wp−1 φ`,n |2 φ`,n φ∗`,n 4 3kb∗` Sn,p
Wp−1 Dn k22 Dn ,
36
DRAFT – December 18, 2017 – 2:26
and taking the summation through returns
X
X
‡
‡
kb∗` Sn,p
Wp−1 Dn k22 Dn ,
E |b∗` Sn,p
Wp−1 φ`,n |2 φ`,n φ∗`,n 4 3
(`,n)∈Γp
(`,n)∈Γp
and thus the operator norm produces the variance
X
(`,n)∈Γp
∗
≤3
E Z`,n Z`,n
X
‡
kb∗` Sn,p
Wp−1 Dn k22 Dn
(`,n)∈Γp
= 3 max
n
X
`∈∆n,p
‡
kb∗` Sn,p
Wp−1 Dn k22 ≤ 3ρ2p−1
L
,
QN
(78)
where the last line follows from the definition of ρ2p in (44). In a similar manner, we can compute
X
(`,n)∈Γp
∗
≤
E Z`,n Z`,n
∗
‡
‡
Wp−1 φ`,n φ∗`,n
E b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n b` b∗` Sn,p
,
X X
n `∈∆n,p
where the first inequality follows for exact same reasoning as before. The summand simplifies to
‡
Wp−1 φ`,n |2 kφ`,n k22 ,
b` b∗` E |b∗` Sn,p
and once the expectation is moved inside, we obtain
‡
Wp−1 φ`,n |2 kφ`,n k22 = 3Kkb∗` Sn,p Wp−1 Dn k22 ,
E |b∗` Sn,p
and using the orthogonality of {b` }` , the operator norm simplifies to
X
X
∗
‡
E Z`,n Z`,n
≤ 3K · max max
kb∗` Sn,p
Wp−1 Dn k22 ·
(`,n)∈Γp
1≤p≤P `∈∆n,p
≤ 3µ2p−1
n
X
b` b∗`
`∈∆n,p
KL
,
Q2
(79)
which follows by the definition of the coherence µ2p in (48). As per (37), the maximum of (78), and
(79) is the variance
KL
L
2
2
2
σZ ≤ 3 µp−1 2 + ρp−1
.
(80)
Q
QN
The fact that Z`,n are sub-exponential can be proven by showing that max`,n kZ`,n kψ1 < ∞. First,
note that
‡
kZ`,n kψ1 = kb` b∗` Sp‡ Wp−1 φ`,n φ∗`,n − b` b∗` Sp‡ Wp−1 Dn kψ1 ≤ 2kb` b∗` Sn,p
Wp−1 φ`,n φ∗`,n kψ1 .
Second, the operator norm of the matrix under consideration is
‡
‡
kb` b∗` Sn,p
Wp−1 φ`,n φ∗`,n k = kb` k2 kφ`,n k2 |b∗` Sn,p
Wp−1 φ`,n |.
37
DRAFT – December 18, 2017 – 2:26
‡
It is well-known that |b∗` Sn,p
Wp−1 φ`,n | is a subgaussian random variable for an arbitrary matrix
p
‡
‡
∗
Wp−1 with kb` Sn,p Wp−1 φ`,n√
kψ2 ≤ Ckb∗` Sn,p
Wp−1 Dn k2 ≤ Cνp−1 L/Q2 N . Also kφ`,n k2 is subgaussian with kkφ`,n k2 kψ2 ≤ K. This implies Uα in Proposition 2 is
∗ ‡
U1 := max max kkφ`,n k2 |b` Sn,p Wp−1 φ`,n |kψ1
n
`∈∆n,p
s
KL
‡
Wp−1 φ`,n kψ2 kkφ`,n kkψ2 ≤ Cνp−1
≤ max max kb∗` Sn,p
,
(81)
n
`∈∆n,p
Q2 N
which means
log
(QN )U12
2
σZ
2 KQN
νp−1
≤ C log M, where M :=
µ2p−1 KN + ρ2p−1 Q
.
Applying the coherence bounds in (45), (47), and (49), we have
M ≤ C(ρ20 Q + µ20 KN ) ≤ CLN,
where the last inequality follows from the fact that ρ20 ≤ N , µ20 ≤ L, and L ≥ K. Combining all the
ingredients gives us the final result and choosing t = (β − 1) log(LN ) in the Bernstein inequality
shows that
kA∗p Ap Sp‡ (Wp−1 ) − E A∗p Ap Sp‡ (Wp−1 )k ≤
s
s
2 ν 2 KL log4 (LN )
β
KL
L
p−1
C max
β µ2p−1 2 + ρ2p−1
log(LN ),
Q
QN
Q2 N
(82)
holds with probability at least 1 − (LN )−β . Using the union bound, it follows from Lemma 9, 7,
and 8 that at least one of the coherence bounds in (45), (47), and (49) fails with probability at
most (P − 1)3 (LN )−β . This means that all of the coherence bounds hold with probability at least
1 − (P − 1)3 (LN )−β ≥ 1 − (LN )−β+3 . Plugging the coherence bounds in (82), and choosing Q, and
N as in Lemma 10 for appropriately large constant C ensures that
kA∗p Ap Sp‡ (Wp−1 ) − E A∗p Ap Sp‡ (Wp−1 )k ≤ 2−p−1 , for each p ∈ {2, 3, . . . , P }
holds with probability at least 1 − (LN )−β+3 . Using union bound for P − 1 choices of p, we can
show that the above conclusion holds for all p ∈ {2, 3, . . . , P } with probability at least 1 − (P −
1)(LN )−β+3 ≥ 1 − (LN )−β+4 .
4.9
Proof of Corollary 2
Proof. The proof follows essentially from the proof of Lemma 10 by taking p = 1, and afterwards
‡
taking S1‡ = (L/Q)P, or equivalently, Sn,1
= (L/Q)P. The final bound is obtained by making the
above changes in (82), and is
kA∗1 A1 (L/Q)P(hm∗ ) − E A∗1 A1 (L/Q)P(hm∗ )k ≤
s
s
4
2
2
L
β ν0 KL log (LN )
KL
β µ20 2 + ρ20
log(LN ),
C max
Q
QN
Q2 N
(83)
for some constant C that may differ from the one in (82). Choosing Q, and N as in Corollary 2 for
a large enough C proves the corollary.
38
DRAFT – December 18, 2017 – 2:26
4.10
Proof of Lemma 11
Proof. Begin by noting that it is clear from (77) that for a fixed Γp , we have
E A∗p Ap Sp‡ (Wp−1 ) 6= Wp−1 ,
and if one takes the random construction of Γp into account, the dependence between Γ1 , Γ2 , . . . , Γp−1 ,
and Γp that in turn means that Wp−1 is dependent on Γp . This means there is no simple way to
write this quantity as a sum of independent random matrices and apply the matrix Bernstein inequality to control the size as before. Fortunately, we can work a uniform bound using restricted
isometry property that works for all matrices Wp−1 , and thus overcome the issues of intricate dependencies between Wp−1 and Γp . By the equivalence between operator, and Frobenius norm, we
have
k E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 k2 ≤ k E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 k2F
h
i 2
h
i
+ P⊥ E A∗p Ap Sp‡ (Wp−1 )
= P E A∗p Ap Sp‡ (Wp−1 ) − Wp−1
F
2
F
,
(84)
where the projection operator P is defined in (14), and P⊥ is the orthogonal complement. Note
that
h
i X X
‡
Wp−1 Dn .
P E A∗p Ap Sp‡ (Wp−1 ) =
Pb` b∗` PSn,p
n
`∈∆n,p
‡
Using the definition of Sn,p
in (23), it is clear that
X
‡
Pb` b∗` PSn,p
= P,
`∈∆n,p
h
i
which implies, using the fact that PWp−1 = Wp−1 , P E A∗p Ap Sp‡ (Wp−1 ) = Wp−1 . This means
(84) reduces to
h
i
k E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 k2 ≤ P⊥ E A∗p Ap Sp‡ (Wp−1 )
2
F
.
As far as the second term in the above expression is concerned, using (77), one obtains
(`,n)∈Γp
‡
P⊥ b` b∗` Sn,p
Wp−1 Dn
2
2
2
X
=
X
‡
b` b∗` Sn,p
Wp−1 Dn
(`,n)∈Γp
F
=
X
(`,n)∈Γp
−
F
‡
kb∗` Sn,p
Wp−1 Dn k22
−
X
(`,n)∈Γp
‡
Pb` b∗` Sn,p
Wp−1 Dn
F
kWp−1 k2F .
The last equality follows from the fact that {b` }` are orthonormal vectors, Wp−1 Dn is orthogonal
‡
to Wp−1 Dn0 for n 6= n0 and by the definition of Sn,p
in (23). The matrix Wp−1 is dependent on
the sets ∆n,1 , ∆n,2 , . . . , ∆n,p−1 that are in turn dependent on ∆n,p by construction. However, we
can avoid this dependence issue here as the result in (22) is uniform in nature in the sense that it
39
DRAFT – December 18, 2017 – 2:26
holds for all S-sparse vectors. Employing this result on every column of Wp−1 ∈ Ω, we obtain
X
(`,n)∈Γp
‡
kb∗` Sn,p
Wp−1 Dn k22 =
≤
X X
n `∈∆n,p
‡
kb∗` Sn,p
Wp−1 Dn k22 ≤
5Q X ‡
kSn,p Wp−1 Dn k2F
4L n
X
5Q
20 L
‡
max kSn,p
k22
kWp−1 Dn k2F ≤
kWp−1 k2F ,
n
4L
9
Q
n
which holds with probability at least 1 − (LN )−β . This implies using the calculation above that
2
X
(`,n)∈Γp
‡
P⊥ b` b∗` Sn,p
Wp−1 Dn
≤
20 L
− 1 kWp−1 k2F .
9 Q
Now the decay rate of kWp kF in the statement of the lemma is sufficient to guarantee that
k E A∗p Ap Sp‡ (Wp−1 ) − Wp−1 k ≤ 2−p−1
for every p ∈ {2, 3, . . . , P }. Using the union bound as before, the statement can be extended to all
p ∈ {2, 3, . . . , P } with probability at least 1 − (LN )−β+1 .
4.11
Proof of Lemma 12
Proof. Using the definition of A1 , and after evaluating the expectation, it is easy to see that
X
E A∗1 A1 (hm∗ ) =
b` b∗` hm∗ Dn .
(`,n)∈Γ1
From Section 3.1, we know that Γ1 = {(∆n,1 , n)}n , where the sets |∆n,1 | = Q for every n, and are
˜ n,1 , n)}n , where ∆
˜ n,1 for each n are independent
chosen uniformly at random. Define a set Γ̃1 := {(∆
Bernoulli sets defined as
˜ n,1 := {` ∈ [L] | δ`,n = 1},
∆
where δ`,n is an independently chosen Bernoulli number for every `, and n that takes value one
with probability Q/L. Since the probability of failure the event
LX X
b` b∗` hm∗ Dn − hm∗ ≤
Q n
`∈∆n,1
for some number > 0 is a nonincreasing function of Q = |∆n,1 | for every n; this follows by
the orthogonality
of {b` }` , and by the fact that increasing |∆n,1 | only increases the range of the
P
projector `∈∆n,1 b` b∗` , and hence the distance in the operator norm above can either decrease or
stay the same. It now follows using Lemma 2.3 in [9] that the probability of failure of the event
above is less than or equal to twice the probability of failure of the event
L
Q
X
(`,n)∈Γ̃1
b` b∗` hm∗ Dn − hm∗ ≤ .
40
DRAFT – December 18, 2017 – 2:26
(85)
Therefore, in the rest of the proof it suffices to only consider the event above where the index sets
˜ n,1 defined above, we can write
˜ n,1 }n . Using the definition of the Bernoulli sets ∆
are {∆
X X
X
b` b∗` hm∗ Dn =
δ`,n b` b∗` hm∗ Dn ,
n `∈∆
˜ n,1
`,n
where on the right hand side the summation is over all ` ∈ [L]. Note that
X
L X
E
δ`,n b` b∗` hm∗ Dn =
b` b∗` hm∗ Dn = hm∗ ,
Q
∀(`,n)
∀(`,n)
where we have used the fact that E δ`,n =
Q
L.
A simple application of matrix Bernstein is now enough to show that the event in (85) holds for the
desired with high probability. To this end, the calculation for the variance is laid out as follows.
Denote the centered random matrices
L
∗
∗
∗
∗
δ`,n b` b` hm Dn − b` b` hm Dn .
Z`,n :=
Q
The matrices Z`,n are independent not only for ` ∈ ∆n,1 but for every n as the set ∆n,1 is chosen
2 is as before the maximum of the operator norms of two
independently for each n. The variance σZ
quantities; firstly,
L2 X 2 ∗ 2
E δ`,n |b` h| kmn k22 b` b∗`
Q2
`,n
`,n
"
#
X
X
S
L
2
∗ 2
kmn k2
≤ max |b` h|
b` b∗` ≤ µ2max ,
Q `
Q
n
X
∗
E Z`,n Z`,n
≤
`
2 =
where the second last inequality follows from the fact that E δ`,n
obtained by applying the definitions in (26), and (10); secondly,
Q
L,
and the last inequality is
L2 X 2 ∗ 2
E δ`,n |b` h| kb` k22 Dn mm∗ Dn
Q2
`,n
`,n
#
"
L X ∗ 2 X
L
Lρ20
≤
Dn mm∗ Dn ≤ max kmn k22 ≤
,
|b` h|
Q
Q n
QN
n
X
∗
Z`,n ≤
E Z`,n
`
where we have used the facts that {b` }` is a complete orthonormal basis, and that the operator
norm of a block diagonal matrix with nth block being mn m∗n is upper bounded by maxn kmn k22 .
2 being the maximum of the operator
The last inequality is the result of (10). Thus, the variance σZ
norm of the two results above is bounded by
2
σZ
≤ µ2max
L
1
S
+ · ρ20 .
Q Q
N
The last ingredient required to apply the Bernstein inequality in Proposition 1 is then
L
L
δ`,n b` b∗` hm∗ Dn − b` b∗` hm∗ Dn ≤ 2 max kb` b∗` hm∗ Dn k
`,n Q
Q
r
L
L
S
1
≤ 2 max(kb` k2 |b∗` h|) · max kmn k2 ≤ 2 · µmax
· ρ0 √ .
n
Q `
Q
L
N
max kZ`,n k = max
`,n
`,n
41
DRAFT – December 18, 2017 – 2:26
With all the ingredients in place, an application of the uniform version of the Bernstein bound with
t = (β − 1) log(LN ) tells us that
k(L/Q) E A∗1 A1 (hm∗ ) − hm∗ k ≤
!
)
(s
r
S
1
L
S
L
µ2max + · ρ20
β log(LN ), 2 µmax ρ0
β log(LN ) .
max
Q Q
N
Q LN
The right hand side can be driven to the desired small number by choosing N ≥ Cρ20 (L/Q) log(LN ),
and L ≥ Cβµ2max S(L/Q) log(LN ) for an appropriately large constant C. The probability that the
above inequlaity holds is 1 − (LN )−β and follows by plugging in the choice t = (β − 1) log(LN ) in
Proposition 1.
5
Supporting Lemmas
This section proves Lemma 13, Corollary 3, and Lemmas 14, 15.
5.1
Proof of Lemma 13
Proof. We start with the proof of Lemma 13 that concerns bounding the quantity 6 kb∗`0 Sn‡ 0 ,p+1 Wp Dn0 k22
for p ≥ 2. Let q ∗ := b∗`0 Sn‡ 0 ,p+1 Wp Dn0 . The quantity can be expanded using the definition of Wp
for p ≥ 2 in (29)
h
i
q ∗ = b∗`0 Sn‡ 0 ,p+1 RA∗p Ap Sp‡ (Wp−1 ) − Wp−1 Dn0
h
i
= b∗`0 Sn‡ 0 ,p+1 RA∗p Ap Sp‡ (Wp−1 ) − R E A∗p Ap Sp‡ (Wp−1 ) Dn0 ,
where the second equality follows from a previously shown fact that E RA∗p Ap Sp‡ (Wp−1 ) = Wp−1 .
Thus q is just a zero-mean random vector that can be expanded further as a sum of zero-mean,
independent random vectors using the definition of map A∗p Ap in (19) as follows
q∗ =
X
(`,n)∈Γp
i
h
‡
‡
Wp−1 E φ`,n φ∗`,n Dn0 .
b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n − R b` b∗` Sn,p
(86)
From here on, we use the matrix Bernstein inequality to find the range in which the `2 -norm of the
random vector q lies with high probability. Let us define random vectors
∗
‡
z`,n
:= b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n Dn0
then by Proposition 2, it suffices to compute the following upper bound on the variance
X
X
X
2
∗
2
σz ≤ max
E z`,n z`,n ,
E kz`,n k2
≤
E kz`,n k22 .
(`,n)∈Γp
(`,n)∈Γp
(`,n)∈Γp
The use of `0 , and n0 as index variables is to avoid conflict with `, and n that are reserved to index the set Γp in
the proof below.
6
42
DRAFT – December 18, 2017 – 2:26
Note that the vectors z`,n can be rewritten as a scalar times a vector as follows
∗
‡
z`,n
= b∗` Sn,p
Wp−1 φ`,n · b∗`0 Sn‡ 0 ,p+1 R b` φ∗`,n Dn0 ,
and using (15), the vector part above can be expanded as
b∗`0 Sn‡ 0 ,p+1 R(b` φ∗`,n )Dn0 = (b∗` h)∗ (b∗`0 Sn‡ 0 ,p+1 h)φ∗`,n Dn0 + (φ∗`,n m)(b∗`0 Sn‡ 0 ,p+1 b` )m∗ Dn0
− (b∗` h)∗ (φ∗`,n m)(b∗`0 Sn‡ 0 ,p+1 h)m∗ Dn0 ,
and its the `2 -norm can then easily be shown to be upper bounded as
kb∗`0 Sn‡ 0 ,p+1 R b` φ∗`,n Dn0 k22
≤ 3|b∗` h|2 |b∗`0 Sn‡ 0 ,p+1 h|2 φ∗`,n Dn0 Dn∗ 0 φ`,n + 3|φ∗`,n m|2 |b∗`0 Sn‡ 0 ,p+1 b` |2 kmn0 k22
+ 3|b∗` h|2 |φ∗`,n m|2 |b∗`0 Sn‡ 0 ,p+1 h|2 kmn0 k22
≤ 3|b∗` h|2 |b∗`0 Sn‡ 0 ,p+1 h|2 φ∗`,n Dn0 Dn∗ 0 φ`,n + 6|φ∗`,n m|2 |b∗`0 Sn‡ 0 ,p+1 b` |2 kmn0 k22 ,
∗ 2
where the last line follows from the fact
that |b2` h| 2≤ 1. For a standard Gaussian vector g, it can
∗
2
∗
2
easily be verified that E |g x| |g y| ≤ 3kxk2 kyk2 . Using this fact, one has
‡
‡
Wp−1 Dn k22 ,
Wp−1 φ`,n |2 ≤ 3kmn k22 kb∗` Sn,p
E |φ∗`,n m|2 |b∗` Sn,p
and
N
X
n=1
`∈∆n,p
#
"
‡
Wp−1 φ`,n |2 ≤ 3K
max E φ∗`,n Dn0 Dn∗ 0 φ`,n |b∗` Sn,p
max kb∗` Sn‡ 0 ,p Wp−1 Dn0 k22 .
`∈∆n0 ,p
Note the change of index variable from n to n0 on the right hand side. Moreover, using (22)
X
`∈∆n,p
|b∗`0 Sn‡ 0 ,p+1 b` |2 ≤
5Q ∗ ‡
5Q
kb`0 Sn0 ,p+1 k22 ≤
kPb`0 k22 kSn‡ 0 ,p+1 k22 ,
4L
4L
where the inclusion of the projection operator P, defined in (14), on the vector b`0 is due to the
‡
fact that the rows of matrix Sn,p
are supported on Ω. Furthermore, using definitions in (9), and
(24), we can finally bound the above result as
X
`∈∆n,p
|b∗`0 Sn‡ 0 ,p+1 b` |2 ≤
20 2 S
µ
.
9 max Q
Putting the above identities together with
N
X
n=1
kmn k22 = kmk22 = 1,
and
X
`∈∆n,p
|b∗` h|2 ≤
5Q
5Q
khk22 =
4L
4L
one directly obtains
X
(`,n)∈Γp
kz`,n k22
=
N
X
X
n=1 `∈∆n,p
"
#
Q
∗ ‡
2
∗ ‡
2
≤ 4 K|b`0 Sn0 ,p+1 h|
max kb` Sn0 ,p Wp−1 Dn0 k2
`∈∆n0 ,p
L
S
2
2
∗ ‡
2
0
+ 14µmax kmn k2 max max kb` Sn,p Wp−1 Dn k2 := Π2p . (87)
n `∈∆n,p
Q
kz`,n k22
43
DRAFT – December 18, 2017 – 2:26
The only ingredient left to apply the Bernstein bound in Proposition 2 is the Orlicz norm of the
‡
summands z`,n . To this end, the ψ2 norm of the vector b∗`0 Sn,p+1
R(b` φ∗`,n )Dn0 can be evaluated as
follows
kb∗`0 Sn‡ 0 ,p+1 R(b` φ∗`,n )Dn0 kψ2 ≤ |b∗` h||b∗`0 Sn‡ 0 ,p+1 h|kφ`,n kψ2 + 2kφ∗`,n mkψ2 |b∗`0 Sn‡ 0 ,p+1 b` |kmn0 k2 .
Since φ`,n are Gaussian vectors, the discussion
on the Orlicz-norms in Section 3.4 tells us that
√
kφ∗`,n mkψ2 ≤ Ckmn k2 , and kφ`,n kψ2 ≤ C K. This means
√
kb∗`0 Sn‡ 0 ,p+1 R(b` φ∗`,n )Dn0 kψ2 ≤ C
K|b∗` h| · |b∗`0 Sn‡ 0 ,p+1 h| + kmn k2 |b∗`0 Sn‡ 0 ,p+1 b` |kmn0 k2
√
K|b∗` h||b∗`0 Sn‡ 0 ,p+1 h| + kPb` k2 kb∗`0 Sn‡ 0 ,p+1 k2 kmn k2 kmn0 k2 ,
≤C
where we have used the fact that kb∗`0 Sn‡ 0 ,p+1 b` k2 = kb∗`0 PSn‡ 0 ,p+1 Pb` k2 ≤ kPb` k2 kb∗`0 Sn‡ 0 ,p+1 k2 .
Again as φ`,n is a Gaussian vector, the Orlicz norm of its inner product with a fixed vector
‡
b∗` Sn,p
Wp−1 is
‡
b∗` Sn,p
Wp−1 φ`,n
ψ2
‡
≤ C b∗` Sn,p
Wp−1 Dn
2
.
Using these facts, we can show that random summands in (86) are sub-exponential vectors by
computing their ψ1 -norm and showing that it is bounded. Note that
i
h
‡
‡
Wp−1 E φ`,n φ∗`,n Dn0
Wp−1 φ`,n φ∗`,n − R b` b∗` Sn,p
b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn,p
ψ1
‡
‡
≤ 2 b∗`0 Sn0 ,p+1 R b` b∗` Sn,p
Wp−1 φ`,n φ∗`,n Dn0
,
ψ1
where the inequality follows by using the identity in (34). Now note that
‡
Wp−1 φ`,n φ∗`,n Dn0
max max b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn,p
n `∈∆n,p
ψ1
≤ max b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn‡ 0 ,p Wp−1 φ`,n0 φ∗`,n0
`∈∆n0 ,p
ψ1
,
and the change of indices from n to n0 is justified as φ`,n [k 0 ] is zero when k 0 ∈
/ {(K−1)n0 +1, . . . , n0 K}.
Now using the result in (35), we can write
b∗` Sn‡ 0 ,p Wp−1 φ`,n0
b∗`0 Sn‡ 0 ,p+1 R b` b∗` Sn‡ 0 ,p Wp−1 φ`,n0 φ∗`,n0
≤ b∗`0 Sn‡ 0 ,p+1 R b` φ∗`,n0
ψ1
ψ2
ψ2
√
∗
∗ ‡
∗ ‡
2
∗ ‡
≤
b` Sn0 ,p Wp−1 Dn0
.
K|b` h||b`0 Sn0 ,p+1 h| + kPb` k2 kb`0 Sn0 ,p+1 k2 kmn0 k2
2
Setting α = 1 in Proposition 2, one obtains the upper bound U1
2
µ0 K ∗ ‡
2
2
2
2
2 S
2 ∗ ‡
2
U1 = max max kz`,n kψ1 ≤ C
|b`0 Sn0 ,p+1 h| + µmax ρ0
kmn0 k2 kb`0 Sn0 ,p+1 k2
n `∈∆n,p
L
LN
#
"
max
`∈∆n0 ,p
≤C
2
b∗` Sn‡ 0 ,p Wp−1 Dn0
µ20 K ∗ ‡
S2
|b`0 Sn0 ,p+1 h|2 + µ4max ρ20 2 kmn0 k22
L
Q N
"
max
`∈∆n0 ,p
2
b∗` Sn‡ 0 ,p Wp−1 Dn0
44
DRAFT – December 18, 2017 – 2:26
2
2
#
:= Λ2p ,
(88)
where the last inequality follows from plugging in the bound for kb∗`0 Sn‡ 0 ,p+1 k22 calculated earlier in
the proof of this lemma. The logarithmic factor in the Bernstein bound can be crudely bounded
as follows
(QN )U12
≤ C log(LN )
log
σz2
for some constant C, which follows from the fact that U12 ≤ |Γp |σz2 . This completes all the ingredients to apply the Bernstein bound with t = β log(LN ) to obtain
kb∗`0 Sn‡ 0 ,p+1 Wp Dn0 k22 ≤ C max βΠ2p log(LN ), β 2 Λ2p log4 (LN ) ,
which holds with probability at least 1 − (LN )−β . This completes the proof of the lemma.
5.2
Proof of Corollary 3
Proof. Note that from (68), we can equivalently write
s∗`0 ,n0 := b∗`0 Sn‡ 0 ,2 [RA∗1 A1 (L/Q)P(hm∗ ) − E RA∗1 A1 (L/Q)P(hm∗ )] Dn0 .
using the fact that hm∗ ∈ Ω. In comparison, Lemma 13 was concerned with bounding the `2 norm
of the term
h
i
b∗`0 Sn‡ 0 ,p+1 RA∗p Ap Sp‡ (Wp−1 ) − E RA∗p Ap Sp‡ (Wp−1 ) Dn0 .
All we need to do over here is to replace Sn0 ,p+1 , Ap , Sp‡ , and Wp−1 with Sn0 ,2 , A1 , (L/Q)P, and
hm∗ and repeat the same argument as in the proof of Lemma 13. This leads to the bound on the
`2 norm of s`,n in the statement of the corollary in terms of Π21 , and Λ21 defined in (66), and (67),
respectively. Compared to Lemma 13 the Π2p , and Λ2p are replaced in the result by Π21 , and Λ21 ,
‡
respectively. The quantities Π21 and Λ21 are just Π2p , and Λ2p evaluated at p = 1, and afterwards Sn,1
therein replaced by (L/Q)P.
5.3
Proof of Lemma 14
Proof. Using the definition (70), and the triangle inequality, we have
kr`∗0 ,n0 k2 ≤
L ∗ ‡
b 0 S 0 R E A∗1 A1 (hm∗ )Dn0
Q ` n ,2
2
+ kb∗`0 Sn‡ 0 ,2 hm∗ Dn0 k2 ,
where using the definition of Sn,1 defined in Section 3.1, we have
"
#
X
L ∗ ‡
L ∗ ‡
∗
∗
∗
b 0 S 0 R E A1 A1 (hm )Dn0 = b`0 Sn0 ,2
R (Sn,1 hm Dn ) Dn0 =
Q ` n ,2
Q
n
"
#
X
L ∗ ‡
∗
∗
∗
2
∗
∗
2
b 0S 0
(h Sn,1 h) hm Dn + Sn,1 hm kmn k2 − (h Sn,1 h) hm kmn k2 Dn0 ≤
Q ` n ,2 n
"
#
X
X
X
L ∗ ‡
5Q
3Q
b 0S 0
hm∗
Dn + max Sn,1 hm∗
kmn k22 −
hm∗
kmn k22 Dn0 ,
n
Q ` n ,2 4L
4L
n
n
n
45
DRAFT – December 18, 2017 – 2:26
(89)
where the second last equality results from the definition of R in (15), and from (24) that directly
imply that
3Q
5Q
≤ h∗ Sn,1 h ≤
,
4L
4L
P
P
for the choice of Q in the lemma. Since n kmn k22 = 1, and n Dn = IKN , the result above
simplifies to
Q
L ∗ ‡
L ∗ ‡
∗
∗
∗
∗
b 0 S 0 R E A1 A1 (hm )Dn0 ≤ b`0 Sn0 ,2
hm + max Sn,1 hm Dn0 .
n
Q ` n ,2
Q
2L
Finally, the operator norm of the above quantity returns
L
L ∗ ‡
1 ∗ ‡
∗
∗
∗ ‡
b 0 S 0 R E A1 A1 (hm )Dn0 ≤
|b 0 S 0 h| + max |b`0 Sn0 ,2 Sn,1 h| kmn0 k2 .
Q ` n ,2
2 ` n ,2
Q n
2
Using (89), we obtain
kr`0 ,n0 k2 ≤
3 ∗ ‡
L
∗ ‡
|b 0 S 0 h| + max |b`0 Sn0 ,2 Sn,1 h| kmn0 k2 ,
2 ` n ,2
Q n
and squaring both sides results in
9 ∗ ‡
L2
kr`0 ,n0 k22 ≤
|b`0 Sn0 ,2 h|2 + 2 2 max |b∗`0 Sn‡ 0 ,2 Sn,1 h|2 kmn0 k22 .
2
Q n
This completes the proof.
Lemma 15. Let φ`,n be as in defined in (5), and X be a fixed matrix. Then
E |hX, b` φ∗`,n i|2 φ`,n φ∗`,n 4 3kb∗` XDn k22 Dn .
Proof. Note that
E |hX, b` φ∗`,n i|2 φ`,n φ∗`,n = E |b∗` Xφ`,n |2 φ`,n φ∗`,n
2
X
= E
{b∗` Xφ`,n }[k] φ`,n φ∗`,n
k∼K n
=
kb∗` XDn k22 Dn
+ 2b∗` XX ∗ b` Dn 4 3kb∗` XDn k22 Dn .
References
[1] A. Aghasi, S. Bahmani, and J. Romberg. A tightest convex envelope heuristic to row sparse
and rank one matrices. In GlobalSIP, page 627, 2013.
[2] A. Ahmed, B. Recht, and J. Romberg. Blind deconvolution using convex programming. IEEE
Trans. Inform. Theory, 60(3):1711–1732, 2014.
46
DRAFT – December 18, 2017 – 2:26
[3] S. Bahmani and J. Romberg. Lifting for blind deconvolution in random mask imaging: Identifiability and convex relaxation. SIAM J. Imag. Sci., 8(4):2203–2238, 2015.
[4] S. Burer and R.D.C. Monteiro. A nonlinear programming algorithm for solving semidefinite
programs via low-rank factorization. Math. Prog., 95(2):329–357, 2003.
[5] E. Candès. The restricted isometry property and its implications for compressed sensing.
Comptes Rendus Mathematique, 346(9):589–592, 2008.
[6] E. Candès, Y. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via matrix completion.
SIAM Review, 57(2):225–251, 2015.
[7] E. Candès and B. Recht. Exact matrix completion via convex optimization. Found. Comput.
Math., 9(6):717–772, 2009.
[8] E. Candès and J. Romberg. Sparsity and incoherence in compressive sampling. Inverse problems, 23(3):969, 2007.
[9] E. Candès, J. Romberg, and T. Tao. Robust uncertainty principles: Exact signal reconstruction
from highly incomplete frequency information. IEEE Trans. Inform. Theory, 52(2):489–509,
February 2006.
[10] E. Candès, J. Romberg, and T. Tao. Stable signal recovery from incomplete and inaccurate
measurements. Commun. Pure Appl. Math., 59(8):1207–1223, 2006.
[11] E. Candès, T. Strohmer, and V. Voroninski. Phaselift: Exact and stable signal recovery from
magnitude measurements via convex programming. Commun. Pure Appl. Math., 66(8):1241–
1274, 2013.
[12] Sunav Choudhary and Urbashi Mitra. Sparse blind deconvolution: What cannot be done. In
IEEE Int’l. Symp. Inform. Theory (ISIT), pages 3002–3006, 2014.
[13] M. Fazel. Matrix rank minimization with applications. PhD thesis, Stanford University, March
2002.
[14] M. Fazel, H. Hindi, and S.P. Boyd. A rank minimization heuristic with application to minimum
order system approximation. In American Control Conf., Proc., volume 6, pages 4734–4739.
IEEE, 2001.
[15] D. Gross. Recovering low-rank matrices from few coefficients in any basis. IEEE Trans. Inform.
Theory, 57(3):1548–1566, 2011.
[16] M. Gürelli, C. Nikias, et al. EVAM: An eigenvector-based algorithm for multichannel blind
deconvolution of input colored signals. IEEE Trans. Sig. Process., 43(1):134–149, 1995.
[17] V. Koltchinskii, K. Lounici, and A. Tsybakov. Nuclear-norm penalization and optimal rates
for noisy low-rank matrix completion. Ann. Stat., 39(5):2302–2329, 2011.
[18] A. Levin, Y. Weiss, F. Durand, and W. Freeman. Understanding blind deconvolution algorithms. IEEE Trans. Patt. Analys. Mach. Intell., 33(12):2354–2367, 2011.
[19] X. Li, S. Ling, T. Strohmer, and K. Wei. Rapid, robust, and reliable blind deconvolution via
nonconvex optimization. arXiv preprint arXiv:1606.04933, 2016.
47
DRAFT – December 18, 2017 – 2:26
[20] Shuyang Ling and Thomas Strohmer. Blind deconvolution meets blind demixing: Algorithms
and performance bounds. IEEE Trans. Inform. Theory, 2017.
[21] H. Liu, G. Xu, L. Tong, and T. Kailath. Recent developments in blind channel equalization:
From cyclostationarity to subspaces. IEEE Trans. Sig. Process., 50(1):83–99, 1996.
[22] S. Oymak, A. Jalali, M. Fazel, Y. Eldar, and B. Hassibi. Simultaneously structured models
with application to sparse and low-rank matrices. IEEE Trans. Inform. Theory, 61(5):2886–
2908, 2015.
[23] B. Recht. A simpler approach to matrix completion. J. Mach. Learn. Res., 12(12):3413–3430,
2011.
[24] B. Recht, M. Fazel, and P. Parrilo. Guaranteed minimum-rank solutions of linear matrix
equations via nuclear norm minimization. SIAM Rev., 52(3):471–501, 2010.
[25] J. Romberg, N. Tian, and K. Sabra. Multichannel blind deconvolution using low rank recovery.
In SPIE Defense, Security, and Sensing, pages 87500E–87500E. Int. Soc. Opt. Phot., 2013.
[26] M. Rudelson and R. Vershynin. On sparse reconstruction from fourier and gaussian measurements. Commun. Pure Appl. Math., 61(8):1025–1045, 2008.
[27] M. Schmidt. minFunc: unconstrained differentiable multivariate optimization in Matlab. http:
//www.di.ens.fr/~mschmidt/Software/minFunc.html, 2012.
[28] L. Tong and S. Perreau. Multichannel blind identification: From subspace to maximum likelihood methods. Proc. IEEE, 86:1951–1968, 1998.
[29] L. Tong, G. Xu, and T. Kailath. Blind identification and equalization based on second-order
statistics: A time domain approach. IEEE Trans. Inform. Theory, 40(2):340–349, 1994.
[30] J. Tropp. User-friendly tail bounds for sums of random matrices. Found. Comput. Math.,
12(4):389–434, 2012.
[31] R. Vershynin. Compressed sensing: theory and applications. Cambridge University Press,
2012.
[32] G. Watson. Characterization of the subdifferential of some matrix norms. Linear Algebra and
its Appl., 170:33–45, 1992.
[33] G. Xu, H. Liu, L. Tong, and T. Kailath. A least-squares approach to blind channel identification. IEEE Trans. Sig. Process., 43(12):2982–2993, 1995.
48
DRAFT – December 18, 2017 – 2:26
| 7cs.IT
|
JSJ decompositions of groups
Vincent Guirardel, Gilbert Levitt
arXiv:1602.05139v2 [math.GR] 28 Jul 2017
Abstract
This is an account of the theory of JSJ decompositions of finitely generated groups,
as developed in the last twenty years or so.
We give a simple general definition of JSJ decompositions (or rather of their BassSerre trees), as maximal universally elliptic trees. In general, there is no preferred JSJ
decomposition, and the right object to consider is the whole set of JSJ decompositions,
which forms a contractible space: the JSJ deformation space (analogous to Outer
Space).
We prove that JSJ decompositions exist for any finitely presented group, without
any assumption on edge groups. When edge groups are slender, we describe flexible vertices of JSJ decompositions as quadratically hanging extensions of 2-orbifold
groups.
Similar results hold in the presence of acylindricity, in particular for splittings
of torsion-free CSA groups over abelian groups, and splittings of relatively hyperbolic
groups over virtually cyclic or parabolic subgroups. Using trees of cylinders, we obtain
canonical JSJ trees (which are invariant under automorphisms).
We introduce a variant in which the property of being universally elliptic is replaced
by the more restrictive and rigid property of being universally compatible. This yields
a canonical compatibility JSJ tree, not just a deformation space. We show that it
exists for any finitely presented group.
We give many examples, and we work throughout with relative decompositions
(restricting to trees where certain subgroups are elliptic).
Introduction
JSJ decompositions first appeared in 3-dimensional topology with the theory of the characteristic submanifold by Jaco-Shalen and Johannson [JS79, Joh79] (the terminology JSJ
was popularized by Sela). We start with a quick review (restricting to manifolds without
boundary).
From 3-manifolds to groups
Let M be a closed orientable 3-manifold. Given a finite collection of disjoint embedded
2-spheres, one may cut M open along them, and glue balls to the boundary of the pieces
to make them boundaryless. This expresses M as a connected sum of closed manifolds Mi .
The prime decomposition theorem (Kneser-Milnor) asserts that one may choose the spheres
so that each Mi is either irreducible (Mi 6= S 3 , and every embedded 2-sphere bounds a
ball) or homeomorphic to S 2 × S 1 ; moreover, up to a permutation, the summands Mi are
uniquely determined up to homeomorphism.
On the group level, one obtains a decomposition of G = π1 (M ) as a free product
G = G1 ∗ · · · ∗ Gp ∗ Fq , with Gi the fundamental group of an irreducible Mi and Fq a free
group of rank q coming from the S 2 × S 1 summands. This decomposition is a Grushko
decomposition of G, in the following sense: each Gi is freely indecomposable (it cannot be
written as a non-trivial free product), non-trivial, and not isomorphic to Z (non-triviality of
Gi is guaranteed by the Poincaré conjecture, proved by Perelman). Any finitely generated
1
group has a Grushko decomposition, with q well-defined and the Gi ’s well-defined up to
conjugacy (and a permutation).
The prime decomposition implies that one should focus on irreducible manifolds. Since
all spheres bound balls, one now considers embedded tori. In order to avoid trivialities (such
as a torus bounding a tubular neighborhood of a curve), tori should be incompressible: the
embedding T 2 → M induces an injection on fundamental groups.
The theory of the characteristic submanifold now says that, given an irreducible M ,
there exists a finite family T of disjoint non-parallel incompressible tori such that each
component Nj of the manifold (with boundary) obtained by cutting M along T is either
atoroidal (every incompressible torus is boundary parallel) or a Seifert fibered space, i.e.
a 3-manifold having a singular fibration by circles over a 2-dimensional surface Σ (better
viewed as a 2-dimensional orbifold).1 Moreover, any incompressible torus may be isotoped
to be disjoint from T .
With groups in mind, let us point out an important feature of this decomposition. If
two incompressible tori cannot be made disjoint by an isotopy, they may be isotoped to be
contained in a Seifert piece. Conversely, in a Seifert fibered space, preimages of intersecting
simple curves on Σ are intersecting tori.
We thus see that the presence of intersecting tori in M forces some surface Σ to appear.
One of the remarkable facts about JSJ theory for groups is that a similar phenomenon
occurs. For instance, if a finitely generated one-ended group admits two splittings over
Z that “intersect” each other in an essential way,2 then it must contain the fundamental
group of a compact surface, attached to the rest of the group along the boundary (see
Theorem 4 below and Section 6, in particular Proposition 6.25).
We also point out that the family T mentioned above is unique up to isotopy. On the
other hand, a family of spheres defining the prime decomposition is usually not unique.
Similarly, the Grushko decompositions of a group usually form a large outer space [CV86,
GL07b], whereas one may often construct canonical splittings of one-ended groups, which
are in particular invariant under automorphisms (see Theorems 6 and 8 below).
These topological ideas were carried over to group theory by Kropholler [Kro90] for
some Poincaré duality groups of dimension at least 3, and by Sela for torsion-free hyperbolic groups [Sel97b]. Constructions of JSJ decompositions were given in more general
settings by many authors (Rips-Sela [RS97], Bowditch [Bow98], Dunwoody-Sageev [DS99],
Fujiwara-Papasoglu [FP06], Dunwoody-Swenson [DS00], Scott-Swarup [SS03], PapasogluSwenson [PS09]. . . ). This has had a vast influence and range of applications, from the isomorphism problem and the structure of the group of automorphisms of hyperbolic groups,
to diophantine geometry over groups, and many others.
In this group-theoretical context, one has a finitely generated group G and a class
of subgroups A (such as cyclic groups, abelian groups, ...), and one tries to understand
splittings (i.e. graph of groups decompositions) of G over groups in A. The family of tori
T of the 3-manifold is replaced by a splitting of G over groups in A. The authors construct
a splitting enjoying a long list of properties, rather specific to each case.
Our first goal is to give a simple general definition of JSJ decompositions stated by
means of a universal maximality property, together with general existence and uniqueness statements in terms of deformation spaces (see below). The JSJ decompositions
constructed in [RS97, Bow98, DS99, FP06] are JSJ decompositions in our sense (see Subsection 2.6).
The regular neighbourhood of [SS03] is of a different nature. In [DS00, SS03], one looks
1
Thurston’s geometrization conjecture, whose proof was completed by Perelman, asserts that each Nj
has a geometric structure. In particular, every atoroidal Nj with infinite fundamental group is hyperbolic
(its interior admits a complete metric with finite volume and constant curvature −1).
2
More precisely: the edge groups of each splitting should be hyperbolic in the Bass-Serre tree of the
other splitting.
2
at almost invariant sets rather than splittings, in closer analogy to a 3-manifold situation
where one wants to understand all immersed tori, not just the embedded ones. One obtains
a canonical splitting of G rather than just a canonical deformation space. See Parts IV
and V for canonical splittings, and [GL10] for the relation between [SS03] and usual JSJ
decompositions.
Definition of JSJ decompositions
To motivate the definition, let us first consider free decompositions of a group G, i.e.
decompositions of G as the fundamental group of a graph of groups with trivial edge
groups, or equivalently actions of G on a simplicial tree T with trivial edge stabilizers.
Let G = G1 ∗ · · · ∗ Gp ∗ Fq be a Grushko decomposition, as defined above (Gi is nontrivial, not Z, freely indecomposable). One may view G as the fundamental group of one
of the graphs of groups pictured on Figure 1. The corresponding Bass-Serre trees have
trivial edge stabilizers, and the vertex stabilizers are precisely the conjugates of the Gi ’s;
we call a tree with these properties a Grushko tree (if G is freely indecomposable, Grushko
trees are points).
t1
G2
t2
..
.
G1
tq
Gp
G1
G2
..
.
b
b
b
G3
b
G5
b
G4
p = 5, q = 6
Figure 1: Graph of groups decompositions corresponding to two Grushko trees.
Since the Gi ’s are freely indecomposable, Grushko trees T0 have the following maximality property: if T is any tree on which G acts with trivial edge stabilizers, Gi fixes a
point in T , and therefore T0 dominates T in the sense that there is a G-equivariant map
T0 → T . In other words, among free decompositions of G, a Grushko tree T0 is as far as
possible from the trivial tree (a point): its vertex stabilizers are as small as possible (they
are conjugates of the Gi ’s). This maximality property does not determine T0 uniquely, as
it is shared by all Grushko trees; we will come back to this key fact later, when discussing
uniqueness.
When more general decompositions are allowed, for instance when one considers splittings over Z, there may not exist a tree with the same maximality property. The fundamental example is the following. Consider an orientable closed surface Σ, and two simple
closed curves c1 , c2 in Σ with non-zero intersection number. Let Ti be the Bass-Serre tree
of the associated splitting of G = π1 (Σ) over Z ' π1 (ci ). Since c1 and c2 have positive
intersection number, π1 (c1 ) is hyperbolic in T2 (it does not fix a point) and vice-versa.
Using the fact that π1 (Σ) is freely indecomposable, it is an easy exercise to check that
there is no splitting of π1 (Σ) which dominates both T1 and T2 . In this case there is no
hope of having a splitting over cyclic groups similar to T0 above.
To overcome this difficulty, one restricts to universally elliptic splittings, defined as
follows.
We consider trees with an action of a finitely generated group G, and we require that
edge stabilizers be in A (a given family of subgroups of G, closed under conjugating and
taking subgroups); we call such a tree an A-tree. By Bass-Serre theory, this corresponds
to splittings of G over groups in A. Unless otherwise indicated, all trees are assumed to
be A-trees.
Definition. An A-tree is universally elliptic if its edge stabilizers are elliptic in every
A-tree.
3
Recall that H is elliptic in T if it fixes a point in T (in terms of graphs of groups, H is
contained in a conjugate of a vertex group). Free decompositions are universally elliptic,
but the trees T1 , T2 introduced above are not.
Definition. A JSJ decomposition (or JSJ tree) of G over A is an A-tree T such that:
• T is universally elliptic;
• T dominates any other universally elliptic tree T 0 .
We call the quotient graph of groups Γ = T /G a JSJ splitting, or a JSJ decomposition.
Recall that T dominates T 0 if there is an equivariant map T → T 0 ; equivalently, any
group which is elliptic in T is also elliptic in T 0 . The second condition in the definition is a
maximality condition expressing that vertex stabilizers of T are as small as possible (they
are elliptic in every universally elliptic tree).
If A consists of all subgroups with a given property (being cyclic, abelian, slender, ...),
we refer to, say, cyclic trees, cyclic JSJ decompositions when working over A.
When A only contains the trivial group, JSJ trees are the same as Grushko trees. If
G = π1 (Σ) as above, and A is the family of cyclic subgroups, the JSJ decomposition is
trivial (the point is the only JSJ tree).
Existence.
JSJ trees do not always exist: the finitely generated inaccessible group D constructed by
Dunwoody [Dun93] has no JSJ tree over finite groups, and D × Z is a one-ended group
with no JSJ decomposition over virtually cyclic subgroups. On the other hand, it follows
rather easily from Dunwoody’s accessibility [Dun85] that a finitely presented group has
JSJ decompositions over any class A of subgroups (we emphasize that no assumption on
A, such as smallness, is needed).
Theorem 1 (Theorem 2.16). Let A be an arbitrary family of subgroups of G, stable under
taking subgroups and under conjugation. If G is finitely presented, it has a JSJ decomposition over A. In fact, there exists a JSJ tree whose edge and vertex stabilizers are finitely
generated.
In Part IV we shall present a different way of constructing JSJ decompositions, based on
Sela’s acylindrical accessibility, which applies in some more general situations (the existence
of such JSJ decompositions for limit groups is mentioned and used in [Sel01], and we give a
complete proof). We will give more details later in this introduction, but we mention two
typical results here. A group is CSA if maximal abelian subgroups are malnormal, small
if it has no free non-abelian subgroup (see Subsection 1.5.2 for variations).
Theorem 2 (Theorem 9.5). Let G be a torsion-free finitely generated CSA group. There
is a JSJ decomposition of G over abelian subgroups.
Theorem 3 (Theorem 9.18). Let G be hyperbolic relative to a finite family of finitely
generated small subgroups. If A is either the family of all virtually cyclic subgroups of G,
or the family of all small subgroups, there is a JSJ decomposition of G over A.
Uniqueness.
JSJ trees are not unique. . . Returning to the example of free decompositions, one
obtains trees with the same maximality property as T0 by precomposing the action of G on
T0 with any automorphism of G. One may also change the topology of the quotient graph
T0 /G (see Figure 1). The canonical object is not a single tree, but the set of all Grushko
trees (trees with trivial edge stabilizers, and non-trivial vertex stabilizers conjugate to the
Gi ’s), a deformation space.
4
Definition (Deformation space [For02]). The deformation space D of a tree T is the set
of trees T 0 such that T dominates T 0 and T 0 dominates T . Equivalently, two trees are in
the same deformation space if and only if they have the same elliptic subgroups.
More generally, given a family of subgroups Ae ⊂ A, one considers deformation spaces
e
over Ae by restricting to trees in D with edge stabilizers in A.
For instance, Culler-Vogtmann’s outer space (the set of free actions of Fn on trees) is
a deformation space. Just like outer space, any deformation space may be viewed as a
complex in a natural way, and it is contractible (see [Cla05, GL07a]).
If T is a JSJ tree, another tree T 0 is a JSJ tree if and only if T 0 is universally elliptic, T
dominates T 0 , and T 0 dominates T . In other words, T 0 should belong to the deformation
space of T over Aell , where Aell is the family of universally elliptic groups in A.
Definition. If non-empty, the set of all JSJ trees is a deformation space over Aell called
the JSJ deformation space (of G over A). We denote it by DJSJ .
The canonical object is therefore not a particular JSJ decomposition, but the JSJ
deformation space. For instance, the JSJ deformation space of Fn over any A is outer
space (see Subsection 3.1).
It is a general fact that two trees belong to the same deformation space if and only if
one can pass from one to the other by applying a finite sequence of moves of certain types,
see [For02, GL07a, For06, CF09] and Remark 2.13 (this may be viewed as a connectedness
statement, but as mentioned above deformation spaces are actually contractible). The
statements about uniqueness of the JSJ decomposition up to certain moves which appear
in [RS97, DS99, FP06], as well as the non-uniqueness results of [For03], are special cases
of this general fact.
Another general fact is the following: two trees belonging to the same deformation
space over A have the same vertex stabilizers, provided one restricts to groups not in A. It
thus makes sense to study vertex stabilizers of JSJ trees. We do so in Part III (see below).
If A is invariant under the group of automorphisms of G (in particular if A is defined
by restricting the isomorphism type), then so is the deformation space DJSJ . As in the
case of outer space, precomposing actions on trees with automorphisms of G yields an
action of Aut(G) on DJSJ , a contractible complex. This action factors through an action
of Out(G), thus providing information about Out(G) [CV86, MM96, GL07b, Cla09].
We stress once again that, in general, the canonical object associated to G and A is
the deformation space consisting of all JSJ trees, and it may be quite large.
. . . but sometimes there is a canonical JSJ tree. In some nice situations one can
construct a canonical JSJ tree T in DJSJ .
By canonical, we essentially mean “defined in a natural, uniform way”. In particular,
given any isomorphism α : G → G0 sending A to A0 , canonicity implies that there is a
unique α-equivariant isomorphism Hα : T → T 0 between the canonical JSJ trees. Applying
this with G0 = G (assuming that A is invariant under automorphisms), one gets an action
of Aut(G) on T . The canonical tree T is a fixed point for the action of Out(G) on the JSJ
deformation space.
The existence of such a canonical splitting gives precise information about Out(G), see
[Sel97b, Lev05a, GL11] for applications. A particularly nice example, due to Bowditch
[Bow98], is the construction of a canonical JSJ decomposition of a one-ended hyperbolic
group over virtually cyclic subgroups, from the structure of local cut points in the Gromov
boundary.
This is not a consequence of the sole fact that G is one-ended, and that one considers
splittings over virtually cyclic groups: the cyclic splittings of G = Fn × Z are in one-toone correspondence with the free splittings of Fn , and the JSJ deformation space of G is
5
the outer space of Fn . Generalized Baumslag-Solitar groups are other striking examples
where a strong non-uniqueness occurs, with surprising algebraic consequences like the
fact (due to Collins-Levin [CL83]) that the group Out(BS(2, 4)) of outer automorphisms
of the Baumslag-Solitar group BS(2, 4) = ha, t | ta2 t−1 = a4 i is not finitely generated
[For06, Cla09].
A method to produce a canonical tree from a deformation space will be given in Part
IV, using a construction called the tree of cylinders [GL11]. In particular, it yields canonical JSJ decompositions of one-ended CSA groups and relatively hyperbolic groups (see
Theorems 6 and 8 below). The compatibility JSJ decomposition introduced below also
yields a canonical tree.
Description: Quadratically hanging vertex groups.
As mentioned above, Grushko trees have a strong maximality property: their vertex stabilizers are elliptic in any free splitting of G. This does not hold any longer when one
considers JSJ decompositions over infinite groups, in particular cyclic groups: a vertex
stabilizer Gv of a JSJ tree may fail to be elliptic in some splitting (over the chosen family
A).
If this happens, we say that the vertex stabilizer Gv (or the corresponding vertex v, or
the vertex group of the quotient graph of groups) is flexible. The other stabilizers (which
are elliptic in every splitting over A) are called rigid. In particular, all vertices of Grushko
trees are rigid (because their stabilizers are freely indecomposable). On the other hand, in
the example of G = π1 (Σ), the unique vertex stabilizer Gv = G is flexible.
Because all JSJ decompositions lie in the same deformation space over Aell , they have
the same flexible vertex stabilizers (and the same rigid vertex stabilizers not in A).
The essential feature of JSJ theory is the description of flexible vertices, in particular
the fact that flexible vertex stabilizers are often “surface-like” ([RS97, DS99, FP06]; see
below, and Theorem 6.2, for more precise statements). In other words, the example of
trees T1 , T2 given above using intersecting curves on a surface is often the only source of
flexible vertices.
This is formalized through the notion of quadratically hanging (QH) groups, a terminology due to Rips and Sela [RS97].
For cyclic splittings of a torsion-free group, a vertex group Gv is QH if it may be viewed
as the fundamental group of a (possibly non-orientable) compact surface with boundary
Σ, in such a way that any incident edge group is trivial or contained (up to conjugacy) in a
boundary subgroup, i.e. the fundamental group B = π1 (C) of a boundary component C of
Σ. The terminology “quadratically hanging” describes the way in which π1 (Σ) is attached
to the rest of the group, since boundary subgroups are generated by elements which are
quadratic words in a suitable basis of the free group π1 (Σ).
In a more general setting, one extends this notion as follows: Gv is an extension
1 → F → Gv → π1 (Σ) → 1, where Σ is a compact hyperbolic 2-orbifold (usually with
boundary), and F is an arbitrary group called the fiber. The condition on the attachment
is that the image of any incident edge group in π1 (Σ) is finite or contained in a boundary
subgroup (see Section 5 for details).
Recall that a group is slender if all its subgroups are finitely generated.
Theorem 4 ([FP06], see Corollary 6.3). Let A be the class of all slender subgroups of a
finitely presented group G. Let Gv be a flexible vertex group of a JSJ decomposition of G
over A. Then Gv is either slender or QH with slender fiber.
One may replace A by a subfamily, provided that it satisfies a suitable stability condition. In particular, A may be the family of cyclic subgroups, virtually cyclic subgroups,
polycyclic subgroups. Failure of this stability condition explains why the result does not
6
apply to JSJ decompositions over abelian groups in general (see Subsection 5.5). On the
other hand, non-abelian flexible vertex groups in Theorem 2 are QH with trivial fiber, and
in Theorem 3 non-small flexible vertex groups are QH with finite fiber.
Theorem 4 says that flexible subgroups of the JSJ decomposition are QH, but one can
say more: they are maximal in the following sense.
Proposition 5 (see Corollary 5.30). Let G be one-ended, and let A be the class of all
virtually cyclic groups. Let Q be a QH vertex stabilizer with finite fiber in an arbitrary
A-tree.
Then Q is contained in a QH vertex stabilizer of any cyclic JSJ decomposition of G.
This does not hold without the one-endedness assumption: non-abelian free groups
contain many non-trivial QH subgroups.
Our proof of Theorem 4 is based on the approach by Fujiwara and Papasoglu using
products of trees [FP06], but with several simplifications. In particular, we do not have to
construct a group enclosing more than two splittings.
The characteristic property of slender groups is that, whenever they act on a tree with
no fixed point, there is an invariant line. Using these lines, one may construct subsurfaces
in the product of two trees and this explains (at least philosophically) the appearance of
surfaces, hence of QH vertex groups, in Theorem 4. This is the content of Proposition
6.25.
This approach does not work if edge groups are not slender (unless there is acylindricity,
as in Theorem 8 below), and the following problem is open.
Problem. Describe flexible vertices of JSJ decompositions of a finitely presented group
over small subgroups.
Relative decompositions
For many applications, it is important not to consider all A-trees, but only those in which
subgroups of G belonging to a given family H are elliptic (i.e. every H ∈ H fixes a point in
the tree). We say that such a tree is relative to H, and we call it an (A, H)-tree. Working
in a relative setting is important for applications (see e.g. [Sel01, Section 9], [Pau04, Per11],
and Theorems 6 and 8 below), and also needed in our proofs. In the proof of Theorem 4
describing flexible vertex groups Gv of slender JSJ decompostions, for instance, we do not
work with splittings of G, but rather with splittings of Gv relative to incident edge groups.
Definitions extend naturally to relative trees: a tree is universally elliptic if its edge
groups are elliptic in every (A, H)-tree, and a JSJ decomposition of G over A relative to
H is an (A, H)-tree which is universally elliptic and maximal for domination.
The theorems stated above remain true, but H must be a finite family of finitely
generated subroups in Theorems 1 and 4, and one must take H into account when defining
QH vertices (see Definition 5.13).
In this text we consistently work in a relative setting, so the reader does not have to
take it as an act of faith that arguments also work for relative trees. For simplicity, though,
we limit ourselves to the non-relative case in this introduction (except in Theorems 6 and
8 where obtaining canonical trees definitely requires working in a relative setting).
Acylindricity
We have explained how Dunwoody’s accessibility may be used to construct the JSJ deformation space, which contains no preferred tree in general. In Part IV we use a different
approach, based on the trees of cylinders of [GL11] and on Sela’s acylindrical accessibility
7
[Sel97a], which yields more precise results when applicable. Unlike Dunwoody’s accessibility, acylindrical accessibility only requires finite generation of G (see Subsection 1.6).
Let G be a one-ended group, and T any tree with (necessarily infinite) virtually cyclic
edge stabilizers. Say that two edges of T are equivalent if their stabilizers are commensurable (i.e. have infinite intersection). One easily checks that equivalence classes are
connected subsets of T , which we call cylinders. Two distinct cylinders intersect in at most
one point. Dual to this partition of T into subtrees is another tree Tc called the tree of
cylinders (it is sometimes necessary to use a collapsed tree of cylinders, see Definition 7.2,
but we neglect this here).
This construction works with other equivalence relations among infinite edge stabilizers.
Here are two examples: G is a relatively hyperbolic group, and the equivalence relation
is co-elementarity among infinite elementary subgroups (A ∼ B if and only if hA, Bi is
elementary, i.e. virtually cyclic or parabolic), or G is a torsion-free CSA group and the
relation is commutation among non-trivial abelian subgroups.
Z
b
Z
b
Z
b
Z
Z
Z
b
Z
Z2
Z
Figure 2: A JSJ splitting of a toral relatively hyperbolic group and its tree of cylinders.
Here is an example (which already appears in [GL11]). Let T be the Bass-Serre tree of
the graph of groups Γ pictured on the left of Figure 2 (all punctured tori have the same
boundary subgroup, equal to the edge groups of Γ). The fundamental group of Γ is a
torsion-free one-ended group G (which is toral relatively hyperbolic and CSA), and T is a
cyclic JSJ decomposition of G. In this case all previous equivalence relations on the set of
edge stabilizers of T (commensurability, co-elementarity, commutation) reduce to equality,
and the quotient graph of groups of the tree of cylinders Tc is pictured on the right of
Figure 2; there is a new vertex group, isomorphic to Z2 .
There are three main benefits in passing from the JSJ tree T to Tc . First, two trees
in the same deformation space always have the same tree of cylinders. In particular, the
tree Tc is invariant under all automorphisms of G (whereas T is only invariant up to
deformation). Second, Tc is acylindrical: all segments of length 3 have trivial stabilizer
(whereas T contains lines with infinite cyclic pointwise stabilizer). Third, trees of cylinders
enjoy nice compatibility properties, which will be important later. There is a small price
to pay in order to replace T by the better tree Tc , namely changing the deformation space
by creating a new vertex group Z2 .
This example is typical of the method we use in Part IV to prove Theorems 2 and 3.
Using trees of cylinders, we show that one may associate to any tree T an acylindrical
tree T ∗ in such a way that groups elliptic in T are also elliptic in T ∗ , and groups elliptic
in T ∗ but not in T are small (T ∗ is smally dominated by T in the sense of Definition
7.10). Applying acylindrical accessibility to trees T ∗ is one of the key ingredients in our
construction of JSJ decompositions, but the proof involves much more than a bound on
the complexity of acylindrical splittings.
In the example, the tree T is a cyclic JSJ tree, but the preferred tree Tc is in a slightly
different deformation space; it is a JSJ tree relative to the Z2 subgroup. In general, we
show:
Theorem 6 (Theorem 9.5). Let G be a torsion-free finitely generated one-ended CSA
group. There is a canonical JSJ tree over abelian subgroups relative to all non-cyclic abelian
subgroups. Its non-abelian flexible vertex stabilizers are QH with trivial fiber.
8
The CSA property (maximal abelian subgroups are malnormal) holds for any torsionfree hyperbolic group Γ, and any Γ-limit group. If Γ is a hyperbolic group with torsion, Γ
and all Γ-limit groups have the weaker property of being K-CSA (for some integer K), as
defined in Subsection 9.2. Theorem 6 generalizes to this setting (see Theorem 9.14) and
implies the following result:
Theorem 7. Let Γ be a hyperbolic group, and let G be a one-ended Γ-limit group. Then
G has a canonical JSJ tree over virtually abelian subgroups relative to all virtually abelian
subgroups which are not virtually cyclic. Its flexible vertex stabilizers are virtually abelian
or QH with finite fiber.
We have a similar statement for relatively hyperbolic groups.
Theorem 8 (Theorem 9.18). Let G be one-ended and hyperbolic relative to a finite family of
finitely generated small subgroups. If A is either the family of all virtually cyclic subgroups
of G, or the family of all small subgroups, there is a canonical JSJ tree over A relative to
all parabolic subgroups. Its flexible vertex stabilizers are small or QH with finite fiber.
The trees produced by these theorems are defined in a uniform, natural way, and are
canonical (as discussed above). In particular, they are invariant under automorphims.
When G is a one-ended hyperbolic group, the canonical JSJ tree (non-relative in this case)
coincides with the tree constructed by Bowditch [Bow98] using the topology of ∂G.
Compatibility JSJ
A refinement T̂ of a tree T is a tree obtained by blowing up vertices of T (beware that
in [FP06] a refinement of T is what we call a tree dominating T ; on the other hand, their
elementary unfoldings are refinements in our sense). There is a map p from T̂ to T , so
T̂ dominates T , but this map is very special: it maps any segment [x, y] of T̂ onto the
segment [p(x), p(y)] of T (in particular, it does not fold). We call such a map a collapse
map, as it is obtained by collapsing certain edges to points.
If a tree T1 is universally elliptic, then given any tree T2 there is a refinement T̂1 of T1
which dominates T2 : there is an equivariant map f : T̂1 → T2 (see Proposition 2.2). If T2
has finitely generated edge stabilizers, one may obtain T2 from T̂1 by a finite sequence of
folds and collapses [BF91]. In this sense, one can read T2 from T1 . In particular, one may
read any tree T2 from any JSJ tree T1 .
In general, the map f : T̂1 → T2 is not a collapse map (there are folds). We say that
T1 and T2 are compatible if there exists a refinement T̂1 with a collapse map f : T̂1 → T2 .
In other words, T1 , T2 are compatible when they have a common refinement. This implies
that edge stabilizers of each tree are elliptic in the other, but is much more restrictive.
For instance, if Σ is a compact surface with boundary, free splittings of π1 (Σ) dual
to properly embedded arcs are always elliptic with respect to each other (edge groups are
trivial), but they are compatible only if the arcs are disjoint (up to isotopy). On the other
hand, splittings of a hyperbolic surface group associated to two simple closed geodesics
γ, γ 0 are compatible if and only if γ and γ 0 are disjoint or equal; in this specific case,
compatibility is equivalent to the trees being elliptic with respect to each other.
In Part V we introduce another type of JSJ decomposition, which encodes compatibility
of splittings rather than ellipticity. The new feature is that, except in degenerate cases, it
will lead to a canonical tree Tco (not just a deformation space). As above, we fix a family
A and all trees are assumed to be A-trees.
We say that a tree T (or the corresponding graph of groups Γ) is universally compatible
if it is compatible with every tree. One may then obtain any tree from T by refining and
collapsing, and view any one-edge splitting of G as coming from an edge of Γ or from a
splitting of a vertex group of Γ.
9
Definition. The compatibility JSJ deformation space Dco is the maximal deformation
space (for domination) containing a universally compatible tree (such a maximal deformation space, if it exists, is unique).
In other words, Dco contains a universally compatible tree T , and T dominates all
universally compatible trees.
Theorem 9 (Theorem 10.3). Let G be finitely presented, and let A be any conjugacyinvariant class of subgroups of G, stable under taking subgroups. Then the compatibility
JSJ deformation space Dco of G over A exists.
Although existence of the usual JSJ deformation space is a fairly direct consequence of
accessibility, proving existence of the compatibility JSJ deformation space is more delicate.
Among other things, we use a limiting argument, and we need to know that a limit of
universally compatible trees is universally compatible. This is best expressed in terms of
R-trees (see the appendix).
As mentioned above, the deformation space Dco contains a canonical element Tco ,
except in degenerate cases. Recall that a tree T is irreducible if G acts on T with no fixed
point, no fixed end, and no invariant line; we say that a deformation space D is irreducible
if some (equivalently, every) T ∈ D is irreducible.
Theorem 10 (Corollary 10.8). If Dco exists and is irreducible, it contains a canonical tree
Tco , the compatibility JSJ tree. In particular, if A is invariant under automorphisms of
G, so is Tco .
This is because a deformation space D containing an irreducible universally compatible
tree has a preferred element.
We develop an analogy with arithmetic, viewing a refinement of T as a multiple of T ,
and one-edge splittings as primes. We define the greatest common divisor (gcd) of two
trees, the least common multiple (lcm) of a family of pairwise compatible trees, and Tco is
the lcm of the reduced universally compatible trees contained in Dco (see Subsection 1.4.2
for the definition of “reduced”).
This tree Tco is similar to the canonical tree TSS constructed by Scott and Swarup
[SS03], which has the property of being compatible with almost-invariant sets (Scott and
Swarup use the word enclosing, which generalizes compatibility). In general, Tco dominates
TSS , but it may be non-trivial when TSS is trivial (this happens for instance for the
Baumslag-Solitar group BS(m, n) when none of m, n divides the other). See [GL10] for
the relation between TSS and JSJ decompositions.
Being invariant under automorphisms sometimes forces Tco to be trivial (a point).
This happens for instance when G is free. On the other hand, we give simple examples
with Tco non-trivial: certain virtually free groups, generalized Baumslag-Solitar groups,
Poincaré duality groups... Trees of cylinders also provide many examples. In particular,
the canonical trees of Theorems 6 and 8 are very closely related to Tco .
Contents of the paper
For the reader’s convenience we now describe the detailed contents of each section. This
includes a few results not directly related to JSJ decompositions, which are of independent
interest: relative finite presentation (4.2.2), small orbifolds (5.1.3), orbifolds with finite
mapping class group (5.1.4), K-CSA groups (9.2), compatibility and length functions (A.3),
arithmetic of trees (A.5).
This is meant as a description, so statements may be imprecise or incomplete.
• In the preliminary section (Section 1), we collect basic facts about groups acting on
trees, and we define (A, H)-trees (trees with edge stabilizers in A relative to H), collapse
10
maps and refinements, compatibility, domination, deformation spaces. We discuss slenderness and smallness of subgroups, one-endedness, and we recall the main accessibility
results. We also define and discuss relative finite generation and presentation (relative
finite presentation of vertex groups is studied in Subsection 4.2.2).
• Section 2 starts with a very useful fact: if all edge stabilizers of T1 are elliptic in T2 ,
there is a refinement of T1 which dominates T2 . After defining universal ellipticity, we define
JSJ trees and the JSJ deformation space by a maximality property, as explained above.
We prove the existence of JSJ decompositions under a finite presentability assumption,
first in the non-relative case and then in general. This relies on a version of Dunwoody’s
accessibility due to Fujiwara-Papasoglu [FP06] which we state and prove. We also explain
why the JSJ decompositions constructed by Rips-Sela [RS97], Dunwoody-Sageev [DS99],
Fujiwara-Papasoglu [FP06] are JSJ decompositions in our sense.
• Section 3 is devoted to simple examples. We first consider Grushko decompositions
(over the trivial group) and Stallings-Dunwoody decompositions (over finite groups), explaining how to interpret them as JSJ decompositions. We also consider small groups, and
locally finite trees (such as those associated to cyclic splittings of generalized BaumslagSolitar groups). All these examples of JSJ decompositions only have rigid vertices. At the
end of the section we work out an example where the JSJ decomposition has QH flexible
vertices.
• Section 4 contains various useful technical results. Given a vertex of a graph of
groups or of a tree, we define the incident edge groups and we point out that any splitting
of the vertex group which is relative to the incident edge groups extends to a splitting of
G. Given a universally elliptic splitting of G, one may obtain a JSJ decomposition of G
from relative JSJ decompositions of vertex groups. In particular, one may usually restrict
to one-ended groups when studying JSJ decompositions.
• Section 5 is devoted to QH groups. We first study 2-dimensional hyperbolic orbifolds
Σ, in particular the relation between splittings of π1 (Σ) and simple closed geodesics on Σ.
We classify orbifolds Σ with no essential simple closed geodesic (their groups do not split
over a cyclic subgroup relative to the fundamental groups of boundary components, so they
do not appear in QH vertices of JSJ decompositions); when Σ is a surface, only the pair
of pants (thrice-punctured sphere) occurs, but the classification is more complicated for
singular, possibly non-orientable, orbifolds. We also classify orbifolds with finite mapping
class group.
We then define QH subgroups, and study their basic properties (universal ellipticity
of the fiber, used boundary components, existence of simple geodesics, universally elliptic
subgroups). In particular, we show that, if Q is a QH vertex group in a JSJ decomposition
over slender groups, and G acts on a tree with slender edge stabilizers, then either Q fixes
a point or its action on its minimal subtree is dual to a family of simple closed geodesics
of Σ. We also prove a more general version of Proposition 5 showing that, under suitable
assumptions, any QH vertex stabilizer Q of any tree is elliptic in JSJ trees.
Using a filling construction, we give examples of possible peripheral structures of QH
vertex groups, and we show that flexible vertex groups of JSJ decompositions over abelian
subgroups do not have to be QH (the filling construction was introduced in Section 4, in
order to provide an alternative construction of relative splittings).
• In Section 6 we study flexible vertices of JSJ decompositions over slender groups;
in particular, we prove Theorem 4. For completeness we also describe slender flexible
subgroups. We allow edge stabilizers which are only “slender in trees”: whenever G acts on
a tree, they fix a point or leave a line invariant. This is useful when working in a relative
setting, as groups in H are automatically slender in trees.
To prove Theorem 4, we follow the approach by Fujiwara-Papasoglu [FP06], with simplifications. In particular, the a priori knowledge that JSJ decompositions exist allows us
to reduce to totally flexible groups. Following [FP06], we define the core, and the enclosing
11
group of two splittings (which we call their regular neighborhood). We then construct a
filling pair of splittings, and we show that their regular neighborhood is the required QH
group. For technical reasons we replace Fujiwara-Papasoglu’s notion of minimal splittings
by the slightly stronger notion of minuscule splittings.
In Theorem 4 the family A does not have to be the family of all slender subgroups of G.
The theorem holds if A is a family of slender groups satisfying one of two stability conditions
(SC) and (SCZ ); the familly of cyclic (resp. virtually cyclic) subgroups satisfies (SCZ )
(resp. (SC)). These conditions ensure that the regular neighborhood of two splittings with
edge groups in A also has edge groups in A.
• Section 7 is devoted to the tree of cylinders. Given an admissible equivalence relation
on the set A∞ of infinite groups in A, one may associate a tree of cylinders Tc to any
tree T with stabilizers in A∞ . This tree only depends on the deformation space of T . We
give conditions ensuring that Tc is acylindrical and smally dominated by T (this means in
particular that groups elliptic in Tc but not in T are small). We also study the compatibility
properties of Tc .
• In Section 8 we show that JSJ decompositions exist, and non-small flexible vertex
groups are QH with finite fiber, under the assumption that one may associate to any tree
T an acylindrical tree T ∗ smally dominated by T . We first construct a relative tree as in
Theorems 6 and 8, and we refine it in order to get the required JSJ tree.
• This is applied in Section 9, with T ∗ the tree of cylinders, and used to prove Theorems
6 and 8. We study torsion-free CSA groups, relatively groups, as well as cyclic splittings
of commutative transitive groups. We introduce K-CSA groups, for K an integer, which
are better suited than CSA groups to study groups with torsion, and we prove Theorem
7. We also discuss a slightly different type of JSJ decompositions of one-ended hyperbolic
groups, where edge groups are required to be maximal virtually cyclic subgroups with
infinite center.
• In Section 10 we define universal compatibility and we show Theorem 9 (existence of
the compatibility JSJ deformation space Dco ). We then construct the compatibility JSJ
tree Tco and we give examples. In particular, we use the compatibility properties of the
tree of cylinders to identify Dco for abelian splittings of CSA groups, elementary splittings
of relatively hyperbolic groups, and cyclic splittings of commutative transitive groups.
• In the appendix we view trees as metric rather than combinatorial objects, and we
actually consider R-trees. We first give simple proofs of two standard results: a tree is
determined by its length function ([AB87], [CM87]), the axes topology agrees with the
equivariant Gromov-Hausdorff topology [Pau89]. We then study compatibility (which is
defined for R-trees using collapse maps): we prove that two R-trees are compatible if and
only if the sum of their length functions is a length function (i.e. it comes from some Rtree). In particular, compatibility is a closed property in the space of trees (this is used in
the proof of Theorem 9).
Using the core introduced by the first author [Gui05], we then show that a finite family
of pairwise compatible R-trees has a common refinement. Going back to simplicial trees,
we develop an analogy with basic arithmetics: we define the prime factors of a tree T (the
one-edge splittings corresponding to edges of the quotient graph T /G), we show that trees
are square-free, and we define gcd’s and lcm’s.
We conclude by some remarks about combining JSJ theory and Rips theory to describe
small actions on R-trees. This gives another, more general, approach to the main result of
[Gui00b].
What is new
JSJ theory was developed by several people, but to the best of our knowledge the following
is original material:
12
• The definition of JSJ decompositions as a deformation space satisfying a maximality
property.
• The systematic study of relative JSJ decompositions.
• Most of Section 4.
• The classification of small orbifolds and orbifolds with finite mapping class group in
Section 5.
• The stability conditions (SC) and (SCZ ) (but compare [DS99]), and the description
of slender flexible groups.
• Everything from Subsection 7.3 on, with the exception of A.1 and A.2. In particular
the detailed proof of existence of the JSJ decomposition under acylindricity assumptions
(compare [Sel99, Sel01]), the compatibility JSJ tree, the arithmetic of trees.
• The K-CSA property (but see also [RW10]).
Acknowledgements
The first author acknowledges support from the Institut universitaire de France, ANR
project ANR-11-BS01-013, and membership to the Henri Lebesgue center 11-LABX-0020.
The second author acknowledges support from ANR-07-BLAN-0141-01 and ANR-2010BLAN-116-03.
Contents
I
Preliminaries
15
1 Preliminaries
1.1 Basic notions and notations . . . . . . . . . . . . . . .
1.2 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 (A, H)-trees, one-endedness . . . . . . . . . . . . . . .
1.4 Maps between trees, compatibility, deformation spaces
1.5 Slenderness, smallness . . . . . . . . . . . . . . . . . .
1.6 Accessibility . . . . . . . . . . . . . . . . . . . . . . . .
1.7 Relative finite generation and presentation . . . . . . .
II
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
The JSJ deformation space
15
15
16
18
18
19
20
21
22
2 Definition and existence
2.1 Standard refinements . . . . . . . . . . . . . . . . . . . . . . .
2.2 Universal ellipticity . . . . . . . . . . . . . . . . . . . . . . .
2.3 The JSJ deformation space . . . . . . . . . . . . . . . . . . .
2.4 Existence of the JSJ deformation space: the non-relative case
2.5 Existence: the relative case . . . . . . . . . . . . . . . . . . .
2.6 Relation with other constructions . . . . . . . . . . . . . . . .
3 Examples of JSJ decompositions
3.1 Free groups . . . . . . . . . . . . . . . . . . . . . . .
3.2 Free splittings: the Grushko deformation space . . .
3.3 Splittings over finite groups: the Stallings-Dunwoody
3.4 Splittings of small groups . . . . . . . . . . . . . . .
3.5 Generalized Baumslag-Solitar groups . . . . . . . . .
3.6 Locally finite trees . . . . . . . . . . . . . . . . . . .
3.7 RAAGs . . . . . . . . . . . . . . . . . . . . . . . . .
13
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . .
. . . . . . .
deformation
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . .
. . . .
space
. . . .
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22
22
24
25
26
27
29
.
.
.
.
.
.
.
30
30
30
31
31
32
32
32
3.8
3.9
Parabolic splittings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Non-rigid examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4 A few useful facts
4.1 Changing edge groups . . . . . . . .
4.2 Incidence structures of vertex groups
4.3 JSJ decompositions of vertex groups
4.4 Relative JSJ decompositions through
III
. . . .
. . . .
. . . .
fillings
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Flexible vertices
6 JSJ
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
decompositions over slender groups
Statement of results . . . . . . . . . . . . . . . . . .
Reduction to totally flexible groups . . . . . . . . . .
Fujiwara-Papasoglu’s core . . . . . . . . . . . . . . .
The regular neighborhood . . . . . . . . . . . . . . .
Constructing a filling pair of splittings . . . . . . . .
Flexible groups are QH when trees are minuscule . .
All splittings of a totally flexible group are minuscule
Slenderness in trees . . . . . . . . . . . . . . . . . . .
Slender flexible groups . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
41
47
50
52
54
.
.
.
.
.
.
.
.
.
55
55
57
59
62
65
68
68
70
71
Acylindricity
7 Trees of cylinders
7.1 Definition . . . .
7.2 Acylindricity . .
7.3 Small domination
7.4 Compatibility . .
34
34
35
37
39
40
5 Quadratically hanging vertices
5.1 2-orbifolds and their splittings . . . . . . . . . . . . . . . . . .
5.2 Definition and properties of quadratically hanging subgroups
5.3 Quadratically hanging subgroups are elliptic in the JSJ . . . .
5.4 Peripheral structure of quadratically hanging vertices . . . . .
5.5 Flexible vertices of abelian JSJ decompositions . . . . . . . .
IV
.
.
.
.
73
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
73
74
76
77
79
8 Constructing JSJ decompositions using acylindricity
80
8.1 Uniform acylindricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.2 Acylindricity up to small groups . . . . . . . . . . . . . . . . . . . . . . . . . 86
9 Applications
9.1 CSA groups . . . . . . . . . . . . .
9.2 Γ-limit groups and K-CSA groups
9.3 Relatively hyperbolic groups . . . .
9.4 Virtually cyclic splittings . . . . .
9.5 The Zmax -JSJ decomposition . . .
V
.
.
.
.
.
Compatibility
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
89
90
91
94
96
97
100
14
10 The compatibility JSJ tree
100
10.1 Existence of the compatibility JSJ space . . . . . . . . . . . . . . . . . . . . 101
10.2 The compatibility JSJ tree Tco . . . . . . . . . . . . . . . . . . . . . . . . . 103
11 Examples
11.1 Free groups . . . . . . . . . . . . . . . . . . . . . .
11.2 Algebraic rigidity . . . . . . . . . . . . . . . . . . .
11.3 Free products . . . . . . . . . . . . . . . . . . . . .
11.4 (Generalized) Baumslag-Solitar groups . . . . . . .
11.5 The canonical decomposition of Scott and Swarup
11.6 Poincaré duality groups . . . . . . . . . . . . . . .
11.7 Trees of cylinders . . . . . . . . . . . . . . . . . . .
A R-trees, length functions, and compatibility
A.1 Metric trees and length functions . . . . . .
A.2 From length functions to trees . . . . . . . .
A.3 Compatibility and length functions . . . . .
A.4 Common refinements . . . . . . . . . . . . .
A.5 Arithmetic of trees . . . . . . . . . . . . . .
A.6 Reading actions on R-trees . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
104
. 104
. 104
. 105
. 105
. 105
. 106
. 106
.
.
.
.
.
.
109
. 109
. 110
. 113
. 115
. 116
. 118
Part I
Preliminaries
1
Preliminaries
In this paper, G will always be a finitely generated group. Sometimes finite presentation
will be needed, for instance to prove existence of JSJ decompositions in full generality.
1.1
Basic notions and notations
Two subgroups H, H 0 of a group G are commensurable if H ∩ H 0 has finite index in H and
in H 0 . The commensurator of H (in G) is the set of elements g ∈ G such that gHg −1 is
commensurable to H.
We denote by Fn the free group on n generators.
A group H is virtually cyclic if it has a cyclic subgroup of finite index: H is finite or has
a finite index subgroup isomorphic to Z (infinite virtually cyclic groups are characterized
as being two-ended). There are two types of infinite virtually cyclic groups: those with
infinite center map onto Z with finite kernel K, those with finite center map onto the
infinite dihedral group D∞ = (Z/2) ∗ (Z/2) with finite kernel K (see [SW79]). The kernel
K is the unique maximal finite normal subgroup of H. In Subsection 7.2 we shall say that
H is C-virtually cyclic for some constant C ≥ 1 if K has cardinality at most C.
A group G is relatively hyperbolic with respect to a family of finitely generated subgroups P = {P1 , . . . , Pn } if it acts properly by isometries on a proper Gromov hyperbolic
space with an invariant collection of disjoint horoballs, the action is cocompact on the complement of these horoballs, and the stabilizers of the horoballs are exactly the conjugates
of the Pi ’s [Bow12, Far98, Osi06, Hru10]. A subgroup is called parabolic if it is conjugate
to a subgroup of some Pi , elementary if it is parabolic or virtually cyclic. Non-elementary
subgroups contain free subgroups acting by hyperbolic isometries.
15
A group is CSA (for conjugately separated abelian) if its maximal abelian subgroups
are malnormal. For example, torsion-free hyperbolic groups are CSA. See Subsection 9.2
for a generalisation in presence of torsion.
1.2
Trees
We consider actions of G on simplicial trees T . We identify two trees if there is an equivariant isomorphism between them. In the appendix we will view trees as metric spaces
and work with R-trees, but in most of the paper trees are considered as combinatorial
objects. Still, it is useful to think of a tree as a geometric object, for instance to define the
arc between any two points, or the midpoint of an edge. See [Ser77, SW79, DD89, Sha91,
Sha87, Chi01] for basic facts about trees.
Given two points a, b ∈ T , there is a unique segment [a, b] joining them; it is degenerate
if a = b. Non-degenerate segments are also called arcs (they are homeomorphic to [0, 1]).
If A, B are disjoint simplicial subtrees, the bridge between them is the unique segment
I = [a, b] such that A ∩ I = {a} and B ∩ I = {b}.
We always assume that G acts without inversion (if e = vw is an edge, no element of G
interchanges v and w), and often that T has no redundant vertex : if a vertex has valence
2, it is the unique fixed point of some element of G. We denote by V (T ) and E(T ) the
set of vertices and (non-oriented) edges of T respectively, by Gv or Ge the stabilizer of a
vertex v or an edge e, by Gx the stabilizer of an arbitrary point x ∈ T .
By Bass-Serre theory, the action of G on T can be viewed as a splitting of G as a
marked graph of groups, i.e. an isomorphism between G and the fundamental group of a
graph of groups Γ. A one-edge splitting (when Γ has one edge) is an amalgam G = A ∗C B
or an HNN-extension G = A∗C . We also denote by Gv or Ge the groups carried by vertices
and edges of Γ.
The action G y T is trivial if G fixes a point (hence a vertex since we assume that
there is no inversion), minimal if there is no proper G-invariant subtree.
Unless otherwise indicated, all trees are endowed with a minimal action of G without
inversion (we allow the trivial case when T is a point). Finite generation of G implies that
the quotient graph Γ = T /G is finite. Note, however, that the restriction of the action to
a subgroup of G does not have to be minimal.
An element g or a subgroup H of G is elliptic in T if it fixes a point. This is equivalent
to g or H being contained in a conjugate of a vertex group of Γ. We denote by Fix g or
Fix H its fixed point set in T . If H is elliptic, any H-invariant subtree meets Fix H. If
H1 ⊂ H2 ⊂ G with H1 of finite index in H2 , then H1 is elliptic if and only if H2 is.
Finite groups, and groups with Kazhdan’s property (T), have Serre’s property (FA):
there is a fixed point in every tree on which they act.
If H is finitely generated, it is elliptic if and only if all its elements are. This follows
from Serre’s lemma [Ser77, 6.5, corollaire 2]: if g1 , . . . , gn , as well as all products gi gj , are
elliptic, then hg1 , . . . , gn i is an elliptic subgroup.
An element g which is not elliptic is hyperbolic, it has a unique axis A(g) on which it
acts by translation. We also denote by A(g) the characteristic set of g: its fixed point set
if it is elliptic, its axis if it is hyperbolic.
Translation lengths and length functions will only be used in Subsections 8.1 and 10.1,
we discuss them in the appendix (for R-trees).
We will need to consider the restriction of the action of G to subgroups. Let therefore
H be an arbitrary group (possibly infinitely generated) acting on a tree T . We only assume
that the action is not trivial (there is no global fixed point).
If H contains no hyperbolic element, it is not finitely generated by Serre’s lemma, and
it fixes a unique end of T : there is a ray ρ such that each finitely generated subgroup of H
fixes a subray of ρ (an end of T is an equivalence class of rays, with ρ, ρ0 equivalent if their
16
intersection is a ray). In this case there are H-invariant subtrees, but no minimal one.
Assume now that H contains a hyperbolic element (we sometimes say that H acts hyperbolically, or is hyperbolic); by Serre’s lemma, this always holds if H is finitely generated
and acts non-trivially. Then there is a unique minimal H-invariant subtree µT (H), namely
the union of axes of hyperbolic elements of H.
The tree T (with the action of H) is irreducible if there exist two hyperbolic elements
g, h ∈ H whose axes are disjoint or intersect in a finite segment. Suitable powers of g and
h then generate a non-abelian free group F2 ⊂ H acting freely on T . It follows that T is
irreducible if and only if there exist two hyperbolic elements g, h whose commutator [g, h]
is hyperbolic.
If T is not irreducible, then H preserves a line of T or fixes a unique end (recall that by
assumption there is no global fixed point, and note that there is an invariant line if there
are two fixed ends).
There are two types of non-trivial actions on a line. If orientation is preserved, the
action is by translations and both ends of the line are invariant. All points and edges
have the same stabilizer. If not, there are reflections. The action is said to be dihedral,
it factors through an action of the infinite dihedral group D∞ = Z/2 ∗ Z/2. There is no
invariant end. All edges have the same stabilizer, but vertex stabilizers may contain the
edge stabilizer with index 2.
If H fixes an end of T , there is an associated homomorphism χ : H → Z measuring
how much an element h pushes towards the end. More precisely, for h ∈ H, one defines
χ(h) as the difference between the number of edges in ρ \ (ρ ∩ hρ) and the number of edges
in hρ \ (ρ ∩ hρ), with ρ any ray going to the end.
The map χ is non-trivial if and only if H contains a hyperbolic element. In this case
the quotient graph of groups Γ is homeomorphic to a circle, and one may orient the circle
so that the inclusion Ge → Gv is onto whenever e = vw is a positively oriented edge.
When Γ is a single edge, it defines an ascending HNN extension and χ is (up to sign) the
exponent sum of the stable letter.
To sum up:
Proposition 1.1. If H acts on a tree, one of the following holds:
(1) there is a global fixed point;
(2) there are hyperbolic elements h ∈ H, and T contains a unique minimal H-invariant
subtree µT (H);
(3) H is infinitely generated and fixes a unique end.
Proposition 1.2. If H acts minimally on a tree T (i.e. there is no proper H-invariant
subtree), there are five possibilities:
(1) T is a point (trivial action);
(2) T is a line, and H acts by translations; the action factors through Z;
(3) T is a line, and some h ∈ H reverses orientation (dihedral action); the action factors
through D∞ = Z/2 ∗ Z/2;
(4) there is a unique invariant end, but no global fixed point; the quotient graph of groups
is homeomorphic to a circle;
(5) T is irreducible (this implies F2 ⊂ H).
In all cases except (1), some h ∈ H is hyperbolic.
In particular:
17
Corollary 1.3. If H acts on T , and T is not irreducible, then there is a fixed point, or a
unique fixed end, or a unique invariant line.
Proof. Existence follows from the propositions. We prove the uniqueness statements. If
there are two invariant ends, the line joining them is invariant. We show that there is a
fixed point if there exist two invariant lines `1 , `2 .
If `1 and `2 are disjoint, the midpoint of the bridge between them is fixed. If their
intersection is a segment of finite length, its midpoint is fixed. If the intersection is a ray,
its origin is a fixed point.
1.3
(A, H)-trees, one-endedness
Besides G, we usually also fix a (nonempty) family A of subgroups of G which is stable
under conjugation and under taking subgroups. An A-tree is a tree T whose edge stabilizers
belong to A. We often say that T , or the corresponding splitting of G, is over A, or over
groups in A. We say cyclic tree (abelian tree, slender tree, . . . ) when A is the family of
cyclic (abelian, . . . ) subgroups.
We also fix an arbitrary set H of subgroups of G, and we restrict to A-trees T such that
each H ∈ H is elliptic in T (in terms of graphs of groups, H is contained in a conjugate
of a vertex group; if H is not finitely generated, this is stronger than requiring that every
h ∈ H be elliptic). We call such a tree an (A, H)-tree, or a tree over A relative to H. The
set of (A, H)-trees does not change if we replace a group of H by a conjugate, or if we
enlarge H by making it invariant under conjugation.
If G acts non-trivially on an (A, H)-tree, we say that G splits over A (or over a group
of A) relative to H.
The group G is freely indecomposable relative to H if it does not split over the trivial
group relative to H. Equivalently (unless G = Z and H is trivial), one cannot write
G = G1 ∗ G2 with G1 , G2 non-trivial, and every group in H contained in a conjugate of G1
or G2 .
One says that G is one-ended relative to H if G does not split over a finite group relative
to H (when H is empty, this is equivalent to G being finite or one-ended by a theorem of
Stallings, see e.g. [SW79]).
1.4
1.4.1
Maps between trees, compatibility, deformation spaces
Morphisms, collapse maps, refinements, compatibility
Maps between trees will always be G-equivariant, send vertices to vertices and edges to
edge paths (maybe a point). By minimality of the actions, they are always surjective; each
edge of T 0 is contained in the image of an edge of T . Any edge stabilizer Ge0 of T 0 contains
an edge stabilizer Ge of T . Also note that any edge or vertex stabilizer of T is contained
in a vertex stabilizer of T 0 .
We mention two particular classes of maps.
A map f : T → T 0 between two trees is a morphism if and only if one may subdivide
T so that f maps each edge onto an edge; equivalently, no edge of T is collapsed to a
point. Folds are examples of morphisms (see [Sta83, BF91]). If f is a morphism, any edge
stabilizer of T is contained in an edge stabilizer of T 0 .
A collapse map f : T → T 0 is a map obtained by collapsing certain edges to points,
followed by an isomorphism (by equivariance, the set of collapsed edges is G-invariant).
Equivalently, f preserves alignment: the image of any arc [a, b] is a point or the arc
[f (a), f (b)]. Another characterization is that the preimage of every subtree is a subtree.
In terms of graphs of groups, one obtains T 0 /G by collapsing edges in T /G. If T 0 is
irreducible, so is T . If T is irreducible, one easily checks that T 0 is trivial (a point) or
irreducible (compare Lemma A.21).
18
A tree T 0 is a collapse of T if there is a collapse map T → T 0 ; conversely, we say that
T refines T 0 . In terms of graphs of groups, one passes from Γ = T /G to Γ0 = T 0 /G by
collapsing edges; for each vertex v of Γ0 , the vertex group Gv is the fundamental group of
the graph of groups Γv occurring as the preimage of v in Γ.
Conversely, suppose that v is a vertex of a splitting Γ0 , and Γv is a splitting of Gv in
which incident edge groups are elliptic. One may then refine Γ0 at v using Γv , so as to
obtain a splitting Γ whose edges are those of Γ0 together with those of Γv (see Lemma
4.12). Note that Γ is not uniquely defined because there is flexibility in the way edges of
Γ0 are attached to vertices of Γv ; this is discussed in [GL16, Section 4.2].
Two trees T1 , T2 are compatible if they have a common refinement: there exists a tree
T̂ with collapse maps gi : T̂ → Ti . There is such a T̂ with the additional property that no
edge of T̂ gets collapsed in both T1 and T2 (this is discussed in Subsection A.5).
1.4.2
Domination, deformation spaces [For02, GL07a]
A tree T1 dominates a tree T2 if there is an equivariant map f : T1 → T2 from T1 to T2 .
We call f a domination map. Equivalently, T1 dominates T2 if every vertex stabilizer of T1
fixes a point in T2 : every subgroup which is elliptic in T1 is also elliptic in T2 . In particular,
every refinement of T1 dominates T1 . Beware that domination is defined by considering
ellipticity of subgroups, not just of elements (this may make a difference if vertex stabilizers
are not finitely generated).
Deformation spaces are defined by saying that two trees belong to the same deformation
space D if they have the same elliptic subgroups (i.e. each one dominates the other). When
we restrict to A-trees, we say that D is a deformation space over A. If a tree in D is
irreducible, so are all others, and we say that D is irreducible.
For instance, all trees with a free action of Fn belong to the same deformation space
D = CVn , Culler-Vogtmann’s outer space [CV86]. Note, however, that only finitely many
trees are compatible with a given T ∈ CVn .
We have defined deformation spaces as combinatorial objects, but (just like outer space)
they may be viewed as geometric objects (see e.g. [GL07a]). We will not use this point of
view.
A deformation space D dominates a space D0 if trees in D dominate those of D0 . Every
deformation space dominates the deformation space of the trivial tree, which is called the
trivial deformation space. It is the only deformation space in which G is elliptic.
A tree T is reduced [For02] if no proper collapse of T lies in the same deformation
space as T (this is different from being reduced in the sense of [BF91]). Observing that the
inclusion from Gv into Gu ∗Ge Gv is onto if and only if the inclusion Ge → Gu is, one sees
that T is reduced if and only if, whenever e = uv is an edge with Ge = Gu , then u and v
belong to the same orbit (i.e. e projects to a loop in Γ = T /G). Another characterization
is that, for any edge uv such that hGu , Gv i is elliptic, there exists a hyperbolic element
g ∈ G sending u to v (in particular the edge maps to a loop in Γ).
If T is not reduced, one obtains a reduced tree T 0 in the same deformation space by
collapsing certain orbits of edges (T 0 is not uniquely defined in general).
1.5
1.5.1
Slenderness, smallness
Slenderness
A group H is slender if H and all its subgroups are finitely generated. Examples of slender
groups include finitely generated virtually abelian groups, finitely generated virtually nilpotent groups, and virtually polycyclic groups. A slender group cannot contain a non-abelian
free group.
19
Slender groups have the characteristic property that, whenever they act on a tree, they
fix a point or there is an invariant line.
Lemma 1.4 ([DS99], Lemma 1.1). Let H be a slender group acting on a tree T . If H does
not fix a point, there is a unique H-invariant line ` ⊂ T .
Since H is finitely generated, there is a minimal subtree µT (H), and with the terminology of Proposition 1.2, only cases (1), (2), (3) are possible for µT (H).
Proof. The action of H cannot be irreducible since H does not contain F2 . If there is no
fixed point and no invariant line, there is a fixed end, and the associated χ : H → Z is
non-trivial because H is finitely generated. Each element of ker χ fixes a ray going to the
fixed end. Being finitely generated, ker χ is elliptic by Serre’s lemma. Its fixed point set is
a subtree T0 , which is H-invariant because ker χ is normal in H. The action of H on T0
factors through an action of the cyclic group H/ ker χ, so T0 contains an H-invariant line.
Uniqueness follows from Corollary 1.3.
It is convenient to use this lemma to define a weaker notion for subgroups of G. We
say that a subgroup H ⊂ G (possibly infinitely generated) is slender in (A, H)-trees if,
whenever G acts on an (A, H)-tree T , there is a point fixed by H or an H-invariant line.
In particular, any slender group, any group contained in a group of H, any group with
property (FA), is slender in (A, H)-trees.
The following lemma will be used in Subsection 6.8.
Lemma 1.5. Let A C A0 be subgroups of G, with A0 /A slender and A slender in (A, H)trees. Then A0 is slender in (A, H)-trees.
Proof. Let T be an (A, H)-tree. If A is elliptic, its fixed point set is an A0 -invariant subtree
because A is normal in A0 . The action of A0 /A on this subtree fixes a point or leaves a line
invariant, so the same is true for A0 . If A is not elliptic, it preserves a unique line, which
is A0 -invariant since A C A0 .
1.5.2
Smallness
One defines an abstract group as being small if it does not contain F2 . Such a group cannot
act irreducibly on a tree. As above, we use trees to give a weaker definition for subgroups
of G. In particular, we want groups of H to be small.
Given a tree T on which G acts, we say (following [BF91] and [GL07a]) that a subgroup
H < G is small in T if its action on T is not irreducible. As mentioned above, H then
fixes a point, or an end, or leaves a line invariant (see Corollary 1.3). We say that H is
small in (A, H)-trees if it is small in every (A, H)-tree on which G acts. Every subgroup
not containing F2 , and every group contained in a group of H, is small in (A, H)-trees.
Moreover, H is small in (A, H)-trees if and only if all its finitely generated subgroups are.
1.6
Accessibility
Constructions of JSJ decompositions are based on accessibility theorems stating that, given
suitable G and A, there is an a priori bound for the number of orbits of edges of A-trees,
under the assumption that there is no redundant vertex (if v has valence 2, it is the unique
fixed point of some g ∈ G). This holds in particular:
(1) if G is finitely generated and all groups in A are finite with bounded order [Lin83];
(2) if G is finitely presented and all groups in A are finite [Dun85];
(3) if G is finitely presented, all groups in A are small, and the trees are reduced in the
sense of [BF91];
20
(4) if G is finitely generated and the trees are k-acylindrical for some k [Sel97a];
(5) if G is finitely generated and the trees are (k, C)-acylindrical [Wei12] ([Del99] for
finitely presented groups).
A tree is k-acylindrical (resp. (k, C)-acylindrical ) if the pointwise stabilizer of any
segment of length > k is trivial (resp. has order ≤ C).
In this paper, we use a version of Dunwoody’s accessibility given in [FP06] (see Proposition 2.17). In Section 8 we use acylindrical accessibility.
1.7
Relative finite generation and presentation
As mentioned above, we always assume that G is finitely generated, or finitely presented.
However, these properties are not always inherited by vertex groups. We therefore consider relative finite generation (or presentation), which behave better in that respect (see
Subsection 4.2.2).
Let G be a group with a finite family of subgroups H = {H1 , . . . , Hp }.
Definition 1.6. One says that G is finitely generated relative to H if there exists a finite
set Ω ⊂ G such that G is generated by Ω ∪ H1 · · · ∪ Hp . Such a subset Ω is a relative
generating set.
Clearly, if G is finitely generated, then it is finitely generated relative to any H. If the
Hi ’s are finitely generated, relative finite generation is equivalent to finite generation.
By adding the conjugators to Ω, one sees that relative finite generation does not change
if one replaces the subgroups Hi by conjugate subgroups.
As in Subsection 1.2, we have:
Proposition 1.7. Suppose that G is finitely generated relative to H, and acts on a tree T
relative to H. If there is no global fixed point, then G contains hyperbolic elements, there is
a unique minimal invariant subtree µT (G), and the quotient µT (G)/G is a finite graph.
Recall that T is relative to H if every Hi is elliptic in T .
We now consider relative finite presentation (see [Osi06]). Note that, if Ω is a relative
finite generating set, then the natural morphism F(Ω)∗H1 ∗· · ·∗Hp → G is an epimorphism
(with F(Ω) the free group on Ω).
Definition 1.8. One says that G is finitely presented relative to H if there exists a finite
relative generating set Ω ⊂ G, such that the kernel of the epimorphism F(Ω)∗H1 ∗· · ·∗Hp
G is normally generated by a finite subset R ⊂ F(Ω) ∗ H1 ∗ · · · ∗ Hp .
In particular, any group which is hyperbolic with respect to a finite family H is finitely
presented relative to H [Osi06].
One easily checks that relative finite presentation does not depend on the choice of Ω,
and is not affected if one replaces the subgroups Hi by conjugate subgroups. If G is finitely
presented, then it is finitely presented relative to any finite collection of finitely generated
subgroups. Note, however, that the free group F2 is not finitely presented relative to an
infinitely generated free subgroup H1 (only finitely many generators of H1 may appear in
R). Conversely, if G is finitely presented relative to any finite collection of finitely presented
subgroups, then G is finitely presented.
The following lemma will be used in Subsection 4.2.2.
Lemma 1.9. Suppose that G is finitely presented relative to {H1 , . . . , Hp }, and Hp ⊂ Hi
for some i < p. Then G is finitely presented relative to {H1 , . . . , Hp−1 }.
21
Proof. We show that Hp is finitely generated. The lemma then follows by applying Tietze
transformations.
Let x1 , . . . , xq be the set of elements of Hp which appear as a letter in one of the relators
of R (expressed as elements of F(Ω)∗H1 ∗· · ·∗Hp ). Each xj is equal to an element yj ∈ Hi .
We define a new finite set of relators by replacing each xj by yj and adding the relations
xj = yj . This new presentation expresses G as the amalgam of hΩ, H1 , . . . , Hp−1 i with Hp
over a finitely generated group H. The inclusion Hp ⊂ Hi now implies Hp = H.
Suppose that a finitely generated group G splits as a finite graph of groups. It is
well-known (see for instance [Coh89], Lemma 8.32 p. 218) that vertex groups are finitely
generated if one assumes that edge groups are finitely generated, but this is false in general
without this assumption. However, vertex groups are always finitely generated relative to
the incident edge groups, and there is a similar statement for relative finite presentation
(see Subsection 4.2.2).
Part II
The JSJ deformation space
We start this part by introducing standard refinements: if edge stabilizers of T1 are elliptic
in T2 , there is a tree T̂1 which refines T1 and dominates T2 . We then define the JSJ
deformation space, and we show that it exists under some finite presentability assumption.
We give examples in cases when there is no flexible vertex (flexible vertices are the subject
of Part III). We conclude this part by collecting a few useful facts. In particular, given
a tree T , we discuss finite presentation of vertex stabilizers, and we relate their splittings
relative to incident edge groups to splittings of G. We also explain why one may usually
restrict to one-ended groups when studying JSJ decompositions.
We fix a finitely generated group G, a family A of subgroups of G (closed under
conjugating and taking subgroups), and another family H. All trees will be minimal
(A, H)-trees (see Subsection 1.3). Whenever we construct a new tree (for instance in
Propositions 2.2 and 2.17) we check that it is a minimal (A, H)-tree.
2
Definition and existence
2.1
Standard refinements
Let T1 , T2 be trees.
Definition 2.1 (Ellipticity of trees). T1 is elliptic with respect to T2 if every edge stabilizer
of T1 fixes a point in T2 .
Note that T1 is elliptic with respect to T2 whenever there is a refinement T̂1 of T1 that
dominates T2 (see Subsection 1.4 for the difference between refinement and domination):
edge stabilizers of T1 are elliptic in T̂1 , hence in T2 . We show a converse statement.
Proposition 2.2. If T1 is elliptic with respect to T2 , there is a tree T̂1 with maps p : T̂1 →
T1 and f : T̂1 → T2 such that:
(1) p is a collapse map;
(2) for each v ∈ T1 , the restriction of f to the subtree Yv = p−1 (v) is injective.
In particular:
(i) T̂1 is a refinement of T1 that dominates T2 ;
22
(ii) the stabilizer of any edge of T̂1 fixes an edge in T1 or in T2 ;
(iii) every edge stabilizer of T2 contains an edge stabilizer of T̂1 ;
(iv) a subgroup of G is elliptic in T̂1 if and only if it is elliptic in both T1 and T2 .
Assertions (ii) and (iv) guarantee that T̂1 is an (A, H)-tree since T1 and T2 are.
Remark 2.3. If edge stabilizers of T2 are finitely generated, then T2 can be obtained from
T̂1 by a finite number of collapses and folds [BF91].
Proof. We construct T̂1 as follows.
For each vertex v ∈ V (T1 ), with stabilizer Gv , choose any Gv -invariant subtree Ỹv of
T2 (for instance, Ỹv can be a minimal Gv -invariant subtree, or the whole of T2 ). For each
edge e = vw ∈ E(T1 ), choose vertices pv ∈ Ỹv and pw ∈ Ỹw fixed by Ge ; this is possible
because Ge is elliptic in T2 by assumption, so has a fixed point in any Gv -invariant subtree.
We make these choices G-equivariantly.
We can now define a tree T̃1 by blowing up each vertex v of T1 into Ỹv ,Fand attaching
edges of T1 using the points pv . Formally, we consider the disjoint union ( v∈V (T1 ) Ỹv ) ∪
F
( e∈E(T1 ) e), and for each edge e = vw of T1 we identify v with pv ∈ Ỹv and w with
pw ∈ Ỹw . We define p̃ : T̃1 → T1 by sending Yv to v, and sending e ∈ E(T1 ) to itself. We
also define a map f˜ : T̃1 → T2 as equal to the inclusion Ỹv ,→ T2 on Ỹv , and sending the
edge e = vw to the segment [pv , pw ] ⊂ T2 .
In general, T̃1 may fail to be minimal, so we define T̂1 ⊂ T̃1 as the unique minimal
G-invariant subtree µT̃1 (G) (the action of G on T̂1 is non-trivial unless T1 and T2 are both
points). We then define p and f as the restrictions of p̃ and f˜ to T̂1 . These maps clearly
satisfy the first two requirements.
Let us check that the other properties follow. Assertion (i) is clear.
If e is an edge of T̂1 that is not collapsed by p, then Ge fixes an edge of T1 . Otherwise,
f maps e injectively to a non-degenerate segment of T2 , so Ge fixes an edge in T2 , and
Assertion (ii) holds.
Assertion (iii) is true for any surjective map f between trees.
To prove the non-trivial direction of Assertion (iv), assume that H is elliptic in T1 and
T2 . Then H ⊂ Gv for some v ∈ T1 , so H preserves the subtree Yv ⊂ T̂1 . Since f is injective
in restriction to Yv , it is enough to prove that H fixes a point in f (Yv ). This holds because
H is elliptic in T2 .
Remark 2.4. One may think of this construction in terms of graphs of groups, as follows.
Starting from the graph of group Γ1 = T1 /G, one replaces each vertex v ∈ Γ1 by the graph
of groups Λv dual to the action of Gv on its minimal subtree in T2 , and one attaches each
edge e of Γ1 incident to v onto a vertex of Λv whose group contains a conjugate of Ge .
Since any vertex group Gv of Γ1 is finitely generated relative to its incident edge groups
(see Lemma 4.7), and these edge groups are elliptic in T2 , there is a minimal Gv -invariant
subtree in T2 by Lemma 1.7. Thus, one may require that Gv act minimally on the preimage
Yv ⊂ T̂1 .
Definition 2.5 (Standard refinement). Any tree T̂1 as in Proposition 2.2 will be called a
standard refinement of T1 dominating T2 .
In general, there is no uniqueness of standard refinements. However, by Assertion (iv)
of Proposition 2.2, all standard refinements belong to the same deformation space, which
is the lowest deformation space dominating the deformation spaces containing T1 and T2
respectively. If T1 dominates T2 (resp. T2 dominates T1 ), then T̂1 is in the same deformation
space as T1 (resp. T2 ). Moreover, there is some symmetry: if T2 also happens to be elliptic
with respect to T1 , then any standard refinement T̂2 of T2 dominating T1 is in the same
deformation space as T̂1 .
23
Lemma 2.6.
(1) If T1 refines T2 and does not belong to the same deformation space, some g ∈ G is
hyperbolic in T1 and elliptic in T2 .
(2) If T1 is elliptic with respect to T2 , and every g ∈ G which is elliptic in T1 is also
elliptic in T2 , then T1 dominates T2 .
(3) If T1 is elliptic with respect to T2 , but T2 is not elliptic with respect to T1 , then G
splits over a group which has infinite index in an edge stabilizer of T2 .
Recall that all trees are assumed to be (A, H)-trees; the splitting obtained in (3) is
relative to H.
Proof. One needs only prove the first assertion when T2 is obtained from T1 by collapsing
the orbit of an edge e = uv. If u and v are in the same orbit, or if Ge 6= Gu and Ge 6= Gv ,
then some hyperbolic element of T1 becomes elliptic in T2 . Otherwise, T1 and T2 are in
the same deformation space (see Subsection 1.4.2).
For the second assertion, assume that T1 does not dominate T2 , and let T̂1 be a standard
refinement of T1 dominating T2 . It does not belong to the same deformation space as
T1 . Since it is a refinement of T1 , we have just seen that some g ∈ G is elliptic in T1
and hyperbolic in T̂1 . By assumption g is elliptic in T2 , contradicting Assertion (iv) of
Proposition 2.2.
For (3) (which is Remark 2.3 of [FP06]), let T̂1 be as in Proposition 2.2. Let Ge be
an edge stabilizer of T2 which is not elliptic in T1 . It contains an edge stabilizer J of T̂1 .
Since J is elliptic in T1 and Ge is not, the index of J in Ge is infinite.
2.2
Universal ellipticity
Definition 2.7 (Universally elliptic). A subgroup H ⊂ G is universally elliptic if it is
elliptic in every (A, H)-tree. A tree T is universally elliptic if its edge stabilizers are
universally elliptic, i.e. if T is elliptic with respect to every (A, H)-tree.
When we need to be specific, we say universally elliptic over A relative to H, or (A, H)universally elliptic. Otherwise we just say universally elliptic, recalling that all trees are
assumed to be (A, H)-trees.
Groups with Serre’s property (FA), in particular finite groups, are universally elliptic.
If H is universally elliptic and H 0 contains H with finite index, then H 0 is universally
elliptic.
Lemma 2.8. Consider two trees T1 , T2 .
(1) If T1 is universally elliptic, then some refinement of T1 dominates T2 .
(2) If T1 and T2 are universally elliptic, any standard refinement T̂1 of T1 dominating T2
is universally elliptic. In particular, there is a universally elliptic tree T̂1 dominating
both T1 and T2 .
(3) If T1 and T2 are universally elliptic and have the same elliptic elements, they belong
to the same deformation space.
Proof. The first two assertions follow directly from Assertions (i) and (ii) of Proposition
2.2. The last one follows from the second assertion of Lemma 2.6.
The following lemma will be used in Subsection 10.1.
24
Lemma 2.9. Let (Ti )i∈I be any family of trees. There exists a countable subset J ⊂ I
such that, if T is elliptic with respect to every Ti (i ∈ I), and T dominates every Tj for
j ∈ J, then T dominates Ti for all i ∈ I.
Proof. Since G is countable, we can find a countable J such that, if an element g ∈ G is
hyperbolic in some Ti , then it is hyperbolic in some Tj with j ∈ J. If T dominates every
Tj for j ∈ J, any g which is elliptic in T is elliptic in every Ti . By (2) of Lemma 2.6, the
tree T dominates every Ti .
For many purposes, it is enough to consider one-edge splittings, i.e. trees with only one
orbit of edges.
Lemma 2.10. Let S be a tree.
(1) S is universally elliptic if and only if it is elliptic with respect to every one-edge
splitting.
(2) S dominates every universally elliptic tree if and only if it dominates every universally
elliptic one-edge splitting.
Proof. For the non-trivial direction, one proves that S is elliptic with respect to T (resp.
dominates T ) by induction on the number of orbits of edges of T , using the following
lemma.
Lemma 2.11. Let T be a tree, and H a subgroup of G. Let E1 t E2 be a partition of E(T )
into two G-invariant sets. Let T1 , T2 be the trees obtained from T by collapsing E1 and E2
respectively.
(1) If a subgroup H is elliptic in T1 and T2 , then H is elliptic in T .
(2) If a tree T 0 dominates T1 and T2 , then it dominates T .
Proof. Let x1 ∈ T1 be a vertex fixed by H. Let Y ⊂ T be its preimage under the collapse
map T → T1 . It is a subtree. Now Y is H-invariant and embeds into T2 . Since H is elliptic
in T2 , it fixes a point in Y , so is elliptic in T . One shows (2) by applying (1) to the vertex
stabilizers of T 0 .
2.3
The JSJ deformation space
Having fixed A, we define Aell ⊂ A as the set of groups in A which are universally elliptic
(over A relative to H); Aell is stable under conjugating and taking subgroups. A tree is
universally elliptic if and only if it is an Aell -tree.
Definition 2.12 (JSJ deformation space). If there exists a deformation space DJSJ of
(Aell , H)-trees which is maximal for domination, it is unique by the second assertion of
Lemma 2.8. It is called the JSJ deformation space of G over A relative to H.
Trees in DJSJ are called JSJ trees (of G over A relative to H). They are precisely
those trees T which are universally elliptic, and which dominate every universally elliptic
tree. We also say that trees T ∈ DJSJ , and the associated graphs of groups Γ = T /G, are
JSJ decompositions.
We will show that the JSJ deformation space exists if G is finitely presented (Theorems
2.16 and 2.20), or in the presence of acylindricity (see Section 8). See Subsection 3.3 for
an example where there is no JSJ deformation space.
In general there are many JSJ trees, but they all belong to the same deformation space
and therefore have a lot in common (see Section 4 of [GL07a]). In particular [GL07a,
Corollary 4.4], they have the same vertex stabilizers, except possibly for vertex stabilizers
in Aell .
25
Remark 2.13. There are results saying that two trees belong to the same deformation space
D if and only if one can pass from one to the other by a finite sequence of moves of a certain
type ([For02], see also [CF09]); in particular, if D is non-ascending as defined in [GL07a,
Section 7], for instance when all groups in A are finite, any two reduced trees in D may be
joined by a finite sequence of slide moves. These results may be interpreted as saying that
a JSJ tree is unique up to certain moves. This is the content of the uniqueness statements
of [Sel97b, DS99, For03, FP06].
Definition 2.14 (Rigid and flexible vertices). Let H = Gv be a vertex stabilizer of a JSJ
tree T (or a vertex group of the graph of groups Γ = T /G). We say that H is rigid if it is
universally elliptic, flexible if it is not. We also say that the vertex v is rigid (flexible). If
H is flexible, we say that it is a flexible subgroup of G (over A relative to H).
The definition of flexible subgroups of G does not depend on the choice of the JSJ tree
T . The heart of JSJ theory is to understand flexible groups. They will be discussed in
Part III.
We record the following simple facts for future reference.
Lemma 2.15. Let T be a JSJ tree, and S any tree.
• There is a tree T̂ which refines T and dominates S.
• If S is universally elliptic, it may be refined to a JSJ tree.
Proof. Since T is elliptic with respect to S, one can construct a standard refinement T̂ of
T dominating S (Proposition 2.2). It satisfies the first assertion.
For the second assertion, since S is elliptic with respect to T , we can consider a standard
refinement Ŝ of S dominating T . It is universally elliptic by the second assertion of Lemma
2.8, and dominates T , so it is a JSJ tree.
There sometimes exists a universally compatible JSJ tree (see Sections 9 and 11). In
this case, one may require that T̂ also be a refinement of S.
2.4
Existence of the JSJ deformation space: the non-relative case
We prove the existence of JSJ decompositions, first assuming H = ∅.
Theorem 2.16. If G is finitely presented, then the JSJ deformation space DJSJ of G over
A exists. It contains a tree whose edge and vertex stabilizers are finitely generated.
There is no hypothesis, such as smallness or finite generation, on the elements of A.
Recall that, G being finitely generated, finite generation of edge stabilizers implies finite
generation of vertex stabilizers.
The existence of DJSJ will be deduced from the following version of Dunwoody’s accessibility, whose proof will be given in the next subsection.
Proposition 2.17 (Dunwoody’s accessibility). Let G be finitely presented. Assume that
T1 ← · · · ← Tk ← Tk+1 ← . . . is a sequence of refinements of trees. There exists a tree S
such that:
(1) for k large enough, there is a morphism S → Tk (in particular, S dominates Tk );
(2) each edge and vertex stabilizer of S is finitely generated.
Note that the maps Tk+1 → Tk are required to be collapse maps.
Recall (Subsection 1.4) that f : S → Tk is a morphism if S may be subdivided so that
f maps edges to edges (a collapse map is not a morphism). In particular, edge stabilizers
of S fix an edge in Tk , so S is an A-tree since every Tk is. It is universally elliptic if every
Tk is.
26
Remark 2.18. Unfortunately, it is not true that the deformation space of Tk must stabilize
as k increases, even if all edge stabilizers are cyclic. For example (as on pp. 449-450 of
[BF91]), let A be a group with a sequence of nested infinite cyclic groups A % C1 % C2 %
. . . , let G = A ∗ B with B non-trivial, and let Tk be the Bass-Serre tree of the iterated
amalgam
G = A ∗C1 C1 ∗C2 C2 ∗C3 · · · ∗Ck−1 Ck−1 ∗Ck hCk , Bi.
These trees refine each other, but are not in the same deformation space since hCk , Bi is
not elliptic in Tk+1 . They are dominated by the tree S dual to the free decomposition
G = A ∗ B, in accordance with the proposition.
Applying Proposition 2.17 to a constant sequence yields the following standard result:
Corollary 2.19. If G is finitely presented, and T is a tree, there exists a morphism f :
S → T where S is a tree with finitely generated edge and vertex stabilizers.
If T is a universally elliptic A-tree, so is S.
Proposition 2.17 is basically Proposition 5.12 of [FP06]. We omit its proof, and refer
to the more general Proposition 2.24.
Proof of Theorem 2.16. Let U be the set of universally elliptic trees with finitely generated
edge and vertex stabilizers, up to equivariant isomorphism. It is non-empty since it contains
the trivial tree. An element of U is described by a finite graph of groups with finitely
generated edge and vertex groups. Since G only has countably many finitely generated
subgroups, and there are countably many homomorphisms from a given finitely generated
group to another, the set U is countable.
By Corollary 2.19, every universally elliptic tree is dominated by one in U, so it suffices
to produce a universally elliptic tree dominating every U ∈ U. Choose an enumeration
U = {U1 , U2 , . . . , Uk , . . . }. We define inductively a universally elliptic tree Tk which refines
Tk−1 and dominates U1 , . . . , Uk (it may have infinitely generated edge or vertex stabilizers). We start with T1 = U1 . Given Tk−1 which dominates U1 , . . . , Uk−1 , we let Tk be a
standard refinement of Tk−1 which dominates Uk (it exists by Proposition 2.2 because Tk−1
is universally elliptic). Then Tk is universally elliptic by the second assertion of Lemma
2.8, and it dominates U1 , . . . , Uk−1 because Tk−1 does.
Apply Proposition 2.17 to the sequence Tk . The tree S is universally elliptic and it
dominates every Tk , hence every Uk . It follows that S is a JSJ tree over A.
2.5
Existence: the relative case
In this section, we prove the existence of a relative JSJ deformation space under a relative
finite presentation assumption.
Theorem 2.20. Assume that G is finitely presented relative to H = {H1 , . . . , Hp }. Then
the JSJ deformation space DJSJ of G over A relative to H exists. It contains a tree with
finitely generated edge stabilizers.
Recalling that a finitely presented group is finitely presented relative to any finite
collection of finitely generated subgroups, we get:
Corollary 2.21. Let G be finitely presented. Let H = {H1 , . . . , Hp } be a finite family of
finitely generated subgroups. Then the JSJ deformation space DJSJ of G over A relative
to H exists. It contains a tree with finitely generated edge (hence vertex) stabilizers.
Remark 2.22. We will give a different approach to existence in Subsection 4.4, by comparing
the relative JSJ decomposition of G to a non-relative JSJ decomposition of a larger group
Ĝ.
27
Remark 2.23. The corollary does not apply when groups in H are infinitely generated. See
Section 8 for existence results with H arbitrary.
The theorem is proved as in the non-relative case (Theorem 2.16), with U the set
of (A, H)-trees with finitely generated edge stabilizers. It is countable because vertex
stabilizers are relatively finitely generated by [Gui08, Lemmas 1.11, 1.12] (this is explained
in Remark 4.8). Proposition 2.17 is replaced by the following result.
Proposition 2.24 (Relative Dunwoody’s accessibility). Let G be finitely presented relative
to H = {H1 , . . . , Hp }. Assume that T1 ← · · · ← Tk ← Tk+1 ← . . . is a sequence of
refinements of (A, H)-trees. There exists an (A, H)-tree S such that:
(1) for k large enough, there is a morphism S → Tk ;
(2) each edge stabilizer of S is finitely generated.
As above, applying the proposition to a constant sequence, we get:
Corollary 2.25. If G is finitely presented relative to H, and T is an (A, H)-tree, there
exists a morphism f : S → T where S is an (A, H)-tree with finitely generated edge
stabilizers.
Before proving the proposition, we recall that G is finitely presented relative to its
subgroups H1 , . . . , Hp if there exists a finite subset Ω ⊂ G such that the natural morphism
F(Ω) ∗ H1 ∗ · · · ∗ Hp → G is onto, and its kernel is normally generated by a finite subset R.
Here is an equivalent definition: G is finitely presented relative to H if and only if
it is the fundamental group of a connected 2-complex X (which may be assumed to be
simplicial) containing disjoint connected subcomplexes Y1 , . . . , Yp (possibly infinite) with
the following properties: X \ (Y1 ∪ · · · ∪ Yp ) contains only finitely many open cells, π1 (Yi )
embeds into π1 (X), and its image is conjugate to Hi .
The fact that G is relatively finitely presented if such a space X exists follows from Van
Kampen’s theorem. Conversely, if G is finitely presented relative to H, one can construct
X as follows. Let (Yi , ui ) be a pointed 2-complex with π1 (Yi , ui ) ' Hi . Starting from
the disjoint union of the Yi ’s, add p edges joining the ui ’s to an additional vertex u, and
#Ω additional edges joining u to itself. We get a complex whose fundamental group is
isomorphic to the free product F(Ω) ∗ H1 ∗ · · · ∗ Hp . Represent each element of R by a loop
in this space, and glue a disc along this loop to obtain the desired space X.
Proof of Proposition 2.24. Let π : X̃ → X be the universal cover of a simplicial 2-complex
X as above, with G acting on X̃ by deck transformations. For i ∈ {1, . . . , p}, consider a
connected component Ỹi of π −1 (Yi ) whose stabilizer is Hi . Also fix lifts v1 , . . . , vq ∈ X̃ of
all the vertices in X \ Y1 ∪ · · · ∪ Yp .
We denote by pk : Tk+1 → Tk the collapse map. Note that the preimage of the midpoint
of an edge ek of Tk is a single point, namely the midpoint of the edge of Tk+1 mapping
onto ek .
We shall now construct equivariant maps fk : X̃ → Tk such that fk maps each Ỹi to a
vertex fixed by Hi , sends each vj to a vertex, and sends each edge of X̃ either to a point
or injectively onto a segment in Tk . We further require that pk (fk+1 (x)) = fk (x) if x is a
vertex of X̃ or if fk (x) is the midpoint of an edge of Tk .
We construct fk inductively. We start with T0 a point, and f0 , p0 the constant maps.
We then assume that fk : X̃ → Tk has been constructed, and we construct fk+1 .
To define fk+1 on Ỹi , note that fk (Ỹi ) is a vertex of Tk fixed by Hi . Since pk preserves
alignment, p−1
k (fk (Ỹi )) is an Hi -invariant subtree. Since Hi is elliptic in Tk+1 , it fixes some
vertex in this subtree, and we map Ỹi to such a vertex. We then define fk+1 (vj ) as any
vertex in p−1
k (fk (vj )), and we extend by equivariance.
28
Now consider an edge e of X not contained in any Yi , and a lift ẽ ⊂ X̃. The map fk+1
is already defined on the endpoints of ẽ, we explain how to define it on ẽ.
The restriction of pk to the segment of Tk+1 joining the images of the endpoints of ẽ
is a collapse map. In particular, the preimage of the midpoint of an edge ek of Tk is the
midpoint of the edge of Tk+1 mapping onto ek . Recalling that fk is constant or injective
on ẽ, this allows us to define fk+1 on ẽ, as a map which is either constant or injective,
and satisfies pk (fk+1 (x)) = fk (x) if fk (x) is the midpoint of an edge of Tk . Doing this
equivariantly, we have now defined fk+1 on the 1-skeleton of X̃.
We then extend fk+1 in a standard way to every triangle abc not contained in a π −1 (Yi );
in particular, if fk+1 is not constant on abc, preimages of midpoints of edges of Tk+1 are
straight arcs joining two distinct sides. This completes the construction of the maps fk .
We now define τ̃k ⊂ X̃ as the preimage (under fk ) of the midpoints of all edges of Tk .
This is a pattern in the sense of Dunwoody [Dun85]. It does not intersect any Ỹi , and the
maps fk were constructed so that τ̃k ⊂ τ̃k+1 . We denote by τk = π(τ̃k ) the projection in
X. It is a finite graph because it is contained in the complement of Y1 ∪ · · · ∪ Yp .
Let Sk be the tree dual to the pattern τ̃k in X̃. We claim that it is an (A, H)-tree
with finitely generated edge stabilizers. By construction, fk induces a map ϕk : Sk → Tk
sending edge to edge, so edge stabilizers of Sk are in A. They are finitely generated because
they are generated by fundamental groups of components of τk . Every Hi is elliptic in Sk
because Ỹi does not intersect τ̃k . This proves the claim.
Let X 0 ⊂ X be the closure of the complement of Y1 ∪ · · · ∪ Yp . By construction, this is a
finite complex. By [Dun85, Theorem 2.2], there is a bound on the number of non-parallel
tracks in X 0 . This implies that there exists k0 such that, for all k ≥ k0 , for every connected
component σ of τk \ τk0 , there exists a connected component σ 0 of τk0 such that σ ∪ σ 0
bounds a product region containing no vertex of X 0 . It follows that, for k ≥ k0 , one can
obtain Sk from Sk0 by subdividing edges. We then take S = Sk0 .
2.6
Relation with other constructions
Several authors have constructed JSJ splittings of finitely presented groups in various
settings. We explain here (in the non-relative case) why those splittings are JSJ splittings
in the sense of Definition 2.12 (results in the literature are often stated only for one-edge
splittings, but this is not a restriction by Lemma 2.10).
In [RS97], Rips and Sela consider cyclic splittings of a one-ended group G (so A consists
of all cyclic subgroups of G, including the trivial group). Theorem 7.1 in [RS97] says that
their JSJ splitting is universally elliptic (this is statement (iv)) and maximal (statement
(iii)). The uniqueness up to deformation is statement (v).
In the work of Dunwoody-Sageev [DS99], the authors consider splittings of a group G
over slender subgroups in a class ZK such that G does not split over finite extensions of
infinite index subgroups of ZK (there are restrictions on the class ZK, but one can typically
take ZK = V P Cn , see [DS99] for details). In our notation, A is the set of subgroups of
elements of ZK. Universal ellipticity of the splitting they construct follows from statement
(3) in the Main Theorem of [DS99], and from the fact that any edge group is contained
in a white vertex group. Maximality follows from the fact that white vertex groups are
universally elliptic (statement (3)) and that black vertex groups either are in ZK (in which
case they are universally elliptic by the non-splitting assumption made on G), or are Kby-orbifold groups and hence are necessarily elliptic in any JSJ tree (see Proposition 5.27
below).
In [FP06], Fujiwara and Papasoglu consider all splittings of a group over the class A of
its slender subgroups. Statement (2) in [FP06, Theorem 5.13] says that the JSJ splitting
they obtain is elliptic with respect to any splitting which is minimal (in their sense). By
Proposition 3.7 in [FP06], any splitting is dominated by a minimal splitting, so universal
29
ellipticity holds. Statement (1) of Theorem 5.15 in [FP06] implies maximality.
As mentioned in the introduction, the regular neighbourhood of Scott-Swarup [SS03]
is closer to the decompositions constructed in Parts IV and V.
3
Examples of JSJ decompositions
Recall that we have fixed A and H, and we only consider (A, H)-trees. Unless otherwise
indicated, G is only assumed to be finitely generated.
At the end of this section we shall give two examples of JSJ decompositions having
flexible vertices, but most examples here will have all vertices rigid. The fact that they are
indeed JSJ decompositions will be a consequence of the following simple fact.
Lemma 3.1. Any tree T with universally elliptic vertex stabilizers is a JSJ tree.
Proof. By assumption, T dominates every tree. In particular, T is universally elliptic and
dominates every universally elliptic tree, so it is a JSJ tree.
We also note:
Lemma 3.2. Assume that all groups in A are universally elliptic. If T is a JSJ tree, then
its vertex stabilizers are universally elliptic.
This applies in particular to splittings over finite groups.
Proof. If a vertex stabilizer Gv of T is flexible, consider T 0 such that Gv is not elliptic in
T . Since T is a JSJ decomposition, it is universally elliptic, so one can consider a standard
refinement T̂ of T dominating T 0 . By our assumption on A, the tree T̂ is universally
elliptic, so by definition of the JSJ deformation space T dominates T̂ . This implies that
Gv is elliptic in T̂ , hence in T 0 , a contradiction.
3.1
Free groups
Let G = Fn be a finitely generated free group, let A be arbitrary, and H = ∅. Then the
JSJ deformation space of Fn over A is the space of free actions (unprojectivized CullerVogtmann’s outer space [CV86]).
More generally, if G is virtually free and A contains all finite subgroups, then DJSJ is
the space of trees with finite vertex stabilizers.
3.2
Free splittings: the Grushko deformation space
Let A consist only of the trivial subgroup of G, and H = ∅. Thus A-trees are trees with
trivial edge stabilizers, also called free splittings. Then the JSJ deformation space exists,
it is the outer space introduced in [GL07b] (see [CV86] when G = Fn , and [MM96] when
no free factor of G is Z). We call it the Grushko deformation space. It consists of trees T
such that edge stabilizers are trivial, and vertex stabilizers are freely indecomposable and
different from Z (one often considers Z as freely decomposable since it splits as an HNN
extension over the trivial group).
Denoting by G = G1 ∗ · · · ∗ Gp ∗ Fq a decomposition of G given by Grushko’s theorem
(with Gi non-trivial and freely indecomposable, Gi 6= Z, and Fq free), the quotient graph
of groups T /G is homotopy equivalent to a wedge of q circles; it has one vertex with group
Gi for each i, and all other vertex groups are trivial (see Figure 1 in the introduction).
If H 6= ∅, the JSJ deformation space is the Grushko deformation space relative to H.
Edge stabilizers of JSJ trees are trivial, groups in H fix a point, and vertex stabilizers are
freely indecomposable relative to their subgroups which are conjugate to a group in H.
30
3.3
Splittings over finite groups: the Stallings-Dunwoody deformation
space
If A is the set of finite subgroups of G, and H = ∅, we call the JSJ deformation space the
Stallings-Dunwoody deformation space. It is the set of trees whose edge groups are finite
and whose vertex groups have 0 or 1 end (one deduces from Stallings’s theorem that a tree
is maximal for domination if and only if its vertex stabilizers have at most one end).
The JSJ deformation space exists if G is finitely presented by Dunwoody’s original
accessibility result [Dun85]. If G is only finitely generated, it exists if and only if G
is accessible. In particular, the inaccessible group constructed in [Dun93] has no JSJ
decomposition over finite groups.
Remark 3.3. Even if G is inaccessible, there is a JSJ deformation space over A if A is a
family of finite subgroups of bounded order. The reason is that Proposition 2.17 remains
true, because Tk+1 is just a subdivision of Tk for k large by Linnell’s accessibility [Lin83].
If H 6= ∅, the JSJ deformation space is the Stallings-Dunwoody deformation space
relative to H. Edge stabilizers are finite, groups in H fix a point, and vertex stabilizers are
one-ended relative to their subgroups which are conjugate to a group in H (one-endedness
is in the sense of Subsection 1.3: they do not split over finite subgroups relative to their
subgroups which are conjugate to a group in H).
As above, the relative JSJ space exists if G is finitely generated and A consists of finite
groups with bounded order (and H is arbitrary). If A contains finite groups of arbitrary
large order, the JSJ space exists if relative accessibility holds.
3.4
Splittings of small groups
Recall that G is small in (A, H)-trees if G has no irreducible action on a tree: there always
is a fixed point, or a fixed end, or an invariant line (see Corollary 1.3 and Subsection 1.5.2).
This is in particular the case if G is small, i.e. contains no non-abelian free group. If G
acts with a fixed end or an invariant line, then every vertex stabilizer has a subgroup of
index at most 2 fixing an edge (the index is 2 if G acts dihedrally on a line).
Lemma 3.4. If G is small in (A, H)-trees, and T is a non-trivial universally elliptic tree,
then T dominates every tree.
Proof. Since the action of G on T is non-trivial but is not irreducible, every vertex stabilizer
contains an edge stabilizer with index at most 2. It follows that every vertex stabilizer is
universally elliptic, so T dominates every tree.
Corollary 3.5. If G is small in (A, H)-trees, there is at most one non-trivial deformation
space containing a universally elliptic tree.
In this situation, the JSJ deformation space always exists: if there is a deformation
space as in the corollary, it is the JSJ space; otherwise, the JSJ space is trivial.
Consider for instance (non-relative) cyclic splittings of solvable Baumslag-Solitar groups
BS(1, n) = ha, t | tat−1 = an i. If n = 1 (so G ' Z2 ), there are infinitely many deformation
spaces (corresponding to epimorphisms G → Z) and there is no non-trivial universally
elliptic tree. If n = −1 (Klein bottle group), there are exactly two non-trivial deformation
spaces: one contains the Bass-Serre tree of the HNN extension ha, t | tat−1 = a−1 i, the
other contains the tree associated to the amalgam hti ∗t2 =v2 hvi, with v = ta. None of
these trees is universally elliptic (t and v are hyperbolic in the HNN extension, and a is
hyperbolic in the amalgam).
Thus for n = ±1 (when G is Z2 or the Klein bottle group) the cyclic JSJ deformation
space of BS(1, n) is the trivial one, and G is flexible (see Subsection 6.9 for generalizations
of these examples). If n 6= ±1, the JSJ space is non-trivial, as we shall now see.
31
3.5
Generalized Baumslag-Solitar groups
Let G be a generalized Baumslag-Solitar group, i.e. a finitely generated group which acts
on a tree T with all vertex and edge stabilizers infinite cyclic. Let A be the set of cyclic
subgroups of G (including the trivial subgroup), and H = ∅. Unless G is isomorphic to
Z, Z2 , or the Klein bottle group, the deformation space of T is the JSJ deformation space
[For03].
Here is a short proof (the arguments are contained in [For03]). We show that every
vertex stabilizer H of T is universally elliptic. The commensurator of H is G because
the intersection of any pair of vertex stabilizers has finite index in both of them. If H
acts hyperbolically in a tree T 0 , its commensurator G preserves its axis, so T 0 is a line.
Edge stabilizers being cyclic, vertex stabilizers are virtually cyclic, hence cyclic since G is
torsion-free. This implies that G is Z, Z2 , or a Klein bottle group.
3.6
Locally finite trees
We generalize the previous example to locally finite trees with small edge stabilizers.
We suppose that G acts irreducibly on a locally finite tree T with small edge stabilizers
(local finiteness is equivalent to edge stabilizers having finite index in neighboring vertex
stabilizers; in particular, vertex stabilizers are small). In [GL07a, Lemma 8.5], we proved
that all such trees T belong to the same deformation space. This happens to be the JSJ
deformation space.
Proposition 3.6. Suppose that all groups of A are small in (A, H)-trees. Then any locally
finite irreducible tree T belongs to the JSJ deformation space.
Proof. We show that every vertex stabilizer H of T is universally elliptic. Since T is locally
finite, H contains an edge stabilizer with finite index, so is small (for simplicity we do not
write “small in (A, H)-trees” in this proof). By way of contradiction, assume that H is not
elliptic in some tree T 0 . Being small, H fixes a unique end of T 0 or preserves a unique line
(see Corollary 1.3). Any finite index subgroup of H preserves the same unique end or line.
As in the previous subsection, local finiteness implies that G commensurates H, so G
preserves the H-invariant end or line of T 0 (in particular, T 0 is not irreducible). We now
define a small normal subgroup G0 ⊂ G. If G does not act dihedrally on a line, there is
a fixed end and we let G0 = [G, G] be the commutator subgroup. It is small because any
finitely generated subgroup pointwise fixes a ray of T 0 , so is contained in an edge stabilizer
Ge ∈ A. If G acts dihedrally, we let G0 be the kernel of the action, so that G/G0 is an
infinite dihedral group.
Consider the action of the normal subgroup G0 on T . If it is elliptic, its fixed point set
is G-invariant, so by minimality the action of G factors through the action of an abelian
or dihedral group; this contradicts the irreducibility of T . Otherwise G0 preserves a unique
end or line; this end or line is G-invariant because G0 is normal, again contradicting the
irreducibility of T .
3.7
RAAGs
Let ∆ be a finite graph. The associated right-angled Artin group A∆ (RAAG, also called
graph group, or partially commutative group) is the group presented as follows: there is
one generator av per vertex, and a relation av aw = aw av if there is an edge between v and
w. See [Cha07] for an introduction.
The decomposition of ∆ into connected components induces a decomposition of A∆
as a free product of freely indecomposable RAAGs (which may be infinite cyclic), so to
study JSJ decompositions of A∆ one may assume that ∆ is connected (see Corollary 4.16
below). M. Clay [Cla14] determines the (non-relative) cyclic JSJ decomposition of A∆ ;
32
he gives a characterization of RAAGs with non-trivial cyclic JSJ decomposition, and he
shows that there is no flexible vertex. See [GH15] for abelian splittings of RAAGs relative
to the generators.
3.8
Parabolic splittings
Assume that G is hyperbolic relative to a family of finitely generated subgroups H =
{H1 , . . . , Hp }. Recall that a subgroup of G is parabolic if it is contained in a conjugate of
an Hi . We let A be the family of parabolic subgroups.
JSJ trees over parabolic subgroups, relative to H (equivalently, to A), exist by Theorem
2.20 because G is finitely presented relative to H [Osi06]. Parabolic subgroups are universally elliptic (all splittings are relative to H!), so JSJ trees do not have flexible vertices by
Lemma 3.2. See [Bow01], where such a JSJ decomposition is related to the cut points of
the boundary of G.
See Subsection 9.3 for the case when virtually cyclic groups are added to A.
3.9
Non-rigid examples
Unlike the previous subsections, we now consider examples with flexible vertices. This may
be viewed as an introduction to Part III. We consider cyclic splittings with H = ∅.
Suppose that G is the fundamental group of a closed orientable hyperbolic surface
Σ. Any simple closed geodesic γ on Σ defines a dual cyclic splitting (an amalgam or an
HNN-extension, depending on whether γ separates or not). A non-trivial element g ∈ G,
represented by an immersed closed geodesic δ, is elliptic in the splitting dual to γ if and
only if δ ∩ γ = ∅ or δ = γ. Since any δ meets transversely some simple γ, this shows that
1 is the only universally elliptic element of G, so the JSJ decomposition of G is trivial and
its vertex is flexible. Similar considerations apply to splittings of fundamental groups of
compact hyperbolic surfaces with boundary which are relative to the fundamental groups
of boundary components (the pair of pants is special because it contains no essential simple
geodesic, see Subsection 5.1).
The content of Section 6 is that this example is somehow a universal example.
ha2 , b2 i
ha1 , b1 i
[a1 ,b1 ]=c
hci
c
c=
[a
]
,b 2
a2
[
=
3,
b3
]
ha3 , b3 i
Figure 3: Three punctured tori attached along their boundaries, and the corresponding
JSJ decomposition.
Now suppose that G is the fundamental group of the space pictured on Figure 3 (consisting of three punctured tori attached along their boundaries). A presentation of G is
ha1 , b1 , a2 , b2 , a3 , b3 | [a1 , b1 ] = [a2 , b2 ] = [a3 , b3 ]i.
It is the fundamental group of a graph of groups Γ with one central vertex v and three
terminal vertices vi . All edges, as well as v, carry the same cyclic group C = hci. We claim
that Γ is a JSJ decomposition of G, with 3 flexible vertices vi .
Let us first show that Γ is universally elliptic, i.e. that C is universally elliptic. Using
Lemma 2.10, we consider a one-edge cyclic splitting Λ of G in which C is not elliptic,
33
and we argue towards a contradiction. The group G12 generated by Gv1 and Gv2 is the
fundamental group of a closed surface (of genus 2), in particular it is one-ended. It follows
that the edge group hai of Λ is non-trivial and has non-trivial intersection with some
conjugate of G12 , so some ak lies in a conjugate of G12 .
We now consider the action of ak on the Bass-Serre tree T of Γ. If ak fixes an edge in
T , then hak i has finite index in some conjugate of C, so C is elliptic in Λ, a contradiction.
Denote by A(ak ) the characteristic set of ak in T , i.e. its unique fixed point or its axis.
Then A(ak ) is contained in the minimal subtree of a conjugate of G12 , so contains no lift
of the vertex v3 . Permuting indices shows that all vertices in A(ak ) are lifts of v, so A(ak )
is a single point (a lift of v). This implies that hak i has finite index in some conjugate of
C, which is again a contradiction.
To prove maximality, consider a universally elliptic tree S dominating T . If the domination is strict (i.e. S and T are in different deformation spaces), some Gvi is non-elliptic
in S. Being universally elliptic, C is elliptic in S, so by a standard fact (see Proposition
5.4) the action of Gvi on its minimal subtree in S is dual to an essential simple closed curve
γi ⊂ Σi (there is only one curve because Σi is a punctured torus). Considering the splitting
of G dual to a curve γ ⊂ Σi intersecting γi non-trivially shows that S is not universally
elliptic, a contradiction.
4
A few useful facts
In this section, we first describe the behavior of the JSJ deformation space when we change
the class A of allowed edge groups. We then introduce the incidence structure inherited
by a vertex group of a graph of groups, and we relate JSJ decompositions of G to JSJ
decompositions of vertex groups relative to their incidence structure. We also discuss
relative finite presentation of vertex groups. Finally, we give an alternative construction
of relative JSJ decompositions, obtained by embedding G into a larger group.
4.1
Changing edge groups
We fix two families of subgroups A and B, with A ⊂ B, and we compare JSJ splittings
over A and over B (universal ellipticity, and all JSJ decompositions, are relative to some
fixed family H). For example:
• A consists of the finitely generated abelian subgroups of G, and B consists of the slender subgroups. This will be useful to describe the abelian JSJ decomposition in Subsection
5.5.
• groups in B are locally slender (their finitely generated subgroups are slender), and
A is the family of slender subgroups.
• G is relatively hyperbolic, A is the family of parabolic subgroups, and B is the family
of elementary subgroups (subgroups which are parabolic or virtually cyclic).
• A consists of the trivial group, or the finite subgroups of G (see Corollary 4.16).
There are now two notions of universal ellipticity, so we shall distinguish between Auniversal ellipticity (being elliptic in all (A, H)-trees) and B-universal ellipticity (being
elliptic in all (B, H)-trees).
Of course, B-universal ellipticity implies A-universal ellipticity.
Recall that two trees are compatible if they have a common refinement.
Proposition 4.1. Assume A ⊂ B. Let TB be a JSJ tree over B.
(1) If there is a JSJ tree over A, there is one which is compatible with TB . It may be
obtained by refining TB , and then collapsing all edges whose stabilizer is not in A.
(2) If every A-universally elliptic A-tree is B-universally elliptic, the tree T̄B obtained
from TB by collapsing all edges whose stabilizer is not in A is a JSJ tree over A.
34
Note that (2) applies if A consists of finite groups (more generally, of groups with
Serre’s property (FA)).
Proof. Let TA be a JSJ tree over A. Since B ⊃ A, the tree TB is elliptic with respect to
TA . Let T̂B be a standard refinement of TB dominating TA . Consider an edge e of T̂B
whose stabilizer is not in A. Then Ge fixes a unique point of TA , so any equivariant map
from T̂B to TA is constant on e. It follows that the tree obtained from T̂B by collapsing all
edges whose stabilizer is not in A dominates T2 . Being A-universally elliptic by Assertion
(ii) of Lemma 2.2, it is a JSJ tree over A.
For (2), first note that T̄B is an A-universally elliptic A-tree. If T 0 is another one, it is
B-universally elliptic, hence dominated by TB . As above, any map from TB to T 0 factors
through T̄B , so T̄B is a JSJ tree over A.
Proposition 4.2. Let A ⊂ B. Assume that every finitely generated group in B belongs to
A. If G is finitely presented relative to H, then any JSJ tree over A (relative to H) is a
JSJ tree over B (relative to H).
Proof. By Corollary 2.25, for every B-tree T (as always, relative to H), there is a morphism
S → T with S a tree with finitely generated edge stabilizers. Edge stabilizers of S fix an
edge in T , so S is an A-tree. The proposition easily follows.
This applies in particular if B is the family of all groups which are locally in A (their
finitely generated subgroups are in A). For instance, B may be the family of all locally
cyclic (resp. locally abelian, resp. locally slender) subgroups, and A the family of cyclic
(resp. finitely generated abelian, resp. slender) subgroups.
4.2
Incidence structures of vertex groups
Given a vertex stabilizer Gv of a tree, it is useful to consider splittings of Gv relative to
incident edge stabilizers, as they extend to splittings of G (Lemma 4.12). In this subsection
we give definitions and we show that Gv is finitely presented relative to incident edge groups
if G is finitely presented and edge stabilizers are finitely generated (Proposition 4.9).
4.2.1
Definitions
Let T be a tree (minimal, relative to H, with edge stabilizers in A). Let v be a vertex,
with stabilizer Gv .
Definition 4.3 (Incident edge groups Incv ). Given a vertex v of a tree T , there are finitely
many Gv -orbits of edges with origin v. We choose representatives ei and we define Incv
(or IncGv ) as the family of stabilizers Gei . We call Incv the set of incident edge groups.
It is a finite family of subgroups of Gv , each well-defined up to conjugacy.
Alternatively, one can define Incv from the quotient graph of groups Γ = T /G as the
image in Gv of the groups carried by all oriented edges with origin v. The peripheral structure, defined in [GL07a] and studied in Subsection 5.4, is a more sophisticated invariant
derived from Incv ; unlike Incv , it does not change when we replace T by another tree in
the same deformation space.
Definition 4.4 (Restriction H|Gv ). Given v, consider the family of conjugates of groups
in H that fix v and no other vertex of T . We define the restriction H|Gv by choosing a
representative for each Gv -conjugacy class in this family.
H
Definition 4.5 (IncH
v ). We define Incv = Incv ∪ H|Gv . We will sometimes write Inc|Q
H
H
and Inc|Q rather than Incv and Incv , with Q = Gv .
35
We also view H|Gv and IncH
v as families of subgroups of Gv , each well-defined up to
conjugacy.
Remark 4.6. We emphasize that H|Gv only contains groups having v as their unique fixed
point. Two such groups are conjugate in Gv if they are conjugate in G. In particular, the
number of Gv -conjugacy classes of groups in H|Gv is bounded by the number of G-conjugacy
classes of groups in H.
Also note that any subgroup of Gv which is conjugate to a group of H is contained
(up to conjugacy in Gv ) in a group belonging to IncH
v , so is elliptic in any splitting of Gv
H
relative to Incv .
4.2.2
Finiteness properties
Assume that G is the fundamental group of a finite graph of groups. It is well-known that,
if G and all edge groups are finitely generated (resp. finitely presented), then so are all
vertex groups.
The goal of this subsection is to extend these results to relative finite generation (resp.
finite presentation), as defined in Subsection 1.7. This is not needed if all groups in A are
assumed to be finitely presented.
Lemma 4.7 ([Gui08, Lemmas 1.11, 1.12]). If a finitely generated group G acts on a tree
T , then every vertex stabilizer Gv is finitely generated relative to the incident edge groups.
More generally, if G is finitely generated relative to H, and T is a tree relative to H,
then Gv is finitely generated relative to IncH
v .
Remark 4.8. If H = {H1 , . . . , Hp }, then the family H|Gv consists of at most p groups, each
conjugate to some Hi . It follows that the set of vertex stabilizers of (A, H)-trees with
finitely generated edge stabilizers is countable. This was used in the proof of Theorem
2.20.
There is a similar statement for relative finite presentation.
Proposition 4.9. If G is finitely presented, and T is a tree with finitely generated edge
stabilizers, then every vertex stabilizer Gv is finitely presented relative to the incident edge
groups.
More generally, if G is finitely presented relative to H, and T is a tree relative to H
with finitely generated edge stabilizers, then Gv is finitely presented relative to IncH
v .
We will use the following fact.
Lemma 4.10. Let X be a cellular complex, and let U ⊂ X be a compact connected subcomplex. Let G = π1 (X), and let H < G be the image of π1 (U ) in G.
Then H is finitely presented relative to the image of the fundamental groups of the
connected components of the topological boundary ∂U .
More generally, if U is not compact but there exist finitely many connected disjoint
subcomplexes Z1 , . . . , Zr ⊂ U , disjoint from ∂U , such that U \ (Z1 ∪ · · · ∪ Zr ) has compact
closure, then H is finitely presented relative to the images of the fundamental groups of
Z1 , . . . , Zr and of the connected components of ∂U .
Proof. Denote by B1 , . . . , Br the connected components of ∂U . In the special case when
each map π1 (Bi ) → π1 (X) is injective, a standard argument shows that the map π1 (U ) →
π1 (X) is also injective, so H ' π1 (U ) is finitely presented, and finitely presented relative to
its finitely generated subgroups π1 (Bi ). One reduces to the special case by gluing (possibly
infinitely many) discs to the Bi ’s. The proof of the second assertion is similar.
36
Proof of Proposition 4.9. Assume that G is finitely presented relative to {H1 , . . . , Hp }. As
in the proof of Proposition 2.24, we consider a 2-complex X and disjoint subcomplexes
Y1 , . . . , Yp with X \ (Y1 , . . . , Yp ) compact and π1 (Yi ) ' Hi . We let π : X̃ → X be the
universal cover, and Ỹi a connected component of π −1 (Yi ) whose stabilizer is Hi . Let
f : X̃ → T be an equivariant map which sends vertex to vertex, sends each Ỹi to a vertex
fixed by Hi , is constant or injective on each edge, and is standard on each triangle.
Consider the pattern τ̃ ⊂ X̃ obtained as the preimage of all midpoints of edges of T ,
and its projection τ ⊂ X. Let W ⊂ X be the closure of the complement of a regular
neighborhood of τ .
Let S be the tree dual to τ̃ . First suppose that S = T . Applying Lemma 4.10 to the
component U of W corresponding to v shows that Gv is finitely presented relative to the
family Hv0 consisting of Incv and those Hi ’s for which Yi is contained in U . If H 6= ∅ this
may not be quite the required family IncH
v : we have to remove Hi if it fixes an edge of T
(see Subsection 4.2.1). But such an Hi is contained (up to conjugacy in Gv ) in a group
belonging to IncH
v , and we can use Lemma 1.9.
Now suppose S 6= T , and recall that edge stabilizers of T are finitely generated. When
H = ∅ we may use [LP97, Theorem 0.6], saying that T is geometric, to construct X̃ and
f such that S = T . In general, we may write the induced map g : S → T as a finite
composition of folds (see the proposition on page 455 of [BF91]). It therefore suffices to
show that, given a factorization g = h◦ρ with ρ a fold, we may change X̃ to a new complex
X̃1 , with f1 : X̃1 → T , so that the associated map g1 equals h.
Let e, e0 be adjacent edges of S which are folded by ρ. They are dual to components
of τ̃ adjacent to a component Z of X̃ \ τ̃ . Let ab and a0 b0 be edges of X̃ mapping to e and
e0 in S respectively, with a and a0 in Z (subdivide X if needed). Let γ be an edge path
joining a to a0 in Z. Note that γ is mapped to a single vertex in S, and that b, b0 have the
same image in T .
Now glue a square [0, 1] × [0, 1] to X̃, with vertical edges glued to e, e0 and the edge
[0, 1] × {0} glued to γ (the edge [0, 1] × {1} is free). The map f : X̃ → T extends to
the square, with all vertical arcs {t} × [0, 1] mapped to the edge f (e) = f (e0 ). Doing this
gluing G-equivariantly yields the desired X̃1 .
4.3
JSJ decompositions of vertex groups
Given an (A, H)-tree T , we compare splittings of G and relative splittings of vertex stabilizers Gv .
Recall that IncH
v is the family of incident edge stabilizers together with H|Gv (see
Subsection 4.2.1), and that Gv is finitely generated relative to IncH
v . In particular, whenever
Gv acts on an (A, H)-tree S relative to Incv with no global fixed point, there is a unique
minimal Gv -invariant subtree µS (Gv ) ⊂ S by Proposition 1.7. We view µS (Gv ) as a tree
with an action of Gv ; if Gv is elliptic in S, we let µS (Gv ) be any fixed point.
Definition 4.11 (Av ). We denote by Av the family consisting of subgroups of Gv belonging
to A. All splittings of Gv will be over groups in Av .
Lemma 4.12. Let Gv be a vertex stabilizer of an (A, H)-tree T . Any splitting of Gv
relative to IncH
v extends (non-uniquely) to a splitting of G relative to H.
More precisely, given an (Av , IncH
v )-tree Sv , there exist an (A, H)-tree T̂ and a collapse
−1
map p : T̂ → T such that p (v) is Gv -equivariantly isomorphic to Sv .
We say that T̂ is obtained by refining T at v using Sv . More generally, one may choose
a splitting for each orbit of vertices of T , and refine T using them. Any refinement of T
may be obtained by this construction (possibly with non-minimal trees Sv ).
37
Proof. We construct T̂ as in the proof of Proposition 2.2 , with Yv = Sv . It is relative
to H because any group in H which is conjugate to a subgroup of Gv is conjugate to a
subgroup of a group belonging to IncH
v . Non-uniqueness comes from the fact that there
may be several ways of attaching edges of T to Sv (see [GL16, Section 4.2]).
Lemma 4.13. Let Gv be a vertex stabilizer of a universally elliptic tree T .
• The groups in IncH
v are elliptic in every (A, H)-tree S.
• A subgroup H < Gv is (A, H)-universally elliptic (as a subgroup of G) if and only if
it is (Av , IncH
v )-universally elliptic (as a subgroup of Gv ).
The second assertion says that H is elliptic in every (A, H)-tree on which G acts if
and only if it is elliptic in every (Av , Incv ∪ H|Gv )-tree on which Gv acts. If this holds, we
simply say that H is universally elliptic.
Proof. The first assertion is clear: S is relative to H|Gv , and also to Incv because T is
universally elliptic.
Suppose that H is (Av , IncH
v )-universally elliptic, as a subgroup of Gv . Let S be any
(A, H)-tree. It is relative to IncH
v by the first assertion. Since Gv is finitely generated
H
relative to Incv , it has a minimal subtree Sv ⊂ S by Proposition 1.7. The action of Gv on
Sv is an (Av , IncH
v )-tree, so by assumption H fixes a point in Sv , hence in S.
We have proved the “if” direction in the second assertion, and the converse follows from
Lemma 4.12.
Corollary 4.14. Let Gv be a vertex stabilizer of a JSJ tree TJ .
(1) Gv does not split over a universally elliptic subgroup relative to IncH
v .
(2) Gv is flexible if it splits relative to IncH
v , rigid otherwise.
Proof. If there is a splitting as in (1), we may use it to refine TJ to a universally elliptic
tree (see Lemma 4.12). This tree must be in the same deformation space as TJ , so the
splitting of Gv must be trivial. (2) follows from Lemma 4.13 applied with H = Gv .
Proposition 4.15. Let T be a universally elliptic (A, H)-tree.
(1) Assume that every vertex stabilizer Gv of T has a JSJ tree Tv relative to IncH
v . One
can then refine T using these decompositions so as to obtain a JSJ tree of G relative
to H.
(2) Conversely, if TJ is a JSJ tree of G relative to H, and Gv is a vertex stabilizer of T ,
one obtains a JSJ tree for Gv relative to IncH
v by considering the action of Gv on its
minimal subtree Tv = µTJ (Gv ) in TJ (with Tv a point if Gv is elliptic).
Proof. To prove (1), let T̂ be the tree obtained by refining T using the Tv ’s as in Lemma
4.12. It is relative to H, and universally elliptic by Lemma 4.13 since its edge stabilizers are
edge stabilizers of T or of some Tv . To show maximality, we consider another universally
elliptic (A, H)-tree T 0 , and we show that any vertex stabilizer H of T̂ is elliptic in T 0 . It is
a vertex stabilizer of some Tv , with v a vertex of T . If Gv is not elliptic in T 0 , its minimal
subtree Yv is a universally elliptic (Av , IncH
v )-tree. Since Tv is a JSJ tree, it dominates Yv
so H is elliptic in T 0 . This proves (1).
Now let TJ and Tv ⊂ TJ be as in (2). By Lemma 4.13 Tv is relative to IncH
v , and it is
H
(Hv , Incv )-universally elliptic because its edge stabilizers are contained in edge stabilizers
of TJ . To prove maximality of Tv , consider another tree Sv with an action of Gv which is
relative to IncH
v and universally elliptic. Use it to refine T to a tree T̂ as in Lemma 4.12.
38
As above, T̂ is relative to H, and universally elliptic by Lemma 4.13. Being a JSJ tree,
TJ dominates T̂ . Vertex stabilizers of Tv are elliptic in TJ , hence in T̂ , hence in Sv , so Tv
dominates Sv . This proves (2).
The following corollary says that one may usually restrict to one-ended groups when
studying JSJ decompositions.
Corollary 4.16. Suppose that A contains all finite subgroups of G.
If H = ∅, refining a Grushko decomposition of G using JSJ decompositions of free
factors yields a JSJ decomposition of G.
Similarly, refining a Stallings-Dunwoody decomposition of G using JSJ decompositions
of the vertex groups Gv yields a JSJ decomposition of G.
If H 6= ∅, one must use relative Grushko or Stallings-Dunwoody decompositions, and
JSJ decompositions of vertex groups Gv relative to H|Gv .
Every flexible subgroup of G is a flexible subgroup of some Gv .
As mentioned in Subsection 3.3, Stallings-Dunwoody decompositions only exist under
some accessibility assumption.
Proof. This follows from the proposition, applied with T in the Grushko or StallingsDunwoody deformation space: finite groups are universally elliptic, and every splitting
of Gv is relative to Incv . The assertion about flexible subgroups follows from Corollary
4.14.
4.4
Relative JSJ decompositions through fillings
Fix a finitely presented group G and a family A. In Subsection 2.5 we have shown the
existence of the JSJ deformation space of G relative to a finite set H = {H1 , . . . , Hp } of
finitely generated subgroups. We now give an alternative construction, using (absolute) JSJ
decompositions of another group Ĝ obtained by a filling construction. This construction
will be used in Subsections 5.4 and 5.5 to provide examples of flexible groups.
G
H1 × R 1
H1
H2
..
.
Hp
H2 × R 2
Hp × R p
Figure 4: The group Ĝ obtained by the filling construction.
The filling construction. Let G and H be as above. For i ∈ {1, . . . , p}, we let Ri
be a non-trivial finitely presented group with property (FA), and we define a group Ĝ by
amalgamating G with Ki = Hi × Ri over Hi for i = 1, . . . , p (see Figure 4); in other words,
Ĝ = ((G ∗H1 K1 ) ∗ . . . ) ∗Hp Kp . It is finitely presented. We denote by T the Bass-Serre
tree of this amalgam, and by v the vertex of T with stabilizer G. The stabilizer of an edge
with origin v is conjugate to one of the Hi ’s in G.
Fix a family B of subgroups of Ĝ such that Bv = A and Ri ∈
/ B, for instance the
family of subgroups of Ĝ having a conjugate in A (note that two subgroups of G which
are conjugate in Ĝ are also conjugate in G because Hi is central in Ki , so this family B
induces A).
A subgroup of G is (A, H)-universally elliptic if it is elliptic in all splittings of G over
A relative to H. A subgroup of Ĝ is B-universally elliptic if it is elliptic in all splittings of
Ĝ with edge groups in B (splittings of Ĝ are non-relative).
39
Lemma 4.17. Hi × Ri is B-universally elliptic. A subgroup J ⊂ G is (A, H)-universally
elliptic if and only if J (viewed as a subgroup of Ĝ) is B-universally elliptic.
Proof. Consider any B-tree. The group Ri fixes a point by property (FA), which is unique
since Ri ∈
/ B. This point is also fixed by Hi since Hi commutes with Ri . This proves the
first assertion. Since T is B-universally elliptic, and H is the family of incident edge groups
at v, the second assertion follows from Lemma 4.13.
Being finitely presented, Ĝ has a JSJ decomposition TJ over B by Theorem 2.16. We
let TG = µTJ (G) be the minimal G-invariant subtree (a point if G is elliptic in TJ ).
Proposition 4.18. The tree TG is a JSJ tree of G over A relative to H.
Proof. The tree TG has edge stabilizers in Bv = A, is relative to H and (A, H)-universally
elliptic by Lemma 4.17. We have to show that it dominates any universally elliptic (A, H)tree SG . Use SG to refine T at v into a tree T̂ , as in Lemma 4.12. The tree T̂ has two
types of edges, those coming from T and those in the Ĝ-orbit of SG .
Define a new tree T 0 by collapsing all edges of T̂ coming from T , and note that a
subgroup of G is elliptic in T 0 if and only if it is elliptic in SG . Indeed, a subgroup H of
Ĝ is elliptic in T̂ if and only if it is elliptic in both T 0 and T ; for H < G (hence elliptic in
T ), being elliptic in T 0 is equivalent to being elliptic in T̂ , i.e. in SG .
The tree T 0 is a B-tree, and it is universally elliptic by Lemma 4.17, so it is dominated
by TJ . Vertex stabilizers of TG are then elliptic in T 0 , hence in SG , so TG dominates
SG .
Part III
Flexible vertices
Flexible vertex groups of JSJ decompositions are most important, as understanding their
splittings conditions the understanding of the splittings of G. The key result is that, in
many cases, flexible vertex groups are “surface-like”.
For instance, first consider cyclic splittings of a torsion-free group G. We will see that,
if Gv is a flexible vertex group of a JSJ decomposition Γ = T /G, then Gv may be viewed as
π1 (Σ), with Σ a compact (possibly non-orientable) surface. Moreover, incident edge groups
are trivial or contained (up to conjugacy) in a boundary subgroup, i.e. the fundamental
group B = π1 (C) of a boundary component C of Σ. Boundary subgroups being generated
by elements which are quadratic words in a suitable basis of the free group π1 (Σ), Rips
and Sela called the subgroup Gv quadratically hanging (QH).
If Σ is not too simple, there are infinitely many isotopy classes of essential two-sided
simple closed curves. Each such curve defines a cyclic splitting of π1 (Σ) relative to incident
edge groups, which extends to a cyclic splitting of G by Lemma 4.12. Two curves which
cannot be made disjoint by an isotopy define two splittings with are not elliptic with
respect to each other, and this makes Gv flexible (see Corollary 4.14). It turns out that
this construction is basically the only source of flexible vertices.
If G is allowed to have torsion, or if non-cyclic edge groups are allowed, the definition
of “surface-like” must be adapted. First, Σ may be a 2-dimensional orbifold rather than a
surface. Second, Gv is not always equal to π1 (Σ); it only maps onto π1 (Σ), with a possibly
non-trivial kernel F (called the fiber).
Various authors have called such groups Gv hanging surface groups, hanging Fuchsian
groups, hanging K-by-orbifold groups, V P C-by-Fuchsian type vertex groups... We choose
to extend Rips and Sela’s initial terminology of QH groups, to emphasize the way in which
Gv is attached to the rest of the group G (in Rips-Sela [RS97], F is trivial and Σ has no
40
mirror, see Theorem 6.6; on the other hand, we insist that a QH group be based on a
hyperbolic orbifold, not on a Euclidean one).
In Section 5 we formalize the definition of QH vertices, and prove general properties of
such vertices. In Section 6 we show that, indeed, flexible vertices of the JSJ deformation
space over nice classes of slender subgroups are QH.
As before, we fix a family A closed under conjugating and passing to subgroups, and
another family H. All trees are assumed to be (A, H)-trees.
5
Quadratically hanging vertices
In this section, after preliminaries about 2-orbifold groups, we give a definition of QH
subgroups. We study their basic properties, in particular we relate their splittings to
families of simple geodesics on the underlying orbifold Σ.
We then show that, under natural hypotheses, any QH subgroup has to be elliptic in
the JSJ deformation space.
In Subsection 5.4 we give examples of possible incident edge groups for QH vertex
groups of a JSJ decomposition. This will be relevant in Section 6, where we show that
flexible subgroups of the slender JSJ decomposition are QH. We also show that flexible
subgroups of abelian JSJ decompositions do not have to be QH.
5.1
5.1.1
2-orbifolds and their splittings
Hyperbolic 2-orbifolds
Most compact 2-dimensional orbifolds Σ, including all those that will concern us, are
Euclidean or hyperbolic (we refer to [Sco83] and [Thu80, Ch. 13] for basic facts about
orbifolds). Euclidean orbifolds whose fundamental group is not virtually cyclic have empty
boundary, so they can only arise from flexible vertices in a trivial way. For instance, in
the case of cyclic splittings of torsion-free groups, Z2 = π1 (T 2 ) and the Klein bottle group
may appear as flexible vertex groups (see Subsections 3.4 and 6.9), but only as free factors
(all incident edge groups must be trivial).
We therefore restrict to hyperbolic orbifolds: Σ is a compact 2-dimensional orbifold,
equipped with a hyperbolic metric with totally geodesic boundary. It is the quotient of
a convex subset Σ̃ ⊂ H2 by a proper discontinuous group of isometries GΣ ⊂ Isom(H2 )
(isometries may reverse orientation); we denote by p : Σ̃ → Σ the quotient map. By
definition, the (orbifold) fundamental group of Σ is π1 (Σ) = GΣ . We may also view Σ as
the quotient of a compact orientable hyperbolic surface Σ0 with geodesic boundary by a
finite group of isometries Λ. A point of Σ is singular if its preimages in Σ̃ (or in Σ0 ) have
non-trivial stabilizer.
If we forget the orbifold structure, Σ is homeomorphic to a surface Σtop (a disc in Figure
5). The boundary of Σtop comes from the boundary ∂ Σ̃, and from mirrors corresponding
to reflections in GΣ (see below). We define the boundary ∂Σ of Σ as the image of ∂ Σ̃ in
Σ (thus excluding mirrors). Equivalently, it is the image of ∂Σ0 . Each component C of
∂Σ is either a component of ∂Σtop (a circle) or an arc contained in ∂Σtop . The (orbifold)
fundamental group of C is Z or an infinite dihedral group D∞ accordingly. A boundary
subgroup is a subgroup B ⊂ π1 (Σ) which is conjugate to the fundamental group of a
component C of ∂Σ. Equivalently, it is the setwise stabilizer of a connected component of
∂ Σ̃.
The closure of the complement of ∂Σ in ∂Σtop is a union of mirrors: a mirror is the
image of a component of the fixed point set of an orientation-reversing element of π1 (Σ)
in Σ̃. Equivalently, a mirror is the image of a component of the fixed point set of an
orientation-reversing element of Λ in Σ0 . Each mirror is itself a circle or an arc contained
41
Z/2
{1}
Z/2
Z/2
Z/2
D4
D4
Z/2
Z/2
Z/2
{1}
Z/2
Figure 5: An orbifold with 4 mirrors (in bold), 2 boundary components, and 2 corner
reflectors carrying D4 . Its fundamental group is the Coxeter group generated by reflections
over 4 sides of a right-angled hexagon in H2 .
in ∂Σtop . Mirrors may be adjacent in ∂Σtop , whereas boundary components of Σ are
disjoint.
Singular points not contained in mirrors are conical points; the stabilizer of their preimages in H2 is a finite cyclic group consisting of orientation-preserving maps (rotations).
Points belonging to two mirrors are corner reflectors; the associated stabilizer is a finite
dihedral group D2r of order 2r.
As in the case of surfaces, hyperbolic orbifolds may be characterized in terms of their
Euler characteristic (see [Thu80, Sco83]).
Definition 5.1 (Euler characteristic). The Euler characteristic χ(Σ) is defined as the
Euler characteristic of the underlying topological surface Σtop , minus contributions coming
from the singularities: a conical point of order q (with isotropy group Z/qZ) contributes
1 − 1q , a corner reflector with isotropy group the dihedral group D2r of order 2r contributes
1
1
1
2 (1 − r ), and a point adjacent to a mirror and a component of ∂Σ contributes 4 .
Proposition 5.2. A compact 2-dimensional orbifold Σ is hyperbolic if and only if χ(Σ) <
0.
5.1.2
Curves and splittings
We now generalize the fact that any essential 2-sided simple closed curve on a surface Σ
defines a cyclic splitting of π1 (Σ).
Let Σ be a hyperbolic 2-orbifold as above. A closed geodesic γ ⊂ Σ is the image in Σ
of a bi-infinite geodesic γ̃ ⊂ Σ̃ whose image in Σ is compact. It is simple if hγ̃ and γ̃ are
equal or disjoint for all h ∈ π1 (Σ). If γ̃ 6⊂ ∂ Σ̃, we say that γ is an essential simple closed
geodesic in Σ (possibly one-sided). For brevity we often just call γ a geodesic.
If γ is an essential simple closed geodesic, then p−1 (γ), the orbit of γ̃ under π1 (Σ), is
a family of disjoint geodesics. There is a simplicial tree Tγ dual to this family: vertices of
Tγ are components of Σ̃ \ p−1 (γ), edges are components of p−1 (γ). The group π1 (Σ) acts
non-trivially on Tγ , but there are inversions if γ is one-sided; in this case we subdivide
edges so as to get an action without inversions (this may viewed as replacing γ by the
boundary of a regular neighborhood, a connected 2-sided simple 1-suborbifold bounding a
Möbius band).
We have thus associated to γ a one-edge splitting of π1 (Σ), which is clearly relative to
the boundary subgroups. We call it the splitting dual to the essential simple closed geodesic
γ (or the splitting determined by γ). Its edge group (well-defined up to conjugacy) is the
subgroup Hγ ⊂ GΣ consisting of elements which preserve γ̃ and each of the half-spaces
bounded by γ̃. It is isomorphic to Z or D∞ (the infinite dihedral group).
42
More generally, there is a splitting dual to any family L of disjoint essential simple
closed geodesics γi . For simplicity, we sometimes just say that the splitting is dual to a
family of geodesics.
Recall that a group is small if it does not contain F2 . The next result says that this
construction yields all small splittings of π1 (Σ) relative to the boundary subgroups (note
that all small subgroups of π1 (Σ) are virtually cyclic).
Remark 5.3. Also note that any subgroup of π1 (Σ) that preserves a line or an end in such
a splitting is virtually cyclic.
Proposition 5.4. Let Σ be a compact hyperbolic 2-orbifold. Assume that π1 (Σ) acts on
a tree T non-trivially, without inversions, minimally, with small edge stabilizers, and with
all boundary subgroups elliptic.
Then T is equivariantly isomorphic to the Bass-Serre tree of the splitting dual to a
family L of disjoint essential simple closed geodesics of Σ.
If edge stabilizers are not assumed to be small, T is still dominated by a tree dual to a
family of geodesics.
Remark 5.5. In this statement, we assume that T has no redundant vertex, and we do
not allow multiple parallel simple closed curves (because we consider geodesics). If T is
allowed to have redundant vertices, then it is only isomorphic to a subdivision of the tree
dual to a family of geodesics.
Proof. When Σ is an orientable surface, this follows from Theorem III.2.6 of [MS84]. If Σ
is a 2-orbifold, we consider a covering surface Σ0 as above. The action of π1 (Σ0 ) on T is
dual to a family of closed geodesics on Σ0 . This family is Λ-invariant and projects to the
required family on Σ. The action of π1 (Σ) on T is dual to this family.
The second statement follows from standard arguments (see the proof of [MS84, Theorem III.2.6]).
Corollary 5.6. π1 (Σ) has a non-trivial splitting relative to the boundary subgroups if and
only if Σ contains an essential simple closed geodesic.
Orbifolds with no essential simple closed geodesic are classified in the next subsection.
Proposition 5.4 implies in particular that π1 (Σ) is one-ended relative to its boundary
subgroups. This does not remain true if we set one boundary component aside.
Lemma 5.7. Let C be a boundary component of a compact hyperbolic 2-orbifold Σ. There
exists a non-trivial splitting of π1 (Σ) over {1} or Z/2Z relative to the fundamental groups
Bk of all boundary components distinct from C.
Proof. Any arc γ properly embedded in Σtop and with endpoints on C defines a free splitting
of π1 (Σ) relative to the groups Bk . In most cases one can choose γ so that this splitting
is non-trivial. We study the exceptional cases: Σtop is a disc or an annulus, and Σ has no
conical point.
If Σtop is a disc, its boundary circle consists of components of ∂Σ and mirrors. Since Σ
is hyperbolic, there must be a mirror M not adjacent to C (otherwise ∂Σtop would consist
of C and one or two mirrors, or two boundary components and two mirrors, and χ(Σ)
would not be negative). An arc γ with one endpoint on C and the other on M defines a
splitting over Z/2Z, which is non-trivial because M is not adjacent to C.
If Σtop is an annulus, there are two cases. If C is an arc, one can find an arc γ from
C to C as in the general case. If C is a circle in ∂Σtop , the other circle contains a mirror
M (otherwise Σ would be a regular annulus) and an arc γ from C to M yields a splitting
over Z/2Z.
43
Remark 5.8. If the splitting constructed is over K = Z/2Z, then Σ contains a mirror and
K is contained in an infinite dihedral subgroup (generated by K and a conjugate).
Definition 5.9 (Filling geodesics). Let Σ be a compact hyperbolic 2-orbifold, and let C be a
non-empty collection of (non-disjoint) essential simple closed geodesics in Σ. We say that
C fills Σ if the following equivalent conditions hold:
(1) For every essential simple closed geodesic α in Σ, there exists γ ∈ C that intersects
α non-trivially (with γ 6= α).
(2) For every element g ∈ π1 (Σ) of infinite order that is not conjugate into a boundary
subgroup, there exists γ ∈ C such that g acts hyperbolically in the splitting of π1 (Σ)
dual to γ.
(3) The full preimage C˜ of C in the universal covering Σ̃ is connected.
The equivalence between these conditions is well known. We include a proof for completeness.
Proof. (2) ⇒ (1) is clear, using a g representing α.
˜ and its convex hull A in
To prove (1) ⇒ (3), consider a connected component C0 of C,
Σ̃ ⊂ H2 . If C˜ is not connected, then A 6= Σ̃; indeed, A is contained in a half space bounded
by a geodesic in C˜ \ C0 . Let α be a connected component of the boundary of A in Σ̃, a
˜ Indeed, if γ ∈ C˜ cuts α, then
bi-infinite geodesic. We note that α cuts no geodesic of C.
γ∈
/ C0 , so all elements of C0 are disjoint from γ, and A is contained in a half-space bounded
by γ, contradicting α ⊂ Ā. Now α projects to a simple closed geodesic in Σ (if gα did
intersect α transversely, then A would be contained in the intersection of the half spaces
bounded by α and gα, a contradiction). Assumption (1) ensures that α is a boundary
component of Σ̃. This implies that A = Σ̃, and that C˜ is connected.
To prove (3) ⇒ (2), consider g ∈ π1 (Σ) of infinite order, and let A(g) be its axis in Σ̃.
˜ By connectedness, C˜ is contained in
If (2) does not hold, it intersects no geodesic in C.
one of the half-spaces bounded by A(g). It follows that the convex hull A of C˜ is properly
contained in Σ̃. Since A is π1 (Σ)-invariant, this is a contradiction.
Corollary 5.10. If Σ contains at least one essential simple closed geodesic, then the set
of simple closed geodesics fills Σ.
Using the first definition of filling, this follows immediately from the following lemma.
Lemma 5.11 (Lemma 5.3 of [Gui00b]). If γ0 is any essential simple closed geodesic, there
exists another essential simple closed geodesic γ1 intersecting γ0 non-trivially.
5.1.3
Small orbifolds
It is well-known that the pair of pants is the only compact hyperbolic surface containing no
essential simple closed geodesic. In this subsection we classify hyperbolic 2-orbifolds which
do not contain an essential geodesic. Their fundamental groups do not split relative to the
boundary subgroups (Corollary 5.6), and they do not appear in flexible vertex groups of
JSJ decompositions (see Subsection 5.2).
As above, we work with compact orbifolds with geodesic boundary, but we could equally
well consider orbifolds with cusps.
Proposition 5.12. A compact hyperbolic 2-orbifold with geodesic boundary contains no
simple closed essential geodesic if and only if it belongs to the following list (see Figure 6):
44
r
2
2
2
p
p
2p
2r
2
2
q
2
2
2p
p
2
2
2
p
2
2
2q
2
q
2p
2
2
2
2
2
2q
2
2
2
2
2
2
2
2
2
2
2
2
Figure 6: Orbifolds with no splittings (mirrors in bold, labels = isotropy size).
(1) a sphere with 3 conical points, a disc with 2 conical points, an annulus with 1 conical
point, a pair of pants (Σ has no mirror);
(2) a disk whose boundary circle is the union of a single mirror with a single boundary
segment, with exactly one conical point;
(3) an annulus with one mirror and no conical point;
(4) a disk whose boundary circle is the union of three mirrors, together with at most 3
boundary segments (no conical point).
Orbifolds Σ in this list are hyperbolic if and only if χ(Σ) < 0 (see Proposition 5.2).
Proof. Let Σ be an orbifold with no closed geodesic, and let Σ∗top be the underlying topological surface with the conical points removed. It has to be orientable, since otherwise it
contains an embedded Möbius band, whose core yields an essential simple closed geodesic.
Similarly, Σ∗top has to be planar, with at most 3 boundary components or punctures.
If Σ has no mirror, we must be in case (1): the total number of conical points and
boundary components must be three because χ(Σ) is negative.
We therefore assume that Σ has a mirror m. Recall that each component c of ∂Σ
is contained in a component b of ∂Σ∗top (a circle). If c 6= b, then c is a segment whose
endpoints belong to (possibly equal) mirrors contained in b.
The classification when there are mirrors relies on the inequality χ(Σ) < 0 and the
following basic observation. Consider a properly embedded arc joining m to a mirror n
(possibly equal to m). Since it cannot be isotopic to an essential simple closed geodesic,
it may be isotoped (in the complement of the conical points, with its endpoints remaining
on m and n respectively) to a boundary segment of Σ or to an arc contained in m.
This implies that the connected component b of ∂Σtop containing m is the only connected component of ∂Σtop containing a mirror, and b cannot contain more than 3 mirrors.
Considering arcs joining m to m, it also implies that Σ∗top is a disc or an annulus (not a
pair of pants).
If Σ∗top is an annulus, the only possibility is that b = m ∪ s with s a boundary segment
of Σ. This corresponds to cases (2) and (3), depending on whether Σ has a conical point
(necessarily of order > 2) or a second boundary component.
The only remaining possibility is that Σtop is a disc with boundary b, and there is
no conical point. Any pair of mirrors in b have to be adjacent, or joined by a boundary
segment of Σ. It follows that b contains at most 3 mirrors, and at most as many boundary
segments. There must be 3 mirrors in order for χ(Σ) to be negative.
45
5.1.4
Orbifold with finite mapping class group
Suppose that Σ is a compact hyperbolic surface. Unless it is a pair of pants, it contains
a simple closed geodesic γ. If γ is 2-sided, it defines a splitting of π1 (Σ) over a maximal
cyclic subgroup Hγ , and the Dehn twist around γ defines an infinite order element of the
mapping class group. If γ is one-sided, the dual splitting is over an index 2 subgroup of
π1 (γ), corresponding to the boundary curve γ̂ of a regular neighborhood; the Dehn twist
around γ̂ is homotopically trivial.
The pair of pants and the twice-punctured projective plane contain no 2-sided geodesic
and have finite mapping class group. All other hyperbolic surfaces contain a 2-sided
geodesic and have infinite mapping class group (note that the once-punctured Klein bottle
contains a unique 2-sided geodesic; like the closed non-orientable surface of genus 3, it has
no pseudo-Anosov mapping class).
In this subsection we generalize this discussion to orbifolds.
There are more examples of simple closed geodesics which do not yield non-trivial
twists. If γ is a mirror which is a full circle in ∂Σtop , its fundamental group is isomorphic
to Z ⊕ Z/2Z; it defines a splitting of π1 (Σ) over the index 2 subgroup isomorphic to Z, but
the associated Dehn twist is once again trivial. If γ is a geodesic arc whose endpoints belong
to mirrors, its fundamental group is the dihedral group D∞ and there is no associated Dehn
twist (because D∞ has trivial center). If γ is a mirror joining two corner reflectors carrying
D4 , it yields a splitting over D∞ but no twist.
We refer to [GL15b] (especially Theorem 7.14) for a general discussion of the relations
between splittings and automorphisms (see also [DG11] or [Lev05a], and the discussion
in Subsection 9.5). Here we limit ourselves to classifying orbifold groups which do not
split over a maximal cyclic subgroup relative to the boundary subgroups; equivalently, we
classify orbifolds Σ such that the interior of Σ∗top contains no 2-sided geodesic. These are
the orbifolds whose mapping class group is finite.
Let Σ be such an orbifold. We first note that Σ must be planar with at most 3 conical
points or boundary components, or be a Möbius band with one conical point, or a Möbius
band with one open disc removed, or a projective plane with 2 conical points.
Now say that a component b of ∂Σtop is simple if b contains no mirror (it is a component
of ∂Σ) or b is a single mirror; we say that b is a circular boundary component or a circular
mirror accordingly. Note that a simple boundary component does not contribute to χ(Σ)
(see Definition 5.1). In particular, if Σ is planar and all boundary components of Σtop are
simple, the total number of boundary components and conical points is 3.
If b is a non-simple boundary component, consider a simple closed curve parallel to it
inside Σtop . Since it is not isotopic to a geodesic, it must be parallel to a simple boundary
component b0 6= b, or bound a Möbius band containing no conical point, or bound a disc
containing at most one conical point.
One can now check that Σ must be in the following list. We write, e.g., (S 2 , 3) to mean
that the surface obtained from Σtop by removing the boundary and the conical points is a
sphere minus 3 points.
• (S 2 , 3): a sphere with a choice of 3 conical points, circular boundary components, or
circular mirrors;
• (S 2 , 2): an annulus one of whose boundary component is non-simple, and the other
is either a conical point or a circular boundary component or a circular mirror;
• (S 2 , 1): a disc whose boundary is non-simple, with no conical point;
• (P 2 , 2): a projective plane with a choice of 2 conical points, circular boundary components, or circular mirrors;
• (P 2 , 1): a Möbius band whose boundary is non-simple (no conical point).
46
5.2
Definition and properties of quadratically hanging subgroups
As usual, we fix G, A, and H. Let Q be a subgroup of G.
Definition 5.13 (QH subgroup, fiber, extended boundary subgroup). We say that Q is a
QH subgroup (over A, relative to H) if:
(1) Q = Gv is the stabilizer of a vertex v of an (A, H)-tree T ;
(2) Q is an extension 1 → F → Q → π1 (Σ) → 1, with Σ a compact hyperbolic 2-orbifold;
we call F the fiber, and Σ the underlying orbifold;
(3) each incident edge stabilizer, and each intersection Q ∩ gHg −1 for H ∈ H, is an
extended boundary subgroup: by definition, this means that its image in π1 (Σ) is
finite or contained in a boundary subgroup B of π1 (Σ).
Condition (3) may be rephrased as saying that all groups in IncH
v (see Definition 4.5)
are extended boundary subgroups.
In full generality, the isomorphism type of Q does not necessarily determine F and Σ;
when we refer to a QH subgroup, we always consider F and Σ as part of the structure. If
F is small, however, it may be characterized as the largest normal subgroup of Q which
is small; in particular, any automorphism of Q leaves F invariant. The group Q does not
have to be finitely generated if F is not, but it is finitely generated relative to IncH
v by
Lemma 4.7, so Proposition 1.7 (guaranteeing the existence of a minimal subtree) applies
to actions of Q relative to IncH
v .
A vertex v as above, as well as its image in Γ = T /G, is called a QH vertex. It is the
only point of T fixed by Q because extended boundary subgroups are proper subgroups
of Q. Also note that the preimage in Q of the finite group carried by a conical point or a
corner reflector of Σ is an extended boundary subgroup.
Any incident edge stabilizer Ge is contained in an extension of F by a virtually cyclic
subgroup of π1 (Σ). But, even if G is one-ended, Ge may meet F trivially, or have trivial
image in π1 (Σ) (see Subsection 5.4). In particular, in full generality, F does not have to
belong to A, or be universally elliptic if T is a JSJ tree.
Definition 5.14 (Dual splitting of G, group Qγ ). A splitting of π1 (Σ) dual to a family
of geodesics L as in Subsection 5.1.2 induces a splitting of Q relative to F . By the third
condition of Definition 5.13, this splitting is also relative to IncH
|Q (see Definition 4.5), so
extends to a splitting of G relative to H by Lemma 4.12. We say that this splitting of G is
dual to L (or determined by L). The edge group associated to γ ∈ L is denoted Qγ .
The edge groups Qγ are extensions of the fiber F by Z or D∞ , and in general they do
not have to be in A. For many natural classes of groups A, though, such extensions are
still in A (this is the content of the stability conditions of Subsection 6.1).
Conversely, we have seen (Proposition 5.4) that small splittings of an orbifold group
relative to the boundary subgroups are dual to families of geodesics. We will prove a similar
statement (Lemma 5.18) for splittings of QH groups relative to incident edge stabilizers
and groups in H|Q (i.e. to IncH
|Q ), but one has to make additional assumptions.
First, the fiber has to be elliptic in the splitting; this is not automatic in full generality,
but this holds as long as F and groups in A are slender, see Lemma 5.23 below. Next, we
need to ensure that boundary subgroups are elliptic; this motivates the following definition.
Definition 5.15 (Used boundary component). A boundary component C of Σ is used if
the group B = π1 (C) (isomorphic to Z or D∞ ) contains with finite index the image of an
incident edge stabilizer or of a subgroup of Q conjugate (in G) to a group in H.
47
Equivalently, C is used if there exists some subgroup H ∈ IncH
v whose image in π1 (Σ)
is infinite and contained in π1 (C) up to conjugacy.
Using Lemma 5.7, we get:
Lemma 5.16. Let Q = Gv be a QH vertex group of a tree T , with fiber F . If some
boundary component is not used, then G splits relative to H over a group F 0 containing F
with index at most 2. Moreover, T can be refined at v using this splitting.
Remark 5.17. F 0 = F whenever the underlying orbifold has no mirror. In general, F 0 does
not have to be in A.
Proof. Let C be a boundary component of Σ. Lemma 5.7 yields a non-trivial splitting of
Q over a group F 0 containing F with index ≤ 2 (with F 0 = F if there is no mirror). If
C is not used, this splitting is relative to IncH
|Q . By Lemma 4.12, one may use it to refine
T at v. One obtains a splitting of G relative to H, which may be collapsed to a one-edge
splitting over F 0 .
Proposition 5.4 implies:
Lemma 5.18. Let Q be a QH vertex group, with fiber F .
(1) Any non-trivial (minimal) splitting of Q relative to F factors through a splitting of
π1 (Σ) = Q/F .
(2) If the splitting is also relative to IncH
|Q , and every boundary component of Σ is used,
then the splitting of π1 (Σ) is dominated by a splitting dual to a family of geodesics.
In particular, Σ contains an essential simple closed geodesic.
(3) If, moreover, the splitting of Q has small edge groups, the splitting of π1 (Σ) is dual
to a family of geodesics.
Remark 5.19. The induced splitting of π1 (Σ) is relative to the boundary subgroups if (2)
holds, but not necessarily if only (1) holds.
Proof. Let S be the Bass-Serre tree of the splitting of Q. The group F acts as the identity
on S: its fixed point set is nonempty, and Q-invariant because F is normal in Q. We
deduce that the action of Q on S factors through an action of π1 (Σ).
Under the assumptions of (2), this action is relative to all boundary subgroups because
all boundary components of Σ are used (by an incident edge stabilizer or a conjugate of a
group in H). We apply Proposition 5.4.
Flexible QH vertex groups Gv of JSJ decompositions have non-trivial splittings relative
to IncH
v (see Corollary 4.14). As in Corollary 5.6 we wish to deduce that Σ contains an
essential simple closed geodesic, so as to rule out the small orbifolds of Proposition 5.12.
Proposition 5.20. Let Q = Gv be a QH vertex stabilizer of a JSJ tree over A relative to
H. Assume that F , and any subgroup of Gv containing F with index 2, belongs to A. Also
assume that F is universally elliptic. Then:
(1) every boundary component of Σ is used;
(2) if Q is flexible, then Σ contains an essential simple closed geodesic;
(3) let T be an (A, H)-tree such that Q acts on T with small edge stabilizers; if Q is not
elliptic in T , then the action of Q on its minimal subtree is dual, up to subdivision,
to a family of essential simple closed geodesics of Σ.
48
Proof. If some boundary component of Σ is not used, Lemma 5.16 yields a refinement of
the JSJ tree. The new edge stabilizers contain F with index at most 2, so belong to A and
are universally elliptic. This contradicts the maximality of the JSJ tree.
If Q = Gv is flexible, it acts non-trivially on an (A, H)-tree. This tree is also relative
to F because F is universally elliptic. This yields a splitting of Gv relative to F , which is
relative to H|Gv and to incident edge stabilizers because they are universally elliptic. Since
every boundary component is used, we obtain a geodesic by Lemma 5.18.
Applying the previous argument to T as in (3) shows that the splitting of Q is dual to
a family of geodesics by the third assertion of Lemma 5.18.
The first part of the following proposition shows that, conversely, the existence of an
essential geodesic implies flexibility.
Proposition 5.21. Let Q be a QH vertex group. Assume that Σ contains an essential
simple closed geodesic, and that, for all essential simple closed geodesics γ, the group Qγ
(see Definition 5.14) belongs to A.
(1) Any universally elliptic element (resp. subgroup) of Q is contained in an extended
boundary subgroup. In particular, Q is not universally elliptic.
(2) If J < Q is small in (A, H)-trees, its image in π1 (Σ) is virtually cyclic.
Remark 5.22. This holds under the weaker assumption that the set of essential simple
closed geodesics γ such that Qγ ∈ A fills Σ (in the sense of Definition 5.9).
Proof. Any subgroup of Q contained in the union of all extended boundary subgroups is
contained in a single extended boundary subgroup, so to prove (1) it suffices to show that,
if an element g does not lie in an extended boundary subgroup, then g is not universally
elliptic.
The image of g in π1 (Σ) has infinite order, so acts non-trivially in a splitting of π1 (Σ)
dual to a geodesic γ (see Definition 5.9). The splitting of G dual to γ is relative to H, and
the edge group Qγ is assumed to be in A, so g is not universally elliptic.
If J as in (2) has infinite image in π1 (Σ), it acts non-trivially in a splitting of G dual
to γ as above. The action preserves a line or an end by smallness of J, so the image of J
in π1 (Σ) is virtually cyclic by Remark 5.3.
Proposition 5.20 requires F to be universally elliptic. We show that this is automatic
for splittings over slender groups.
Lemma 5.23. Let Q be a QH vertex group. If F and all groups in A are slender, then F
is universally elliptic.
Proof. Suppose that F is slender but not universally elliptic. It acts non-trivially on some
tree, and there is a unique F -invariant line `. This line is Q-invariant because F is normal
in Q. It follows that Q is an extension of a group in A by its image in Isom(`), a virtually
cyclic group. If groups in A are slender, we deduce that Q is slender, a contradiction
because Q maps onto π1 (Σ) with Σ hyperbolic.
Corollary 5.24. Let Q be a QH vertex group Gv of a JSJ tree T . Assume that all groups
in A are slender, and that every extension of the fiber F by a virtually cyclic group belongs
to A. Then:
(1) F is universally elliptic; it is the largest slender normal subgroup of Q;
(2) all boundary components of the underlying orbifold Σ are used;
49
(3) if G acts on a tree S, and Q does not fix a point, then the action of Q on its minimal
subtree is dual to a family of essential simple closed geodesics of Σ;
(4) Q is flexible if and only if Σ contains an essential simple closed geodesic;
(5) if Q is flexible, than any universally elliptic subgroup of Q is an extended boundary
subgroup.
This follows from results proved above.
5.3
Quadratically hanging subgroups are elliptic in the JSJ
The goal of this subsection is to prove that, under suitable hypotheses, any QH vertex
group is elliptic in the JSJ deformation space (if we do not assume existence of the JSJ
deformation space, we obtain ellipticity in every universally elliptic tree).
We start with the following fact:
Lemma 5.25. If G splits over a group K ∈ A, but does not split over any infinite index
subgroup of K, then K is elliptic in the JSJ deformation space.
Proof. Apply Assertion 3 of Lemma 2.6 with T1 a JSJ tree and T2 a one-edge splitting
over K.
Remark 5.26. If K is not universally elliptic, it fixes a unique point in any JSJ tree. Also
note that being elliptic or universally elliptic is a commensurability invariant, so the same
conclusions hold for groups commensurable with K.
In [Sel97b, RS97, DS99] it is proved that, if Q is a QH vertex group in some splitting
(in the class considered), then Q is elliptic in the JSJ deformation space.
This is not true in general, even if A is the class of cyclic groups: Fn contains many
QH subgroups, none of them elliptic in the JSJ deformation space (which consists of free
actions, see Subsection 3.1). This happens because G = Fn splits over groups in A having
infinite index in each other, something which is prohibited by the hypotheses of the papers
mentioned above (in [FP06], G is allowed to split over a subgroup of infinite index in a
group in A, but Q has to be the enclosing group of minimal splittings, see [FP06, Definition
4.5 and Theorem 5.13(3)]).
A different counterexample will be given in Example 5.39 of Subsection 5.5, with A
the family of abelian groups. In that example the QH subgroup Q only has one abelian
splitting, which is universally elliptic, so Q is not elliptic in the JSJ space.
These examples explain the hypotheses in the following result.
Theorem 5.27. Let Q be a QH vertex group. Assume that, if Jˆ ⊂ Q is the preimage of
a virtually cyclic subgroup J ⊂ π1 (Σ), then Jˆ belongs to A and G does not split over a
ˆ If one of the following conditions holds, then Q is elliptic
subgroup of infinite index of J.
in the JSJ deformation space:
(1) Σ contains an essential simple closed geodesic γ;
(2) H = ∅, and the boundary of Σ is nonempty;
(3) the fiber F is elliptic in the JSJ deformation space;
(4) all groups in A are slender.
Note that Jˆ contains F , so F ∈ A.
50
Proof. Fix a JSJ tree TJ . We start by proving ellipticity of Q in TJ assuming the existence
of an essential simple closed geodesic γ.
Given a lift γ̃ ⊂ Σ̃, we have denoted by Hγ the subgroup of π1 (Σ) consisting of elements
which preserve γ̃ and each of the half-spaces bounded by γ̃, and by Qγ the preimage of Hγ
in Q. We now write Qγ̃ rather than Qγ because we want it to be well-defined, not just up
to conjugacy.
The group Qγ̃ belongs to A, and G splits over Qγ̃ . By Proposition 5.21, Qγ̃ is not
universally elliptic. By Lemma 5.25 and Remark 5.26, Qγ̃ fixes a unique point cγ̃ in TJ .
Lemma 5.28. Let γ̃, γ̃ 0 ⊂ Σ̃ \ ∂ Σ̃ be lifts of simple geodesics. If γ̃ and γ̃ 0 intersect, then
cγ̃ = cγ̃ 0 .
Proof. We can assume γ̃ 6= γ̃ 0 . Let T 0 be the Bass-Serre tree of the splitting of G determined
by γ̃ 0 . It contains a unique edge e0 with stabilizer Qγ̃ 0 . Since γ̃ and γ̃ 0 intersect, the group
Qγ̃ acts hyperbolically on T 0 , and its minimal subtree M (a line) contains e0 . Let T1 be a
refinement of TJ which dominates T 0 , as in Lemma 2.2, and let M1 ⊂ T1 be the minimal
subtree of Qγ̃ .
The image of M1 in TJ consists of the single point cγ̃ (because T1 is a refinement of
TJ ), and its image by any equivariant map f : T1 → T 0 contains M . Let e1 be an edge of
M1 such that f (e1 ) contains e0 . The stabilizer Ge1 of e1 is contained in Qγ̃ 0 , so Ge1 fixes
cγ̃ 0 . Since G does not split over infinite index subgroups of Qγ̃ 0 , the index of Ge1 in Qγ̃ 0 is
finite. The group Qγ̃ 0 is not universally elliptic (it acts non-trivially in the splitting dual to
γ), so cγ̃ 0 is the unique fixed point of Ge1 in TJ by Remark 5.26. But Ge1 fixes cγ̃ because
e1 is mapped to cγ̃ in TJ , so cγ̃ = cγ̃ 0 .
We can now conclude. Since the set of all essential simple geodesics of Σ fills Σ (Corollary 5.10), the union of their lifts is a connected subset of Σ̃ (Definition 5.9). In particular,
given any pair γ̃, γ̃ 0 of such lifts, there exists a finite sequence γ̃ = γ̃0 , γ̃1 , . . . , γ̃p = γ̃ 0 where
γ̃i intersects γ̃i+1 . Lemma 5.28 implies that cγ̃ = cγ̃ 0 , so cγ̃ does not depend on γ̃. It is
fixed by Q, so Q is elliptic.
We have proved the theorem in case (1) (when Σ contains a geodesic). We reduce cases
(2) and (3) to this one, using Lemma 5.18 to find a geodesic (case (4) reduces to (3) by
Lemma 5.23).
We first show that every boundary component of Σ is used (see Definition 5.15). If not,
Lemma 5.16 yields a splitting of G over a group F 0 containing F with index 1 or 2. By
Remark 5.8, the group F 0 is contained in the preimage Jˆ of a 2-ended subgroup J ⊂ π1 (Σ),
ˆ
and this contradicts the assumptions of the theorem since F 0 has infinite index in J.
If H = ∅ and ∂Σ 6= ∅, every boundary component C is used by an incident edge
stabilizer Ge whose image in π1 (Σ) is contained in B = π1 (C) with finite index. Let
B̂ ⊂ Q be the preimage of B. Folding yields a non-trivial splitting of G over B̂ (if for
instance G = R ∗Ge Q, then G = (R ∗Ge B̂) ∗B̂ Q). Since B̂ ∈ A, Lemma 5.25 implies that
B̂ is elliptic in the JSJ deformation space; in particular, F is elliptic. If Q is not elliptic,
Lemma 5.18 implies that Σ contains a geodesic.
Now suppose that F is elliptic in the JSJ space, but Q is not. The action of Q on
any JSJ tree then factors through π1 (Σ) by Assertion (1) of Lemma 5.18. Every boundary
subgroup B = π1 (C) is elliptic: this follows from the previous argument if C is used by
an incident edge stabilizer, and holds if C is used by a group in H because splittings are
relative to H. Proposition 5.4 yields a geodesic.
Remark 5.29. Under the assumptions of Theorem 5.27, assume moreover that all groups
in A are slender, and that G does not split over a subgroup of Q whose image in π1 (Σ) is
finite. Then Q fixes a unique point vJ ∈ TJ , so Q ⊂ GvJ . We claim that, if GvJ (hence
also Q) is universally elliptic, then vJ is a QH vertex of TJ . This is used in [GL10] (proof
of Theorem 4.2).
51
Let T be a tree in which Q is a QH vertex group Gw . Note that Q = GvJ because GvJ
is elliptic in T . We have to show that, if e is an edge of TJ containing v, then Ge is an
extended boundary subgroup of Q. Let T̂ be a refinement of TJ which dominates T . Let
ŵ be the unique point of T̂ fixed by Q, and let f : T̂ → T be an equivariant map. Let ê
be the lift of e to T̂ .
If f (ê) 6= {w}, then Ge fixes an edge of T adjacent to w, so is an extended boundary
subgroup of Q. Otherwise, consider a segment xŵ, with f (x) 6= w, which contains ê.
Choose such a segment of minimal length, and let ε = xy 6= ê be its initial edge (so
f (y) = w). We have Gε ⊂ Gy ⊂ Gw = Q, and Gε fixes an edge of T adjacent to w.
Since G does not split over groups mapping to finite groups in π1 (Σ), the image of Gε in
π1 (Σ) is a finite index subgroup of a boundary subgroup B ⊂ π1 (Σ). But we also have
Gε ⊂ Gw = Gŵ , so that Gε ⊂ Gê = Ge . Being slender and containing a finite index
subgroup of B, the image of Ge in π1 (Σ) has to be contained in B.
Corollary 5.30. Let A = V P C≤n be the class of all virtually polycyclic groups of Hirsch
length ≤ n. Assume that G does not split over a group in V P C≤n−1 . Let Q be a QH vertex
group with fiber in V P C≤n−1 , in some splitting of G over V P C≤n .
If T is any JSJ tree of G over A, then Q is contained in a QH vertex stabilizer of T .
Proposition 5 of the introduction is the case n = 1.
Proof. The group Q fixes a point v in T by Theorem 5.27. If Gv is universally elliptic,
we apply Remark 5.29. If not, v is flexible and we will prove in Theorem 6.2 that Gv is
QH.
5.4
Peripheral structure of quadratically hanging vertices
Suppose that Q is a QH vertex group of a JSJ decomposition. The incident edge groups
are extended boundary subgroups (see Definition 5.13). However, the collection of incident
edge groups (the family Incv of Definition 4.3) may change when the JSJ tree T varies in
the JSJ deformation space (though the collection of extended boundary subgroups usually
does not change, see Assertion (6) of Theorem 6.2).
In Section 4 of [GL07a] we introduced a collection M0 of subgroups of Q, which is
related to the incident edge groups, but which does not depend on the tree in the JSJ
deformation space. We called this collection the peripheral structure of Q.
The goal of this subsection is to show that the peripheral structure contains more
information than the collection of all extended subgroups, and may be fairly arbitrary.
In the examples, A is the class of slender groups, and no slender group is conjugate to
a proper subgroup of itself. In this context, the peripheral structure M0 of a non-slender
vertex stabilizer Gv of a tree T may be determined as follows. We first collapse edges so
as to make T reduced (see Subsection 1.4.2). The image of v has the same stabilizer, and
we still denote it by v. Then M0 is the set of conjugacy classes of incident edge stabilizers
Ge ⊂ Gv that are not properly contained in another incident edge stabilizer of Gv (see
[GL07a] for details).
We have seen (Proposition 5.20) that M0 often uses every boundary component of Σ.
Apart from that, the peripheral structure of Q may be fairly arbitrary. We shall now give
examples. In particular, Example 5.34 will show the following proposition:
Proposition 5.31. If Q is a QH vertex group of a slender JSJ decomposition of a oneended group, it is possible for an incident edge group to meet F trivially, or to have trivial
image in π1 (Σ).
We work with H = ∅ and A the family of slender subgroups.
52
H1 × R 1
H1
H2 × R 2 H
2
H3
H3 × R3 ..
.
Hp × R p H p
Figure 7: An example of a QH group in a JSJ decomposition.
To give examples, we use the filling construction described in Subsection 4.4, see Figure
7. We start with an extension 1 → F → Q → π1 (Σ) → 1, with F slender and Σ a
compact orientable surface (with genus ≥ 1, or with at least 4 boundary components). Let
H1 , . . . , Hp be a finite family of infinite extended boundary subgroups of Q as defined in
Definition 5.13 (note that they are slender). We impose that, for each boundary subgroup
B of π1 (Σ), there is an i such that Hi maps onto a finite index subgroup of B (i.e. every
boundary component is used by some Hi in the sense of Definition 5.15). Let Ri be a nonslender finitely presented group with Serre’s property (FA) (it has no non-trivial action on
a tree), for instance SL(3, Z). As in Subsection 4.4, we define a finitely presented group Ĝ
by amalgamating Q with Ki = Hi × Ri over Hi for each i.
Lemma 5.32. The Bass-Serre tree T of the amalgam defining Ĝ is a slender JSJ tree, Q
is a flexible QH subgroup, and Ĝ is one-ended. If no Hi is conjugate in Q to a subgroup of
Hj for i 6= j, the peripheral structure M0 consists of the conjugacy classes of the Hi ’s.
Proof. Let T 0 be any tree. Each Ri fixes a unique point because it is not slender, and this
point is also fixed by Hi . In particular, Hi × Ri , Hi and T are universally elliptic. To prove
that T is a JSJ tree, it suffices to see that Q is elliptic in any universally elliptic tree T 0 .
By Lemma 5.23, F is universally elliptic. If Q is not elliptic in T 0 , then by Lemma
5.18 the action of Q on its minimal subtree TQ ⊂ T 0 factors through a nontrivial action of
π1 (Σ) with slender (hence cyclic) edge stabilizers. Since every Hi , hence every boundary
subgroup of π1 (Σ), is elliptic, this action is dual to a system of disjoint geodesics on Σ
by Proposition 5.4. By Proposition 5.21, no edge stabilizer of TQ is universally elliptic,
contradicting universal ellipticity of T 0 .
This shows that T is a JSJ tree, and Q is flexible because Σ was chosen to contain
intersecting simple closed curves (see Corollary 5.24).
By Proposition 4.1, one obtains a JSJ tree of Ĝ over finite groups by collapsing all
edges of T with infinite stabilizer. Since each Hi is infinite, this JSJ tree is trivial, so Ĝ is
one-ended.
The assertion about M0 follows from the definition of M0 given in [GL07a].
Example 5.33. Let Σ be a punctured torus, with fundamental group ha, bi. Write u = [a, b].
Let Q = F × ha, bi, with F finite and non-trivial. Let H1 = hF, u2 i and H2 = hui. The
peripheral structure of Q in the JSJ tree T consists of two elements, though Σ only has one
boundary component. There is a JSJ tree T 0 such that incident edge groups are conjugate
to hF, ui (the quotient T 0 /G is a tripod), but it does not display the peripheral structure
of Q.
Example 5.34. Let Σ, a, b, u be as above. Again write Q = F × ha, bi, but now F = hti is
infinite cyclic. Let H1 = hui and H2 = hti. Then H1 meets F trivially, while H2 maps
trivially to π1 (Σ).
53
Example 5.35. Assume that the orbifold Σ has a conical point x carrying a finite cyclic
group Fx and that the fiber F is infinite. Then one can attach an edge to x: one chooses
any infinite subgroup Hx of the preimage of Fx in Q, and one constructs an amalgam
Q ∗Hx (Hx × Rx ). Similar constructions are possible with x a corner reflector, a point on
a mirror, or even a point with trivial isotropy.
5.5
Flexible vertices of abelian JSJ decompositions
We shall see in Section 6 that, if A is the family of cyclic subgroups, or virtually cyclic
subgroups, or slender subgroups, then non-slender flexible vertex groups Q of JSJ decompositions over A are QH with slender fiber (we say that Q is slender-by-orbifold). Things
are more complicated when A is the family of abelian subgroups (or equivalently of finitely
generated abelian subgroups, see Proposition 4.2).
The basic reason is the following: if a group Q is an extension 1 → F → Q → π1 (Σ) → 1
with F a finitely generated abelian group and Σ a surface, a splitting of π1 (Σ) dual to a
simple closed curve induces a splitting of Q over a subgroup which is slender (indeed
polycyclic) but not necessarily abelian.
Using the terminology of Definition 6.23, the regular neighbourhood of two abelian
splittings is not necessarily an abelian splitting.
In fact, we shall now construct examples showing:
Proposition 5.36. (1) Flexible subgroups of abelian JSJ trees are not always slenderby-orbifold groups.
(2) One cannot always obtain an abelian JSJ tree by collapsing edges of a slender JSJ
tree.
By Proposition 4.1, one can obtain an abelian JSJ tree by refining and collapsing a
slender JSJ tree. The point here is that collapsing alone is not always sufficient. It may
be shown that collapsing suffices when G is torsion-free, finitely presented, and CSA (see
Proposition 8.12 of [GL09]).
We use the same construction as in the previous subsection, but now π1 (Σ) will act
non-trivially on the fiber F .
Example 5.37. In this example F ' Z. Let Σ be obtained by gluing a once-punctured torus
to one of the boundary components of a pair of pants. Let M be a circle bundle over Σ
which is trivial over the punctured torus but non-trivial over the two boundary components
of Σ. Let Q = π1 (M ), and let H1 , H2 be the fundamental groups of the components of
∂M (homeomorphic to Klein bottles). Note that H1 , H2 are non-abelian. Construct Ĝ by
amalgamation with Hi × Ri as above. We claim that the abelian JSJ decomposition of Ĝ
is trivial, and Ĝ is flexible (but not slender-by-orbifold).
We argue as in the proof of Lemma 5.32. We know that H1 × R1 and H2 × R2 (hence
also F ) are universally elliptic. If T is any tree with abelian edge stabilizers, the action
of Q on its minimal subtree factors through π1 (Σ), and the action of π1 (Σ) is dual to a
system of simple closed curves. But not all simple closed curves give rise to an abelian
splitting of Q: they have to be “positive”, in the sense that the bundle is trivial over them.
To prove that none of these splittings is universally elliptic, hence that the abelian JSJ
space of Ĝ is trivial and Ĝ is flexible, it suffices to see that any positive curve intersects
(in an essential way) some other positive curve. This is true for the curve δ separating the
pair of pants from the punctured torus (one easily constructs a positive curve meeting δ
in 4 points). It is also true for curves meeting δ. Curves disjoint from δ are contained in
the punctured torus, and the result is true for them.
Every map from SL(3, Z) to a 2-orbifold group has finite image, so Ĝ is not isomorphic
to a slender-by-orbifold group if we choose Ri = SL(3, Z).
54
Remark 5.38. If one performs the construction adding a third group H3 = F , then Ĝ
becomes a flexible vertex group in a group whose JSJ decomposition is non-trivial.
Example 5.39. Now F = Z2 . Let Σ be a surface of genus ≥ 2 with two boundary components C1 , C2 . Let γ be a simple closed curve separating C1 from C2 . Let Σ0 be the
space obtained from Σ by collapsing γ to a point. Map π1 (Σ) to SL(2, Z) ⊂ Aut(Z2 ) by
projecting to π1 (Σ0 ) and embedding the free group π1 (Σ0 ) into SL(2, Z). Let Q be the
associated semi-direct product Z2 o π1 (Σ), and Hi = Z2 o π1 (Ci ). Construct Ĝ as before.
Abelian splittings of Ĝ now come from simple closed curves on Σ belonging to the
kernel of ρ : π1 (Σ) → SL(2, Z). But it is easy to see that γ is the only such curve. It
follows that the one-edge splitting dual to γ is an abelian JSJ decomposition of Ĝ. It has
two rigid vertex groups. It cannot be obtained by collapsing a slender JSJ splitting.
6
JSJ decompositions over slender groups
The main result of this section is the description of JSJ decompositions over slender groups.
Recall (Subsection 1.5) that a subgroup A ⊂ G is slender if A and all its subgroups are
finitely generated. Whenever G acts on a tree, A fixes a point or leaves a line invariant.
Our approach essentially follows Fujiwara and Paposoglu [FP06], but with simplifications. In particular, we do not have to “deal with a third splitting” (see below for further
discussion).
6.1
Statement of results
Let G be a finitely generated group, A a family of subgroups stable under conjugation and
taking subgroups, and H a finite set of finitely generated subgroups of G such that G is
finitely presented relative to H.
The goal of this section is to show that non-slender flexible vertex groups Q of JSJ
decompositions over A relative to H are QH (see Subsection 5.2). We need two assumptions on A. First, groups in A should be slender (or at least slender in (A, H)-trees, see
Subsection 6.8). The second is a stability condition involving a subfamily F ⊂ A (we will
show that fibers of QH flexible vertex groups belong to F).
Definition 6.1 (Stability Condition (SC)). We say that A satisfies the stability condition
(SC), with fibers in a family of subgroups F, if the following hold for every short exact
sequence
1→F →A→K→1
with A < G:
(1) if A ∈ A, and K is isomorphic to Z or D∞ , then F ∈ F;
(2) if F ∈ F and K is isomorphic to a quotient of Z or D∞ , then A ∈ A.
The group Z acts on the line in an orientation-preserving way. The infinite dihedral
group D∞ also acts on the line, but orientation is not preserved. In (2), the group K may
be finite (cyclic or dihedral) or infinite (isomorphic to Z or D∞ ).
If Q is a QH subgroup with fiber F , recall that any simple geodesic γ on Σ defines
a splitting of G over a group Qγ which is an extension of F by Z or D∞ . If there is a
geodesic γ0 such that Qγ0 ∈ A, the stability condition ensures first that F ∈ F, and then
that Qγ ∈ A for every γ (compare the assumptions of Proposition 5.21, Corollary 5.24 and
Theorem 5.27). Failure of the stability condition explains why Theorem 6.2 below does
not apply to abelian JSJ splittings (see Subsection 5.5).
On the other hand, one easily checks that the stability condition holds in the following
cases:
55
• A={slender} (i.e. A consists of all slender subgroups of G), with F = A;
• A={virtually cyclic}, with F={finite};
• A={virtually polycyclic}, with F = A;
• A = {V P C≤n }, the virtually polycyclic subgroups of Hirsch length at most n, with
F={VPC≤n−1 };
• G is a torsion-free CSA group, A={finitely generated abelian}, with F = A (recall
that a group is CSA if all maximal abelian subgroups are malnormal).
See Definition 6.4 for a different condition, which applies to A={cyclic}. Also recall that,
since G is relatively finitely presented, a JSJ decomposition over A is also one over groups
locally in A (Proposition 4.2).
Theorem 6.2. Suppose that all groups in A are slender, and A satisfies the stability
condition (SC), with fibers in a family F. Let H be a finite family of finitely generated
subgroups. Let G be finitely presented (or only finitely presented relative to H).
If Q is a non-slender flexible vertex group of a JSJ decomposition of G over A relative
to H, then:
(1) Q is QH with fiber in F (it maps onto π1 (Σ), where Σ is a compact hyperbolic 2orbifold, with kernel F ∈ F; the image of an incident edge group in π1 (Σ) is finite
or contained in a boundary subgroup);
(2) if G acts on a tree and Q does not fix a point, the action of Q on its minimal subtree
is dual to a family of geodesics of Σ;
(3) F and extended boundary subgroups are universally elliptic;
(4) every boundary component of Σ is used;
(5) Σ contains an essential closed geodesic;
(6) every universally elliptic subgroup of Q is an extended boundary subgroup.
Once (1) is known, Assertions (2)-(6) are direct consequences of the results of Subsection
5.2. See Subsection 5.1.3 for the list of orbifolds containing no essential closed geodesic,
and Subsection 6.9 for a description of slender flexible vertex groups.
Corollary 6.3. Let A be the class of all slender subgroups of G. Let H be a finite family of
finitely generated subgroups. Let G be finitely presented (or only finitely presented relative
to H).
Then every flexible vertex of any JSJ decomposition over A relative to H is either
slender or QH with slender fiber.
In view of the examples given above, one has a similar description of JSJ decompositions
over virtually cyclic groups (non-slender flexible groups are QH with finite fiber), over
V P C≤n -groups (non-slender flexible groups are QH with V P C≤n−1 fiber), etc.
The result by Fujiwara-Papasoglu [FP06] is the case when H = ∅ and A is the class of
all slender groups. Dunwoody-Sageev [DS99] consider (a generalisation of) the case when
H = ∅ and A = {V P C≤n }; if G does not split over a V P C≤n−1 -subgroup, flexible vertex
groups are QH with V P Cn−1 fiber.
If groups in A are only assumed to be slender in (A, H)-trees, the conclusions of the
theorem apply to groups Q which are not slender in (A, H)-trees (see Subsection 6.8).
The class of all (finite or infinite) cyclic groups does not satisfy the stability condition (SC) if G contains dihedral subgroups. To recover Rips and Sela’s description of
JSJ decompositions over cyclic groups, we introduce a modified stability condition (SCZ )
preventing groups in A from acting dihedrally on a line.
56
Definition 6.4 (Stability Condition (SCZ )). We say that A satisfies the stability condition
(SCZ ), with fibers in F, if the following hold:
(1) no group of A maps onto D∞ ; given a short exact sequence 1 → F → A → Z → 1
with A ∈ A, we have F ∈ F;
(2) given a short exact sequence 1 → F → A → K → 1 with F ∈ F and K isomorphic
to a quotient of Z, we have A ∈ A.
This condition is satisfied by the classes consisting of
• all cyclic subgroups, with F = {1},
• all subgroups which are finite or cyclic, with F = {1},
• all virtually cyclic subgroups which do not map onto D∞ (i.e. are finite or have
infinite center), with F ={finite},
and any class satisfying (SC) and consisting of groups which do not map onto D∞ .
Theorem 6.5. Theorem 6.2 holds if A satisfies the stability condition (SCZ ) rather than
(SC). In this case the underlying orbifold Σ has no mirror (all singular points are conical
points).
In particular:
Theorem 6.6 ([RS97]). Let G be finitely presented relative to a finite family H of finitely
generated subgroups. Let A be the class of all finite or cyclic subgroups of G.
If Q is a flexible vertex group of a JSJ decomposition of G over A relative to H, and
Q is not virtually Z2 , then Q is QH with trivial fiber. Moreover, the underlying orbifold Σ
has no mirror, every boundary component of Σ is used, and Σ contains an essential simple
closed geodesic.
Indeed, it follows from (RN1) of Proposition 6.25, or from Proposition 6.38, that slender
flexible groups are virtually Z2 .
In the next subsection we shall reduce Theorems 6.2 and 6.5 to Theorem 6.8, which
will be proved in Subsections 6.3 through 6.7.
In Subsections 6.3 through 6.5, and 6.7, we only assume that G is finitely generated (not
finitely presented), and H may be arbitrary (we use finite presentability only in Lemma
6.12 and Subsection 6.6). This will be useful in Subsection 8.1, where finite presentability
will not be assumed. In particular, we will see (Theorem 8.6) that Theorem 6.6 is true if
G is only finitely generated and H is an arbitrary family of subgroups.
6.2
Reduction to totally flexible groups
Unlike Fujiwara-Papasoglu, we know in advance that JSJ decompositions exist, and we
only need to show that flexible vertex groups Gv are QH. This allows us to forget G and
concentrate on Gv , but we have to remember the incident edge groups and we therefore
consider splittings of Gv that are relative to IncH
v (see Definition 4.3): since incident edge
groups are universally elliptic, these are exactly the splittings that extend to splittings
of G relative to H (see Lemma 4.12). Even if we are interested only in non-relative JSJ
decompositions of G (H = ∅), it is important here that we work in a relative context. In
Subsection 8.1.2 we will even have to allow groups in H to be infinitely generated.
The fact that Gv is a flexible vertex of a JSJ decomposition says that Gv splits relative
to IncH
v , but not over a universally elliptic subgroup (Corollary 4.14). This motivates the
following general definition.
57
Definition 6.7 (Totally flexible). G is totally flexible (over A relative to H) if it admits
a non-trivial splitting, but none over a universally elliptic subgroup. Equivalently, the JSJ
decomposition of G is trivial, and G is flexible.
The example to have in mind is the fundamental group of a compact hyperbolic surface
other than a pair of pants, with A the class of cyclic groups and H consisting of the
fundamental groups of the boundary components.
We shall deduce Theorems 6.2 and 6.5 from the following result, which says that totally
flexible groups are QH.
Theorem 6.8. Let G be finitely presented relative to a finite family H of finitely generated
subgroups. Let A be a class of slender groups satisfying the stability condition (SC) or
(SCZ ) with fibers in F (see Definitions 6.1 and 6.4).
Assume that G is totally flexible over A relative to H, and not slender. Then G is QH
with fiber in F, and Σ has no mirror in the SCZ case.
Since there are no incident edge groups, being QH means that G is an extension of a
group F ∈ F by the fundamental group of a hyperbolic orbifold Σ, and the image of each
group H ∈ H in π1 (Σ) is either finite or contained in a boundary subgroup. By Corollary
5.24, every component of ∂Σ is used by a group of H (if H = ∅, then Σ is a closed orbifold).
This theorem implies Theorems 6.2 and 6.5: we apply it to Gv , with A = Av (the
H
family of subgroups of Gv belonging to A) and H = IncH
v (Definition 4.3). Note that Incv
is a finite family of finitely generated subgroups by Theorem 2.20 and Remark 4.6, that
Gv is finitely presented relative to IncH
v by Proposition 4.9, and that Av satisfies (SC) or
(SCZ ) with fibers in Fv .
There are three main steps in the proof of Theorem 6.8.
• Given two trees T1 , T2 such than no edge stabilizer of one tree is elliptic in the other,
we follow Fujiwara-Papasoglu’s construction of a core C ⊂ T1 × T2 . This core happens to
be a surface away from its vertices. More precisely, if one removes from C its cut vertices
and the vertices whose link is homeomorphic to a line, one gets a surface whose connected
components are simply connected. The decomposition of C dual to its cut points yields
a tree R having QH vertices coming from the surface components. Borrowing Scott and
Swarup’s terminology [SS03], we call this tree R the regular neighborhood of T1 and T2 (see
Example 6.24 for an explanation of this name). The stability condition is used to ensure
that edge stabilizers of R are in A.
• Given a totally flexible G, we construct two splittings U and V of G which “fill” G,
and we show that their regular neighborhood is the trivial splitting of G. We deduce that
G itself is QH, as required (in the case of cyclic splittings of a torsion-free group, G is
the fundamental group of a compact surface and one should think of U and V as dual to
transverse pair of pants decompositions, see Example 6.29).
• The previous steps require that splittings of Q be minuscule. This is a condition
which controls the way in which Q may split over subgroups A, A0 with A0 ⊂ A, and we
check that it is satisfied.
Our first step is the same as [FP06]. But total flexibility allows us to avoid the more
complicated part of their paper where, given a QH vertex group of a tree (obtained for
instance as the regular neighborhood of two splittings), one has to make it larger to enclose
a third splitting.
In the second step, the first splitting U is obtained by a maximality argument which
requires finite presentability.
Fujiwara and Papasoglu work with a minimality condition for splittings that allows the
construction of regular neighborhoods. This condition is not sufficient for our purpose,
and we replace it by a stronger condition (minuscule).
58
6.3
Fujiwara-Papasoglu’s core
As mentioned earlier, we let H be arbitrary and we only assume that G is finitely generated.
Groups in A are slender, and one of the stability conditions (SC) or (SCZ ) is satisfied.
6.3.1
Minuscule splittings
Definition 6.9 (Minuscule). Given A, A0 ∈ A, we say that A0 A if A0 ⊂ A and there
exists a tree S such that A0 is elliptic in S but A is not. We also write A A0 .
We say that A ∈ A is minuscule if, whenever a subgroup A0 A fixes an edge e of a
tree, A0 has infinite index in the stabilizer Ge . Equivalently, A is minuscule if and only if
no A0 A is commensurable with an edge stabilizer.
A tree is minuscule if its edge stabilizers are minuscule. We say that all trees (or all
splittings) are minuscule if all (A, H)-trees are minuscule.
If A0 A, then A0 has infinite index in A. If A0 ⊂ A B ⊂ B 0 , then A0 B 0 . In
particular, the relation is transitive.
Being minuscule is a commensurability invariant. It is often used in the following way:
if an edge stabilizer Ge of a tree T is elliptic in another tree S, then any minuscule A
containing Ge is also elliptic in S.
Remark 6.10. If G does not split (relative to H) over a subgroup commensurable with a
subgroup of infinite index of a group in A, then every A ∈ A is minuscule (because, if
A is not minuscule, some Ge has a finite index subgroup A0 contained in A with infinite
index). This holds for instance under the assumptions of [DS99], in particular for splittings
of one-ended groups over virtually cyclic subgroups. The reader interested only in this case
may therefore ignore Subsection 6.7, where we prove that all splittings of a totally flexible
group are minuscule.
The following lemma says that ellipticity is a symmetric relation among minuscule trees
(in the terminology of [FP06], minuscule one-edge splittings are minimal ).
Lemma 6.11. If T1 is elliptic with respect to T2 , and T2 is minuscule, then T2 is elliptic
with respect to T1 .
Proof. Let T̂1 be a standard refinement of T1 dominating T2 (Definition 2.5). Let e2 be an
edge of T2 , and e an edge of T̂1 with Ge ⊂ Ge2 . The group Ge is elliptic in T1 , and since
Ge2 is minuscule Ge2 itself is elliptic in T1 . This argument applies to any edge of T2 , so T2
is elliptic with respect to T1 .
Lemma 6.12. Assume that H is a finite family of finitely generated subgroups, G is finitely
presented relative to H, and all trees are minuscule. There exists a tree T which is maximal
for domination: if T 0 dominates T , then T dominates T 0 (so T and T 0 are in the same
deformation space).
Finite presentation is necessary as evidenced by Dunwoody’s inaccessible group [Dun93],
with H = ∅ and A the family of finite subgroups.
Example 6.13. If we consider cyclic splittings of the fundamental group of a closed orientable surface Σ, then T is maximal if and only if it is dual to a pair of pants decomposition
of Σ.
Proof. Let T be the set of all trees with finitely generated edge and vertex stabilizers, up
to equivariant isomorphism. By Dunwoody’s accessibility (Corollary 2.25) every tree is
dominated by a tree in T , so it suffices to find a maximal element in T .
As pointed out in the proof of Theorem 2.16, the set T is countable, so it suffices
to show that, given any sequence Tk such that Tk+1 dominates Tk , there exists a tree T
dominating every Tk .
59
We produce inductively a tree Sk in the same deformation space as Tk which refines
Sk−1 . Start with S0 = T0 , and assume that Sk−1 is already defined. Since Tk dominates
Tk−1 , hence Sk−1 , it is elliptic with respect to Sk−1 . All trees being assumed to be minuscule, Sk−1 is elliptic with respect to Tk by Lemma 6.11. We may therefore define Sk as a
standard refinement of Sk−1 dominating Tk . It belongs to the same deformation space as
Tk by Assertion (iv) of Proposition 2.2.
By Dunwoody’s accessibility (Proposition 2.24), there exists a tree T ∈ T dominating
every Sk , hence every Tk .
6.3.2
Definition of the core
Let T1 and T2 be trees. Recall that T1 is elliptic with respect to T2 if every edge stabilizer
of T1 is elliptic in T2 (Definition 2.1). If T1 has several orbits of edges, it may happen that
certain edge stabilizers are elliptic in T2 and others are not (being slender, they act on T2
hyperbolically, leaving a line invariant). This motivates the following definition.
Definition 6.14 (Fully hyperbolic). T1 is fully hyperbolic with respect to T2 if every edge
stabilizer of T1 acts hyperbolically on T2 .
This implies that edge stabilizers of T1 are infinite, and no vertex stabilizer of T1 is
elliptic in T2 (except if T1 and T2 are both trivial).
Example 6.15. Suppose that G is a surface group and T1 , T2 are dual to families of disjoint
geodesics L1 , L2 as in Subsection 5.1. Then T1 is elliptic with respect to T2 if each curve
in L1 is either contained in L2 or disjoint from L2 . It is fully hyperbolic with respect to
T2 if each curve in L1 meets L2 and every intersection is transverse.
Let T1 be fully hyperbolic with respect to T2 . We consider the product T1 × T2 . We
view it as a complex made of squares e1 × e2 , with the diagonal action of G. An edge of
the form e1 × {v2 } is horizontal, and an edge {v1 } × e2 is vertical.
Following [FP06], we define the asymmetric core C(T1 , T2 ) as follows. For each edge e
and each vertex v of T1 , let µT2 (Ge ) and µT2 (Gv ) be the minimal subtrees of Ge and Gv
respectively in T2 (with µT2 (Ge ) a line since Ge is slender).
ˇ 1 , T2 )). Let T1 be fully hyperbolic with
Definition 6.16 (Asymmetric cores C(T1 , T2 ), C(T
respect to T2 . The asymmetric core C(T1 , T2 ) ⊂ T1 × T2 is
[
[
C(T1 , T2 ) =
{v} × µT2 (Gv ) ∪
e × µT2 (Ge ) .
v∈V (T1 )
e∈E(T1 )
ˇ 1 , T2 ) ⊂
If we also assume that T2 is fully hyperbolic with respect to T1 , we denote by C(T
T1 × T2 the opposite construction:
[
[
ˇ 1 , T2 ) =
C(T
µT1 (Gv ) × {v} ∪
µT1 (Ge ) × e .
v∈V (T2 )
e∈E(T2 )
ˇ 1 , T2 ).
When no confusion is possible, we use the notations C, Cˇ instead of C(T1 , T2 ), C(T
Note that C consists of all (x, y) ∈ T1 × T2 such that y belongs to the minimal subtree of
Gx .
Every µT2 (Gv ), µT2 (Ge ) being a non-empty subtree, with µT2 (Ge ) ⊂ µT2 (Gv ) if v is an
endpoint of e, a standard argument shows that C(T1 , T2 ) is simply connected.
The group G acts diagonally on T1 ×T2 , and C is G-invariant. Since Ge acts cocompactly
on the line µT2 (Ge ) for e ∈ E(T1 ), there are finitely many G-orbits of squares in C.
Remark 6.17. If H ⊂ G is elliptic in T1 and T2 , it fixes a vertex in C (because H fixes some
v1 ∈ T1 , and being elliptic in T2 it fixes a point in the Gv1 -invariant subtree µT2 (Gv1 ) ⊂ T2 ).
60
6.3.3
Symmetry of the core
The goal of this subsection is to prove that the asymmetric core of minuscule splittings is
actually symmetric: Cˇ = C. Before doing so, we note the following basic consequence of
symmetry.
Lemma 6.18. Let T1 , T2 be fully hyperbolic with respect to each other. Assume that Cˇ = C.
Then C is a pseudo-surface: if V (C) is its set of vertices, then C \ V (C) is a surface (which
does not have to be connected or simply connected).
Proof. By construction, C ∩ (˚
e1 × T2 ) = e˚1 × µT2 (Ge1 ) ' e˚1 × R if e1 is an edge of T1 and
e˚1 denotes the open edge. It follows that all horizontal edges of C are contained in exactly
two squares. The symmetric argument shows that all vertical edges of Cˇ are contained in
ˇ Since Cˇ = C, all open edges of C are contained in exactly two
exactly two squares of C.
squares, so C \ V (C) is a surface.
Proposition 6.19. Let T1 , T2 be two minuscule trees that are fully hyperbolic with respect
to each other. Then Cˇ = C.
In other words, the relation “y belongs to the minimal subtree of Gx ” is symmetric.
The remainder of this subsection is devoted to the proof of the proposition. We always
assume that T1 and T2 are fully hyperbolic with respect to each other, but they are only
assumed to be minuscule when indicated.
We denote by C (2) ⊂ C the union of all closed squares e1 × e2 of T1 × T2 which are
contained in C, or equivalently such that e2 ⊂ µT2 (Ge1 ). It contains all horizontal edges
e × {v2 } of C, but only contains those vertical edges {v} × e2 which bound a square in C.
Any open edge {v} × e̊2 which does not bound a square in C disconnects C.
We define Cˇ(2) analogously.
Remark. The minimality condition of [FP06] is weaker than requiring trees to be minuscule,
and they only conclude Cˇ(2) = C (2) .
(2)
If y is a point of T2 , define Yy = C ∩ (T1 × {y}) and Yy = C (2) ∩ (T1 × {y}). They
are invariant under Gy (we do not claim that they are connected). If y is not a vertex,
(2)
(2)
Yy \ Yy is a union of isolated vertices, so any connected component of Yy is a connected
component of Yy .
Lemma 6.20. Let m be the midpoint of an edge e2 of T2 . If Ym is not connected, then
(2)
T2 is not minuscule. Moreover, if Ym is not connected, then some edge stabilizer of T1 is
hyperbolic in a one-edge splitting over a group A Ge2 .
The “moreover” will only be needed in Subsection 6.7.2.
Proof. Assuming that Ym is not connected, let Z be a connected component that does not
contain µT1 (Ge2 ) × {m} (there exists one since µT1 (Ge2 ) × {m} is connected; we do not
(2)
(2)
claim µT1 (Ge2 ) × {m} ⊂ C). If Ym is disconnected, we choose Z ⊂ Ym . Let GZ ⊂ Ge2
be the global stabilizer of Z.
We first show that GZ is elliptic in T1 , so in particular, GZ Ge2 . If GZ contains an
element h which is hyperbolic in T1 , the projection of Z in T1 contains the axis µT1 (h).
Since Ge2 is slender, µT1 (h) = µT1 (Ge2 ), contradicting our choice of Z.
To prove that T2 is not minuscule, it is therefore enough to construct a (minimal) tree S
in which GZ is an edge stabilizer. Consider Z̃ = G.Z ⊂ C, and note that Z is a connected
component of Z̃. Since C is simply connected, each connected component of Z̃ separates C
(Z is a track in C). Let S be the tree dual to Z̃: its vertices are the connected components
of C \ Z̃, and its edges are the connected components of Z̃.
61
By construction, S is a one-edge splitting, GZ is the stabilizer of the edge corresponding
to Z, so in particular S is an A-tree. Since any H ∈ H fixes a vertex in C (see Remark
6.17), S is an (A, H)-tree.
There remains to check that S is non-trivial (hence minimal). Let {v1 } × {m} be a
vertex of Z. By definition of C, the point m belongs to µT2 (Gv1 ), so there is g ∈ Gv1
whose axis in T2 contains m. The line {v1 } × µT2 (g) ⊂ C naturally defines an embedded
g-invariant line in S, on which g acts as a non-trivial translation. This proves that g is
hyperbolic in S, so S is non-trivial. It follows that T2 is not minuscule.
(2)
Under the stronger assumption that Ym is disconnected, there is an edge e1 × {m} in
Z, and we can choose the element g in Ge1 . The tree S provides the required splitting,
with edge group GZ Ge2 .
Lemma 6.21. If Ym is connected whenever m is the midpoint of an edge of T2 , then Yy
is connected for every y ∈ T2 .
Proof. Clearly Yy is homeomorphic to some Ym if y belongs to the interior of an edge, so
assume that y is a vertex of T2 . We sketch the argument, which is standard. Consider
a, b ∈ Yy , and join them by a piecewise linear path γ in C. The projection of γ to T2 is a loop
based at y. If this projection is not {y}, then γ has a subpath γ0 whose endpoints project
to y and whose initial and terminal segments project to the same edge e2 = yv. Using
connectedness of Ym , for m the midpoint of e2 , we may replace γ0 by a path contained in
Yy . Iterating yields a path joining a and b in Yy .
Lemma 6.22. Assume that T1 , T2 are fully hyperbolic with respect to each other, and that
T2 is minuscule. Then Cˇ ⊂ C.
Proof. Lemmas 6.20 and 6.21 imply that Yy is connected for every y ∈ T2 . Being connected
ˇ
and Gy -invariant, Yy contains µT1 (Gy ) × {y}, so C contains C.
Proposition 6.19 follows immediately from Lemma 6.22 by symmetry.
6.4
The regular neighborhood
In this subsection we assume that T1 , T2 are non-trivial, fully hyperbolic with respect to
each other, and that Cˇ = C. We use the core to construct a tree R, which we call the
regular neighborhood of T1 and T2 . Its main properties are summarized in Proposition
6.25.
We have seen (Lemma 6.18) that C is a surface away from its vertices. It follows
that the link of any vertex v ∈ C is a one-dimensional manifold, i.e. a disjoint union of
lines and circles. Since C is simply connected, the vertices whose link is disconnected are
precisely the cut points of C. We define the regular neighborhood R as the tree dual to
the decomposition of C by its cut points.
Definition 6.23 (Regular neighborhood). The regular neighborhood R = RN (T1 , T2 ) =
RN (T2 , T1 ) is the bipartite tree with vertex set V t S, where V is the set of cut vertices x
of C, and S is the set of connected components Z of C \ V. There is an edge between x and
Z in R if and only if v is in the closure Z̄ of Z.
V may be empty, but S is always non-empty (unless T1 , T2 are trivial).
Example 6.24. The basic example is the following. As in Example 6.15, let T1 , T2 be cyclic
splittings of the fundamental group of a closed orientable surface Σ dual to families of
geodesics L1 , L2 . Then R is dual to the family L defined as follows (more precisely, R
is a subdivision of the tree dual to L): consider the boundary of a regular neighborhood
of L1 ∪ L2 , disregard homotopically trivial curves, and isotope each non-trivial one to a
geodesic. Orbits of vertices in S correspond to components of Σ \ L meeting L1 ∪ L2 . The
other components correspond to orbits of vertices in V.
62
Proposition 6.25. Suppose that all groups in A are slender, and A satisfies (SC) or
(SCZ ). Let T1 , T2 be non-trivial trees that are fully hyperbolic with respect to each other,
with Cˇ = C.
The bipartite tree R is a minimal (A, H)-tree satisfying the following properties:
(RN1) If v ∈ S, then Gv acts hyperbolically in T1 and T2 . It is an extension 1 → F → Gv →
O → 1, where the fiber F is in F, is contained in every incident edge stabilizer, and
fixes an edge in T1 and in T2 . There are two possibilities.
If Gv is not slender, it is QH with fiber F . The underlying orbifold Σ is hyperbolic,
with fundamental group O, and contains an essential simple closed geodesic. There
is no mirror if (SCZ ) holds. Extended boundary subgroups of Gv are elliptic in T1
and T2 .
If Gv is slender, then O is virtually Z2 (it is the fundamental group of a Euclidean
orbifold Σ without boundary). Incident edge stabilizers are finite extensions of F .
(RN2) The stabilizer of a vertex v ∈ V is elliptic in T1 and T2 ; in particular, R is elliptic
with respect to T1 and T2 . Conversely, any group H ⊂ G elliptic in T1 and T2 is
elliptic in R; more precisely, H fixes a vertex v ∈ V, or H fixes a vertex v ∈ S and
the image of H in O is finite or contained in a boundary subgroup.
(RN3) For i = 1, 2, one passes from R to Ti by refining R at vertices v ∈ S using families
of essential simple closed geodesics in the underlying orbifolds, and then collapsing
all original edges of R. In particular, R is compatible with Ti . Any edge stabilizer of
Ti fixes some v ∈ S, and no other vertex of R; it contains the associated fiber.
The tree R may be trivial (in Example 6.24, this happens precisely when L1 and L2
fill Σ). In this case, G itself is slender or QH with fiber in F.
Proof. The action of G on C induces an action of G on R. It is not obvious that R is
minimal or has edge stabilizers in A. Note, however, that R is relative to H. In fact, any
subgroup H which is elliptic in T1 and T2 is elliptic in R, because H fixes a vertex of C by
Remark 6.17.
The stabilizer of a vertex v ∈ V fixes a point in C, so is elliptic in T1 and T2 . In
particular, edge stabilizers of R are elliptic in T1 and T2 ; in other words, R is elliptic with
respect to T1 and T2 .
The heart of the proof is to show that Gv is QH or slender for v ∈ S (RN1). Along
the way we will show that R is an A-tree and complete the proof of (RN2). We will then
prove (RN3) and minimality of the action of G on R.
If v ∈ S, the group Gv acts on Z̄, the closure of the connected component Z of C \ V
associated to v. Thus Z̄ may be viewed as one of the pieces one obtains when cutting C
open at its cut points. Stabilizers of edges of R incident to v fix a vertex in Z̄. The group
Gv contains an edge stabilizer of T1 (resp. T2 ), so acts hyperbolically in T2 (resp. T1 ).
Recall that Z̄ is made of squares. Since G acts on T1 and T2 without inversions, any
g ∈ Gv leaving a square S invariant is the identity on S, hence on adjacent squares (because
Z̄ is a pseudo-surface), and therefore on the whole of Z̄. We let F be the pointwise stabilizer
of Z̄, so that O = Gv /F acts on Z̄.
We have pointed out in Subsection 6.3.2 that G acts on C with finitely many orbits
of squares, so the same is true for the action of Gv (and O) on Z̄. The action of O on
Z̄ is proper in the complement of vertices, and free in the complement of the 1-skeleton
(an element may swap two adjacent squares). We consider this action near a vertex x.
Since we now view x as a vertex of Z̄ rather than one of C, its link Lx is connected. The
stabilizer Ox of x for the action of O acts on Lx with trivial edge stabilizers.
63
If Lx is a circle (so Z̄ is a surface near x), Ox is a finite group (cyclic or dihedral) and
the action is proper near x. If Lx is a line, the stabilizer Ox acts on it by translations or
dihedrally, and the image of x in Z̄/O must be viewed as a puncture. Since we want a
compact orbifold with boundary, we remove an Ox -invariant open neighborhood of x from
Z̄.
After doing this Gv -equivariantly near all vertices of Z̄ whose link is a line, we get
an effective proper action of O on a simply connected surface, with quotient a compact
orbifold Σ. The fundamental group of Σ is isomorphic to O.
If H is a subgroup of Gv which is elliptic in T1 and T2 (in particular if H ∈ H), it fixes
a vertex x ∈ Z̄, so its image in π1 (Σ) is finite (if the link of x is a circle) or contained in a
boundary subgroup of π1 (Σ) (if the link is a line). Conversely, any subgroup of Gv whose
image in π1 (Σ) is finite or contained in a boundary subgroup of π1 (Σ) is elliptic in T1 and
T2 . This completes the proof of (RN2).
Before concluding that Gv is slender or QH, we use (SC) or (SCZ ) to prove that edge
stabilizers of R are in A.
We first show F ∈ F. Given any square e1 × e2 ⊂ Z̄, the group F is the stabilizer of
e2 for the action of Ge1 on the line µT2 (Ge1 ); in particular, F fixes an edge in T2 (and in
T1 ). Moreover, F is the kernel of an epimorphism from Ge1 to Z or D∞ . The stability
condition implies F ∈ F. If (SCZ ) holds, Ge1 acts on µT2 (Ge1 ) by translations, and Ge2
acts on µT1 (Ge2 ) by translations. This is true for all squares e1 × e2 ⊂ Z̄, so the orbifold
Z̄/O contains no mirror.
Any incident edge stabilizer of v in R is the stabilizer Gx of a vertex x ∈ Z̄ for the
action of Gv on Z̄ (as pointed out earlier, it is elliptic in T1 and T2 ). It contains F , and
Ox = Gx /F acts on the link of x with trivial edge stabilizers, so is cyclic or dihedral (finite
or infinite). The stability condition (SC) implies Gx ∈ A. If (SCZ ) holds, Ox is cyclic
because there is no mirror and we also get Gx ∈ A. The tree R being bipartite, it is an
A-tree.
Given v ∈ S, we would have proved that Gv is a QH subgroup (in the sense of Definition
5.13) if we knew that the orbifold Σ is hyperbolic. Assume otherwise. Since Σ is the
quotient of a simply connected surface by the infinite group O = π1 (Σ), it is Euclidean.
Its fundamental group is not virtually cyclic because it contains the image of Ge1 and Ge2
for any square e1 × e2 in Z̄. This implies that Σ is a quotient of the torus. In particular,
Σ has empty boundary, and incident edge stabilizers of v in R are finite extensions of F .
We also deduce that π1 (Σ) contains Z2 with finite index, so Gv is slender because it is an
extension of F by π1 (Σ).
To sum up: for v ∈ S, the group Gv is slender (if Σ is Euclidean) or QH (if Σ is
hyperbolic).
We now prove (RN3) for i = 1 (the proof for i = 2 is the same). In the process we
will show that all orbifolds Σ contain an essential geodesic, thus completing the proof of
(RN1), and that the action of G on R is minimal.
Let M1 ⊂ C be the preimage of all midpoints of edges of T1 under the first projection
p1 : C → T1 . It is a G-invariant collection of disjoint properly embedded lines containing
no vertex. One may view T1 as the tree dual to M1 , with vertices the components of
C \ M1 and edges the components of M1 : the projection p1 induces a map from the dual
tree to T1 , which is an isomorphism because point preimages of p1 are connected.
It immediately follows that any edge stabilizer Ge1 of T1 fixes a vertex v ∈ S (associated
to the component of C \V containing the preimage by p1 of the midpoint of e1 ) and contains
the associated fiber; in particular, T1 is elliptic with respect to R. Since Ge1 is hyperbolic
in T2 and edge stabilizers of R are elliptic, v is the only vertex of R fixed by Ge1 . This
shows that, for any v ∈ S, there exists a subgroup of G (an edge stabilizer of T1 ) having
v as its unique fixed point. Minimality of R easily follows from this observation, since no
vertex in V is terminal.
64
We define T̂1 refining both T1 and R as the tree dual to the decomposition of C given
by its cut points and M1 . Its vertices are the elements of V together with the components
of C \ (V ∪ M1 ), its edges are those of R and components of M1 . One obtains T1 from T̂1
by collapsing edges coming from edges of R, and T̂1 from T1 by refining at vertices v ∈ S.
There remains to show that the refinement at v is dual to a family of geodesics on the
associated orbifold Σ.
Recall that v is associated to a component Z of C \ V. Consider a component `1 of
M1 contained in Z. Its stabilizer (an edge stabilizer Ge1 of T1 ) acts cocompactly on `1 , so
the image of `1 in the orbifold Σ associated to v is a simple 1-suborbifold (simple closed
curve if there is no mirror). This suborbifold is not homotopically trivial, so is isotopic
to a geodesic; it is not boundary parallel because Ge1 is hyperbolic in T2 but extended
boundary subgroups of Gv are elliptic. Components of M1 contained in Z thus yield the
required non-empty family of disjoint essential simple closed geodesics in Σ.
Remark 6.26. Note that a given orbifold Σ is filled by the images of M1 (defined above)
and M2 (defined similarly using T2 ). The proof also shows that the number of orbits of
vertices in S is bounded by the number of orbits of edges of T1 . If T1 and T2 are one-edge
decompositions, then R only has one orbit of vertices in S, and T1 , T2 are dual to single
geodesics.
Proposition 6.27. Under the assumptions of Proposition 6.25, suppose furthermore that
T1 is minuscule and G is not slender. If v ∈ S, then Gv is not slender; it is QH, and every
boundary component of the underlying orbifold is used (see Definition 5.15).
Proof. Recall that F fixes an edge ei in Ti . Since T1 is fully hyperbolic with respect to T2 ,
we have F Ge1 . If Gv is slender but G is not, then R 6= {v}, so there exists an edge in
R incident to v. Its stabilizer is a finite extension of F , so Ge1 cannot be minuscule. This
contradiction implies that Gv is QH.
By Lemma 5.16, if some boundary component is unused, G splits over a group containing F with index ≤ 2 (hence in A by the stability condition), again a contradiction.
6.5
Constructing a filling pair of splittings
In this subsection, as well as the next one, we assume that all trees are minuscule. This
guarantees symmetry of the core (Proposition 6.19), so any two trees fully hyperbolic with
respect to each other have a regular neighborhood. It also implies that ellipticity is a
symmetric relation among trees (Lemma 6.11).
The goal of this subsection is the following result showing the existence of a pair of
splittings U , V that fill G.
Proposition 6.28. Assume that G is totally flexible and all trees are minuscule. Given
a tree U , there exists a tree V such that U and V are fully hyperbolic with respect to each
other.
In the next subsection we will apply this proposition to a maximal U (given by Lemma
6.12).
Example 6.29. In the surface case (Examples 6.13 and 6.15), if U is dual to a family L
decomposing Σ into pairs of pants, then V is dual to a family meeting (transversely) every
curve in L.
More generally, suppose that G is the fundamental group of a compact hyperbolic 2orbifold Σ, and H is the set of boundary subgroups, with A consisting of all virtually
cyclic subgroups. Then U is dual to a family of geodesics L by Proposition 5.4, and the
proposition claims that there exists a family L0 such that every geodesic in L meets L0
and every geodesic in L0 meets L (but no geodesic belongs to both L and L0 ). To prove
65
this, consider a maximal family L0 such that each geodesic in L0 meets L (transversely).
Applying Lemma 5.11 to the orbifold obtained by cutting Σ along L0 shows that every
geodesic in L meets L0 .
Proof. The difficulty is to find V with U fully hyperbolic with respect to V . Once this
is done, we redefine V by collapsing all edges which do not belong to the axis of an edge
stabilizer of U . This makes U hyperbolic with respect to each one-edge splitting underlying
V . By Lemma 6.11, each one-edge splitting underlying V is hyperbolic with respect to U
and therefore V is fully hyperbolic with respect to U .
We argue by induction on the number n of orbits of edges of U . For n = 1, the existence
of V is just the fact that, G being totally flexible, U is not universally elliptic.
U
|
G1
{z
Gn−1 Gn
}| {z }
hyp. in W
ell. in W
hyp. in T
T
W
q
WE + WH
| {z } | {z }
R = RN (WH , T )
| {z }
hyp. in WH
ell. in T hyp. in T
ell. in WE
Q
Figure 8: The first players in the proof of Proposition 6.28.
Denote by G1 , . . . , Gn representatives of the edge stabilizers of U (see Figure 8, where
we represent the quotient graphs of groups). By induction, there is a tree W in which
G1 , . . . , Gn−1 are hyperbolic. We can assume that Gn is elliptic in W . By the case n = 1
there exists T in which Gn is hyperbolic, and we can assume that T has a single orbit
of edges. If W is elliptic with respect to T , we take for V a standard refinement of W
dominating T .
Otherwise, let WE , WH be obtained from W by collapsing edges, keeping only those
edges whose stabilizer is elliptic or hyperbolic in T respectively (WE may be trivial, but
WH is not). Note that WH is fully hyperbolic with respect to T . By symmetry of the
ellipticity relation among minuscule trees, T is (fully) hyperbolic with respect to WH . On
the other hand, WE and T are elliptic with respect to each other, and so are WE and WH
(they are both collapses of W ).
Let R = RN (WH , T ) be the regular neighborhood. Recall that the set of vertices of
R is bipartite, with V the set of cut points of C = C(WH , T ) ⊂ WH × T , and S the set
of connected components of C \ V. Since T has a single orbit of edges, there is a single
orbit of vertices in S. We fix q ∈ S, and denote by Q its stabilizer. It is a QH vertex by
Proposition 6.27.
Lemma 6.30. The QH vertex group Q is elliptic in WE .
ε⊂Axis(Ge )
ŴE
ê′ ⊂Axis(Ge )
R̂T
e′ ⊂Axis(Ge )
WE
| {z }
ce ,ce′ ∈
ell. in T, WH , R, R̂T
T
Ge′
| {z }
R = RN (WH , T )
hyp. in WH
Figure 9: Proof that Q is elliptic in WE ( represents a collapse map).
66
Proof. In spirit, the proof is somewhat analogous to that of Theorem 5.27.
Let e be an edge of WH , and e0 an edge of T . Since WH is elliptic with respect to WE ,
the stabilizer Ge fixes a point ce ∈ WE . This point is unique because otherwise Ge would
fix an edge of WE . But WE is elliptic with respect to T , so Ge would be elliptic in T ,
contradicting the fact that WH is fully hyperbolic with respect to T . Similarly, Ge0 fixes a
unique point ce0 ∈ WE .
As above, consider the core C = C(WH , T ) ⊂ WH × T . We claim that ce = ce0 if the
square e × e0 is contained in C. Assuming this, we associate to any square e × e0 ⊂ C the
point ce = ce0 ∈ WE . Since adjacent squares are mapped to the same point, all squares
in a given connected component of C \ V are mapped to the same point. Since Q is the
stabilizer of such a component, it fixes a point in WE , as required.
To conclude the proof, we have to prove our claim that ce = ce0 if e × e0 ⊂ C. Let R̂T
be a common refinement of R and T as in (RN3) (see Figure 9). By definition of the core,
e0 lies in the axis (=minimal subtree) of Ge in T . Since R̂T collapses to T , the edge ê0 of
R̂T projecting to e0 lies in the axis of Ge in R̂T . Note that Gê0 = Ge0 .
Since WE is elliptic with respect to WH and T , it is elliptic with respect to R by
(RN2), hence also with respect to R̂T . Let therefore ŴE be a standard refinement of WE
dominating R̂T as in Proposition 2.2.
Let M ⊂ ŴE be the axis of Ge . Since R̂T does not have to be a collapse of ŴE , we
cannot claim that M maps to the axis of Ge in R̂T , but we can find an edge ε ⊂ M whose
image in R̂T contains ê0 , so that Gε ⊂ Gê0 = Ge0 . Since Ge fixes ce and only this point in
WE , the image of M under the collapse map ŴE → WE is {ce }. In particular, Gε fixes ce .
Since Gε ⊂ Ge0 fixes ce0 , it is enough to prove that Gε fixes no edge in WE . If it does, then
Gε is elliptic in WH (because WE is elliptic with respect to WH ), but Ge0 is not (because
T is hyperbolic with respect to WH ), so Gε Ge0 . This contradicts the fact that T is
minuscule and proves the claim.
V
U
|
G1
Gn−1
} Gn
| {z }
in W
{z
hyp.
U
| {zE }
ell. in R̂
dual to L′
R̂
hyp. in T
WE
Q
L ⋔ L′
R = RN (WH , T )
W = WE + WH
Q
Figure 10: End of the proof of Proposition 6.28.
Using the lemma, we can construct a standard refinement R̂ of R dominating WE (see
Figure 10) without refining at vertices in the orbit of q. In particular, q is still a QH vertex
of R̂. The tree V will be obtained by refining R̂ at q. Recall that we want every edge
stabilizer Gi of U to be hyperbolic in V . Since this will be automatic if Gi is hyperbolic
in R̂, we consider the tree UE obtained by collapsing edges of U , keeping only those edges
whose stabilizer is elliptic in R̂.
If UE is trivial, we simply take V = R̂, so we assume the contrary. We claim that
Q cannot be elliptic in UE , nor in any collapse ŪE which is not a point. Otherwise, the
edge stabilizers of T and WH , which are conjugate to subgroups of Q by (RN3), would be
elliptic in ŪE , so ŪE would be elliptic with respect to T and WH . This is a contradiction
because every Gi is hyperbolic in T or W , hence in T or WH because it is elliptic in WE
(which is dominated by R̂).
67
We now consider the action of Q on its minimal subtree µUE (Q). Note that µUE (Q)
meets every G-orbit of edges (otherwise Q would be elliptic in some ŪE ), so every Gi which
is elliptic in R̂ contains (up to conjugacy) an edge stabilizer of the action of Q on µUE (Q).
It follows from previous results that the action of Q on µUE (Q) is (up to subdivision)
dual to a family L of geodesics on the underlying orbifold Σ. To see this, view Q as a QH
vertex stabilizer of R or R̂. Every boundary component of Σ is used (Proposition 6.27),
the fiber is universally elliptic (Lemma 5.23), and incident edge groups are elliptic in UE
because UE and R̂ are elliptic with respect to each other. The existence of L then follows
from Lemma 5.18.
Now let L0 be a family of geodesics transverse to L as in Example 6.29. Define V by
refining R̂ at q using the splitting of Q dual to L0 , and observe that every edge stabilizer
Gi of U is hyperbolic in V , as required. This is clear if Gi is hyperbolic in R̂. Otherwise,
it contains (up to conjugacy) an edge stabilizer J ⊂ Q of µUE (Q), and J is hyperbolic in
V because every curve in L meets L0 .
6.6
Flexible groups are QH when trees are minuscule
We can now prove Theorem 6.8 under the assumption that all trees are minuscule. We
will prove in the next subsection that this assumption is always fulfilled.
Proof of Theorem 6.8, assuming that all trees are minuscule. Using finite presentability (which
is assumed in Theorem 6.8), we fix a maximal splitting U as in Lemma 6.12. Proposition
6.28 yields V such that U and V are fully hyperbolic with respect to each other. We may
therefore consider the regular neighborhood R of U and V . We claim that R is trivial (a
point). This implies that G is QH, as required.
Assume that R is non-trivial. Let R̂ be a common refinement of R and U (it exists by
Proposition 6.25(RN3)). Since R is elliptic with respect to V , while U is fully hyperbolic
with respect to V , each edge of R̂ must be collapsed in R or in U (possibly in both). Being
non-trivial, R contains an edge ε, and we let ε̂ be the edge of R̂ that is mapped to ε. Define
U 0 by collapsing all edges of R̂ which are not collapsed in U , except those in the orbit of ε.
Since U 0 collapses to U , maximality of U (Lemma 6.12) implies that U and U 0 belong
to the same deformation space. Thus ε̂ (viewed as an edge of U 0 ) has an endpoint v ∈ U 0
with Gv = Gε̂ , and its other endpoint is in a different orbit. Since the action of G on U 0
is minimal, there is an edge e of U 0 with origin v which does not belong to the orbit of ε̂.
This edge has the same stabilizer as an edge of U , so Ge is hyperbolic with respect to V .
This is a contradiction since Ge ⊂ Gv = Gε̂ = Gε , and Gε is elliptic in V because R is
elliptic with respect to V by (RN2).
6.7
All splittings of a totally flexible group are minuscule
In this subsection, we complete the proof of Theorems 6.2 and 6.5 by showing that all
splittings of G are minuscule if G is totally flexible (Proposition 6.35). As mentioned in
Remark 6.10, this is sometimes not needed.
In a few words, the proof goes as follows. Given an edge stabilizer Ge of a tree T , we
first find a minuscule tree T1 with an edge stabilizer Ge1 commensurable with a subgroup
of Ge . The goal is to show that Ge and Ge1 are in fact commensurable. This is proved
by showing that Ge1 and Ge are slender subgroups of a QH vertex group, both containing
the fiber. So ultimately, the argument relies on the fact that two nested infinite slender
subgroups of a hyperbolic orbifold group are commensurable. To embed Ge1 and Ge into
a QH group, we use a tree T2 in which Ge1 is hyperbolic and we construct the core. Since
we do not know that T2 is minuscule, we have to choose it carefully, to ensure that the
core is symmetric (hence a surface).
68
6.7.1
Complexity of edge groups
Recall that A consists of slender groups; in particular, they are finitely generated.
Lemma 6.31. Let A0 ∈ A. There is a bound, depending only on A0 , for the length n of
a chain A0 A1 A2 · · · An . In particular, there is no infinite descending chain.
Proof. Since Ai−1 Ai , we can find a tree Ti in which Ai is elliptic, and Ai−1 is not.
Let B0 be the intersection of all subgroups of index 2 of A0 . For each i, the action of A0
on Ti has an invariant line, and B0 acts on this line as a translation, given by a non-zero
homomorphism ϕi : B0 → Z. Letting Bi = Ai ∩ B0 , we have ϕi (Bj ) = 0 if and only if
j ≥ i. This implies that the ϕi ’s are linearly independent, and the lemma follows since B0
is finitely generated.
Using this lemma, we can associate a complexity c(A0 ) to any A0 ∈ A: it is the maximal
length n of a chain A0 A1 · · · An where A1 , . . . , An are commensurable with edge
stabilizers of trees T1 , . . . , Tn . Thus A0 is minuscule if and only if c(A0 ) = 0. Note that
c(B) ≤ c(A) if B ⊂ A, and c(B) = c(A) if A, B are commensurable.
Lemma 6.32. Given an edge stabilizer Ge of a tree T , there exists a tree T 0 with an edge
stabilizer Ge0 such that Ge0 is minuscule and Ge0 ∩ Ge has finite index in Ge0 .
Proof. The result is trivial if c(Ge ) = 0, with T 0 = T . If c(Ge ) > 0, we argue by induction.
By definition of the complexity, there exists A Ge with A commensurable with an edge
stabilizer Ge1 . We have c(Ge1 ) = c(A) < c(Ge ), so by the induction hypothesis there exists
a minuscule edge stabilizer Ge0 such that Ge0 ∩ Ge1 has finite index in Ge0 . Since A and
Ge1 are commensurable, and A ⊂ Ge , the index of Ge0 ∩ Ge in Ge0 is finite.
6.7.2
Symmetry of the core
Recall (Proposition 6.19) that the core of minuscule trees is symmetric.
In order to prove that all trees are minuscule, we need to establish some weaker symmetry statements. We use the same notations as in Subsection 6.3.
Lemma 6.33. Let T1 , T2 be fully hyperbolic with respect to each other. If Cˇ ⊂ C and
C (2) ⊂ Cˇ(2) , then Cˇ = C.
Proof. Note that Cˇ ⊂ C implies Cˇ(2) ⊂ C (2) , so we have Cˇ(2) = C (2) . Assuming Cˇ =
6 C, we
get C (2) = Cˇ(2) ⊆ Cˇ C. As pointed out in Subsection 6.3.3, C \ C (2) is a union of open
vertical edges and each such edge disconnects C. Our assumption Cˇ 6= C says that C \ Cˇ
contains such an edge ε̊ = {v} × e̊2 .
The set of edges e of µT2 (Gv ) such that {v} × e bounds a square in C is a non-empty
Gv -invariant set which does not contain e2 . By minimality of the action of Gv on µT2 (Gv ),
this set intersects both components of µT2 (Gv ) \ e̊2 . In particular, there exist a, b ∈ C (2)
that are not in the same connected component of C \ ε̊. But a and b may be joined by a
ˇ a contradiction since Cˇ ⊂ C \ ε̊.
path in C,
Lemma 6.34. Let T1 be a minuscule one-edge splitting. Assume that T1 is not universally
elliptic. Among all one-edge splittings T such that T1 is not elliptic with respect to T ,
consider T2 whose edge stabilizers have minimal complexity.
ˇ 1 , T2 ).
Then C(T1 , T2 ) = C(T
Note that T1 and T2 are fully hyperbolic with respect to each other by Lemma 6.11, so
ˇ 1 , T2 ) are defined.
C(T1 , T2 ) and C(T
69
Proof. Let m be the midpoint of an edge e2 of T2 . Since edge stabilizers of T2 have minimal
(2)
complexity, the second assertion of Lemma 6.20 shows that Ym = C (2) ∩ (T1 × {m}) is
(2)
connected. In particular, Ym contains µT1 (Ge2 ) × {m}, and C (2) therefore contains all
squares e1 × e2 with e1 ⊂ µT1 (Ge2 ). This shows Cˇ(2) ⊂ C (2) .
ˇ We conclude by
On the other hand, T1 being minuscule, Lemma 6.22 implies C ⊂ C.
Lemma 6.33.
6.7.3
All splittings are minuscule
Proposition 6.35. If G is totally flexible and not slender, then every tree T is minuscule.
Proof. We may assume that T is a one-edge splitting. Denoting by Ge an edge stabilizer,
Lemma 6.32 yields a minuscule one-edge tree T1 with an edge stabilizer Ge1 such that
Ge1 ∩ Ge has finite index in Ge1 . Since G is totally flexible, T1 is not universally elliptic,
so we can choose T2 as in Lemma 6.34. The core of T1 and T2 being symmetric, we can
consider the regular neigbourhood decomposition R = RN (T1 , T2 ).
By (RN3), we know that Ge1 fixes a unique vertex v of R, and v ∈ S. Its finite index
subgroup Ge1 ∩ Ge also has v as its unique fixed point (it is hyperbolic in T2 , while edge
stabilizers of R are elliptic by (RN2)).
First assume that Gv is not slender, so that it is QH with fiber F . Consider any slender
group A containing Ge1 ∩ Ge , for instance Ge . We claim that Ge1 ∩ Ge has finite index in
A. First, A is elliptic in R: otherwise Ge1 ∩ Ge acts on the axis of A, and (being elliptic)
has a subgroup of index ≤ 2 fixing the axis, a contradiction since Ge1 ∩ Ge fixes only v.
We deduce A ⊂ Gv . Since Gv is QH, its slender subgroups are contained in extensions
of F by virtually cyclic groups. As Ge1 contains F with infinite index, and Ge1 ∩ Ge has
finite index in Ge1 , the index of Ge1 ∩ Ge in A is finite, so A is commensurable with Ge1 .
Applied to A = Ge , this argument shows that Ge is minuscule.
If Gv is slender, there has to be at least one edge incident to v since we assume that
G is not slender. By (RN1), the stabilizer of this edge is commensurable with F . On the
other hand, F is contained in Ge1 ; it is elliptic in T2 , but Ge1 is not, so F Ge1 . This
shows that Ge1 is not minuscule, a contradiction.
6.8
Slenderness in trees
We have proved that non-slender flexible vertex groups of JSJ decompositions are QH
when A satisfies a stability condition and consists of slender groups. We now consider
edge groups which are only slender in trees.
Recall that A is slender in (A, H)-trees if, whenever G acts on a tree, A fixes a point
or leaves a line invariant.
We now have the following generalization of Theorems 6.2 and 6.5:
Theorem 6.36. Let G be finitely presented relative to a finite family H of finitely generated
subgroups. Suppose that all groups in A are slender in (A, H)-trees, and A satisfies (SC)
or (SCZ ), with fibers in a family F.
If Q is a flexible vertex group of a JSJ decomposition of G over A relative to H, and
Q is not slender in (A, H)-trees, then (as in Theorems 6.2 and 6.5) Q is QH with fiber in
F (and the orbifold has no mirror in the (SCZ ) case).
This applies whenever groups in A are assumed to be slender or contained in a group
of H (up to conjugacy), for instance to splittings of relatively hyperbolic groups over
elementary subgroups relative to parabolic subgroups.
Theorem 6.36 is proved as in the slender case, with just a few changes which we now
describe.
70
The arguments in Subsections 6.3 through 6.5 extend directly, replacing “slender” by
“slender in (A, H)-trees” and using Lemma 1.5. The proof of Proposition 6.28 uses Lemma
5.23, which we replace by the following statement:
Lemma 6.37. Let Q be a QH vertex group. If F and all groups in A are slender in
(A, H)-trees, but Q is not, then F is universally elliptic.
Proof. If F is not universally elliptic, then as in the proof of Lemma 5.23 Q is an extension
of a group in A by a virtually cyclic group, so is slender in (A, H)-trees by Lemma 1.5.
In Subsection 6.7, Lemma 6.31 requires A0 to be finitely generated, but the other groups
Ai may be arbitrary. Complexity may be infinite, but Lemmas 6.32 and 6.34 remain valid
because every tree is dominated by a tree with finitely generated stabilizers (see Corollary
2.25).
At the end of the proof of Proposition 6.35 we need to know that a subgroup A ⊂ Gv
which is slender in (A, H)-trees is contained in an extension of F by a virtually cyclic
group. More generally, suppose that Q is QH and the underlying orbifold Σ contains an
essential simple geodesic. We claim that, if A ⊂ Q is slender in (A, H)-trees, then its
image in π1 (Σ) is virtually cyclic.
To see this, assume that A is not an extended boundary subgroup. As in Proposition
5.21, we consider a splitting of Q dual to a geodesic γ, in which A is not elliptic. The
group A/(A ∩ F ) acts on a line with virtually cyclic edge stabilizers, so is virtually abelian,
hence virtually cyclic because it embeds into π1 (Σ). This proves the claim.
6.9
Slender flexible groups
In this subsection we consider a slender flexible vertex group Q of a JSJ tree. Whenever G
acts on a tree and Q does not fix a point, there is a unique Q-invariant line and the action
of Q on this line gives rise to a map ϕ : Q → D∞ ' Isom(Z) whose image is isomorphic to
Z (if orientation is preserved) or to D∞ .
A natural analogue of Theorem 6.2 would be the following statement:
(1) Q maps onto the fundamental group of a compact Euclidean 2-orbifold Σ, with fiber
in F;
(2) incident edge groups have finite image in π1 (Σ) (note that Euclidean 2-orbifolds
whose fundamental group is not virtually cyclic have empty boundary);
(3) if G acts on a tree and Q does not fix a point, the action of Q on its invariant line
factors through π1 (Σ).
Unfortunately, Assertion (3) of this statement is not correct (consider splittings of Zn ,
with n ≥ 3). We replace it by two assertions.
Proposition 6.38. Let G, H, A be as in Theorem 6.2 or Theorem 6.5. Let Q be a slender
flexible vertex group of a JSJ tree.
(1) Q is an extension 1 → F → Q → O → 1, with F ∈ F and O virtually Z2 . Incident
edge groups are contained in finite extensions of F (they are finite if groups in A are
virtually cyclic).
(2) there is a map ψ : Q → (D∞ )n , for some n ≥ 2, whose image has finite index;
incident edge groups have finite image; if G acts on a tree and Q does not fix a point,
the action of Q on its invariant line factors through ψ.
71
In the first assertion, we do not claim that actions of Q factor through the fundamental
group of the 2-orbifold O (the fiber does not have to be universally elliptic). In the second
one, Q maps to an n-dimensional Euclidean orbifold, but we do not claim that the kernel
is in F.
The first assertion only requires the first half of the stability condition, the second one
requires no stability condition.
The proposition remains true if Q is only assumed to be slender in (A, H)-trees.
Example 6.39. The group Zn has infinitely many slender splittings, given by its morphisms
to Z. The Klein bottle group ha, t | tat−1 = a−1 i = hti ∗t2 =v2 hvi has exactly two splittings, corresponding to these two presentations (see Subsection 3.4). They correspond to
morphisms to D∞ and Z respectively.
In [Bee14], Beeker classified all flexible groups that occur in the abelian JSJ decomposition of the fundamental group of a graph of free abelian groups. In particular, he
exhibits a twisted Klein bottle group K 0 = hb0 , b1 | [b0 , b21 ] = 1, [b20 , b1 ] = 1i which splits as
an amalgam hb0 , b21 i ∗hb20 ,b21 i hb20 , b1 i and also as an HNN extension (see Proposition 4.1 in
[Bee14]).
Proof. We sketch the proof of (1). Being flexible, Q admits two different splittings relative
to IncH
|Q . They are given by epimorphisms ϕi : Q → Ki , with Ki equal to Z or D∞ . The
image O of ϕ : Q → K1 × K2 is not virtually cyclic, so is virtually Z2 . The kernel F is in
F because it is also the kernel of ϕ| ker ϕ1 , with ker ϕ1 ∈ A. If Ge is an incident edge group,
its image by ϕi has order at most 2 because Ge is universally elliptic, hence elliptic in the
corresponding splitting of Q.
For (2), we consider all non-trivial actions of Q on the line which are restrictions of
actions of G on trees. We view them as non-trivial homomorphisms ϕα : Q → D∞ (there
may be infinitely many of them). Let J = ∩ ker ϕα be the subgroup of Q consisting of
elements always acting as the identity, and let D = Q/J. We shall show that D embeds
with finite index in some (D∞ )n .
The maps ϕα induce maps ϕ̄α : D → D∞ , and ∩ ker ϕ̄α is trivial. Let D0 be the
subgroup of D consisting of elements always acting as translations. It is abelian, torsionfree, has finite index (it contains the intersection of all subgroups of index 2), and contains
the commutator subgroup. Let n be its rank.
Each ϕ̄α induces pα : D0 → Z. The subgroup of Hom(D0 , Z) generated by the pα ’s has
finite index because ∩ ker pα is trivial, and we choose a finite family ϕ̄αi (1 ≤ i ≤ n) such
that the corresponding pαi ’s are a basis over Q. We show that the product map
Φ=
n
Y
i=1
ϕ̄αi : D → (D∞ )n
is injective and its image has finite index.
Let x ∈ D. If it has infinite order, there is a non-trivial xk in D0 . This xk is mapped
non-trivially by some pα , hence by some pαi , hence by ϕ̄αi , hence by Φ, so x ∈
/ ker Φ. If x
has finite order and x 6= 1, the order is 2 and some ϕ̄α maps it to a nontrivial reflection.
We can find y ∈ D such that z = [x, y] is mapped by ϕ̄α to a non-trivial translation. The
element z is in D0 , and as in the previous case z ∈
/ ker Φ. It follows that Φ is injective. Its
image has finite index because D and (D∞ )n are both virtually Zn .
Composing Φ with the quotient map Q → D yields a map ψ : Q → (D∞ )n whose
image has finite index. As in Assertion (1), incident edge groups are universally elliptic,
so their image by any ϕα has order at most 2. This implies that their image by ψ is finite.
The last claim is obvious from the way D was defined.
72
Remark 6.40. Similar arguments show that a finitely generated group is residually D∞ if
and only if it is a finite index subgroup of a direct product whose factors are isomorphic
to D∞ or Z/2Z.
Part IV
Acylindricity
In the previous chapters we have studied the JSJ deformation space. Its existence was
guaranteed by Dunwoody’s accessibility, which requires finite presentability. In this chapter
we propose a construction based on the idea of acylindrical accessibility.
Sela [Sel97a] defined a tree to be k-acylindrical if fixed point sets of non-trivial elements
have diameter bounded by k. Since we allow G to have torsion, the following definition is
better adapted [Del99].
Definition (Acylindrical). A tree T is (k, C)-acylindrical if the pointwise stabilizer of
every arc of length ≥ k + 1 is finite, of cardinality ≤ C.
Example. Let Σ be a 2-orbifold as in Subsection 5.1. The splitting of π1 (Σ) dual to a
family L of geodesics is 2-acylindrical (1-acylindrical if all geodesics are 2-sided). If Σ is
the underlying orbifold of a QH vertex group Q, the splitting of Q dual to L is (2, C)acylindrical if and only if the fiber F is finite with order ≤ C.
Acylindrical accessibility bounds the number of orbits of edges in any (k, C)-acylindrical
tree, under the assumption that G is finitely generated [Sel97a, Del99, Wei12].
Since finite presentability is no longer required, this allows us for instance to construct
JSJ decompositions of hyperbolic groups relative to an infinite collection of infinitely generated subgroups, and JSJ decompositions of finitely generated torsion-free CSA groups.
In order for our approach to work, we must be able to produce a (k, C)-acylindrical
tree, for some uniform constants k and C, out of an arbitrary tree. We do so using the
tree of cylinders introduced in [GL11].
As an additional benefit, this construction (when available) produces a canonical tree
T out of a deformation space D. If D is invariant under automorphisms of G, so is T (see
[GL10, GL15b] for applications of this construction, [Bow98, SS03, KM05, PS09] for other
constructions of canonical decompositions, and [Lev05a, DG08, BKM07, DG11, DT13] for
other uses of such a canonical decomposition). A general construction of another invariant
tree, based on compatibility of splittings, will be given in Part V (compatibility JSJ tree).
The tree of cylinders is introduced in Section 7, and Section 8 describes our construction
of JSJ decompositions based on acylindricity. Applications are given in Section 9.
Unless indicated otherwise, we only assume that G is finitely generated, and H may be
arbitrary.
7
Trees of cylinders
In the previous chapters we have defined and studied the JSJ deformation space, consisting
of all JSJ trees. It is much better to be able to find a canonical tree. A key example is provided by one-ended hyperbolic groups: Bowditch [Bow98] constructs a virtually cyclic JSJ
tree using only the topology of the boundary ∂G; it is Out(G)-invariant by construction.
Unfortunately, it is not always possible to find an invariant JSJ tree. If for instance G is
free, the JSJ deformation space consists of all trees with a free action of G (see Subsection
3.1), but (see Subsection 11.1) it is easy to check that the only Out(G)-invariant tree is
the trivial one (a point).
73
In this section we describe a construction, the (collapsed) tree of cylinders [GL11],
which, under certain conditions, associates a new, nicer tree to a given tree T . The first
feature of this new tree is that it depends only on the deformation space D of T . Its second
feature of interest to us here is that, under suitable assumptions, it is acylindrical with
uniform constants, and it lies in D or at least in a deformation space not too different from
D (the new tree is smally dominated by T in the sense of Definition 7.10). Its third feature
lies in its compatibility properties (in the sense of common refinements, see Subsection
1.4.1); this will be used in Chapter V to provide examples of compatibility JSJ trees.
The results of this section and the next (where JSJ decompositions are constructed) will
be summed up in Corollary 9.1, which gives conditions ensuring that there is a canonical
JSJ tree, and that its flexible vertices are QH with finite fiber.
7.1
Definition
We recall the definition and some basic properties of the tree of cylinders (see [GL11] for
details).
As usual, we fix A and H and we restrict to (A, H)-trees. We let A∞ be the family of
infinite groups in A. In applications we assume that G is one-ended relative to H, so all
trees have edge stabilizers in A∞ . The family A∞ is not stable under taking subgroups,
but it is sandwich closed : if A, B ∈ A∞ and A < H < B, then H ∈ A∞ .
Definition 7.1 (Admissible equivalence relation). An equivalence relation ∼ on A∞ is
admissible (relative to H) if the following axioms hold for any A, B ∈ A∞ :
(1) If A ∼ B and g ∈ G, then gAg −1 ∼ gBg −1 (invariance under conjugation).
(2) If A ⊂ B, then A ∼ B (nesting implies equivalence).
(3) Let T be a tree (relative to H) with infinite edge stabilizers. If A ∼ B, and A, B fix
a, b ∈ T respectively, then for each edge e ⊂ [a, b] one has Ge ∼ A ∼ B.
The equivalence class of A ∈ A∞ will be denoted by [A]. We let G act on A∞ /∼ by
conjugation, and the stabilizer of A will be denoted by G[A] .
Here are a few examples. They will be studied in detail later (see Section 9).
(1) If G is a torsion free CSA group (for instance a limit group, or more generally a toral
relatively hyperbolic group), we can take for A∞ the set of infinite abelian subgroups,
and for ∼ the commutation relation: A ∼ B if hA, Bi is abelian. The group G[A] is
the maximal abelian subgroup containing A.
(2) If G is a relatively hyperbolic group with small parabolic subgroups, we can take
for A∞ the set of infinite elementary subgroups (a subgroup is elementary if it is
virtually cyclic or parabolic; in this case, this is equivalent to being small). The
relation ∼ is co-elementarity: A ∼ B if and only if hA, Bi is elementary. The group
G[A] is the maximal elementary subgroup containing A. We may also allow nonsmall parabolic subgroups, provided that we include them in H (i.e. we only consider
splittings relative to the parabolic groups).
(3) A∞ is the set of infinite virtually cyclic subgroups, and ∼ is the commensurability
relation (A ∼ B if and only if A ∩ B has finite index in A and B). The group G[A]
is the commensurator of A.
Given an admissible equivalence relation ∼ on A∞ , we now associate a tree of cylinders
Tc to any tree T with infinite edge stabilizers.
We declare two edges e, f to be equivalent if Ge ∼ Gf (these groups are assumed to
be in A, and they are in A∞ because they are infinite). The union of all edges in an
74
equivalence class is a subtree by axiom (3). Such a subtree Y is called a cylinder of T ; two
distinct cylinders meet in at most one point. The equivalence class in A∞ /∼ containing
the stabilizers of edges of Y will be denoted by [Y ].
Definition 7.2 (Tree of cylinders). Given a tree T with edge stabilizers in A∞ , its tree of
cylinders Tc is the bipartite tree with vertex set V0 (Tc ) t V1 (Tc ), where V0 (Tc ) is the set of
vertices v of T which belong to at least two cylinders, V1 (Tc ) is the set of cylinders Y of
T , and there is an edge ε = (v, Y ) between v and Y in Tc if and only if v ∈ Y in T .
Equivalently, one obtains Tc from T by replacing each cylinder Y by the cone over the
set of vertices v ∈ Y belonging to another cylinder.
The collapsed tree of cylinders Tc∗ is the tree obtained from Tc by collapsing all edges
whose stabilizer does not belong to A (Warning: Tc is not always an A-tree).
It is clear from the equivalent definition that Tc is a tree, but its edge stabilizers do not
always belong to A; this is why we also consider Tc∗ . The next lemma will say that Tc∗ is
relative to H (i.e. it is an (A, H)-tree).
The trees Tc and Tc∗ are minimal. We claim that they are always irreducible or trivial
(i.e. they consist of a single point). Indeed, if Tc or Tc∗ is non-trivial and not irreducible,
then so is T because it is compatible with them (Lemma 7.14 below). This implies that
there is only one cylinder in T , so Tc and Tc∗ are trivial, a contradiction.
The stabilizer of a vertex v ∈ V0 (Tc ) is the stabilizer of v, viewed as a vertex of T ; it
does not belong to A. The stabilizer GY of a vertex Y ∈ V1 (Tc ) is the stabilizer of the
equivalence class [Y ]. The stabilizer of an edge ε = (v, Y ) of Tc is Gε = Gv ∩ GY ; it is
elliptic in T , and infinite because it contains Ge if e ⊂ T is any edge of Y with origin v. If
Gε lies in A∞ , it is a representative of [Y ].
Lemma 7.3.
(1) T dominates Tc and Tc∗ , so Tc∗ is an (A, H)-tree.
(2) If H < G is a vertex stabilizer of Tc which is not elliptic in T , it is the stabilizer of
the equivalence class [Y ] associated to some cylinder Y ⊂ T . If e is an edge of Y ,
the equivalence class of Ge is H-invariant (hGe h−1 ∼ Ge for h ∈ H).
(3) Tc and Tc∗ only depend on the deformation space D containing T (we sometimes say
that Tc is the tree of cylinders of D).
(4) Suppose that the stabilizer of every equivalence class [A] belongs to A. Then edge
stabilizers of Tc belong to A, and therefore Tc = Tc∗ .
(5) The tree Tc∗ is equal to its collapsed tree of cylinders: (Tc∗ )∗c = Tc∗ .
Assertion (4) applies in particular to Examples (1) and (2) above (CSA groups, relatively hyperbolic groups).
Proof. Consider a vertex v of T . If v belongs to two cylinders, it defines a vertex in V0 (Tc ),
and this vertex is fixed by Gv . If v belongs to a single cylinder Y , then Gv fixes the vertex
of V1 (Tc ) corresponding to Y . This shows that T dominates Tc , hence also its collapse Tc∗ .
The second assertion follows from remarks made above: the stabilizer of a vertex in
V0 (Tc ) is a vertex stabilizer of T , the stabilizer of a vertex in V1 (Tc ) is the stabilizer of an
equivalence class [Y ]. Also note that (4) is clear since the stabilizer of an edge ε = (v, Y )
of Tc is contained in the stabilizer of [Y ]. Assertion (5) is Corollary 5.8 of [GL11].
[GL11] contains three proofs of the third assertion. We sketch one. A domination map
f : T → T 0 induces a map fc : Tc → Tc0 mapping each edge onto a vertex or an edge. The
image of x ∈ V0 (Tc ) by fc is the unique point of T 0 fixed by Gx , viewed as a vertex of
V0 (Tc0 ); the image of Y ∈ V1 (Tc ) is either the unique cylinder whose edge stabilizers are
75
equivalent to those of Y , viewed as a vertex of V1 (Tc0 ), or the unique point of T 0 fixed by
stabilizers of edges of Y , viewed as a vertex of V0 (Tc0 ). If T and T 0 belong to the same
deformation space, the map gc : Tc0 → Tc induced by a map g : T 0 → T is the inverse of fc ,
so Tc0 = Tc .
Assertion (3) implies that trees of cylinders are canonical elements of their deformation
space. In particular:
Corollary 7.4. If T is a JSJ tree over A relative to H, then Tc and Tc∗ are invariant
under any automorphism of G which preserves A and H.
Z
b
Z
b
Z
b
Z
Z
Z
b
Z
Z2
Z
Figure 11: A tree and its tree of cylinders.
In the following examples, ∼ may be viewed at will as commutation, or co-elementarity,
or commensurability. Indeed, two edge stabilizers will be equivalent if and only if they are
equal.
Example 7.5. The Bass-Serre tree of the graph of groups Γ of Subsection 3.9 (three punctured tori glued along their boundaries, see Figure 3 in the introduction) is equal to its
tree of cylinders (cylinders are tripods projecting bijectively onto Γ).
Example 7.6. Now let T be the Bass-Serre tree of the graph of groups pictured on the left
hand side of Figure 11 (reproducing Figure 2). There are four QH vertices vi , which are
fundamental groups of punctured tori Σi . There is an edge between vi and vi+1 (mod 4),
each carrying the same group Ge , equal to the boundary subgroup of Σi . Any cylinder Y
is a line, with vertices ui (i ∈ Z) which are lifts of vi (mod 4). The setwise stabilizer of Y ,
which we shall call P , is isomorphic to Z2 , it acts on Y by translations with edge stabilizers
equal to Ge (up to conjugacy). The group G is hyperbolic relative to this subgroup P (it
is a toral relatively hyperbolic group).
In Tc∗ (which is equal to Tc ), each line Y is collapsed to a point u (a vertex of type
V1 ) and is replaced by edges joining that point to each of the ui ’s. The stabilizer of u is
isomorphic to P , which is the centralizer (and the commensurator) of Ge . Unlike T , the
tree Tc∗ is invariant under all automorphisms of G.
7.2
Acylindricity
Recall that T is (k, C)-acylindrical if the pointwise stabilizer of every arc of length ≥ k + 1
has cardinality ≤ C (see the beginning of part IV).
Lemma 7.7. Assume that there exists C > 0 such that, if two groups of A∞ are inequivalent, then their intersection has order ≤ C. If T is any (A, H)-tree, then Tc∗ is a
(2, C)-acylindrical (A, H)-tree.
This applies in particular in Examples (1) and (2) above (CSA groups, relatively hyperbolic groups); note that in a relatively hyperbolic group there is a bound C for the order
of finite non-parabolic subgroups (see Lemma 9.15). In Example 7.6, the tree T is not
acylindrical (cylinders are lines fixed by an infinite cyclic group) but Tc∗ is 2-acylindrical.
76
Proof. We have seen that Tc∗ is an (A, H)-tree. Let I be a segment of length 3 in Tc∗ . Its
edges are images of edges εi = (xi , Yi ) of Tc . Let ei be an edge of Yi with origin xi . The
stabilizer Gεi of εi belongs to A and contains Gei , so Gεi belongs to A∞ and is equivalent
to Gei .
Since Tc∗ is a collapse of Tc , we can find i, j with Yi 6= Yj . This implies Gei 6∼ Gej , so
Gεi 6∼ Gεj . The stabilizer of I fixes εi and εj , so has cardinality at most C.
Virtually cyclic groups play a particular role in this context. An infinite group H is
virtually cyclic if some finite index subgroup is infinite cyclic. Such a group maps onto Z
or D∞ (the infinite dihedral group Z/2 ∗ Z/2) with finite kernel.
Definition 7.8 (C-virtually cyclic). Given C ≥ 1, we say that H is C-virtually cyclic if
it maps onto Z or D∞ with kernel of order at most C. Equivalently, H acts non-trivially
on a simplicial line with edge stabilizers of order ≤ C.
An infinite virtually cyclic group is C-virtually cyclic if the order of its maximal finite
normal subgroup is ≤ C. On the other hand, finite subgroups of a C-virtually cyclic group
have cardinality bounded by 2C.
Recall (Subsection 1.5.2) that J < G is small in T (resp. in (A, H)-trees) if it fixes a
point, or an end, or leaves a line invariant in T (resp. in all (A, H)-trees on which G acts).
A group not containing F2 , or contained in a group of H, is small in (A, H)-trees.
The following lemma is simple, but conceptually important. It says that subgroups
which are small but not virtually cyclic are elliptic in acylindrical trees.
Lemma 7.9. If a subgroup J ⊂ G is small in a (k, C)-acylindrical tree T , but J is not
elliptic in T , then J is C-virtually cyclic.
Proof. By smallness, J preserves a line or fixes an end. The result is clear if it acts on
a line, since edge stabilizers for this action have order ≤ C. If it fixes an end, the set
of elliptic elements of J is the kernel J0 of a homomorphism from J to Z. Every finitely
generated subgroup of J0 is elliptic. It fixes a ray, so has order ≤ C by acylindricity. Thus
J0 has order ≤ C.
7.3
Small domination
We have seen that T dominates Tc and Tc∗ (Lemma 7.3), and that Tc∗ is (2, C)-acylindrical
under the assumptions of Lemma 7.7. This domination may be strict: in particular, as
shown by Lemma 7.9, small groups which are not virtually cyclic tend to be elliptic in
Tc and Tc∗ . For instance, in Example 7.6, the subgroup P = Z2 becomes elliptic in Tc∗ .
This is unavoidable according to Lemma 7.9, but P (together with its subgroups and their
conjugates) is the only group that is elliptic in Tc∗ and not in T .
We want the deformation space of Tc∗ to be as close to that of T as possible, as in this
example, and this motivates the following definition; it applies to an arbitrary pair of trees
(T, T ∗ ).
Definition 7.10 (smally dominates). Let T, T ∗ be (A, H)-trees. We say that T smally
dominates T ∗ if:
(i) T dominates T ∗ ;
(ii) edge stabilizers of T ∗ are elliptic in T .
(iii) any group which is elliptic in T ∗ but not in T is small in (A, H)-trees.
More generally, let S be a family of subgroups closed under conjugation and taking
subgroups, with every J ∈ S small in (A, H)-trees. If T smally dominates T ∗ , and every
group which is elliptic in T ∗ but not in T belongs to S, we say that T S-dominates T ∗ .
77
Remark 7.11. S-domination will be useful to describe flexible vertex stabilizers of JSJ trees
(as in the proof of Theorem 9.5).
Note that the first two conditions of the definition are always satisfied if T is a tree
and T ∗ is its collapsed tree of cylinders Tc∗ . The following proposition basically says that
the third condition holds when the groups G[A] are small in (A, H)-trees. In this case we
can make every tree acylindrical without changing its deformation space more than forced
by Lemma 7.9.
Proposition 7.12. Let ∼ be an admissible equivalence relation on A∞ , and let C be an
integer. Assume that:
(1) if two groups of A∞ are inequivalent, their intersection has order ≤ C;
(2) every stabilizer G[A] is small in (A, H)-trees;
(3) one of the following holds:
(a) every stabilizer G[A] belongs to A;
(b) if A ⊂ A0 has index 2, and A ∈ A, then A0 ∈ A;
(c) no group G[A] maps onto D∞ .
If T is any (A, H)-tree with infinite edge stabilizers, then Tc∗ is a (2, C)-acylindrical
(A, H)-tree smally dominated by T .
Assume furthermore that all subgroups which are small in (A, H)-trees but not virtually
cyclic are elliptic in T . Then Tc∗ belongs to the same deformation space as T .
Proof. Acylindricity comes from Lemma 7.7, and conditions (i) and (ii) of small domination
are always satisfied by pairs (T, Tc ) and (T, Tc∗ ).
Vertex stabilizers of Tc are vertex stabilizers of T or equal to some G[A] , so by Assumption (2) subgroups elliptic in Tc satisfy condition (iii). In other words, T would smally
dominate Tc if Tc were an A-tree. To conclude that Tc∗ is smally dominated by T , we show
that Tc and Tc∗ belong to the same deformation space under Assumptions (2) and (3) (see
[GL11, Remark 5.11]).
Note that Tc∗ = Tc when (3a) holds, so we may assume that (3b) or (3c) holds. Let
ε = (x, Y ) be an edge of Tc such that Gε ∈
/ A. It suffices to prove that Gε = GY (so
that collapsing the orbit of ε does not change the deformation space) and that Gε0 ∈ A for
every edge ε0 = (x0 , Y ) with x0 6= x (so no further collapse occurs in the star of the vertex
Y of Tc ).
The group GY is small in (A, H)-trees by (2). We claim that it is elliptic in T . Assume
otherwise, and consider its subgroup Gε , which is elliptic in T . If GY fixes an end of T ,
then Gε fixes a ray, so Gε ∈ A, a contradiction. The remaining possibility (which is ruled
out if (3c) holds) is that GY acts dihedrally on a line. In this case some subgroup of Gε of
index at most 2 fixes an edge, so is in A, contradicting (3b).
We have proved that GY is elliptic in T , hence fixes a (unique) vertex v ∈ Y ⊂ T . We
claim x = v. If x 6= v, let e be the initial edge of the segment [x, v]. Then
Gε = Gx ∩ GY ⊂ Gx ∩ Gv ⊂ Ge ,
contradicting Gε ∈
/ A. Thus x = v, so ε = (v, Y ) is the only edge of Tc incident to Y ∈
V1 (Tc ) with stabilizer not in A. Moreover, since GY fixes x, we have Gε = Gx ∩ GY = GY .
This proves that Tc and Tc∗ lie in the same deformation space, and that T smally dominates
Tc∗ .
For the furthermore, we need to show that every GY is elliptic in T . By Assumption
(2) it is small in (A, H)-trees, so the only case to consider is when GY is virtually cyclic. If
e is any edge of T contained in Y , its stabilizer Ge is infinite so is a finite index subgroup
of GY . It follows that GY is elliptic in T .
78
Remark 7.13. Given S as in Definition 7.10, assume that every G[A] belongs to S. Then
T S-dominates Tc∗ . If all groups of S which are not virtually cyclic are elliptic in T , then
Tc∗ belongs to the same deformation space as T . The proof is exactly as above.
7.4
Compatibility
The tree of cylinders has strong compatibility properties, which will be useful to construct
the compatibility JSJ tree (see Theorem 11.4).
The following fact is a general property of the tree of cylinders. Recall that two trees
are compatible if they have a common refinement.
Lemma 7.14 ([GL11, Proposition 8.1]). Tc and Tc∗ are compatible with any (A, H)-tree
dominated by T .
We will also need the following more technical statement.
Lemma 7.15. Let G be one-ended relative to H (see Subsection 1.3). Let ∼ be an admissible equivalence relation on A∞ such that every stabilizer G[A] (in particular, every
group in A) is small in (A, H)-trees. Let C be an integer, and suppose that A contains all
C-virtually cyclic groups.
Let S, T be (A, H)-trees, with S refining T . Assume that each vertex stabilizer Gv of T
is small in S (possibly elliptic) or QH with finite fiber of cardinality ≤ C. Then Sc∗ refines
Tc∗ .
The assumption on vertex stabilizers of T holds in particular if T is a JSJ decomposition
whose flexible vertices are small or QH with fiber of cardinality ≤ C.
When S dominates T , any map f : S → T sends cylinder to cylinder and induces a
cellular map f ∗c from Sc∗ to T ∗c (it maps a vertex to a vertex, an edge to a vertex or an
edge, and is independent of f ) [GL11, Lemma 5.6]. The point of the lemma is that f ∗c is
a collapse map if f is.
Proof. One passes from S to T by successively collapsing orbits of edges, in an arbitrary
order. Starting from S, perform collapses which do not change the deformation space
as long as possible. Since trees in the same deformation space have the same tree of
cylinders, this allows us to assume that no proper collapse of S refining T belongs to the
same deformation space as S. This ensures that, for each vertex v of T , the action of Gv
on the preimage Sv of v in S is minimal (Sv may be a point).
Fix v, and consider the tree S 0 obtained from S by collapsing all edges mapped to a
vertex of T not in the orbit of v. Thus Sv embeds in S 0 , and we view Sv as the minimal
subtree of Gv in both S or S 0 . We show that Sc0 ∗ refines Tc∗ , and vertex stabilizers of S 0 are
small in S or QH with fiber of cardinality ≤ C. The lemma then follows by an inductive
argument, considering S and S 0 . We may assume that Sv is not a point (otherwise S 0 = T ).
First suppose that Gv is small in S. This implies that Sv is a line or a subtree with a Gv fixed end, so all edges in Sv belong to the same cylinder of S 0 . In particular, Gv is elliptic
f1
f2
in the tree of cylinders of S 0 , hence in Sc0∗ . We thus have domination maps S 0 −→ T −→ Sc0∗ .
As recalled above, by [GL11, Lemma 5.6], these maps induce (equivariant) cellular maps
f∗
f∗
0∗
Sc0∗ −−1c
→ Tc∗ −−2c
→ (Sc0∗ )0∗
c = Sc between collapsed trees of cylinders (see Lemma 7.3 for
0∗
∗
∗
the equality (Sc0∗ )0∗
c = Sc ). The maps f1c and f2c cannot increase translation lengths,
so it follows from Theorem A.5 that they are isomorphisms (one may also show directly
that they are injective on any segment joing two vertices with stabilizer not in A). Thus
Tc∗ = Sc0∗ . Vertex stabilizers of S 0 are small in S or QH because they are vertex stabilizers
of T or contained in a conjugate of Gv .
The second case is when v is QH, with finite fiber F and underlying orbifold Σ. Consider
the action of Gv on Sv . By one-endedness, every component of ∂Σ is used (Lemma 5.16),
79
so by Lemma 5.18 the action is dominated by an action dual to a family of geodesics L on
Σ. It is in fact equal to such an action because edge stabilizers are small in (A, H)-trees
by assumption, hence virtually cyclic by Assertion (2) of Proposition 5.21 (the assumption
that C-virtually cyclic groups are in A ensures that the groups Qγ are in A). In particular,
vertices in Sv are QH with fiber F .
We claim that any cylinder of S 0 containing an edge of the preimage Sv of v is entirely
contained in Sv . This implies that Sc0 refines Tc by Remark 4.13 of [GL11], and therefore
Sc0∗ refines Tc∗ , thus completing the proof.
Let e be an arbitrary edge in Sv . Given an edge f of S 0 with Gf ∼ Ge , we have to
prove f ⊂ Sv . Suppose not. Cylinders being connected, we can assume that f has an
endpoint in Sv (so is not collapsed in T ). Since v is QH with finite fiber F , and Gf is
infinite by one-endedness, the image of Gf in the orbifold group π1 (Σ) = Gv /F is infinite,
and therefore contained with finite index in a boundary subgroup.
There is a geodesic γ ∈ L such that Ge contains the preimage Gγ of the fundamental
group of γ. Since Ge ∼ Gf , we have hGγ , Gf i ⊂ hGe , Gf i ⊂ G[Ge ] , so hGγ , Gf i is small in
(A, H)-trees. This contradicts the second assertion of Proposition 5.21 since hGγ , Gf i is
not virtually cyclic. This proves the claim, hence the lemma.
8
Constructing JSJ decompositions using acylindricity
Using acylindrical accessibility we show (Subsection 8.1) that one may construct the JSJ
deformation space of G over A relative to H, and describe its flexible subgroups, provided
that every deformation space contains an acylindrical tree (with uniform constants). In
Subsection 8.2 we prove the same results under the weaker assumption that every tree T
smally dominates an acylindrical tree T ∗ . In Section 9 we will combine this with Proposition 7.12, which ensures the existence of such trees T ∗ .
As already mentioned, acylindrical accessibility bounds the number of orbits of edges of
acylindrical trees. This does not prevent the existence of infinite sequences of refinements.
For example, consider G = A∗B with A, B torsion-free hyperbolic groups. Fix b ∈ B \{1},
k
and define Ck = hb2 i and Ak = hA, Ck i. Let Tk be the Bass-Serre tree of the amalgam G =
Ak ∗Ck B. It is 2-acylindrical and can be refined into the splitting G = Ak+1 ∗Ck+1 Ck ∗Ck B.
This refinement is not 2-acylindrical but it belongs to the same deformation space as the
2-acylindrical tree Tk+1 . The trees Tk all lie in distinct deformation spaces. Although this
is true in this example, it is not obvious in general that G splits over the intersection of
the groups Ck (or some conjugates).
The proof in Subsection 8.1.1 therefore uses not only acylindrical accessibility but also
arguments from Sela’s proof [Sel97a], involving actions on R-trees obtained by taking limits
of splittings similar to the Tk ’s considered above. We refer the reader to the appendix for
some basic facts about R-trees. In particular, we will use compactness of the space of
projectivized length functions [CM87], and Paulin’s theorem [Pau89] that the Gromov
topology and the axes topology agree on the space of irreducible R-trees (and even on the
space of semisimple trees, see Theorem A.8).
As usual, A is a family stable under conjugation and taking subgroups. The family H
may be arbitrary, and G is only assumed to be finitely generated (it would be enough to
assume that G is finitely generated relative to a finite collection of subgroups, and these
subgroups are in H).
8.1
Uniform acylindricity
Recall that a tree T is (k, C)-acylindrical if all segments of length ≥ k + 1 have stabilizer of
cardinality ≤ C. If T 0 belongs to the same deformation space as T , it is (k 0 , C)-acylindrical
for some k 0 (see [GL07a]), but in general there is no control on k 0 .
80
The main assumption of this section is the following. We fix k and C, and we assume
that for each (A, H)-tree T there is a (k, C)-acylindrical tree T ∗ in the same deformation
space. Our goal is to deduce the existence of the JSJ deformation space. This section is
a first step towards Theorem 8.7, where we allow T ∗ not to lie in the same deformation
space as T .
For example, suppose that G is a one-ended torsion-free CSA group (for instance, a
toral relatively hyperbolic group), and A is the class of abelian subgroups. One can take
T ∗ := Tc = Tc∗ to be the tree of cylinders as in the first example of Subsection 7.1, provided
that G contains no non-cyclic abelian subgroups, or more generally that these subgroups
are contained in a group of H (up to conjugacy): this guarantees that Tc lies in the same
deformation space as T (see Proposition 7.12). The general case (no assumption on abelian
subgroups) will be taken care of in Subsection 8.2.
Recall (Definition 7.8) that a group is C-virtually cyclic for some C > 0 if it maps to
Z or D∞ with kernel of cardinality ≤ C. Also recall that G is only assumed to be finitely
generated, and there is no restriction on H (it can be any collection of subgroups).
Theorem 8.1. Given A and H, suppose that there exist numbers C and k such that:
• A contains all C-virtually cyclic subgroups of G, and all subgroups of order ≤ 2C;
• for any (A, H)-tree T , there is an (A, H)-tree T ∗ in the same deformation space which
is (k, C)-acylindrical.
Then the JSJ deformation space of G over A relative to H exists.
Moreover, if all groups in A are small in (A, H)-trees, then the flexible vertices are QH
with fiber of cardinality at most C.
See Subsections 1.5, 2.3 and 5.2 for the definitions of small, flexible and QH.
Theorem 8.1 will be proved in Subsections 8.1.1 and 8.1.2. We start with a general
lemma.
Lemma 8.2. If a finitely generated group G does not split over subgroups of order ≤ C
relative to a family H, there exists a finite family H0 = {H1 , . . . , Hp }, with Hi a finitely
generated group contained in a group of H, such that G does not split over subgroups of
order ≤ C relative to H0 .
A special case of this lemma is proved in [Per11].
Proof. All trees in this proof will have stabilizers in the family A(C) consisting of all
subgroups of order ≤ C. Note that over A(C) all trees are universally elliptic, so having
no splitting is equivalent to the JSJ deformation space being trivial.
Let H1 , . . . , Hn , . . . be an enumeration of all finitely generated subgroups of G contained
in a group of H. We have pointed out in Subsection 3.3 that by Linnell’s accessibility G admits JSJ decompositions over A(C), so let Tn be a JSJ tree relative to Hn = {H1 , . . . , Hn }.
We show the lemma by proving that Tn is trivial for n large.
By Lemma 2.15 the tree Tn , which is relative to Hn−1 , may be refined to a JSJ tree
relative to Hn−1 . If we fix k, we may therefore find trees S1 (k), . . . , Sk (k) such that Si (k)
is a JSJ tree relative to Hi and Si (k) refines Si+1 (k). By Linnell’s accessibility theorem,
there is a uniform bound for the number of orbits of edges of S1 (k) (assumed to have no
redundant vertices). As Ti and Si (k) are in the same deformation space, this number is an
upper bound for the number of i0 s such that Ti and Ti+1 belong to different deformation
spaces, so for n large the trees Tn all belong to the same deformation space D (they have
the same elliptic subgroups).
Since every Hn is elliptic in D, so is every H ∈ H (otherwise H would fix a unique end,
and edge stabilizers would increase infinitely many times along a ray going to that end).
The non-splitting hypothesis made on G then implies that D is trivial.
81
We also note:
Lemma 8.3. It suffices to prove Theorem 8.1 under the additional hypothesis that G does
not split over subgroups of order ≤ 2C relative to H (“one-endedness condition”).
Proof. Let A(2C) ⊂ A denote the family of subgroups of order ≤ 2C. As mentioned in
the previous proof, Linnell’s accessibility implies the existence of a JSJ tree over A(2C)
relative to H. We can now apply Proposition 4.15.
8.1.1
Existence of the JSJ deformation space
Because of Lemma 8.3, we assume from now on that G does not split over subgroups of
order ≤ 2C relative to H.
In this subsection we prove the first assertion of Theorem 8.1 (flexible vertices will be
studied in the next subsection). We have to construct a universally elliptic tree TJ which
dominates every universally elliptic tree (of course, all trees are (A, H)-trees and universal
ellipticity is defined with respect to (A, H)-trees). Countability of G allows us to choose a
sequence of universally elliptic trees Ui such that, if g ∈ G is elliptic in every Ui , then it is
elliptic in every universally elliptic tree. By Assertion (2) of Lemma 2.6, it suffices that TJ
dominates every Ui . Inductively replacing each Ui by its standard refinement dominating
Ui−1 , as in the proof of Theorem 2.16, we may assume that Ui dominates Ui−1 for all i. In
particular, we are free to replace Ui by a subsequence when needed.
Let Ti be a (k, C)-acylindrical tree in the same deformation space as Ui , with distance
function denoted by di (we view Ti as a metric tree, with all edges of length 1). Let
`i : G → Z be the translation length function of Ti (see the appendix). The proof has
two main steps. First we assume that, for all g, the sequence `i (g) is bounded, and we
construct a universally elliptic (A, H)-tree TJ which dominates every Ui . Such a tree is
a JSJ tree. In the second step, we deduce a contradiction from the assumption that the
sequence `i is unbounded.
• If `i (g) is bounded for all g, we pass to a subsequence so that `i has a limit ` (possibly
0). Since the set of length functions of trees is closed [CM87, Theorem 4.5], ` is the length
function associated to the action of G on an R-tree T . It takes values in Z, so the tree T
is simplicial (possibly with edges of length 12 , see Example A.4). In general T is not an
A-tree, but we will show that it is relative to H.
We can assume that all trees Ti are non-trivial. By Lemma 7.9, Ti is irreducible except
if G is virtually cyclic, in which case the theorem is clear. If g ∈ G is hyperbolic in some
Ti0 , then `i (g) ≥ 1 for all i ≥ i0 (because Ti dominates Ti0 ), so g is hyperbolic in T . Since
every Ti is irreducible, there exist g, h ∈ G such that g, h, and [g, h] are hyperbolic in Ti
(for some i), hence in T , so T is also irreducible. By [Pau89], Ti converges to T in the
equivariant Gromov-Hausdorff topology (see Theorem A.6).
We will not claim anything about the edge stabilizers of T , but we study its vertex
stabilizers. We claim that a subgroup H ⊂ G is elliptic in T if and only if it is elliptic in
every Ti ; in particular, every H ∈ H is elliptic in T , and T dominates every Ti .
The claim is true if H is finitely generated, since g ∈ G is elliptic in T if and only if it
is elliptic in every Ti . If H is infinitely generated, fix a finitely generated subgroup H 0 of
cardinality > C. If H is elliptic in T , it is elliptic in Ti because otherwise it fixes a unique
end and H 0 fixes an infinite ray of Ti , contradicting acylindricity of Ti . Conversely, if H is
elliptic in every Ti but not in T , the group H 0 fixes an infinite ray in T . By convergence in
the Gromov topology, H 0 fixes a segment of length k + 1 in Ti for i large. This contradicts
acylindricity of Ti , thus proving the claim.
We now return to the trees Ui . They are dominated by T since the trees Ti are. Their
edge stabilizers are in Aell , the family of groups in A which are universally elliptic. Since
Aell is stable under taking subgroups, any equivariant map f : T → Ui factors through the
82
tree TJ obtained from T by collapsing all edges with stabilizer not in Aell . The tree TJ is
a universally elliptic (A, H)-tree dominating every Ui , hence every universally elliptic tree.
It is a JSJ tree.
• We now suppose that `i (g) is unbounded for some g ∈ G, and we work towards
a contradiction. Since the set of projectivized non-zero length functions on a finitely
generated group is compact [CM87, Theorem 4.5], we may assume that `i /λi converges to
the length function ` of a non-trivial R-tree T∞ , for some sequence λi → +∞. By Theorem
A.8, Ti /λi converges to T in the Gromov topology (all Ti ’s are irreducible, and we take
T∞ to be a line if it is not irreducible).
Lemma 8.4.
(1) Any subgroup H < G of order > C which is elliptic in every Ti fixes a unique point
in T∞ . In particular, elements of H are elliptic in T∞ .
(2) Tripod stabilizers of T∞ have cardinality ≤ C.
(3) If I ⊂ T∞ is a non-degenerate arc, its stabilizer GI has order ≤ C or is C-virtually
cyclic.
(4) Let J ⊂ I ⊂ T∞ be two non-degenerate arcs. If GI is C-virtually cyclic, then
GI = GJ .
Proof. Recall that Ti+1 dominates Ti , so a subgroup acting non-trivially on Ti also acts
non-trivially on Tj for j > i.
To prove (1), we may assume that H is finitely generated. It is elliptic in T∞ because
all of its elements are. But it cannot fix an arc in T∞ : otherwise, since Ti /λi converges to
T∞ in the Gromov topology, H would fix a long segment in Ti for i large, contradicting
acylindricity.
Using the Gromov topology, as in Example A.7, one sees that a finitely generated
subgroup fixing a tripod of T∞ fixes a long tripod of Ti for i large, so has cardinality at
most C by acylindricity. This proves (2).
To prove (3), consider a group H fixing a non-degenerate arc I = [a, b] in T∞ . We write
|I| for the length of I. It suffices to show that H is small: depending on whether H is
elliptic in every Ti or not, Assertion (1) or Lemma 7.9 then gives the required conclusion.
If H contains a non-abelian free group, choose elements {h1 , . . . , hn } generating a free
subgroup Fn ⊂ H of rank n C, and choose ε > 0 with ε |I|. For i large, there
exist approximations ai , bi of a, b in Ti at distance at least (|I| − ε)λi from each other, and
contained in the characteristic set (axis or fixed point set) of each hj . Additionally, the
translation length of every hj in Ti is at most ελi if i is large enough. Then all commutators
[hj1 , hj2 ] fix most of the segment [ai , bi ], contradicting acylindricity of Ti if n(n − 1)/2 > C.
We now prove (4). By Assertion (1), we know that GI acts non-trivially on Ti for i
large, so we can choose a hyperbolic element h ∈ GI . We suppose that some g ∈ GJ
does not fix an endpoint, say a, of I = [a, b], and we argue towards a contradiction. Let
ai , bi ∈ Ti be in the axis of h as above, with di (ai , gai ) ≥ δλi for some δ > 0. For i large
i
the translation lengths of g and h in Ti are small compared to |J|λ
C , and the elements
2
C+1
[g, h], [g, h ], . . . , [g, h
] all fix a common long arc in Ti . By acylindricity, there exist
j1 6= j2 such that [g, hj1 ] = [g, hj2 ], so g commutes with hj1 −j2 . It follows that g preserves
the axis of h, and therefore moves ai by `i (g), a contradiction since `i (g)/λi goes to 0 as
i → ∞.
Sela’s proof of acylindrical accessibility [Sel97a] now comes into play to describe the
structure of T∞ . We use the generalization given by Theorem 5.1 of [Gui08], which allows
non-trivial tripod stabilizers. Lemma 8.4 shows that stabilizers of unstable arcs and tripods
have cardinality at most C, and we have assumed that G does not split over a subgroup of
83
cardinality at most 2C relative to H, hence also relative to a finite family H0 by Lemma
8.2. It follows that T∞ is a graph of actions as in Theorem 5.1 of [Gui08]. In order to
reach the desired contradiction, we have to rule out several possibilities.
First consider a vertex action Gv y Yv of the decomposition of T∞ given by [Gui08].
If Yv is a line on which Gv acts with dense orbits through a finitely generated group, then,
by Assertion (3) of Lemma 8.4, Gv contains a finitely generated subgroup H mapping
onto Z2 with finite or virtually cyclic kernel, and acting non-trivially. The group H acts
non-trivially on Ti for i large, contradicting Lemma 7.9.
Now suppose that Gv y Yv has kernel Nv , and the action of Gv /Nv is dual to an
arational measured foliation on a 2-orbifold Σ (with conical singularities). Then Nv has
order ≤ C since it fixes a tripod. Consider a one-edge splitting S of G (relative to H)
dual to a simple closed curve on Σ. This splitting is over a C-virtually cyclic group Ge . In
particular, S is an (A, H)-tree. Since Ge is hyperbolic in T∞ , it is also hyperbolic in Ti ,
hence in Ui , for i large enough. On the other hand, being universally elliptic, Ui is elliptic
with respect to S. By Remark 2.3 of [FP06] (see Assertion (3) of Lemma 2.6), G splits
relative to H over an infinite index subgroup of Ge , i.e. over a group of order ≤ 2C (in fact
≤ C in this case), contradicting our assumptions.
By Theorem 5.1 of [Gui08], the only remaining possibility is that T∞ itself is a simplicial
tree, and all edge stabilizers are C-virtually cyclic. Then T∞ is an (A, H)-tree, and its edge
stabilizers are hyperbolic in Ti for i large. This leads to a contradiction as in the previous
case, and concludes the proof of the first assertion of Theorem 8.1.
8.1.2
Description of flexible vertices
Now that we know that the JSJ decomposition exists, we prove the second assertion of
Theorem 8.1: if all groups in A are small in (A, H)-trees, then the flexible vertices are QH
with fiber of cardinality at most C.
The arguments are similar to those used in Section 6, with one key difference: since we
do not assume finite presentability of G, we do not have Lemma 6.12 constructing a tree
which is maximal for domination using Dunwoody’s accessibility. We shall use acylindrical
accessibility instead.
As in Subsection 6.2 we may assume that G is totally flexible, i.e. that there exist nontrivial (A, H)-trees, but none of them is universally elliptic. Indeed, to prove that a flexible
vertex Gv of the JSJ decomposition of G is QH, it is enough to study (Gv , IncH
v ) instead
of (G, H). Note that we do not know in advance that JSJ trees have finitely generated
edge stabilizers, so we have to allow infinitely generated groups in IncH
v (even if groups in
H are finitely generated).
The fact that G is totally flexible implies that any edge stabilizer A of any tree is
C-virtually cyclic: since A is small in (A, H)-trees, this follows from Lemma 7.9 applied
to the action of A on T ∗ , where T is some (A, H)-tree in which A is not elliptic. We
may therefore replace A by the family consisting of all C-virtually cyclic subgroups and
all finite subgroups. It satisfies the stability condition (SC) of Definition 6.1, with F the
family of subgroups of order ≤ C, so we are free to use the results of Subsections 6.3 and
6.4, as well as Proposition 6.28 (splittings are clearly minuscule, so Subsection 6.7 is not
needed).
By acylindrical accessibility [Wei12], there is a bound on the number of orbits of edges
of a (k, C)-acylindrical tree. Among all (k, C)-acylindrical trees with no redundant vertex,
consider a tree U whose number of orbits of edges is maximal (this is a substitute for the
maximal tree provided by Lemma 6.12). Proposition 6.28 yields V such that U and V
are fully hyperbolic with respect to each other, and we let R = RN (U, V ) be their regular
neighborhood (Proposition 6.25).
Recall (Proposition 6.25 (RN1)) that the stabilizer Gv of a vertex v ∈ S is QH or
84
slender. If slender, it maps onto a group that is virtually Z2 . On the other hand, it must
be virtually cyclic by Lemma 7.9 because it acts hyperbolically on U . This contradiction
shows that Gv must be QH (with fiber of cardinality ≤ C) if v ∈ S. As in the proof of
Theorem 6.8, it thus suffices to show that R is a point.
By (RN3) one obtains U from R by refining R at vertices v ∈ S and collapsing all other
edges. We let S be a common refinement of R and U such that no edge of S is collapsed in
both R and U (it is the lcm S = R ∨ U of Subsection A.5). Let S ∗ be a (k, C)-acylindrical
tree with no redundant vertex in the deformation space of S.
We now consider commensurability classes of edge stabilizers.
Since U is dual to geodesics in QH vertices of R by (RN3), and has no redundant vertex,
edges of U in different orbits have non-commensurable stabilizers. Every edge stabilizer
of U is also an edge stabilizer of S. Next observe that, for any edge e of S, there exists
an edge e0 of S ∗ such that Ge and Ge0 are commensurable. Indeed, since S ∗ dominates
S, there exists e0 with Ge0 ⊂ Ge . By one-endedness (Lemma 8.3), the cardinality of Ge0
is greater than 2C. Since Ge is C-virtually cyclic, this implies that Ge0 is infinite, hence
commensurable to Ge .
Given a tree T , denote by |T | the number of orbits of edges, and by c(T ) the number of
equivalence classes of orbits of edges, where two orbits are equivalent if they contain edges
with commensurable stabilizers. We have just proved |U | ≤ c(U ) ≤ c(S) ≤ c(S ∗ ) ≤ |S ∗ |.
The maximality property of U now implies that these inequalities are equalities, so
every edge stabilizer of S is commensurable to an edge stabilizer of U . If R is not a point,
let A be an edge stabilizer. It is elliptic in V by (RN2). On the other hand, A is an edge
stabilizer of S, so is commensurable to an edge stabilizer of U . This contradicts the full
hyperbolicity of U with respect to V , thus completing the proof of Theorem 8.1.
8.1.3
Splittings over virtually cyclic groups
Before generalizing Theorem 8.1 in the next subsection, we give an application of the
previous arguments. In Section 6 we proved that certain flexible groups are QH when H
is a finite family of finitely generated subgroups and G is finitely presented relative to H.
Acylindricity will allow us to remove these assumptions for splittings over virtually cyclic
groups. This is based on the following lemma. The stability conditions and total flexibility
are defined in Section 6.
Lemma 8.5. Assume that all groups in A are virtually cyclic, and A satisfies one of the
stability conditions (SC) or (SCZ ). If G is totally flexible, then any tree T is (2, C)acylindrical for some C (depending on T ).
Proof. Since groups in A are virtually cyclic, we may assume (by making F smaller if
needed) that all groups in F are finite. Total flexibility implies that G is one-ended
relative to H, so all trees are minuscule (of course, this also follows from Proposition 6.35).
By Proposition 6.28, there exists a tree T 0 such that T and T 0 are fully hyperbolic with
respect to each other. Let R = RN (T, T 0 ) be their regular neighborhood. By (RN3), the
tree T is dual to families of geodesics in the orbifolds underlying the QH vertex groups of
R. It follows that T is (2, C)-acylindrical, for C the maximum order of the corresponding
fibers (see the example in the beginning of Part IV).
This lemma allows us to argue as in the previous subsection, provided that there is a
bound for the order of groups in F. In particular, we get the following strengthening of
Theorem 6.6:
Theorem 8.6. Let G be a finitely generated group, and let H be an arbitrary family of
subgroups. Let A be the class of all finite or cyclic subgroups of G. Assume that there
exists a JSJ tree T over A relative to H.
85
If Q is a flexible vertex stabilizer of T , then Q is virtually Z2 or QH with trivial fiber.
Moreover, the underlying orbifold Σ has no mirror, every boundary component of Σ is used,
and Σ contains an essential simple closed geodesic.
Proof. The argument is the same as in Subsection 8.1.2. One first reduces to the case
when G is totally flexible. Since F = {1}, all trees are 2-acylindrical by Lemma 8.5, so
acylindrical accessibility applies.
8.2
Acylindricity up to small groups
In this section we generalize Theorem 8.1. Instead of requiring that every deformation
space contains a (k, C)-acylindrical tree, we require that every tree smally dominates some
(k, C)-acylindrical tree (Definition 7.10).
Recall that G is only assumed to be finitely generated, and there is no restriction on
H (it can be any collection of subgroups).
Theorem 8.7. Given A and H, suppose that there exist numbers C and k such that:
• A contains all C-virtually cyclic subgroups, and all subgroups of cardinal ≤ 2C;
• every (A, H)-tree T smally dominates some (k, C)-acylindrical (A, H)-tree T ∗ .
Then the JSJ deformation space of G over A relative to H exists.
Assume further that all groups in A are small in (A, H)-trees. Then the flexible vertex
groups that are not small in (A, H)-trees are QH with fiber of cardinality at most C.
More generally, if S is as in Definition 7.10 and T always S-dominates T ∗ , then flexible
vertex groups that do not belong to S are QH with fiber of cardinality at most C.
Z
Z
b
Z
b
Z
b
Z
b
A non-relative abelian JSJ
decomposition T
Z
Z
Z2
Z
Z
Z
Z Z Z
Tr , the canonical relative
abelian JSJ
Z
Ta , the non-relative abelian JSJ
constructed from Tr
Figure 12: JSJ decompositions of a CSA group; relative means relative to non-cyclic abelian
subgroups.
Example 8.8. This applies for instance if G is a toral relatively hyperbolic group (more
generally, a torsion-free CSA group), A is the family of abelian subgroups, and we take for
T ∗ the tree of cylinders Tc (equal to Tc∗ ), with k = 1 and C = 2 (see Section 9 for more
examples).
In particular, let us return to Example 7.6, with H = ∅. Figure 12 shows (the quotient
graphs of groups of) three trees. The first one is an abelian JSJ tree T ; the next one is
its tree of cylinders Tc = Tc∗ = T ∗ , which is also the JSJ tree Tr relative to non-cyclic
abelian subgroups constructed in the proof of Theorem 8.7 (see Corollary 8.12); the last
one is the JSJ tree Ta constructed in the proof, obtained by refining Tr at vertices with
stabilizer Z2 (Lemma 8.14). It is another JSJ tree, in the same deformation space as T ,
and Tr = Tc∗ = (Ta )∗c .
The rest of this section is devoted to the proof of Theorem 8.7. As above, we assume
(thanks to Lemma 8.3) that G does not split over groups of order ≤ 2C relative to H
(“one-endedness” assumption).
86
We write S for the family of all subgroups of G which are small in (A, H)-trees, or for a
family as in Definition 7.10 if T S-dominates T ∗ . We write Snvc for the groups in S which
are not C-virtually cyclic.
Lemma 8.9. Suppose that T S-dominates a (k, C)-acylindrical tree T ∗ .
(1) If a vertex stabilizer Gv of T ∗ is not elliptic in T , it belongs to Snvc (in particular, it
is not C-virtually cyclic).
(2) Let H be a subgroup. It is elliptic in T ∗ if and only if it is elliptic in T or contained
in a group K ∈ Snvc . In particular, all (k, C)-acylindrical trees S-dominated by T
belong to the same deformation space.
(3) Assume that T ∗ is reduced. Then every edge stabilizer Ge of T ∗ has a subgroup of
index at most 2 fixing an edge in T . In particular, if T is universally elliptic, so is
T ∗.
Recall (see Subsection 1.4) that a tree S is reduced if no proper collapse of S lies in
the same deformation space as S. Equivalently, any edge e = uv with u, v in different
orbits satisfies Ge 6= Gu and Ge 6= Gv . Since one may obtain a reduced tree in the same
deformation space as S by collapsing edges, there is no loss of generality in assuming that
the smally dominated trees T ∗ are reduced.
Proof. Clearly Gv ∈ S because T S-dominates T ∗ . Assume that it is C-virtually cyclic.
Stabilizers of edges incident to v have infinite index in Gv since they are elliptic in T and
Gv is not, so they have order ≤ 2C. This contradicts the “one-endedness” assumption (note
that T ∗ is not trivial because then G = Gv would be C-virtually cyclic, also contradicting
one-endedness).
The “if” direction of Assertion (2) follows from Lemma 7.9. Conversely, assume that
H is elliptic in T ∗ but not in T . If v is a vertex of T ∗ fixed by H, we have H ⊂ Gv and
Gv ∈ Snvc by Assertion (1).
For Assertion (3), let u and v be the endpoints of an edge e of T ∗ . First suppose that
Gu is not elliptic in T . It is small in T , so it preserves a line or fixes an end. Since Ge
is elliptic in T , some subgroup of index at most 2 fixes an edge. If Gu fixes two distinct
points of T , then Ge fixes an edge. We may therefore assume that Gu and Gv each fix a
unique point in T .
If these fixed points are different, Ge = Gu ∩Gv fixes an edge of T . Otherwise, hGu , Gv i
fixes a point x in T , and is therefore elliptic in T ∗ . Since T ∗ is reduced, some g ∈ G acting
hyperbolically on T ∗ maps u to v and conjugates Gu to Gv (unless there is such a g,
collapsing the edge uv yields a tree in the same deformation space as T ∗ ). This element g
fixes x, so is elliptic in T ∗ , a contradiction.
Corollary 8.10.
(1) T ∗ is an (A, H ∪ Snvc )-tree.
(2) If T1 dominates T2 , then T1∗ dominates T2∗ .
(3) If T is an (A, H ∪ Snvc )-tree, then T ∗ lies in the same deformation space as T . In
particular, Theorem 8.1 applies to (A, H ∪ Snvc )-trees.
We also deduce:
Corollary 8.11.
• If a subgroup J is small in (A, H ∪ Snvc )-trees, it is also small in (A, H)-trees.
87
• If J ∈
/ S is elliptic in every tree relative to H ∪ Snvc , it is elliptic in every tree which
is only relative to H.
Proof. If J is not small in an (A, H)-tree T , it does not belong to S, and is not elliptic in
T ∗ by Assertion (2) of Lemma 8.9. By Lemma 7.9, it is not small in the (A, H ∪ Snvc )-tree
T ∗.
If J ∈
/ S is not elliptic in some (A, H)-tree T , then by Lemma 8.9 it is not elliptic in
T ∗.
Thanks to the third assertion of Corollary 8.10, we may apply Theorem 8.1 to get:
Corollary 8.12. There exists a JSJ tree Tr over A relative to H ∪ Snvc .
We can assume that Tr is reduced. We think of Tr as relative, as it is relative to Snvc
(not just to H). In Example 8.8, Snvc is the class of all non-cyclic abelian subgroups and
Tr is a JSJ tree over abelian groups relative to all non-cyclic abelian subgroups.
Lemma 8.13. If Tr is reduced, then it is (A, H)-universally elliptic.
Proof. We let e be an edge of Tr such that Ge is not elliptic in some (A, H)-tree T , and we
argue towards a contradiction. We may assume that T only has one orbit of edges. The
first step is to show that Tr dominates T ∗ .
Since T ∗ is relative to H ∪ Snvc , the group Ge fixes a vertex u ∈ T ∗ by (A, H ∪ Snvc )universal ellipticity of Tr . This u is unique because edge stabilizers of T ∗ are elliptic in T .
Also note that Gu ∈ Snvc by Assertion (1) of Lemma 8.9.
We may assume that T ∗ is not a point (if it is, then G ∈ Snvc and Tr itself is a point), so
Gu contains an edge stabilizer of T ∗ , hence the stabilizer of some edge f ⊂ T since there is
an equivariant map T → T ∗ . The group Gf is (trivially) (A, H ∪ Snvc )-universally elliptic
because Gf ⊂ Gu ∈ Snvc . Since T has a single orbit of edges, it is (A, H ∪ Snvc )-universally
elliptic (but it is not relative to H ∪ Snvc ). On the other hand, T ∗ is an (A, H ∪ Snvc )-tree,
and it is (A, H ∪ Snvc )-universally elliptic by Assertion (3) of Lemma 8.9. By maximality
of the JSJ, Tr dominates T ∗ .
Recall that u is the unique fixed point of Ge in T ∗ . Denote by a, b the endpoints of e in
Tr . Since Tr dominates T ∗ , the groups Ga and Gb fix u, so hGa , Gb i ⊂ Gu ∈ Snvc is elliptic
in Tr . As in the proof of Lemma 8.9, some g ∈ G acting hyperbolically on Tr maps a to b
(because Tr is reduced). This g fixes u, so belongs to Gu ∈ Snvc , a contradiction since Tr
is relative to H ∪ Snvc .
We shall now construct a JSJ tree Ta relative to H by refining Tr (a reduced JSJ tree
relative to H ∪ Snvc ) at vertices with small stabilizer. This JSJ tree Ta is thought of as
absolute as it is not relative to Snvc .
Lemma 8.14. There exists a JSJ tree Ta over A relative to H. It may be obtained by
refining Tr at vertices with stabilizer in S (in particular, the set of vertex stabilizers not
belonging to S is the same for Ta as for Tr ). Moreover, Ta∗ lies in the same deformation
space as Tr .
Proof. Let v be a vertex of Tr . We shall prove the existence of a JSJ tree Tv for Gv relative
to the family IncH
v consisting of incident edge groups and subgroups conjugate to a group of
H (see Definition 4.5). By Proposition 4.15, which applies because Tr is (A, H)-universally
elliptic (Lemma 8.13), one then obtains a JSJ tree Ta for G relative to H by refining Tr
using the trees Tv .
If Gv is elliptic in every (A, H)-universally elliptic tree T , its JSJ is trivial (see Lemma
4.13) and no refinement is needed at v. Assume therefore that Gv is not elliptic in such a
T . Consider the (A, H ∪ Snvc )-tree T ∗ . It is (A, H)-universally elliptic by Assertion (3) of
88
Lemma 8.9, hence (A, H ∪ Snvc )-universally elliptic, so it is dominated by Tr . In particular,
Gv is elliptic in T ∗ , so belongs to S.
Let Tv ⊂ T be the minimal Gv -invariant subtree (it exists by Proposition 1.7 and
Lemma 4.7, since the incident edge stabilizers of v in Tr are elliptic in T ). Being small
in (A, H)-trees, Gv has at most one non-trivial deformation space containing a universally
elliptic tree (Proposition 3.4). Applying this to splittings of Gv relative to IncH
v , we deduce
H
that Tv is a JSJ tree of Gv relative to Incv . This shows the first two assertions of the lemma.
We now show the “moreover”. Since Ta∗ is an (A, H ∪ Snvc )-tree, which is (A, H ∪ Snvc )universally elliptic by the third assertion of Lemma 8.9, it is dominated by Tr . Conversely,
Ta dominates Tr and therefore Ta∗ dominates Tr∗ by Corollary 8.10(2), so Ta∗ dominates Tr
since Tr and Tr∗ lie in the same deformation space by Corollary 8.10(3).
Remark 8.15. By Corollary 8.11, the type of vertex stabilizers not in S (rigid or flexible)
is the same in Ta (relative to H) and Tr (relative to H ∪ Snvc ).
We can now conclude the proof of Theorem 8.7. The JSJ deformation space relative to
H exists by Lemma 8.14. The description of flexible vertex groups follows from Theorem
8.1, since JSJ trees relative to H and H ∪ Snvc have the same vertex stabilizers not in S
by Lemma 8.14.
9
Applications
Recall that A∞ is the family of infinite groups in A. Combining Proposition 7.12 and
Theorem 8.7 yields:
Corollary 9.1. Let G be a finitely generated group. Given A and H, let ∼ be an admissible
equivalence on A∞ . Let S be the family of groups contained in some G[A] , with A ∈ A∞ .
Assume that G is one-ended relative to H, and there exists an integer C such that:
(1) A contains all C-virtually cyclic subgroups, and all subgroups of cardinal ≤ 2C;
(2) if two groups of A∞ are inequivalent, their intersection has order ≤ C;
(3) every stabilizer G[A] is small in (A, H)-trees (hence so is every element of S);
(4) one of the following holds:
(a) every stabilizer G[A] belongs to A;
(b) if A ⊂ A0 has index 2, and A ∈ A, then A0 ∈ A;
(c) no group G[A] maps onto D∞ .
Then:
(1) there is a JSJ tree Ta over A relative to H; its collapsed tree of cylinders (Ta )∗c is
a JSJ tree relative to H ∪ Snvc (with Snvc the family of groups in S which are not
C-virtually cyclic);
(2) Ta and (Ta )∗c have the same vertex stabilizers not in S; flexible vertex stabilizers that
do not belong to S are QH with fiber of cardinality at most C;
(3) (Ta )∗c is a canonical JSJ tree relative to H ∪ Snvc ; in particular, it is invariant under
any automorphism of G preserving A and H;
(4) (Ta )∗c is compatible with every (A, H)-tree.
89
Proof. By the one-endedness assumption, all (A, H)-trees T have edge stabilizers in A∞ ,
so the collapsed tree of cylinders Tc∗ is defined. By Proposition 7.12 and Remark 7.13,
it is a (2, C)-acylindrical (A, H)-tree S-dominated by T , and Theorem 8.7 applies taking
T ∗ := Tc∗ (groups in A are small in (A, H)-trees by Assumption (3)): there exists a JSJ
tree Ta , obtained as in Lemma 8.14, and its flexible vertex stabilizers are in S or QH with
fiber of cardinality at most C.
Lemma 8.14 states that (Ta )∗c belongs to the deformation space of Tr (a JSJ tree relative
to H ∪ Snvc ), and Ta is obtained by refining Tr at vertices with stabilizer in S. This proves
the first two assertions of the corollary. The third one follows from Corollary 7.4.
For the fourth assertion, let T be any (A, H)-tree. Since Ta is universally elliptic, there
exists a refinement S of Ta dominating T by Proposition 2.2. By Lemma 7.14, Sc∗ and T
have a common refinement R. Since Sc∗ is a refinement of (Ta )∗c by Lemma 7.15, the tree
R is a common refinement of T and (Ta )∗c .
Remark 9.2. The result remains true if we enlarge S (keeping it invariant under conjugating
and taking subgroups), as long as all groups in S are small in (A, H)-trees.
Remark 9.3. By Corollary 4.16, the one-endedness assumption is usually not necessary for
the assertions about Ta , provided that the relative Stallings-Dunwoody space exists.
The first assumption of the corollary ensures that Propositions 5.20 and 5.21 apply to
flexible vertex stabilizers Gv of Ta and (Ta )∗c that do not belong to S. In particular:
Corollary 9.4. If Gv is a flexible vertex stabilizer not belonging to S, then:
(1) the underlying orbifold Σ contains an essential simple geodesic;
(2) every boundary component of Σ is used;
(3) every universally elliptic subgroup of Gv is contained in an extended boundary subgroup;
(4) if T is an (A, H)-tree in which Gv does not fix a point, the action of Gv on its minimal
subtree µT (Gv ) is dual to a family of geodesics of Σ.
Proof. This follows directly from Propositions 5.20 and 5.21, noting that edge stabilizers
for the action of Gv on µT (Gv ) are virtually cyclic by the second assertion of Proposition
5.21.
In the following sections, we are going to describe examples where Corollaries 9.1 and
9.4 apply. We first treat the case of abelian splittings of CSA groups. To allow torsion,
we introduce K-CSA groups in Subsection 9.2, and describe their JSJ decomposition over
virtually abelian groups. We then consider elementary splittings of relatively hyperbolic
groups, and splittings over virtually cyclic subgroups under the assumption that these subgroups have small commensurators. We conclude by defining the Zmax -JSJ decomposition
of one-ended hyperbolic groups.
9.1
CSA groups
In our first application, G is a torsion-free CSA group, and we consider abelian or cyclic
splittings. Recall that G is CSA if the commutation relation is transitive on G \ {1},
and maximal abelian subgroups are malnormal. Toral relatively hyperbolic groups, in
particular limit groups and torsion-free hyperbolic groups, are CSA. See Example 8.8 and
Figure 12 for an illustration.
We let A be either the family of abelian subgroups of G, or the family of cyclic subgroups. If G is freely indecomposable relative to H, commutation is an admissible equivalence relation on A∞ (see [GL11], or Lemma 9.13 below), and we can define trees of
90
cylinders Tc . The groups G[A] are maximal abelian subgroups, so are small in all trees.
Over abelian groups (i.e. when A is the class of abelian subgroups), all edge stabilizers of
Tc belong to A since every G[A] is abelian, so Tc∗ = Tc . Over cyclic groups, Tc may have
non-cyclic edge stabilizers, so we have to use Tc∗ , obtained from Tc by collapsing edges with
non-cyclic stabilizers.
Theorem 9.5. Let G be a finitely generated torsion-free CSA group, and H any family of
subgroups. Assume that G is freely indecomposable relative to H.
(1) There is an abelian (resp. cyclic) JSJ tree Ta relative to H. Its collapsed tree of
cylinders (Ta )∗c (for commutation) is a JSJ tree relative to H and all non-cyclic
abelian subgroups.
(2) Ta and (Ta )∗c have the same non-abelian vertex stabilizers; non-abelian flexible vertex
stabilizers are QH (they are fundamental groups of compact surfaces).
(3) (Ta )∗c is invariant under all automorphisms of G preserving H. It is compatible with
every (A, H)-tree.
Proof. We apply Corollary 9.1, with A consisting of all abelian (resp. cyclic) subgroups,
S the family of abelian subgroups, and C = 1. Since G is torsion-free, QH vertex groups
have trivial fiber, and the underlying orbifold is a surface.
9.2
Γ-limit groups and K-CSA groups
The notion of CSA groups is not well-adapted to groups with torsion. This is why we shall
introduce K-CSA groups, where K is an integer. Every hyperbolic group Γ is K-CSA
for some K. Being K-CSA is a universal property; in particular, all Γ-limit groups are
K-CSA.
We say that a group is K-virtually abelian if it contains an abelian subgroup of index
≤ K (note that the infinite dihedral group D∞ is 1-virtually cyclic, in the sense of Definition
7.8, but only 2-virtually abelian). As usual, a group is locally K-virtually abelian if its
finitely generated subgroups are K-virtually abelian.
Lemma 9.6. If a countable group J is locally K-virtually abelian, then J is K-virtually
abelian.
Proof. Let g1 , . . . , gn , . . . be a numbering of the elements of J. Let An ⊂ hg1 , . . . , gn i be
an abelian subgroup of index ≤ K. For a given k, there are only finitely many subgroups
of index ≤ K in hg1 , . . . , gk i, so there is a subsequence Ani (k) such that Ani (k) ∩ hg1 , . . . gk i
is independent of i. By a diagonal argument, one produces an abelian subgroup A of J
whose intersection with each hg1 , . . . , gn i has index ≤ K, so A has index ≤ K in J.
Definition 9.7 (K-CSA). Say that G is K-CSA for some K > 0 if:
(1) Any finite subgroup has cardinality at most K (in particular, any element of order
> K has infinite order).
(2) Any element g ∈ G of infinite order is contained in a unique maximal virtually
abelian group M (g), and M (g) is K-virtually abelian.
(3) M (g) is its own normalizer.
A 1-CSA group is just a torsion-free CSA group. The Klein bottle group is 2-CSA but
not 1-CSA. Any hyperbolic group Γ is K-CSA for some K since finite subgroups of Γ have
bounded order, and there are only finitely many isomorphism classes of virtually cyclic
groups whose finite subgroups have bounded order (see Lemma 2.2 of [GL16] for a proof).
Corollary 9.10 will say that Γ-limit groups also are K-CSA.
91
Lemma 9.8. Let G be a K-CSA group.
(1) If g, h ∈ G have infinite order, the following conditions are equivalent:
(a) g and h have non-trivial commuting powers;
(b) g K! and hK! commute;
(c) M (g) = M (h);
(d) hg, hi is virtually abelian.
(2) Any infinite virtually abelian subgroup H is contained in a unique maximal virtually
abelian group M (H). The group M (H) is K-virtually abelian and almost malnormal:
if M (H) ∩ M (H)g is infinite, then g ∈ M (H).
Proof. (c) ⇒ (b) ⇒ (a) in Assertion (1) is clear since g K! ∈ A if A ⊂ M (g) has index
≤ K. We prove (a) ⇒ (c). If g m commutes with hn , then g m normalizes M (hn ), so
M (g m ) = M (hn ) and M (g) = M (g m ) = M (hn ) = M (h). Clearly (c) ⇒ (d) ⇒ (a). This
proves Assertion (1).
Being virtually abelian, H contains an element h0 of infinite order, and we define
M (H) = M (h0 ). By Assertion (1), M (H) does not depend on the choice of h0 . To
prove that H ⊂ M (H), consider h ∈ H. Since hh0 h−1 ∈ H has infinite order, we have
M (h0 ) = M (hh0 h−1 ) = hM (h0 )h−1 , so h ∈ M (h0 ) because M (h0 ) equals its normalizer.
A similar argument shows almost malnormality. There remains to prove uniqueness. If
A is any virtually abelian group containing H, then M (A) is defined and coincides with
M (h0 ), so A ⊂ M (A) = M (h0 ) = M (H).
One easily checks that any subgroup of a K-CSA group is still K-CSA. This is in fact
a consequence of the following proposition saying that K-CSA is a universal property. We
refer to [CG05] for the topological space of marked groups, and its relation with universal
theory.
Proposition 9.9. For any fixed K > 0, the class of K-CSA groups is defined by a set of
(coefficient-free) universal sentences (K-CSA is a universal property). In particular, the
class of K-CSA groups is stable under taking subgroups, and closed in the space of marked
groups.
Proof. For any finite group F = {a1 , . . . , an }, the fact that G does not contain a subgroup
isomorphic to F is equivalent to a universal sentence saying that for any n-tuple (x1 , . . . , xn )
satisfying the multiplication table of F , not all xi ’s are distinct. Thus, the first property
of K-CSA groups is defined by (infinitely many) universal sentences.
Now consider the second property. We claim that, given m and n, the fact that
hg1 , . . . , gn i is m-virtually abelian may be expressed by the disjunction VAm,n of finitely
many finite systems of equations in the elements g1 , . . . , gn . To see this, let π : Fn → G be
the homomorphism sending the i-th generator xi of the free group Fn = hx1 , . . . xn i to gi .
If A ⊂ hg1 , . . . , gn i has index ≤ m, so does π −1 (A) in Fn . Conversely, if B ⊂ Fn has index
≤ m, so does π(B) in hg1 , . . . , gn i. To define VAm,n , we then enumerate the subgroups of
index ≤ m of Fn . For each subgroup, we choose a finite set of generators wi (x1 , . . . , xn )
and we write the system of equations [wi (g1 , . . . , gn ), wj (g1 , . . . , gn )] = 1. This proves the
claim.
By Lemma 9.6 (and Zorn’s lemma), any g is contained in a maximal K-virtually abelian
subgroup. The second property of Definition 9.7 can be restated as follows: any finitely
generated virtually abelian group is K-virtually abelian, and if hg, hi and hg, g1 , . . . , gn i
are K-virtually abelian, with g of order > K, then hg, h, g1 , . . . , gn i is K-virtually abelian.
This is defined by a set of universal sentences constructed using the VAm,n ’s.
92
If the first two properties of the definition hold, the third one is expressed by saying
that, if g has order > K and hg, hgh−1 i is K-virtually abelian, so is hg, hi. This is a set of
universal sentences as well.
Recall that a Γ-limit group is defined as a limit of subgroups of Γ in the space of marked
groups. Proposition 9.9 implies that, if Γ is K-CSA, then any Γ-limit group is K-CSA. In
particular:
Corollary 9.10. Let Γ be a hyperbolic group. There exists K such that any Γ-limit group
is K-CSA.
Moreover, any subgroup of a Γ-limit group G contains a non-abelian free subgroup or
is K-virtually abelian.
Remark 9.11. We will not use the “moreover”. There are additional restrictions on the
virtually abelian subgroups. For instance, there exists N ≥ 1 such that, if hgh−1 = g −1
for some g of infinite order, then hg 0N h−1 = g 0−N for all g 0 of infinite order in M (g).
Proof. The first assertion is immediate from Proposition 9.9.
Now let H be an infinite subgroup of G not containing F2 . By [Kou98, Proposition
3.2], there exists a number M such that, if x1 , . . . , xM are distinct elements of Γ, some
element of the form xi or xi xj has infinite order (i.e. order > K). This universal statement
also holds in G, so H contains an element g of infinite order. Recall that there exists a
number N such that, if x, y ∈ Γ, then xN and y N commute or generate F2 (see [Del96]).
The same statement holds in G since, for each non-trivial word w, the universal statement
[xN , y N ] 6= 1 ⇒ w(xN , y N ) 6= 1 holds in Γ hence in G. Thus, for all h ∈ H, the elements
g N and hg N h−1 commute. By Lemma 9.8, H normalizes M (g), so H ⊂ M (g) and H is
K-virtually abelian.
Let G be a K-CSA group. We now show how to define a tree of cylinders for virtually
abelian splittings of G (hence also for virtually cyclic splittings).
Definition 9.12 (Virtual commutation). Let A be the family of all virtually abelian subgroups of G, and A∞ the family of infinite subgroups in A. Given H, H 0 ∈ A∞ , define
H ∼ H 0 if M (H) = M (H 0 ). This is an equivalence relation, which we call virtual commutation.
Equivalently, H ∼ H 0 if and only if hH, H 0 i is virtually abelian. The stabilizer G[H] of
the equivalence class of any H ∈ A∞ (for the action of G by conjugation) is the virtually
abelian group M (H).
Lemma 9.13. If G is one-ended relative to H, the equivalence relation ∼ on A∞ is admissible (see Definition 7.1).
Proof. By one-endedness, all (A, H)-trees have edge stabilizers in A∞ . The first two properties of admissibility are obvious. Consider A, B ∈ A∞ with A ∼ B, and an (A, H)-tree T
in which A fixes some a and B fixes some b. Since the group generated by two commuting
elliptic groups is elliptic, there are finite index subgroups A0 ⊂ A and B0 ⊂ B such that
hA0 , B0 i fixes a point c ∈ T . Given any edge e in the segment [a, b], it is contained in [a, c]
or [c, b], so, say, A0 ⊂ Ge and Ge ∼ A0 ∼ A as required.
Theorem 9.14. Let G be a K-CSA group, and H any family of subgroups. Assume that
G is one-ended relative to H.
(1) There is a JSJ tree Ta relative to H over virtually abelian (resp. virtually cyclic)
subgroups. Its collapsed tree of cylinders (Ta )∗c (for virtual commutation) is a JSJ
tree relative to H and all virtually abelian subgroups which are not virtually cyclic.
93
(2) Ta and (Ta )∗c have the same non-virtually abelian vertex stabilizers; flexible vertex
stabilizers which are not virtually abelian are QH with finite fiber.
(3) (Ta )∗c is invariant under all automorphisms of G preserving H. It is compatible with
every (A, H)-tree.
Proof. We apply Corollary 9.1, with A the family of all virtually abelian (resp. virtually
cyclic) subgroups, ∼ virtual commutation, S the family of virtually abelian subgroups, and
C = K.
9.3
Relatively hyperbolic groups
In this subsection we assume that G is hyperbolic relative to a family of finitely generated
subgroups P = {P1 , . . . , Pp }. Recall that a subgroup is parabolic if it is conjugate to a
subgroup of some Pi , elementary if it is virtually cyclic (possibly finite) or parabolic. Any
infinite elementary subgroup is contained in a unique maximal elementary subgroup.
The following lemma is folklore.
Lemma 9.15. Let G be a relatively hyperbolic group.
(1) There exists C > 0 such that any elementary subgroup A < G of cardinality > C is
contained in unique maximal elementary subgroup E(A); moreover, E(A) is parabolic
if A is finite.
(2) If A < G is virtually cyclic but not parabolic, then it is finite of cardinality ≥ C or
C-virtually cyclic.
(3) If A, B < G are elementary subgroups such that A ∩ B has cardinality > C, then
E(A) = E(B).
Proof. The first assertion is contained in [GL15b, Lemma 3.1]. If A < G is virtually cyclic
but not C-virtually cyclic, its maximal finite normal subgroup F has cardinality > C. Thus
F is parabolic by (1), and so is A since E(A) = E(F ). This proves the second assertion.
The third assertion immediately follows from the first.
Definition 9.16 (Co-elementary). We say that two infinite elementary subgroups A, B
are co-elementary if hA, Bi is elementary, or equivalently if E(A) = E(B). This is an
equivalence relation ∼ on the set of infinite elementary subgroups.
We let A be either the class of elementary subgroups, or the class of virtually cyclic
groups. In both cases, co-elementarity is an equivalence relation on A∞ .
Lemma 9.17. Let A be the family of elementary subgroups (resp. of virtually cyclic subgroups), and let H be any family of subgroups. If every Pi is small in (A, H)-trees, the
co-elementarity equivalence relation on A∞ is admissible (relative to H).
Proof. We fix an (A, H)-tree T with infinite edge stabilizers. We assume that A ∼ B, and
that A, B fix a, b respectively in T . We must show Ge ∼ A for every edge e ⊂ [a, b]. This is
clear if hA, Bi is not parabolic (hence is virtually cyclic), since then A ∩ B is infinite (it has
finite index in A and B) and contained in Ge , so A ∼ (A ∩ B) ∼ Ge . Assume therefore that
hA, Bi is contained in some Pi . By assumption Pi is small in T . We distinguish several
cases.
If Pi fixes a point c, the edge e is contained in [a, c] or [b, c], and Ge contains A or B
so is equivalent to A and B. The argument is the same if Pi fixes an end of T (with c at
infinity). The last case is when Pi acts dihedrally on a line L. Let a0 , b0 be the projections
of a, b on L. They are fixed by A and B respectively. If e is contained in [a, a0 ] or [b, b0 ],
then Ge contains A or B, so we may assume e ⊂ [a0 , b0 ]. Now a subgroup A0 ⊂ A of index
at most 2 fixes L pointwise, so A0 ⊂ Ge and Ge ∼ A0 ∼ A.
94
The lemma allows us to define trees of cylinders. The stabilizer G[A] of the equivalence
class of any A ∈ A∞ (for the action of G by conjugation) is E(A), it is small in (A, H)-trees
if every Pi is. Note that Pi is small in (A, H)-trees if it does not contain F2 , or is contained
in a group of H.
When A is the class of elementary subgroups, no collapsing is necessary: (Tc )∗ = Tc .
On the other hand, if A is the class of virtually cyclic subgroups, and some of the Pi ’s are
not virtually cyclic, one may have G[A] ∈
/ A and (Tc )∗ may be a proper collapse of Tc .
Theorem 9.18. Let G be hyperbolic relative to a family of finitely generated subgroups
P = {P1 , . . . , Pp }, with no Pi virtually cyclic. Let A be the class of all elementary subgroups
of G (resp. of all virtually cyclic subgroups). Let H be any family of subgroups.
If G is one-ended relative to H, and every Pi is small in (A, H)-trees, then:
(1) there is a JSJ tree Ta relative to H over elementary (resp. virtually cyclic) subgroups;
its collapsed tree of cylinders (Ta )∗c (for co-elementarity) is a JSJ tree relative to
H ∪ P.
(2) Ta and (Ta )∗c have the same non-elementary vertex stabilizers; flexible vertex stabilizers which are not elementary are QH with finite fiber;
(3) (Ta )∗c is invariant under all automorphisms of G preserving P and H, and is compatible with every (A, H)-tree.
When G is hyperbolic and H = ∅, the tree (Ta )∗c = (Ta )c is the virtually cyclic JSJ tree
constructed by Bowditch [Bow98] using the topology of ∂G.
Removing virtually cyclic groups from P does not destroy relative hyperbolicity; the
assumption that no Pi is virtually cyclic makes statements simpler and causes no loss of
generality.
Proof. We apply Corollary 9.1, with C as in Lemma 9.15 and S the family of all elementary
subgroups (we use Remark 9.2 if we work over virtually cyclic groups and some Pi is a
torsion group). All groups in Snvc are parabolic, and every Pi is in Snvc . In particular,
a tree is relative to Snvc if and only if it is relative to P. Automorphisms preserving P
preserve the set of elementary subgroups, so Corollary 7.4 applies.
The assumption that parabolic groups Pi are small in (A, P)-trees is automatic as soon
as H contains P, since we consider splittings relative to H. We therefore get:
Corollary 9.19. Let G be hyperbolic relative to a finite family of finitely generated subgroups P = {P1 , . . . , Pp }. Let A be the family of elementary subgroups of G. Let H be any
family of subgroups containing P.
If G is one-ended relative to H, there is a JSJ tree over A relative to H which is equal
to its tree of cylinders, invariant under automorphisms of G preserving H, and compatible
with every (A, H)-tree. Its non-elementary flexible vertex stabilizers are QH with finite
fiber.
In particular:
Corollary 9.20. Let G be hyperbolic relative to a finite family of finitely generated subgroups P = {P1 , . . . , Pp }. Let A be the family of elementary subgroups of G.
If G is one-ended relative to P, there is a JSJ tree over A relative to P which is equal
to its tree of cylinders, invariant under automorphisms of G preserving P, and compatible
with every (A, P)-tree. Its non-elementary flexible vertex stabilizers are QH with finite
fiber.
Note that G is finitely presented relative to P, so existence of a JSJ tree also follows
from Theorem 2.20.
95
9.4
Virtually cyclic splittings
In this subsection we consider splittings of G over virtually cyclic groups, assuming smallness of their commensurators.
Let A be the family of virtually cyclic (possibly finite) subgroups of G, and A∞ the
family of all infinite virtually cyclic subgroups. Recall that two subgroups A and B of G
are commensurable if A ∩ B has finite index in A and B. The commensurability relation
∼ is an admissible relation on A∞ (see [GL11]), so one can define a tree of cylinders.
The stabilizer G[A] of the equivalence class of a group A ∈ A∞ is its commensurator
Comm(A), consisting of elements g such that gAg −1 is commensurable with A.
Corollary 9.1 yields:
Theorem 9.21. Let A be the family of virtually cyclic subgroups, and let H be any set
of subgroups of G, with G one-ended relative to H. Let S be the set of subgroups of
commensurators of infinite virtually cyclic subgroups. Assume that there is a bound C for
the order of finite subgroups of G, and that all groups of S are small in (A, H)-trees. Then:
(1) There is a virtually cyclic JSJ tree Ta relative to H. Its collapsed tree of cylinders
(Ta )∗c (for commensurability) is a virtually cyclic JSJ tree relative to H and the groups
of S which are not virtually cyclic.
(2) Ta and (Ta )∗c have the same vertex stabilizers not in S; their flexible subgroups commensurate some infinite virtually cyclic subgroup, or are QH with finite fiber.
(3) (Ta )∗c is invariant under all automorphisms of G preserving H. It is compatible with
every (A, H)-tree.
Remark 9.22. This applies if G is a torsion-free CSA group, or a K-CSA group, or any relatively hyperbolic group whose finite subgroups have bounded order as long as all parabolic
subgroups are small in (A, H)-trees.
If G is K-CSA, the trees of cylinders of a given T for commutation and for commensurability belong to the same deformation space (this follows from Lemma 9.8).
We also have:
Theorem 9.23. Let G be torsion-free and commutative transitive. Let H be any family of
subgroups. If G is freely indecomposable relative to H, then:
(1) There is a cyclic JSJ tree Ta relative to H. Its collapsed tree of cylinders (Ta )∗c
(for commensurability) is a JSJ tree relative to H and all subgroups isomorphic to a
solvable Baumslag-Solitar group BS(1, s).
(2) Ta and (Ta )∗c have the same non-solvable vertex stabilizers. Their flexible subgroups
are QH (they are surface groups), unless G = Z2 .
(3) (Ta )∗c is invariant under all automorphisms of G preserving H. It is compatible with
every (A, H)-tree.
Recall that G is commutative transitive if commutation is a transitive relation on G\{1}.
Remark 9.24. We cannot apply Corollary 9.1 directly, because we cannot claim that G[A] ,
the commensurator of a cyclic subgroup A = hai, is a BS(1, s). Note however that G[A]
is metabelian: if g commensurates A, there is a relation g ap g −1 = aq ; mapping g to p/q
defines a map from G[A] to Q∗ whose kernel is the centralizer of A, an abelian group.
Proof. By Proposition 6.5 of [GL11], if T is any tree with cyclic edge stabilizers, a vertex stabilizer of its collapsed tree of cylinders Tc∗ which is not elliptic in T is a solvable
Baumslag-Solitar group BS(1, s) (with s 6= −1 because of commutative transitivity). In
96
particular, T S-dominates Tc∗ , with S consisting of all groups contained in a BS(1, s)
subgroup.
We may therefore apply Theorem 8.7, taking T ∗ := Tc∗ , and argue as in the proof
of Corollary 9.1, using Lemmas 8.14 and Lemma 7.15 (which applies since G[A] cannot
contain F2 ). No group in S can be flexible, except if Gv = G ' Z2 , so all flexible groups
are QH surface groups.
9.5
The Zmax -JSJ decomposition
In this section, G is a one-ended hyperbolic group. We consider splittings of G over
virtually cyclic subgroups (necessarily infinite), and for simplicity we assume H = ∅.
Theorem 9.18 yields a tree T = (Ta )∗c = (Ta )c , which is the tree of cylinders of any
JSJ tree. This tree is itself a JSJ tree, it is canonical (in particular, invariant under
automorphisms), and its flexible vertex stabilizers are QH with finite fiber. It is in fact the
tree constructed by Bowditch in [Bow98].
It has been noticed by several authors [Sel97b, DG08, DG11] that it is sometimes useful
to replace T by a slightly different tree whose edge stabilizers are maximal virtually cyclic
subgroups with infinite center.
To motivate this, recall that there is a strong connection between splittings and automorphisms. By Paulin’s theorem [Pau91] combined with Rips’s theory of actions on
R-trees [BF95], G splits over a virtually cyclic subgroup whenever Out(G) is infinite.
Conversely, suppose G = A ∗C B, with C virtually cyclic (there is a similar discussion
for HNN extensions). If c belongs to the center of C, it defines a Dehn twist τc : the
automorphism of G which is conjugation by c on A and the identity on B.
But this does not always imply that Out(G) is infinite (see [MNS99]). There are two
reasons for this. First, even though C is infinite by one-endedness, its center may be
finite, for instance if C is infinite dihedral (see Subsection 1.1). Second, if some power of
c centralizes A or B, then the image of τc in Out(G) has finite order.
We therefore consider the set Z of subgroups C < G that are virtually cyclic with
infinite center, and the family Zmax consisting of the maximal elements of Z (for inclusion).
It is now true that Out(G) is infinite if and only if G splits over a group C ∈ Zmax
([Sel97b, DG11, Car11, GL11]).
We say that a subgroup is a Z-subgroup or a Zmax -subgroup if it belongs to Z or
Zmax . For C ∈ Z, we denote by Ĉ the unique Zmax -subgroup of G containing C (it is the
pointwise stabilizer of the pair of points of ∂G fixed by C).
A tree is a Z-tree, or a Zmax -tree, if its edge stabilizers are in Z or Zmax .
Definition 9.25 (Zmax -JSJ tree). A Zmax -tree is a Zmax -JSJ tree if it is elliptic with
respect to every Zmax -tree, and maximal (for domination) for this property.
Beware that Z and Zmax are not stable under taking subgroups, so this does not fit in
our usual setting.
All Zmax -JSJ trees belong to the same deformation space, the Zmax -JSJ deformation
space. As in Subsection 2.3, this follows from Proposition 2.2, but we need to know that,
given Zmax -trees T1 , T2 , any standard refinement T̂1 of T1 dominating T2 is a Zmax -tree.
To see this, consider an edge e of T̂1 . If its image in T1 is an edge, then Ge is an edge
stabilizer of T1 so is Zmax . If not, e is contained in the preimage Yv of a vertex v of T1 .
The group Ĝe is elliptic in T1 (because it contains Ge with finite index). If it fixes some w,
the segment between v and w is fixed by Ge , hence by Ĝe because T1 is a Zmax -tree. We
deduce that Ĝe fixes v, and therefore leaves Yv invariant. The fact that Yv maps injectively
into the Zmax -tree T2 now implies that Ĝe fixes e, since it fixes every edge in the image of
e in T2 .
97
Remark 9.26. This argument is based on the following useful fact: if a subgroup H < G
fixes a vertex v in a Zmax -tree, and H contains a group C ∈ Z, then Ĉ fixes v.
In this section we shall construct and describe a canonical Zmax -JSJ tree. Because
of the relation between Zmax -splittings and infiniteness of Out(G) mentioned above, it is
algorithmically computable (for computability of the usual JSJ decomposition, see [Tou09]
and [Bar16]).
Lemma 9.27. Given any Z-tree T , one can construct a Zmax -tree TZmax with the following
properties:
(1) T dominates TZmax ;
(2) every Zmax -tree S dominated by T is dominated by TZmax ;
(3) every edge stabilizer of TZmax has finite index in some edge stabilizer of T .
Example 9.28. It may happen that TZmax is trivial even though T is not. This occurs for
instance for T corresponding to a splitting of the form A ∗a=ck hci.
Proof. Let T 0 be the quotient of T by the smallest equivalence relation such that, for all
edges e ⊂ T and all h ∈ Ĝe , we have h.e ∼ e. We shall give an alternative description
of T 0 , which shows that it is a Zmax -tree satisfying all the required properties; however,
it may happen that T 0 is not minimal (it may even be trivial), so we define TZmax as the
minimal G-invariant subtree of T 0 , and the lemma follows.
We construct T 0 by folding. We argue by induction on the number of G-orbits of edges
e with Ge 6= Ĝe . Let e be such an edge (if there is none, T 0 = T is a Zmax -tree). Since
[Ĝe : Ge ] < ∞, the group Ĝe is elliptic in T . If one of the endpoints of e is fixed by Ĝe ,
let T1 be the tree obtained by folding together all the edges in the Ĝf -orbit of Gf , for
every edge f in the G-orbit of e. If not, let e0 be the first edge in the shortest path joining
Fix Ĝe to e. Since this path is fixed by Ge , one has Ge ⊂ Ge0
Ĝe = Ĝe0 , and one can
fold the edges in the Ĝe0 -orbit of e0 . In both cases we obtain a tree T1 having fewer orbits
of edges with Ge 6= Ĝe , and any map T → S factors through T1 . The lemma now follows
by induction.
Remark 9.29. A similar construction is used in Section 5 of [GL15a]: G is a toral relatively
hyperbolic group, T is a tree with abelian edge stabilizers, and it is replaced by a tree
whose edge stabilizers are abelian and stable under taking roots.
We denote by A the family consisting of all subgroups C ∈ Z and all their finite
subgroups (alternatively, one could include all finite subgroups of G). It is stable under
taking subgroups, and since G is one-ended A-trees have edge stabilizers in Z. Let T =
(Ta )∗c be the canonical JSJ tree over A provided by Theorem 9.18. Its flexible vertex
stabilizers are QH with finite fiber, and the underlying orbifold has no mirrors (see Theorem
6.5).
Let TZmax be the tree associated to T by Lemma 9.27. It is a Zmax -tree, which is elliptic
with respect to every Zmax -tree by Assertion (3) of the lemma. Unfortunately, it is not
always a Zmax -JSJ tree. We illustrate this on the Klein bottle group K (of course K is
not hyperbolic, see hyperbolic examples below).
As pointed out in Subsection 3.4, the cyclic JSJ decomposition of K is trivial, and K is
flexible because it has two cyclic splittings, corresponding to the presentations K = ha, t |
tat−1 = a−1 i = hti ∗t2 =v2 hvi. Thus T and TZmax are trivial trees. Note, however, that the
amalgam is not over a Zmax -subgroup (the group generated by t2 is not a maximal cyclic
subgroup). It follows that the Bass-Serre tree of the HNN extension is elliptic with respect
to every Zmax -tree, and is a Zmax -JSJ tree.
98
Geometrically, let Σ be a flat Klein bottle, or a compact hyperbolic surface, or more
generally a hyperbolic orbifold as in Theorem 6.5 (we allow cone points, but not mirrors).
An essential simple closed geodesic γ on Σ defines a splitting over a cyclic group, which is
in Zmax if and only if γ is 2-sided (see Subsection 5.1.4). The Zmax -JSJ tree is trivial if
and only if every essential 2-sided simple closed geodesic crosses (transversely) some other
2-sided geodesic.
This happens in almost all cases, but there are exceptions: the flat Klein bottle, the
Klein bottle with one conical point, and the Klein bottle with one open disc removed. On a
flat Klein bottle, all essential 2-sided simple closed geodesics are isotopic. In the other two
exceptional cases (which are hyperbolic), there is a unique essential 2-sided simple closed
geodesic.
We can now construct a canonical Zmax -JSJ tree T̂Zmax . If no orbifold Σv underlying
a QH vertex v of the canonical JSJ tree T is an exceptional one, we let T̂Zmax be the tree
obtained by applying Lemma 9.27 to T . Now suppose that there are vertices v with Σv a
Klein bottle with one conical point (in which case Gv = G because ∂Σ = ∅), or a Klein
bottle with one open disc removed. We then refine T at these vertices, using the splitting of
Σv dual to the unique essential 2-sided geodesic, and we apply the construction of Lemma
9.27 to the tree T̂ thus obtained.
Proposition 9.30. Let G be a one-ended hyperbolic group. The tree T̂Zmax constructed
above is a canonical Zmax -JSJ tree.
Proof. First suppose that T has no exceptional QH vertex. Then T̂Zmax = TZmax , and by
Assertion (3) of Lemma 9.27 it is elliptic with respect to every Zmax -tree. It is canonical
because T is canonical and the definition of TZmax given in the first paragraph of the proof
of Lemma 9.27 does not involve choices. We only need to prove maximality: any vertex
stabilizer of TZmax is elliptic in every tree S which is elliptic with respect to every Zmax -tree.
Recall that T is a tree of cylinders, so is bipartite (see Subsection 7.1). If v ∈ V1 (T ), its
stabilizer in TZmax is the same as in T (a maximal virtually cyclic subgroup). It is elliptic
in S because it contains an edge stabilizer of T with finite index.
If two edges of T are folded when passing from T to TZmax , they have commensurable
stabilizers, so belong to the same cylinder of T (which is the star of a vertex in V1 ). This
implies that, if v ∈ V0 (T ), its stabilizer in TZmax is a multiple amalgam (i.e. a tree of
groups) Ĝv = Gv ∗B1 B̂1 · · · ∗Bk B̂k , where B1 , . . . , Bk are representatives of conjugacy
classes of incident edge stabilizers. The group Gv is clearly elliptic in S if it is rigid, and
also if it is QH with non-exceptional underlying orbifold because no Zmax -splitting of Gv
relative to its incident edge group is universally elliptic. It follows that Ĝv is elliptic by
Remark 9.26.
The argument when there are exceptional QH vertices is similar. Refining T replaces
the exceptional QH vertices by QH vertices whose underlying orbifold is a pair of pants or
an annulus with a conical point; their stabilizers do not split over a Zmax -subgroup relative
to the boundary subgroups.
Remark 9.31. The proof shows that flexible vertex groups of T̂Zmax are QH with sockets
(also called sockets [Sel97b], or orbisockets [DG11]), i.e. QH groups with roots added to
the boundary subgroups. More precisely, they are of the form Ĝv = Gv ∗B1 B̂1 · · · ∗Bk B̂k
with Gv a QH vertex group of T and B1 , . . . , Bk representatives of the conjugacy classes of
boundary subgroups of Gv . The incident edge groups of Ĝv in T̂Zmax are the B̂i ’s (some of
them may be missing when the tree T 0 defined in the proof of Lemma 9.27 is not minimal).
Remark 9.32. In applications to model theory, hyperbolic surfaces Σ which do not carry
a pseudo-Anosov diffeomorphism play a special role. There are four of them: the pair of
pants, the twice-punctured projective plane, the once-punctured Klein bottle, the closed
99
non-orientable surface of genus 3. The first two have finite mapping class group, but the
other two do not and this causes problems (see the proof of Proposition 5.1 in [GLS17]).
When a once-punctured Klein bottle appears in a QH vertex, one may refine the splitting as explained above, using the unique essential 2-sided simple closed geodesic γ (this
creates a QH vertex based on a pair of pants). The Dehn twist Tγ around γ generates a
finite index subgroup of the mapping class group of Σ. It acts trivially on every vertex
group of the refined splitting.
Similarly, if Σ is a closed non-orientable surface of genus 3, there is a unique 1-sided
simple geodesic γ whose complement is orientable (see Proposition 2.1 of [GAnMB06]). It
is the core of a Möbius band whose complement is a once-punctured torus Σ0 (unlike Σ, it
carries a pseudo-Anosov diffeomorphism). The mapping class group of Σ leaves γ invariant
and preserves the cyclic splitting of π1 (Σ) given by decomposing Σ as the union of Σ0 and
a Möbius band. It is isomorphic to GL(2, Z), the mapping class group of Σ0 ,
These refinements give a canonical way of modifying the JSJ decomposition of a torsionfree one-ended hyperbolic group (described in Subsection 9.1) so that all surfaces appearing
in QH vertices have a mapping class group which is finite or contains a pseudo-Anosov map.
Part V
Compatibility
As usual, we fix a family A of subgroups which is stable under conjugating and taking subgroups, and another family H, and we only consider (A, H)-trees. We work with simplicial
trees, but we often view them as metric trees (with every edge of length 1) in order to
apply the results from the appendix (for instance, the fact that compatibility passes to the
limit). We will freely use some concepts from the appendix, in particular the arithmetic
of trees (Subsection A.5).
In Section 2 we have defined a JSJ tree (of G over A relative to H) as a tree which
is universally elliptic and dominates every universally elliptic tree. Its deformation space
is the JSJ deformation space DJSJ . In the next section we define the compatibility JSJ
deformation space Dco and the compatibility JSJ tree Tco . The deformation space Dco
contains a universally compatible tree and dominates every universally compatible tree.
The tree Tco is a preferred universally compatible tree in Dco . In particular, it is invariant
under automorphisms of G preserving A and H. In Section 11 we give examples, provided
in particular by trees of cylinders (see Section 7).
10
The compatibility JSJ tree
Recall (Subsection 1.4.1) that two trees T1 and T2 are compatible if they have a common
refinement. In other words, there exists a tree T with collapse maps T → Ti .
Definition 10.1 (Universally compatible). A tree T is universally compatible (over A
relative to H) if it is compatible with every tree.
In particular, this means that any tree T 0 can be obtained from T by refining and
collapsing. When T 0 is a one-edge splitting, either T 0 coincides with the splitting associated
to one of the edges of T /G, or one can obtain T 0 /G by refining T /G at some vertex v using a
one-edge splitting of Gv relative to the incident edge groups, and collapsing all the original
edges of T /G.
Definition 10.2 (Compatibility JSJ deformation space). If, among deformation spaces
containing a universally compatible tree, there is one which is maximal for domination, it
100
is unique. It is denoted by Dco and it is called the compatibility JSJ deformation space of
G over A relative to H.
To prove uniqueness, consider universally compatible trees T1 , T2 . By Corollary A.22
we may assume that they are irreducible. They are compatible with each other, and T1 ∨T2
(see Lemma A.24 in the appendix) is universally compatible by Proposition A.26 (2). If
T1 and T2 belong to maximal deformation spaces, we get that T1 , T1 ∨ T2 , and T2 lie in the
same deformation space, proving uniqueness.
Clearly, a universally compatible tree is universally elliptic. This implies that Dco
is dominated by DJSJ . Also note that, if T is universally compatible and J is an edge
stabilizer in an arbitrary tree, then J is elliptic in T (i.e. any tree is elliptic with respect
to T ).
10.1
Existence of the compatibility JSJ space
Theorem 10.3. If G is finitely presented relative to a family H = {H1 , . . . , Hp } of finitely
generated subgroups, the compatibility JSJ space Dco of G (over A relative to H) exists.
The heart of the proof of Theorem 10.3 is the following proposition.
Proposition 10.4. Let G be finitely presented relative to a family H = {H1 , . . . , Hp } of
finitely generated subgroups. Let T0 ← T1 · · · ← Tk ← · · · be a sequence of refinements
of irreducible universally compatible trees. There exist collapses T k of Tk , in the same
deformation space as Tk , such that the sequence T k converges to a universally compatible
simplicial (A, H)-tree T which dominates every Tk .
We view trees as metric, with each edge of length 1, and convergence is in the space
of R-trees (unlike the proof of Theorem 8.1, no rescaling of the metric is necessary here).
The tree T may have redundant vertices.
Proof of Theorem 10.3 from the proposition. We may assume that there is a non-trivial
universally compatible tree. We may also assume that all such trees are irreducible: otherwise it follows from Corollary A.22 (see the appendix) that there is only one deformation
space of trees, and the theorem is trivially true.
Let (Sα )α∈A be the set of (isomorphism classes of) universally compatible trees. We
have to find a universally compatible tree T which dominates every Sα . By Lemma 2.9,
we only need T to dominate all trees in a countable set Sk , k ∈ N.
Let Tk be the lcm S0 ∨ · · · ∨ Sk (see Definition A.27); it is universally compatible by
Assertion (2) of Proposition A.26, and refines Tk−1 . Proposition 10.4 yields the desired
tree T : it dominates every Tk , hence every Sk , and it is universally compatible by Corollary
A.12.
Proof of Proposition 10.4. By Dunwoody’s accessibility (see Proposition 2.24), there exists
a tree S which dominates every Tk (this is where we use finite presentability). But of course
we cannot claim that it is universally compatible.
We may assume that Tk and S are minimal, that Tk+1 is different from Tk , and that
the gcd S ∧ Tk (see Definition A.23) is independent of k. We define Sk = S ∨ Tk ; it has
no redundant vertices (see Remark A.25). We denote by ∆k , Γ, Γk the quotient graphs of
groups of Tk , S, Sk , and we let πk : Γk → Γ be the collapse map (see Figure 13). We denote
by ρk the collapse map Γk+1 → Γk .
The trees Sk all belong to the deformation space of S, and Sk+1 strictly refines Sk . In
particular, the number of edges of Γk grows. The idea now is the following. Accessibility
holds within a given deformation space (see [GL07a] page 147; this is an easy form of
accessibility, which requires no smallness or finite presentability hypothesis, and in any
101
cv
A
Bk
C1
C2
...
Tk /G = ∆k
Ck
Sk /G = Γk
A
C1
C2
A
Bk
Ck
T k /G = ∆k
...
u
B
C
Ck
A
πk
B
Ck
C
Sk′ /G
A
B
C
S/G = Γ
Figure 13: Segments in the quotients of the trees Tk , S, Sk , T k , Sk0 , with C1 ⊃ C2 ⊃ · · · ⊃
Ck ⊃ C, and Bk = B ∗C Ck .
case is not used directly here). This implies that the growth of Γk occurs through the
creation of a bounded number of long segments whose interior vertices have valence 2,
with one of the incident edge groups equal to the vertex group (but the other edge group
is smaller than the vertex group, since Sk has no redundant vertices). We now make this
precise.
Fix k. For each vertex v ∈ Γ, define Yv = πk−1 ({v}) ⊂ Γk . The Yv ’s are disjoint, and
edges of Γk not contained in ∪v Yv correspond to edges of Γ.
Since Sk and S are in the same deformation space, Yv is a tree of groups, and it contains
a vertex cv whose vertex group equals the fundamental group of Yv (which is the vertex
group of v in Γ). This cv may fail to be unique, but we can choose one for every k in a
way which is compatible with the maps ρk . We orient edges of Yv towards cv . The group
carried by such an edge is then equal to the group carried by its initial vertex.
Say that a vertex u ∈ Yv is peripheral if u = cv or u is adjacent to an edge of Γk which
is not in Yv (i.e. is mapped onto an edge of Γ by πk ). By minimality of Sk , each terminal
vertex u0 of Yv is peripheral (because it carries the same group as the initial edge of the
segment u0 cv ).
In each Γk , the total number of peripheral vertices is bounded by 2|E(Γ)| + |V (Γ)|. It
follows that the number of points of valence 6= 2 in ∪v Yv is bounded. Cutting each Yv at
its peripheral vertices and its points of valence ≥ 3 produces the segments of Γk mentioned
earlier. On the example of Figure 13, there is one segment cv u in Γk , corresponding to the
edges labelled C1 , . . . , Ck . The point cv ∈ Γk is the vertex labelled by A. The vertex v of Γ
to which the segment corresponds is the vertex of Γ labelled A. The vertex u is peripheral.
Having defined segments for each k, we now let k vary. The preimage of a segment
of Γk under the map ρk is a union of segments of Γk+1 . Since the number of segments is
bounded independently of k, we may assume that ρk maps every segment of Γk+1 onto a
segment of Γk . In particular, the number of segments is independent of k.
Recall that we have oriented the edges of Yv towards cv . Each edge contained in ∪v Yv
carries the same group as its initial vertex, and edges in a given segment are coherently
oriented. Segments are therefore oriented.
There are various ways of performing collapses on Γk . Collapsing all edges contained
in segments yields Γ (this does not change the deformation space). On the other hand,
one obtains ∆k = Tk /G from Γk by collapsing some of the edges which are not contained
in any segment (all of them if S ∧ Tk is trivial).
The segments of Γk may be viewed as segments in ∆k , but collapsing the initial edge
of a segment of ∆k may now change the deformation space (if the group carried by the
initial point of the segment has increased when Γk is collapsed to ∆k ).
We define a graph of groups ∆k by collapsing, in each segment of ∆k , all edges but
the initial one. The corresponding tree T k is a collapse of Tk which belongs to the same
deformation space as Tk . Moreover, the number of edges of ∆k (prime factors of T k ) is
102
constant: there is one per segment, and one for each common prime factor of Tk and S.
Let `k : G → Z be the length function of T k .
Lemma 10.5. The sequence `k is non-decreasing (i.e. every sequence `k (g) is non-decreasing)
and converges.
Proof. The difference between `k and `k−1 comes from the fact that initial edges of segments
of ∆k may be collapsed in ∆k−1 . Fix a segment L of ∆k . Let ek be its initial edge. We
assume that ek is distinct from the edge fk mapping onto the initial edge of the image of
L in ∆k−1 .
Assume for simplicity that ek and fk are adjacent (the general case is similar). The
group carried by fk is equal to the group carried by its initial vertex vk . A given lift ṽk
of vk to Tk is therefore adjacent to only one lift of fk (but to several lifts of ek ). On
any translation axis in Tk , every occurrence of a lift of fk is immediately preceded by
an occurrence of a lift of ek . The length function of the prime factor of Tk and T k−1
corresponding to fk is therefore bounded from above by that of the prime factor of Tk and
T k corresponding to ek . Since this is true for every segment, we get `k−1 ≤ `k as required.
Let Sk0 = S ∨T k . It collapses to S, belongs to the same deformation space as S (because
it is a collapse of Sk ), and the number of edges of Sk0 /G is bounded. By an observation
due to Forester (see [GL07a, p. 169]), this implies an inequality `(Sk0 ) ≤ C`(S), with C
independent of k. Since `k ≤ `(Sk0 ), we get convergence.
We call ` the limit of `k . It is the length function of a tree T because the set of length
functions of trees is closed [CM87]. This tree is simplicial because ` takes values in Z (see
Example A.4), and irreducible because `k is non-decreasing. It is universally compatible
as a limit of universally compatible trees, by Corollary A.12. Since ` ≥ `k , every g ∈ G
elliptic in T is elliptic in Tk , and T dominates Tk by Lemma 2.6. Each Hi is elliptic in T
because it is finitely generated and its elements are elliptic.
There remains to prove that every edge stabilizer Ge of T belongs to A. If Ge is finitely
generated, there is a simple argument using the equivariant Gromov topology. In general,
we argue as follows. We may find hyperbolic elements g, h such that Ge is the stabilizer of
the bridge between A(g) and A(h) (the bridge might be e ∪ e0 as in the proof of Lemma
A.21 if an endpoint of e is a valence 2 vertex). Choose k so that the values of `k and `
coincide on g, h, gh. In particular, the axes of g and h in T k are disjoint.
Any s ∈ Ge is elliptic in T k since `k ≤ `. Moreover, `k (gs) ≤ `(gs) ≤ `(g) = `k (g).
The fixed point set of s in T k must therefore intersect the axis of g, since otherwise
`k (gs) > `k (g). Similarly, it intersects the axis of h. It follows that Ge fixes the bridge
between the axes of g and h in T k , so Ge ∈ A. This concludes the proof of Proposition
10.4.
10.2
The compatibility JSJ tree Tco
We shall now deduce from [GL07a] that Dco , if irreducible, contains a canonical tree Tco ,
which we call the compatibility JSJ tree. It is fixed under any automorphism of G that
leaves Dco invariant. Note that Tco may be refined to a JSJ tree (Lemma 2.15).
Lemma 10.6. An irreducible deformation space D can only contain finitely many reduced
universally compatible trees.
Recall (see Subsection 1.4) that T is reduced if no proper collapse of T lies in the same
deformation space as T . If T is not reduced, one may perform collapses so as to obtain a
reduced tree T 0 in the same deformation space (and T 0 is universally compatible if T is).
Proof. This follows from results in [GL07a]. We refer to [GL07a] for definitions not
given here. Suppose that there are infinitely many reduced universally compatible trees
103
T1 , T2 , . . . . Let Sk = T1 ∨ T2 ∨ · · · ∨ Tk . It is an (A, H)-tree and belongs to D by Assertion
(3) of Proposition A.26 .
As pointed out on page 172 of [GL07a], the tree Sk is BF-reduced, i.e. reduced in the
sense of [BF91], because all its edges are surviving edges (they survive in one of the Ti ’s),
and the space D is non-ascending by Assertion (4) of Proposition 7.1 of [GL07a]. Since
there is a bound CD for the number of orbits of edges of a BF-reduced tree in D ([GL07a,
Proposition 4.2]), the sequence Sk is eventually constant.
Remark 10.7. The proof shows that D only contains finitely many reduced trees which are
compatible with every tree in D.
Corollary 10.8. If D is irreducible and contains a universally compatible tree, it has a
preferred element: the lcm of its reduced universally compatible trees.
This preferred element is universally compatible by Assertion (2) of Proposition A.26.
Definition 10.9 (Compatibility JSJ tree Tco ). If the compatibility JSJ deformation space
Dco exists and is irreducible, its preferred element is called the compatibility JSJ tree Tco
of G (over A relative to H). If Dco is trivial, we define Tco as the trivial tree (a point).
It may happen that Dco is neither trivial nor irreducible. It then follows from Remark
A.22 that it is the only non-trivial deformation space of trees. If there is a unique reduced
tree T in Dco (in particular, if Dco consists of actions on a line), we define Tco = T .
Otherwise we do not define Tco . See Subsection 11.4 for an example where Dco consists of
trees with exactly one fixed end.
11
Examples
We start with various examples, and we then explain in Subsection 11.7 that the tree of
cylinders (Ta )∗c of Section 7 belongs to the compatibility deformation space if all groups
G[A] belong to A.
For simplicity we assume H = ∅ in Subsections 11.1 through 11.6.
11.1
Free groups
When A is Aut(G)-invariant, the compatibility JSJ tree Tco is Out(G)-invariant. This
sometimes forces it to be trivial. Suppose for instance that G has a finite generating set ai
such that all elements ai and ai a±1
j (i 6= j) belong to the same Aut(G)-orbit. Then the only
Out(G)-invariant length function ` is the trivial one. This follows from Serre’s lemma (see
Subsection 1.2) if the generators are elliptic, from the inequality max(`(ai aj ), `(ai a−1
j )) ≥
`(ai ) + `(aj ) (see Lemma A.1) if they are hyperbolic. In particular:
Proposition 11.1. If G is a free group and A is Aut(G)-invariant, then Tco is trivial.
11.2
Algebraic rigidity
The following result provides simple examples with Tco non-trivial.
Proposition 11.2. Assume that there is only one reduced JSJ tree TJ ∈ DJSJ , and that
G does not split over a subgroup contained with infinite index in a group of A. Then Tco
exists and equals TJ .
Proof. Let T be any A-tree. The second assumption implies that T is elliptic with respect
to TJ by Remark 2.3 of [FP06] or Assertion (3) of Lemma 2.6, so we can consider a standard
refinement T̂ of T dominating TJ as in Proposition 2.2. Any equivariant map f : T̂ → TJ
must be constant on any edge whose stabilizer is not universally elliptic, hence factors
104
through the tree T 0 obtained from T̂ by collapsing these edges. In particular, T 0 dominates
TJ hence is a JSJ tree because it is universally elliptic. Since TJ is the unique reduced JSJ
tree, T 0 is a refinement of TJ , so TJ is compatible with T . This shows that TJ is universally
compatible. Thus Tco = TJ .
A necessary and sufficient condition for a tree to be the unique reduced tree in its
deformation space is given in [Lev05b] (see also [CF09]).
The proposition applies for instance to free splittings and splittings over finite groups,
whenever there is a JSJ tree with only one orbit of edges. This provides examples of
virtually free groups with Tco non-trivial: any amalgam F1 ∗F F2 with F1 , F2 finite and
F 6= F1 , F2 has this property (with A the set of finite subgroups).
11.3
Free products
Let A consist only of the trivial group. Let G = G1 ∗· · ·∗Gp ∗Fq be a Grushko decomposition
(Gi is non-trivial, not Z, and freely indecomposable, Fq is free of rank q). If p = 2 and
q = 0, or p = q = 1, there is a JSJ tree with one orbit of edges and Tco is a one-edge
splitting as explained above. We now show that Tco is trivial if p + q ≥ 3 (of course it is
trivial also if G is freely indecomposable or free of rank ≥ 2).
Assuming p + q ≥ 3, we actually show that there is no non-trivial tree T with trivial
edge stabilizers which is invariant under a finite index subgroup of Out(G). By collapsing
edges, we may assume that T only has one orbit of edges. Since p + q ≥ 3, we can write
G = A ∗ B ∗ C where A, B, C are non-trivial and A ∗ B is a vertex stabilizer of T . Given
a non-trivial c ∈ C and n 6= 0, the subgroup cn Ac−n ∗ B is the image of A ∗ B by an
automorphism but is not conjugate to A ∗ B. This contradicts the invariance of T .
11.4
(Generalized) Baumslag-Solitar groups
We consider cyclic splittings of generalized Baumslag-Solitar groups (see Subsection 3.5).
First consider a solvable Baumslag-Solitar group BS(1, s), with |s| ≥ 2. In this case
Dco is trivial if s is not a prime power. If s is a prime power, Dco is the JSJ deformation
space (it is not irreducible).
When G = BS(r, s) with none of r, s dividing the other, Proposition 11.2 applies
by [Lev05b]. In particular, Dco is non-trivial. This holds, more generally, when G is
a generalized Baumslag-Solitar group defined by a labelled graph with no label dividing
another label at the same vertex. See [Bee13] for a systematic study of Dco for generalized
Baumslag-Solitar groups.
11.5
The canonical decomposition of Scott and Swarup
Recall that a group is VPCn (resp. VPC≤n ) if it is virtually polycyclic of Hirsch length n
(resp. ≤ n). Let G be a finitely presented group, and n ≥ 1. Assume that G does not split
over a VPCn−1 subgroup, and that G is not VPCn+1 . Let A consist of all subgroups of
VPCn subgroups.
We have shown in [GL10] that the tree of cylinders (for commensurability) of the JSJ
deformation space is (up to subdivision) the Bass-Serre tree TSS of the regular neighbourhood Γn = Γ(Fn : G) constructed by Scott-Swarup in Theorem 12.3 of [SS03].
Since TSS is universally compatible ([SS03, Definition 6.1(1)], or [GL11, Corollary 8.4]
and [GL10, Theorem 4.1]), it is dominated by the compatibility deformation space Dco .
The domination may be strict: if G = BS(r, s), the tree TSS is always trivial but, as
pointed out above, Dco is non-trivial when none of r, s divides the other.
105
11.6
Poincaré duality groups
Let G be a Poincaré duality group of dimension n (see also work by Kropholler on this
subject [Kro90]). Although such a group is not necessarily finitely presented, it is almost
finitely presented [Wal04, Proposition 1.1], which is sufficient for Dunwoody’s accessiblity,
so the JSJ deformation space and the compatibility JSJ deformation space exist. By
[KR89b, Theorem A], if G splits over a virtually solvable subgroup H, then H is VPCn−1 .
We therefore consider the family A consisting of VPC≤n−1 -subgroups.
By [KR89a, Corollary 4.3], for all VPCn−1 subgroups H, the number of coends ẽ(G, H)
is 2. By [KR89a, Theorem 1.3], if G is not virtually polycyclic, then H has finite index in its
commensurator. By Corollary 8.4(2) of [GL11], this implies that the JSJ deformation space
contains a universally compatible tree (namely its tree of cylinders for commensurability),
so equals Dco .
But one has more in this context: any universally elliptic tree is universally compatible.
Indeed, since VPCn−1 -subgroups of G have precisely 2 coends, Proposition 7.4 of [SS03]
implies that any two one-edge splittings T1 , T2 of G over A with edge stabilizers of T1 elliptic
in T2 are compatible. Indeed, strong crossing of almost invariant subsets corresponding to
T1 and T2 occurs if and only edge stabilizers of T1 are not elliptic in T2 ([Gui05, Lemme
11.3]), and the absence of (weak or strong) crossing is equivalent to compatibility of T1
and T2 [SS00]. To sum up, we have:
Corollary 11.3. Let G be a Poincaré duality group of dimension n, with G not virtually
polycyclic. Let A the family of VPC≤n−1 -subgroups. Then Tco exists and lies in the JSJ
deformation space of G over A.
In particular, G has a canonical JSJ tree over A.
11.7
Trees of cylinders
In Sections 8 and 9 we have used trees of cylinders to construct a universally compatible
tree (Ta )∗c (see the last assertion of Corollary 9.1); we always denote by Ta a JSJ tree over
A relative to H, as in Lemma 8.14. We now show that (Ta )∗c belongs to the compatibility
JSJ deformation space under the additional assumption that every stabilizer G[A] belongs
to A (this implies that trees of cylinders have edge stabilizers in A, so (see Subsection 7.1)
collapsing is not needed: Tc = Tc∗ ).
Theorem 11.4. Given A and H, let ∼ be an admissible equivalence on A∞ . Assume that
G is one-ended relative to H, and there exists an integer C such that:
(1) A contains all C-virtually cyclic subgroups, and all subgroups of cardinal ≤ 2C;
(2) if two groups of A∞ are inequivalent, their intersection has order ≤ C;
(3) every stabilizer G[A] belongs to A, but G ∈
/ A;
(4) every G[A] is small in (A, H)-trees.
Then the compatibility JSJ deformation space Dco exists and contains (Ta )c , the tree of
cylinders of JSJ trees (see Corollary 9.1). It is trivial or irreducible, so the JSJ compatibility
tree Tco is defined. Flexible vertex stabilizers of Tco belong to A or are QH subgroups with
finite fiber.
Proof. Corollary 9.1, applied with S = A, shows that (Ta )c is universally compatible and
its flexible vertex stabilizers are in A or QH. The point is to show that (Ta )c is maximal
(for domination) among universally compatible trees. This will prove that Dco exists
106
and contains (Ta )c . It is trivial or irreducible because any tree of cylinders is trivial or
irreducible (see Subsection 7.1).
We consider a universally compatible tree T , and we show that (Ta )c dominates T .
Replacing T by T ∨ (Ta )c (which is universally compatible by Assertion (2) of Proposition
A.26), we can assume that T refines (Ta )c .
We have to show that each vertex stabilizer Gv of (Ta )c is elliptic in T . If Gv is not a
G[A] , then it is elliptic in Ta , hence in T because T , being universally elliptic, is dominated
by Ta . By Assumptions (3) and (4), we can therefore assume that Gv ∈ A, and also that
Gv is small in T .
Since G ∈
/ A, the quotient graph (Ta )c /G is not a point (equivalently, there are at least
two cylinders in Ta ). There are two cases. If the image of v in (Ta )c /G has valence at least
2, we can refine (Ta )c to a minimal tree T 0 (in the same deformation space) having Gv as
an edge stabilizer. Since Gv ∈ A, this is an (A, H)-tree. Its edge group Gv is elliptic in T
because T is universally compatible.
The remaining case is when the image of v in (Ta )c /G has valence 1. We assume that
Gv is not elliptic in T , and we argue towards a contradiction.
Let e be an edge of (Ta )c containing v. We are going to prove that Gv contains a
subgroup G0 of index 2 with Ge ⊂ G0 . Assuming this fact, we can refine (Ta )c to a
minimal (A, H)-tree T 0 in which G0 is an edge stabilizer. As above, G0 is elliptic in T ,
and so is Gv . This is the required contradiction, proving that (Ta )c dominates T .
We now construct G0 . Since we have assumed that T refines (Ta )c , Proposition 1.7 and
Lemma 4.7 imply that Gv contains a hyperbolic element. We know that Gv is small in T ,
so there are only two possibilities.
If there is a fixed end, the action defines a homomorphism χ : Gv → Z (see Subsection
1.2). This homomorphism vanishes on Ge , which is elliptic in T , but is non-trivial because
there is a hyperbolic element, so we define G0 as the preimage of the index 2 subgroup of
the image of ϕ.
If the action is dihedral, we get an epimorphism χ : Gv → Z/2 ∗ Z/2. Since Ge is
elliptic in T , its image under χ is trivial or contained in a conjugate of a Z/2 factor. One
constructs G0 as the preimage of a suitable index 2 subgroup of the image of χ.
This theorem applies directly to abelian splittings of CSA groups (Subsection 9.1),
virtually abelian splittings of K-CSA groups (Subsection 9.2), elementary splittings of
relatively hyperbolic groups (Subsection 9.3), because the condition G[A] ∈ A is satisfied.
In each of these cases we conclude that (Ta )∗c = (Ta )c belongs to Dco . We get for instance:
Corollary 11.5. Let G be a finitely generated one-ended torsion-free CSA group, A the
class of abelian subgroups, and H a family of subgroups. Let ∼ be the commutation relation
among infinite abelian subgroups.
Then the compatibility JSJ deformation space Dco exists and contains the tree of cylinders of any JSJ tree.
Corollary 11.6. Let G be hyperbolic relative to a family of finitely generated subgroups
P = {P1 , . . . , Pp }. Let A be the class of elementary subgroups, and let H any family of
subgroups containing all Pi ’s which contain F2 . Let ∼ be the co-elementary relation on
A∞ .
If G is one-ended relative to H, then the compatibility JSJ deformation space Dco exists
and contains the tree of cylinders of any JSJ tree.
We now consider cyclic splittings, with ∼ the commensurability relation.
Example. Let H be a torsion-free hyperbolic group with property (FA) (it has no nontrivial action on a tree), and hai a maximal cyclic subgroup. Consider the HNN extension
G = hH, t | tat−1 = ai, a one-ended torsion-free CSA group. The Bass-Serre tree T0
107
is a JSJ tree over abelian groups. Its tree of cylinders T1 is the Bass-Serre tree of the
amalgam G = H ∗hai ha, ti, it is also the compatibility JSJ tree over abelian groups by
Theorem 11.4. Over cyclic groups, T0 is a JSJ tree, its (collapsed) tree of cylinders is T1 ,
but the compatibility JSJ tree is T0 (this follows from Proposition 11.2 and [Lev05b]; the
non-splitting assumption of Proposition 11.2 holds over cyclic groups, but not over abelian
groups).
In this example Dco strictly dominates (Ta )∗c . One obtains a tree in Dco by refining
T1 = (Ta )∗c at vertices with group Z2 . This is a general fact.
Theorem 11.7. Let G be a finitely generated torsion-free group. Let A be the family of
cyclic subgroups, and let H be any family of subgroups of G, with G one-ended relative to
H. Assume that commensurators of infinite cyclic subgroups are small in (A, H)-trees.
Then the cyclic compatibility JSJ space Dco relative to H exists. If furthermore G is
not a solvable Baumslag-Solitar group BS(1, s), one obtains a tree in Dco by refining (Ta )∗c
(the collapsed tree of cylinders for commensurability) at some vertices v with Gv virtually
Z2 .
Remark 11.8. If G has torsion, but there is a bound for the order of finite subgroups, a
similar theorem holds for virtually cyclic splittings (with the same proof). In the furthermore, one must assume that G is not virtually BS(1, s); one has to prove that an ascending
HNN extension of an infinite virtually cyclic group is virtually BS(1, s), we leave this as
an exercise to the reader (compare [FM99]).
Proof. By Theorem 9.21, (Ta )∗c is universally compatible, so Dco , if it exists, dominates
(Ta )∗c and is dominated by Ta . By [GL11, Remark 5.11], smallness of commensurators
implies that (Ta )c and (Ta )∗c are in the same deformation space. It follows that any group
elliptic in (Ta )∗c but not in Ta is contained in the commensurator of some A ∈ A∞ , hence
is small in (A, H)-trees.
We shall now show that universally elliptic trees S dominating (Ta )∗c and dominated
by Ta belong to only finitely many deformation spaces. This will prove the existence of
Dco . To determine such a tree S up to deformation, one needs to know the action on S of
vertex stabilizers Gv of (Ta )∗c (up to deformation). This action has universally elliptic edge
stabilizers by Lemma 4.13, and Gv is small in (A, H)-trees or elliptic in Ta (hence in S),
so only two deformation spaces are possible for the action of a given Gv on S by Corollary
3.5. This shows the required finiteness, hence the existence of Dco .
One may obtain a tree T 0 ∈ Dco by refining (Ta )∗c at vertices v with Gv not elliptic
in T 0 . As explained above such a Gv is small in T 0 . There are two possibilities. If Gv
fixes exactly one end, then Dco is an ascending deformation space (as defined in [GL07a,
Section 7]). By Proposition 7.1(4) of [GL07a], Dco cannot be irreducible, so G = Gv is an
ascending HNN extension of a cyclic group, hence isomorphic to some BS(1, s). The other
possibility is that Gv acts on a line, hence is virtually Z2 because edge stabilizers are cyclic
(Gv is isomorphic to Z2 or the Klein bottle group).
Corollary 11.9. Let G be torsion-free and commutative transitive. Let H be any family
of subgroups. If G is freely indecomposable relative to H, and is not a solvable BaumslagSolitar group, the cyclic compatibility JSJ deformation space Dco relative to H exists and
may be obtained by (possibly) refining (Ta )∗c at vertices with stabilizer isomorphic to Z2 .
Proof. The theorem applies because commensurators of non-trivial cyclic subgroups are
metabelian (see Remark 9.24) hence small.
108
A
R-trees, length functions, and compatibility
In this appendix we view a simplicial tree as a metric space (by giving length 1 to every
edge), and more generally we consider R-trees. See [Sha91, Chi01] for basic facts on R-trees.
Recall that two simplicial trees T1 , T2 are compatible if there is a tree that collapses
onto both T1 and T2 . In the context of R-trees (simplicial or not), collapse maps have
a natural generalisation as maps preserving alignment: the image of an arc [a, b] is the
segment [f (a), f (b)] (possibly a point). Compatibility of R-trees thus makes sense.
The length function of an R-tree T with an isometric action of G is the map ` : G → R
defined by `(g) = minx∈T d(x, gx).
The first main result of this appendix is Theorem A.10, saying that two R-trees are
compatible if and only if the sum of their length functions is again a length function. This
has a nice consequence: the set of R-trees compatible with a given tree is closed.
As a warm-up, we give a proof of the following classical facts: a minimal irreducible
R-tree is determined by its length function; the equivariant Gromov-Hausdorff topology
and the axes topology (determined by length functions) agree on the space of irreducible
R-trees (following a suggestion by M. Feighn, we extend this to the space of semi-simple
trees). Our proof does not use based length functions and extends to a proof of Theorem
A.10.
After proving Theorem A.10, we show that pairwise compatibility for a finite set of Rtrees implies the existence of a common refinement. We then define prime factors, greatest
common divisors (gcd’s), least common multiples (lcm’s) for irreducible simplicial trees.
We conclude by explaining how to obtain actions on R-trees by blowing up vertices of JSJ
trees.
We assume that G is finitely generated, but Subsections A.1 to A.4 apply to any
infinitely generated group (hypotheses such as irreducibility ensure that G contains enough
hyperbolic elements). We leave details to the reader.
A.1
Metric trees and length functions
When endowed with a path metric making each edge isometric to a closed interval, a
simplicial tree becomes an R-tree (we usually declare each edge to have length 1). An Rtree is a geodesic metric space T in which any two distinct points are connected by a unique
topological arc (which we often call a segment). Most considerations of the preliminary
section apply to R-trees as well as simplicial trees.
We denote by d, or dT , the distance in a tree T . All R-trees are equipped with an
isometric action of G, and considered equivalent if they are equivariantly isometric. If
λ > 0, we denote by T /λ the tree T equipped with the distance d/λ.
A branch point is a point x ∈ T such that T \ {x} has at least three components. A
non-empty subtree is degenerate if it is a single point, non-degenerate otherwise. If A, B
are disjoint closed subtrees, the bridge between them is the unique arc I = [a, b] such that
A ∩ I = {a} and B ∩ I = {b}.
We say that a map f : T → T 0 preserves alignment, or is a collapse map, if the image
of any segment [a, b] is the segment [f (a), f (b)] (possibly a point). The restriction of f
to [a, b] is then continuous. A map is a collapse map if and only if its restriction to any
segment is continuous, and the preimage of any point is a subtree. Two trees T1 and T2
are compatible if there exists a tree T with collapse maps fi : T → Ti .
If g ∈ G, we denote by `(g) its translation length `(g) = minx∈T d(x, gx). There is no
parabolic isometry in an R-tree, so the minimum is achieved on a non-empty subset of T ,
the characteristic set A(g): the fixed point set if g is elliptic (`(g) = 0), the axis if g is
hyperbolic (`(g) > 0).
The map ` : G → R is the length function of T ; we denote it by `T if there is a risk of
confusion. We say that a map ` : G → R is a length function if there is a tree T such that
109
` = `T .
The action is minimal if there is no proper G-invariant subtree. If G contains a hyperbolic element, there is a unique minimal subtree: the union of translation axes of hyperbolic
elements.
As in Proposition 1.2, if a group H (possibly infinitely generated) acts on an R-tree T ,
one of the following must occur:
• the action is irreducible (there are two hyperbolic elements g, h whose axes have intersection of finite length; for n large, g n and hn generate an F2 acting freely and discretely);
• there is a fixed point in T , or only in its metric completion (trivial action);
• there is an invariant line;
• there is a fixed end (an end of an R-tree is defined as an equivalence class of geodesic
rays up to finite Hausdorff distance).
If H is finitely generated (or finitely generated relative to finitely many elliptic subgroups), and fixes a point in the metric completion of T , then it fixes a point in T . When
there is a fixed end, the length function is the absolute value of a homomorphism χ : H → R
(such length functions are usually called abelian; we do not use this terminology, as it may
cause confusion).
As in [CM87], we say that a minimal T is semi-simple if there is a hyperbolic element,
and either there is an invariant line in T , or the action is irreducible.
We will use the following facts, with T an R-tree with a minimal action of G.
Lemma A.1 ([Pau89]). Let g, h be hyperbolic elements.
(1) If their axes A(g), A(h) are disjoint, then
`(gh) = `(g −1 h) = `(g) + `(h) + 2d(A(g), A(h)) > `(g) + `(h).
The intersection between A(gh) and A(hg) is the bridge between A(g) and A(h).
(2) If their axes meet, then
min(`(gh), `(g −1 h)) ≤ max(`(gh), `(g −1 h)) = `(g) + `(h).
The inequality is an equality if and only if the axes meet in a single point.
Lemma A.2 ([CM87, Theorem 2.7]). T is irreducible if and only if there exist hyperbolic
elements g, h with [g, h] hyperbolic.
Lemma A.3 ([Pau89, Lemma 4.3]). If T is irreducible, any arc [a, b] is contained in the
axis of some g ∈ G.
Example A.4. We use these lemmas to show that, if T has a minimal irreducible action of
G, and ` takes values in Z, then T is a simplicial tree. It suffices to prove that the distance
between any two branch points lies in 12 Z. Given two branch points a, b ∈ T , by Lemma A.3,
one can find hyperbolic elements g, h ∈ G with disjoint axes such that the bridge between
the axes of g and h is precisely [a, b]. By Lemma A.1(1), d(a, b) = 12 (`(gh) − `(g) − `(h)).
A.2
From length functions to trees
Let G be a finitely generated group. Let T be the set of minimal isometric actions of
G on R-trees modulo equivariant isometry. Let Tirr ⊂ T be the set of irreducible R-trees.
The following are classical results:
Theorem A.5 ([AB87, CM87]). Two minimal irreducible R-trees T, T 0 with the same
length function are equivariantly isometric.
110
Theorem A.6 ([Pau89]). The equivariant Gromov-Hausdorff topology and the axes topology agree on Tirr .
By Theorem A.5, the assignment T 7→ `T defines an embedding Tirr → RG . The axes
topology is the topology induced by this embedding. The set of length functions is closed
in RG (even if G is not finitely generated), and when G is finitely generated it is projectively
compact [CM87, Theorem 4.5].
The equivariant Gromov-Hausdorff topology (or just Gromov topology) on T is defined
by the following neighbourhood basis. Given T ∈ T , a number ε > 0, a finite subset
A ⊂ G, and x1 , . . . , xn ∈ T , define Nε,A,{x1 ,...xn } (T ) as the set of trees T 0 ∈ T such that
there exist x01 , . . . , x0n ∈ T 0 with
|dT 0 (x0i , ax0j ) − dT (xi , axj )| ≤ ε
for all a ∈ A and i, j ∈ {1, . . . , n}. We call x0i an approximation of xi in T 0 .
Example A.7. As an illustration of this definition, let us explain why, for a given g ∈ G, the
set of trees T ∈ T such that g fixes a tripod in T is open in the Gromov topology. Recall
that a tripod is the convex hull of 3 points that do not lie in a segment. Let x1 , x2 , x3
(with indices modulo 3) be the endpoints of a tripod fixed by g, and p the center of this
tripod. Let ε be very small compared to the distances dT (p, xi ), and take A = {1, g}. If
T 0 lies in Nε,A,{x1 ,x2 ,x3 } , consider approximation points x0i of xi . Since dT (xi , gxi ) = 0, we
have dT 0 (x0i , gx0i ) < ε. It follows that the midpoint m0i of [x0i , gx0i ] in T 0 is at distance at
most ε/2 from x0i . Now
dT (xi−1 , xi ) + dT (xi , xi+1 ) − dT (xi−1 , xi ) = 2dT (p, xi ) ε,
so
dT 0 (x0i−1 , x0i ) + dT 0 (x0i , x0i+1 ) − dT 0 (x0i−1 , x0i ) ε
and
dT 0 (m0i−1 , m0i ) + dT 0 (m0i , m0i+1 ) − dT 0 (m0i−1 , m0i ) > 0.
It follows that the three points m01 , m02 , m03 do not lie in a segment. But the midpoint of
[x, gx] always belongs to the characteristic set of g, so the characteristic set of g cannot be
a line. Thus g is elliptic, and therefore fixes m01 , m02 , m03 .
The fact that the axes topology is finer than the Gromov topology (which is the harder
half of Theorem A.6) should be viewed as a version with parameters of Theorem A.5: the
length function determines the tree, in a continuous way. As a preparation for the next
subsection, we now give quick proofs of these theorems. Unlike previous proofs, ours does
not use based length functions.
Proof of Theorem A.5. Let T, T 0 be minimal irreducible R-trees with the same length function `. We denote by A(g) the axis of a hyperbolic element g in T , by A0 (g) its axis in T 0 .
By Lemma A.1, A(g) ∩ A(h) is empty if and only if A0 (g) ∩ A0 (h) is empty.
We define an isometric equivariant map f from the set of branch points of T to T 0 ,
as follows. Let x be a branch point of T , and y 6= x an auxiliary branch point. By
Lemmas A.1 and A.3, there exist hyperbolic elements g, h whose axes in T do not intersect,
such that [x, y] is the bridge between A(g) and A(h), with x ∈ A(g) and y ∈ A(h).
Then {x} = A(g) ∩ A(gh) ∩ A(hg). The axes of g and h in T 0 do not intersect, so
A0 (g) ∩ A0 (gh) ∩ A0 (hg) is a single point which we call f (x).
Note that f (x) = ∩k A0 (k), the intersection being over all hyperbolic elements k whose
axis in T contains x: if k is such an element, its axis in T 0 meets all three sets A0 (g),
A0 (gh), A0 (hg), so contains f (x). This gives an intrinsic definition of f (x), independent
of the choice of y, g, and h. In particular, f is G-equivariant. It is isometric because
dT 0 (f (x), f (y)) and dT (x, y) are both equal to 1/2 `(gh) − `(g) − `(h) .
111
We then extend f equivariantly and isometrically first to the closure of the set of branch
points of T , and then to each complementary interval. The resulting map from T to T 0 is
onto because T 0 is minimal.
Proof of Theorem A.6. Given g ∈ G, the map T 7→ `T (g), from Tirr to R, is continuous in
the Gromov topology: this follows from the formula `(g) = max(d(x, g 2 x) − d(x, gx), 0).
This shows that the Gromov topology is finer than the axes topology.
For the converse, we fix ε > 0, a finite set of points xi ∈ T , and a finite set of elements
ak ∈ G. We have to show that, if the length function `0 of T 0 is close enough to ` on a
suitable finite subset of G, there exist points x0i ∈ T 0 such that
|dT 0 (x0i , ak x0j ) − dT (xi , ak xj )| < ε
for all i, j, k.
First assume that each xi is a branch point. For each i, choose elements gi , hi as
in the previous proof, with xi an endpoint of the bridge between A(gi ) and A(hi ). If `0
is close to `, the axes of gi and hi in T 0 are disjoint and we can define x0i as A0 (gi ) ∩
A0 (gi hi ) ∩ A0 (hi gi ). A different choice g̃i , h̃i may lead to a different point x̃0i . But the
distance between x0i and x̃0i goes to 0 as `0 tends to ` because all pairwise distances between
A0 (gi ), A0 (gi hi ), A0 (hi gi ), A0 (g̃i ), A0 (g̃i h̃i ), A0 (h̃i g̃i ) go to 0. It is then easy to complete the
proof.
If some of the xi ’s are not branch points, one can add new points so that each such xi
is contained in an arc bounded by branch points xbi , xci . One then defines x0i as the point
dividing [x0bi , x0ci ] in the same way as xi divides [xbi , xci ].
As suggested by M. Feighn, one may extend the previous results to reducible trees. Let
Tss consist of all minimal trees which are either irreducible or isometric to R (we only rule
out trivial trees and trees with exactly one fixed end).
Every non-zero length function is the length function of a tree in Tss .
Theorem A.8. Two minimal trees T, T 0 ∈ Tss with the same length function are equivariantly isometric. The equivariant Gromov-Hausdorff topology and the axes topology agree
on Tss .
In other words, the assignment T 7→ `T induces a homeomorphism between Tss ,
equipped with the Gromov topology, and the space of non-zero length functions.
We note that the results of [CM87] are stated for all trees in Tss , those of [Pau89] for
trees which are irreducible or dihedral.
Proof. We refer to [CM87, page 586] for a proof of the first assertion when the actions are
not irreducible. Since the set of irreducible length functions is open, it suffices to show the
following fact:
Claim A.9. If Tn is a sequence of trees in Tss whose length functions `n converge to the
length function ` of an action of G on T = R, then Tn converges to T in the Gromov
topology.
To prove the claim, we denote by An (g) the characteristic set of g ∈ G in Tn , and we
fix h ∈ G hyperbolic in T (hence in Tn for n large). We denote by In (g) the (possibly
empty or degenerate) segment An (g) ∩ An (h).
The first case is when G acts on T by translations. To show that
T Tn converges to T , it
suffices to show that, given elements g1 , . . . , gk in G, the length of i In (gi ) goes to infinity
with n. By a standard argument using Helly’s theorem, we may assume k = 2.
We first show that, for any g, the length |In (g)| goes to infinity. Let N ∈ N be arbitrary.
Since g −1 hN gh−N is elliptic in T , the distance between In (hN gh−N ) and In (g) goes to 0
as n → ∞. But In (hN gh−N ) is the image of In (g) by hN , so lim inf n→∞ |In (g)| ≥ N `(h).
112
To show that the overlap between In (g1 ) and In (g2 ) goes to infinity, we can assume
that the relative position of In (g1 ) and In (g2 ) is the same for all n’s. If they are disjoint,
In (g1 g2 g1−1 ) or In (g2 g1 g2−1 ) is empty, a contradiction. Since every |In (g)| goes to infinity,
the result is clear if In (g1 ) and In (g2 ) are nested. In the remaining case, up to changing
gi to its inverse, we can assume that g1 , g2 translate in the same direction along An (h) if
they are both hyperbolic. Then In (g1 ) ∩ In (g2 ) equals In (g1 g2 ) or In (g2 g1 ), so its length
goes to infinity.
Now suppose that the action of G on T is dihedral. Suppose that g ∈ G reverses
orientation on T . For n large, the axes of h, g −1 hg, ghg −1 in Tn have a long overlap by the
previous argument. On this overlap h translates in one direction, g −1 hg and ghg −1 in the
other (because `n (hg −1 hg) is close to 0 and `n (h−1 g −1 hg) is not). It follows that g acts as
a central symmetry on a long subarc of An (h). Moreover, if g, g 0 both reverse orientation,
the distance between their fixed points on An (h) is close to 2`(gg 0 ). The convergence of Tn
to T easily follows from these observations. This proves the claim, hence the theorem.
A.3
Compatibility and length functions
Recall that two R-trees T1 , T2 are compatible if they have a common refinement: there
exists an R-tree T̂ with (equivariant) maps gi : T̂ → Ti preserving alignment (the image of
a segment is a segment, possibly a point, see Subsections 1.4 and A.1); we call such maps
collapse maps.
If T1 and T2 are compatible, they have a standard common refinement Ts constructed
as follows.
We denote by di the distance in Ti , and by `i the length function. Let T̂ be any common
refinement. Given x, y ∈ T̂ , define
δ(x, y) = d1 (g1 (x), g1 (y)) + d2 (g2 (x), g2 (y)).
This is a pseudo-distance satisfying δ(x, y) = δ(x, z) + δ(z, y) if z ∈ [x, y] (this is also
a length measure, as defined in [Gui00a]). The associated metric space (Ts , d), obtained
by identifying x, y when δ(x, y) = 0, is an R-tree which refines T1 and T2 , with maps
fi : Ts → Ti satisfying d(x, y) = d1 (f1 (x), f1 (y))+d2 (f2 (x), f2 (y)). The fi ’s are 1-Lipschitz,
hence continuous.
The length function of Ts is ` = `1 +`2 (this follows from the formula `(g) = limn→∞ n1 d(x, g n x)).
In particular, `1 + `2 is a length function. We now prove the converse.
Theorem A.10. Two minimal irreducible R-trees T1 , T2 with an action of G are compatible if and only if the sum ` = `1 + `2 of their length functions is a length function.
Remark A.11. If T1 and T2 are compatible, then λ1 l1 + λ2 l2 is a length function for all
λ1 , λ2 ≥ 0.
Corollary A.12. Compatibility is a closed relation on Tirr × Tirr . In particular, the set of
irreducible R-trees compatible with a given T0 is closed in Tirr .
Proof. This follows from the fact that the set of length functions is a closed subset of RG
[CM87].
Proof of Theorem A.10. We have to prove the “if” direction. Let T1 , T2 be irreducible
minimal R-trees with length functions `1 , `2 , such that ` = `1 + `2 is the length function
of a minimal R-tree T . We denote by A(g), A1 (g), A2 (g) axes in T , T1 , T2 respectively.
By Lemma A.2, T is irreducible: hyperbolic elements g, h with [g, h] hyperbolic exist
in T since they exist in T1 . We want to prove that T is a common refinement of T1 and T2 .
In fact, we show that T is the standard refinement Ts mentioned earlier (which is unique
by Theorem A.5). The proof is similar to that of Theorem A.5, but we first need a few
lemmas.
113
Lemma A.13. [[Gui05, lemme 1.3]] Let S ⊂ G be a finitely generated semigroup such
that no point or line in T is invariant under the subgroup hSi generated by S. Let I be an
arc contained in the axis of a hyperbolic element h ∈ S.
Then there exists a finitely generated semigroup S 0 ⊂ S with hS 0 i = hSi such that every
element g ∈ S 0 \ {1} is hyperbolic in T , its axis contains I, and g translates in the same
direction as h on I.
Lemma A.14. Let T1 , T2 , T be arbitrary irreducible minimal trees. Given an arc I ⊂ T ,
there exists g ∈ G which is hyperbolic in T1 , T2 and T , and whose axis in T contains I.
Proof. Apply Lemma A.13 with S = G and any h whose axis in T contains I (if G is not
finitely generated, one takes for S the group generated by s1 , t1 , s2 , t2 , h, with si , ti , [si , ti ]
hyperbolic in Ti ). Since S 0 generates S as a group, it must contain an element h0 which
is hyperbolic in T1 : otherwise S would have a global fixed point in T1 by Serre’s lemma
(see Subsection 1.2). Applying Lemma A.13 to the action of S 0 on T1 , we get a semigroup
S 00 ⊂ S 0 whose non-trivial elements are hyperbolic in T1 . Similarly, S 00 contains an element
g which is hyperbolic in T2 . This element g satisfies the conclusions of the lemma.
Remark A.15. More generally, one may require that g be hyperbolic in finitely many trees
T1 , . . . , Tn .
We again assume that T1 , T2 , T are as described at the beginning of the proof of
Theorem A.10.
Lemma A.16. Let g, h be hyperbolic in T1 and T2 (and therefore in T ).
• If their axes in T meet, so do their axes in Ti .
• If their axes in T do not meet, their axes in Ti meet in at most one point. In
particular, the elements gh and hg are hyperbolic in Ti .
Proof. Assume that A(g) and A(h) meet, but A1 (g) and A1 (h) do not. Then `1 (gh) >
`1 (g)+`1 (h). Since `(gh) ≤ `(g)+`(h), we get `2 (gh) < `2 (g)+`2 (h). Similarly, `2 (g −1 h) <
`2 (g) + `2 (h). But these inequalities are incompatible by Lemma A.1.
Now assume that A(g) and A(h) do not meet, and A1 (g), A1 (h) meet in a nondegenerate arc. We may assume
`1 (gh) < `1 (g −1 h) = `1 (g) + `1 (h).
Since
`(gh) = `(g −1 h) > `(g) + `(h),
we have
`2 (gh) > `2 (g −1 h) > `2 (g) + `2 (h),
contradicting Lemma A.1.
We can now complete the proof of Theorem A.10. It suffices to define maps fi : T → Ti
such that
d(x, y) = d1 (f1 (x), f1 (y)) + d2 (f2 (x), f2 (y)).
Such maps are collapse maps (they are 1-Lipschitz, and if three points satisfy a triangular
equality in T , then their images under fi cannot satisfy a strict triangular inequality), so
T is the standard common refinement Ts .
The construction of fi is the same as that of f in the proof of Theorem A.5. Given
branch points x, y ∈ T , we use Lemma A.14 to get elements g and h hyperbolic in all
three trees, and such that the bridge between A(g) and A(h) is [x, y]. Then Lemma A.16
guarantees that Ai (g) ∩ Ai (gh) ∩ Ai (hg) is a single point of Ti , which we define as fi (x);
the only new phenomenon is that Ai (g) and Ai (h) may now intersect in a single point.
114
The relation between d, d1 , d2 comes from the equality ` = `1 + `2 , using the formula
di (fi (x), fi (y)) = 1/2(`i (gh) − `i (g) − `i (h)).
Having defined fi on branch points, we extend it by continuity to the closure of the set
of branch points of T (it is 1-Lipschitz) and then linearly to each complementary interval.
The relation between d, d1 , d2 still holds.
A.4
Common refinements
The following result is proved for almost-invariant sets in [SS03, Theorem 5.16].
Proposition A.17. Let G be a finitely generated group, and let T1 , . . . , Tn be irreducible
minimal R-trees such that Ti is compatible with Tj for i 6= j. Then there exists a common
refinement T of all Ti ’s.
Remark A.18. This statement may be interpreted as the fact that the set of projectivized
trees satisfies the flag condition for a simplicial complex: whenever one sees the 1-skeleton
of an n-simplex, there is indeed an n-simplex. Two compatible trees Ti , Tj define a 1simplex t`i + (1 − t)`j of length functions. If there are segments joining any
P pair of length
functions `i , `j , the proposition says that there is an (n − 1)-simplex
ti `i of length
functions.
To prove Proposition A.17, we need some terminology from [Gui05]. A direction in an
R-tree T is a connected component δ of T \ {x} for some x ∈ T . A quadrant in T1 × T2 is
a product Q = δ1 × δ2 of a direction of T1 by a direction of T2 . A quadrant Q = δ1 × δ2
is heavy if there exists h ∈ G hyperbolic in T1 and T2 such that δi contains a positive
semi-axis of h (equivalently, for all x ∈ Ti one has hn (x) ∈ δi for n large). We say that h
makes Q heavy. The core C(T1 ×T2 ) ⊂ T1 ×T2 is the complement of the union of quadrants
which are not heavy (this is not the same core as in Subsection 6.3).
By [Gui05, Théorème 6.1], T1 and T2 are compatible if and only if C(T1 × T2 ) contains
no non-degenerate rectangle (a product I1 × I2 where each Ii is an arc not reduced to a
point).
We first prove a technical lemma.
Lemma A.19. Let T1 , T2 be irreducible and minimal. Let f : T1 → T10 be a collapse map,
with T10 irreducible. Let δ10 × δ2 be a quadrant in T10 × T2 , and δ1 = f −1 (δ10 ). If the quadrant
δ1 × δ2 ⊂ T1 × T2 is heavy, then so is δ10 × δ2 .
Note that δ1 is a direction because f preserves alignment.
Proof. Consider an element h making δ1 × δ2 heavy. If h is hyperbolic in T10 , then h makes
δ10 × δ2 heavy and we are done. If not, assume that we can find some g ∈ G, hyperbolic
in T10 and T2 (hence in T1 ), such that for i = 1, 2 the axis Ai (g) of g in Ti intersects Ai (h)
in a compact set. Then for n > 0 large enough the element hn gh−n makes δ1 × δ2 heavy.
Since this element is hyperbolic in T10 and T2 , it makes δ10 × δ2 heavy.
We now prove the existence of g. Consider a line l in T2 , disjoint from A2 (h), and the
bridge [x, y] between l and A2 (h). Let I ⊂ l be an arc containing x in its interior. By
Lemma A.14, there exists g hyperbolic in T10 and T2 whose axis in T2 contains I, hence
is disjoint from A2 (h). Being hyperbolic in T10 , the element g is hyperbolic in T1 . Its
axis intersects A1 (h) in a compact set because A1 (h) is mapped to a single point in T10
(otherwise, h would be hyperbolic in T10 ).
Proof of Proposition A.17. We assume n = 3, as the general case then follows by a straightforward induction. Let T12 be the standard common refinement of T1 , T2 (see Subsection
A.3). Let C be the core of T12 × T3 . By [Gui05, Théorème 6.1], it is enough to prove that
115
C does not contain a product of non-degenerate arcs [a12 , b12 ] × [a3 , b3 ]. Assume otherwise. Denote by a1 , b1 , a2 , b2 the images of a12 , b12 in T1 , T2 . Since a12 6= b12 , at least one
inequality a1 6= b1 or a2 6= b2 holds. Assume for instance a1 6= b1 .
We claim that [a1 , b1 ]×[a3 , b3 ] is contained in the core of T1 ×T3 , giving a contradiction.
We have to show that any quadrant Q = δ1 × δ3 of T1 × T3 intersecting [a1 , b1 ] × [a3 , b3 ] is
heavy. Denote by f : T12 → T1 the collapse map. The preimage Q̃ = f −1 (δ1 ) × δ3 of Q in
T12 × T3 is a quadrant intersecting [a12 , b12 ] × [a3 , b3 ]. Since this rectangle in contained in
C(T12 × T3 ), the quadrant Q̃ is heavy, and so is Q by Lemma A.19.
A.5
Arithmetic of trees
In this subsection, we work with simplicial trees. We let Sirr be the set of simplicial trees
T which are minimal, irreducible, with no redundant vertices and no inversion. We also
view such a T as a metric tree, by declaring each edge to be of length 1. This makes Sirr
a subset of Tirr . By Theorem A.5, a tree T ∈ Sirr is determined by its length function `.
Definition A.20 (Prime factors). The prime factors of T are the one-edge
splittings Ti
P
obtained from T by collapsing edges in all orbits but one. Clearly ` = i `i , where `i is the
length function of Ti .
We may view a prime factor of T as an orbit of edges of T , or as an edge of the quotient
graph of groups Γ = T /G. Since G is assumed to be finitely generated, there are finitely
many prime factors (by Proposition 1.7, this remains true if G is only finitely generated
relative to a finite collection of elliptic subgroups).
Lemma A.21. Let T ∈ Sirr .
(1) Any non-trivial tree T 0 obtained from T by collapses (in particular, its prime factors)
belongs to Sirr .
(2) The prime factors of T are distinct (T is “squarefree”).
Proof. Let e be any edge of T which is not collapsed in T 0 . Since T has no redundant
vertex and is not a line, either the endpoints of e are branch points u, v, or there are
branch points u, v such that [u, v] = e ∪ e0 with e0 in the same orbit as e. Using Lemma
A.3, we can find elements g, h hyperbolic in T , whose axes are not collapsed to points in
T 0 , and such that [u, v] is the bridge between their axes. Since g, h are hyperbolic with
disjoint axes in T 0 , the tree T 0 is irreducible. It is easy to check that collapsing cannot
create redundant vertices, so T 0 ∈ Sirr .
Now suppose that e, hence [u, v], gets collapsed in some prime factor T 00 . Then `0 (gh) >
0
` (g) + `0 (h) holds in T 0 but not in T 00 , so T 0 6= T 00 .
Because of this lemma, a tree of Sirr is determined by its prime factors. In particular,
T refines T 0 if and only if every prime factor of T 0 is also a prime factor of T .
Corollary A.22. Assume that T and T 0 are compatible non-trivial trees. If T is irreducible, so is T 0 . If T is not irreducible, T 0 belongs to the same deformation space as
T.
Proof. The lemma shows that performing a collapse on an irreducible simplicial tree yields
an irreducible tree (or a point). If T is not irreducible, the quotient graph of groups is a
circle and every edge e has an endpoint v such that the inclusion Ge → Gv is onto (see
Subsection 1.2). This implies that performing a collapse on a non-irreducible tree yields a
minimal non-irreducible tree belonging to the same deformation space (or a point). The
lemma follows.
116
If T1 and T2 are compatible, the standard refinement Ts with `Ts = `T1 +`T2 constructed
in Subsection A.3 is a metric tree which should be viewed as the “product” of T1 and T2 .
We shall now define the lcm T1 ∨ T2 of simplicial trees T1 and T2 . To understand the
difference between the two, suppose T1 = T2 . Then Ts is obtained from T1 by subdividing
each edge, its length function is 2`1 . On the other hand, T1 ∨ T1 = T1 .
Definition A.23 (gcd). Consider two trees T1 , T2 ∈ Sirr , with length functions `1 , `2 .
We define `1 ∧ `2 as the sum of all length functions which appear as prime factors in
both T1 and T2 . It is the length function of a tree T1 ∧ T2 (possibly a point) which is a
collapse of both T1 and T2 . We call T1 ∧ T2 the gcd of T1 and T2 .
We define `1 ∨ `2 = `1 + `2 − `1 ∧ `2 as the sum of all length functions which appear as
prime factors in T1 or T2 (or both).
Lemma A.24. Let T1 and T2 be compatible trees in Sirr . There is a tree T1 ∨ T2 ∈ Sirr
whose length function is `1 ∨ `2 . It is a common refinement of T1 and T2 , and no edge of
T1 ∨ T2 is collapsed in both T1 and T2 .
Proof. Let T be any common refinement. We modify it as follows. We collapse any edge
which is collapsed in both T1 and T2 . We then restrict to the minimal subtree and remove
redundant vertices. The resulting tree T1 ∨ T2 belongs to Sirr (it is irreducible because it
refines T1 ). It is a common refinement of T1 and T2 , and no edge is collapsed in both T1
and T2 .
We check that T1 ∨T2 has the correct length function by finding its prime factors. Since
no edge is collapsed in both T1 and T2 , a prime factor of T1 ∨ T2 is a prime factor of T1 or
T2 . Conversely, a prime factor of Ti is associated to an orbit of edges of Ti , and this orbit
lifts to T1 ∨ T2 .
Remark A.25. Unlike the standard refinement Ts , the tree T1 ∨ T2 does not have redundant
vertices.
Proposition A.26. Let T1 , . . . , Tn be pairwise compatible trees of Sirr . There exists a tree
T1 ∨ · · · ∨ Tn in Sirr whose length function is the sum of all length functions which appear
as a prime factor in some Ti . Moreover:
(1) A tree T ∈ Sirr refines T1 ∨ · · · ∨ Tn if and only if it refines each Ti .
(2) A tree T ∈ Sirr is compatible with T1 ∨ · · · ∨ Tn if and only if it is compatible with
each Ti .
(3) A subgroup H is elliptic in T1 ∨ · · · ∨ Tn if and only if it is elliptic in each Ti . If T1
dominates each Ti , then T1 ∨ · · · ∨ Tn belongs to the deformation space of T1 .
Proof. First suppose n = 2. We show that T1 ∨ T2 satisfies the additional conditions.
If T refines T1 and T2 , it refines T1 ∨ T2 because every prime factor of T1 ∨ T2 is a prime
factor of T . This proves Assertion (1).
If T1 , T2 , T are pairwise compatible, they have a common refinement T̂ by Proposition
A.17 or [SS03, Theorem 5.16] (where one should exclude ascending HNN extensions). This
T̂ refines T1 ∨ T2 by Assertion (1), so T and T1 ∨ T2 are compatible.
Assertion (3) follows from the fact that no edge of T1 ∨ T2 is collapsed in both T1 and
T2 , as in the proof of Proposition 2.2: if H fixes a point v1 ∈ T1 and is elliptic in T2 , it
fixes a point in the preimage of v1 in T1 ∨ T2 .
The case n > 2 now follows easily by induction. By Assertion (2), the tree T1 ∨· · ·∨Tn−1
is compatible with Tn , so we can define T1 ∨ · · · ∨ Tn = (T1 ∨ · · · ∨ Tn−1 ) ∨ Tn .
Definition A.27 (lcm). We call T1 ∨ · · · ∨ Tn the lcm of the compatible trees Ti .
117
A.6
Reading actions on R-trees
Rips theory gives a way to understand stable actions on R-trees, by relating them to actions
on simplicial trees. Therefore, they are closely related to JSJ decompositions. We consider
first the JSJ deformation space, then the compatibility JSJ tree.
For simplicity, we assume H = ∅.
A.6.1
Reading R-trees from the JSJ deformation space
Proposition A.28. Let G be finitely presented. Let TJ be a JSJ tree over the family of
slender subgroups. If T is an R-tree with a stable action of G whose arc stabilizers are
slender, then edge stabilizers of TJ are elliptic in T .
Recall [BF95] that an arc I ⊂ T is stable if any g ∈ G that pointwise fixes some subarc
of I also fixes I. The action of G on T is stable if every arc is stable.
Proof. By [Gui98], T is a limit of simplicial trees Tk with slender edge stabilizers. Since
TJ is universally elliptic, each edge stabilizer Ge of TJ is elliptic in every Tk . Passing to
the limit, we deduce that each element of Ge is elliptic in T . Since Ge is finitely generated,
Ge is elliptic in T .
Remark A.29. More generally, suppose that G is finitely presented and A is stable under
extension by finitely generated free abelian groups: if H < G is such that 1 → A → H →
Zk → 1, with A ∈ A, then H ∈ A. Let TJ be a JSJ tree over A with finitely generated
edge stabilizers (this exists by Theorem 2.16). If T is a stable R-tree with arc stabilizers
in A, then edge stabilizers of TJ are elliptic in T .
Recall (Theorem 6.2) that, when G is finitely presented, flexible vertices of the slender
JSJ deformation space are either slender or QH with slender fiber.
Proposition A.30. Let G, TJ , T be as in the previous proposition. There exists an R-tree
T̂ obtained by blowing up each flexible vertex v of TJ into
(1) an action by isometries on a line if Gv is slender,
(2) an action dual to a measured foliation on the underlying 2-orbifold of Gv if v is QH,
which resolves (or dominates) T in the following sense: there exists a G-equivariant map
f : T̂ → T which is piecewise linear: every segment of T̂ can be decomposed into finitely
many subsegments, in restriction to which f preserves alignment.
Proof. Using ellipticity of TJ with respect to T , we argue as in the proof of Proposition
2.2, with T1 = TJ and T2 = T . If v ∈ V (TJ ) and Gv is elliptic in T , we let Yv ⊂ T be a
fixed point. If Gv is not elliptic in T , we let Yv be its minimal subtree. It is a line if Gv
is slender. If v is a QH vertex, then Yv is dual to a measured foliation of the underlying
orbifold by Skora’s theorem [Sko96] (applied to a covering surface Σ0 ).
Remark A.31. The arguments given above may be applied in more general situations. For
instance, assume that G is finitely generated, that all subgroups of G not containing F2
are slender, and that G has a JSJ tree TJ over slender subgroups whose flexible subgroups
are QH. Let T be an R-tree with slender arc stabilizers such that G does not split over a
subgroup of the stabilizer of an unstable arc or of a tripod in T . Then, applying [Gui08]
and the techniques of [Gui98], we see that T is a limit of slender trees, so Propositions
A.28 and A.30 apply.
118
A.6.2
Reading R-trees from the compatibility JSJ tree
In [Gui00b], the first author explained how to obtain all small actions of a one-ended
hyperbolic group G on R-trees from a JSJ tree. The proof was based on Bowditch’s
construction of a JSJ tree from the topology of ∂G. Here we give a different, more general,
approach, based on Corollary A.12 (saying that compatibility is a closed condition) and
results of Subsection 11.7 (describing the compatibility JSJ space). Being universally
compatible, Tco is compatible with any R-tree which is a limit of simplicial A-trees. We
illustrate this idea in a simple case.
Let G be a one-ended finitely presented torsion-free CSA group. Assume that G is not
abelian, and let Tco be its compatibility JSJ tree over the class A of abelian groups (see
Definition 10.9 and Subsection 11.7). Let G y T be an action on an R-tree with trivial
tripod stabilizers, and abelian arc stabilizers. By [Gui98], T is a limit of simplicial A-trees.
Since Tco is compatible with all A-trees, it is compatible with T by Corollary A.12. Let
T̂ be the standard common refinement of T and Tco with length function `T + `Tco (see
Subsection A.3). Let fco : T̂ → Tco and f : T̂ → T be maps preserving alignment such
that
dT̂ (x, y) = dTco (fco (x), fco (y)) + dT (f (x), f (y)).
−1 (v)
To each vertex v and each edge e of Tco there correspond closed subtrees T̂v = fco
and T̂e = f −1 (e̊) of T̂ . By minimality, T̂e is an arc of T̂ containing no branch point except
maybe at its endpoints. The relation between dT̂ , dTco , and dT shows that the restriction
of f to T̂v is an isometric embedding. In particular, T can be obtained from T̂ by changing
the length of the arcs T̂e (possibly making the length 0).
We shall now describe the action of Gv on T̂v . Note that Gv is infinite. Its action on
T̂v need not be minimal, but it is finitely supported, see [Gui08]. Given an edge e of Tco
containing v, we denote by xe the endpoint of T̂e belonging to T̂v . If T̂v is not minimal, it
is the convex hull of the set of points xe which are extremal (i.e. T̂v \ xe is connected).
First suppose that Gv fixes some x ∈ T̂v and that T̂v is not a point. Note that, if xe 6= x
is extremal, the stabilizer of the arc [x, xe ] contains Ge , so is infinite. We claim that, if e and
f are edges of Tco containing v with xe 6= xf both extremal, then [x, xe ] ∩ [x, xf ] = {x}. If
not, the intersection is an arc [x, y]. The stabilizer of [x, y] contains hGe , Gf i and is abelian.
It follows that any point in the Gf -orbit of xe is fixed by Ge . Since tripod stabilizers are
trivial, we deduce that Gf fixes xe , a contradiction. We have thus proved that T̂v is a cone
on a finite number of orbits of points.
If Gv does not fix a point in T̂v , then it is flexible. If it is abelian, triviality of tripod
stabilizers implies that T̂v is a line. If Gv is not abelian, it is a surface group by Theorems
9.5 and 11.4. Skora’s theorem [Sko96] asserts that the minimal subtree µTv (Gv ) is dual to a
measured lamination on a compact surface. By triviality of tripod stabilizers, Tv \ µTv (Gv )
is a disjoint union of segments, and the pointwise stabilizer of each such segment has index
at most 2 in a boundary subgroup of Gv .
It follows in particular from this analysis that T̂ and T are geometric (see [LP97]).
References
[AB87]
Roger Alperin and Hyman Bass. Length functions of group actions on Λtrees. In Combinatorial group theory and topology (Alta, Utah, 1984), pages
265–378. Princeton Univ. Press, Princeton, NJ, 1987.
[Bar16]
Benjamin Barrett. Computing JSJ decompositions of hyperbolic groups,
2016. arXiv:1611.00652.
[Bee13]
Benjamin Beeker. Compatibility JSJ decomposition of graphs of free abelian
groups. Internat. J. Algebra Comput., 23(8):1837–1880, 2013.
119
[Bee14]
Benjamin Beeker. Abelian JSJ decomposition of graphs of free abelian groups.
J. Group Theory, 17(2):337–359, 2014.
[BF91]
Mladen Bestvina and Mark Feighn. Bounding the complexity of simplicial
group actions on trees. Invent. Math., 103(3):449–469, 1991.
[BF95]
Mladen Bestvina and Mark Feighn. Stable actions of groups on real trees.
Invent. Math., 121(2):287–321, 1995.
[Bow98]
Brian H. Bowditch. Cut points and canonical splittings of hyperbolic groups.
Acta Math., 180(2):145–186, 1998.
[Bow01]
Brian H. Bowditch. Peripheral splittings of groups. Trans. Amer. Math. Soc.,
353(10):4057–4082 (electronic), 2001.
[Bow12]
Brian H. Bowditch. Relatively hyperbolic groups. Internat. J. Algebra Comput., 22(3):1250016, 66, 2012.
[BKM07]
Inna Bumagin, Olga Kharlampovich, and Alexei Miasnikov. The isomorphism
problem for finitely generated fully residually free groups. J. Pure Appl.
Algebra, 208(3):961–977, 2007.
[Car11]
Mathieu Carette. The automorphism group of accessible groups. J. Lond.
Math. Soc. (2), 84(3):731–748, 2011.
[CG05]
Christophe Champetier and Vincent Guirardel. Limit groups as limits of free
groups. Israel J. Math., 146:1–75, 2005.
[Cha07]
Ruth Charney. An introduction to right-angled Artin groups. Geom. Dedicata, 125:141–158, 2007.
[Chi01]
Ian Chiswell. Introduction to Λ-trees. World Scientific Publishing Co. Inc.,
River Edge, NJ, 2001.
[Cla05]
Matt Clay. Contractibility of deformation spaces of G-trees. Algebr. Geom.
Topol., 5:1481–1503 (electronic), 2005.
[Cla09]
Matt Clay. Deformation spaces of G-trees and automorphisms of BaumslagSolitar groups. Groups Geom. Dyn., 3(1):39–69, 2009.
[Cla14]
Matt Clay. When does a right-angled Artin group split over Z? Internat. J.
Algebra Comput., 24(6):815–825, 2014.
[CF09]
Matt Clay and Max Forester. Whitehead moves for G-trees. Bull. Lond.
Math. Soc., 41(2):205–212, 2009.
[Coh89]
Daniel E. Cohen. Combinatorial group theory: a topological approach, volume 14 of London Mathematical Society Student Texts. Cambridge University
Press, Cambridge, 1989.
[CL83]
Donald J. Collins and Frank Levin. Automorphisms and Hopficity of certain
Baumslag-Solitar groups. Arch. Math. (Basel), 40(5):385–400, 1983.
[CM87]
Marc Culler and John W. Morgan. Group actions on R-trees. Proc. London
Math. Soc. (3), 55(3):571–604, 1987.
[CV86]
Marc Culler and Karen Vogtmann. Moduli of graphs and automorphisms of
free groups. Invent. Math., 84(1):91–119, 1986.
120
[DG08]
François Dahmani and Daniel Groves. The isomorphism problem for toral
relatively hyperbolic groups. Publ. Math. Inst. Hautes Études Sci., 107:211–
290, 2008.
[DG11]
François Dahmani and Vincent Guirardel. The isomorphism problem for all
hyperbolic groups. Geom. Funct. Anal., 21(2):223–300, 2011.
[DT13]
François Dahmani and Nicholas Touikan. Isomorphisms using Dehn fillings:
the splitting case, 2013. arXiv:1311.3937 [math.GR].
[Del96]
Thomas Delzant. Sous-groupes distingués et quotients des groupes hyperboliques. Duke Math. J., 83(3):661–682, 1996.
[Del99]
Thomas Delzant. Sur l’accessibilité acylindrique des groupes de présentation
finie. Ann. Inst. Fourier (Grenoble), 49(4):1215–1224, 1999.
[DD89]
Warren Dicks and M. J. Dunwoody. Groups acting on graphs, volume 17 of
Cambridge Studies in Advanced Mathematics. Cambridge University Press,
Cambridge, 1989.
[Dun85]
M. J. Dunwoody. The accessibility of finitely presented groups. Invent. Math.,
81(3):449–457, 1985.
[DS99]
M. J. Dunwoody and M. E. Sageev. JSJ-splittings for finitely presented groups
over slender groups. Invent. Math., 135(1):25–44, 1999.
[DS00]
M. J. Dunwoody and E. L. Swenson. The algebraic torus theorem. Invent.
Math., 140(3):605–637, 2000.
[Dun93]
Martin J. Dunwoody. An inaccessible group. In Geometric group theory, Vol.
1 (Sussex, 1991), pages 75–78. Cambridge Univ. Press, Cambridge, 1993.
[Far98]
B. Farb. Relatively hyperbolic groups. Geom. Funct. Anal., 8(5):810–840,
1998.
[FM99]
Benson Farb and Lee Mosher. Quasi-isometric rigidity for the solvable
Baumslag-Solitar groups. II. Invent. Math., 137(3):613–649, 1999.
[For02]
Max Forester. Deformation and rigidity of simplicial group actions on trees.
Geom. Topol., 6:219–267 (electronic), 2002.
[For03]
Max Forester. On uniqueness of JSJ decompositions of finitely generated
groups. Comment. Math. Helv., 78(4):740–751, 2003.
[For06]
Max Forester. Splittings of generalized Baumslag-Solitar groups. Geom. Dedicata, 121:43–59, 2006.
[FP06]
K. Fujiwara and P. Papasoglu. JSJ-decompositions of finitely presented
groups and complexes of groups. Geom. Funct. Anal., 16(1):70–125, 2006.
[GAnMB06] Francisco Javier González-Acuña and Juan Manuel Márquez-Bobadilla. On
the homeotopy group of the non orientable surface of genus three. Rev. Colombiana Mat., 40(2):75–79, 2006.
[GH15]
Daniel Groves and Michael Hull. Abelian splittings of right-angled Artin
groups, 2015. arXiv:1502.00129.
[Gui98]
Vincent Guirardel. Approximations of stable actions on R-trees. Comment.
Math. Helv., 73(1):89–121, 1998.
121
[Gui00a]
Vincent Guirardel. Dynamics of Out(Fn ) on the boundary of outer space.
Ann. Sci. École Norm. Sup. (4), 33(4):433–465, 2000.
[Gui00b]
Vincent Guirardel. Reading small actions of a one-ended hyperbolic group
on R-trees from its JSJ splitting. Amer. J. Math., 122(4):667–688, 2000.
[Gui05]
Vincent Guirardel. Cœur et nombre d’intersection pour les actions de groupes
sur les arbres. Ann. Sci. École Norm. Sup. (4), 38(6):847–888, 2005.
[Gui08]
Vincent Guirardel. Actions of finitely generated groups on R-trees. Ann. Inst.
Fourier (Grenoble), 58(1):159–211, 2008.
[GL07a]
Vincent Guirardel and Gilbert Levitt. Deformation spaces of trees. Groups
Geom. Dyn., 1(2):135–181, 2007.
[GL07b]
Vincent Guirardel and Gilbert Levitt. The outer space of a free product.
Proc. Lond. Math. Soc. (3), 94(3):695–714, 2007.
[GL09]
Vincent Guirardel and Gilbert Levitt. JSJ decompositions: definitions, existence and uniqueness. I: The JSJ deformation space. arXiv:0911.3173 v2
[math.GR], 2009.
[GL10]
Vincent Guirardel and Gilbert Levitt. Scott and Swarup’s regular neighbourhood as a tree of cylinders. Pacific J. Math., 245(1):79–98, 2010.
[GL11]
Vincent Guirardel and Gilbert Levitt. Trees of cylinders and canonical splittings. Geom. Topol., 15(2):977–1012, 2011.
[GL15a]
Vincent Guirardel and Gilbert Levitt. McCool groups of toral relatively hyperbolic groups. Algebr. Geom. Topol., 15(6):3485–3534, 2015.
[GL15b]
Vincent Guirardel and Gilbert Levitt. Splittings and automorphisms of relatively hyperbolic groups. Groups Geom. Dyn., 9(2):599–663, 2015.
[GL16]
Vincent Guirardel and Gilbert Levitt. Vertex finiteness for splittings of relatively hyperbolic groups. Israel J. Math., 212(2):729–755, 2016.
[GLS17]
Vincent Guirardel, Gilbert Levitt, and Rizos Sklinos. Elementary equivalence
vs commensurability for hyperbolic groups, 2017. arXiv:1701.08853.
[Hru10]
G. Christopher Hruska. Relative hyperbolicity and relative quasiconvexity
for countable groups. Algebr. Geom. Topol., 10(3):1807–1856, 2010.
[JS79]
William H. Jaco and Peter B. Shalen. Seifert fibered spaces in 3-manifolds.
Mem. Amer. Math. Soc., 21(220):viii+192, 1979.
[Joh79]
Klaus Johannson. Homotopy equivalences of 3-manifolds with boundaries,
volume 761 of Lecture Notes in Mathematics. Springer, Berlin, 1979.
[KM05]
Olga Kharlampovich and Alexei G. Myasnikov. Effective JSJ decompositions.
In Groups, languages, algorithms, volume 378 of Contemp. Math., pages 87–
212. Amer. Math. Soc., Providence, RI, 2005.
[Kou98]
Malik Koubi. Croissance uniforme dans les groupes hyperboliques. Ann. Inst.
Fourier (Grenoble), 48(5):1441–1453, 1998.
[Kro90]
P. H. Kropholler. An analogue of the torus decomposition theorem for certain
Poincaré duality groups. Proc. London Math. Soc. (3), 60(3):503–529, 1990.
122
[KR89a]
P. H. Kropholler and M. A. Roller. Relative ends and duality groups. J. Pure
Appl. Algebra, 61(2):197–210, 1989.
[KR89b]
P. H. Kropholler and M. A. Roller. Splittings of Poincaré duality groups. III.
J. London Math. Soc. (2), 39(2):271–284, 1989.
[Lev05a]
Gilbert Levitt. Automorphisms of hyperbolic groups and graphs of groups.
Geom. Dedicata, 114:49–70, 2005.
[Lev05b]
Gilbert Levitt. Characterizing rigid simplicial actions on trees. In Geometric
methods in group theory, volume 372 of Contemp. Math., pages 27–33. Amer.
Math. Soc., Providence, RI, 2005.
[LP97]
Gilbert Levitt and Frédéric Paulin. Geometric group actions on trees. Amer.
J. Math., 119(1):83–102, 1997.
[Lin83]
P. A. Linnell. On accessibility of groups. J. Pure Appl. Algebra, 30(1):39–46,
1983.
[MM96]
Darryl McCullough and Andy Miller. Symmetric automorphisms of free products. Mem. Amer. Math. Soc., 122(582):viii+97, 1996.
[MNS99]
C. F. Miller, III, Walter D. Neumann, and G. A. Swarup. Some examples of
hyperbolic groups. In Geometric group theory down under (Canberra, 1996),
pages 195–202. de Gruyter, Berlin, 1999.
[MS84]
John W. Morgan and Peter B. Shalen. Valuations, trees, and degenerations
of hyperbolic structures. I. Ann. of Math. (2), 120(3):401–476, 1984.
[Osi06]
Denis V. Osin. Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems. Mem. Amer. Math. Soc.,
179(843):vi+100, 2006.
[PS09]
Panos Papasoglu and Eric Swenson. Boundaries and JSJ decompositions of
CAT(0)-groups. Geom. Funct. Anal., 19(2):559–590, 2009.
[Pau89]
Frédéric Paulin. The Gromov topology on R-trees. Topology Appl., 32(3):197–
221, 1989.
[Pau91]
Frédéric Paulin. Outer automorphisms of hyperbolic groups and small actions
on R-trees. In Arboreal group theory (Berkeley, CA, 1988), pages 331–343.
Springer, New York, 1991.
[Pau04]
Frédéric Paulin. Sur la théorie élémentaire des groupes libres (d’après Sela).
Astérisque, 294:ix, 363–402, 2004.
[Per11]
Chloé Perin. Elementary embeddings in torsion-free hyperbolic groups. Ann.
Sci. Éc. Norm. Supér. (4), 44(4):631–681, 2011.
[RW10]
Cornelius Reinfeldt and Richard Weidmann. Makanin-Razborov diagrams for
hyperbolic groups, 2010. preprint http://www.math.uni-kiel.de/algebra/
de/weidmann/research/.
[RS97]
E. Rips and Z. Sela. Cyclic splittings of finitely presented groups and the
canonical JSJ decomposition. Ann. of Math. (2), 146(1):53–109, 1997.
[Sco83]
Peter Scott. The geometries of 3-manifolds.
15(5):401–487, 1983.
123
Bull. London Math. Soc.,
[SS00]
Peter Scott and Gadde A. Swarup. Splittings of groups and intersection
numbers. Geom. Topol., 4:179–218 (electronic), 2000.
[SS03]
Peter Scott and Gadde A. Swarup. Regular neighbourhoods and canonical decompositions for groups. Astérisque, 289:vi+233, 2003. Corrections available
at http://www.math.lsa.umich.edu/~pscott/.
[SW79]
Peter Scott and Terry Wall. Topological methods in group theory. In Homological group theory (Proc. Sympos., Durham, 1977), pages 137–203. Cambridge Univ. Press, Cambridge, 1979.
[Sel97a]
Z. Sela. Acylindrical accessibility for groups. Invent. Math., 129(3):527–565,
1997.
[Sel97b]
Z. Sela. Structure and rigidity in (Gromov) hyperbolic groups and discrete
groups in rank 1 Lie groups. II. Geom. Funct. Anal., 7(3):561–593, 1997.
[Sel99]
Z. Sela. Endomorphisms of hyperbolic groups. I. The Hopf property. Topology,
38(2):301–321, 1999.
[Sel01]
Zlil Sela. Diophantine geometry over groups. I. Makanin-Razborov diagrams.
Publ. Math. Inst. Hautes Études Sci., 93:31–105, 2001.
[Ser77]
Jean-Pierre Serre. Arbres, amalgames, SL2 . Société Mathématique de France,
Paris, 1977. Rédigé avec la collaboration de Hyman Bass, Astérisque, No. 46.
[Sha87]
Peter B. Shalen. Dendrology of groups: an introduction. In Essays in group
theory, pages 265–319. Springer, New York-Berlin, 1987.
[Sha91]
Peter B. Shalen. Dendrology and its applications. In Group theory from a
geometrical viewpoint (Trieste, 1990), pages 543–616. World Sci. Publishing,
River Edge, NJ, 1991.
[Sko96]
Richard K. Skora. Splittings of surfaces. J. Amer. Math. Soc., 9(2):605–616,
1996.
[Sta83]
John R. Stallings. Topology of finite graphs. Invent. Math., 71(3):551–565,
1983.
[Thu80]
William P. Thurston. The geometry and topology of three-manifolds. Princeton lecture notes, 1980.
[Tou09]
Nicholas W.M. Touikan. Detecting geometric splittings in finitely presented
groups. arXiv:0906.3902v2 [math.GR]. To appear in Transactions of the
A.M.S., 2009.
[Wal04]
C. T. C. Wall. Poincaré duality in dimension 3. In Proceedings of the Casson
Fest, volume 7 of Geom. Topol. Monogr., pages 1–26 (electronic). Geom.
Topol. Publ., Coventry, 2004.
[Wei12]
Richard Weidmann. On accessibility of finitely generated groups. Q. J. Math.,
63(1):211–225, 2012.
124
Index
A(g): axis of g . . . . . . . . . . . . . . . . . . . . . . . . 16
[A]: the equivalence class of A ∈ A∞ . . 74
A: allowed edge stabilizers . . . . . . . . . . . . 18
Aell : universally elliptic groups in A . . 25
A∞ : infinite groups in A . . . . . . . . . . . . . 74
A(g): characteristic set of g . . . . . . . . . . 109
(A, H)-tree . . . . . . . . . . . . . . . . . . . . . . . . . 7, 18
A-tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3, 18
Av : groups of A in Gv . . . . . . . . . . . . . . . . 37
c(A): complexity of an edge group A . . 69
C(T1 , T2 ): core of T1 × T2 . . . . . . . . . . . . . 60
ˇ 1 , T2 ): flipped core of T1 × T2 . . . . . 60
C(T
C (2) : union of squares in C(T1 , T2 ) . . . . 61
C-virtually cyclic . . . . . . . . . . . . . . . . . 15, 77
Dco : compatibility JSJ space . . . . . 10, 101
DJSJ : the JSJ deformation space . . . . . 25
E(T ): edge set of T . . . . . . . . . . . . . . . . . . 16
(FA): Serre’s property . . . . . . . . . . . . . . . . 16
Fn , the free group on n generators . . . . 15
Gv , Ge : vertex or edge stabilizer/group
16
G[A] : the stabilizer of [A] . . . . . . . . . . . . . 74
H: relative structure . . . . . . . . . . . . . . . . . . 18
H|Gv . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
H
Incv , IncH
v , Inc|Q , Inc|Q : incident structures
35
K-CSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
(k, C)-acylindrical . . . . . . . . . . . . . . . . . 73, 86
`(g): the translation length of g . . . . . 109
`, `T : the length function of T . . . . . . . 110
`1 ∧ `2 : gcd of two length functions . . 117
`1 ∨ `2 : lcm of two length functions . . 117
µT (H): minimal H-invariant subtree . . 17
M (g), M (H): the maximal virtually abelian
group containing g, H . . . 91, 92
RN (T1 , T2 ): the regular neighbourhood of two
trees . . . . . . . . . . . . . . . . . . . . . . . . 62
Sirr : set of irreducible simplicial trees
116
(SC), (SCZ ): stability conditions . 55, 57
S: a family of small subgroups . . . . . . . . 78
Snvc : groups in S not C-virtually cyclic . . .
87
Tc : tree of cylinders . . . . . . . . . . . . . . . . . . 75
Tc∗ : collapsed tree of cylinders . . . . . . . . 75
Tco : compatibility JSJ tree . . . . . . . 10, 104
T1 ∧ T2 : gcd of two trees . . . . . . . . . . . . . 117
T1 ∨ T2 : lcm of compatible trees . . . . . 117
T : set of actions on R-trees . . . . . . . . . . 110
Tirr : set of irreducible actions on R-trees . .
110
V (T ): vertex set of T . . . . . . . . . . . . . . . . . 16
Z: virtually cyclic subgroups with infinite
center . . . . . . . . . . . . . . . . . . . . . . . 97
Zmax : maximal virtually cyclic subgroups with
infinite center . . . . . . . . . . . . . . . 97
abelian tree . 8, 18, 54, 55, 86, 91, 107, 119
accessibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
accessibility (Dunwoody) . . . . . . . . . . 26, 28
acylindrical accessibility . . . . . . 4, 7, 21, 83
acylindrical tree, splitting
8, 21, 73, 76, 86
admissible equivalence relation on A∞
74
alignment preserving map . . . . . . . . 18, 109
approximation point . . . . . . . . . . . . . . . . . 111
arc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
axes topology . . . . . . . . . . . . . . . . . . . . . . . . 111
axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Baumslag-Solitar group
6, 10, 31, 96, 105
boundary (in an orbifold) . . . . . . . . . . . . . 41
boundary subgroup . . . . . . . . . . . . . . . . . 6, 41
Bowditch . . . . . . . . . . 2, 5, 9, 73, 95, 97, 119
bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16, 109
canonical tree . . . . . . 5, 10, 73, 99, 103, 105
characteristic set . . . . . . . . . . . . . . . . . 16, 109
co-elementary subgroups . . . . . . . . 8, 74, 94
collapse map . . . . . . . . . . . . . . . . . . . 9, 18, 109
collapse of a tree . . . . . . . . . . . . . . . . . . . . . . 19
collapsed tree of cylinders . . . . . . . . . . . . . 75
commensurable . . . . . . . . . . . . 15, 74, 96, 107
commensurator . . . . . . . . . . . . 15, 32, 74, 108
common refinement . . . . . . . . . . . . . . . 19, 115
commutative transitive . . . . . . . . . . . 96, 108
compatibility JSJ deformation space . 10,
100
compatibility JSJ tree . . . . . . . . . . . 10, 104
compatible trees . . 9, 19, 79, 100, 109, 113
conical point . . . . . . . . . . . . . . . . . . . . . . . . . . 42
core in the product of two trees . . . 58, 60,
115
corner reflector . . . . . . . . . . . . . . . . . . . . . . . . 42
CSA group
6, 8, 9, 16, 56, 74, 86, 91, 107,
119
cyclic tree . . . . . . . . . . . . . . . . . 18, 57, 91, 107
cylinder in a tree . . . . . . . . . . . . . . . . . . . . 8, 75
deformation space . . . . . . . . . . . . . . . . . . 5, 19
degenerate segment . . . . . . . . . . . . . . 16, 109
dihedral action . . . . . . . . . . . . . . . . . . . . . . . . 17
dihedral group . . . . . . . . . . . . . . . . . . . . . . . . 15
domination . . . . . . . . . . . . . . . . . . . . . . . . . 3, 19
dual splitting . . . . . . . . . . . . . . . . . . . . . . 42, 47
Dunwoody-Sageev . . . . . . . . . 2, 6, 20, 29, 50
elementary subgroup . . . . . . . . . . . . . . 15, 94
elliptic element or subgroup . . . . . . . . 4, 16
elliptic with respect to a tree . . . . . . . . . 22
end of a tree . . . . . . . . . . . . . . . . . . . . . . 16, 110
Euler characteristic . . . . . . . . . . . . . . . . . . . 42
extended boundary subgroup . . . . . . . . . 47
125
fiber (of a QH subgroup) . . . . . . . 6, 47, 55
filling construction . . . . . . . . . . . . . . . . . . . . 39
flexible vertex, group, stabilizer
6, 26, 33,
38, 40, 54, 56, 71, 86, 91, 95
fold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
free splitting . . . . . . . . . . . . . . . . . . . 3, 30, 105
freely indecomposable relative to H . . . 18
Fujiwara-Papasoglu . . . . . . . 2, 6, 29, 50, 55
fully hyperbolic tree . . . . . . . . . . . . . . . . . . 60
QH, quadratically hanging .
56, 86
R-tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
RAAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
reduced tree . . . . . . . . . . . . . . . . . . . . . . 19, 103
redundant vertex . . . . . . . . . . . . . . . . . . . . . . 16
refinement . . . . . . . . . . . . . . . . . . . 9, 19, 22, 38
refining T at a vertex v . . . . . . . . . . . . . . . 37
regular neighbourhood . 2, 54, 58, 62, 105
relative finite generation, presentation
21,
36
relative generating set . . . . . . . . . . . . . . . . 21
relative tree or splitting . . . . . . . . . . . . 7, 18
relatively hyperbolic group . . 6, 9, 15, 74,
94, 107
restriction of H . . . . . . . . . . . . . . . . . . . . . . . 35
rigid vertex, group, stabilizer . . . . . . . 6, 26
Rips-Sela . . . . . . . . . . . . . . . . . 2, 6, 29, 50, 56
Γ-limit group . . . . . . . . . . . . . . . . . . . . . . 9, 93
gcd of two trees . . . . . . . . . . . . . . . . . . 10, 117
generalized Baumslag-Solitar group
6, 10,
32, 105
geodesic (in an orbifold)
closed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
essential simple closed . . . . . . . . . . . . 42
filling the orbifold . . . . . . . . . . . . . . . . 44
simple . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
graph of groups . . . . . . . . . . . . . . . . . . . . . . . 16
(equivariant) Gromov-Hausdorff topology .
82, 111
Grushko decomposition, deformation space
1, 3, 30, 39, 105
sandwich closed family of subgroups . . 74
Scott-Swarup . . . . . . . . . . . . . . . 2, 10, 29, 105
segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
semi-simple action on a tree . . . . . . . . . 110
Serre’s property (FA), Serre’s lemma . . 16
slender group 6, 18, 19, 49, 55, 56, 71, 118
slender in (A, H)-trees . . . . . . . . . . . . 20, 70
small domination, S-domination . . 77, 86
small orbifold . . . . . . . . . . . . . . . . . . . . . . . . . 44
small, small in (A, H)-trees . 8, 20, 31, 32,
77
socket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
dual to a family of geodesics . . 42, 47
splitting over A . . . . . . . . . . . . . . . . . . . . . . . 18
stability condition . . . . . . . . . . . . . . 6, 47, 85
stability condition (SC) . . . . . . . . . . . . . . 55
stability condition (SCZ ) . . . . . . . . . . . . . 57
stable action on an R-tree . . . . . . . . . . . 118
Stallings-Dunwoody deformation space . . . .
31, 39
standard common refinement . . . . . . . . 113
standard refinement . . . . . . . . . . . . . . . . . . 23
horizontal edge of the core . . . . . . . . . . . . 60
hyperbolic element, subgroup . . . . . 16, 17
incidence structure . . . . . . . . . . . . . . . . . . . 35
incident edge groups . . . . . . . . . . . . . . . . . . 35
irreducible tree, deformation space
10, 17,
19, 103, 110
JSJ decomposition, tree, deformation space
4, 25
K-CSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
(k, C)-acylindrical . . . . . . . . . . . . . . . . . 73, 86
lcm of pairwise compatible trees . 10, 104,
117
length function . . . . . . . . . . . . . . . . . . . . . . . 109
map between trees . . . . . . . . . . . . . . . . . . . . . 18
minimal action . . . . . . . . . . . . . . . . . . . 16, 110
minimal subtree . . . . . . . . . . . . . . . . . . . 17, 38
minuscule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
mirror . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41, 57
morphism between trees . . . . . . . . . . . . . . 18
one-edge splitting . . . . . . . . . . . . . . . . . 16,
one-ended relative to H . . . . . . . . . . . 18,
orbifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6,
small orbifolds . . . . . . . . . . . . . . . . . . .
with finite mapping class group . .
outer space . . . . . . . . . . . . . . . . . . . 2, 5, 19,
6, 40, 47, 50,
topology
axes topology . . . . . . . . . . . . . . . . . . . 111
equivariant Gromov-Hausdorff topology
82, 111
totally flexible group . . . . . . . . . . . . . . . . . . 58
translation length . . . . . . . . . . . . . . . . . . . 109
tree of cylinders . . 6, 8, 10, 73, 75, 79, 106
trivial action . . . . . . . . . . . . . . . . . . . . . . . . . . 16
trivial deformation space . . . . . . . . . . . . . . 19
25
39
41
44
46
30
universally compatible tree . . . . . . . . 9, 100
universally elliptic subgroup, tree . . 3, 24,
38
used boundary component . . . . . . . . . . . . 47
parabolic . . . . . . . . . . . . . . . . . . . . . . . 15, 33, 94
peripheral structure . . . . . . . . . . . . . . . . . . 52
prime factors of a tree . . . . . . . . . . . . . . . 116
vertical edge of the core . . . . . . . . . . . . . .
126
60
virtual commutation . . . . . . . . . . . . . . . . . . 93
K-virtually abelian . . . . . . . . . . . . . . . . . . . 91
virtually cyclic . . . . . . . . . . . . . 15, 56, 74, 96
C-virtually cyclic . . . . . . . . . . . . . . . . . 15, 77
VPC . . . . . . . . . . . . . . . . . . . . . . . . . . 52, 56, 105
127
Vincent Guirardel
Institut de Recherche Mathématique de Rennes
Université de Rennes 1 et CNRS (UMR 6625)
263 avenue du Général Leclerc, CS 74205
F-35042 RENNES Cédex
e-mail: [email protected]
Gilbert Levitt
Laboratoire de Mathématiques Nicolas Oresme (LMNO)
Université de Caen et CNRS (UMR 6139)
(Pour Shanghai : Normandie Univ, UNICAEN, CNRS, LMNO, 14000 Caen, France)
e-mail: [email protected]
128
| 4math.GR
|
ON THE COMPRESSION OF RECURRENT NEURAL NETWORKS WITH AN APPLICATION
TO LVCSR ACOUSTIC MODELING FOR EMBEDDED SPEECH RECOGNITION
Rohit Prabhavalkar†
Ouais Alsharif †
Antoine Bruguier
Ian McGraw
arXiv:1603.08042v2 [cs.CL] 2 May 2016
Google Inc.
{prabhavalkar,oalsha,tonybruguier,imcgraw}@google.com
ABSTRACT
We study the problem of compressing recurrent neural networks (RNNs). In particular, we focus on the compression
of RNN acoustic models, which are motivated by the goal
of building compact and accurate speech recognition systems
which can be run efficiently on mobile devices. In this work,
we present a technique for general recurrent model compression that jointly compresses both recurrent and non-recurrent
inter-layer weight matrices. We find that the proposed technique allows us to reduce the size of our Long Short-Term
Memory (LSTM) acoustic model to a third of its original size
with negligible loss in accuracy.
Index Terms— model compression, LSTM, RNN, SVD,
embedded speech recognition
1. INTRODUCTION
Neural networks (NNs) with multiple feed-forward [1, 2] or
recurrent hidden layers [3, 4] have emerged as state-of-theart acoustic models (AMs) for automatic speech recognition
(ASR) tasks. Advances in computational capabilities coupled
with the availability of large annotated speech corpora have
made it possible to train NN-based AMs with a large number
of parameters [5] with great success.
As speech recognition technologies continue to improve,
they are becoming increasingly ubiquitous on mobile devices:
voice assistants such as Apple’s Siri, Microsoft’s Cortana,
Amazon’s Alexa and Google Now [6] enable users to search
for information using their voice. Although the traditional
model for these applications has been to recognize speech
remotely on large servers, there has been growing interest
in developing ASR technologies that can recognize the input speech directly “on-device” [7]. This has the promise to
reduce latency while enabling user interaction even in cases
where a mobile data connection is either unavailable, slow
or unreliable. Some of the main challenges in this regard
are the disk, memory and computational constraints imposed
by these devices. Since the number of operations in neural
† Equal contribution. The authors would like to thank Haşim Sak and
Raziel Alvarez for helpful comments and suggestions on this work, and Chris
Thornton and Yu-hsin Chen for comments on an earlier draft.
networks is proportional to the number of model parameters,
compressing the model is desirable from the point of view of
reducing memory usage and power consumption.
In this paper, we study techniques for compressing recurrent neural networks (RNNs), specifically RNN acoustic
models. We demonstrate how a generalization of conventional inter-layer matrix factorization techniques (e.g., [8,
9]), where we jointly compress both recurrent and inter-layer
weight matrices, allows us to compress acoustic models up
to a third of their original size with negligible loss in accuracy. While we focus on acoustic modeling, the techniques
presented can be applied to RNNs in other domains, e.g.,
handwriting recognition [10] and machine translation [11] inter alia. The technique presented in this paper encompasses
both traditional recurrent neural networks (RNNs) as well as
Long Short-Term Memory (LSTM) neural networks.
In Section 2, we review previous work that has focussed
on techniques for compressing neural networks. Our proposed compression technique is presented in Section 3. We
examine the effectiveness of proposed techniques in Sections 4 and 5. Finally, we conclude with a discussion of our
findings in Section 6.
2. RELATED WORK
There have been a number of previous proposals to compress
neural networks, both in the context of ASR as well as in the
broader field of machine learning. We summarize a number
of proposed approaches in this section.
It has been noted in previous work that there is a large
amount of redundancy in the parameters of a neural network.
For example, Denil et al. [12] show that the entire neural network can be reconstructed given the values of a small number
of parameters. Caruana and colleagues show that the output
distribution learned by a larger neural network can be approximated by a neural network with fewer parameters by training the smaller network to directly predict the outputs of the
larger network [13, 14]. This approach, termed “model compression” [13] is closely related to the recent “distillation”
approach proposed by Hinton et al. [15]. The redundancy in
a neural network has also been exploited in the HashNet approach of Chen et al. [16], which imposes parameter tying in
network based on a set of hash functions.
In the context of ASR, previous approaches to acoustic
model compression have focused mainly on the case of feedforward DNNs. One popular technique is based on sparsifying the weight matrices in the neural network, for example, by
setting weights whose magnitude falls below a certain threshold to zero [1] or based on the second-derivative of the loss
function in the “optimal brain damage” procedure [17]. In
fact, Seide et al. [1] demonstrate that up to two-thirds of the
weights of the feed-forward network can be set to zero without incurring any loss in performance. Although techniques
based on sparsification do decrease the number of effective
weights, encoding the subset of weights which can be ‘zeroed
out’ requires additional memory. Further, if the weight matrices are represented as dense matrices for efficient computation, then the parameter savings on disk will not translate in to
savings of runtime memory. Other techniques to reduce the
number of model parameters is based on changing the neural network architecture, e.g., by introducing bottleneck layers [18] or through a low-rank matrix factorization layer [19].
We also note recent work by Wang et al. [20] which uses a
combination of singular value decomposition (SVD) and vector quantization to compress acoustic models.
The methods investigated in our work are most similar to
previous work that has examined using SVD to reduce the
number of parameters in the network in the context of feedforward DNNs [8, 9, 21]. As we describe in Section 3, our
methods can be thought of as an extension of the techniques
proposed by Xue et al. [8], wherein we jointly factorize both
recurrent and (non-recurrent) inter-layer weight matrices in
the network.
3. MODEL COMPRESSION
In this section, we present a general technique for compressing individual recurrent layers in a recurrent neural network,
thus generalizing the methods proposed by Xue et al. [8].
We describe our approach in the most general setting of
a standard RNN. We denote the activations of the l-th hidden
l
layer, consisting of N l nodes, at time t by hlt ∈ RN . The
inputs to this layer at time t – which are in turn the activations
from the previous layer or the input features – are denoted by
l−1
hl−1
∈ RN . We can then write the following equations
t
which define the output activations of the l-th and (l + 1)-th
layers in a standard RNN:
hlt
hl+1
t
=
=
σ(Wxl−1 hl−1
t
l l
σ(Wx ht
l
+
+
Whl hlt−1
Whl+1 hl+1
t−1
l
+b )
l+1
+b
Fig. 1. The initial model (Figure (a)) is compressed by jointly
factorizing recurrent (Whl ) and inter-layer (Wxl ) matrices, using a shared recurrent projection matrix (P l ) [3] (Figure (b)).
weight matrices, respectively1 . Since our proposed approach
can be applied independently for each recurrent hidden layer,
we only describe the compression operations for a particular layer l. We jointly compress the recurrent and inter-layer
matrices corresponding to a specific layer l by determining
a suitable recurrent projection matrix [3], denoted by P l ∈
l
l
Rr ×N , of rank rl < N l such that, Whl = Zhl P l and Wxl =
l l
Zx P , thus allowing us to re-write (1) and (2) as,
hlt = σ(Wxl−1 hl−1
t
hl+1
t
=
+ Zhl P l hlt−1
σ(Zxl P l hlt
l
+
l+1
l
(2)
l+1
+b
(3)
)
(4)
l
where, Zhl ∈ RN ×r and Zxl ∈ RN ×r . This compression
process is depicted graphically in Figure 1.
We note that sharing P l across the recurrent and interlayer matrices allows for more efficient parameterization of
the weight matrices; as shown in Section 5, this does not result in a significant loss of performance. Thus, the degree
of compression in the model can be controlled by setting the
ranks rl of the projection matrices in each of the layers of the
network.
We determine the recurrent projection matrix P l , by first
computing an SVD of the recurrent weight matrix, which we
then truncate, retaining only the top rl singular values (defl ) and the corresponding singular vectors from U l
noted by Σ
h
h
fl and V
fl , respectively):
and V l (denoted by U
h
h
h
T
T
l l
fl Σ
fl fl
Whl = Uhl Σlh Vhl ≈ U
h h Vh = Zh P
(1)
)
Whl+1 hl+1
t−1
+ bl )
(5)
T
fl Σ
fl
fl
l
where Zhl = U
h h and P = Vh . Finally, we determine
l
Zx , as the solution to the following least-squares problem:
l+1
where, bl ∈ RN and bl+1 ∈ RN
represent bias vectors, σ(·) denotes a non-linear activation function, and Wxl ∈
l+1
l
l
l
RN ×N and Whl ∈ RN ×N denote weight matrices that
we refer to respectively as the inter-layer and the recurrent
Zxl = arg min kY P l − Wxl k2F
(6)
Y
1 The equations are slightly more complicated when using LSTM cells in
the recurrent layer, but the basic form remains the same. See Section 3.1.
where, kXkF denotes the Frobenius norm of the matrix. In
pilot experiments we found that the proposed SVD-based initialization performed better than training a model with recurrent projection matrices (i.e., same model architecture) but
with random initialization of the network weights.
3.1. Applying our technique to LSTM RNNs
Generalizing the procedure described above in the context
of standard RNNs to the case of LSTM RNNs [3, 22, 23]
is straightforward. Using the notation in [3], note that the
recurrent-weight matrix Whl in the case of the LSTM is the
concatenation of the four gate weight matrices, obtained by
stacking them vertically:
[Wim , Wom , Wf m , Wcm ]
T
which represent respectively, recurrent connections to the input gate, the output gate, the forget gate and the cell state.
Similarly, the inter-layer matrix Wxl is the concatenation of
the matrices:
T
[Wix , Wf x , Wox , Wcx ]
which correspond to the input gate, the forget gate, the output
gate and the cell state (of the next layer). With these definitions, compression can be applied as described in Section 3.
Note that we do not compress the “peep-hole” weights, since
they are already narrow, single column matrices and do not
contribute significantly to the total number of parameters in
the network.
4. EXPERIMENTAL SETUP
In order to determine the effectiveness of the proposed RNN
compression technique, we conduct experiments on a openended large-vocabulary dictation task.
As we mentioned in Section 1, one of our primary motivations behind investigating acoustic model compression is to
build compact acoustic models that can be deployed on mobile devices. In recent work, Sak et al. have demonstrated
that deep LSTM-based AMs trained to predict either contextindependent (CI) phoneme targets [22] or context-dependent
(CD) phoneme targets [23] approach state-of-the-art performance on speech tasks. These systems have two important
characteristics: in addition to the CI or CD phoneme labels,
the system can also hypothesize a “blank” label if it is unsure
of the identity of the current phoneme, and the systems are
trained to optimize the connectionist temporal classification
(CTC) criterion [24] which maximizes the total probability
of correct label sequence conditioned on the input sequence.
More details can be found in [22, 23].
Following [22], our baseline model is thus a CTC model:
a five hidden layer RNN with 500 LSTM cells in each layer,
which predicts 41 CI phonemes (plus “blank”). As a point
of comparison, we also present results obtained using a much
larger state-of-the-art ‘server-sized’ model which is too large
to deploy on embedded devices but nonethless serves as an
upper-bound performance for our models on this dataset. This
model consists of five hidden layers with 600 LSTM cells per
layer, and is trained to predict one of 9287 context-dependent
phonemes (plus “blank”).
Our systems are trained using distributed asynchronous
stochastic gradient descent with a parameter server [25].
The systems are first trained to convergence to optimize the
CTC criterion, following which these are discriminatively
sequence trained to optimize the state-level minimum Bayes
risk (sMBR) criterion [26, 27]. As discussed in Section 5,
after applying the proposed compression scheme, we further
fine-tune the network: first with the CTC criterion, followed
by sequence discriminative training with the sMBR criterion.
This additional fine-tuning step was found to be necessary
to achieve good performance, particularly as the amount of
compression was increased.
The language model used in this work is a 5-gram model
trained on ∼100M sentences of in-domain data, with entropybased pruning applied to reduce the size of the LM down to
roughly 1.5M n-grams (mainly bigrams) with a 64K vocabulary. Since our goal is to build a recognizer to run efficiently
on mobile devices, we minimize the size of the decoder graph
used for recognition, following the approach outlined in [7]:
we perform an additional pruning step to generate a much
smaller first-pass language model (69.5K n-grams; mainly
unigrams), which is composed with the lexicon transducer
to construct the decoder graph. We then perform on-the-fly
rescoring with the larger LM. The resulting models, when
compressed for use on-device, total about 20.3 MB, thus enabling them to be run many times faster than real-time on
recent mobile devices [28].
We parameterize the input acoustics by computing 40dimensional log mel-filterbank energies over the 8Khz range,
which are computed every 10ms over 25ms windowed speech
segments. The server-sized system uses 80-dimensional features computed over the same range since this resulted in
slightly improved performance. Following [23], we stabilize CTC training by stacking together 8 consecutive speech
frames (7 right context frames); only every third stacked
frame is presented as an input to the network.
4.1. Training and Evaluation Data
Our systems are trained on ∼3M hand-transcribed anonymized
utterances extracted from Google voice search traffic (∼2000
hours). We create “multi-style” training data by synthetically distorting utterances to simulate background noise and
reverberation using a room simulator with noise samples
extracted from YouTube videos and environmental recordings of everyday events; 20 distorted examples are created
for each utterance in the training set. Systems are additionally adapted using the sMBR criterion [26, 27] on a set of
∼1M anonymized hand-transcribed (in-domain) dictation
utterances extracted from Google traffic, processed to generate “multi-style” training data as described above, which
improves performance on our dictation task. All results are
reported on a set of 13.3K hand-transcribed anonymized utterances extracted from Google traffic from an open-ended
dictation domain.
5. RESULTS
In our experiments, we seek to determine the impact of the
proposed joint SVD-based compression technique on system
performance. In particular, we are interested in determining
how system performance varies as a function of the degree of
compression, which is controlled by setting the ranks of the
recurrent projection matrices rl as described in Section 3.
Notice that since the proposed compression scheme is applied to all hidden layers of the baseline system, there are
numerous settings of the ranks rl for the projection matrices
in each layer which result in the same number of total parameters in the compressed network. In order to avoid this ambiguity, we set the various projection ranks using the following
criterion: Given a threshold τ , for each layer l, we set the rank
rl of the corresponding projection matrix such that it corresponds to retaining a fraction of at most τ of the explained
variance after the truncated SVD of Whl . More specifically,
if the singular values in Σlh in (5) are sorted in non-increasing
l
, we set each rl as:
order as σ1l ≥ σ2l ≥ · · · ≥ σN
( Pk
rl = arg max
1≤k≤N
j=1
PN
j=1
σjl
2
σjl
2
)
≤τ
(7)
Choosing the projection ranks using (7) allows us to control
the degree of compression, and thus compressed model size
by varying a single parameter, τ . In pilot experiments we
found that this scheme performed better than setting ranks to
be equal for all layers (given the same total parameter budget). Once the projection ranks rl have been determined for
the various projection matrices we fine-tune the compressed
models by first optimizing the CTC criterion, followed by sequence training with the sMBR criterion and adaptation on
in-domain data as described in Section 4.1. The results of our
experiments are presented in Table 1.
As can be seen in Table 1, the baseline system which predicts CI phoneme targets is only ∼10% relative worse than
the larger server-sized system, although it has half as many
parameters. Since the ranks rl are all chosen to retain a given
fraction of the explained variance in the SVD operation, we
also note that earlier hidden layers in the network appear to
have lower ranks than later layers, since most of the variance
is accounted for by a smaller number of singular values. It
can be seen from Table 1 that word error rates increase as the
amount of compression is increased, although performance of
System
server
baseline
τ = 0.95
τ = 0.90
τ = 0.80
τ = 0.70
τ = 0.60
τ = 0.50
τ = 0.40
Projection ranks, rl
350, 375, 395, 405, 410
270, 305, 335, 345, 350
175, 215, 245, 260, 265
120, 150, 180, 195, 200
80, 105, 130, 145, 150
50, 70, 90, 100, 110
30, 45, 55, 65, 75
Params
20.1M
9.7M
8.6M
7.2M
5.4M
4.1M
3.1M
2.3M
1.7M
WER
11.3
12.4
12.3
12.5
12.5
12.6
12.9
13.2
14.4
Table 1. Word error rates (%) on the test set as a function
of the percentage of explained variance retained (τ ) after the
SVDs of the recurrent weight matrices Whl in the hidden layers of the RNN.
the compressed systems are close to the baseline for moderate compression (τ ≥ 0.7). Using a value of τ = 0.6, enables
the model to be compressed to a third of its original size, with
only a small degradation in accuracy. However, performance
begins to degrade significantly for τ ≤ 0.5. Future work will
consider alternative techniques for setting the projection ranks
rl in order to examine their impact on system performance.
6. CONCLUSIONS
We presented a technique to compress RNNs using a joint factorization of recurrent and inter-layer weight matrices, generalizing previous work [8]. The proposed technique was applied to the task of compressing LSTM RNN acoustic models
for embedded speech recognition, where we found that we
could compress our baseline acoustic model to a third of its
original size with negligible loss in accuracy. The proposed
techniques, in combination with weight quantization, allow us
to build a small and efficient speech recognizer that run many
times faster than real-time on recent mobile devices [28].
7. REFERENCES
[1] F. Seide, G. Li, and D. Yu, “Conversational speech
transcription using context-dependent deep neural networks,” in Proc. of Interspeech, 2011, pp. 437–440.
[2] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-R. Mohamed,
N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. N.
Sainath, and B. Kingsbury, “Deep neural networks for
acoustic modeling in speech recognition: The shared
views of four research groups,” IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 82–97, 2012.
[3] H. Sak, A. Senior, and F. Beaufays, “Long short-term
memory recurrent neural network architectures for large
scale acoustic modeling,” in Proc. of Interspeech, 2014,
pp. 338–342.
[4] T. N. Sainath, O. Vinyals, A. Senior, and H. Sak, “Convolutional, long short-term memory, fully connected
deep neural networks,” in Proc. of ICASSP, 2015, pp.
4580–4584.
[5] L. Deng and D. Yu, “Deep learning: methods and applications,” Foundations and Trends in Signal Processing,
vol. 7, no. 3–4, pp. 197–387, 2014.
[6] J. Schalkwyk, D. Beeferman, F. Beaufays, B. Byrne,
C. Chelba, M. Cohen, M. Kamvar, and B. Strope,
““Your Word is my Command”: google search by voice:
A case study,” in Advances in Speech Recognition, pp.
61–90. Springer US, 2010.
[7] X. Lei, A. Senior, A. Gruenstein, and J. Sorensen, “Accurate and compact large vocabulary speech recognition
on mobile devices,” in Proc. of Interspeech, 2013, pp.
662–665.
[8] J. Xue, J. Li, and Y. Gong, “Restructuring of deep neural
network acoustic models with singular value decomposition,” in Proc. of Interspeech, 2013, pp. 2365–2369.
[9] J. Xue, J. Li, D. Yu, M. Seltzer, and Y. Gong, “Singular
value decomposition based low-footprint speaker adaptation and personalization for deep neural network,” in
Proc. of ICASSP, 2014, pp. 6359–6363.
[10] A. Graves, M. Liwicki, S. Fernández, R. Bertolami,
H. Bunke, and J. Schmidhuber, “A novel connectionist system for unconstrained handwriting recognition,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 31, no. 5, pp. 855–868, 2009.
[11] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to
sequence learning with neural networks,” in Proc. of
NIPS, 2014, pp. 3104–3112.
[12] M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and
N. de Freitas, “Predicting parameters in deep learning,”
in Proc. of NIPS, 2013, pp. 2148–2156.
[13] C. Buciluă, R. Caruana, and A. Niculescu-Mizil,
“Model compression,” in Proc. of the 12th ACM
SIGKDD international conference on Knowledge discovery and data mining, 2006, pp. 535–541.
[14] L. J. Ba and R. Caruana, “Do deep nets really need to
be deep?,” in Proc. of NIPS, 2014, pp. 2654–2662.
[15] G. Hinton, O. Vinyals, and J. Dean,
“Distilling
the knowledge in a neural network,” arXiv preprint
arXiv:1503.02531, 2015.
[16] W. Chen, J. T. Wilson, S. Tyree, K. Q. Weinberger, and
Y. Chen, “Compressing neural networks with the hashing trick,” in Proc. of ICML, 2015, pp. 2285–2294.
[17] Y. LeCun, J. S. Denker, and S. A. Solla, “Optimal brain
damage,” in Proc. of NIPS, 1989, pp. 598–605.
[18] F. Grézl and P. Fousek, “Optimizing bottle-neck features for LVCSR,” in Proc. of ICASSP, March 2008, pp.
4729–4732.
[19] T. N. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy,
and B. Ramabhadran, “Low-rank matrix factorization
for deep neural network training with high-dimensional
output targets,” in Proc. of ICASSP, 2013, pp. 6655–
6659.
[20] Y. Wang, J. Li, and Y. Gong, “Small-footprint highperformance deep neural network-based speech recognition using split-VQ,” in Proc. of ICASSP, 2015, pp.
4984–4988.
[21] P. Nakkiran, R. Alvarez, R. Prabhavalkar, and C. Parada,
“Compressing deep neural networks using a rankconstrained topology,” in Proc. of Interspeech, 2015,
pp. 1473–1477.
[22] H. Sak, A. Senior, K. Rao, O. İrsoy, A. Graves, F. Beaufays, and J. Schalkwyk, “Learning acoustic frame labeling for speech recognition with recurrent neural networks,” in Proc. of ICASSP, 2015, pp. 4280–4284.
[23] H. Sak, A. Senior, K. Rao, and F. Beaufays, “Fast and
accurate recurrent neural network acoustic models for
speech recognition,” in Proc. of Interspeech, 2015, pp.
1468–1472.
[24] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification: Labelling
unsegmented sequence data with recurrent neural networks,” in Proc. of ICML, 2006, pp. 369–376.
[25] J. Dean, G. S. Corrado, R. Monga, K. Chen, M. Devin,
Q. V. Le, M. Z. Mao, M. Ranzato, A. Senior, P. Tucker,
K. Yang, and A. Y. Ng, “Large scale distributed deep
networks,” in Proc. of NIPS, 2012, pp. 1223–1231.
[26] B. Kingsbury, “Lattice-based optimization of sequence
classification criteria for neural-network acoustic modeling,” in Proc. of ICASSP, 2009, pp. 3761–3764.
[27] H. Sak, O. Vinyals, G. Heigold, A. Senior, E. McDermott, R. Monga, and M. Mao, “Sequence discriminative distributed training of long short-term memory recurrent neural networks,” in Proc. of Interspeech, 2014,
pp. 1209–1213.
[28] I. McGraw, R. Prabhavalkar, R. Alvarez, M. Gonzalez Arenas, K. Rao, D. Rybach, O. Alsharif, H. Sak,
A. Gruenstein, F. Beaufays, and C. Parada, “Personalized speech recognition on mobile devices,” in Proc. of
ICASSP, 2016.
| 9cs.NE
|
Towards Statistical Reasoning in Description Logics
over Finite Domains (Full Version)
Rafael Peñaloza1 and Nico Potyka2
arXiv:1706.03207v1 [cs.AI] 10 Jun 2017
1
KRDB Research Centre, Free University of Bozen-Bolzano, Italy
[email protected]
2
University of Osnabrück, Germany
[email protected]
Abstract. We present a probabilistic extension of the description logic ALC for
reasoning about statistical knowledge. We consider conditional statements over
proportions of the domain and are interested in the probabilistic-logical consequences of these proportions. After introducing some general reasoning problems
and analyzing their properties, we present first algorithms and complexity results
for reasoning in some fragments of Statistical ALC.
1 Introduction
Probabilistic logics enrich classical logics with probabilities in order to incorporate uncertainty. In [5], probabilistic logics have been classified into three types that differ in
the way how they handle probabilities. Type 1 logics enrich classical interpretations
with probability distributions over the domain and are well suited for reasoning about
statistical probabilities. This includes proportional statements like “2% of the population suffer from a particular disease.” Type 2 logics consider probability distributions
over possible worlds and are better suited for expressing subjective probabilities or degrees of belief. For instance, a medical doctor might say that she is 90% sure about her
diagnosis. Type 3 logics combine type 1 and type 2 logics allow to reason about both
kinds of uncertainty.
One basic desiderata of probabilistic logics is that they generalize a classical logic.
That is, the probabilistic interpretation of formulas with probability 1 should agree with
the classical interpretation. However, given that first-order logic is undecidable, a probabilistic first-order logic that satisfies our basic desiderata will necessarily be undecidable. In order to overcome the problem, we can, for instance, restrict to Herbrand interpretations over a fixed domain [2,9,13] or consider decidable fragments like description
logics [3, 8, 10].
Probabilistic type 2 extensions of description logics have been previously studied
in [11]. In the unpublished appendix of this work, a type 1 extension of ALC is presented along with a proof sketch for E XP T IME-completeness of the corresponding satisfiability problem. This type 1 extension enriches classical interpretations with probability distributions over the domain as suggested in [5]. We consider a similar, but more
restrictive setting here. We are interested in an ALC extension that allows statistical reasoning. However, we do not impose a probability distribution over the domain. Instead,
we are only interested in reasoning about the proportions of a population satisfying
some given properties. For instance, given statistical information about the relative frequency of certain symptoms, diseases and the relative frequency of symptoms given
diseases, one can ask the relative frequency of a disease given a particular combination
of symptoms. Therefore, we consider only classical ALC interpretations with finite domains and are interested in the relative proportions that are true in these interpretations.
Hence, interpretations in our framework can be regarded as a subset of the interpretations in [11], namely those with finite domains and a uniform probability distribution
over the domain. These interpretations are indeed sufficient for our purpose. In particular, by considering strictly less interpretations, we may be able to derive tighter answer
intervals for some queries. Our approach bears some resemblance to the random world
approach from [4]. However, the authors in [4] consider possible worlds with a fixed
domain size N and are interested in the limit of proportions as N goes to infinity. We
are interested in all finite possible worlds that satisfy certain proportions and ask what
statistical statements must be true in all these worlds.
We begin by introducing Statistical ALC in Section 2 together with three relevant
reasoning problems. Namely, the Satisfiability Problem, the l-Entailment problem and
the p-Entailment problem. In Section 3, we will then discuss some logical properties of
Statistical ALC. In Section 4 and 5, we present first computational results for fragments
of Statistical ALC.
2 Statistical ALC
We start by revisiting the classical description logic ALC. Given two disjoint sets NC
of concept names and NR of role names, ALC concepts are built using the grammar
rule C ::= ⊤ | A | ¬C | C ⊓ C | ∃r.C, where A ∈ NC and r ∈ NR . One can
express disjunction, universal quantification and subsumption through the usual logical
equivalences like C1 ⊔ C2 ≡ ¬(¬C1 ⊓ ¬C2 ). For the semantics, we focus on finite
interpretations. An ALC interpretation I = (∆I , ·I ) consist of a non-empty, finite
domain ∆I and an interpretation function ·I that maps concept names A ∈ NC to sets
AI ⊆ ∆I and roles names r ∈ NR to binary relations rI ⊆ ∆I × ∆I . Two ALC
concepts C1 , C2 are equivalent (C1 ≡ C2 ) iff C1I = C2I for all interpretations I.
Here, we consider a probabilistic extension of ALC. Statistical ALC knowledge
bases consist of probabilistic conditionals that are built up over ALC concepts.
Definition 1 (Conditionals, Statistical KB). A probabilistic ALC conditional is an
expression of the form (C | D)[ℓ, u], where C, D are ALC concepts and ℓ, u ∈ Q are
rational numbers such that 0 ≤ ℓ ≤ u ≤ 1. A statistical ALC knowledge base (KB) is
a set K of probabilistic ALC conditionals.
For brevity, we usually call probabilistic ALC conditionals simply conditionals.
Example 2. Let Kflu = {(∃has.fever | ∃has.flu)[0.9, 0.95], (∃has.flu | ⊤)[0.01, 0.03]}.
Kflu states that 90 to 95 percent of patients who have the flu have fever, and that only 1
to 3 percent of patients have the flu.
Intuitively, a conditional (C | D)[ℓ, u] expresses that the relative proportion of elements
of D that also belong to C is between ℓ and u. In order to make this more precise, consider a finite ALC interpretation I, and an ALC concept X. We denote the cardinality
of X I by [X]I , that is, [X]I := |X I |. The interpretation I satisfies (C | D)[ℓ, u],
written as I |= (C | D)[ℓ, u], iff either [D]I = 0 or
[C ⊓ D]I
∈ [ℓ, u].
[D]I
(1)
I satisfies a statistical ALC knowledge base K iff it satisfies all conditionals in K. In this
case, we call I a model of K and write I |= K. We denote the set of all models of K by
Mod(K). As usual, K is consistent if Mod(K) 6= ∅ and inconsistent otherwise. We call
two knowledge bases K1 , K2 equivalent and write K1 ≡ K2 iff Mod(K1 ) = Mod(K2 ).
Example 3. Consider again the KB Kflu from Example 2. Let I be an interpretation
with 1000 individuals. 10 of these have the flu and 9 have both the flu and fever. Then
I ∈ Mod(Kflu ).
In classical ALC, knowledge bases are defined by a set of general concept inclusions (GCIs) C ⊑ D that express that C is a subconcept of D. An interpretation I
satisfies C ⊑ D iff C I ⊆ DI . As shown next, GCIs can be seen as a special kind of
conditionals, and hence statistical ALC KBs are a generalization of classical ALC KBs.
Proposition 4. For all statistical ALC interpretations I, we have I |= C ⊑ D iff
I |= (D | C)[1, 1].
Proof. If I |= C ⊑ D then C I ⊆ DI and C I ∩ DI = C I . If C I = ∅, we have
I
= 1. Hence, I |= (D | C)[1, 1].
[C]I = 0. Otherwise [C⊓D]
[C]I
Conversely, assume I |= (D | C)[1, 1]. If [C]I = 0, then C I = ∅ and I |= C ⊑ D.
I
= 1, that is, [C ⊓ D]I = [C]I . If there was a d ∈ C I \ DI , we had
Otherwise, [C⊓D]
[C]I
[C ⊓ D]I < [C]I , hence, we have C I ⊆ DI and I |= C ⊑ D.
⊓
⊔
Given a statistical ALC knowledge base K, the first problem that we are interested in
is deciding consistency of K. We define the satisfiability problem for statistical ALC
knowledge bases as usual.
Satisfiability Problem: Given a knowledge base K, decide whether Mod(K) 6= ∅.
Example 5. Consider again the knowledge base Kflu from Example 2. The conditional
(∃has.flu | ⊤)[0.01, 0.03] implies that [∃has.flu]I ≥ 0.01 for all models I ∈ Mod(Kflu ).
(∃has.fever | ∃has.flu)[0.9, 0.95] implies [∃has.fever ⊓ ∃has.flu]I ≥ 0.9[∃has.flu]I . Therefore, [∃has.fever]I ≥ [∃has.fever ⊓ ∃has.flu]I ≥ 0.9[∃has.flu]I ≥ 0.009. Hence, adding
the conditional (∃has.fever | ⊤)[0, 0.005]} renders Kflu inconsistent.
If K is consistent, we are interested in deriving (implicit) probabilistic conclusions. We
can think of different reasoning problems in this context. First, we can define an entailment relation analogously to logical entailment. Then, the probabilistic conditional
(C | D)[ℓ, u] is an l-consequence of the KB K iff Mod(K) ⊆ Mod({(C | D)[ℓ, u]}).
In this case, we write K |=l (C | D)[ℓ, u]. In the context of type 2 probabilistic conditionals, this entailment relation has also been called just logical consequence [9].
l-Entailment Problem: Given a knowledge base K and a conditional (C | D)[ℓ, u],
decide whether K |=l (C | D)[ℓ, u].
Example 6. Consider again the KB Kflu from Example 2. As explained in Example 5,
[∃has.fever]I ≥ 0.009 holds for all models I ∈ Mod(K). Therefore, it follows that
Kflu |=l (∃has.fever | ⊤)[0.009, 1]. That is, our statistical information suggests that at
least 9 out of 1, 000 of our patients have fever.
Example 7. Consider a domain with birds (B), penguins (P) and flying animals (F). We
let Kbirds = {(B | ⊤)[0.5, 0.6], (F | B)[0.85, 0.9], (F | P )[0, 0]}. Note that the conditional (F | B)[0.85, 0.9] is actually equivalent to (¬F | B)[0.1, 0.15]. Furthermore,
for all I ∈ Mod(Kbirds ), (F | P )[0, 0] implies [P ⊓ F ]I = 0. Therefore, we have
[P ⊓ B]I = [B ⊓ P ⊓ F ]I + [B ⊓ P ⊓ ¬F ]I ≤ 0 + [B ⊓ ¬F ]I ≤ 0.15[B]I . Hence,
Kbirds |=l (P | B)[0, 0.15]. That is, our statistical information suggests that at most 15
out of 100 birds in our population are penguins.
As usual, the satisfiability problem can be reduced to the l-entailment problem.
Proposition 8. K is inconsistent iff K |=l (⊤ | ⊤)[0, 0].
Proof. If K is inconsistent, then Mod(K) = ∅ and so K |=l (⊤ | ⊤)[0, 0].
I
Conversely, assume K |=l (⊤ | ⊤)[0, 0]. We have [⊤]I > 0 and [⊤⊓⊤]
= 1 for all
[⊤]I
interpretations I. Hence, Mod({(⊤ | ⊤)[0, 0]}) = ∅ and since K |=l (⊤ | ⊤)[0, 0], we
must have Mod(K) = ∅ as well.
⊓
⊔
Often, we do not want to check whether a specific conditional is entailed, but rather
deduce tight probabilistic bounds for a statement. This problem is often referred to
as the probabilistic entailment problem in other probabilistic logics, see [6, 9, 13] for
instance. Consider a query of the form (C | D), where C, D are ALC concepts. We
define the p-Entailment problem similar to the probabilistic entailment problem for
type 2 probabilistic logics.
p-Entailment Problem: Given knowledge base K and a query (C | D), find minimal
and maximal solutions of the optimization problems
inf
I∈Mod(K)
subject to
I
/
sup
I∈Mod(K)
[C ⊓ D]I
[D]I
[D]I > 0
is bounded from below by 0 and from above by
Since the objective function [C⊓D]
[D]I
1, the infimum m and the maximum M are well-defined whenever there is a model
I ∈ Mod(K) such that [D]I > 0. In this case, we say that K p-entails (C | D)[m, M ]
and write K |=p (C | D)[m, M ]. In the context of type 2 probabilistic conditionals,
this entailment relation has also been called tight logical consequence [9]. If [D]I = 0
for all I ∈ Mod(K), the p-Entailment problem is infeasible, that is, there exists no
solution.
Example 9. In Example 7, we found that Kbirds |=l (P | B)[0, 0.15]. This bound is
actually tight. Since 0 is always a lower bound and we showed that 0.15 is an upper
bound, it suffices to give examples of interpretations that take these bounds. For the
lower bound, let I0 be an interpretation with 200 individuals. 100 of these individuals
are birds and 85 are birds that can fly. There are no penguins. Then I0 is a model of
Kbirds with [B]I0 > 0 that satisfies (P | B)[0, 0]. Construct I1 from I0 by letting the
15 non-flying birds be penguins. Then I1 is another model of Kbirds and I1 satisfies
(P | B)[0.15, 0.15]. Hence, we also have Kbirds |=p (P | B)[0, 0.15].
If K |=p (C | D)[m, M ], one might ask whether the values between m and M are
actually taken by some model of K or whether there can be large gaps in between. For
the probabilistic entailment problem for type 2 logics, we can show that the models of
K do indeed yield a dense interval by noting that each convex combination of models is
a model and applying the Intermediate Value Theorem from Real Analysis. However,
in our framework, we do not consider probability distributions over possible worlds, but
the worlds themselves, which are discrete in nature. We therefore cannot apply the same
tools here. However, for each two models that yield different probabilities for a query,
we can find another model that takes the probability in the middle of these probabilities.
Lemma 10 (Bisection Lemma). Let C, D be two arbitrary ALC concepts. If there
I0
I1
exist I0 , I1 ∈ Mod(K) such that r0 = [C⊓D]
< [C⊓D]
= r1 , then there is an
[D]I0
[D]I1
I0.5 ∈ Mod(K) such that
[C⊓D]I0.5
[D]I0.5
=
r0 +r1
2 .
Proof. Given an interpretation I and n ∈ N, we construct the interpretation I (n) as
(n)
follows. We set ∆I = {d1 , . . . , dn | d ∈ ∆I }; that is, we make n different copies of
(n)
the domain. For all A ∈ NC , we set AI = {d1 , . . . , dn | d ∈ AI }, and for all r ∈ NR ,
(n)
we set rI
= {(d1 , e1 ), . . . , (d1 , en ), . . . (dn , e1 ), . . . , (dn , en ) | (d, e) ∈ rI }. By
(n)
induction on the shape of ALC concepts, we can show that [F ]I = n[F ]I for all
concepts F .
Let now (Ci | Di )[ℓi , ui ], i = 1, . . . , |K| be all the conditionals from K. Let ℓ be
the least common multiple of all values from [D]I0 , [D]I1 , [D1 ]I0 , [D1 ]I1 , . . . , [Dn ]I0 ,
[Dn ]I1 that are non-zero, and k, K, k1 , K1 , . . . , kn , Kn be such that k[D]I0 = ℓ,
K[D]I1 = ℓ, . . . , kn [Dn ]I0 = ℓ, Kn [Dn ]I1 = ℓ. Assume w.l.o.g. that I0 and I1 have
different domains (just rename the elements of one domain if necessary). For n, N ∈ N,
(n)
(N )
let In,N be the interpretation that is obtained from I0 and I1 by taking the union
(n)
(N )
of the domains, concept and role interpretations. That is, ∆In,N = ∆I0 ∪ ∆I1 ,
Qn
(n)
(N )
(n)
(N )
I1
. Consider k ′ = k i=1 ki ,
and rIn,N = rI0 ∪ rQ
AIn,N =QAI0 ∪ AI1 Q
n
n
′
′
= K i6=j Ki . Then, for i = 1, . . . , |K|,
k−j
= k i6=j ki , K ′ = K i=1 Ki and K−j
(k′ n)
(K ′ N )
[Ci ⊓ Di ]I0
+ [Ci ⊓ Di ]I1
[Ci ⊓ Di ]Ik′ n,K ′ N
=
(k′ n)
(K ′ N )
Ik′ n,K ′ N
[Di ]
[Di ]I0
+ [Di ]I1
k ′ n[Ci ⊓ Di ]I0 + K ′ N [Ci ⊓ Di ]I1
k ′ n[Di ]I0 + K ′ N [Di ]I1
k ′ n[Ci ⊓ Di ]I0 + K ′ N [Ci ⊓ Di ]I1
=
′ n + K ′ N )l
(k−i
−i
=
=
k′ n
[Ci ⊓ Di ]I0
[Ci ⊓ Di ]I1
K ′N
+
′ n + K ′ N k [D ]I0
′ n + K ′ N K [D ]I1
k−i
k−i
i
i
i
i
−i
−i
=
′
K−i
N
k′ n
[Ci ⊓ Di ]I0
[Ci ⊓ Di ]I1
+
.
′
′
′
′
I
0
k−i n + K−i N [Di ]
k−i n + K−i N [Di ]I1
The last equality shows that
I1
[Ci ⊓Di ]
[Di ]I1
[Ci ⊓Di ]
[Di
I ′
k n,K ′ N
I
] k′ n,K ′ N
is a convex combination of
[Ci ⊓Di ]I0
[Di ]I0
and
. Since, I0 and I1 satisfy the i-th conditional, Ik′ n,K ′ N satisfies the condi-
tional as well. In case that both [Di ]I0 = 0 and [Di ]I1 = 0, we have [Di ]Ik′ n,K ′ N = 0
as well and so the conditional is still satisfied. If only [Di ]I0 = 0, we can see from
the second inequality that
[Ci ⊓Di ]
[Di ]
I ′
k n,K ′ N
=
I ′
k n,K ′ N
k′ n0+K ′ N [Ci ⊓Di ]I1
k′ n0+K ′ N [Di ]I1
=
[Ci ⊓Di ]I1
[Di ]I1
and
I1
the conditional is still satisfied. The case [Di ] = 0 is analogous of course. Hence,
Ik′ n,K ′ N ∈ Mod(K) for all choices of n and N .
Qn
Qn
Let k0 = i=1 ki and K0 = i=1 Ki . Then we can show completely analogously
that
[C⊓D]
I ′
k n,K ′ N
I
[D] k′ n,K ′ N
N = K ′ , we have
=
[Ci ⊓Di ]I0
k0 n
k0 n+K0 N [Di ]I0
[C⊓D]
I ′ ′
k K ,K ′ k′
I
[D] k′ K ′ ,K ′ k′
=
+
[Ci ⊓Di ]I1
K0 N
k0 n+K0 N [Di ]I1
I
1 [Ci ⊓Di ] 0
2 [Di ]I0
+
I
1 [Ci ⊓Di ] 1
2 [Di ]I1
. Letting n = K ′ and
=
r0 +r1
2 .
⊓
⊔
We can now show that for each value between the lower and upper bound given by
p-entailment, we can find a model that gives a probability arbitrarily close to this value.
Proposition 11 (Intermediate Values). Let K |=p (C | D)[m, M ]. Then for every
x ∈ (m, M ) (where (m, M ) denotes the open interval between m and M ) and for all
Ix,ǫ
ǫ > 0, there is a Ix,ǫ ∈ Mod(K) such that | [C⊓D]
− x| < ǫ.
[D]Ix,ǫ
Proof. Since K |=p (C | D)[m, M ], there must exist an I0 ∈ Mod(K) such that
m≤
[C⊓D]I0
[D]I0
[C⊓D]I1
≤ M.
[D]I1
⊥
I0 = I0 , I0⊤ =
≤ x and an I1 ∈ Mod(K) such that x ≤
Consider the following bisection algorithm: we let
I1 . Then start⊥
⊤
ing from i = 1, we let Ii0.5 be the model of K that is obtained from Ii−1
and Ii−1
0.5
as explained in the bisection lemma. If
0.5
[C⊓D]Ii
0.5
[D]Ii
⊥
0.5
[D]Ii
= x, we are done. Otherwise, if
⊥
< x, we let Ii⊥ = Ii−1
and Ii⊤ = I 0.5 . Otherwise, we have
and we let
[C⊓D]Ii
[C⊓D]Ii
Ii⊥
= Ii0.5 and
⊤
[C⊓D]Ii
0.5
[C⊓D]Ii
0.5
[D]Ii
⊤
= Ii−1
. By construction, we maintain the invariant
⊥
Ii⊤
[C⊓D]Ii
we have [C⊓D]
−
≤ M−m
⊤
⊥
2n . Hence, after at
[D]Ii
[D]Ii
≤x≤
and
⊤
i
[D]I
M−m
most i = log
iterations, Ii0.5 is a model of K that proves the claim.
ǫ
⊥
[D]Ii
< x,
Ii⊤
⊓
⊔
3 Logical Properties
We now discuss some logical properties of Statistical ALC. We already noted that Statistical ALC generalizes classical ALC in Proposition 4. Furthermore, p-entailment
yields a tight and dense (Proposition 11) answer interval for all queries whose condition can be satisfied by models of the knowledge base. Let us also note that statistical
ALC is language invariant. That is, increasing the language by adding new concept or
role names does not change the semantics of ALC. This can be seen immediately by
observing that the interpretation of conditionals in (1) depends only on the concept and
role names that appear in the conditional.
Statistical ALC is also representation invariant in the sense that for all concepts
C1 , D1 and C2 , D2 , if C1 ≡ C2 and D1 ≡ D2 then (C1 | D1 )[l, u] ≡ (C2 | D2 )[l, u].
Hence, changing the syntactic representation of conditionals does not change their semantics. In particular, entailment results are independent of such changes.
Both l- and p-entailment satisfy the following independence property: whether or
not K |=l (C | D)[ℓ, u] (K |=p (C | D)[m, M ]) depends only on the conditionals
in K that are connected with the query. This may simplify answering the query by
reducing the size of the KB. In order to make this more precise, we need some additional
definitions. For an arbitrary ALC concept C, Sig(C) denotes the set of all concept and
role names appearing in C. The conditionals (C1 | D1 )[ℓ1 , u1 ] and (C2 | D2 )[ℓ2 , u2 ]
are directly connected (written (C1 | D1 )[ℓ1 , u1 ] ⇋ (C2 | D2 )[ℓ2 , u2 ]) if and only
if (Sig(C1 ) ∪ Sig(D1 )) ∩ (Sig(C2 ) ∪ Sig(D2 )) 6= ∅. That is, two conditionals are
directly connected iff they share concept or role names. Let ⇋∗ denote the transitive
closure of ⇋. We say that (C1 | D1 )[ℓ1 , u1 ] and (C2 | D2 )[ℓ2 , u2 ] are connected iff
(C1 | D1 )[ℓ1 , u1 ] ⇋∗ (C2 | D2 )[ℓ2 , u2 ]. The restriction of K to conditionals connected
to (C | D)[ℓ, u] is the set {κ ∈ K | κ ⇋∗ (C | D)[ℓ, u]}. Using an analogous definition
for queries (qualitative conditionals) (C1 | D1 ) and (C2 | D2 ), we get the following
result.
Proposition 12 (Independence). If K is consistent, we have
1. K |=l (C | D)[ℓ, u] iff {κ ∈ K | κ ⇋∗ (C | D)[ℓ, u]} |=l (C | D)[ℓ, u].
2. K |=p (C | D)[m, M ] iff {κ ∈ K | κ ⇋∗ (C | D)} |=p (C | D)[m, M ].
Proof. For both claims, it suffices to show that for each model I1 of K, there is a
model I2 of {κ ∈ K | κ ⇋∗ (C | D)} ({κ ∈ K | κ ⇋∗ (C | D)[ℓ, u]}) such that
[D]I1 = [D]I2 and [C ⊓ D]I1 = [C ⊓ D]I2 and vice versa.
If I1 is a model of K, let I2 be the restriction of I1 to the concept and role names
in {κ ∈ K | κ ⇋∗ (C | D)}. Then I2 is still a model of {κ ∈ K | κ ⇋∗ (C | D)}. In
particular, [D]I1 = [D]I2 and [C ⊓ D]I1 = [C ⊓ D]I2 .
Conversely, let I2 be a model of {κ ∈ K | κ ⇋∗ (C | D)}. By consistency of K,
there is a model I0 of K. Let I1 be the interpretation defined as the disjoint union of I0
and I2 . Since {κ ∈ K | κ ⇋∗ (C | D)} and K \ {κ ∈ K | κ ⇋∗ (C | D)} do not share
any concept and role names by definition of connectedness, I1 satisfies conditionals in
{κ ∈ K | κ ⇋∗ (C | D)} iff I2 does and conditionals in K \ {κ ∈ K | κ ⇋∗ (C | D)}
iff I0 does. Hence, I1 is a model of K. In particular, it holds that [D]I1 = [D]I2 and
⊓
⊔
[C ⊓ D]I1 = [C ⊓ D]I2 .
Another interesting property of probabilistic logics is continuity. Intuitively, continuity
states that minor changes in the knowledge base do not yield major changes in the
derived probabilities. However, as demonstrated by Courtney and Paris, this condition is
too strong when reasoning with the maximum entropy model of the knowledge base [14,
p. 90]. The same problem arises for the probabilistic entailment problem [16, Example
4]. While these logics considered subjective probabilities, the same problem occurs in
our setting for statistical probabilities as we demonstrate now.
Example 13. Consider the knowledge base
K = {(B | A)[0.4, 0.5], (C | A)[0.5, 0.6], (B | C)[1, 1], (C | B)[1, 1]}.
The interpretation I = ({a, b}, ·I ) with AI = {a, b}, B I = C I = {b} is a model of
K, i.e., K is consistent. In particular, since A is interpreted by the whole domain of I
we know that
K |=p (A | ⊤)[m, 1]
for some m ∈ [0, 1]. As explained in Proposition 4, deterministic conditionals correspond to concept inclusions and so (B | C)[1, 1] and (C | B)[1, 1] imply that
′
I′
′
I′
B I = C I for all models I ′ of K. Therefore, [B⊓A]
= [C⊓A]
. Let K′ denote
[A]I ′
[A]I ′
the knowledge base that is obtained from K by decreasing the upper bound of the first
conditional in K by an arbitrarily small ǫ > 0. That is,
K′ = {(B | A)[0.4, 0.5 − ǫ], (C | A)[0.5, 0.6], (B | C)[1, 1], (C | B)[1, 1]}.
Then the only way to satisfy the first two conditionals in K′ is by interpreting A by the
empty set. Indeed, the interpretation I∅ that interprets all concept names by the empty
set is a model of K′ . So K′ is consistent and
K′ |=p (A | ⊤)[0, 0].
Hence, a minor change in the probabilities in the knowledge base can yield a severe
change in the entailed probabilities. This means that the p-entailment relation that we
consider here is not continuous in this way either.
As an alternative to this strong notion of continuity, Paris proposed to measure the
difference between KBs by the Blaschke distance between their models. Blaschke continuity says that if KBs are close with respect to the Blaschke distance, the entailed
probabilities are close. Blaschke continuity is satisfied by some probabilistic logics under maximum entropy and probabilistic entailment [14, 16]. In [14, 16], probabilistic
interpretations are probability distributions over a finite number of classical interpretations and the distance between two interpretations is the distance between the corresponding probability vectors. We cannot apply this definition here because we interpret
conditionals by means of classical interpretations. It is not at all clear what a reasonable
definition for the distance between two classical interpretations is. We leave the search
for a reasonable topology on the space of classical interpretations for future work.
4 Statistical EL
Proposition 4 and the fact that reasoning in ALC is E XP T IME-complete, show that our
reasoning problems are E XP T IME-hard. However, we did not find any upper bounds on
the complexity of reasoning in ALC so far. We will therefore focus on some fragments
of ALC now.
To begin with, we will focus on the sublogic EL [1] of ALC that does not allow
for negation and universal quantification. Formally, EL concepts are constructed by the
grammar rule C ::= A | ⊤ | C⊓C | ∃r.C, where A ∈ NC and r ∈ NR . A statistical EL
KB is a statistical ALC KB where conditionals are restricted to EL concepts. Notice that,
due to the upper bounds in conditionals, statistical EL KBs are capable of expressing
some weak variants of negations. For instance, a statement (C | ⊤)[ℓ, u] with u < 1
restricts every model I = (∆I , ·I ) to contain at least one element δ ∈ ∆I \ C I . Thus,
contrary to classical EL, statistical EL KBs may be inconsistent.
Example 14. Consider the KB K1 = (∅, C1 ), where
C1 = {(A | ⊤)[0, 0.2], (A | ⊤)[0.3, 1]}.
Since ⊤I = ∆I 6= ∅, every model I = (∆I , ·I ) of K1 must satisfy
[A]I ≤ 0.2[⊤]I < 0.3[⊤]I ≤ [A]I ,
which is clearly a contradiction. Thus, K1 is inconsistent.
More interestingly, though, it is possible to simulate valuations over a finite set of propositional formulas wit the help of conditional statements. Thus, the satisfiability problem
is at least NP-hard even for Statistical EL.
Theorem 15. The satisfiability problem for Statistical EL is NP-hard.
Proof. We provide a reduction from the well-known
coNP-complete problem of decidW
ing validity of a 3DNF formula. Let ϕ = ni=1 κi be a 3DNF formula; that is, each
κi , 1 ≤ i ≤ n is a conjunction of three literals κi = λ1i ∧ λ2i ∧ λ3i . We construct a
statistical EL KB as follows. Let V be the set of all variables appearing in ϕ. For every
x ∈ V, we use two concept names Ax and A¬x . In addition, for every clause κi we
introduce a concept name Bi , and create an additional concept name C.
Consider the KB Kϕ = (Tϕ , Cϕ ), where
Tϕ := {
3
l
j=1
Aλj ⊑ Bi , Bi ⊑ C | 1 ≤ i ≤ n}
i
Cϕ := {(Ax | ⊤)[0.5, 1], (A¬x | ⊤)[0.5, 1], (A¬x | Ax )[0, 0]} ∪ {(C | ⊤)[0, 0.5]}.
Then it holds that ϕ is valid iff Kϕ is inconsistent.
⊓
⊔
On the other hand, consistency can be decided in non-deterministic exponential
time, through a reduction to integer programming. Before describing the reduction in
detail, we introduce a few simplifications.
Recall from Proposition 4 that a conditionals of the form (D | C)[1, 1] is equivalent
to the classical GCI C ⊑ D. Thus, in the following we will often express statistical EL
KBs as pairs K = (T , C), where T is a classical TBox (i.e., a finite set of GCIs), and C
is a set of conditionals. A statistical EL KB K = (T , C) is said to be in normal form if
all the GCIs in T are of the form
A1 ⊓ A2 ⊑ B,
A ⊑ ∃r.B,
∃r.A ⊑ B
and all its conditionals are of the form
(A | B)[ℓ, u]
where A, B ∈ NC ∪ {⊤}, and r ∈ NR . Informally, a KB is in normal form if at
most one constructor is used in any GCI, and all conditionals are atomic (i.e., between
concept names). Every KB can be transformed to an equivalent one (w.r.t. the original
signature) in linear time using the normalization rules from [1], and introducing new
concept names for complex concepts appearing in conditionals. More precisely, we
replace any conditional of the form (C | D)[ℓ, u] with the statement (A | B)[ℓ, u],
where A, B are two fresh concept names, and extend the TBox with the axioms A ≡ C,
and B ≡ D.
The main idea behind our consistency algorithm is to partition the finite domain of
a model into the different types that they define, and use integer programming to verify
that all the logical and conditional constraints are satisfied. Let NC (K) denote the set of
all concept names appearing in the KB K. We call any subset θ ⊆ NC (K) a type for K.
Intuitively, such a type θ represents all the elements of the domain that are interpreted
to belong to all concept names A ∈ θ and no concept name A ∈
/ θ. We denote as Θ(K)
the set of all types of K. To simplify the presentation, in the following we treat ⊤ as a
concept name that belongs to all types.
Given a statistical EL KB K = (T , C) in normal form, we consider an integer
variable xθ for every type θ ∈ Θ(K). These variables will express the number of domain
elements that belong to the corresponding type. In addition, x⊤ will be used to represent
the total size of the domain. We build a system of linear inequalities over these variables
as follows. First, we require that all variables have a value at least 0, and that the sizes
of all types add exactly the size of the domain.
X
xθ = x⊤
(2)
θ∈Θ(K)
0 ≤ xθ
for all θ ∈ Θ(K)
(3)
Then, we ensure that all the conditional statements from the KB are satisfied by adding,
for each statement (A | B)[ℓ, u] ∈ C the constraint
X
X
X
ℓ·
xθ ≤
xθ ≤ u ·
xθ ,
(4)
B∈θ
A,B∈θ
B∈θ
Finally, we must ensure that the types satisfy all the logical constraints introduced by
the TBox. The GCI A1 ⊓ A2 ⊑ B states that every element that belongs to both A1 and
A2 must also belong to B. This means that types containing A1 , A2 but excluding B
should not be populated. We thus introduce the inequality
xθ = 0
if A1 ⊓ A2 ⊑ B ∈ T , A1 , A2 ∈ θ, and B ∈
/θ
(5)
Dealing with existential restrictions requires checking different alternatives, which we
solve by creating different linear programs. The GCI A ⊑ ∃r.B implies that, whenever
there exists an element in A, there must also exist P
at least one elementP
in B. Thus, to
satisfy this axiom, either A should be empty (i.e., A∈θ xθ = 0), or B∈θ xθ ≥ 1.
Hence, for every existential restriction of the form A ⊑ ∃r.B, we define the set
X
X
EA,B := {
xθ = 0,
xθ ≥ 1}
A∈θ
B∈θ
To deal with GCIs of the form ∃r.A ⊑ B, we follow a similar approach, together with
the ideas of the completion algorithm for classical EL. For every pair of existential
restrictions A ⊑ ∃r.B, ∃r.C ⊑ D, we define the set
X
X
xθ ≥ 1}
xθ = 0,
FA,B,C,D := {
B∈θ,C ∈
/θ
A∈θ,D∈θ
/
P
Intuitively, A∈θ,D∈θ
/ xθ ≥ 1 whenever there exists an element that belongs to A but
not to D. If this is the case, and the GCIs A ⊑ ∃r.B, ∃r.C ⊑ D belong to the TBox T ,
then there must exist some element that belongs to B but not to C.
We call the hitting sets of
{EA,B | A ⊑ ∃r.B ∈ T } ∪ {FA,B,C,D | A ⊑ ∃r.B, ∃r.C ⊑ D ∈ T }
choices for T . A program for K is an integer program containing all the inequalities (2)–
(5) and a choice for T . Then we get the following result.
Lemma 16. K is consistent iff there exists a program for K that is satisfiable.
Proof. The “only if” direction is straight-forward since the inequalities are sound w.r.t.
the semantics of statistical KBs. We focus on the “if” direction only.
Given a solution of the integer program, we construct an interpretation I = (∆, ·I )
as follows. We create a domain ∆ with x⊤ elements, and partition it such that for every
type θ ∈ Θ(K), there is a class [[θ]] containing exactly xθ elements. For every nonempty class, select a representative element δθ ∈ [[θ]].
The interpretation function ·I maps every concept name A to the set
[
AI :=
[[θ]].
A∈θ
Given a non-empty class [[θ]] such that A ∈ θ and A ⊑ ∃r.B ∈ T , let τ be a type such
that B ∈ τ , xτ > 0, and for every ∃r.C ⊑ D ∈ T , if D ∈
/ θ, then C ∈
/ τ . Notice
that such a τ must exist because the solution must satisfy at least one restriction in each
θ
FA,B,C,D . We define rA,B
:= θ × {δτ } and set
[
θ
rI :=
rA,B
.
A∈θ,A⊑∃r.B∈T
It remains to be shown that I is a model of K.
S
Notice that for two concept names A, B, it holds that (A ⊓ B)I = A,B∈θ [[θ]] and
P
hence [A ⊓ B]I | = A,B∈θ xθ . Given a conditional statement (A | B)[ℓ, u] ∈ C, since
the solution must satisfy the inequality (4), it holds that
ℓ · [B]I ≤ [A ⊓ B]I ≤ u · [B]I .
For a GCI A1 ⊓ A2 ⊑ B ∈ T , by the inequality (5) it follows that for every type
θ containing both A1 , A2 , but not B, [[θ]] = ∅. Hence AI1 ∩ AI2 ⊆ B I . For every
A ⊑ ∃r.B ∈ T , and every γ ∈ ∆, if γ ∈ AI then by construction there is an element
γ ′ such that (γ, γ ′ ) ∈ rI .
Finally, if (γ, γ ′ ) ∈ rI , then by construction there exists a type θ and an axiom
A ⊑ ∃r.B ∈ T such that γ ∈ [[θ]] and γ ′ = δτ . Then, for every GCI ∃r.C ⊑ D ∈ T ,
γ ′ ∈ C I implies C ∈ τ and hence D ∈ θ which means that γ ∈ DI .
⊓
⊔
Notice that the construction produces exponentially many integer programs, each of
which uses exponentially many variables, measured on the size of the KB. Since satisfiability of integer linear programs is decidable in non-deterministic polynomial time
on the size of the program, we obtain a non-deterministic exponential time upper bound
for deciding consistency of statistical EL KBs.
Theorem 17. Consistency of statistical EL KBs is in NE XP T IME.
5 Reasoning with Open Minded KBs
In order to regain tractability, we now further restrict statistical EL KBs by disallowing upper bounds in the conditional statements. We call such knowledge bases open
minded.
Definition 18 (Open Minded KBs). A statistical EL KB K = (T , C) is open minded
iff all the conditional statements (C | D)[ℓ, u] ∈ C are such that u = 1.
For the scope of this section, we consider only open minded KBs. The first obvious
consequence of restricting to this class of KBs is that negations cannot be simulated. In
fact, every open minded KB is consistent and, as in classical EL, can be satisfied in a
simple universal model.
Theorem 19. Every open minded KB is consistent.
Proof. Consider the interpretation I = ({δ}, ·I ) where the interpretation function maps
every concept name A to AI := {δ} and every role name r to rI := {(δ, δ)}. It is easy
to see that this interpretation is such that C I = {δ} holds for every EL concept C.
Hence, I satisfies all EL GCIs and in addition [C ⊓ D]I = [C]I = 1 which implies that
all conditionals are also satisfied.
⊓
⊔
Recall that, intuitively, conditionals specify that a proportion of the population satisfies
some given properties. One interesting special case of p-entailment is the question how
likely it is to observe an individual that belongs to a given concept.
Table 1. Rules for deciding m-necessity
C1
C2
C3
if {(X, A1 ), (X, A2 )} ⊆ S and A1 ⊓ A2 ⊑ B ∈ T then add (X, B) to S
if (X, A) ∈ S
and A ⊑ ∃r.B ∈ T
then add (A, r, B) to S
if {(X, r, Y ), (Y, A)} ⊆ S and ∃r.A ⊑ B ∈ T
then add (X, B) to S
L1
L2
L3
if
if
if (B, A) ∈ S
(A | B)[ℓ, 1] ∈ C then L(A) ← ℓ · L(B)
A1 ⊓ A2 ⊑ B ∈ T then L(B) ← L(A1 ) + L(A2 ) − 1
then L(A) ← L(B)
Definition 20. Let K be an open minded KB, C a concept, and m ∈ [0, 1]. C is m-necessary in K if K p-entails (C | ⊤)[m, 1]. The problem of m-necessity consists in
deciding whether C is m-necessary in K.
We show that this problem can be solved in polynomial time. As in the previous section, we assume that the KB is in normal form and additionally, that all conditional
statements (A | B)[ℓ, 1] ∈ C are such that ℓ < 1. This latter assumption is made
w.l.o.g. since the conditional statement (A | B)[1, 1] can be equivalently replaced by
the GCI B ⊑ A (see Proposition 4). Moreover, checking m-necessity of a complex
concept C w.r.t. the KB (T , C) is equivalent to deciding m-necessity of a new concept
name A w.r.t. the KB (T ∪ {A ≡ C}, C). Thus, in the following we consider w.l.o.g.
only the problem of deciding m-necessity of a concept name w.r.t. to a KB in normal
form.
Our algorithm extends the completion algorithm for classification of EL TBoxes to
in addition keep track of the lower bounds of necessity for all relevant concept names.
The algorithm keeps as data structure a set S of tuples of the form (A, B) and (A, r, B)
for A, B ∈ NC ∪ {⊤}. These intuitively express that the TBox T entails the subsumptions A ⊑ B and A ⊑ ∃r.B, respectively. Additionally, we keep a function L that
maps every element A ∈ NC ∪ {⊤} to a number L(A) ∈ [0, 1]. Intuitively, L(A) = n
expresses that K p-entails (A | ⊤)[n, 1].
The algorithm initializes the structures S and L as
S := {(A, A), (A, ⊤) | A ∈ NC (K) ∪ {⊤}}
(
0 if A ∈ NC (K)
L(A) :=
1 if A = ⊤.
These structures are then updated using the rules from Table 1. In each case, a rule
is only applied if its execution extends the available knowledge; that is, if either S is
extended to include one more tuple, or a lower bound in L is increased. In the latter
case, only the larger value is kept through the function L.
The first three rules in Table 1 are the standard completion rules for classical EL.
The remaining rules update the lower bounds for the likelihood of all relevant concept
names, taking into account their logical relationship, as explained next.
Rule L1 applies the obvious inference associated to conditional statements: from
all the individuals that belong to B, (A | B)[ℓ, 1] states that at least 100ℓ% belong also
to A. Thus, assuming that L(B) is the lowest proportion of elements in B possible, the
proportion of elements in A must be at least ℓ ·L(B). L3 expresses that if every element
of B must also belong to A, then there must be at least as many elements in A as there
are in B. Finally, L2 deals with the fact that two concepts that are proportionally large
must necessarily overlap. For example, if 60% of all individuals belong to A and 50%
belong to B, then at least 10% must belong to both A and B; otherwise, together they
would cover more than the whole domain.
The algorithm executes all the rules until saturation; that is, until no rule is applicable. Once it is saturated, we can decide m-necessity from the function L as follows:
A is m-necessary iff m ≤ L(A). Before showing the correctness of this algorithm, we
show an important property.
Notice that the likelihood information from L is never transferred through roles.
The reason for this is that an existential restriction ∃r.B only guarantee the existence
of one element belonging to the concept B. Proportionally, the number of elements that
belong to B tends to 0.
Example 21. Consider the KB ({⊤ ⊑ ∃r.A}, ∅). For any n ∈ N, construct the interpretation In := ({0, . . . , n}, ·In ), where AIn = {0} and rIn = {(k, 0) | 0 ≤ k ≤ n}. It
I
is easy to see that In is a model of the KB and [A] n/[⊤]In < 1/n. Thus, the best lower
bound for m-necessity of A is 0, as correctly given by the algorithm.
Theorem 22 (correctness). Let L be the function obtained by the application of the
rules until saturation and A0 ∈ NC . Then A0 is m-necessary iff m ≤ L(A).
Proof (sketch). It is easy to see that all the rules are sound, which proves the “if” direction. For the converse direction, we consider a finite domain ∆ and an interpretation
I
·I of the concept names such that [A] /|∆| = L(A) and the post-conditions of the rules
L1 –L3 are satisfied. Such interpretation can be obtained recursively by considering the
last rule application that updated L(A). Assume w.l.o.g. that the domain is large enough
so that c/|∆| < m − L(A0 ), where c is the number of concept names appearing in K.
It is easy to see that this interpretation satisfies all conditional statements and the GCIs
A1 ⊓ A2 ⊑ B ∈ T . For every concept name A, create a new domain element δA and
extend the interpretation I such that δA ∈ B iff (A, B) ∈ S. Given a role name r, we
define rI := {(γ, δB ) | A ⊑ ∃r.B, γ ∈ AI }. Then, this interpretation satisfies the KB
K, and [A0 ]I /|∆| ≤ L(A0 ) + c/|∆| < m.
⊓
⊔
Thus, the algorithm can correctly decide m-necessity of a given concept name. It remains only to be shown that the process terminates after polynomially many rule applications. To guarantee this, we impose an ordering in the rule applications. First, we
apply all the classical rules C1 –C3 , and only when no such rules are applicable, we
update the function L through the rules L1 –L3 . In this case, the rule that will update
to the largest possible value is applied first. It is known that only polynomially many
classical rules (on the size of T ) can be applied [1]. Deciding which bound rule to apply
next requires polynomial time on the number of concept names in K. Moreover, since
the largest update is applied first, the value of L(A) is changed at most once for every
concept name A. Hence, only linearly many rules are applied. Overall, this means that
the algorithm terminates after polynomially many rule applications, which yields the
following result.
Theorem 23. Deciding m-necessity is in P.
6 Related Work
Over the years, various probabilistic extensions of description logics have been investigated, see, for instance, [3, 7, 8, 10, 12, 15, 17]. The one that is closest to our approach is
the type 1 extension of ALC proposed in the appendix of [11]. Briefly, [11] introduces
probabilistic constraints of the form P (C | D) ≤ p, P (C | D) = p, P (C | D) ≥ p for
ALC concepts C, D. These correspond to the conditionals (C | D)[0, p], (C | D)[p, p],
(C | D)[p, 1], respectively. Conversely, each conditional can be rewritten as such a
probabilistic constraint. However, there is a subtle but fundamental difference in the
semantics. While the definition in [11] allows for probability distributions over arbitrary domains, we do not consider uncertainty over the domain. This comes down to
allowing only finite domains and only the uniform distribution over this domain; that is,
our approach further restricts the class of models of a KB. One fundamental difference
between the two approaches is that Proposition 4 does not hold in [11]: the reason is
that the conditional (C | D)[1, 1] can be satisfied by an interpretation I that contains
an element x ∈ (C ⊓ ¬D)I , where x has probability 0.
This difference is the main reason why the E XP T IME algorithm proposed by Lutz
and Schröder cannot be transferred to our setting. It does not suffice to consider the
satisfiable types independently, but other implicit subsumption relations may depend
on the conditionals only.
Example 24. Consider the statistical EL KB K = (T , C) with
T := {⊤ ⊑ ∃r.A, ∃r.B ⊑ C}
C := {(B | ⊤)[0.5, 1], (A | B)[0.5, 1],
(A | ⊤)[0, 0.25]}
From C it follows that every element of A must also belong to B, and hence every
domain element must be an element of C. However, ¬C defines a satisfiable type (w.r.t.
T ) which will be interpreted as non-empty in the model generated by the approach
in [11].
7 Conclusions
We have introduced Statistical ALC, a new probabilistic extension of the description
logic ALC for statistical reasoning. We analyzed the basic properties of this logic and
introduced some reasoning problems that we are interested in. As a first step towards
effective reasoning in Statistical ALC, we focused on EL, a well-known sublogic of
ALC that, in its classical form, allows for polynomial-time reasoning. We showed that
upper bounds in conditional constraints make the satisfiability problem in statistical
EL NP-hard and gave an NE XP T IME algorithm to decide satisfiability. We showed
that tractability can be regained by disallowing strict upper bounds in the conditional
statements.
We are going to provide more algorithms and a more complete picture of the complexity of reasoning for Statistical ALC and its fragments in future work. A combination
of integer programming and the inclusion-exclusion principle may be fruitful to design
first algorithms for reasoning in full Statistical ALC.
References
1. Baader, F., Brandt, S., Lutz, C.: Pushing the E L envelope. In: Kaelbling, L.P., Saffiotti, A.
(eds.) Proc. of the 19th Int. Joint Conf. on Artificial Intelligence (IJCAI’05). pp. 364–369.
Morgan-Kaufmann (2005)
2. Beierle, C., Kern-Isberner, G., Finthammer, M., Potyka, N.: Extending and completing
probabilistic knowledge and beliefs without bias. KI-Künstliche Intelligenz 29(3), 255–262
(2015)
3. Ceylan, İ.İ., Peñaloza, R.: The bayesian ontology language BE L. J. Autom. Reasoning 58(1),
67–95 (2017)
4. Grove, A.J., Halpern, J.Y., Koller, D.: Random worlds and maximum entropy. In: Logic in
Computer Science, 1992. LICS’92., Proceedings of the Seventh Annual IEEE Symposium
on. pp. 22–33. IEEE (1992)
5. Halpern, J.Y.: An analysis of first-order logics of probability. Artificial intelligence 46(3),
311–350 (1990)
6. Hansen, P., Jaumard, B.: Probabilistic satisfiability. In: Kohlas, J., Moral, S. (eds.) Handbook of Defeasible Reasoning and Uncertainty Management Systems, vol. 5, pp. 321–367.
Springer Netherlands (2000)
7. Klinov, P., Parsia, B.: Pronto: A practical probabilistic description logic reasoner. In: Uncertainty Reasoning for the Semantic Web II, pp. 59–79. Springer (2013)
8. Koller, D., Levy, A., Pfeffer, A.: P-classic: a tractable probablistic description logic.
AAAI/IAAI 1997, 390–397 (1997)
9. Lukasiewicz, T.: Probabilistic logic programming with conditional constraints. ACM Trans.
Comput. Logic 2(3), 289–339 (Jul 2001)
10. Lukasiewicz, T., Straccia, U.: Managing uncertainty and vagueness in description logics for
the semantic web. JWS 6(4), 291–308 (2008)
11. Lutz, C., Schröder, L.: Probabilistic description logics for subjective uncertainty. In: Proc.
KR 2010. AAAI Press (2010)
12. Niepert, M., Noessner, J., Stuckenschmidt, H.: Log-linear description logics. In: IJCAI. pp.
2153–2158 (2011)
13. Nilsson, N.J.: Probabilistic logic. Artificial Intelligence 28, 71–88 (February 1986)
14. Paris, J.B.: The Uncertain Reasoner’s Companion – A Mathematical Perspective. Cambridge
University Press (1994)
15. Peñaloza, R., Potyka, N.: Probabilistic reasoning in the description logic ALCP with the
principle of maximum entropy. In: International Conference on Scalable Uncertainty Management. pp. 246–259. Springer (2016)
16. Potyka, N., Thimm, M.: Probabilistic reasoning with inconsistent beliefs using inconsistency
measures. In: IJCAI. pp. 3156–3163 (2015)
17. Riguzzi, F., Bellodi, E., Lamma, E., Zese, R.: Probabilistic description logics under the distribution semantics. Semantic Web 6(5), 477–501 (2015)
| 2cs.AI
|
Finding the optimal nets for self-folding Kirigami
N. A. M. Araújo,1, 2, ∗ R. A. da Costa,3 S. N. Dorogovtsev,3, 4 and J. F. F. Mendes3
arXiv:1710.00852v1 [cs.DS] 2 Oct 2017
1
Departamento de Fı́sica, Faculdade de Ciências,
Universidade de Lisboa, 1749-016 Lisboa, Portugal
2
Centro de Fı́sica Teórica e Computacional, Universidade de Lisboa, 1749-016 Lisboa, Portugal
3
Department of Physics & I3N, University of Aveiro, 3810-193 Aveiro, Portugal
4
A. F. Ioffe Physico-Technical Institute, 194021 St. Petersburg, Russia
Three-dimensional shells can be synthesized from the spontaneous self-folding of two-dimensional
templates of interconnected panels, called nets. The yield is maximized following sequentially two
design rules: (i) maximum number of vertices with a single-edge cut and (ii) minimum radius of
gyration of the net. Previous methods to identify the optimal net are based on random search
and thus limited to very simple shell structures and not guaranteeing a unique solution. Here, we
show that the optimal net can be found using a deterministic algorithm. We map the connectivity
of the shell into a shell graph, where the nodes and links of the graph represent the vertices and
edges of the shell, respectively. Applying the design rule (i) corresponds then to finding the set of
maximum leaf spanning trees of the shell graph, to which (ii) can be applied straightforwardly. This
method allows not only to designing the self-assembly of much larger shell structures but also to
apply additional design rules, as a complete catalog of the maximum leaf spanning trees is obtained.
PACS numbers: 05.65.+b,64.60.aq
The synthesis of three-dimensional polyhedral shells
at the micron and nano scales is key for encapsulation and drug delivery [1–3]. Inspired by the Japanese
art of Kirigami, where hollowed structures are obtained
from cutting and folding a sheet of paper, lithographic
methods have been developed to form shells from twodimensional nanometer templates of interconnected panels [4–8]. The potential is enormous, for a wide range of
shapes and sizes can be obtained. Ideally, the unfolded
templates (nets) should spontaneously self-fold into the
target structure to reduce production costs and achieve
large-scale parallel production.
Many nets fold into the same structure, but the time
and effectiveness of their self-folding pathways may differ by orders of magnitude. Finding the optimal net that
maximizes yield is not simple; The self-folding process depends on the geometry of the net, its physical properties,
and its interactions with the surrounding medium [9–11].
Experiments by Pandey et al. suggest that maximum
yield is obtained when, from the entire set of nets, one
first picks the nets with the maximum number of vertices
with a single-edge cut and, from them, the nets with the
lowest radius of gyration [12]. A vertex with a single-edge
cut is one whose the number of adjacent faces is the same
in the net and in the polyhedral shell. The implementation of this global search in Ref. [12] implies considering
all possible cuts of the shell. This inefficient procedure
is time consuming, technically demanding, and for most
shapes impossible, since, as we show here, the number
of possible cuts rapidly grows with the number of edges.
For example, with only twelve edges, a cube has 384 possible cuts, while a dodecahedron, with thirty edges, has
more than 5 million possible cuts. Consequently, previous methods to identify the optimal net of a sufficiently
large shell have to be based on random searches in the
configuration space. They consider only a subset (actually a small subset) of possible nets and so they do not
guarantee a unique, globally optimal solution.
We propose a deterministic procedure to identify the
optimal net that only requires generating a subset of all
possible cuts. Let us exemplarily consider the case of a
cubic shell. As shown in Fig. 1(a), the structure of the
shell can be mapped into a shell graph (black nodes and
links in the figure), where nodes represent the vertices
and links represent the edges. A second graph can also
be defined, the face graph, whose nodes are the faces and
links connect pairs of adjacent faces (blue graph). Every
net of the cubic shell corresponds to a spanning tree of its
face graph, i.e., a connected sub-graph that includes all
nodes but the minimum number of links (see Fig. 1(c)).
The nets can be obtained by a set of cuts along the edges
of the shell graph, under the constraint that the set of
nodes in the face graph remains connected. The cut is
defined as the sub-graph of the shell graph that contains
all removed links (cut edges), as represented in red in
Figs. 1(b) and (c). It consists of all nodes and it is a
spanning tree of the shell graph. The main advantage
of the mapping proposed here is that it makes possible
to implement in a systematic and deterministic way the
two design rules. Below, we discuss how to apply such
mapping to 1) identify the cuts that maximize the number of vertices with a single-edge cut and 2) rank them
by increasing radius of gyration.
First criterion. The vertices with a single-edge cut
are the nodes of unitary degree in the cut, known as
the leaves. Since the cut is a spanning tree of the shell
graph, the cuts that maximize the number of vertices
with a single-edge cut are the maximum leaf spanning
trees (MLST) of the shell graph. To identify the full set
of MLSTs, we first identify all minimum connected dom-
2
a)
b)
Archimedean solids with up to 150 shell edges. The obtained dependence is consistent with the predicted linear
dependence (solid line). From the simple relation (1), one
can estimate an upper bound for the number of MLSTs,
NMLST , as a function of E. The exact number of spanning trees, NST , is given by the Kirchhoff’s matrix-tree
theorem, which states that the total number of labeled
spanning trees is given by the product of the eigenvalues
λn of the Laplacian matrix, i.e.,
c)
NST =
V −1
1 Y
λi ,
V i=1
(2)
where V is the number of vertices and λV = 0 is excluded
from the product. Nevertheless, we can also get an upper
bound for NST as a function of E. Accordingly, as we
show in the Supplemental Material [13], an estimate the
upper bound for the ratio NMLST /NST is given by
NMLST /NST ∼ 2−E/2+3/2 ,
FIG. 1. Net of a cubic shell. The cubic shell is mapped into a
shell graph (black), where the nodes and links of the graph are
the shell’s vertices and edges, respectively. A net is obtained
by any sequence of edge cuts (red links in (b) and (c)). The
cut is defined as a sub-graph of the shell graph, consisting of
the set of removed links and all the nodes (in red). A face
graph is defined such that nodes are the faces of the cube and
the links connect pairs of its adjacent faces (blue nodes and
links). The cut and the net are spanning trees of the shell
and face graphs, respectively.
inant sets of the shell graph. These are minimum sets of
shell-graph nodes such that all nodes are either part of
the set or directly connected to it. Once these minimum
sets are identified, obtaining the MLSTs is straightforward, as all the remaining nodes are the leaves (see details in the Supplemental Material [13]). This algorithm
to identify the MLST is simpler than the most efficient
ones (see, e.g. Ref. [14–16]) but it does find the full list
of MLST while the others do not.
Figures 2(a)-(c) show three examples of threedimensional shell structures and one of their nets, corresponding to cuts that are MLSTs. For these examples,
the number of spanning trees and MLSTs increase with
the number of shell edges (see figure caption for values).
However, the fraction of spanning trees that are MLST
decays exponentially with the number of shell edges, E,
as shown in Fig. 3. To understand this result, we first
estimate how the number of leaves, L, in a MLST scales
with E. As shown in detail in the Supplemental Material [13], assuming that the shell is a convex polyhedral
with regular faces and a simplified (approximated) procedure to identify the MLST, we can predict that
L ∼ E/4 + 2 .
(1)
Figure 4 shows L as a function of E for all Platonic and
(3)
what is also consistent over more than 20 orders of magnitude with the observed exponential dependence shown
in Fig. 3. This fast decay with E reinforces the necessity
of a deterministic method, since the chances of obtaining
a MLST from a random search is given by this ratio. For
example, for the dodecahedron, with only twelve faces,
less than 0.04% of its more than 5 million spanning trees
are MLSTs, i.e., to obtain a MLST one would need to
randomly sample 2500 configurations on average. For
the largest shell that we have considered (E = 150), this
number is larger than 1026 . Thus, identifying the optimal net from random methods is practically impossible
for such large shells.
With the strategy proposed here, we can also consider open structures, without one or more faces, as the
ones shown in Figs. 2(d) and (e). These shells might be
relevant for several applications involving, for example,
encapsulation, drug delivery, or key-lock mechanism [1–
3, 11]. The shell graph for such structures is equal to
the corresponding polyhedron; However, every cut includes all edges adjacent to the missing face (connecting
two nodes of that face). So, to identify the optimal net,
we follow the same procedure as before, but under the
constraint that the edges of the missing face of the polyhedron are always in the cut. Note that, in this case, the
cut is no longer a tree, as the edges of the missing face
form a loop, but this is the only possible loop in the cut as
any other will split the net into pieces (see Supplemental
Materials for further details [13]).
Second criterion. The second criterion proposed in
Ref. [12] is to select, among all possible cuts with the
maximum number of vertices with a single-edge cut, the
one corresponding to the net with the lowest radius of
gyration. To apply this criterion, the label of the individual nodes in the MLST is irrelevant and so we need to
identify first the subset of non-isomorphic cuts, i.e., the
3
a)
b)
c)
d)
e)
FIG. 2. Five examples of shells and of one of their nets corresponding to a cut that is a maximum leaf spanning trees:
a) tetrahedron, with four faces and nine edges, it has four maximum leaf spanning trees, but only one non-isomorphic; b)
dodecahedron, with twelve faces and thirty edges, it has 1980 maximum leaf spanning trees, but only 21 non-isomorphic; c)
small rhombicuboctahedron, with 26 faces and 48 edges, it has 1536 maximum leaf spanning trees, but only 32 non-isomorphic;
d) open cubic shell, with five faces and twelve edges, it has only one maximum leaf spanning tree e) small rhombicuboctahedron
with the top nine faces removed and 17 faces, 36 edges and 20 nodes remaining, it has 720 maximum leaf spanning trees, but
only 90 non-isomorphic. The red circles in the nets indicate the vertices with a single-edge cut.
10
10
0
-20
Platonic
Archimedean
other
E/4+2
30
-10
NMLST
NST
10
40
L 20
Platonic
Archimedean
other
- E/2 + 3/2
2
10
-30
10 0
100
50
150
E
FIG. 3. Fraction of spanning trees (NST ) that are maximum
leaf spanning trees (NMLST ) as a function of the number of
shell edges (E). This ratio was calculated for a total of 21
shells, including all Platonic solids and all Archimedean solids
with up to 150 shell edges (see Table S1 in the Supplemental
Material [13]). The solid line corresponds to the estimation
given by Eq. (3).
subset of cuts that one cannot get one from another by
relabeling the nodes (or faces). This is a much smaller
subset. For example, for the cubic shell, 120 MLSTs were
identified, but only four are non-isomorphic. To identify
them, we rely on the concept of adjacency matrix Aij of
a graph with V nodes, defined as V × V matrix, where
Aij is either unity, if nodes i and j are connected, and
zero otherwise. Two graphs are isomorphic if the adjacency matrix of one can be made equal to the other by a
0
0
100
50
150
E
FIG. 4. Number of leaves (L) as a function of the number of
shell edges (E). This number was calculated for a total of 21
shells, including all Platonic solids and all Archimedean solids
with up to 150 shell edges (see Table S1 in the Supplemental
Material [13]). The solid line corresponds to the estimation
given by Eq. (1).
set of line and column swaps. Note that, each swap corresponds to a relabeling of the nodes and thus the same
swap needs to be done for the lines and the columns.
Figure 5 shows the study for the truncated icosahedron, also known as soccer ball or buckyball. With 60
vertices, 32 faces, and 90 edges, this shell has more than
1020 possible cuts. By mapping the shell into a graph,
we can identify the 484800 cuts with the maximum num-
4
a)
1.4
1.3
RG
RGmin 1.2
1.03
RG 1.02
RGmin 1.01
1.1
1
0
b)
1
0
1000
c)
2000
rank
20
rank
3000
40
4000
d)
FIG. 5. For the truncated icosahedron, also known as soccer
ball or buckyball: (a) spectrum of the radii of gyration for all
the 4114 non-isomorphic MLSTs, where MLSTs were ordered
by increasing radius of gyration; the MLST with the (b) optimal (rank 1); (c) intermediate (rank 2057: RG/RGmin ≈
1.23); and (d) largest (rank 4114: RG/RGmin ≈ 1.40) radius
of gyration.
ber of leaves (30 leaves each) and identify the 4114 nonisomorphic corresponding nets. To calculate the radius
of gyration Rg of each net α ⊂ R2 , we first obtain its
centroid, defined as
Z
1
(x̄, ȳ) =
(x, y) dA ,
(4)
A α
where A is the total area of the shell faces. We then calculate the radius of gyration with respect to the centroid
as
s Z
1
[(x − x̄)2 + (y − ȳ)2 ] dA .
(5)
Rg =
A α
Figure 5(a) shows the spectrum of the radii of gyration
for all the 4114 non-isomorphic nets of the truncated
icosahedron, ranked by increasing radius of gyration; Figures 5(b)-(c) show from those nets, the ones with the
lowest (optimal), intermediate, and maximum Rg . The
radius of gyration rapidly increases with the position in
the rank. The less optimal net (Fig. 5(d)) has a radius
of gyration that is 40% higher than the one of the optimal net (Fig. 5(b)). Previous methods to identify the
optimal net are based on random search. Using such
methods, even if the obtained net corresponds to one
with the maximum number of vertices with a single-edge
cut, the probability that its radius of gyration differs by
less that 3% from the optimal one is below 1.5% (see inset in Fig. 5(a)). Since the timescale and yield of the
self-folding depends strongly on the radius of gyration,
the self-folding efficiency of an approximated solution obtained with previous methods based on a random search
is likely far from optimal.
Conclusions. We proposed a method to identify the
optimal net that spontaneously self-folds into a closed or
open polyhedral shell structures. The method consists
of mapping the shell structure into a shell graph and sequentially apply two design rules: (i) to identify the cuts
that maximize the number of vertices with a single-edge
cut and, (ii) among those, the one with the minimum radius of gyration. Adapting concepts and methods from
Graph Theory, we show that the optimal solution can be
obtained in a deterministic and systematic manner. Previous methods are based on random search and thus not
providing a unique solution. As we showed, the fraction
of cuts that are identified by rule (i) decays exponentially
with the number of edges in the polyhedron, reinforcing
the necessity of a deterministic method. Also, with the
method proposed here, since a complete list of possible
cuts respecting rule (i) is obtained, other design rules
(alternative to rule (ii)) can be implemented straightforwardly.
It is conjectured that the nets for all convex shells are
non-overlapping, a necessary condition for obtaining it
from a two-dimensional template. In all examples studied here, all optimal shells are non-overlapping but nets
for concave shells are more likely to overlap. In fact, this
conjecture is not necessary, for once the optimal net is
identified, self-overlap can be tested. If the selected net
overlaps, one should proceed through the rank of increasing radius of gyration and pick the first net that does not
self-overlap. If all nets for the maximum leaf spanning
trees obtained with rule (1) overlap, one should proceed
iteratively considering spanning trees with less and less
leaves until we found the first that does not overlap.
Identifying the maximum leaf spanning tree is a NP
problem and so the numerical complexity will still grow
rapidly with the number of shell vertices. If the number
of vertices is too large for the straightforward implementation of our deterministic algorithm then, in the spirit
of our approach, approximated algorithms can be used
that identify spanning trees with a number of leaves that
is close to the maximum [17, 18].
We suggest that our deterministic algorithm and
its variations could be used when searching for optimal design and production of even more complex selfassembling systems.
We acknowledge financial support from the Portuguese Foundation for Science and Technology (FCT)
under Contract no. UID/FIS/00618/2013 and grant
5
SFRH/BPD/123077/2016.
∗
[email protected]
[1] R. Fernandes and D. H. Gracias, “Self-folding polymeric
containers for encapsulation and delivery of drugs,” Adv.
Drug Deliv. Rev. 64, 1579 – 1589 (2012).
[2] J. Shim, C. Perdigou, E. R. Chen, K. Bertoldi, and P. M.
Reis, “Buckling-induced encapsulation of structured elastic shells under pressure,” Proc. Natl. Acad. Sci. 109,
5978–5983 (2012).
[3] M. Filippousi, T. Altantzis, G. Stefanou, M. Betsiou,
D. N. Bikiaris, M. Angelakeris, E. Pavlidou, D. Zamboulis, and G. Van Tendeloo, “Polyhedral iron oxide
core-shell nanoparticles in a biodegradable polymeric matrix: preparation, characterization and application in
magnetic particle hyperthermia and drug delivery,” RSC
Adv. 3, 24367–24377 (2013).
[4] D. M. Sussman, Y. Cho, T. Castle, X. Gong, E. Jung,
S. Yang, and R. D. Kamien, “Algorithmic lattice
kirigami: A route to pluripotent materials,” Proc. Natl.
Acad. Sci. 112, 7449–7453 (2015).
[5] Y. Zhang, Z. Yan, K. Nan, D. Xiao, Y. Liu, H. Luan,
H. Fu, X. Wang, Q. Yang, J. Wang, W. Ren, H. Si, F. Liu,
L. Yang, H. Li, J. Wang, X. Guo, H. Luo, L. Wang,
Y. Huang, and J. A. Rogers, “A mechanically driven
form of kirigami as a route to 3d mesostructures in micro/nanomembranes,” Proc. Natl. Acad. Sci. 112, 11757–
11764 (2015).
[6] A. Lamoureux, K. Lee, M. Shlian, S. R. Forrest, and
M. Shtein, “Dynamic kirigami structures for integrated
solar tracking,” Nat. Commun. 6, 8092 (2015).
[7] G. P. Collins, “Science and culture: Kirigami and technology cut a fine figure, together,” Proc. Natl. Acad. Sci.
113, 240–241 (2016).
[8] W. M. Jacobs and D. Frenkel, “Self-assembly of structures with addressable complexity,” J. Am. Chem. Soc.
138, 2457–2467 (2016).
[9] G. M. Whitesides and B. Grzybowski, “Self-assembly at
all scales,” Science 295, 2418–2421 (2002).
[10] T. G. Leong, P. A. Lester, T. L. Koh, E. K. Call, and
D. H. Gracias, “Surface tension-driven self-folding polyhedra,” Langmuir 23, 8747–8751 (2007).
[11] A. Azam, K. E. Laflin, M. Jamal, R. Fernandes, and
D. H. Gracias, “Self-folding micropatterned polymeric
containers,” Biomed. Microdevices 13, 51–58 (2011).
[12] S. Pandey, M. Ewing, A. Kunas, N. Nguyen, D. H. Gracias, and G. Menon, “Algorithmic design of self-folding
polyhedra,” Proc. Natl. Acad. Sci. 108, 19885–19890
(2011).
[13] See Supplemental Material at.
[14] H. Fernau, J. Kneis, D. Kratsch, A. Langer, M. Liedloff,
D. Raible, and P. Rossmanith, “An exact algorithm for
the maximum leaf spanning tree problem,” Theor. Comput. Sci. 412, 6290–6302 (2011).
[15] T. Fujie, “An exact algorithm for the maximum leaf spanning tree problem,” Comput. Oper. Res. 30, 1931–1944
(2003).
[16] A. Lucena, N. Maculan, and L. Simonetti, “Reformulations and solution algorithms for the maximum leaf
spanning tree problem,” Comput. Manag. Sci. 7, 289–
311 (2010).
[17] H.-I. Lu and R. Ravi, “Approximating maximum leaf
spanning trees in almost linear time,” J. Algorithms 29,
132–141 (1998).
[18] R. Solis-Oba, P. Bonsma,
and S. Lowski, “A 2approximation algorithm for finding a spanning tree with
maximum number of leaves,” Algorithmica 77, 374–388
(2017).
S1
Finding the optimal nets for self-folding Kirigami – Supplemental Material
N. A. M. Araújo, R. A. da Costa, S. N. Dorogovtsev, and J. F. F. Mendes
S1. ONE-TO-ONE CORRESPONDENCE
BETWEEN NETS AND SPANNING TREES
To each net of a polyhedron corresponds a cut along
the edges of a spanning tree of the shell graph:
i) To unfold a polyhedral shell into a 2D net, the cut
must reach every vertex of the shell graph, and must be
connected;
ii) For the polyhedron faces to remain connected as a
single component, the cut cannot contain any loops.
The only subgraphs that span to every vertex, are connected, and contain no loops are spanning trees. Therefore, maximizing the number of single-edge cut vertices
of a net is equivalent to maximizing the number of leaves
of a spanning tree.
S2.
THE MAXIMUM LEAF SPANNING TREE
The Maximum Leaf Spanning Tree (MLST) problem
has been extensively studied in the scope of graph theory and computer science [S1–S7]. It consist of finding a spanning tree with the largest possible number of
leaves in a given undirected unweighted graph. Finding
a MLST, or just determining the number of leaves of the
MLSTs of a generic graph is a well known NP-complete
problem [S1].
Here, we describe a simple (exact) algorithm to find
the full set of labeled MLSTs of an arbitrary (undirected
and unweighted) graph. Notice that while the algorithms
for the MLST problem typically find a single optimal tree
[S3–S5], our algorithm provides all possible labeled MLSTs. The number of spanning trees and of MLSTs grow
both exponentially with the graph size and the computation time of an algorithm that lists all MLSTs grows
at least as quickly as the number of MLSTs.
A dominating set of a graph is a subset of vertices, such
that all vertices of the graph either belong to the set, or
are connected to at least one vertex in the set. If, in
addition, those vertices (together with the edges among
them) form a connected subgraph, this set is called a
connected dominating set. Clearly, the non-leaf vertices
of any spanning tree form a connected dominating set.
Therefore, finding the maximum number of leaves in a
spanning tree is equivalent to determining the minimum
size of a connected dominating set. In the rest of this
section, a subtree of the graph whose vertices form a
connected dominating set will be called dominating subtree. (Note that the vertices of a subtree are connected
by definition.)
A.
Algorithm
In order to list all MLSTs, we use a search algorithm
for finding the full set of dominating subtrees with exactly nS vertices (i.e., subtrees with nS vertices that are
a connected dominating set of the graph). We start by
checking if there are any dominating subtrees of nS = 1
vertices, which only exist when a single vertex is connected to all the other vertices in the original graph. If
there is no such tree, we iteratively increase nS by 1 and
search again. The search stops when the set of dominating subtrees with nS vertices is non-empty. These minimum dominating subtrees are the interiors of the MLSTs,
that is, the MLSTs without the leaf vertices and their respective edges.
To finalize the construction of the MLSTs, we attach
the remaining vertices to the obtained minimum dominating subtree. These vertices are the leaves of the
MLST. If every leaf vertex only has one edge of the original graph connecting it to the dominating subtree, then
there is only one possible MLST with that particular interior subtree. However, some leaf vertices may have
multiple edges of the original graph linking them to the
dominating subtree, and so there are multiple MLSTs
with the same interior subtree. In these cases, we have to
chose one of the possibilities for each leaf vertex. Since
those choices are independent from each other, the total number of MLSTs that share that particular interior
subtree equals the product of the numbers of possibilities
of each leaf vertex (i.e., the number of different ways of
connecting the leaf vertices to a particular dominating
subtree).
The algorithm recursively grows subtrees with nS vertices and nS − 1 edges of the graph starting from a single
root vertex and enumerates all MLSTs where the root
is a non-leaf vertex. There is a set of vertices, say R,
that the algorithm uses as roots (one at each time), and
consists of a specific arbitrary vertex and all of its neighbors. For any dominating subtree, all the vertices of the
original graph must either be in the subtree or have at
least one neighbor in the subtree. To find all MLSTs, it
is sufficient to consider the roots in R. In the present
work, we use the vertex with the smallest degree and its
neighbors as the set of roots R.
Given a size of the subtrees nS , the algorithm performs
a separate search for each root vertex in R. So, we need
an additional constraint to avoid multiple counts of the
MLSTs that have more than one non-leaf vertices in R.
We introduce a set of vertices, Vexcl , that are explicitly
forbidden from joining the dominating subtree. In the
first search, rooted in the first vertex of R, say r1 , the set
S2
Vexcl is empty and the algorithm enumerates all MLSTs
where r1 is a non-leaf vertex. Since the vertex r1 must
be a leaf in all the other MLSTs still not found, we add
r1 to Vexcl . In the second search, rooted in r2 , the vertex
r1 is never included in dominating subtree, insuring that
the algorithm only returns the MLSTs where r1 is a leaf
and r2 is a non-leaf vertex. Then, r2 is added to Vexcl ,
the third search is made, and so on.
Given a graph with V vertices labeled i = 1, ..., V , let
us denote the edge that connects the vertices i and j by
eij . In the following Algorithms 1-3 the set of vertices
that are connected to i in the original graph is denoted
Ai , and the set A = {Ai : i = 1, .., V } provides complete
information about the graph.
Algorithm 1 Enumeration of all MLSTs. This procedure initializes the necessary variables, and calls the function recursive of Algorithm 2. Each time it is called,
recursive enumerates all MLSTs with L − nS leaves,
where the supplied root, r, in a non-leaf vertex and the
vertices in Vexcl are leaves.
procedure list all MLSTs(A)
Input:
List of neighbors, Ai , for every vertex i (A = {Ai }).
v = arbitrary vertex . For example, a vertex with the
minimum degree
R = append(v, Av )
M LST s = ∅
nS = 1
while M LST s = ∅ do
Vexcl = ∅
for all r ∈ R do
EA = {eri : i ∈ Ar }
M LST s0 =recursive(A,{r},∅,EA ,Vexcl ,nS )
M LST s = append(M LST s, M LST s0 )
Vexcl = append(Vexcl , r)
end for
nS = nS + 1
end while
end procedure
The procedure list all MLSTs, shown in Algorithm 1, initializes the set of roots R as described above,
and the current size of the searched subtrees, nS , is initialized to 1. The set M LST s will store the collection of
found MLSTs, and starts as an empty set. The search of
MLSTs with larger nS , and a smaller number of leaves
L = V − nS , will only proceed if the set M LST s remains
empty. The list of vertices Vexcl stores the roots of R that
were already used for each particular value of nS , and is
initialized as an empty set every time that nS is incremented. The search itself, is performed by the recursive
function recursive, Algorithm 2. For each considered
nS , the function recursive is called by the procedure
list all MLSTs once for each root in R.
The function recursive of Algorithm 2 starts with a
single vertex (root) and recursively grows subtrees T up
to a predetermined size (nS vertices), while keeping track
of the elements already in the tree and at its border. Let
VT and ET be the lists of vertices and edges, respectively,
currently in T , and EA be the list of edges that are not in
T and are connected to at least one vertex in T (EA is the
exterior boarder of T , which the algorithm is currently
exploring). Furthermore, Vexcl is a specific set of vertices
that are forbidden to participate in the subtree T (these
vertices are the roots of the previous searches for the same
nS ). When T reaches the target size (nS vertices) it stops
increasing. If the vertices VT form a dominating set, then
the algorithm enumerates the possible ways of joining
each vertex outside of VT to one and only one vertex in
VT by an edge. Each different way of making these last
connections represents a different (labeled) spanning tree
whose leaves are the vertices outside of VT .
In the first stage, while the number of vertices in the
growing subtree |VT | < nS , the function recursive considers all possibilities for the next edge addition to the
subtree T from the set of adjacent edges EA . For each
of those possibilities, recursive is called again with the
updated VT 0 , ET 0 , and EA 0 . Note that, when an edge
in EA connects two vertices already in VT it cannot be
added to the tree (because it would close a loop in T ).
In order to keep the vertices in the set Vexcl outside of T ,
the edges that lead to those vertices are not added to EA
in the update. In this way, recursive finds all configurations of subtrees with nS vertices that include the root
and exclude all the vertices in Vexcl . In the second stage,
when finally |VT | = nS , recursive checks if the vertices
in VT form a dominating set: if so, it finishes the construction of the spanning trees by connecting the leaves
in every possible way, otherwise it returns an empty set.
In the search algorithm presented here, we consider
only connected sets of vertices (subtrees), and check if
they are dominating. Furthermore, we only consider subtrees that include either a specific arbitrary vertex or at
least one of its neighbors (if a subtree is dominating, every vertex of the graph fulfills this requirement). The
combination of these two strategies drastically reduces
the configurational space and computation time.
S3
Algorithm 2 Recursive search function. This function
generates all possible subtrees T with nS vertices that
include the root vertex and exclude all the vertices in
Vexcl . If T is a dominating subtree, then the function
lists all possible spanning trees that can be obtained by
joining the remaining vertices of the graph to T .
function recursive(A, VT , ET , EA , Vexcl , nS )
Input:
List of neighbors, Ai , for every vertex i (A = {Ai }).
Lists of vertices, VT , and edges, ET , currently in T .
List of edges, EA , currently adjacent to T .
List of vertices, Vexcl , excluded from T .
Final number of vertices, nS , in T .
Output:
List of spanning trees, ST s, with at least V − nS leaves
(each spanning tree is returned as the list of its edges).
ST s = ∅
if |VT | < nS then
. Add one edge to T .
for all ejk ∈ EA do
EA = EA \{ejk }
if k ∈
/ VT then
i=k
else if j ∈
/ VT then
i=j
else
continue
end if
ET 0 = append(ET , ejk )
VT 0 = append(VT , i)
EA 0 = EA
for all l ∈ Ai do
if eil ∈
/ ET and l ∈
/ Vexcl then
EA 0 = append (EA 0 , eil )
end if
end for
ST s0 = recursive(A, VT 0, ET 0, EA 0, Vexcl , nS )
ST s = append(ST s, ST s0 )
end for
return ST s
else if VT is a dominating set then . Expand T into
corresponding spanning trees.
ST s = {ET }
for all vertices i ∈
/ VT do
ST s0 = ∅
for all vertices j ∈ Ai ∩ VT do
for all trees U ∈ ST s do
ST s0 =append (ST s0 , append(U, eij ))
end for
end for
ST s = ST s0
end for
return ST s
else
return ∅
end if
end function
The order by which the edges of EA are picked in the
outermost for-cycle of the search of recursive is not
specified in Algorithm 2 because any order will produce
the same result. In our implementation of the function
recursive, we picked the edges in EA using a first-infirst-out method and effectively performed a breadth-first
search. If, for instance, the edges of EA are picked in a
last-in-first-out fashion, then the algorithm will perform
depth-first search instead.
The algorithm described in this section for listing all
MLSTs is defined for labeled graphs, and finds the set
of labeled MLSTs. If the original graph has automorphisms, i.e., a relabeling that results in the same labeled
graph, then the set of labeled MLSTs may have multiple
‘copies’ of the same unlabeled MLST with different labelings (isomorphic MLSTs). In this work, we consider the
nets of polyhedra with high regularity, such as Platonic
and Archimedean solids where all vertices are equivalent.
The polyhedral graphs of polyhedra where all vertices are
equivalent, or with sets of equivalent vertices, have automorphisms. However, each geometrically distinct net
is entirely determined by a cut of the polyhedral graph
along the edges of an unlabeled spanning tree. So, to determine the number of distinct nets, we need to disregard
the labels of MLSTs, and search for isomorphisms in the
set of MLSTs found by the algorithm. To check if any
two labeled MLSTs are isomorphic, we simply check if
there is an automorphic relabeling of the original graph
that maps one labeled MLST into the other, see Section
for details.
Typically the algorithms for the MLST problem are
designed to find the number of leaves in a MLST, and
not the full set of MLSTs [S2–S5]. Some of these algorithms include multiple stages of optimization, which
include heuristics, the use of approximated algorithms to
make initial guesses, and other sophisticated approaches.
Due to the complexity of this problem, and to the best
of our knowledge, the fastest existing algorithms can find
one MLST only in graphs with up to roughly 200 vertices
[S5].
In this work, we find not just one MLST, but we list
also the full set of optimal nets (unlabeled MLSTs) for
polyhedral graphs with up to 60 vertices. The number of
different optimal nets strongly depends on the details of
the polyhedral graph. For instance, the Truncated Icosahedron and the Truncated Dodecahedron each have 60
vertices, 32 edges, and 90 faces, however their numbers
of optimal nets are 4 114 and 3 719 677 167, respectively.
Table S1 summarizes the exact results obtained in this
work with the algorithm presented here, namely the number of leaves in a MLST and the number of optimal nets
of each of the 21 polyhedron considered. Table S1 also
shows a figure of the optimal net which minimizes the
radius of gyration for each polyhedra.
Finally, it should be mentioned that there are several
approximated algorithms that find spanning trees with a
high number of leaves, close to the maximum possible. If
a polyhedral graph is too large to solve by exact methods,
these approximated algorithms can find a near-optimal
S4
V F
E
L
Nopt nets
Tetrahedron
4
6
4
3
1
Octahedron
6
8
12
4
2
Cube
8
6
12
4
4
12 20 30
8
21
20 12 30 10
21
Octogonal
Piramid
9
16
8
1
Octogonal
Dipiramid
10 16 24
7
3
12 8
18
6
4
12 14 24
7
34
Truncated Cube 24 14 36 10
399
solution in linear or almost linear time [S6, S7].
S3.
SHELLS WITH HOLES
We consider now the problem of finding optimal nets
for shells that contain holes, i.e., shells consisting of all
the faces of a polyhedron except for one. The graph of
the vertices and edges of the shell is the same as the polyhedral graph of the complete polyhedron. The difference
is that all the edges adjacent to the missing face will be
in every cut, effectively detaching that face from the rest
of the net, as intended. (An edge adjacent to a face is an
edge between two vertices of that face.) For this reason,
a vertex adjacent to the missing face cannot be a singleedge cut vertex in the net, since it always has two edges
included in the cut.
The subgraph of the cut edges in the presence of a
hole is not a pure tree, because it contains a single loop
formed by the edges adjacent to the hole. For the shell
to unfold into a 2D net, the cut subgraph must reach
all vertices (spanning) and be connected. Also, for the
shell’s faces to remain connected in a single component,
the cut subgraph cannot have any other loop apart from
one surrounding the hole. Then, the cut subgraph consists of the loop adjacent to the hole, and some loopless
branches connected to it.
We use the following Algorithm 3 to maximize the
number of single-edge cut vertices in nets of shells that
contain holes. Algorithm 3 is a simple adaptation of Algorithm 1 that calls the same recursive function of Algorithm 2, gradually increasing the allowed number of
non-leaf vertices nS . The recursive function remains unchanged in this procedure. In this version of the algorithm, in addition to the lists of adjacencies, A = {Ai },
we supply the list of vertices adjacent to the hole, Vh .
Instead of single root vertex, the search starts from a
subgraph already containing all the vertices and edges
adjacent to the hole (these vertices and edges must be
present in all the cuts, optimal or not). Then, in the first
call of the function recursive for each nS , VT = Vh and
ET is the set edges adjacent to the hole. The set EA is
initialized with the edges connected to the vertices of Vh
that are not in ET , as in Algorithm 1. For the sake of
clarity, while in Algorithm 1 we denote the set of edges of
the optimal cuts by M LST s, in Algorithm 3 we denote
it by Cuts because in the presence of holes the cuts contain a loop surrounding the hole, and they are no longer
trees. When Algorithm 3 calls the function recursive,
it passes an empty set as the fifth input argument. This
set is the list of vertices that are forced to be leaves,
called Vexcl in recursive, which we do not use in this
algorithm.
Icosahedron
Dodecahedron
Truncated
Tetrahedron
Cuboctahedron
9
Snub Cube
24 38 60 16
600
Rhombicuboctahedron
24 26 48 15
32
Truncated
Octahedron
24 14 36 12
56
Icosidodecahedron 30 32 60 16
308 928
Truncated
Cuboctahedron
48 26 72 24
244
Truncated
Icosahedron
(soccer ball)
60 32 90 30
4 114
Truncated
Dodecahedron
Rhombicosidodecahedron
Snub
Dodecahedron
Triakis
Icosahedron
Pentakis
Dodecahedron
Min RG
–
60 32 90 22 3 719 677 167
–
60 62 120 37
77 952
–
60 92 150 39
13 436 928
–
32 60 90 26
664 128
–
32 60 90 22
845 280
–
TABLE S1. Optimal nets. The number of leaves, L, in the
MLSTs, and the number of distinct optimal nets (unlabeled
MLSTs), Nopt nets , were obtained for each polyhedron with
our algorithm. The numbers of vertices V , faces F , and edges
E are also shown, as well as the optimal net with the smallest
radius of gyration (only for cases with Nopt nets < 10 000).
The red circles indicate the single-edge cut vertices.
S5
Algorithm 3 Enumeration of all optimal cuts of shells
with a hole. This procedure initializes the necessary variables and calls the recursive function of Algorithm 2. The
set Cuts stores the spanning subgraphs that include all
edges adjacent to the hole and maximize the number of
leaves.
procedure list cuts with hole(A, Vh )
Input:
Lists of neighbors, Ai , of each vertex i (A = {Ai }).
Vertices of the hole Vh .
Cuts = ∅
nS = |Vh |
while Cuts = ∅ do
ET = eij : i, j ∈ Vh
EA = eij : j ∈ Ai and i ∈ Vh and j ∈
/ Vh
Cuts0 = recursive(A, Vh , ET , EA , ∅, nS )
Cuts = append(Cuts, Cuts0 )
nS = nS + 1
end while
end procedure
S4.
A.
ESTIMATIONS
Maximum number of leaves
The aim of this section is to estimate the number of
leaves L in a MLST in terms of the simplest possible
polyhedral parameters, preferably as a function of the
number of vertices V , or of edges E. The exact number
L depends on the details of the graph, and its determination requires to actually solve the maximum leaf spanning
tree problem. However, we can obtain a simple estimate
for L by considering a local optimization algorithm for
finding approximated maximum leaf spanning trees.
This algorithm iteratively grows a tree by progressively
attaching to it vertices of the graph, until the tree becomes spanning (i.e. reaches all vertices in the polyhedral
graph), in the following way:
(i) To seed the iterative process, connect the highest
degree vertex to all its neighbors.
(ii) If the current tree is a spanning tree the algorithm
reaches its end. Otherwise, from the vertices already in
the tree, select the one with the highest number of neighbors still not in the tree, connect it to those neighbors,
and repeat step (ii).
In this algorithm, except for the highest degree vertex
of step (i), every vertex added to the tree starts as a
leaf attached to a non-leaf vertex. The neighbors of nonleaf vertices of the intermediate (non-spanning) trees are
guaranteed to be in the tree as well. Furthermore, the
number of leaves in the final spanning tree is L = V −
niter − 1, where niter is the total number of iterations,
and the −1 is due to the initial step (i). The number
niter depends on how many vertices are added to the tree
in each iteration.
As we show in the following, the number of vertices
added to the tree at each iteration turns out to be close to
2, regardless of the details of the polyhedral (shell) graph.
If the polyhedron has no triangular faces, a vertex of
degree k selected in step (ii) contributes at most with k−1
new leaves, because each vertex in the tree has at least
one neighbor also in the tree. However, if the faces are
triangular, any two vertices connected to each other share
two common neighbors. In this case, the selected vertex
has at least 3 neighbors already in the tree, namely, the
parent non-leaf vertex plus two vertices that are common
neighbors with the parent vertex, therefore it contributes
at most with k − 3 new leaves.
In convex polyhedra with regular faces (equilateral triangles, squares, regular pentagons, etc), the sum of the
internal angles attached to a vertex must be smaller than
2π. This strongly constrains the types of faces that can
be attached to a vertex of degree k, and in particular
the number of triangles. Notice that k cannot be smaller
than 3, and that k = 6 is not feasible even with just triangles (in that case the sum of angles is equal to 2π).
Then, let us consider the number of new vertices added
to the tree when the vertex selected in step (ii) has degree k = 3, 4, and 5. A vertex with k = 5 is mainly
surrounded by triangles (at least 4 out of 5 faces must
be triangles), so, when selected in step (ii), most of the
times it can contribute with k − 3 = 2 new leaves to the
tree. A vertex with k = 4 must have between 1 and 4
triangles among its faces, and it can contribute with 1 to
3 new leaves depending on the particular configuration
of the faces. Finally, a vertex with k = 3 can have any
number of triangles between 0 and 3. On the one hand, if
it has 3 triangles, this vertex is never selected in step (ii)
because it has 0 = k − 3 neighbors still not in the tree.
On the other hand, if the vertex has no triangular faces
attached, it can contribute with k − 1 = 2 new leaves.
Due to these geometrical constraints, the average number of new vertices added to the tree at each iteration is
essentially independent of the size and local details of
the polyhedral graph, and is close to 2. Furthermore,
we expect these arguments to qualitatively hold also for
irregular convex polyhedra. Even when the internal angles of a face are different from each other, their sum is
the same, and so, on average, the internal angles are the
same as for a regular face.
Assuming that each iteration adds approximately 2
new vertices on average to the growing tree, and that
the tree becomes spanning after niter iterations, we can
write:
k0 + 1 + 2niter ∼ V,
(S1)
where k0 is the highest degree in the polyhedral graph.
For simplicity, let us use k0 = 4. Replacing niter = V −
S6
L − 1 in Eq. (S1) we get
L ∼ (V + 3)/2.
(S2)
This formula fits well with the general trend observed for
L vs. V , as shown in the inset of Fig. S1.
40
Platonic
Archimedean
other
E/4+2
V/2+3/2
30
40
L 20
L 20
10
0
0
20
40
60
V
0
0
50
100
150
E
and of spanning trees, NST , in terms of the number of
edges, E, of a labeled polyhedral graph. These numbers
grow very quickly with the size of the graph and are not
uniquely determined by E – they depend on the details
of the graph. To obtain a simple estimate, we calculate
upper bounds for NST and NMLST , and use their ratio as
an estimator.
Each spanning tree has V − 1 edges out of a total of
E edges. Therefore, the upper bound of the number of
spanning trees NST is the number of possible combinationsof V − 1 edges, given by the binomial coefficient
E
V −1 . Similarly, each maximum leaf spanning tree has
L leaves of a total of V vertices, and the upper bound for
NMLST is the number of combinations of L vertices, VL .
To obtain an estimation for NMLST /NST , we take the
ratio of the upper bounds and replace the approximations
for L and V of Eqs. (S2) and (S3), respectively:
FIG. S1. Number of leaf vertices in a maximum leaf spanning
tree, L, vs.the number of edges E of the polyhedra in Table S1.
Different symbols represent different sets of polyhedra. The
solid line is the estimation provided by Eq (S4). The inset
shows L vs. the number of vertices V , the dashed line is the
estimation provided by Eq (S2).
Interestingly, the main panel of Fig. S1 shows that the
dispersion of the points is significantly smaller if we plot
L vs E. Let us recall Euler’s polyhedron formula V +
F = E + 2, where V , F , and E are the numbers of
vertices, faces, and edges of the polyhedron, respectively.
For a given E there are different polyhedra with different
combinations of V and F such that V + F = E + 2. On
the one hand, if all faces are triangles we have F = 2E/3
and in this case V = E/3 + 2. On the other hand, if
most faces of the polyhedron have many edges (which
implies wide internal angles) most vertices are attached
to only three faces, since the sum of the angles must be
smaller than 2π for convex polyhedra. Then, in this case
V ≈ 2E/3. Taking this into account, and because we
consider a variety of polyhedra with different types of
faces, we use the middle point,
V ∼ E/2 + 1,
(S3)
as an estimation of the number of vertices V of a polyhedron with E edges. We replace E/2+1 for V in Eq. (S2),
and finally obtain
L ∼ E/4 + 2.
where
we used Stirling’s approximation n!
∼
√
2πn(n/e)n . Figure S2 clearly shows that the ratio NMLST /NST has an exponential-like decay with the
number of edges of the polyhedron. The simple form of
Eq. (S5) fits well with this decay.
10
10
0
-10
NMLST
NST
10
-20
Platonic
Archimedean
other
- E/2 + 3/2
2
-30
10 0
100
50
150
E
FIG. S2. Ratio between the total number of spanning trees,
NST , and the number of maximum leaf spanning trees, NMLST
vs. the number of edges E of the polyhedra in Table S1.
Different symbols represent different sets of polyhedra. The
ratio NMLST /NST decays exponentially with the polyhedron
size, as predicted by Eq. (S5).
(S4)
This formula, plotted in the main panel of Fig. S1 as a
solid line, shows a remarkable agreement with our results.
B.
NMLST /NST ∼
E/2 + 1
E
/
∼ 2−E/2+3/2 (S5)
E/4 + 2
E/2
Ratio NMLST /NST
We now estimate the ratio NMLST /NST , between the
the numbers of maximum leaf spanning trees, NMLST ,
S5.
NON-ISOMORPHIC CUTS
To determine the optimal nets of a polyhedron, which
have no labels, we need to find the set of distinct unlabeled MLSTs. However, the algorithm of Section distinguishes each node by its label, and does not consider
symmetries that may exist in the polyhedral graph (automorphisms). An automorphism of a labeled graph is
S7
a relabeling that results in the same graph. If the polyhedral graph contains automorphisms, then the set of
MLSTs may contain isomorphisms, i.e., multiple ‘copies’
(differently labeled) of the same unlabeled MLST [S8].
These isomorphic cuts correspond to nets that are indistinguishable from each other and have the same radius of
gyration, which is our second criterion of optimization.
Therefore, we need only one member of each set of isomorphic MLSTs in the list of optimal cuts. To determine
if two MLSTs are isomorphic we employ an adjacency
matrix approach. (Note that this approach is valid for
any isomorphic subgraphs, not just MLSTs.)
The adjacency matrix, A, is a convenient representation of labeled graphs. Each element Aij , is 1 if the
vertices with labels i and j are connected by an edge,
otherwise Aij is 0. To switch the labels of any pair of
vertices, say i and j, we simply switch the rows i and j
and the columns i and j in matrix A. Any permutation
of V labels can be mapped into any other by a series of
at most V − 1 switches between pairs of vertices (V is
the number of vertices). We find the complete set of automorphisms of the polyhedral graph by comparing the
relabeled adjacency matrix A0 with the original matrix
A: when A0 = A the relabeling is automorphic.
Similarly to the polyhedral graph, we can represent a
MLST by an adjacency matrix B where each element
Bij is 1 if vertices i and j are connected by an edge in
the MLST and 0 otherwise. Two MLSTs, or cuts, are
isomorphic if, and only if, there is an automorphism of
the polyhedral graph that maps one MLST into the other.
That is, if two cuts, with adjacency matrices B1 and B2 ,
unfold into the same net, then there are automorphic
relabelings of A that map B1 into B2 and vice-versa.
We apply each of the previously obtained automorphisms
of A to one of the matrices, say B2 , and compare the
relabeled matrix B20 with B1 . If one of those relabeling
gives B20 = B1 , then the two cuts are isomorphic, and
we may discard one of them. We systematically compare
each of the MLSTs remaining in the list with all the
others to eliminate any isomorphisms, and obtain the full
set of Nopt net distinct optimal nets of the polyhedron.
S6.
NUMBER OF LABELED SPANNING TREES
Kirchhoff’s matrix-tree theorem allows us to calculate
the exact number of spanning trees of a labeled graph,
NST , in terms of the spectrum of the Laplacian matrix.
The Laplacian matrix of a graph is defined as L = D −A,
where D is the degree matrix (i.e., the diagonal matrix
with each entry dii equal to the degree of vertex i), and
A is the adjacency matrix. For a graph of V vertices the
matrix L has V eigenvalues λi ≥ 0, the smallest of which
is λV = 0. The matrix-tree theorem states that the total
number of spanning trees, NST , is given by the product
of eigenvalues of its Laplacian matrix
NST =
V −1
1 Y
λi ,
V i=1
(S6)
where λV is excluded from the product, and λi > 0 for
connected graphs and i < V [S8, S9]. The values of NST
used to plot the points in Fig. S2 (Fig. 3 of the main text)
were obtained with Eq. (S6) (while the values of NMLST
were obtained by the algorithm described in Section ).
Tetrahedron
Cube
Octahedron
Icosahedron
Dodecahedron
5-cell
8-cell
16-cell
24-cell
120-cell
600-cell
V
E
Nnets
4
6
2
8
12
11
6
12
11
12 30
43 380
20 30
43 380
5
10
3
16 32
261
8
24
110 912
24 96 1.79×1016
600 1200 2.76×10119
∗
120 720 1.20×10307
NST /Naut
∆
2/3
0.66
8
0.27
8
0.27
43 200
0.0041
43 200
0.0041
25/24
0.65
216
0.17
110 592
0.0029
1.79×1016 1.9×10−11
2.76×10119 3.9×10−61
1.20×10307
0
TABLE S2. Comparison of the exact number of nets Nnets
[S8] and the estimation NST /Naut for all regular convex polytopes in 3D (top 5 rows) and 4D (bottom 6 rows). The
right-hand side column of the table is the relative difference
∆ = (Nnets − NST /Naut ) /Nnets . The numbers of vertices,
and edges are denoted V and E, respectively. We include the
4D polytopes in this table to show that ∆ approaches 0 very
quickly as the size (V and E) of the polytope increases. Note
that everywhere else in this paper we consider only 3D shells
and their 2D nets.
∗
The number of nets of the 600-cell shown in Ref. [S8]
is wrong, due to a mistake in the calculation of the
graph’s spectrum. Correcting the mistake gives exactly
2182 ·3102 ·520 ·736 ·1148 ·2348 ·2930 nets for the 600-cell. In this
particular case, the lower-bound and the actual number of
nets coincide exactly.
S7.
NUMBER OF NON-ISOMORPHIC CUTS
The number of non-isomorphic cuts (i.e. nets), Nnets ,
of a polyhedron with no automorphisms is equal to
NST . When the polyhedral graph has automorphisms,
the number of distinct nets is actually smaller than NST ,
see Section . In that case, the exact number of nets,
Nnets , can be obtained using the approach of Ref. [S8],
which involves a detailed, case by case, analysis for each
polyhedron. In Ref. [S8] the Nnets are obtained only for
the five platonic solids. Nevertheless, we can estimate
Nnets for other polyhedron with automorphisms (e.g.,
Archimedean solids) with high precision by taking the ratio of NST to the number of automorphisms of the graph,
Naut . The ratio NST /Naut is in fact a lower-bound of
Nnets ; This ratio assumes that each unlabeled spanning
tree contributes with Naut differently labeled ‘copies’ to
S8
the set of labeled spanning trees, however, some spanning
trees have a smaller number of isomorphic ‘copies’, due
to the existence of symmetries in their branches. Since
Naut can be found with linear time algorithms [S10], the
calculation of NST /Naut is straightforward.
Table S2 clearly shows that this lower-bound actually
approaches Nnets very quickly for large graphs. This happens because the fraction of spanning trees that have
some symmetry in their structure quickly approaches
0 as the size of the graph increases. Figure S3(a)
demonstrates that the number of distinct nets, Nnets ≈
NST /Naut , grows exponentially with the size of the polyhedral graph.
NST / Naut
10
10
10
30
Platonic
Archimedean
other
20
10
a
0
10 0
100
50
150
E
Nopt nets Naut / NST
10
0
b
10
10
duced in this Supplemental Material as Fig. S2. If we
were to generate nets with a method that avoids the
sampling of isomorphic cuts completely, thus reducing
the search space, the probability that a randomly sampled net has the maximum number of single-edge cut vertices, Nopt nets /Nnets , would still be essentially the same
as NMLST /NST .
Figure S3(b) shows the probability that a net sampled at random from the set of non-isomorphic nets has
the maximum number of single-edge cut vertices, i.e.,
Nopt nets /Nnets ≈ Nopt nets Naut /NST . We observe no significant differences between Figs. S2 and S3(b), which
means that searches over unlabeled and labeled configurations have similar performances when only a small
fraction of the configurations is sampled.
-10
-20
Platonic
Archimedean
other
- E/2 + 3/2
2
-30
10 0
100
50
150
E
FIG. S3. Precise estimations of (a) the number of distinct
nets Nnets ≈ NST /Naut , and of (b) the fraction of optimal nets
Nopt nets /Nnets ≈ Nopt nets Naut /NST vs. the number of edges
E of the 3D polyhedra in Table S1. The number Nnets grows
exponentially with E, and shows a remarkably low level of dispersion. The solid line in the plot of panel (b) is Eq. (S5). The
fraction of unlabeled optimal cuts Nopt nets /Nnets is essentially
equal to the fraction of labeled optimal cuts, NMLST /NST , see
Fig. S2.
The probability that a randomly sampled labeled spanning tree is a MLST is equal to the ratio NMLST /NST ,
which is shown is Fig. 3 of the main text, and repro-
∗
[email protected]
[S1] M. R. Garey and D. S. Johnson, Computers and intractability, Vol. 29 (W. H Freeman and Company, New
York, 2002).
[S2] F. Rosamond, “Max leaf spanning tree,” in Encyclopedia of Algorithms (Springer, 2008) pp. 1–99.
[S3] H. Fernau, J. Kneis, D. Kratsch, A. Langer, M. Liedloff,
D. Raible, and P. Rossmanith, “An exact algorithm
for the maximum leaf spanning tree problem,” Theor.
Comput. Sci. 412, 6290–6302 (2011).
[S4] T. Fujie, “An exact algorithm for the maximum leaf
spanning tree problem,” Comput. Oper. Res. 30, 1931–
1944 (2003).
[S5] A. Lucena, N. Maculan, and L. Simonetti, “Reformulations and solution algorithms for the maximum leaf
spanning tree problem,” Comput. Manag. Sci. 7, 289–
311 (2010).
[S6] R. Solis-Oba, P. Bonsma, and S. Lowski, “A 2approximation algorithm for finding a spanning tree
with maximum number of leaves,” Algorithmica 77,
374–388 (2017).
[S7] H.-I. Lu and R. Ravi, “Approximating maximum leaf
spanning trees in almost linear time,” J. Algorithms 29,
132–141 (1998).
[S8] F. Buekenhout and M. Parker, “The number of nets of
the regular convex polytopes in dimension ≤ 4,” Discrete Math. 186, 69–94 (1998).
[S9] D. M. Cvetković, M. Doob, and H. Sachs, Spectra of
Graphs: Theory and Applications, 3rd Ed. (Wiley, New
York, 1998).
[S10] J. E. Hopcroft and J.-K. Wong, “Linear time algorithm
for isomorphism of planar graphs (preliminary report),”
in Proc. 6th Annual ACM Symp. Theory of Computing
(ACM, 1974) pp. 172–184.
| 8cs.DS
|
arXiv:1403.1324v1 [math.AC] 6 Mar 2014
Classification of the linearly reductive finite
subgroup schemes of SL2
Mitsuyasu Hashimoto
Department of Mathematics, Okayama University
Okayama 700–8530, JAPAN
E-mail: [email protected]
Dedicated to Professor Ngo Viet Trung
on the occasion of his sixtieth birthday
Abstract
We classify the linearly reductive finite subgroup schemes G of
SL2 = SL(V ) over an algebraically closed field k of positive characteristic, up to conjugation. As a corollary, we prove that such G is in oneto-one correspondence with an isomorphism class of two-dimensional
F -rational Gorenstein complete local rings with the coefficient field k
by the correspondence G 7→ ((Sym V )G )b.
1.
Introduction
The classification of the finite subgroups of SL2 (C) is well-known ([Dor, Section 26], [LW, (6.2)], see Theorem 3.2), and such a group corresponds to a
Dynkin diagram of type A, D, or E. A two-dimensional singularity is Gorenstein and rational if and only if it is a quotient singularity by a finite subgroup
of SL2 (C), and such singularities (also called Kleinian singularities) are classified via these subgroups, see [Dur]. Indeed, a two-dimensional singularity
2010 Mathematics Subject Classification. Primary 14L15; Secondary 13A50. Key
Words and Phrases. group scheme, Kleinian singularity, invariant theory,
1
is Gorenstein and rational if and only if it is a quotient singularity by a finite
subgroup of SL2 .
It is known that the F -rationality is the characteristic p version of the
rational singularity. More precisely, a finite-type algebra over a field of characteristic zero has rational singularities if and only if its modulo p reduction is
F -rational for almost all prime numbers p [Sm], [Har]. The two-dimensional
complete local F -rational Gorenstein rings over an algebraically closed field k
of characteristic p > 0 is classified using Dynkin diagrams A, D, and E, based
on Artin’s classification of rational double points [Art], see [WY], [HL]. Then
we might well ask whether such a ring is obtained as an invariant subring
k[[x, y]]G with G a finite subgroup of SL2 = SL(V ), where V = kx ⊕ ky.
Before considering this question, we have to consider several things.
First, any finite subgroup of SL2 (C) is small in the sense that it does not
have a pseudo-reflection, where an element g of GL(V ) is called a pseudoreflection if rank(g − 1V ) = 1. This is important in studying the ring of
invariants. If G is a small finite subgroup of GL(V ) (V = C2 ), then G can be
recovered from R̂ = Ŝ G , where Ŝ is the completion of S = Sym V , in the sense
that the fundamental group of Spec R̂ \ {0} is G, where 0 is the unique closed
point. Moreover, the category of maximal Cohen–Macaulay modules of R̂ is
canonically equivalent to the category of Ŝ-finite Ŝ-free (G, Ŝ)-modules [Yos,
(10.9)]. However, this is not the case for SL2 (k) with char(k) > 0. Indeed,
a finite subgroup of SL2 (k) may have a transvection, where g ∈ GL(V ) is
called a transvection if it is a pseudo-reflection and g − 1V is nilpotent. Even
if G is a non-trivial subgroup of SL2 , Ŝ G may be a formal power series ring
again, see [KS, Proposition 4.6].
Next, even if G is a finite subgroup of SL(V ), the ring of invariants
R = (Sym V )G may not be F -regular. Indeed, Singh [Sin] proved that if G is
the alternating group An acting canonically on V = k n , then R = (Sym V )G
is strongly F -regular if and only if p = char(k) does not divide the order
(n!)/2 of G = An . More generally, Yasuda [Yas] proved that if G is a small
subgroup of GL(V ), then the ring of invariants (Sym V )G is strongly F regular if and only if p = char(k) does not divide the order of G.
So we want to classify the subgroups G ⊂ SL2 with the order of G is not
divisible by p = char(k). It is easy to see that such G must be small. The
classification is known (see Theorem 3.2), and the result is the same as that
over C, except that small p which divides the order |G| of G is not allowed.
More precisely, for the type (An ), p must not divide n + 1, for (Dn ), p must
not divide 4n − 8, and we must have p ≥ 5, p ≥ 5, p ≥ 7 for type (E6 ), (E7 ),
2
and (E8 ), respectively. However, the restriction on p for the classification of
two-dimensional F -rational Gorenstein complete local rings is different [HL],
and it is p arbitrary for (An ), p ≥ 3 for (Dn ), and p ≥ 5, p ≥ 5, p ≥ 7 for
type (E6 ), (E7 ), and (E8 ), respectively.
The purpose of this paper is to show the gap occuring on the type (An )
and (Dn ) comes from the non-reduced group schemes, as shown in Theorem 3.8. As a corollary, we show that all the two-dimensional F -rational
Gorenstein complete local rings with the algebraically closed coefficient field
appear as the ring of invariants under the action of a linearly reductive finite
subgroup scheme of SL2 , see Corollary 3.10. This is already pointed out by
Artin [Art] for the type (An ), and is trivial for (E6 ), (E7 ), and (E8 ) because
of the order of the group and the restriction on p. What is new in this paper
is the case (Dn ). At this moment, the author does not know how to recover the group scheme G from R = S G . So although the classification of R,
the two-dimensional F -rational Gorenstein singularities are well-known, the
classification of G seems to be nontrivial for the author. As a result, we can
recover G from R in the sense that the correspondence from G to R̂ = Ŝ G is
one-to-one.
The key to the proof is Sweedler’s theorem (Theorem 2.8) which states
that a connected linearly reductive group scheme over a field of positive
characteristic is abelian.
The author thanks Professor Kei-ichi Watanabe for valuable advice.
2.
Preliminaries
(2.1) Let k be a field. For a k-scheme X, we denote the ring H 0 (X, OX )
by k[X].
We say that an affine algebraic k-group scheme G is linearly reductive if
any G-module is semisimple.
Lemma 2.2. Let
1→N →G→H→1
be an exact sequence of affine algebraic k-group schemes. Then G is linearly
reductive if and only if H and N are linearly reductive.
Proof. We prove the ‘if’ part. If M is a G-module, then the LyndonHochschild-Serre spectral sequence [Jan, (I.6.6)]
E2p,q = H p (H, H q (N, M)) ⇒ H p+q (G, M)
3
degenerates, and E2p,q = 0 for (p, q) 6= (0, 0) by assumption. Thus H n (G, M) =
0 for n > 0, as required.
We prove the ‘only if’ part. First, given a short exact sequence of Hmodules, it is also a short exact sequence of G-modules by restriction. By
assumption, it G-splits, and hence it H-splits. Thus any short exact sequence
of H-modules H-splits, and H is linearly reductive. Next, we prove that N
is linearly reductive. Let M be a finite dimensional N-module. Then there
is a spectral sequence
p+q
(N, M),
E2p,q = H p (G, Rq indG
N (M)) ⇒ H
see [Jan, (I.4.5)]. As G/N ∼
= H is affine, Rq indG
N (M) = 0 (q > 0) by [Jan,
(I.5.13)]. As G is linearly reductive by assumption, E2p,q = 0 for (p, q) 6= (0, 0).
Thus H n (N, M) = 0 for n > 0, and thus N is linearly reductive.
(2.3) Let C = (C, ∆, ε) be a k-coalgebra. An element c ∈ C is said to
be group-like if c 6= 0 and ∆(c) = c ⊗ c [Swe]. If so, ε(c) = 1. The set of
group-like elements of C is denoted by X (C). Note that X (C) is linearly
independent.
Let H be a k-Hopf algebra. Then for h ∈ X (H), S(h) = h−1 , where
S is the antipode. Note that X (H) is a subgroup of the unit group H × .
We denote GL1 = Spec k[t, t−1 ] with t group-like by Gm , and its subgroup
scheme Spec k[t]/(tr − 1) by µr for r ≥ 0. Note that µr represents the group
of the rth roots of unity, but it is not a reduced scheme if char(k) = p divides
r.
(2.4) In the rest of this paper, let k be algebraically closed. For an affine
algebraic group scheme G over k, let X (G) denote the group of characters
(one-dimensional representations) of G. Note that X (G) is canonically identified with X (k[G]), see [Wat, (2.1)].
Lemma 2.5. Let G be an affine algebraic k-group scheme. Then the following are equivalent.
1 G is abelian (that is, the product is commutative) and linearly reductive.
2 G is linearly reductive, and any simple G-module is one-dimensional.
3 G is diagonalizable. That is, a closed subgroup scheme of a torus Gnm .
4 The coordinate ring k[G] is group-like as a coalgebra. That is, k[G] is
the group ring kΓ, where Γ = X (G).
4
5 G is a finite direct product of Gm and µr with r ≥ 2.
Proof. 1⇒2 Follows easily from [Swe, (8.0.1)].
2⇒3 Take a finite dimensional faithful G-module V (this is possible
[Wat, (3.4)]). Take a basis v1 , . . . , vn of V such that each kvi is a onedimensional G-submodule of V . Then the embedding G → GL(V ) factors
through GL(kv1 ) × · · · × GL(kvn ) ∼
= Gnm .
n
3⇒4 Let G ⊂ Gm = T . Then k[T ] is a Laurent polynomial ring
λ1
±1
λn
k[t1 , . . . , t±1
n ]. As each Laurent monomial t1 · · · tn is group-like, k[T ] is
generated by its group-like elements. This property is obviously inherited by
its quotient Hopf algebra k[G], and we are done.
4⇒5 Apply the fundamental theorem of abelian groups on Γ.
5⇒4⇒2⇒1 is easy.
(2.6) The category of finitely generated abelian groups and the category
of diagonalizable k-group schemes are contravariantly equivalent with the
equivalences Γ 7→ Spec(kΓ) and G 7→ X (G). For a diagonalizable k-group
scheme G, a G-module is identified with an X (G)-graded k-vector space. A
G-algebra is nothing but a X (G)-graded k-algebra.
(2.7) For a diagonalizable group scheme G = Spec kΓ, the closed subgroup
schemes H of G is in one-to-one correspondence with the quotient groups M
of Γ with the correspondence H 7→ X (H) and M 7→ Spec kM. In particular,
the only closed subgroup schemes of Gm is µr with r ≥ 0, since the only
quotient groups of Z are Z/rZ.
The following is due to Sweedler [Swe2].
Theorem 2.8. Let G be a connected linearly reductive affine algebraic kgroup scheme over an algebraically closed field of positive characteristic p.
Then G is an abelian group (and hence is diagonalizable). So G is, up to
isomorphisms, of the form
Grm × µpe1 × · · · × µpes
for some r ≥ 0, s ≥ 0, and e1 ≥ · · · es ≥ 1.
(2.9) Let G be an affine algebraic k-group scheme. Note that Spec k, Gred
and Gred × Gred are all reduced. Hence the unit map e : Spec k → G, the
inverse ι : Gred → G, and the product µ : Gred × Gred → G all factor through
Gred ֒→ G, and so Gred is a closed subgroup scheme of G. Thus Gred is
k-smooth.
5
(2.10) We denote the identity component (the connected component containing the identity element) of G by G◦ . As Gred ֒→ G is a homeomorphism
and Gred is k-smooth, each connected component of G is irreducible, and
is isomorphic to G◦ . As Spec k, G◦ , and G◦ × G◦ are all irreducible, it is
easy to see that the unit map, the inverse, the product from them all factor
through G◦ ֒→ G, and hence G◦ is a closed open subgroup of G. If C is any
irreducible component of G, then the image of the map C × G◦ → G given
by (g, n) 7→ gng −1 is contained in G◦ . Thus G◦ is a normal subgroup scheme
of G. That is, the map G × G◦ → G given by (g, n) 7→ gng −1 factors through
G◦ .
(2.11) As the inclusion G◦ · Gred ֒→ G is a surjective open immersion, we
have that G◦ · Gred = G. As G◦ is an open subscheme of G, G◦ ∩ Gred = G◦red .
So if G is finite, then G is a semidirect product G = G◦ ⋊ Gred .
3.
The classification
(3.1) Throughout this section, let k be an algebraically closed field of characteristic p > 0.
The purpose of this section is to classify the linearly reductive finite subgroup schemes of SL2 over k, up to conjugation. Our starting point is the
reduced case, which is well-known. Unfortunately, the author does not know
the proof of the theorem below exactly as stated, but the proof in [Dor,
Section 26] also works for the case of positive characteristic. See also [LW,
Chapter 6, Section 2].
Theorem 3.2. Let k be an algebraically closed field of characteristic p > 0,
and G a finite nontrivial subgroup of SL2 . Assume that the order |G| of G
is not divisible by p. Then G is conjugate to one of the following, where ζr
denotes a primitive rth root of unity.
(An ) (n ≥ 1) The cyclic group generated by
ζn+1
0
.
−1
0 ζn+1
(Dn ) (n ≥ 4) The binary dihedral group generated by (A2n−5 ) and
0 ζ4
.
ζ4 0
6
(E6 ) The binary tetrahedral group generated by (D4 ) and
1 ζ87 ζ87
√
.
5
2 ζ8 ζ8
(E7 ) The binary octahedral group generated by (E6 ) and (A7 ).
(E8 ) The binary icosahedral group generated by (A9 ),
1
0 1
ζ5 + ζ5−1
1
, and
.
−1 0
1
−(ζ5 + ζ5−1 )
ζ52 − ζ53
Conversely, if g = n + 1 (resp. 4n − 8, 24, 48, and 120) is not zero in k, then
(An ) (resp. (Dn ), (E6 ), (E7 ), and (E8 )) above is defined, and is a linearly
reductive finite subgroup of SL2 of order g.
(3.3) Let G be a linearly reductive finite subgroup scheme of SL2 = SL(V ).
As the sequence
1 → G◦ → G → Gred → 1
is exact, both G◦ and Gred are linearly reductive by Lemma 2.2.
(3.4) First, consider the case that G is abelian. Then the vector representation V is the direct sum of two one-dimensional G-modules, say V1 and V2 ,
and hence we may assume that G is diagonalized. As G ⊂ SL2 , V2 ∼
= V1∗ .
Thus G → GL(V1 ) = Gm is also a closed immersion, and G ∼
= µm is
a 0
| a ∈ µm .
0 a−1
(3.5) So assume that G is not abelian. If G◦ is trivial, then G = Gred , and
the classification for this case is done in Theorem 3.2. So assume further that
G◦ is non-trivial.
G◦ is diagonalized as above, since G◦ is linearly reductive and connected
(and hence is also abelian by Theorem 2.8). We have G◦ ∼
= µr with r = pe
for some e ≥ 0.
(3.6) We consider the case that G◦ is contained in the group of scalar matrices. In this case, r = 2 (so p = 2), as G ⊂ SL2 . Then by Maschke’s
theorem, the order of Gred is odd. According to the classification in Theorem 3.2, Gred must be of type (An ) and is cyclic. This shows that G is
abelian, and this is a contradiction.
7
(3.7) So G◦ is not contained in the group of scalar matrices. Note that if
a, b, c, d ∈ k with ad − bc = 1 and
a b
ζr 0
λ 0
a b
=
0 µ
c d
c d
0 ζr−1
for some λ, µ ∈ A = k[T, T −1 ]/(T r − 1), where ζr is the image of T in A, then
(1) λ = ζr , µ = ζr−1 and b = c = 0, or (2) λ = ζr−1 , µ = ζr and a = d = 0. This
is because ζr 6= ζr−1. Then it is easy to see that the centralizer C = ZG (G◦ ) is
contained in the subgroup of diagonal matrices in SL2 . As we assume that G
is not abelian, C 6= NG (G◦ ) = G. Clearly, Cred has index two in Gred . This
shows that the order of Gred is divided by 2. By Maschke’s theorem, p 6= 2.
There exists some matrix
0
b
−b−1 0
in Gred for some b ∈ k × . After taking conjugate by
−1/2
b
ζ8
0
,
0
b1/2 ζ8−1
we obtain the group scheme of type (Dn ) below (see Theorem 3.8) for appropriate n.
In conclusion, we have the following.
Theorem 3.8. Let k be an algebraically closed field of arbitrary characteristic
p (so p is a prime number, or ∞). Let G be a linearly reductive finite subgroup
scheme of SL2 . Then, up to conjugation, G agrees with one of the following,
where ζr denotes a primitive rth root of unity.
(An ) (n ≥ 1) The group scheme µn+1 lying in SL2 as
a 0
. | a ∈ µn+1 .
0 a−1
(Dn ) (n ≥ 4) p ≥ 3. The subgroup scheme generated by (A2n−5 ) and
0 ζ4
.
ζ4 0
8
(E6 ) p ≥ 5. The binary tetrahedral group generated by (D4 ) and
1 ζ87 ζ87
√
.
5
2 ζ8 ζ8
(E7 ) p ≥ 5. The binary octahedral group generated by (E6 ) and (A7 ).
(E8 ) p ≥ 7. The binary icosahedral group generated by (A9 ),
1
ζ5 + ζ5−1
1
0 1
, and
.
−1 0
1
−(ζ5 + ζ5−1 )
ζ52 − ζ53
Conversely, any of above is a linearly reductive finite subgroup scheme of SL2 ,
and a different type gives a non-isomorphic group scheme.
(3.9) For a finite k-group scheme G over k, we define |G| := dimk k[G].
Then in the theorem, |G| is n + 1 for (An ), 4n − 8 for (Dn ), and 24, 48,
and 120 for (E6 ), (E7 ), and (E8 ), respectively. This is independent of p, and
hence is the same as the case for p = ∞.
Corollary 3.10. Let k be an algebraically closed field of positive characteristic. Let R̂ be a two-dimensional F -rational Gorenstein complete local
ring with the coefficient field k. Then there is a linerly reductive finite subgroup scheme G of SL2 = SL(V ), where V = k 2 , such that the completion
of (Sym V )G with respect to the irrelevant maximal ideal is isomorphic to R̂.
Conversely, if G is such a group scheme, then the completion of (Sym V )G
is a two-dimensional F -rational Gorenstein complete local ring with the coefficient field k.
Proof. This follows from the theorem and the list in [HL, Example 18].
Let u, v be the standard basis of V = k 2 and G be as in the list of the
theorem. Let S = k[u, v] and R = S G .
The case that G = (An ). Then a G-algebra is nothing but a X (G) =
Z/(n + 1)Z-graded k-algebra. S is a G-algebra with deg u = 1 and deg v =
−1, and R = S0 , the degree 0 component with respect to this grading. Set
x = un+1, y = −v n+1 and z = uv. Then it is easy to see that R = k[x, y, z].
Obviously, it is a quotient of R1 = k[X, Y, Z]/(XY + Z n+1 ). As R1 is a
normal domain of dimension two, R1 = R. So R̂ is of type (An ).
9
The case that G = (Dn ). Set x = uv(u2n−4 − (−1)n−2 v 2n−4 ), y =
−2
u v and z = 2−1/(n−1) (u2n−4 + (−1)n−2 v 2n−4 ). Then
2/(n−1) 2 2
′
′
′
k[x, y, z] ⊂ R = S G = (S G )G/G ⊂ k[u2n−4 , uv, v 2n−4] = S G ⊂ S,
where G′ is the group scheme of type (A2n−5 ). Note that k[x, y, z] is a quotient
of R1 = k[x, y, z] = k[X, Y, Z]/(X 2 +Y Z 2 +Y n−1 ). As R1 is a two-dimensional
normal domain, R1 → k[x, y, z] is an isomorphism, and hence k[x, y, z] is
′
normal. It is easy to see that Q(S G ) = k(x, y, z, uv) and [k(x, y, z, uv) :
k(x, y, z)] ≤ 2. As |G/G′ | = 2, R1 = k[x, y, z] → R is finite and birational.
As R1 is normal, R1 = R. Thus R̂ is of type (Dn ).
The cases of constant groups G = (E6 ), (E7 ), (E8 ) are well-known [LW],
and we omit the proof.
Remark 3.11. Note that the converse in the corollary is also checked theoretically. As G is linearly reductive, R = (Sym V )G is a direct summand subring
of S = Sym V , and hence is strongly F -regular. Thus its completion is also
strongly F -regular, see for example, [Has2, (3.28)]. Gorenstein property of
R is a consequence of [Has, (32.4)].
Nevertheless, at this moment, the author does not know a theoretical
reason why G can be recovered from the isomorphism class of R̂ (this is true,
as can be seen from the result of the classification).
Remark 3.12. Let V and G be as above. Set S := (Sym V )G , and let Ŝ be its
completion with respect to the irrelevant maximal ideal so that Ŝ ∼
= k[[x, y]].
◦
◦
G◦
As G is infinitesimal, Ŝ → Ŝ is purely inseparable. So Spec Ŝ G \ 0 is
◦
simply connected. As Spec Ŝ G \ 0 → Spec Ŝ G \ 0 is a Galois covering of the
Galois group G/G◦ = Gred , the fundamental group of Spec Ŝ G \ 0 is Gred ,
which is linearly reductive, as stated in [Art].
References
[Art]
M. Artin, Coverings of rational double points in characteristic p,
in Complex Analysis and Algebraic Geometry, dedicated to K. Kodaira, W. L. Baily and T. Shioda (eds.), Cambridge (1977), pp. 11–
22.
[Dor]
L. Dornhoff, Group representation theory. Part A: Ordinary representation theory, Dekker (1971).
10
[Dur]
A. Durfee, Fifteen characterizations of rational double points and
simple critical points, L’Enseignement Math. 25 (1979), 131–163.
[Har]
N. Hara, A characterization of rational singularities in terms of
injectivity of Frobenius maps, Amer. J. Math. 120 (1998), 981–
996.
[Has]
M. Hashimoto, Equivariant twisted inverses, Foundations of
Grothendieck Duality for Diagrams of Schemes (J. Lipman, M.
Hashimoto), Lecture Notes in Math. 1960, Springer (2009),
pp. 261–478.
[Has2]
M. Hashimoto, F -pure homomorphisms, strong F -regularity, and
F -injectivity, Comm. Algebra 38 (2010), 4569–4596.
[HL]
C. Huneke and G. J. Leuschke, Two theorems about maximal
Cohen–Macaulay modules, Math. Ann. (2002), 391–404.
[LW]
G. J. Leuschke and R. Wiegand, Cohen–Macaulay Representations,
AMS (2012).
[Jan]
J. C. Jantzen, Representations of algebraic groups, Second edition,
AMS (2003).
[KS]
D. B. Karagueuzian and P. Symonds, The module structure of a
group action on a polynomial ring: A finiteness theorem, J. Amer.
Math. Soc. 20 (2007), 931–967.
[Sin]
A. K. Singh, Failure of F -purity and F -regularity in certain rings
of invariants, Illinois J. Math. 42 (1998), 441–448.
[Sm]
K. E. Smith, F -rational rings have rational singularities, Amer. J.
Math. 119 (1997), 159–180.
[Swe]
M. Sweedler, Hopf Algebras, Benjamin (1969).
[Swe2]
M. E. Sweedler, Connected fully reducible affine group schemes in
positive characteristic are Abelian, J. Math. Kyoto Univ. 11 (1971),
51–70.
[Wat]
W. C. Waterhouse, Introduction to Affine Group Schemes, Springer
(1979).
11
[Yas]
T. Yasuda, Pure subrings of regular local rings, endomorphism rings
and Frobenius morphisms, J. Algebra 370 (2012), 15–31.
[WY]
K.-i. Watanabe and K.-i. Yoshida, Hilbert–Kunz multiplicity and
an inequality between multiplicity and colength, J. Algebra 230
(2000), 295–317.
[Yos]
Y. Yoshino, Cohen–Macaulay Modules over Cohen–Macaulay
Rings, Cambridge (1990).
12
| 0math.AC
|
arXiv:cs/9906013v1 [cs.LO] 14 Jun 1999
Combining Inclusion Polymorphism
and Parametric Polymorphism
Sabine Glesner
Karl Stroetmann
Institut für Programmstrukturen
und Datenorganisation
Lehrstuhl Prof. Goos
Universität Karlsruhe
Postfach 6980, D 76128 Karlsruhe
Tel.: +49 / 721 - 608 7399
Fax: +49 / 721 - 300 47
[email protected]
Siemens AG
ZT SE 4
Otto-Hahn-Ring 6
D-81739 München
Tel.: +49 / 89 - 636 49 555
Fax: +49 / 89 - 636 42 284
[email protected]
Abstract
We show that the question whether a term is typable is decidable for type systems combining inclusion polymorphism with parametric polymorphism provided
the type constructors are at most unary. To prove this result we first reduce the
typability problem to the problem of solving a system of type inequations. The
result is then obtained by showing that the solvability of the resulting system of
type inequations is decidable.
1
1
Introduction
As a common agreement, a flexible type system needs to contain inclusion as well
as parametric polymorphism. Unfortunately, such a flexibility of the type system
causes type inference to become hard or even undecidable. In this paper, we investigate the problem of checking the well-typedness of terms in the presence of
inclusion polymorphism combined with parametric polymorphism. We show that
in this case typability is decidable, provided that the type constructors are at most
unary. This result has not been stated before.
The result of our paper can be used for the design of new type systems that
combine both inclusion polymorphism and parametric polymorphism. Type systems
of this kind are of interest for object-oriented programming languages. In particular,
our result is applicable for the programming language Java which up to now does
not allow for parametric polymorphism but will probably do so in future versions
[MBL97, OW97].
Another area where our result is applicable is logic programming. A number of
type systems have been designed in this area, e. g. [AM94, HT92, Pfe92, YFS92],
but the systems that have been implemented so far either offer no inclusion polymorphism at all [HL94, SHC95] or impose stronger restrictions [Bei95b] than a type
system that would be based on our result.
As it stands, our result cannot be applied to functional programming languages
because these languages allow for the binary type constructor → which takes two
types σ and τ and returns the type σ → τ of all functions mapping σ to τ .
In general, Tiuryn and Urzyczyn have shown that type inference for a type
system which combines inclusion polymorphism and parametric polymorphism is
undecidable for second-order types [TU96]. On the other hand, type inference
for inclusion polymorphism combined with nullary type constructors is decidable:
[Mit84, Mit91] presents an algorithm called MATCH, which solves type inequations
in the case that only inequations between nullary type constructors are allowed.
Fuh and Mishra [FM90] introduce a similar algorithm to solve the same problem.
The logic programming language Protos-L [Bei95a] is based on this type system.
This paper is organized as follows: Section 2 contains a definition of the type
language. In section 3 we define well-typed terms. Moreover, we show how the
question whether a term is well-typed can be reduced to the problem of solving a
system of type inequations. The solvability of these systems is shown to be decidable
in section 4. Section 5 concludes.
2
Type Language
In this section we first introduce a language for describing types. Since types behave
in many ways like terms, there is also a notion of substitution. This notion is defined
in subsection 2.2.
2.1
Types
Types are constructed from type constructors and type parameters. The set of type
constructors is partially ordered. This ordering is extended to types.
Definition 1 (Ordered Type Alphabet) An ordered type alphabet is a tuple
A = (A, #, ≤) such that
1. The type alphabet A is a finite set of type constructors. Elements of A are
denoted by K, L, M, · · · .
2
2. (A, ≤) is a partial order.
3. # : A → N is a function assigning an arity, #K, to every type constructor
K ∈ A.
⋄
Definition 2 (Types) To define types, we assume that an ordered type alphabet
A = (A, #, ≤) and a set P = {αi : i ∈ N} of type parameters are given. Then the
set of types T = T(A, P) is defined inductively:
• α ∈ T for all α ∈ P.
• If K ∈ A, #K = n, and σi ∈ T for all i = 1, . . . , n, then K(σ1 , . . . , σn ) ∈ T.
If #K = 0, then we write K instead of K().
⋄
Types are denoted by π, ̺, σ, and τ , while parameters are denoted by α and β. A
monotype is a type constructed without type parameters. If τ is a type, then Par(τ )
denotes the set of type parameters in τ .
Next, we extend the relation ≤ from A to the set of types T(A, P).
Definition 3 (Subtype Relation) Let A = (A, #, ≤) be an ordered type alphabet and let T(A, P) be the set of types constructed from A. Then the subtype
relation on T(A, P) is defined inductively:
1. If α ∈ P, then α ≤ α.
2. If K, L ∈ A, #K = m, and #L = n, then K(σ1 , . . . , σm ) ≤ L(τ1 , . . . , τn )
holds iff K ≤ L and σi ≤ τi for all i = 1, . . . , min(m, n).
⋄
Without further provisons, (T, ≤) is not a partial order. This is shown by the
counter example given next.
Example 1 Assume that A = {K1 , K2 , L1 , L2 } where #Ki = 0 and #Li = 1 for
i = 1, 2. The ordering ≤ on A is defined by the following chain of inequations:
L1 ≤ K1 ≤ K2 ≤ L2 .
Then we have L1 (K2 ) ≤ K1 and K1 ≤ L2 (K1 ). However, L1 (K2 ) 6≤ L2 (K1 ).
⋄
This problem is caused by an incompatibility between the arity function # : A → N
and the ordering of the type alphabet.
Definition 4 (Compatible) Assume a type alphabet A = (A, #, ≤) is given.
Then the arity # : A → N is compatible with the ordering ≤ iff the following
condition is satisfied for all type constructors K, L, and M :
K ≤ L ∧ L ≤ M ⇒ min(#K, #M ) ≤ #L.
⋄
Convention:
For the rest of this paper we assume the following: If an ordered
type alphabet (A, #, ≤) is given, then # is compatible with ≤.
Lemma 1 If A = (A, #, ≤) is an ordered type alphabet, then T(A, P), ≤ is a
partial order.
Proof: We need to show that the relation ≤ is reflexive, antisymmetric, and transitive. In order to prove the reflexivity, we have to show σ ≤ σ for all types σ. This
is done via a trivial induction on σ. To prove the antisymmetry, assume σ ≤ τ and
τ ≤ σ. We have to show σ = τ . The proof proceeds by induction on σ.
3
1. σ is a parameter α. Because of σ ≤ τ we know that τ = α.
2. σ = L(σ1 , . . . , σl ). Then τ = M (τ1 , . . . , τm ) and we must have L ≤ M and
M ≤ L. Since ≤ is a partial order on A, we have L = M and l = m. Further,
we have
σi ≤ τi
for all i = 1, . . . , l,
τi ≤ σi
for all i = 1, . . . , l.
and
The induction hypothesis yields σi = τi for all i = 1, . . . , l and then σ = τ is
immediate.
To prove the transitivity, assume that ̺, σ, τ ∈ T(A, P) are given such that ̺ ≤ σ
and σ ≤ τ . We need to prove ̺ ≤ τ . The proof proceeds by induction on σ.
1. σ is a parameter α. Then ̺ is α and, similarly, τ is α. Obviously, ̺ ≤ τ .
2. σ is L(σ1 , . . . , σl ). Then ̺ = K(̺1 , . . . , ̺k ) and τ = M (τ1 , . . . , τm ). The
assumption ̺ ≤ σ yields K ≤ L and
̺i ≤ σi
for all i = 1, . . . , min(k, l)
and, similarly, the assumption σ ≤ τ yields L ≤ M and
σi ≤ τi
for all i = 1, . . . , min(l, m).
Since ≤ is a partial order on A, we have K ≤ M . Further, the induction
hypothesis shows that
̺i ≤ τi
for all i = 1, . . . , min(k, l, m).
Since the arity # is compatible with ≤, we have min(k, m) ≤ l. Therefore,
min(k, l, m) = min(k, m). But then ̺ ≤ τ is immediate.
✷
2.2
Parameter Substitutions
Types behave in many ways like terms. Therefore there is also a notion of substitution. Since type parameters are substituted rather than variables, these substitutions are called parameter substitutions. Parameter substitutions are denoted by
the capital Greek letters Θ, Φ, and Ψ.
Definition 5 (Parameter Substitution) A parameter substitution Θ is a finite
set of pairs of the form
α1 7→ τ1 , . . . , αn 7→ τn
where α1 , . . . , αn are distinct parameters and τ1 , .., τn are types. It is interpreted
as a function mapping type parameters to types:
τi if α = αi ;
Θ(α) :=
α otherwise.
This function is extended to types homomorphically:
Θ F (σ1 , . . . , σn ) := F Θ(σ1 ), . . . , Θ(σn ) .
We use a postfix notation to denote the result of evaluating Θ on a type τ , i.e. we
write τ Θ instead of Θ(τ ).
The domain of Θ is defined as dom(Θ) := {α | α 6= αΘ}. The set of parameters
appearing in the range of a parameter substitution Φ is defined as
S
Par(Φ) := {Par(αΦ) | α ∈ dom(Φ) .
A parameter substitution is called a parameter renaming iff it has the form
4
[α1 7→ απ(1) , . . . , αn 7→ απ(n) ]
where π is a permutation of the set {1, . . . , n}.
If Θ1 and Θ2 are parameter substitutions, then their composition Θ1 ◦ Θ2 is
defined such that α(Θ1 ◦ Θ2 ) = (αΘ1 )Θ2 holds for all type parameters α.
⋄
Parameter substitutions respect the ordering ≤ on T.
Lemma 2 If Θ is a parameter substitution and σ, τ ∈ T, then
σ ≤ τ ⇒ σΘ ≤ τ Θ.
Proof: The proof is done by an induction following the definition of σ ≤ τ .
1. The case α ≤ α is obvious.
2. If σ = K(σ1 , . . . , σm ) ≤ L(τ1 , . . . , τn ) = τ , then K ≤ L and σi ≤ τi for
i = 1, . . . , min(m, n). Using the induction hypothesis we have σi Θ ≤ τi Θ for
all relevant i. Therefore K(σ1 Θ, . . . , σm Θ) ≤ L(τ1 Θ, . . . , τn Θ).
⋄
3
Well-Typed Terms
We define the set of well-typed terms in the first subsection. Then in subsection 3.2
we reduce the question whether a term is well-typed to the solvability of a system
of type inequations.
3.1
Definition of Well-Typed Terms
We assume a set of functions symbols F and a set of variables V to be given. Every
function symbol f ∈ F is supposed to have an arity.
Definition 6 (Terms) The set of terms T (F , V) is defined inductively:
1. If v ∈ V, then v ∈ T (F , V).
2. If f ∈ Σ, f is n-ary, and t1 , . . . , tn ∈ T (F , V), then f (t1 , . . . , tn ) ∈ T (F , V).
The set of variables occurring in a term t is defined by an obvious inductive definition
and denoted by Var(t). If this set is empty, then t is called a closed term. The set
of closed terms is denoted by T (F ).
⋄
Definition 7 (Signature) If f is n-ary, then its signature is a string of n+1 types.
If σ1 . . . σn τ is the signature of f , then this is communicated by writing
f : σ1 × . . . × σn → τ .
In the following, we assume that every function symbol f has a signature.
A signature ̺1 . . . ̺n π is appropriate for a function symbol f iff
f : σ1 × . . . × σn → τ
and there exists a parameter substitution Θ such that π = τ Θ and ̺i = σi Θ for
i = 1, . . . , n.
⋄
Definition 8 (Type Assignment) A type annotation is a pair written as t : τ
where t is a term and τ is a type. The type annotation t : τ is called a variable
annotation if t is a variable. If Γ = {x1 : τ1 , . . . , xn : τn } is a finite set of variable
annotations such that the variables xi are pairwise distinct, then we call Γ a type
5
assignment. If Γ = {x1 : τ1 , . . . , xn : τn } is a type assignment, then we regard Γ
as a function with domain {x1 , . . . , xn } mapping the variables xi to the types τi ,
i.e. we have Γ(xi ) = τi for i = 1, . . . , n and dom(Γ) = {x1 , . . . , xn }.
⋄
Definition 9 (Well-Typed Term) The notion of a well-typed term is defined via
a binary relation ⊢ taking as its first argument a type assignment and as its second
argument a type annotation. The definition of ⊢ is done inductively:
1. If Γ(x) ≤ π, then
Γ ⊢ x : π.
2. If we have
(a)
(b)
(c)
(d)
Γ ⊢ si : ̺i for all i = 1, . . . , n,
σ1 × . . . × σn → τ is appropriate for f ,
̺i ≤ σi for all i = 1, . . . , n, and
τ ≤ π,
then
Γ ⊢ f (s1 , . . . , sn ) : π.
A term t is well-typed iff there exist a type assignment Γ and a type τ such that
Γ ⊢ t : τ . We read Γ ⊢ t : τ as “Γ entails t : τ ”. We call Γ ⊢ t : τ a type judgement.
⋄
3.2
Type Checking
In this subsection, we reduce the question whether a term is well-typed to the
solvability of a system of type inequations. Here, a type inequation is a pair of
types written as σ τ . A parameter substitution Θ solves a type inequation σ τ
(denoted Θ |= σ τ ) if σΘ ≤ τ Θ. A system of type inequations is a set of type
inequations. A parameter substitution Θ solves a system of type inequations I
(denoted Θ |= I) iff Θ solves every type inequation in I.
Assume that Γ is a type assignment and t : τ is a type annotation such that
Var(t) ⊆ dom(Γ). We define a function ineq(Γ, t : τ ) by induction on t such that
ineq(Γ, t : τ ) is a system of type inequations. A parameter substitution Θ will solve
ineq(Γ, t : τ ) iff ΓΘ ⊢ t : τ Θ. The inductive definition of ineq(Γ, t : τ ) is given as
follows:
1. ineq(Γ, x : τ ) := {Γ(x) τ }
2. Assume the signature of f is given as f : σ1 × . . . × σn → σ, where the type
parameters have been appropriately renamed so that they are new, i.e. the
new parameters may occur neither in Γ nor in τ nor in any of the signatures
used to construct ineq(Γ, si : σi ) for some i = 1, . . . , n. Then
n
S
ineq(Γ, si : σi ).
ineq Γ, f (s1 , . . . , sn ) : τ := {σ τ } ∪
i=1
Before starting with the proofs of the soundness and completeness for the above
transformation, we state some definitions: If Γ is a type assignment and t : τ is a
type annotation, then Γ ⊲ t : τ is called a hypothetical type judgement. A parameter
substitution Θ solves a hypothetical type judgement Γ ⊲ t : τ iff ΓΘ ⊢ t : τ Θ holds.
A type constraint is either a type inequation or a hypothetical type judgement. A
parameter substitution Θ solves a set of type constraints C iff it solves every type
inequation and every hypothetical type judgement in C. This is written Θ |= C.
We define a rewrite relation on sets of type constraints. It is the least transitive
relation ❀ such that:
6
1. C ∪ {Γ ⊲ x : τ } ❀ C ∪ {Γ(x) τ }
2. Assume that the signature of f is given as f : σ1 ×· · ·×σn → σ where the type
parameters have been appropriately renamed so that they are new. Then
Sn
C ∪ {Γ ⊲ f (s1 , . . . , sn ) : τ } ❀ C ∪ {σ τ } ∪ i=1 {Γ ⊲ si : σi }.
If a hypothetical type judgement Γ ⊲ t : τ is given, then the two rewrite rules can be
used repeatedly until the set ineq(Γ, t : τ ) is derived. This is easily seen by induction
on t. Furthermore, the rewrite relation ❀ satisfies the following invariants:
1. (Θ |= C2 ) ∧ (C1 ❀ C2 ) ⇒ (Θ |= C1 )
(I1 )
2. (Θ |= C1 ) ∧ (C1 ❀ C2 ) ⇒ ∃Ψ.(Θ ⊆ Ψ ∧ Ψ |= C2 )
(I2 )
Before proving these invariants, we show that they suffice to verify the soundness
and completeness of our transformation.
Theorem 1 (Soundness of the Transformation) Assume Γ is a type assignment and t : τ is a type annotation. If Θ |= ineq(Γ, t : τ ), then ΓΘ ⊢ t : τ Θ.
Proof: Since the assumption is Θ |= ineq(Γ, t : τ ) and we know that {Γ ⊲ t : τ } ❀
ineq(Γ, t : τ ), the invariant (I1 ) shows that Θ |= {Γ ⊲ t : τ }. By definition, this
implies ΓΘ ⊢ t : τ Θ.
✷
Theorem 2 (Completeness of the Transformation) Assume Γ is a type assignment, t : τ is a type annotation, and Θ is a parameter substitution such that
ΓΘ ⊢ t : τ Θ. Then, Θ can be extended to a parameter substitution Φ that is a
solution of ineq(Γ, t : τ ).
Proof: ΓΘ ⊢ t : τ Θ implies Θ |= Γ ⊲ t : τ . Since {Γ ⊲ t : τ } ❀ ineq(Γ, t : τ ), the
invariant (I2 ) shows that Θ can be extended to a parameter substitution Φ such
that Φ |= ineq(Γ, t : τ ).
✷
Proof of (I1 ): According to the definition of the rewrite relation ❀, it suffices to
consider the following two cases:
1. C1 = C ∪{Γ⊲ x : τ } ❀ C ∪{Γ(x) τ } = C2 . The assumption is that Θ |= C2 .
Then Θ |= C and Γ(x)Θ ≤ τ Θ. Therefore, ΓΘ ⊢ x : τ Θ showing Θ |= C1 .
Sn
2. C1 = C ∪ {Γ ⊲ f (s1 , . . . , sn ) : τ } ❀ C ∪ {σ τ } ∪ i=1 {Γ ⊲ si : σi } = C2 ,
where f : σ1 × · · · σn → σ. According to the assumption, we have Θ |= C,
σΘ ≤ τ Θ, and Θ |= Γ ⊲ si : σi for i = 1, . . . , n. Then ΓΘ ⊢ si : σi Θ for
i = 1, . . . , n. Therefore, ΓΘ ⊢ f (s1 , . . . , sn ) : τ Θ and that yields the claim.
✷
To prove the invariant (I2 ) we need the following lemma, which follows directly
from Defs. 7 and 9.
Lemma 3 Suppose that t = f (s1 , . . . , sn ) and f : σ1 × · · · × σn → σ . Then
Γ ⊢ t : τ iff there is a parameter substitution Θ such that σΘ ≤ τ and Γ ⊢ si : σi Θ
for all i = 1, . . . , n.
Proof of (I2 ): Again, it suffices to consider the following two cases corresponding
to the definition of the relation ❀:
1. C1 = C ∪{Γ⊲ x : τ } ❀ C ∪{Γ(x) τ } = C2 . The assumption is that Θ |= C1 .
Then Θ |= C and ΓΘ ⊢ x : τ Θ. Therefore, Γ(x)Θ ≤ τ Θ. Define Ψ := Θ.
7
Sn
2. C1 = C ∪ {Γ ⊲ f (s1 , . . . , sn ) : τ } ❀ C ∪ {σ τ } ∪ i=1 {Γ ⊲ si : σi } = C2 ,
where f : σ1 × · · · σn → σ. W.l.o.g. we assume that the type parameters
occurring in this signature do not occur in dom(Θ), since the type parameters
in the signature can be renamed. According to our assumption, we have Θ |=
C and Θ |= {Γ ⊲ f (s1 , . . . , sn ) : τ }. The latter implies ΓΘ ⊢ f (s1 , . . . , sn ) :
τ Θ. Lemma 3 shows that there is a parameter substitution Φ such that
ΓΘ ⊢ si : σi Φ for all i = 1, . . . , n and σΦ ≤ τ Θ. We can assume that
dom(Φ) contains only type parameters occurring in the signature of f . Then
dom(Θ) ∩ dom(Φ) = ∅. Define Ψ := Θ ∪ Φ.
✷
When checking whether a term t is well-typed we want to compute a type
assignment Γ and a type τ such that Γ ⊢ t : τ holds. To this end, we define a
most general type assignment S
Γinit and a most general type τinit : Let Var(t) be the
variables in t. Define Γinit = x∈Var(t) {x : αx } and τinit = α where αx and α are
distinct new type parameters. The claim now
is that t is well-typed if and only if
the set of type constraints ineq Γinit , t : τinit is solvable.
Proof: “⇒”: Assume t is well-typed. Then there exists a type assignment Γ
and a type τ such that Γ ⊢ t : τ . Define a parameter substitution Θ by setting
Θ(αx ) = Γ(x) for x ∈ Var(t) and
Θ(α) = τ . Then we have Γ(x) = Γinit (x)Θ and
τ = τinit Θ and therefore Θ |= Γinit ⊲ t : τinit . Since
Γinit ⊲ t : τ ❀ ineq Γinit , t : τinit ,
the invariant (I2 ) shows that there exists a parameter substitution Ψ such that
Ψ |= ineq(Γinit , t : τinit ).
“⇐”: On the other hand, if Ψ |= ineq Γinit , t : τinit , then Theorem 1 shows that
Γinit Ψ ⊢ t : τinit Ψ holds.
✷
Therefore, the problem whether a term t is well-typed is reduced to the problem
of solving systems of type inequations.
4
Solving Systems of Type Inequations
In this section, we assume that type constructors are at most unary, i.e., given an
ordered type alphabet A = (A, #, ≤) we have that #K ≤ 1 for all K ∈ A. We show
that then it is decidable whether a system S of type inequations is solvable. To
this end we present an algorithm which effectively tests all possible instantiations
for the type parameters in the type inequations. The fact that the type constructors are at most unary enables us to guarantee three important properties during
this instantiation process: We do not create any additional parameters; we do not
increase the overall number of inequations; and the depth of the terms in the type
inequations does not increase. Therefore we can generate only finitely many systems of instantiated type inequations. If one of these systems is solvable, then we
can construct a solution for S.
4.1
Some Definitions
We start with some definitions necessary to formulate the algorithm for checking
the solvability of systems of type inequations.
4.1.1
Solvability and Equivalence of Type Inequations
A system of type inequations I is solvable (denoted ✸I) iff there is a parameter
substitution Φ such that Φ |= I. Two type inequations I1 and I2 are equivalent
(denoted I1 ≈ I2 ) iff a parameter substitution Φ solves I1 if and only if Φ solves I2 :
8
def
I1 ≈ I2 ⇐⇒ ∀Φ · (Φ |= I1 ⇔ Φ |= I2 )
A type inequation I is equivalent to true (denoted I ≈ true) iff every parameter
substitution solves I, it is equivalent to false (denoted I ≈ false) iff no parameter
substitution solves I. Two systems of type inequations I1 and I1 are equivalent
(denoted I1 ≈ I2 ) iff a parameter substitution Φ solves I1 if and only if Φ solves
I2 :
def
I1 ≈ I2 ⇐⇒ ∀Φ · (Φ |= I1 ⇔ Φ |= I2 )
Next, a system of type inequations I is equivalent to a set of systems of type
inequations I (denoted I ≈ I) iff I is solvable if and only if there is a system J ∈ I
such that J is solvable:
def
I ≈ I ⇐⇒ ✸I ⇔ ∃J ∈ I · ✸J .
To proceed, we define the depth of a type inductively:
1. depth(α) := 0 for all type parameters α.
2. depth(K) := 1 for all nullary type constructors K.
3. depth(K(σ)) := 1 + depth(σ).
The depth of a type inequation is defined by taking the maximum:
depth(σ τ ) := max depth(σ), depth(τ ) .
Furthermore, we define depth(true) := depth(false) := 0. The function depth is
then extended to systems of type inequations:
depth(I) := max depth(I) | I ∈ I .
The depth of a parameter substitution Φ is defined as
depth(Φ) := max depth(αΦ) | α ∈ dom(Φ) .
We define the depth of the empty parameter substitution as 0. A system of inequations I is solvable at depth k (denoted ✸k I) iff there is a closed parameter
substitution Φ such that Φ |= I and depth(Φ) ≤ k.
4.1.2
Definition of nf
The function nf takes a type inequation as input and either produces an equivalent
type inequation or yields true or false. The function is defined inductively.
1. nf(α σ) := α σ and nf(σ α) := σ α for every type parameter α.
true
iff K ≤ L;
2. nf(K L) :=
false else.
true
iff K ≤ L;
3. nf K L(τ ) :=
false else.
true
iff K ≤ L;
4. nf K(σ) L :=
false else.
nf(σ τ ) iff K ≤ L;
5. nf K(σ) L(τ ) :=
⋄
false
else.
It is easy to see that nf(I) ≈ I holds for every inequation I. We extend the function
nf to systems of type inequations. First, we define an auxiliary function nfaux :
9
nfaux (I) :=
nf(I) | I ∈ I ∧ nf(I) 6= true .
Then, the function nf(I) is defined as
{false} if false ∈ nfaux (I);
nf(I) :=
nfaux (I) otherwise.
It is easy to see that I ≈ nf(I) for any system of type inequations I.
4.1.3
Definition of AllParSubst
Next, we define the function AllParSubst. The input to AllParSubst is a finite set
A of type parameters. The output is the set of parameter substitutions Φ such
that dom(Φ) ⊆ A, depth(Φ) ≤ 1, and Par(αΦ) ⊆ {α} but αΦ 6= α for all α ∈ A.
Therefore, AllParSubst(A) is equal to the set
Φ | dom(Φ) ⊆ A ∧ depth(Φ) ≤ 1 ∧ ∀α ∈ A · Par(αΦ) ⊆ {α} ∧ αΦ 6= α .
The function AllParSubst has the following properties:
1. AllParSubst(A) is finite.
This is true because the type alphabet is assumed to be finite. Therefore,
given a finite set A of type parameter there are only finitely many types τ
such that depth(τ ) ≤ 1 and Par(τ ) ⊆ A. But then AllParSubst(A) must be
finite, too.
2. If Ψ is a parameter substitution such that depth(Ψ) = n ≥ 1 and Par(Ψ) = ∅,
then there exist parameter substitutions Φ1 and Φ2 such that
(a) Φ1 ∈ AllParSubst(dom(Ψ)),
(b) depth(Φ2 ) = n − 1, and
(c) Φ = Φ1 ◦ Φ2 .
To prove this, assume Ψ = [α1 7→ τ1 , . . . , αn 7→ τn ]. For those τi such that
depth(τi ) > 1, we must have τi = Li (σi ) for some type constructor Li and
some type σi with depth(σi ) < depth(τi ). W.l.o.g. assume that depth(τi ) ≤ 1
for all i = 1, . . . , m − 1 and depth(τi ) > 1 for all i = m, . . . , n. Then define
Φ1 := [α1 7→ τ1 , . . . , αm−1 7→ τm−1 , αm 7→ Lm (αm ), . . . , αn 7→ Ln (αn )]
Φ2 := [αm 7→ σm , . . . , αn 7→ σn ].
and
Then the claim is obvious.
3. I ≈ IΦ | Φ ∈ AllParSubst Par(I) .
Assume Φ |= I where w.l.o.g. dom(Φ) ⊆ Par(I). Then the previous property
shows that Φ can be written as Φ1 ◦ Φ2 where Φ1 ∈ AllParSubst Par(I) . But
then Φ2 |= IΦ1 .
Conversely, if Ψ |= IΦ for a substitution Φ ∈ AllParSubst Par(I) , then
Φ ◦ Ψ |= I.
4. If Φ ∈ AllParSubst Par(I) , then depth nf(IΦ) ≤ depth(I).
Assume σ τ is an inequation in I of maximal depth. First, assume σ =
K(σ ′ ) and τ = L(τ ′ ). When going from I to nf(IΦ) this inequation either
disappears or it has the form nf(σ ′ Φ τ ′ Φ). But the depth of this inequation
is not greater than the depth of the original inequation.
Next, σ = α for a parameter α and τ = L(τ ′ ). But then σΦ must have
either of the forms K or K(σ ′ ). When going from I to nf(IΦ) the inequation
σ τ either disappears or it has the form nf(σ ′ Φ τ ′ Φ). Again the depth of
10
this inequation is not greater than the depth of the original inequation. The
remaining cases are similar.
4.1.4
Definition of Inst
The function Inst transforms a single system of type inequations into an equivalent
set of systems of type inequations. It is defined as
Inst(I) := nf(IΦ) | Φ ∈ AllParSubst Par(I) ∧ nf(IΦ) 6= {false} .
The function Inst has the following properties:
1. Inst(I) is finite.
2. I ≈ Inst(I).
3. If ✸k I and k ≥ 1, then there is a J ∈ Inst(I) such that ✸k−1 J .
4. If ✸k J and J ∈ Inst(I), then ✸k+1 I.
5. If J ∈ Inst(I), then Par(J ) ⊆ Par(I).
6. If J ∈ Inst(I), then depth(J ) ≤ depth(I).
These properties are immediate consequences of the definition of Inst and the properties of the function AllParSubst.
4.2
Deciding Type Inequations
We present an algorithm for solving (or refuting) systems of type inequations. The
algorithm maintains two sets of systems of inequations. Call theses sets M and
A. M serves as a memory of systems of type inequations that have already been
encountered, while A contains systems of type inequations that can be derived from
I by application of the function Inst. The algorithm initializes both M and A to the
singleton {I}, where I is the system of type inequations that is to be solved. After
this initialization, the algorithm enters a loop. In this loop, we compute Inst(J )
for all J ∈ A. Then, we update A as follows:
S
A :=
Inst(J ) | J ∈ A − M
that is, we apply Inst to all systems in A and we discard those systems that appear
already in the memory M. If ∅ ∈ A, then I is solvable and the algorithm halts
with success. If A becomes empty, the algorithm halts with failure. Otherwise, we
update M as
M := M ∪ A
and reenter the loop. Figure 1 specifies the algorithm formally.
Lemma 4 (Termination) The algorithm given in Figure 1 terminates.
Proof: For every system of inequations J ∈ M the number of inequations in
J is less or equal than the number of inequations in I, Par(J ) ⊆ Par(I), and
depth(J ) ≤ depth(I). Since the type alphabet is finite, the size of M must therefore
be bounded.
Now assume the algorithm given in Figure 1 does not terminate. Then the set
An+1 can never be empty. Therefore, every time the loop is executed, the statement
M := M ∪ An+1 increases the number of elements of the set M. But then the size
of M would increase beyond every bound.
✷
11
Input: I
% system of type inequations to be solved
M := {I};
A0 := {I};
n := 0;
Loop:
S
An+1 :=
Inst(J ) | J ∈ An − M
if ∅ ∈ An+1 then
return true;
end-if;
if An+1 = ∅ then
return false;
end-if;
M := M ∪ An+1 ;
n := n + 1;
goto Loop;
Figure 1: An algorithm for deciding solvability of type inequations.
Lemma 5 (Soundness) Assume n, k ∈ N, J ∈ An and ✸k J . Then ✸k+n I.
Proof: The proof is given by induction on n.
1. n = 0: Since A0 = {I} we must have J = I and the claim is trivial.
2. n → n + 1: Assume J ∈ An+1 with ✸k J . Then there is a K ∈ An such that
J ∈ Inst(K). This implies ✸k+1 K. By i.h. we have ✸(k+1)+n I.
✷
Lemma 6 Assume that ✸k I and k is minimal with this property. Then for all
n ≤ k there is a J ∈ An such that ✸k−n J .
Proof: The proof is done by induction on n.
1. n = 0: Obvious.
2. n → n + 1: Assume ✸k I and that k is minimal with this property. By
i.h. there is a J ∈ An such that ✸k−n J . Then there is a K ∈ Inst(J ) such
that ✸k−n−1 K. Assume K ∈ M. Since
M=
n
S
Ai ,
i=1
there is an i ≤ n such that K ∈ Ai . Therefore Lemma 5 shows ✸k−n−1+i I.
Since k − n − 1 + i < k this contradicts the minimality of k. This shows
that the assumption K ∈ M is wrong and we have K ∈ An+1 . Because of
✸k−(n+1) K the proof is complete.
✷
Theorem 3 The algorithm given in Figure 1 is correct.
Proof: Assume that I is solvable. Then ✸n I for some n ∈ N. By Lemma 6 we find
a J ∈ An such that ✸0 J holds. But then J = ∅ and the algorithm returns true.
Assume now that I is not solvable. If the algorithm would return true, then
∅ ∈ An for some n ∈ N. Since ✸0 ∅ Lemma 5 would then give ✸n I. Therefore the
algorithm cannot return true. Since it terminates, it must return false.
✷
12
5
Conclusion
In this paper we have presented a type system that supports both inclusion polymorphism and parametric polymorphism. We were able to prove that for this type
system typability is decidable, provided we use at most unary type constructors.
In practice, many interesting type constructors are either nullary or unary. Unary
type constructors occur naturally when dealing with container types, e. g. types
that are interpreted as sets, lists, or bags. It is convenient to be able to cast, for example, lists to sets. This cannot be done with the type system proposed by Mitchell
[Mit84], but is possible with the type system introduced in this paper.
Previously, it has been know that type inference is decidable for a system that
restricts inclusion polymorphism to nullary type constructors [FM90, Mit84, Mit91].
On the negative side, Tiuryn and Urzyczyn [TU96] have shown that the type inference problem for second-order types is undecidable. We have shown in this paper,
that typability is decidable for type systems with at most unary type constructors.
It is still an open question whether typability is decidable in the case of binary type
constructors.
Acknowledgement: The authors would like to thank Pawel Urzyczyn for pointing
out some technical weaknesses in an earlier version of this paper.
References
[AM94]
Krzysztof R. Apt and Elena Marchiori. Reasoning about Prolog programs:
From modes through types to assertions. Formal Aspects of Computing,
6A:743–764, 1994.
[Bei95a] Christoph Beierle. Concepts, implementation, and applications of a typed
logic programming language. In Christoph Beierle and Lutz Plümer, editors, Logic Programming: Formal Methods and Practical Applications,
chapter 5, pages 139–167. Elsevier Science B.V./North-Holland, 1995.
[Bei95b] Christoph Beierle. Type inferencing for polymorphic order-sorted logic
programs. In Leon Sterling, editor, Proceedings of the 1995 International
Conference on Logic Programming. MIT Press, 1995.
[FM90]
You-Chin Fuh and Prateek Mishra. Type inference with subtypes. Theoretical Computer Science, 73(2):155–175, 1990.
[HL94]
Patricia M. Hill and John W. Lloyd. The Gödel Programming Language.
MIT Press, 1994.
[HT92]
P. M. Hill and R. W. Topor. A semantics for typed logic programs. In
Pfenning [Pfe92], pages 1–62.
[MBL97] Andrew C. Myers, Joseph A. Bank, and Barbara Liskov. Parameterized
Types for Java. In Proceedings of the 24th Symposium on Principles of
Programming Languages, pages 132–145. ACM Press, 1997.
[Mit84]
John C. Mitchell. Coercion and type inference. In 11th Annual ACM Symposium on Principles of Programming Languages, pages 175–185, 1984.
[Mit91]
John C. Mitchell. Type inference with simple subtypes. Journal of Functional Programming, 1:245–285, 1991.
[OW97] Martin Odersky and Philip Wadler. Pizza into Java: Translating theory
into practice. In Proceedings of the 24th Symposium on Principles of
Programming Languages, pages 146–159. ACM Press, 1997.
13
[Pfe92]
Frank Pfenning, editor. Types in Logic Programming. The MIT Press,
1992.
[SHC95] Zoltan Somogyi, Fergus J. Henderson, and Thomas Conway. Mercury: an
efficient purely declarative logic programming language. In Proceedings
of the Australian Computer Science Conference, pages 499–512, Glenelg,
Australia, February 1995.
[TU96]
Jerzy Tiuryn and Pawel Urzyczyn. The subtyping problem for secondorder types is undecidable. In Proceedings of the IEEE Symposion on
Logic in Computer Science (LICS 96), pages 74–85, 1996.
[YFS92] Eyal Yardeni, Thom Frühwirth, and Ehud Shapiro. Polymorphically
typed logic programs. In Pfenning [Pfe92], pages 63–90.
14
| 6cs.PL
|
IMAG
arXiv:cs/0412012v1 [cs.PL] 3 Dec 2004
Institut d’Informatique et de
Mathématiques Appliquées
de Grenoble
LSR
Laboratoire Logiciels, Systèmes, Réseaux
RAPPORT DE RECHERCHE
Jartege: a Tool for Random Generation of
Unit Tests for Java Classes
Catherine Oriat
RR 1069
Juin 2004
B.P. 72 – 38402 SAINT MARTIN D’HERES CEDEX – France
Centre National de la Recherche Scientifique
Institut National Polytechnique de Grenoble
Université Joseph Fourier Grenoble I
Jartege: a Tool for Random Generation of Unit Tests for
Java Classes
Catherine Oriat
LSR-IMAG, Grenoble
email: [email protected]
Résumé
Ce rapport présente Jartege, un outil qui permet la génération aléatoire de tests unitaires
pour des classes Java spécifiées en JML. JML (Java Modeling Language) est un langage
de spécification pour Java qui permet d’écrire des invariants pour des classes, ainsi que
des pré- et des post-conditions pour des opérations. Comme dans l’outil JML-JUnit, nous
utilisons les spécifications JML d’une part pour éliminer des cas de test non pertinents,
et d’autre part comme oracle de test. Jartege génère de façon aléatoire des cas de test,
qui consistent en une séquence d’appels de constructeurs et de méthodes des classes sous
test. L’aspect aléatoire de l’outil peut être paramétré en associant des poids aux classes
et aux opérations, et en contrôlant le nombre d’instances créées pour chaque classe sous
test. L’utilisation pratique de Jartege est illustrée par une petite étude de cas.
Mots-clés
Test, test unitaire, génération aléatoire de cas de test, Java, JML
Abstract
This report presents Jartege, a tool which allows random generation of unit tests for Java
classes specified in JML. JML (Java Modeling Language) is a specification language for
Java which allows one to write invariants for classes, and pre- and postconditions for
operations. As in the JML-JUnit tool, we use JML specifications on the one hand to
eliminate irrelevant test cases, and on the other hand as a test oracle. Jartege randomly
generates test cases, which consist of a sequence of constructor and method calls for the
classes under test. The random aspect of the tool can be parameterized by associating
weights to classes and operations, and by controlling the number of instances which are
created for each class under test. The practical use of Jartege is illustrated by a small case
study.
Keywords
Testing, unit testing, random generation of test cases, Java, JML
Jartege: a Tool for Random Generation
of Unit Tests for Java Classes
Catherine Oriat
LSR-IMAG, Grenoble
email: [email protected]
1
Introduction
Main validation technique in software engineering, program testing aims at ensuring that
the program is correct, i.e. conforms to its specifications. As the input domain of a
program is usually very large or infinite, exhaustive testing, which consists in testing the
program for all its possible inputs, is in general impossible. The objective of testing is
thus rather to improve the software quality by finding faults in it. Testing is an important
activity of software development, whose cost is usually estimated to about 40% of the total
cost of software development, exceeding the cost of code writing.
A test campaign for a program requires several steps: design and development of test
sets, execution and results examination (or oracle). Considering the cost of testing, it is
interesting to automate some of these steps.
For Java programs, the JUnit framework [JUn, BG98] allows the developer to write
an oracle for each test case, and to automatically execute test sets. JUnit in particular
permits to automatically regression test several test sets.
If a formal specification is available, it can be translated into assertions which can be
checked at runtime, and thus serve as a test oracle. For instance, the DAISTS system
[GMH81] compiles algebraic axioms of an abstract data type into consistency checks;
Rosenblum’s APP pre-processor allows the programmer to write assertions for C programs
[Ros92]; the Eiffel design and contract approach integrates assertions in the programming
language [Mey88, Mey92].
It is also interesting to automate the development of tests. We can distinguish between
two groups of strategies to produce test sets: random and systematic strategies. Systematic strategies, such as functional testing or structural testing, consist in decomposing the
input domain of the program in several subdomains, often called “partitions”.
Many systematic strategies propose to derive test cases from a formal specification
[Gau95]. For instance, the Dick and Faivre method [DF93], which consists in constructing
a finite state automaton from the formal specification and in selecting test cases as paths
in this automaton, has been used as a basis by other approaches, in particular Casting
[ABM97] or BZTT [LPU02]. BZTT uses B or Z specifications to generate test cases which
consist in placing the system in a boundary state and calling an operation with a boundary
value.
In contrast to systematic methods, random testing generally does not use the program
nor the specification to produce test sets. It may use an operational profile of the program,
which describe how the program is expected to be used. The utility of random testing is
controversial in the testing community. It is usually presented as the poorest approach for
selecting test data [Mye94]. However, random testing has a few advantages which make
us think that it could be a good complement to systematic testing:
• random testing is cheap and rather easy to implement. In particular, it can produce
large or very large test sets;
3
• it can detect a substantial number of errors at a low cost [DN84, HT90, Nta01].
Moreover, if an operational profile of the program is available,
• random testing allows early detection of the failures that are most likely to appear
when using the program [FHLS98];
• it can be used to evaluate the program reliability [HT90, Ham94].
However, partition testing can be much more effective at finding failures, especially when
the strategy defines some very small subdomains with a high probability to cause failures
[WJ91].
Among the best practices of extreme programming, or XP [Bec99, Bec00], are continuous testing and code refactoring. While they are writing code, developers should write
corresponding unit tests, using a testing framework such as JUnit. Unit testing is often
presented as a support to refactoring: it gives the developer confidence that the changes
have not introduced new errors. However, code refactoring often requires to change some
of the corresponding unit tests as well. If a class has to be tested intensively, the amount
of code corresponding to the tests often exceeds the amount of code of the class. Both
practices can therefore be hard to conciliate.
This report proposes to use random generation of tests to facilitate countinuous testing
and code refactoring in the context of extreme programming. We presents Jartege, a tool
for random generation of unit tests for Java classes specified in JML, which aims at easily
producing numerous test cases, in order to detect a substantial number of errors at a low
cost.
The rest of the report is organized as follows. Section 2 introduces the approach.
Section 3 presents a case study which consists in modeling bank accounts. This case study
will serve to illustrate the use of our testing tool. Section 4 presents our tool Jartege
and how it can be used to test the bank account case study. Section 5 introduces more
advanced features of Jartege, which allow one to parameterize its random aspect. Section
6 shows the errors which are detected by test cases generated by Jartege in the case study.
Section 7 presents and compares related approaches. Section 8 discusses some points about
random generation of tests and draw future work we intend to undertake around Jartege.
2
Approach
JML (Java Modeling Language) is a specification language for Java inspired by Eiffel,
VDM and Larch, which was designed by Gary Leavens and his colleagues [JML, LBR03,
LPC+ 03]. Several teams are currently still working on JML design and tools around JML
[BCC+ 03]. JML allows one to specify various assertions in particular invariants for classes
as well as pre- and postconditions for methods.
The JML compiler (jmlc) [CL02a] translates JML specifications into assertions checked
at runtime. If an assertion is violated, then a specific exception is raised. In the context
of a given method call, the JML compiler makes a useful difference between an entry
precondition which is a precondition of the given method, and an internal precondition,
which is a precondition of an operation being called, at some level, by the given method.
The JML-JUnit tool [CL02b] generates JUnit test cases for a Java program specified
in JML, using the JML compiler to translate JML specifications into test oracles. Test
cases are produced from a test fixture and parameter values supplied by the user.
4
Our approach is inspired by the JML-JUnit tool: we propose to generate random tests
for Java programs specified in JML, using this specification as a test oracle, in the JMLJUnit way. Our aim is to produce a big number of tests at a low cost, in order to facilitate
unit testing.
To implement these ideas, we started developing a prototype tool, called Jartege for
Java Random Test Generator. Jartege is designed to generate unit tests for Java classes
specified with JML. By unit tests, we here mean tests for some operations in a single class
or a small cluster of classes. In our context, a test case is a Java method which consists of
a sequence of operation (constructor or method) calls.
As in the JML-JUnit tool, we use the JML specification to assist test generation in
two ways:
1. It permits the rejection of test sequences which contain an operation call which
violates the operation entry precondition. We consider that these test sequences are
not interesting because they detect errors which correspond to a fault in the test
program.
(Although such sequences could be used to detect cases when a precondition is too
strong, this goal would prevent us from producing long sequences of calls. Thus,
we choose to trust the specification rather than the code. We can note that if an
operation uses a method whose precondition is too strong, this will produce an
internal precondition error and the sequence will not be rejected.)
2. The specification is also used as a test oracle: the test detects an error when another
assertion (e.g. an invariant, a postcondition or an internal precondition) is violated.
Such an error corresponds to a fault in the Java program or in the JML specification.
Our work has been influenced by the Lutess tool [Par96, dBORZ99], which aims at deriving test data for synchronous programs, with various generation methods, in particular
a purely random generation and a generation guided by operational profiles. The good
results obtained by Lutess, which won the best tool award of the first feature interaction
detection contest [dBZ99], have encouraged us to consider random generation of tests as
a viable approach.
3
Case Study
In this section, we present a case study to illustrate the use of Jartege. This case study
defines bank accounts and some operations on these accounts.
3.1
Informal Specification of the Bank Accounts
We describe here an informal specification of bank accounts:
1. An account contains a certain available amount of money (its balance), and is associated with a minimum amount that this account may contain (the minimum
balance).
2. It is possible to credit or debit an account. A debit operation is only possible if there
is enough money on the account.
3. One or several last credit or debit operations may be cancelled.
5
4. The minimum balance of an account may be changed.
3.2
Bank Account Modeling
To represent accounts, we define a class Account with two attributes: balance and min
which respectively represent the balance and the minimum balance of the account, and
three methods: credit, debit and cancel. In order to implement the cancel operation, we
associate with each account an history, which is a linked list of the previous balances of
the account. The class History has one attribute, balance, which represents the balance
of its associated account before the last credit or debit operation. With each history is
associated its preceding history. Figure 1 shows the UML class diagram of bank accounts.
hist
Account
History
0..1
balance
min
balance
0..1
credit(amount: int)
debit(amount: int)
cancel()
prec
Figure 1: UML diagram of the case study
3.3
JML Specification and Java Implementation
We implement both class Account and History in Java, and specify them with JML. We
choose to write lightweight public specifications to emphasize that these specifications are
destined for clients and need not be complete.
In order to forbid uncontrolled modifications of the attributes, we declare them as
private and define associated access methods: getBalance, getMin and getHist in class
Account and getBalance and getPrec in class History. These methods are specified as pure
methods in JML because they are side effect free, which allows us to use them in JML
assertions.
The class Account has an invariant which specifies that the balance of an account must
always be greater than the minimum balance.
/* Class of bank accounts. */
public class Account {
/* Invariant of class Account. */
/*@ public invariant getBalance( ) >= getMin( ); */
private int balance; // The balance of this account
private int min;
// The minimum balance
private History hist; // The history list of this account
/* The balance of this account. */
public /*@ pure */ int getBalance( ) { return balance; }
6
/* The history list of this account. */
public /*@ pure */ History getHist( ) { return hist; }
/* The minimum balance of this account. */
public /*@ pure */ int getMin( ) { return min; }
The constructor of class Account constructs an account with the specified balance and the
specified minimum balance. Its precondition asserts that the specified balance is greater
than the specified minimum balance.
/* Constructs an account with the specified balance and
* minimum balance. */
/*@ requires balance >= min; */
public Account (int balance, int min) {
this.balance = balance;
this.min = min;
this.hist = null ;
}
As the minimum balance min is a private attribute, we have to introduce a method to
change its value. The method setMin (int min) sets the minimum balance to the specified
value. Its precondition asserts that the balance is greater than the specified minimum
value.
/* Sets the minimum balance to the specified value. */
/*@ requires getBalance ( ) >= min; */
public void setMin (int min) { this.min = min; }
The method credit (int amount) credits the account with the specified amount. Its precondition requires the amount to be positive. Its postcondition asserts that the new balance is
the former balance augmented by the specified amount, that a new history is created with
balance the former balance of the account and with previous history the former history of
the account. Its exceptional postcondition asserts that the method should never terminate
abruptly.
/* Credits this account with the specified amount. */
/*@ requires amount >= 0;
*@ ensures getBalance ( ) == \old (getBalance ( )) + amount &&
*@
\fresh (getHist ( )) &&
*@
getHist ( ).getBalance ( ) == \old (getBalance ( )) &&
*@
getHist ( ).getPrec ( ) == \old (getHist ( ));
*@ signals (Exception e) false;
*/
public void credit(int amount) {
hist = new History (balance, getHist ( ));
balance = balance + amount;
}
The debit operation, which is very similar to the credit operation, is not detailed here.
It has the additional precondition that the balance decreased by the specified amount is
greater than the minimum balance.
7
The method cancel cancels the last credit or debit operation. Its precondition requires
that the history is not null, which means that at least one operation of credit or debit has
taken place since the account was created. Its postcondition ensures that the balance and
the history of the account have been updated with their former values.
/* Cancels the last credit or debit operation. */
/*@ requires getHist ( ) != null ;
*@ ensures getHist ( ) == \old (getHist ( ).getPrec ( )) &&
*@
getBalance ( ) == \old (getHist ( ).getBalance ( ));
*@ signals (Exception e) false;
*/
public void cancel ( ) {
balance = hist.getBalance ( );
hist = hist.getPrec ( );
}
} // End of class Account
We do not define any JML assertion for the class History.
/* Class of histories. */
public class History {
private int balance; // The balance of this history.
private History prec; // The preceding history.
/* Constructs a history with the specified balance and preceding history. */
public History (int balance, History prec) {
this.balance = balance; this.prec = prec;
}
/* The balance of this history. */
public /*@ pure */ int getBalance ( ) { return balance; }
/* The preceding history. */
public /*@ pure */ History getPrec ( ) { return prec; }
} // End of class History
4
Jartege
Jartege (Java Random Test Generator) is a framework for automatic random generation
of unit tests for Java classes specified with JML. This approach consists in producing test
programs which are composed of test cases, each test case consisting of randomly chosen
sequences of method calls for each class under test. Each generated test program can be
executed to test the classes, and re-executed later on either after having corrected some
faults or for regression test.
The tool is designed to produce unit tests, i.e. tests composed of calls of some methods
which belong to a few classes. As noticed in [LTFWD00], because of complex dependences
that exist between classes in object-oriented programs, it is usually not possible to test a
method or a class in complete isolation. Jartege thus is able to generate test cases which
allow the integration of several classes.
8
4.1
Practical Use of Jartege
Suppose we wish to generate tests for the classes Account and History. We write the
following Java program:
import jartege.*;
/** Jartege test cases generator for classes Account and History. */
class TestGen {
public static void main (String[ ] args) {
// Creates a class tester
ClassTester t = new ClassTester ( );
// Adds the specified classes to the set of classes under test
t.addClass ("Account");
t.addClass ("History");
// Generates a test class TestBank, made of 100 test cases.
// For each test case, the tool tries to generate 50 method calls.
t.generate ("TestBank", 100, 50);
}
}
The main class of the Jartege framework is ClassTester. This class must be instantiated
to allow the creation of test programs.
The method addClass (String className) adds the class className to the set of classes
under test. In this example, we wish to generate tests for the classes Account and History.
The method generate (String className, int numberOfTests, int numberOfMethodCalls)
generates a file className.java which contains a class called className. This class is composed of numberOfTests test cases. For each test case, the tool makes numberOfMethodCalls attempts to generate a method call of one of the classes under test. Using the
accessible constructors, the tool constructs objects which serve as parameters for these
method calls.
When the program TestGen is executed, it produces a file TestGen.java which contains
a main program. This main program calls successively 100 test methods test1 ( ), test2 ( )
... test100 ( ). Each test method contains about 50 method calls.
While the program is generated, Jartege executes on the fly each operation call, which
allows it to eliminate calls which violate the operation precondition. When this precondition is strong, it may happen that the tool does not succeed in generating a call for a
given method, which explains that about 50 method calls are generated.
4.2
Test Programs Produced by Jartege
A test program produced by Jartege is a class with a main method which consists in calling
sequentially all generated test cases. Each test case consists of a sequence of constructor
and method calls of the classes under test. Here is an example of such a test case:
// Test case number 1
public void test1 ( ) throws Exception {
try {
Account ob1 = new Account (1023296578, 223978640);
ob1.debit (152022897);
9
History ob2 = new History (1661966075, (History) null);
History ob3 = new History (-350589348, ob2);
History ob4 = ob2.getPrec ( );
int ob5 = ob3.getBalance ( );
ob1.cancel ( );
// ...
} catch (Throwable except) {
error ( except, 1);
}
}
For each test method, if a JML exception is raised, then an error message (coming from
jmlc) is printed. The test program terminates by printing an assessment of the test. As an
example, here is an excerpt of what is printed with a generated program TestBank.java:
1) Error detected in class TestBank by method test2:
org.jmlspecs.jmlrac.runtime.JMLInvariantError:
By method ”[email protected]:79:18i” of class ”Account”
for assertions specified at Account.java:11:32 [...]
at TestBank.test2(TestBank.java:138)
[...]
Number of tests: 100
Number of errors: 71
Number of inconclusive tests: 0
The program has detected 71 errors. The first error detected comes from a violation of
the invariant of class Account (specified line 11), which happened after a credit operation.
The test program also indicates the number of inconclusive tests. A test case is inconclusive when it does not allow one to conclude whether the program behaviour is correct
or not. A test program generated by Jartege indicates that a test case is inconclusive
when it contains an operation call whose entry precondition is violated. As Jartege is
designed to eliminate such operation calls, this situation may only arise when the code or
the specification of one of the classes under test has been modified after the test file was
generated. A high number of inconclusive tests indicates that the test file is no longer
relevant.
5
Controlling Random Generation
If we leave everything to chance, Jartege might not produce interesting sequences of calls.
Jartege thus provides a few possibilities to parameterize its random aspect. These features
can be useful for stress testing, for instance if we want to test more intensively a given
method. More generally, they allow us to define an operational profile for the classes under
test, which describe how these classes are likely to be used by other components.
5.1
Weights
With each class and operation of a class is associated a weight, which defines the probability
that a class will be chosen, and that an operation of this class will be called. In particular,
it is possible to forbid to call some operation by associating a null weight with it. By
10
default, all weights are equal to 1. A weight can be modified by a weight change method.
In particular:
• changeAllMethodsWeight (String className, double weight) changes the weight of all
methods in the class className to the specified weight.
• changeMethodWeight (String className, String methodName, double weight) changes
the weight of the specified method(s) in the class className to the specified weight.
• changeMethodWeight (String className, String methodName, String [ ] signature, double weight) changes the weight of the method by its name and its signature to the
specified weight.
5.2
Creation of Objects
Objects creation is commanded by creation probability functions, which define the probability of creating a new object according to the number of existing objects of the class
against that of reusing an already created object. If this probability is low, Jartege is more
likely to reuse an already created object than to construct a new one. This allows the user
either to create a predefined number of instances for a given class, or on the opposite, to
create numerous instances for a class.
In the example of bank accounts, it is not very interesting to create many accounts.
It is possible to test the class Account more efficiently for example by creating a unique
account and by applying numerous method calls to it.
The function changeCreationProbability(String className, CreationProbability creationProbabilityFunction) changes the creation probability function associated with the specified
class to the the specified creation probability function.
The interface CreationProbability contains a unique method
double theFunction(int nbCreatedObjects)
which must satisfy the condition
theFunction(0) = 1;
theFunction(n) ∈ [0, 1], ∀n ≥ 1.
The class ThresholdProbability allows one to define threshold probability functions whose
value is 1 under some threshold s and 0 above.
theFunction(n) = 1, if n < s;
theFunction(n) = 0, otherwise.
A threshold probability function with threshold s allows one to define at most s instances
of a given class. We can for instance forbid the creation of more than one instance of
Account by adding the following statement in the test generator:
t.changeCreationProbability ("Account", new ThresholdProbability (1));
11
5.3
Parameter Generation of Primitive Types
When a method has a strong precondition, the probability that Jartege, without any further indication, will generate a call to this method which does not violate this precondition
is low. For primitive types, Jartege provides the possibility to define generators for some
parameters of a given method.
For example, the precondition of the debit operation requires the parameter amount to
be in range
[0, getBalance ( ) − getMin ( )].
Let us suppose the range is small, in other words that the balance is closed to the minimum
balance. If Jartege chooses an amount to be debited entirely randomly, this amount is not
likely to satisfy the method precondition.
Jartege provides a way of generating parameter values of primitive types for operations.
For this, we define a class JRT Account as follows.
import jartege.RandomValue;
public class JRT Account {
private Account theAccount; // The current account
/* Constructor. */
public JRT Account (Account theAccount) { this.theAccount = theAccount; }
/** Generator for the first parameter of operation debit (int). */
public int JRT debit int 1 ( ) {
return RandomValue.intValue (0,
theAccount.getBalance ( ) - theAccount.getMin ( ));
}
}
The class JRT Account must contain a private field of type Account which will contain
the current object on which an operation of class Account is applied. A constructor allows
Jartege to initialize this private field. The class also contains one parameter generation
method for each parameter for which we specify the generation of values. In the example,
to specify the generation of the first parameter of operation debit (int amount), we define
the method int JRT debit int 1 ( ). We use the signature of the operation in the name of
the method to allow overloading. The method RandomValue.intValue (int min, int max)
chooses a random integer in range [min, max].
5.4
Fixtures
If we want to generate several test cases which operate on a particular set of objects, we
can write a test fixture, in a similar way to JUnit. A test fixture is a class which contains:
• attributes corresponding to the objects a test operates on;
• an optional setUp method which defines the preamble of a test case (which typically
constructs these objects);
• an optional tearDown method which defines the postamble of a test case.
12
6
Applying Jartege to the Case Study
The 100 test cases generated by Jartege, showing 71 failures, revealed three different errors:
one error caused by a credit operation, and two errors caused by a cancel operation. We
extracted the shorter sequence of calls which resulted in each failure and obtained the
following results. We also changed the parameter values and added some comments for
more readability.
Error 1. The credit operation can produce a balance inferior to the previous balance
because of an integer overflow.
public void test1 ( ) {
Account ob1 = new Account (250000000, 0);
ob1.credit (2000000000); // Produces a negative balance,
}
// below the minimum balance.
Error 2. The cancel operation can produce an incorrect result if it is preceded by a
setMin operation which changes the minimum balance of the account to a value which is
superior to the balance before cancellation.
public void test11 ( ) {
Account ob1 = new Account (-50, -100);
ob1.credit (100);
ob1.setMin (0);
ob1.cancel ( ); // Restores the balance to a value
}
// inferior to the minimum balance.
Error 3. The third error detected is a combination of an overflow on a debit operation
(similar to Error 1, which comes from an overflow on a credit operation) and of the second
error.
public void test13 ( ) {
Account ob1 = new Account (-1500000000, -2000000000);
ob1.debit (800000000); // Produces a positive balance.
ob1.setMin (0);
ob1.cancel ( ); // Restores the balance to a value
}
// inferior to the minimum balance.
We have the feeling that the three errors detected with test cases generated by Jartege
are not totally obvious, and could have easily been forgotten in a manually developed test
suite. Errors 2 and 3 in particular require three method calls to be executed in a specific
order and with particular parameter values.
It must be noted that the case study was originally written to show the use of JML to
undergraduate students, without us being aware of the faults.
7
Comparison with Related Work
Our work has been widely inspired by the JML-JUnit approach [CL02b]. The JML-JUnit
tool generates test cases for a method which consist of a combination of calls of this method
13
with various parameter values. The tester must supply the object invoking the method
and the parameter values. With this approach, interesting values could easily be forgotten
by the tester. Moreover, as a test case only consists of one method call, it is not possible
to detect errors which result of several calls of different methods. At last, the JML-JUnit
approach compels the user to construct the test data, which may require the call of several
constructors. Our approach thus has the advantage of being more automatic, and of being
able to detect more potential errors.
Korat [BKM02] is a tool also based on the JML-JUnit approach, which allows exhaustive testing of a method for all objects of a bounded size. The tools automatically
construct all non isomorphic test cases and execute the method on each test case. Korat
therefore has the advantage over JML-JUnit of being able to construct the objects which
invoke the method under test. However, test cases constructed by Korat only consist of
one object construction and one method invocation on this object.
Tobias [Led02, MLBdB02] is a combinatorial testing tool for automatic generation of
test cases derived from a “test pattern”, which abstractly describes a test case. Tobias was
first designed to produce test objectives for the TGV tool [JM99] and was then adapted
to produce test cases for Java programs specified in JML and for programs specified in
VDM. The main problem of Tobias is the combinatorial explosion which happens if one
tries to generate test cases which consist of more than a couple of method calls. Jartege
was designed to allow the generation of long test sequences without facing the problem of
combinatorial explosion.
8
Discussion and Future Work
Jartege is only in its infancy and a lot of work remains to be done.
Primitive values generation for methods parameters is currently done manually by
writing primitive parameters generating methods. Code for these methods could be automatically constructed from the JML precondition of the method. This could consist in
extracting range constraints from the method precondition and automatically produce a
method which could generate meaningful values for the primitive parameters.
Jartege easily constructs test cases which consist of hundreds of constructors and methods calls. It would be useful to develop a tool for extracting a minimum sequence of calls
which results in a given failure.
We developed Jartege in Java, and we specified some of its classes with JML. We
applied Jartege to these classes to produce test cases, which allowed us to experiment
our tool on a larger case study and to detect a few errors. We found much easier to
produce tests with Jartege than to write unit tests with JUnit or JML-JUnit. We intend
to continue our work of specifying Jartege in JML and testing its classes with itself. We
hope that this real case study will help us to evaluate the effectiveness and scalability of
the approach.
A comparison of our work with other testing strategies still remains to be done. We can
expect systematic methods, using for instance boundary testing such as BZTT [LPU02],
to be able to produce more interesting test cases that ours. Our goal is certainly not to
pretend that tests produced randomly can replace tests produced by more sophisticated
methods, nor a carefully designed test set written by an experienced tester.
Our first goal in developing Jartege was to help the developer to write unit tests for
unstable Java classes, thus for “debug unit testing”. It would also be interesting to use
Jartege to evaluate the reliability of a stable component before it is released. Jartege
14
provides some features to define an operational profile of a component, which should allow
statistical testing. However, the definition of a correct operational profile, especially in
the context of object-oriented programming, is a difficult task. Moreover, the relation
between test sets generated by Jartege and the reliability of a component requires more
theoretical work, one difficult point being to take into account the state of the component.
9
Conclusion
This report presents Jartege, a tool for random generation of unit tests for Java classes
specified in JML. The aim of the tool is to easily produce numerous test cases, in order to
detect a substantial number of errors without too much effort. It is designed to produce
automated tests, which can in part replace tests written by the developer using for instance JUnit. We think that the automatic generation of such unit tests should facilitate
continuous testing as well as code refactoring in the context of extreme programming.
The JML specifications are used on the one hand to eliminate irrelevant test cases,
and on the other hand as a test oracle. We think that the additional cost of specification
writing should be compensated by the automatic oracle provided by the JML compiler, as
long as we wish to intensively test the classes. Moreover, this approach has the advantage
of supporting the debugging of a specification along with the corresponding program.
This allows the developer to increase his confidence in the specification and to use this
specification in other tools.
Most test generation methods are deterministic, while our approach is statistical. We
do not wish to oppose both approaches, thinking that they both have their advantages
and drawbacks, and that a combination of both could be fruitful.
At last, we found JML to be good language to start learning formal methods. Its
Java-based syntax makes it easy to learn for Java programmers. As JML specifications
are included in Java source code as comments, it is easy to develop and debug a Java
program along with its specification. Moreover, automatic test oracles as well as automatic
generation of test cases are good reasons of using specification languages such as JML.
References
[ABM97]
Lionel Van Aertryck, Marc Benveniste, and Daniel Le Métayer. Casting: a formally based software test generation method. In Proceedings of
the First IEEE Internatinal Conference on Formal Engineering Methods —
ICFEM’97, Hiroshima, Japan, pages 101–111, November 1997.
[BCC+ 03]
Lilian Burdy, Yoonsik Cheon, David R. Cok, Michael Ernst, Joe Kiniry,
Gary T. Leavens, K. Rustan, M. Leino, and Erik Poll. An overview of JML
tools and applications. Technical Report NIII-R0309, Department of Computer Science, University of Nijmegen, March 2003.
[Bec99]
Kent Beck. Embracing change with extreme programming. IEEE Computer,
32(10):70–77, October 1999.
[Bec00]
Kent Beck. Extreme Programming Explained. Addison Wesley, 2000.
[BG98]
Kent Beck and Erich Gamma. Test infected: Programmers love writing tests.
Java Reports, 3(7):51–56, 1998.
15
[BKM02]
Chandrasekhar Boyapati, Safraz Khurshid, and Darko Marinov. Korat: Automated testing based on Java predicates. In Proceedings of the International
Symposium on Software Testing and Analysis — ISSTA’02, Rome, pages
123–133, July 2002.
[CL02a]
Yoonsik Cheon and Gary T. Leavens. A runtime assertion checker for the Java
Modeling Language (JML). In Hamid R. Arabnia and Youngsong Mun (eds.),
International Conference on Software Engineering Research and Practice —
SERP’02, Las Vegas, Nevada, pages 322–328. CSREA Press, 2002.
[CL02b]
Yoonsik Cheon and Gary T. Leavens. A simple and practical approach to unit
testing: The JML and JUnit way. In Boris Magnusson (ed.), 16th European
Conference on Object-Oriented Programming — ECOOP’02, Malaga, Spain,
number 2374 in Lecture Notes in Computer Science, pages 231–255. SpringerVerlag, June 2002.
[dBORZ99] Lydie du Bousquet, Farid Ouabdesselam, Jean-Luc Richier, and Nicolas
Zuanon. Lutess: a specification-driven testing environment for synchronous
software. In International Conference on Software Engineering — ICSE’99,
Los Angeles, USA. ACM Press, May 1999.
[dBZ99]
Lydie du Bousquet and Nicolas Zuanon. An overview of Lutess: a
specification-based tool for testing synchronous software. In 14th IEEE International Conference on Automated Software Engineering — ASE’99, pages
208–215, October 1999.
[DF93]
Jeremy Dick and Alain Faivre. Automating the generation and sequencing
of test cases from model-based specifications. In Proceedings of FME’93,
number 670 in LNCS, pages 268–284. Springer-Verlag, April 1993.
[DN84]
Joe W. Duran and Simeon C. Ntafos. An evaluation of random testing. IEEE
Transactions on Software Engineering, 10(4):438–444, 1984.
[FHLS98]
Phyllis G. Frankl, Richard G. Hamlet, Bev LittleWood, and Lorenzo Strigini.
Evaluating testing methods by delivered reliability. IEEE Transactions on
Software Engineering, 24(8):586–601, August 1998.
[Gau95]
Marie-Claude Gaudel. Testing can be formal too. In Proceedings of TAPSOFT’95, Aarhus, Denmark, number 915 in LNCS, pages 82–96. SpringerVerlag, May 1995.
[GMH81]
John Gannon, Paul McMullin, and Richard Hamlet. Data-abstraction implementation, specification, and testing. ACM Transactions on Programming
Languages and Systems, 3(3):211–233, July 1981.
[Ham94]
Richard Hamlet. Random testing. In J. Marciniak, editor, Encyclopedia of
Software Engineering, pages 970–978. Wiley, 1994.
[HT90]
Dick Hamlet and Ross Taylor. Partition testing does not inspire confidence.
IEEE Transactions on Software Engineering, 16(12), December 1990.
16
[JM99]
Thierry Jéron and Pierre Morel. Test generation derived from modelchecking. In Proceedings of the 11th International Conference on Computer
Aided Verification — CAV’99, Trento, Italy, number 1633 in LNCS, pages
108–121. Springer-Verlag, July 1999.
[JML]
The Java Modeling Language (JML) Home Page,
http://www.cs.iastate.edu/˜leavens/JML.
[JUn]
The JUnit Home Page, http://www.junit.org.
[LBR03]
Gary T. Leavens, Albert L. Baker, and Clyde Ruby. Preliminary Design of
JML: A Behavioral Interface Specification Language for Java. Department
of Computer Science, Iowa State University, 1998–2003.
[Led02]
Yves Ledru. The TOBIAS test generator and its adaptation to some ASE
challenges (position paper). In Workshop on the State of the Art in Automated Software Engineering, ICS Technical Report UCI-ICS-02-17, University of California, Irvine, USA, 2002.
[LPC+ 03]
Gary T. Leavens, Erik Poll, Curtis Clifton, Yoonsik Cheon, and Clyde Ruby.
JML Reference Manual. (Draft), April 2003.
[LPU02]
Bruno Legeard, Fabien Peureux, and Mark Utting. Automated boundary
testing from Z and B. In Proceedings of FME’02, Formal Methods Europe, Copenhaguen, Denmark, number 2391 in LNCS, pages 21–40. SpringerVerlag, July 2002.
[LTFWD00] Yvan Labiche, Pascale Thévenod-Fosse, Hélène Waeselynck, and M.-H. Durand. Testing levels for object-oriented software. In Proceedings of the 22nd
International Conference on Software Engineering — ICSE’00, Limerick,
Ireland, pages 136–145. ACM, June 2000.
[Mey88]
Bertrand Meyer. Object-Oriented Software Construction. Prentice Hall, 1988.
[Mey92]
Bertrand Meyer. Applying design by contract. IEEE Computer, 25(10):40–
51, October 1992.
[MLBdB02] Olivier Maury, Yves Ledru, Pierre Bontron, and Lydie du Bousquet. Using
TOBIAS for the automatic generation of VDM test cases. In Third VDM
Workshop (at FME’02), Copenhaguen, Denmark, July 2002.
[Mye94]
G. J. Myers. The Art of Software Testing. John Wiley and Sons, New York,
1994.
[Nta01]
Simeon C. Ntafos. On comparisons of random, partition, and proportional
partition testing. IEEE Transactions on Software Engineering, 27(10):949–
960, July 2001.
[Par96]
Ioannis Parissis. Test de logiciels synchrones spécifiés en Lustre. PhD thesis,
Grenoble, France, Septembre 1996.
[Ros92]
David S. Rosenblum. Towards a method of programming with assertions. In
International Conference on Software Engineering — ICSE’92. IEEE Computer Society Press, 1992.
17
[WJ91]
Elaine J. Weyuker and Bingchiang Jeng. Analyzing partition testing strategies. IEEE Transactions on Software Engineering, 17(7):703–711, July 1991.
18
| 2cs.AI
|
1
Runlength-Limited Sequences and
Shift-Correcting Codes
arXiv:1803.06117v1 [cs.IT] 16 Mar 2018
Mladen Kovačević
Abstract—This work is motivated by the problem of error
correction in bit-shift channels with the so-called (d, k) input
constraints (where successive 1’s are required to be separated
by at least d and at most k zeros, 0 ≤ d < k ≤ ∞). Bounds
on the size of optimal (d, k)-constrained codes correcting a
fixed number of bit-shifts are derived. The upper bound is
obtained by a packing argument, while the lower bound follows
from a construction based on a family of integer lattices.
Several properties of (d, k)-constrained sequences that may be
of independent interest are established as well; in particular,
the capacity of the noiseless channel with (d, k)-constrained
constant-weight inputs is characterized. The results are relevant
for magnetic and optical storage systems, reader-to-tag RFID
channels, and other communication models where bit-shift errors
are dominant and where (d, k)-constrained sequences are used
for modulation.
Index Terms—Bit-shift channel, peak shift, timing errors,
runlength-limited code, integer compositions, Manhattan metric,
asymmetric distance, magnetic recording, inductive coupling.
I. I NTRODUCTION
HIFT AND TIMING errors are a dominant type of
noise in several communication and information storage
scenarios, examples of which include magnetic and optical
recording devices [11], [30], inductively coupled channels
such as the reader-to-tag RFID channel [28], parallel asynchronous communications [7], various types of timing channels [2], [17], [25], etc. Designing codes that are able to
correct these types of errors and studying their fundamental
limits is therefore important for all these applications, in
addition to being an interesting theoretical challenge. The
problem is further complicated by the fact that, in many of
the mentioned applications, particularly magnetic, optical, and
emerging DNA storage systems, the codewords are required to
satisfy modulation constraints that are introduced to alleviate
inter-symbol interference and other impairing effects [12],
[13], [14], [31], [32]. Perhaps the best-known example of such
constraints are runlength constraints where a minimum and a
maximum number of zero symbols between two consecutive
non-zero symbols is specified. Motivated by these communication settings, we study in the present paper the errorcorrection problem for channels with shift and timing errors,
and with runlength input constraints. The precise channel
model we have in mind and our contributions are described in
the following two subsections.
S
Date: March 16, 2018.
The author is with the Department of Electrical & Computer Engineering, National University of Singapore, Singapore 117583 (emails:
[email protected], [email protected]).
This work was supported by the Singapore Ministry of Education (MoE)
Tier 2 grant “Network Communication with Synchronization Errors: Fundamental Limits and Codes” (Grant number R-263-000-B61-112).
A. The Channel Model
Fix d, k ∈ Z ∪ {∞} with 0 ≤ d < k ≤ ∞. We assume
that the channel inputs are binary strings of length n that are
composed of blocks from the set {0d 1, . . . , 0k 1}, where 0j is
a string of j zeros. In other words, the set of inputs is
n
∗
Sd,k (n) := 0, 1 ∩ 0d 1, . . . , 0k 1 .
(1)
S
∞
(Here A∗ is the usual notation for the set i=0 Ai .)P
For a string
n
x = x1 · · · xn ∈ {0, 1}n, denote by wt(x) =
i=1 xi its
Hamming weight. It will be convenient to introduce a special
notation for the set of all channel inputs of a given weight W :
n
o
Sd,k (n, W ) := x ∈ Sd,k (n) : wt(x) = W .
(2)
In other words, Sd,k (n, W ) contains all those input strings
that are composed of exactly W blocks from the set
{0d 1, . . . , 0k 1}.
The definition (1) ensures that (a) consecutive 1’s in any
input string are separated by at least d and at most k zeros,
and (b) every input string starts with a string of zeros of
length j, d ≤ j ≤ k, and ends with a 1. The property (a)
is the defining property of the so-called (d, k)-constrained
sequences. However, the boundary conditions (b) are not
universally adopted in the literature [12]. We shall nevertheless
find it convenient to work under definition (1), as in, e.g., [22];
it will be evident later on that different boundary conditions
would not affect the analysis in any significant way.
For a given input string x ∈ Sd,k (n, W ), the channel
outputs another binary string y of length n and weight W .
We think of 1’s in x as being “shifted” in the channel, each
for a number of positions to the left or to the right of its
original position, thus producing y at the output.
PW We say that
t bit-shifts have occurred in the channel if i=1 |x̄i − ȳi | = t,
where x̄i (resp. ȳi ) is the position of the i’th 1 in x (resp. y),
1 ≤ i ≤ W.
Example 1. Consider an input string x ∈ S2,4 (15, 4) and the
corresponding output string y:
x = 001000010010001
y = 0 1 0 0 0 1 0 0 0 0 0 1 0 0 1.
(3)
We think of y as obtained from x by shifting the first 1 in
x one position to the left, the second 1 two positions to the
left, and the third 1 one position to the right. We then say that
the total number of bit-shifts that occurred in the channel is
t = 1 + 2 + 1 = 4.
N
We emphasize once more that the output string y is always
assumed to be of the same length and the same weight as the
corresponding input string x. Note, however, that y may in
general violate the (d, k)-constraints.
2
B. Main Results
Our main object of study in the present paper are errorcorrecting codes for the above-described channel model. In
particular, we shall derive explicit bounds on the cardinality of
optimal (d, k)-constrained shift-correcting codes, with a focus
on their asymptotic behavior in the regime of growing blocklength (n → ∞). Despite a sizable body of literature on the
bit-shift channel and related models1 , such bounds, to the best
of our knowledge, have not been obtained before, even for a
single bit-shift.
We consider two scenarios. The first one, analyzed in
Section III, corresponds to the situation where shifts to the
right (right-shifts) and shifts to the left (left-shifts) are treated
independently, and separate requirements on their correctability are imposed. More precisely, codes are in this case required
to have the capability of correcting t→ right-shifts and t← leftshifts, for fixed (but arbitrary) t→ and t← . The second scenario,
analyzed in Section IV, corresponds to the situation where
right-shifts and left-shifts are treated in a symmetric way, i.e.,
where codes are required to have the capability of correcting
t shifts, regardless of the direction of each individual shift.
In Section II we also demonstrate several properties of the
code space Sd,k (n). In particular, we characterize the capacity
of the noiseless channel with (d, k)-constrained constantweight inputs. These results are needed to prove the bounds on
the cardinality of optimal shift-correcting codes in Sections III
and IV, but are also of possible interest in other applications
of (d, k)-constrained sequences and in other channel models.
C. Notation
Z denotes the integers and Z+ the non-negative integers.
log is the base-2
P logarithm.
If we write i si , it is understood that i ranges over all
possible values, which will be clear from the context.
Given two non-negative real sequences (an ) and (bn ), we
adopt the following asymptotic notation:
a
• an ∼ bn means limn→∞ b n = 1;
n
a
• an & bn means lim inf n→∞ b n ≥ 1;
n
• an ∼
˙ bn means log an ∼ log bn (i.e., the exponents of an
and bn have the same asymptotic behavior);
a
• an = O(bn ) means lim supn→∞ b n < ∞;
n
an
• an = o(bn ) means limn→∞ b = 0.
n
For x ∈ Sd,k (n, W ), we denote by x̄ = (x̄1 , . . . , x̄W ) the
vector indicating the positions of 1’s in the string x, meaning
that x̄i is the position of the i’th 1 in x. For example, for
x = 0100101 ∈ S1,3 (7, 3) we have x̄ = (2, 5, 7). The
mapping x 7→ x̄ is clearly injective. With this correspondence
in mind, we define
n
S̄d,k (n, W ) := x̄ ∈ ZW : 0 < x̄1 < · · · < x̄W = n,
o
d + 1 ≤ x̄i − x̄i−1 ≤ k + 1 for i = 1, . . . , W , (4)
where x̄0 = 0 by convention. Hence, S̄d,k (n, W ) ⊂ ZW is
just a different representation of the set of all channel inputs
of
S length n and weight W , namely Sd,k (n, W ), and S̄d,k (n) :=
W S̄d,k (n, W ) is the corresponding representation of the set
of all channel inputs of length n, namely Sd,k (n).
The space S̄d,k (n), or S̄d,k (n, W ) in the constant-weight
case, seems to be more convenient for describing constructions
of codes for the bit-shift channel.
B. Combinatorial Description and Asymptotics
Denote
= |S¯d,k (n)|,
Sd,k (n, W ) := |Sd,k (n, W )| = |S¯d,k (n, W )|.
Sd,k (n)
:= |Sd,k (n)|
(5a)
(5b)
We see from the definition of the code space (1) that Sd,k (n)
is in fact the number of compositions2 of the number n with
parts restricted to the set {d + 1, . . . , k + 1}. This number can
be expressed in the recursive form
Sd,k (n) =
k+1
X
Sd,k (n − i)
(6)
i=d+1
with initial conditions Sd,k (0) = 1 and Sd,k (n) = 0 for n < 0.
It follows that Sd,k (n) ∼ c · 2−n log ρ [35], where ρ is the
Pk+1
unique positive solution to i=d+1 xi = 1 and the constant c
can be obtained from (6).
Similarly, Sd,k (n, W ) is the number of compositions of the
number n with exactly W parts, each part belonging to the
set {d + 1, . . . , k + 1}. Again, this quantity can be expressed
in the recursive form [10]
Sd,k (n, W ) =
k+1
X
Sd,k (n − i, W − 1)
(7)
i=d+1
II. T HE C ODE S PACE
In this section we demonstrate some properties of the set
Sd,k (n) that will be used in the derivations to follow, but are
also of potential interest in other applications. We also describe
another representation of this space that is of course equivalent
to the one given in (1)–(2), but that may be preferable to it,
depending on the problem being analyzed.
A. Equivalent Representation
Another representation of the set of all channel inputs that
is useful for analyzing bit-shift errors is based on specifying
the positions of 1’s in the input string.
1 See,
e.g., [1], [8], [11], [16], [23], [24], [28], [30], [36].
with initial conditions Sd,k (n, 0) = 1{n = 0} and
Sd,k (n, W ) = 0 for n < 0. In the following lemma we
characterize the asymptotic behavior
1 of1 Sd,k (n, W ) as n → ∞
, d+1 . (For simplicity, we
and W = wn, for fixed w ∈ k+1
write wn instead of, e.g., ⌊wn⌋, ignoring the fact that the
former is not necessarily an integer.)
1
1
, d+1
define the function
Lemma 1. For w ∈ k+1
σd,k (w) := w log
k+1
X
1
i− w
ρw
,
(8)
i=d+1
2 A composition of an integer n is a tuple of positive integers (called parts)
summing to n [34]. For a study of compositions with parts restricted to a
subset of Z+ see, e.g., [10].
3
where
positive real root of the polynomial
Pk+1 ρw is the unique
1
1
i
let
σd,k k+1
= σd,k d+1
= 0.
i=d+1 (wi−1)x , and
1 also
1
For any fixed w ∈ k+1 , d+1
, as n → ∞ we have
Sd,k (n, wn) ∼
˙ 2
nσd,k (w)
.
(9)
The exponent σd,k (w) is a strictly concave function of w.
It attains its maximal value at
!−1
k+1
X
w∗ =
iρi
(10)
i=d+1
and this value is σd,k (w∗ ) = − log ρ, where ρ is the unique
Pk+1
positive solution to i=d+1 xi = 1.
Proof: The asymptotics of Sd,k (n, wn) as n → ∞
is most easily determined using the machinery of analytic
combinatorics [27]. The generating function of the bivariate
sequence (Sd,k (n, W ))n,W is obtained from (7) as:
Fd,k (x, y) :=
∞ X
∞
X
Sd,k (n, W )xn y W
n=0 W =0
1
,
=
1 − y(xd+1 + · · · + xk+1 )
(11)
wherefrom one verifies that this sequence is a Riordan array3
and satisfies
of [27, Thm 12.2.2]. Denoting
P the conditions
i
v(x) = k+1
x
(the
polynomial
appearing in the denomi=d+1
inator of the generating function (11)), we conclude from
[27, Thm 12.2.2] that
!wn
k+1
X i− 1
wn −n
w
, (12)
Sd,k (n, wn) ∼
˙ v(ρw ) ρw =
ρw
i=d+1
x
∂
v(x)
= w1 . This
where ρw is the unique positive solution to ∂x
v(x)
proves (9).
The second part of the claim is obtained by carefully
analyzing the involved functions. The root ρw is a function
weight w and is implicitly defined by
Pk+1of the relative
i
(wi
−
1)ρ
=
0, ρw > 0. Differentiating this equation
w
i=d+1
∂
∂
ρw < 0. Also, we have ∂w
σd,k (w) =
w.r.t. w we find that ∂w
∂
Pk+1
2
∂
i
∂w ρw
log i=d+1 ρw and (∂w)2 σd,k (w) = log e · wρw < 0, implying that σd,k (w) is concave. The weight w∗ that maximizes
∂
σd,k (w) = 0,
the exponent σd,k (w) is the one for which ∂w
Pk+1
i
i.e., v(ρw ) = i=d+1 ρw = 1.
The quantity σd,k (w) defined in (8) is the maximal information rate (i.e., the capacity) that can be achieved in the
noiseless channel with (d, k)-constrained inputs of relative
weight w. This is a refinement of the well-known result
that states that the capacity of the noiseless channel with
(d, k)-constrained inputs (but with no weight constraints) is
limn→∞ n1 log Sd,k (n) = − log ρ [35]. Namely, the latter
statement can be recovered from Lemma 1 by noting that
∗
Sd,k (n) ∼
˙ Sd,k (n, w∗ n) ∼
˙ 2nσd,k (w ) = 2−n log ρ . The following lemma is a strengthening of this result; it asserts that the
input strings of weight (approximately) equal to w∗ n account
3 Bivariate
sequences with generating functions of the form
called (generalized) Riordan arrays; see [27, Sec. 12.2].
φ(x)
1−yv(x)
are
for most of the space Sd,k (n). Informally, we say that the
“typical” input strings have relative weight w∗ .
Lemma 2. There exists a sublinear function f (n) = o(n)
such that, as n → ∞,
w ∗ n+f (n)
Sd,k (n) ∼
X
Sd,k (n, W ).
(13)
W =w ∗ n−f (n)
Proof: It follows from Lemma 1 (in particular, the fact
that σd,k (w) is uniquely maximized at w∗ ) that the contribution of the termsP
W ≤ (w∗ − ǫ)n and W ≥ (w∗ + ǫ)n in the
sum Sd,k (n) = W Sd,k (n, W ) is asymptotically negligible
for any fixed ǫ > 0.
We shall also need in the sequel an estimate of the number
of blocks 0j 1, for fixed j ∈ {d, . . . , k}, in typical input
strings. For the purpose of formally stating this result, denote
(j)
by Sd,k (n, W, ℓ) the number of input strings consisting of
W blocks from {0d1, . . . , 0k 1}, exactly ℓ of which are 0j 1.
(j)
Equivalently, Sd,k (n, W, ℓ) is the number of compositions
of the number n with W parts, the parts taking values in
{d + 1, . . . , k + 1}, and exactly ℓ of them having value j + 1.
PW (j)
We then have Sd,k (n, W ) =
ℓ=0 Sd,k (n, W, ℓ), for every
j ∈ {d, . . . , k}.
Lemma 3. Fix j ∈ {d, . . . , k} and denote λ∗j = ρj+1 w∗ =
−1
Pk+1
iρi
, where ρ is the unique positive solution
ρj+1
Pk+1 i=d+1
to i=d+1 xi = 1. Then, for every j ∈ {d, . . . , k}, as n → ∞
we have
(j)
(14)
Sd,k (n) ∼
˙ Sd,k n, w∗ n, λ∗j n .
Moreover, there exists a sublinear function f (n) = o(n) such
that, for every j ∈ {d, . . . , k}, as n → ∞ we have
w ∗ n+f (n)
λ∗
j n+f (n)
X
X
Sd,k (n) ∼
W =w ∗ n−f (n)
(j)
Sd,k (n, W, ℓ).
(15)
ℓ=λ∗
j n−f (n)
Proof: Let us take j = d and denote λ := λd for
simplicity; the proof for general j is analogous. The following
relation is valid:
W
(d)
Sd,k (n, W, ℓ) =
(16)
Sd+1,k n − ℓ(d + 1), W − ℓ
ℓ
(the ℓ partsof value d + 1 can be distributed among the W
parts in Wℓ ways, and the remaining W − ℓ parts, which are
all from {d + 2, . . . , k + 1}, form a composition of the number
n − ℓ(d + 1)), and therefore
X X W
Sd,k (n) =
Sd+1,k n − ℓ(d + 1), W − ℓ . (17)
ℓ
W
ℓ
Since the sum in (17) has polynomially many terms, we know
that it grows exponentially with the same exponent as one of
1
1
, d+1
its summands W = wn, ℓ = λn, for some w ∈ k+1
4
and λ ∈ (0, w). By using Stirling’s approximation and (9),
this exponent can be expressed in the form
#
"
wn
1
log
Sd+1,k n − λn(d + 1), (w − λ)n
n
λn
(18a)
k+1
λ
X
i− 1−λ(d+1)
w−λ
,
ρw,λ
+ (w − λ) log
∼ wH
w
i=d+2
where H(·) is the binary entropy function and ρw,λ > 0 is a
function implicitly defined by
k+1
X
1 − λ(d + 1)
i−
ρiw,λ = 0.
(18b)
w−λ
i=d+2
By calculating the derivatives of the exponent (the function
on the right-hand side of (18a)) with respect to w and λ, one
finds that it is uniquely maximized for λ∗ = ρd+1 w∗ , w∗ =
Pk+1
P
i −1
i
, where k+1
i=d+1 iρ
i=d+1 ρ = 1. This implies that
(d)
Sd,k (n) ∼
˙ Sd,k n, w∗ n, λ∗ n
(19)
and proves (14). It also implies that, for every ǫ > 0, the part
of the sum in (17) where W ≤ (w∗ − ǫ)n or W ≥ (w∗ + ǫ)n
or ℓ ≤ (λ∗ − ǫ)n or ℓ ≥ (λ∗ + ǫ)n is asymptotically negligible
compared to the remaining part (exponentially smaller in fact).
This proves (15).
In words, input strings of length n typically contain ∼ λ∗j n
4
blocks 0j 1, i.e., runs of zeros
P
Pk of ∗length ∗j. Notice that, since
k+1
i
i=d+1 ρ = 1, we have
j=d λj = w , which is expected
because w∗ n is the total (typical) number of blocks in the
input strings of length n.
In the derivations of the asymptotic bounds in the following
two sections, we shall restrict our analysis to input strings
of weight ∼ w∗ n having ∼ λ∗j n runs of zeros of length j.
Lemmas 2 and 3 will enable us to show that this restriction
incurs no loss in generality.
III. C ODES C ORRECTING A SYMMETRIC S HIFTS
We now turn to the analysis of the bit-shift channel with
(d, k) input constraints. The scenario we consider in this
section is the channel in which only shifts to the right are
possible. We are interested in codes enabling the receiver to
reconstruct the transmitted string whenever the total shift of
its 1’s does not exceed a specified threshold. In particular,
we derive bounds on the cardinality of optimal codes for this
setting. We shall not make an attempt to optimize the bounds
for every block-length n; rather, the focus is put on their
asymptotic behavior as n → ∞.
A. Geometric Characterization
Suppose that x̄ ∈ S̄d,k (n, W ) is the transmitted vector and
z the corresponding received vector (see Section II-A). If the
i’th 1 in x is shifted to the right by ki positions in the channel,
then zi = x̄i + ki . Therefore, the bit-shift channel can be seen
4 A run of zeros is a block of contiguous zeros of maximal length, i.e., such
that it is delimited on both sides either by a 1, or by the end of the string.
as an additive noise channel with input alphabet
PW Z+ . If the 1’s
in x are shifted by t positions in total, then i=1 (zi − x̄i ) = t.
We say that a code C¯ ⊆ S̄d,k (n) corrects t right-shifts if no
two different codewords x̄, ȳ ∈ C¯ can produce the same output
after being impaired with arbitrary patterns of ≤ t right-shifts.
¯ x̄ 6= ȳ, wt(x) = wt(y) = W ,
In symbols, for every x̄, ȳ ∈ C,
PW
and all noise vectors f , g ∈ ZW with fi , gi ≥ 0, i=1 fi ≤ t,
PW
i=1 gi ≤ t, we have x̄ + f 6= ȳ + g. Such a code C is said
to be optimal if there is no other code C¯′ ⊆ S̄d,k (n) correcting
¯
t right-shifts and such that |C¯′ | > |C|.
Remark 4 (Shifts vs. Right-Shifts). It is straightforward to
show that a code C¯ can correct t right-shifts if and only if it can
correct t→ right-shifts and t← left-shifts, for any t→ , t← ≥ 0
with t→ +t← = t. In other words, C¯ can correct t right-shifts if
¯ x̄ 6= ȳ, wt(x) = wt(y) = W ,
and only if for every x̄, ȳ ∈ C,
P
and
all
noise
vectors
f
,
g
∈
ZW suchP
that i:fi >0 fi ≤ t→ ,
P
P
i:gi <0 |gi | ≤ t← , we
i:gi >0 gi ≤ t→ ,
i:fi <0 |fi | ≤ t← ,
have x̄ + f 6= ȳ + g.
Therefore, the results of this section apply also to channels
where both left- and right-shifts are allowed. The difference
with respect to the model that we shall analyze in Section IV is
that these two types of errors are here treated independently of
each other, i.e., separate requirements on their “correctability”
are imposed (t← and t→ ).
N
Consider the following metric on S̄d,k (n, W ):
W
W
X
X
(ȳi − x̄i ) .
(x̄i − ȳi ),
da (x̄, ȳ) := max
i=1
i=1
x̄i >ȳi
(20)
x̄i <ȳi
This distance is of importance in the theory of codes for
asymmetric channels [15] (hence the subscript ‘a’). For vectors
of different dimensions (corresponding to strings of different
weights), x̄ ∈ S̄d,k (n, W1 ), ȳ ∈ S̄d,k (n, W2 ), W1 6= W2 ,
we define da (x̄, ȳ) = ∞. The minimum distance of a code
¯
C¯ ⊆ S̄d,k (n) with respect to the metric da is denoted da (C).
The following proposition gives a metric characterization of
right-shift-correcting codes.
Proposition 5. A code C¯ ⊆ S̄d,k (n) can correct t right-shifts
¯ > t.
if and only if da (C)
Proof. Suppose that da (x̄, ȳ) ≤ t for two distinct codewords
¯ Define f = max{ȳ− x̄, 0} and g = max{x̄− ȳ, 0},
x̄, ȳ ∈ C.
where the maximum is taken
PW coordinate-wise.
PW Then x̄ + f =
ȳ + g and fi , gi ≥ 0, i=1 fi ≤ t, i=1 gi ≤ t (the last
two inequalities together are equivalent to our assumption that
da (x̄, ȳ) ≤ t). This means that C¯ cannot correct t right-shifts.
The other direction is similar. If C¯ cannot correct t rightshifts, then there exist two distinct P
codewords x̄, P
ȳ and two
W
noise vectors f , g with fi , gi ≥ 0, W
i=1 fi ≤ t,
i=1 gi ≤
t, such
that
x̄
+
f
=
ȳ
+
g.
This
implies
that
d
(x̄,
ȳ) ≤
a
PW
PW
¯ ≤ t.
≤
t,
and
so
d
(
C)
g
f
,
max
a
i
i
i=1
i=1
Note that balls in the metric space S̄d,k (n, W ), da do not
have uniform sizes—the size of a ball of radius r depends
on its center. For that reason,
when studying the properties
of codes in S̄d,k (n, W ), da it is sometimes more convenient
5
ሺͳǡʹሻ
L, and represents the average number of lattice points (from
L) per one point of the ambient space (ZW ). The quantity of
interest to us is the present context is the maximum density
a lattice L ⊆ ZW can have when its minimum distance is
required to satisfy da (L) > t, namely
ሺͳǡ͵ሻ ሺʹǡ͵ሻ
ሺͳǡͶሻ ሺʹǡͶሻ ሺ͵ǡͶሻ
n
o
µa (W, t) := max µ(L) : L ⊆ ZW a lattice, da (L) > t .
(22)
ሺͳǡͷሻ ሺʹǡͷሻ ሺ͵ǡͷሻ ሺͶǡͷሻ
Lemma 6. For every fixed t ∈ Z+ , as W → ∞ we have
µa (W, t) & W −t .
ሺͳǡሻ ሺʹǡሻ ሺ͵ǡሻ ሺͶǡሻ ሺͷǡሻ
ሺͳǡሻ ሺʹǡሻ ሺ͵ǡሻ ሺͶǡሻ ሺͷǡሻ ሺǡሻ
ሺͳǡͻሻ
ሺʹǡͺሻ ሺ͵ǡͺሻ ሺͶǡͺሻ ሺͷǡͺሻ ሺǡͺሻ ሺǡͺሻ
ሺ͵ǡͻሻ ሺͶǡͻሻ ሺͷǡͻሻ ሺǡͻሻ ሺǡͻሻ ሺͺǡͻሻ
Fig. 1: The space S̄0,5 (9, 2) representing the set of all binary
strings of length n = 9 and weight W = 2 satisfying
(0, 5)-constraints, and a code correcting t = 2 right-shifts.
Codewords are depicted as black dots; gray regions illustrate
balls (under the metric da ) of radius ⌊t/2⌋ = 1 around the
codewords.
W
to consider the unrestricted metric space Z , da where this
ሺͳǡʹሻ
effect does not occur5 . For example, we have the following
expression for the cardinality of a ball of radius r in ZW , da
[21,
Lem.ሺʹǡ͵ሻ
1]:
ሺͳǡ͵ሻ
X W r r + W − i
.
(21)
Ba (W, r) =
i
i
W −i
i
ሺͳǡͶሻ ሺʹǡͶሻ ሺ͵ǡͶሻ
The cardinality of an arbitrary ball of radius r in
S̄d,k (n, W ), da is upper bounded by Ba (W, r).
ሺͳǡͷሻ ሺʹǡͷሻ ሺ͵ǡͷሻ ሺͶǡͷሻ
B. Construction and Bounds
ሺ͵ǡሻ
ሺͳǡሻ
Denoteሺʹǡሻ
by Md,k
(n; t) ሺͶǡሻ
(resp. Mሺͷǡሻ
d,k (n, W ; t)) the cardinality
→
Proof: It was shown in [21, Thm 7] that every sublattice
L ⊆ ZW with da (L) > t corresponds to a Sidon set6 of order
t and cardinality W + 1 in an Abelian group, and vice versa.
Consequently, the largest possible density of such a sublattice
1
can be expressed as µa (W, t) = φ(W,t)
, where φ(W, t) is the
cardinality of the smallest Abelian group containing a Sidon
set of order t and cardinality W + 1. The statement that
µa (W, t) & W −t then follows from the constructions of Sidon
sets by Singer [33] (t = 2) and Bose–Chowla [4] (arbitrary t)
which imply that φ(W, t) ≤ W t + o(W t ) as W → ∞.
Theorem 7. Let 0 ≤ d < k ≤ ∞. For every fixed t ∈ Z+ , as
n → ∞ we have
Sd,k (n)
Md,k (n; t) &
nt
→
Sd,k (n)
Md,k (n; t) .
nt
→
k+1
X
iρ
i
!t
iρ
i
!t
i=d+1
k+1
X
i=d+1
,
(23a)
⌈t/2⌉! ⌊t/2⌋!
t
t ,
1 − ρd+1 1 − ρk+1
(23b)
where ρ is the unique positive solution to
Pk+1
i=d+1
xi = 1.
For example, in the unconstrained case (d = 0, k = ∞) we
have S0,∞ (n) = 2n , ρ = 1/2, and the bounds (23) reduce to:
→
2n t
2n
2 . M0,∞ (n; t) . t 4t ⌈t/2⌉! ⌊t/2⌋!.
t
n
n
(24)
→
of an optimal code in Sd,k (n) (resp. Sd,k (n, W )) correcting
t right-shifts. The symbol → in the superscript indicates that
ሺʹǡሻ ሺ͵ǡሻ
ሺͶǡሻ
ሺͷǡሻ
ሺǡሻ
ሺͳǡሻ
only
right-shifts
are allowed
in the
model.
Since the channel
does not affectPthe weight of the transmitted string, we have
→
→
Md,k (n; t) = W Md,k (n, W ; t).
ሺʹǡͺሻ ሺ͵ǡͺሻ ሺͶǡͺሻ→ ሺͷǡͺሻ ሺǡͺሻ ሺǡͺሻ
The lower bound on Md,k (n; t) given in Theorem 7
below is obtained
by constructing a family of codes in
,
which
in turnሺͷǡͻሻ
is done
by using
S̄
(n,
W
),
d
ሺͶǡͻሻ
ሺǡͻሻ
ሺǡͻሻ“good”
ሺͺǡͻሻ
ሺͳǡͻሻ
d,k
a ሺ͵ǡͻሻ
codes in ZW , da and restricting them to S̄d,k (n, W ). For that
reason, the best known lower bound for codes in ZW , da is
given first (Lemma 6). A few definitions are needed to state it
W
W
precisely. We say that L ⊆
Z is a sublattice of ZW if (L, +)
W
is a subgroup of Z , + . The density of L in Z is defined
−1
as µ(L) := ZW /L , where ZW /L is the quotient group of
5 See [21, Sec. III] for a study of the geometry of ZW , d .
a
Proof: We first derive the lower bound (23a). Consider a
class of codes in S̄d,k (n, W ) obtained in the following way:
Take a lattice L ⊆ ZW of minimum distance da (L) = t + 1,
and let Cu := (u + L) ∩ S̄d,k (n, W ) for an arbitrary u ∈ ZW
(here u + L = {u + x : x ∈ L}). The code Cu has minimum
distance da (Cu ) ≥ t + 1 and hence it can correct t rightshifts. To give a lower bound on its cardinality notice that
there are ZW /L = µ(L)−1 different translates u + L that
are disjoint and whose union is all of ZW , so for at least
one such u we will have |Cu | = |(u + L) ∩ S̄d,k (n, W )| ≥
µ(L) · Sd,k (n, W ). This establishes the existence of a code
in S̄d,k (n, W ) of minimum distance > t and cardinality ≥
µa (W, t) · Sd,k (n, W ). From Lemma 6 we then conclude that
6 A Sidon set of order t in an Abelian group (G, +) is any subset
{b0 , b1 , . . . , bW } ⊆ G having the property that all its t-sums (bi1 +· · ·+bit )
are distinct, up to the order of the summands; see [26], [20].
6
S
→
(n,wn)
Md,k (n, wn; t) & d,kwt nt . Finally, to get the desired lower
→
bound on Md,k (n; t) write:
X →
→
Md,k (n; t) =
Md,k (n, W ; t)
(25)
W
w ∗ n+f (n)
>
X
Md,k (n, W ; t)
(26)
X
µa (W, t) · Sd,k (n, W )
(27)
→
W =w ∗ n−f (n)
w ∗ n+f (n)
≥
W =w ∗ n−f (n)
1
&
∗
(w n + f (n))t
∼
Sd,k (n)
.
(w∗ n)t
w ∗ n+f (n)
X
Sd,k (n, W ) (28)
W =w ∗ n−f (n)
(29)
where w∗ is the optimizing weight given in Lemma 1 and f (n)
is the sublinear function from Lemma 2. Here (27) follows
from our code construction, (28) follows from Lemma 6, and
(29) follows from Lemma 2 and the fact that f (n) = o(n).
We now turn to the derivation of the upper bound (23b).
Our approach is essentially a packing argument; however, due
to the structure of the code space and the fact that balls in it
do not have uniform sizes, some care is needed in making the
argument work. Let C ⊆ Sd,k (n) be an optimal code correcting
t right-shifts (or, equivalently, ⌊t/2⌋ right-shifts and ⌈t/2⌉ left→
shifts; see Remark 4), |C| = Md,k (n; t). Consider a codeword
x ∈ C of weight W , and let Λj (resp. j Λ) denote the number
of 1’s in x that are followed (resp. preceded) by exactly j
zeros, and Λ6=j (resp. 6=j Λ) the number of 1’s in x that are
followed (resp. preceded) by a run of zeros whose length is
not j. Let also i Λj denote the number of 1’s in x that are
preceded by exactly i zeros and followed by exactly j zeros;
i Λ6=j the number of 1’s in x that are preceded by exactly
i zeros and followed by a run of zeros whose length is not
j; and similarly for 6=i Λj and 6=i Λ6=j . We next show that the
number of strings in Sd,k (n) that can be obtained after x is
impaired by ⌊t/2⌋ right-shifts and ⌈t/2⌉ left-shifts is at least
6=k Λ6=d − 1
6=d Λ6=k − 2t
.
(30)
⌊t/2⌋
⌈t/2⌉
To see this, first count the number of strings that can be
obtained by shifting ⌊t/2⌋ 1’s one position to the right. In other
words, pick ⌊t/2⌋ out of W 1’s, and shift each of them one
position to the right. Notice that not all such choices will result
in a string that belongs to our code space Sd,k (n). Namely,
right-shifting a 1 that (a) is preceded by exactly k zeros, or
(b) is followed by exactly d zeros, or (c) is the last symbol
in the string, would result in either a string that violates the
(d, k)-constraints, or is of length n + 1. Excluding the 1’s
satisfying (a)–(c) leaves us with at least 6=k Λ6=d − 1 1’s to
choose from, which gives the left-hand term in (30). The righthand term is obtained in an analogous way by counting the
number of strings that can be obtained after picking ⌈t/2⌉ out
of 6=d Λ6=k −1 1’s and shifting each of them one position to the
left. The difference here is that, after choosing ⌊t/2⌋ 1’s for
the right-shifts in the first step, we exclude additional 3⌊t/2⌋
1’s in the second step. Namely, if the i’th 1 has been chosen
for the right-shift in the first step, then the (i − 1)’th, the i’th,
and the (i + 1)’th 1 are excluded in the second step: the i’th
because right-shifting and then left-shifting the same 1 would
potentially result in the same string we started with, and the
(i−1)’th (resp. (i+1)’th) because right-shifting the i’th 1 and
then left-shifting the (i − 1)’th (resp. (i + 1)’th) could result in
a run of zeros of length k+1 (resp. d−1) in between these 1’s.
We are thus left with at least 6=d Λ6=k −3⌊t/2⌋−1 ≥ 6=d Λ6=k −2t
1’s to choose from, which yields the right-hand term in (30).
This proves our claim that the expression in (30) is a lower
bound on the number of strings in Sd,k (n) that x can produce
after being impaired by ⌊t/2⌋ right-shifts and ⌈t/2⌉ left-shifts.
Let us now give the asymptotic form of (30), for fixed t
and n → ∞, that we shall need to conclude the proof. We
know from Lemma 3 that, for the “typical” strings in Sd,k (n),
W ∼ w∗ n and j Λ ∼ Λj ∼ λ∗j n = ρj+1 w∗ n, d ≤ j ≤ k.
This implies that a given ∗block 0j 1 is preceded by a block
λ
0i 1 with probability ≈ wi∗ = ρi+1 , or in other words, of
the Λj blocks 0j 1, a fraction of ≈ ρi+1 is preceded by a
block 0i 1. It follows that, for the typical strings in Sd,k (n),
∗
d+1
)(1 − ρk+1 )n, and therefore the
6=d Λ6=k ∼ w (1 − ρ
expression (30) has the following asymptotic form:
t
w∗ (1 − ρd+1 ) (1 − ρk+1 ) nt
,
(31)
∼
⌈t/2⌉! ⌊t/2⌋!
nm
n
where we have used the fact that m
∼ m! for fixed m and
n → ∞.
Finally, due to our assumption that C corrects ⌊t/2⌋ rightshifts and ⌈t/2⌉ left-shifts, the sets of outputs that can be
obtained in the above-described way from any two different
codewords x, y ∈ C have to be disjoint. This implies that
t
w∗ (1 − ρd+1 ) (1 − ρk+1 ) nt
→
. Sd,k (n)
Md,k (n; t) ·
⌈t/2⌉! ⌊t/2⌋!
(32)
and proves the upper bound in (23b). We have used in (32) the
fact that in the asymptotic analysis we can safely ignore the
non-typical inputs (see Lemma 3), as we did in the derivation
of the lower bound as well.
From Theorem 7 we can find the asymptotic scaling of the
redundancy of optimal codes correcting t right-shifts.
Corollary 8. For every fixed t ∈ Z+ , as n → ∞ we have
→
log Sd,k (n) − log Md,k (n; t) = t log n + O(1).
(33)
IV. C ODES C ORRECTING S YMMETRIC S HIFTS
In this section we discuss a slightly different scenario—the
one which is usually considered in the literature on bit-shift
channels—where both left-shifts and right-shifts are allowed
in the model and are treated in a symmetric way. Our object
of study are codes enabling the receiver to reconstruct the
transmitted string whenever the total shift of its 1’s does not
exceed a specified threshold, regardless of the direction of the
individual shifts.
7
A. Geometric Characterization
ሺͳǡʹሻ
Suppose that x̄ ∈ S̄d,k (n) is the transmitted codeword, and
z the corresponding received vector. If the i’th 1 in x is shifted
by ki ∈ Z positions, we have zi = x̄i + ki . Therefore, we
ሺͳǡ͵ሻ
ሺʹǡ͵ሻ
can again
think of the bit-shift channel as an additive noise
channel with input alphabet
PWZ+ . If the 1’s in x̄ are shifted by
t
positions
in
total,
then
i=1 |zi − xi | = t.
ሺͳǡͶሻ ሺʹǡͶሻ ሺ͵ǡͶሻ
We say that a code C¯ ⊆ S̄d,k (n) can correct t shifts if no
two different codewords x̄, ȳ ∈ C¯ can produce the same output
after being
arbitrary patterns of ≤ t shifts. In
ሺͳǡͷሻ
ሺʹǡͷሻimpaired
ሺ͵ǡͷሻ with
ሺͶǡͷሻ
¯ x̄ 6= ȳ, wt(x) = wt(y) =
symbols, for every x̄, ȳ ∈ C,
PW
W , and all noise vectors f , g ∈ ZW with
i=1 |fi | ≤ t,
P
W ሺʹǡሻ ሺ͵ǡሻ ሺͶǡሻ ሺͷǡሻ
ሺͳǡሻ
|g
|
≤
t,
we
have
x̄
+
f
=
6
ȳ
+
g.
Such
a
code
C¯ is said
i
i=1
′
¯
to be optimal if there is no other code C ⊆ S̄d,k (n) correcting
¯
t
shiftsሺʹǡሻ
and such
that |ሺͶǡሻ
C¯′ | > |C|.
ሺͳǡሻ
ሺͷǡሻ ሺǡሻ
ሺ͵ǡሻ
Let d1 denote the ℓ1 (or Manhattan) distance on S̄d,k (n, W ):
W
X
ሺʹǡͺሻ ሺ͵ǡͺሻ
|x̄ −ሺǡͺሻ
ȳ |. ሺǡͺሻ
d (x̄,ሺͶǡͺሻ
ȳ) := ሺͷǡͺሻ
1
i
i
(34)
i=1
For vectors of ሺ͵ǡͻሻ
differentሺͶǡͻሻ
dimensions
strings
ሺͳǡͻሻ
ሺͷǡͻሻ(corresponding
ሺǡͻሻ ሺǡͻሻto ሺͺǡͻሻ
of different weights), x̄ ∈ S̄d,k (n, W1 ), ȳ ∈ S̄d,k (n, W2 ),
W1 6= W2 , we define d1 (x̄, ȳ) = ∞. The minimum distance
of a code C¯ ⊆ S̄d,k (n) with respect to the metric d1 is denoted
¯
d1 (C).
Proposition 9. A code C¯ ⊆ S̄d,k (n) can correct t shifts if and
¯ > 2t.
only if d1 (C)
Proof: Analogous to the proof of Proposition 5.
ሺͳǡʹሻ
ሺͳǡ͵ሻ ሺʹǡ͵ሻ
ሺͳǡͶሻ ሺʹǡͶሻ ሺ͵ǡͶሻ
ሺͳǡͷሻ ሺʹǡͷሻ ሺ͵ǡͷሻ ሺͶǡͷሻ
ሺͳǡሻ ሺʹǡሻ ሺ͵ǡሻ ሺͶǡሻ ሺͷǡሻ
ሺͳǡሻ ሺʹǡሻ ሺ͵ǡሻ ሺͶǡሻ ሺͷǡሻ ሺǡሻ
ሺͳǡͻሻ
ሺʹǡͺሻ ሺ͵ǡͺሻ ሺͶǡͺሻ ሺͷǡͺሻ ሺǡͺሻ ሺǡͺሻ
ሺ͵ǡͻሻ ሺͶǡͻሻ ሺͷǡͻሻ ሺǡͻሻ ሺǡͻሻ ሺͺǡͻሻ
Fig. 2: The space S̄0,5 (9, 2) representing the set of all binary
strings of length n = 9 and weight W = 2 satisfying (0, 5)constraints, and a code correcting t = 1 shift. Codewords are
depicted as black dots; gray regions illustrate balls (under the
metric d1 ) of radius t = 1 around the codewords.
Again, balls in the metric space S̄d,k (n, W ), d1 do not
have uniform sizes—the size of a ball of radius r depends on
its center. In the unrestricted metric space ZW , d1 this effect
does not occur and we have the following expression [9] for
the cardinality of a ball of radius r:
X W r
.
(35)
2i
B1 (W, r) =
i
i
i
The cardinality of an arbitrary ball of radius r in
S̄d,k (n, W ), d1 is upper bounded by B1 (W, r).
B. Construction and Bounds
↔
↔
Let Md,k (n; t) (resp. Md,k (n, W ; t)) denote the cardinality
of an optimal code in Sd,k (n) (resp. Sd,k (n, W )) correcting t
shifts. As in the asymmetric case, the channel does not affect
↔
the weight of the transmitted string, so we have Md,k (n; t) =
P
↔
W Md,k (n, W ; t).
In analogy with (22) we define the maximum density a
lattice L ⊆ ZW of minimum distance d1 (L) > 2t can have as
n
o
µ1 (W, t) := max µ(L) : L ⊆ ZW a lattice, d1 (L) > 2t .
(36)
Lemma 10. For every fixed t ∈ Z+ , as W → ∞ we have
µ1 (W, t) & c(t) · W −t , where
(
2−t ,
t≤2
c(t) :=
.
(37)
(2t + 1)−1 , t ≥ 3
Proof: Consider the so-calledP
AW −1 lattice defined by
AW −1 := (u1 , . . . , uW ) ∈ ZW : W
i=1 ui = 0 , and let L0
be the densest sublattice of AW −1 with d1 (L0 ) > 2t. The
metric space (AW −1 , d1 ) is in fact isometric to (ZW −1 , 2da )
[21, Thm 4], so the density of L0 in AW −1 can be expressed
as µ(L0 ) := |AW −1 /L0 |−1S = µa (W − 1, t). Now define a
lattice L ⊆ ZW by L := k∈Z L0 + k(2t + 1)e1 , where
e1 = (1, 0, . . . , 0) is a unit vector in ZW . In words, L comprises infinitely many translates of L0 separated by a multiple
of 2t + 1. It follows from the construction that d1 (L) > 2t
1
1
µ(L0 ) = 2t+1
µa (W − 1, t).
and µ(L) := |ZW /L|−1 = 2t+1
Therefore, recalling from Lemma 6 that µa (W − 1, t) & W −t ,
we conclude that µ1 (W, t) & (2t + 1)−1 W −t .
The lower bound just given can be improved for t = 1, 2
to µ1 (W, t) & 2−t W −t . In fact, for t = 1 the optimal density
is known exactly for every W , namely µ1 (W, 1) = 2W1+1 .
This follows from the existence of perfect codes of radius
t = 1 in (ZW , d1 ), for every W [9]. For t = 2 (or, indeed, for
any t), one can construct codes in (ZW , d1 ) by periodically
extending the codes in the torus ZW
m correcting t = 2 errors
in the Lee metric7 (here Zm := Z/(mZ)). Such a periodic
extension of Berlekamp’s codes for the Lee metric [3, Ch. 9]
gives µ1 (W, 2) & 14 W −2 .
7 To the best of our knowledge, no known construction of codes in the
Lee metric gives a lower bound on the density µ1 (W, t) better than the one
obtained above, except for t = 1, 2. For example, Berlekamp’s construction
[3, Ch. 9] (see also [6]) gives µ1 (W, t) & 2−t W −t , and the BCH-like
−t , where
construction of Roth and Siegel [29] gives µ1 (W, t) & p−1
2t+3 W
p2t+3 is the smallest prime greater than or equal to 2t + 3.
8
Theorem 11. Let 0 ≤ d < k ≤ ∞. For every fixed t ∈ Z+ ,
as n → ∞ we have
!t
k+1
X
↔
Sd,k (n)
(38a)
iρi c(t),
Md,k (n; t) &
nt
i=d+1
!t
k+1
X
↔
Sd,k (n)
t! 2−t
i
Md,k (n; t) .
iρ
t ,
t
nt
1 − ρd+1 1 − ρk+1
i=d+1
(38b)
where c(t) P
is defined in (37) and ρ is the unique positive
k+1
solution to i=d+1 xi = 1.
For example, in the unconstrained case (d = 0, k = ∞) we
have S0,∞ (n) = 2n , ρ = 1/2, and the bounds (38) reduce to:
↔
2n t
2n t
2
c(t)
.
M
(n;
t)
.
2 t!.
(39)
0,∞
nt
nt
Proof: The proof is analogous to the proof of Theorem
7 for the asymmetric case. The only difference in proving the
lower bound (38a) is that we need to use the metric d1 instead
of da (see Proposition 9). Following the same steps as in (25)–
↔
(29) we get Md,k (n; t) & µ1 (w∗ n, t) · Sd,k (n), and then the
result follows by applying Lemma 10.
Now for the upper bound (38b). Let C ⊆ Sd,k (n) be
↔
an optimal code correcting t shifts, |C| = Md,k (n; t), and
consider a codeword x ∈ C. Notice that every pattern of t
shifts that can impair x in the channel consists of r rightshifts and t − r left-shifts, for some r ∈ {0, 1, . . . , t}. By a
reasoning identical to that used in the proof of Theorem 7 we
then conclude that the number of strings in Sd,k (n) that can
be produced after x is impaired by t shifts is at least
t
X
6=k Λ6=d − 1
6=d Λ6=k − 3r − 1
(40)
r
t−r
r=0
(see equation (30) and the paragraph following it). Recalling
that 6=d Λ6=k ∼ 6=k Λ6=d ∼ w∗ (1−ρd+1 )(1−ρk+1 )n as n → ∞,
we find the asymptotics of the expression (40) in the form:
t−r
r
t
X
t
2t
6=d Λ6=k
6=k Λ6=d
∼
∼
6=k Λ6=d
r!
(t
−
r)!
t!
r=0
t
2t
w∗ (1 − ρd+1 ) (1 − ρk+1 ) nt .
(41)
t!
Since C corrects t shifts by assumption, we must have
t
2t w∗ (1 − ρd+1 ) (1 − ρk+1 ) nt
↔
Md,k (n; t) ·
. Sd,k (n),
t!
(42)
∼
which proves the upper bound (38b).
Corollary 12. For every fixed t ∈ Z+ , as n → ∞ we have
↔
log Sd,k (n) − log Md,k (n; t) = t log n + O(1).
↔
(43)
The only bounds on Md,k (n; t) appearing in the literature
that we are aware of are those in [1] (and [36] for t = 1).
However, these bounds are not explicit and are difficult to
compare to (38).
V. C ONCLUDING R EMARKS
We conclude the paper with a few remarks on error models
related to those studied in this paper.
In some applications it is reasonable to assume that the shifts
are limited in the sense that each 1 in an input string x can be
shifted by at most s positions [1], [22], for some s ∈ Z+ . More
precisely, if x̄ is the transmitted vector and z the received
vector, the assumption here is that 0 ≤ zi − x̄i ≤ s, 1 ≤ i ≤ n,
in the asymmetric case (when only right-shifts are allowed in
the model), or |zi − x̄i | ≤ s in the symmetric case. For such
models, the lower bound on the cardinality of optimal shiftcorrecting codes can possibly be improved by using known
constructions of codes for limited-magnitude errors; see, e.g.,
[5]. For example, in the asymmetric case with t = 2, s = 1,
the lower bound in (23a) can be improved by a factor of 2
by using a construction of codes correcting 2 asymmetric 1limited-magnitude errors [5, Sec. IV.C]. Note that we have
used implicitly the assumption that s = 1 in our derivation
of the upper bounds (23b) and (38b). Therefore, these upper
bounds are not likely to be improved in limited-shift models
by using the approach we have used.
In this context one may also be interested in codes correcting all possible patterns of shifts such that each shift is
bounded by s. Such codes, usually called zero-error codes,
have been studied in several related settings: bit-shift channels
[30], [22], timing channels [17], [19], skew-tolerant parallel
asynchronous communications [7], [18], etc. In many cases,
the optimal codes have been found and the zero-error capacity
of the corresponding channel determined.
ACKNOWLEDGMENT
The author would like to thank Vincent Y. F. Tan (NUS), for
his detailed reading and helpful comments on a preliminary
version of this work; Mehul Motani (NUS), for several discussions on a model related to the one studied in this paper; and
Anshoo Tandon (NUS), for helpful discussions on constrained
codes and related notions.
R EFERENCES
[1] K. A. S. Abdel-Ghaffar and J. H. Weber, “Bounds and Constructions
for Runlength-Limited Error-Control Block Codes,” IEEE Trans. Inform.
Theory, vol. 37, no. 3, pp. 789–800, May 1991.
[2] V. Anantharam and S. Verdú, “Bits Through Queues,” IEEE Trans. Inf.
Theory, vol. 42, no. 1, pp. 4–18, Jan. 1996.
[3] E. R. Berlekamp, Algebraic Coding Theory, revised edition, World
Scientific, Singapore, 2015.
[4] R. C. Bose and S. Chowla, “Theorems in the Additive Theory of
Numbers,” Comment. Math. Helv., vol. 37, no. 1, pp. 141–147, Dec.
1962.
[5] Y. Cassuto, M. Schwartz, V. Bohossian, and J. Bruck, “Codes for
Asymmetric Limited-Magnitude Errors With Application to Multilevel
Flash Memories,” IEEE Trans. Inform. Theory, vol. 56, no. 4, pp. 1582–
1595, Apr. 2010.
[6] J. C. Chiang and J. K Wolf, “On Channels and Codes for the Lee Metric,”
Inform. and Control, vol. 19, no. 2, 159–173, Sep. 1971.
[7] S. Engelberg and O. Keren, “Reliable Communications Across Parallel
Asynchronous Channels With Arbitrary Skews,” IEEE Trans. Inform.
Theory, vol. 63, no. 2, pp. 1120–1129, Feb. 2017.
[8] H. C. Ferreira and S. Lin, “Error and Erasure Control (d, k) Block
Codes,” IEEE Trans. Inform. Theory, vol. 37, no. 5, pp. 1399–1408,
Sep. 1991.
9
[9] S. W. Golomb and L. R. Welch, “Perfect Codes in the Lee Metric and
the Packing of Polyominoes,” SIAM J. Appl. Math., vol. 18, no. 2, pp.
302–317, Mar. 1970.
[10] S. Heubach and T. Mansour, “Compositions of n with Parts in a Set,”
Congr. Numer., vol. 168, pp. 127–143, 2004.
[11] H. M. Hilden, D. G. Howe, and E. J. Weldon, Jr., “Shift Error Correcting
Modulation Codes,” IEEE Trans. Magn., vol. 27, no. 6, pp. 4600–4605,
Nov. 1991.
[12] K. A. S. Immink, “Runlength-Limited Sequences,” Proc. IEEE, vol. 78,
no. 11, pp. 1745–1759, Nov. 1990.
[13] K. A. S. Immink, P. H. Siegel, and J. K. Wolf, “Codes for Digital
Recorders,” IEEE Trans. Inform. Theory, vol. 44, no. 6, pp. 2260–2299,
Oct. 1998.
[14] K. A. S. Immink and K. Cai, “Design of Capacity-Approaching Constrained Codes for DNA-Based Storage Systems,” IEEE Commun. Lett.,
vol. 22, no. 2, pp. 224–227, Feb. 2018.
[15] T. Kløve, “Error Correcting Codes for the Asymmetric Channel,” Technical Report, Dept. of Informatics, University of Bergen, 1981. (Updated
in 1995.)
[16] T. Kløve, “Codes Correcting a Single Insertion/Deletion of a Zero or
a Single Peak-Shift,” IEEE Trans. Inform. Theory, vol. 41, no. 1, pp.
279–283, Jan. 1995.
[17] M. Kovačević and P. Popovski, “Zero-Error Capacity of a Class of
Timing Channels,” IEEE Trans. Inform. Theory, vol. 60, no. 11, pp.
6796–6800, Nov. 2014.
[18] M. Kovačević, “A Note on Parallel Asynchronous Channels With
Arbitrary Skews,” IEEE Trans. Inform. Theory, vol. 63, no. 11, pp.
7320–7321, Nov. 2017.
[19] M. Kovačević, M. Stojaković, and V. Y. F. Tan, “Zero-Error Capacity of
P -ary Shift Channels and FIFO Queues,” IEEE Trans. Inform. Theory,
vol. 63, no. 12, pp. 7698–7707, Dec. 2017.
[20] M. Kovačević and V. Y. F. Tan, “Improved Bounds on Sidon Sets via
Lattice Packings of Simplices,” SIAM J. Discrete Math., vol. 31, no. 3,
pp. 2269–2278, 2017.
[21] M. Kovačević and V. Y. F. Tan, “Codes in the Space of Multisets—
Coding for Permutation Channels with Impairments,” IEEE Trans.
Inform. Theory, to appear in vol. 64, 2018. Published online at:
https://doi.org/10.1109/TIT.2017.2789292.
[22] V. Yu. Krachkovsky, “Bounds on the Zero-Error Capacity of the InputConstrained Bit-Shift Channel,” IEEE Trans. Inform. Theory, vol. 40,
no. 4, pp. 1240–1244, Jul. 1994.
[23] A. V. Kuznetsov and A. J. Han Vinck, “A Coding Scheme for Single
Peak-Shift Correction in (d, k)-Constrained Channels,” IEEE Trans.
Inform. Theory, vol. 39, no. 4, pp. 1444–1450, Jul. 1993.
[24] V. I. Levenshtein and A. J. Han Vinck, “Perfect (d, k)-Codes Capable
of Correcting Single Peak-Shifts,” IEEE Trans. Inform. Theory, vol. 39,
no. 2, pp. 656–662, Mar. 1993.
[25] T. Nakano, A. W. Eckford, and T. Haraguchi, Molecular Communication, Cambridge University Press, 2013.
[26] K. O’Bryant, “A Complete Annotated Bibliography of Work Related to
Sidon Sequences,” Electron. J. Combin., #DS11, 39 pp, 2004.
[27] R. Pemantle and M. C. Wilson, Analytic Combinatorics in Several
Variables, Cambridge University Press, 2013.
[28] E. Rosnes, Á. I. Barbero, and Ø. Ytrehus, “Coding for Inductively
Coupled Channels,” IEEE Trans. Inform. Theory, vol. 58, no. 8, pp.
5418–5436, Aug. 2012.
[29] R. M. Roth and P. H. Siegel, “Lee-Metric BCH Codes and their
Application to Constrained and Partial-Response Channels,” IEEE Trans.
Inform. Theory, vol. 40, no. 4, pp. 1083–1096, Jul. 1994.
[30] S. Shamai (Shitz) and E. Zehavi, “Bounds on the Capacity of the BitShift Magnetic Recording Channel,” IEEE Trans. Inform. Theory, vol.
37, no. 3, pp. 863–872, May 1991.
[31] P. H. Siegel, “Recording Codes for Digital Magnetic Storage,” IEEE
Trans. Magn., vol. 21, no. 5, pp. 1344–1349, Sep. 1985.
[32] P. H. Siegel and J. K. Wolf, “Modulation and Coding for Information
Storage,” IEEE Commun. Mag., vol. 29, no. 12, pp. 68–86, Dec. 1991.
[33] J. Singer, “A Theorem in Finite Projective Geometry and Some Applications to Number Theory,” Trans. Amer. Math. Soc., vol. 43, pp. 377–385,
1938.
[34] R. P. Stanley, Enumerative Combinatorics, Vol I, Cambridge University
Press, 1997.
[35] D. T. Tang and L. R. Bahl, “Block Codes for a Class of Constrained
Noiseless Channels,” Inform. and Control, vol. 17, no. 5, pp. 436–461,
Dec. 1970.
[36] Ø. Ytrehus, “Upper Bounds on Error-Correcting Runlength-Limited
Block Codes,” IEEE Trans. Inform. Theory, vol. 37, no. 3, pp. 941–
945, May 1991.
| 7cs.IT
|
1
Fine-tuning the Ant Colony System algorithm
through Particle Swarm Optimization.
D. Gómez-Cabrero+, D. N. Ranasinghe*
*University of Colombo School of Computing
35, Reid Avenue, Colombo 7, Sri Lanka
Email: [email protected]
+
Departamento de Estadística e Investigación Operativa,
Universitat de Valencia, Calle Doctor Moliner s/n, Burjassot, España
Email: [email protected]
Abstract—Ant Colony System (ACS) is a distributed (agentbased) algorithm which has been widely studied on the
Symmetric Travelling Salesman Problem (TSP). The optimum
parameters for this algorithm have to be found by trial and
error. We use a Particle Swarm Optimization algorithm (PSO)
to optimize the ACS parameters working in a designed subset
of TSP instances. First goal is to perform the hybrid PSO-ACS
algorithm on a single instance to find the optimum parameters
and optimum solutions for the instance. Second goal is to
analyze those sets of optimum parameters, in relation to
instance characteristics. Computational results have shown
good quality solutions for single instances though with high
computational times, and that there may be sets of parameters
that work optimally for a majority of instances.
I. INTRODUCTION
H
EURISTICS algorithms, and their higher forms metaheuristics, have been widely used to find reasonable
good solutions for NP-Hard problems. The performance of
those heuristics is based on the optimum set-up of a set of
parameters. The problem of fine-tuning those parameters is
also a hard problem.
Fine-Tuning is an unavoidable task that needs a correct
design of experiments. For statistical experimental design
and analysis for heuristics see [4]. Heuristics as CALIBRA
[1] and F-Race [3] have been designed as procedures for
finding optimal parameters. Fine-tuning can be understood
as “to find the set of parameters performing well on a wide
set of instances”.
We have designed an algorithm, using a Particle Swarm
Optimization (PSO) framework, to optimize the parameters
of the ACS algorithm working on a “single” Symmetric
Travelling Salesman Problem (TSP) instance. For each
instance the algorithm computes an “optimal” set of ACS
parameters. A second goal is to analyze jointly those sets of
parameters, their performance on all instances (not only
their related instance) and the characteristics of their related
instances for the purpose of finding correlations.
Ant Colony Optimization (ACO) is a generic framework
for ant-based optimization heuristic algorithms. In ACO
algorithms ants are agents which, in the TSP case, construct
tours by moving from city to city on the graph problem.
Those ants are sharing information using a pheromone trail.
The first ACO algorithm, called Ant-System, was
proposed in [5], [7] and [8]. A full review of ACO
algorithms and applications can be found in [9]. ACS is a
version of the Ant System that modifies the updating of the
pheromone trail, see [6] and [11]. We have chosen this ACS
algorithm to work with because of the theoretical
background we have found on it, see [6] and [9], and the
previous fine-tuning research on the parameters by [16].
PSO is a Swarm Intelligence method for global
optimization. Given a domain D, there is defined a function
f : D → ℜ assigning to each point of D a fitness value.
In PSO there is a population (swarm) of individuals, named
particles, moving on the domain and adjusting their
trajectory to its own previous best position and the previous
best position of the neighbourhood. We will use the Global
PSO version which considers the neighbourhood as all the
swarm. For a further review of the PSO [10], [13] and [14].
An example of PSO applications can be found in [12] and
[15]. We have chosen PSO because it has an easy
implementation for integer and real parameters, and, as
genetic algorithms, it performs a “blind” search on all the
possible sets of parameters.
In our algorithm the domain of the PSO will be all
possible sets of parameters for ACS. For a position of a
particle we compute the fitness by running the ACS
algorithm with the parameters given by the position on a
TSP instance.
Section II describes ACS and PSO algorithms. Section III
describes the PSO-ACS algorithm, the parameters used on
the PSO and the initial population of particles (set of
feasible parameters for the ACS algorithm). Computational
2
results are given in section IV and finally in section V
conclusions are set.
II. PSO AND ACS ALGORITHMS
Let us introduce some notation about a TSP graph, where
for a given instance denote V as the set of vertices, E as the
set of edges (shortest paths between the vertices), ce cost of
traversing edge e∈E.
A. Ant Colony System (ACS)
The ACS works as follow: it has a population of n ants.
Let denote for each arc e=(i,j) in the TSP-instance graph an
initial heuristic value ηe and an initial pheromone value τe.
ηe is originally set to the inverse of the cost of traversing the
edge e. τe is initially set to τ0=1\Lnn for all edge e, where Lnn
is equal to the inverse of the tour length computed by the
nearest-neighbour-heuristic algorithm.
Let q0, α, ρ ∈ [0,1], be real values and ϕ, β integer values
between 0 and 8. For each vertex s ∈V a neighbour set is
defined among the nearest vertices, N(s). For a given ant r
let be NV(r) the set of non-visited vertices. We denote
Jr(s)=N(s)∩NV(r) the set of non-visited vertices among the
neighbour set for a given vertex s and a given ant r.
At each iteration, each ant constructs a tour solution for
the TSP-instance. The constructions phase works as follows:
Each ant is initially set in a randomly vertex, then at each
step the entire ants make a movement to a non-visited
vertex. Given and ant r with an actual position (vertex) s,
pk(rs) is computed as a reference value for visiting or not
vertex k, where:
pkrs
⎧ [τ ( s ,k ) ]ϕ [η ( s ,k ) ]β
⎪⎪
ϕ
β
= ⎨ ∑ [τ ( s ,u ) ] [η( s ,u ) ]
u∈J ( s )
⎪
⎪⎩
0
⎫
if k ∈ J(s)⎪
⎪
⎬
⎪
otherwise ⎪⎭
(1)
This formula includes a small modification respect to
the original ACS algorithm including ϕ as exponent of the
pheromone level, this will allow us a deeper research on the
effects of the possible combinations of ϕ, β parameters.
A sample random value q is computed. If q≤q0 we visit
the city k∈V with maximum pk(rs) (exploitation of the
knowledge) otherwise ACS follows a random-proportional
rule on pk(rs) for all k∈V (biased exploration). If there are no
non-visited vertex on the neighbour of vertex s, we extend
(1) to all vertices in NV(r)/N(s) (those not visited by ant r
and not included in the neighbour of s) and visits the vertex
with maximum pk(rs).
After an arc is inserted into a route (a new vertex is
visited), its pheromone trail is updated. This phase is called
Local Update, and for a inserted e∈E:
τ e = (1 − α )τ e + ατ 0
(2)
This reduces the pheromone level in the visited arcs and
the exploration in the set of possible tours is increased.
When all the tours have been computed a Global Update
phase is done. For each edge e pertaining to the global-besttour found:
τ e = (1 − α )τ e + α∆τ e
∆τ e = 1/L gb
(3)
(4)
where Lgb is the length of the global-best-tour found.
In the original ant algorithm, and in most of the later
versions, pheromone global update is performed in all the
edges; ACS only updates pheromone level in the set of
edges pertaining to the best tour.
We consider a trial as a performance of 1000 iterations.
The lowest length tour found after all iterations are finished
is the best solution found by the trial.
A feasible set of parameters for running ACS is a
combination of feasible q0, ϕ, β, α, ρ, number of ants (na)
and a concrete neighbour definition.
B. Particle Swarm Optimization (PSO)
As described by Eberhert and Kennedy [10], [13], [14]
PSO is an adaptative algorithm based on a social
environment where a set of particles, called population, are
visiting possible “positions” of a given domain. Each
position has a fitness value (and it can be computed). At
each iteration particles will move returning stochastically
toward its previous best fitness position and to the
population best fitness position. All the particles of the
population are sharing information of the best areas to
search.
Let denote P as the set of parameters and let define PO as
the population of particles. At each iteration xfp and vfp
denotes respectively the actual position and the actual
velocity of parameter p∈P of the particle f ∈ PO.
The movement of the particles are defined by the
following equations:
v fp = wv fp + r1c1 ( x fp − bl fp ) + r2 c 2 ( x fp − bg p ) (5)
x fp = x fp + χv fp
(6)
Where c1, c2 are integer non-negative values, named
cognitive and social respectively, r1, r2 are sample random
values in [0,1], w and χ are non-negative real values, named
respectively inertia weight and constriction factor, bgp is the
value of parameter p pertaining to the best set of parameters
found by the population (social knowledge) and blfp is the
value of parameter p pertaining to the best parameters set
found by particle f (self-knowledge). In (5) the first factor
refers to the previous velocity, second and third factors are
related respectively to the distance to the best set of
parameters found by the particle and to the distance to the
3
III. PSO-ACO ALGORITHM
A. PSO Parameters, initial particles and fitness value.
The algorithm is run each time on a single TSP-instance.
The set of parameters of ACS that define a point in the PSO
domain are q0, ϕ, β, φ , α, ρ and the number of ants (na).
Most of them have been previously explained in the section
II. φ denotes the percentage of vertices that will be included
into N(v) for any vertex v∈V, so given a v∈V and φ=0.5,
|N(v)|=⎡φ*|V|⎤ .The ranges of each parameter are shown in
Table I, where each parameter pertains to its related
]Minimum, Maximum].
TABLE I
RANGE OF ACS PARAMETERS
q0
ϕ
β
ρ
α
φ
Minimum
0
-1
-1
0
0
0
Maximum
1
8
8
1
1
1
n
a
0
4
0
q0, ϕ, β, φ , α, ρ and na are the parameters used in
ACS. φ is explained in the beginning of section III
DPSO = ]0,1]x]-1,8]x]-1,8]x]0,1]x]0,1]x]0,1]x]0,40]⊂ℜ7
is the domain of the PSO. We define the fitness value of a
given position (point) as the length of the best tour
computed by an ACS using the related parameters in the
given instance. If comparing two different positions they
have the same length value then computing time is
considered. We consider better of those parameters that
minimize the length of the tour and secondly the time of
computing. For computing the fitness of a given position,
first integer parameters (na, ϕ and β) are rounded up as
shown in Fig. 1, secondly the algorithm runs five trials of
the ACS algorithm using the rounded parameters in the
TSP-instance and returns the best value obtained from the
trials.
Point in the PSO domain (reflects a set of
parameters):
(0.1 2.3 8 0.5 0.88 0.34 32.4)
Modified values to run on ACS:
(0.1 3 8 0.5 0.88 0.34 33 )
Fig. 1. Modification of PSO points. In bold are the
modified values
set initially to 1 and gradually decreasing from 1 to 0.1 (at
each PSO iteration w=0.99w). Maximum number of
iterations is set to 500 due to computing time constraints
(for 1000 PSO-iterations more than 1 day of computing time
was necessary).
B. Algorithm
The algorithm PSO-ACS’s pseudo-code is as follows:
Select TSP-Instance.
Initialize particles.
Do 500
For all the set of particles
Position_Fitness (PF)=INFINITE
Do 5
Perform a trial ACS with particle parameters.
If NewValue < PF
PF=NewValue
End if
End Do
End for
Compute w=0.99w
Update Best Parameters Found by each Particle
Update Best Parameters Found by the Population
Compute Velocity
Movement of Particles
End Do
Return the set of parameters related to the best tour length
found and the tour length.
The algorithm is based in a PSO framework, where
particles are initialized and iteratively are moving though
the domain of the set of parameters. The goal of the
algorithm is, for a given instance, to compute the tour with
lowest length and to compute the set of ACS parameters,
among those in DPSO, which gets the best ACS
performance. Those final parameters are related with the
TSP-instance selected.
eil101
rat99
50,00
AVG_TIME_IT
best set of parameters found by the population.
40,00
30,00
20,00
0,00
For each particle of the population its initial velocity is
set randomly. For half of the population the initial position
is set using predefined parameters assuring that for every
parameter there will be a particle containing a value
covering the full range. The positions of the other half of the
initial population is set randomly.
Parameters for the PSO (see Section II) have been set
following [13], [14]: c1=c2=2, χ=0.729, the inertia weight is
25,00
50,00
75,00
100,00
0,00
25,00
50,00
ITERATION_PSO
ITERATION_PSO
(a)
(b)
75,00
100,00
Fig 2. Average time of the swarm at first 100 iterations for instances eil101
(a) and rat99 (b). AVG_TIME is average time of the iteration (given a
fixed number of particles) and ITERATION_PSO is the number of the
iteration in the PSO-ACS algorithm.
4
kroA100
kroB100
22800,00
MIN_VALUE_ITE
MIN_VALUE_ITE
22800,00
22400,00
22000,00
22600,00
22400,00
21600,00
22200,00
0,00
25,00
50,00
75,00
100,00
0,00
ITERATION_PSO
25,00
50,00
75,00
100,00
ITERATION_PSO
(a)
(b)
kroA100
eil51
480,00
30000,00
AVG_VALUE_IT
AVG_VALUE_IT
470,00
460,00
450,00
27500,00
25000,00
440,00
22500,00
430,00
0,00
25,00
50,00
75,00
100,00
0,00
25,00
50,00
75,00
100,00
ITERATION_PSO
ITERATION_PSO
(c)
(d)
Fig. 3. First 100 iterations of the PSO-ACS algorithm. (a) and (b) are related to the minimum tour obtained at each
iteration, (c) and (d) are related to the average of the particles´ fitness values. (a) and (d) are related to the instance kroA100;
(b) is related to kroB100; and (c) to eil51. Those are examples of typical behaviors in the 100 first iterations of the algorithm.
TABLE II.
SETS OF PARAMETERS
q0
α,
β
ρ
ϕ na
P_eil51
0.36 7 0.40 1
1 0.54
P_eil76
0.21 5 0.40 1
5 0.58
P_eil101
0.71 7 0.23 2
7 0.78
P_kroA100 0.64 4 0.24 1
4 0.64
P_kroB100 0.71 1 0.08 1
39 0.86
P_rat99
0.15 3 0.28 1
9 0.95
ACS
0.10 2 0.10 1
10 0.9
ACS_GA 0.20 6 0.20 1
10 0.7
“a” have been tested for φ = 0.1 0.2 … 0.9 1 “b”
value related. Values in bold mean optimums.
Fitness
φ
0.18
426
0.20
538
0.12
629
0.12
21282
0.12
22141
0.00
1211
a
b
a
b
there is no fitness
IV. COMPUTATIONAL RESULTS
Algorithm was coded in C++ in a Pentium M 1.7 with
1GB RAM. Algorithm has been run on six of the most
widely used TSP-instances. Computational results are given
in four parts: PSO-ACS behavior, PSO-ACS optimum
values obtained, best set of parameters and comparison
among sets of parameters´ performance.
Computationally each PSO-ACS iteration shows a clear
convergence: when the “optimum” (defined by the
algorithm) number of ants and φ are nearly fixed, the
computational time is also fixed (see fig. 2). In less than 100
iterations algorithm computes an “optimum” for integer
parameters and in 200 iterations there are small differences
among the optimum found and the particles´ position for
real parameters. In Fig. 3 we can see the evolution of the
algorithm in the first 100 iterations. For the average of the
fitness of the swarm (in a given iteration), there is a
decreasing global tendency, and after iteration 75 we can
see the average of the fitness is kept on a fixed range, the
size of this range is variable as shown in (c) and (d). For the
minimum value obtained by the swarm in a given iteration,
computational results show that at the beginning there are
increasing and decreasing phases, when the particles are
exploring their local optimums and moving also to the
global one, but near iteration 100 the minimum is
5
TABLE IV.
MINIMUM TOUR LENGTH FOUND. BY PARAMETER_SET / INSTANCE
TABLE III.
AVERAGE PERFORMANCE. BY PARAMETER_SET / INSTANCE
eil51
eil76
eil101
kroA100
kroB100
rat99
P_eil51
548.04
653.68 22147.6
22718.92
1241.64
428.6ab
P_eil76
22357.24
1223.8b
428.52 544.36ab 647.28 21600.36
P_eil101
24648.12
1251.32
433.92
552.88
651.2a 24382.36
P_kroA100
430.28
548.52
645.6 21566.12a 22437.28
1234
b
P_kroB100
545.68 643.52b 21495.4b 22340.56ab
1228.8
428.6
P_rat99
429.72
553.12
662.16 21915.24
22908.16
1242.16a
PACS-1
430.28
547.92
643.8 21625.08
22348.16
1226.4
PACS-2
428.56
546.28
645.2 21640.12
22377.28
1234.16
PACS-3
429.52
546.32
649.2 21641.12
22436.6
1231.88
PACS-4
429.64
547.84
647.4 21621.08
22446.4
1242.72
PACS-5
430.6
546.28
648.24 21589.24
22412
1237.12
PACS-6
430.44
546.8
647.88 21624.2
22440.8
1241.12
PACS-7
430.2
549.84
649.04 21584.28
22457.44
1238.96
PACS-8
430.2
549
649.76 21594.56
22578.88
1245
PACS-9
430.04
546.56
651.72 21590.68
22429.6
1249.64
ACS_GA-1
428.56
547.32
644.08 21674.48
22345.48
1227.48
ACS_GA-2 427.84
544.28
643.92 21669.24
22366.84
1219.88
ACS_GA-3 427.84
643.84 21594.36
22376.16
1228.28
543.56
ACS_GA-4 427.84
543.96
643.92 21514.24
22358.48
1219.2
ACS_GA-5 428.32
643.58 21589.88
22385
1222.84
543.56
ACS_GA-6
428.4
544.52
22353.16
1224.52
642.68 21524.84
ACS_GA-7 428.52
548.24
644.76 21586.16
22373.72
1225.96
ACS_GA-8
428.8
544.44
643.56 21514.8
22363.6
1222.52
ACS_GA-9 428.24
542.84
643.57 21582.36
22354.68
1223.84
“a” expected best value. “b” best value obtained among the new sets of parameters.
Values in bold are the best values obtained.
maintained as in (a) or frequently visited as in (b). This fast
convergence can be an advantage as well as a drawback
because it can lead to a fast non-desirable convergence. We
set the reasons of this fast convergence in the PSO
framework used and mainly in the method for evaluating a
set of parameters: in a stochastic algorithm there is the
probability that a bad set of parameters could perform well,
if all the particles move into this area and the number of
iterations in this area increases leading to probably good
solutions that will cause the algorithm to remain in this nonoptimal area.
Table II shows the optimum set of parameters found
running PSO-ACS on each one of the instances selected;
P_eil51 is the set of parameters obtained running PSO-ACS
on the instance eil51, similarly for P_eil76, P_eil101..and so
on.. PACS and ACS_GA are the set of parameters proposed
by [6] and [16] respectively, considering several values of φ
(parameter that defines the neighbourhood). “Fitness” is the
minimal length tour obtained by the PSO-ACS algorithm in
the related instance. There is no clear cut rule to define the
optimal parameters even if the recommended ranges can be
guessed, also β is normally bigger than ϕ; q0 is always
bigger than 0.5; φ, percentage of vertices in the neighbour,
eil51
eil76
eil101
kroA100
kroB100
rat99
P_eil51
539
639
21470
22404
1213
426a
P_eil76
637
21318
22237
426
538a
1211
P_eil101
23764
23774
1214
428
544
638a
P_kroA100
22237
1212
427
637
21343a
538
P_kroB100
539
22199a
426
629
21282
1211
P_rat99
427
547
644
21560
22422
1211a
PACS-1
427
540
634
22236
1212
21282
PACS-2
630
22200
1212
426
538
21282
PACS-3
427
634
22179
1212
538
21282
PACS-4
427
540
630
22251
1213
21282
PACS-5
540
630
22179
1214
426
21282
PACS-6
427
638
21305
22244
1213
538
PACS-7
427
540
636
22284
1213
21282
PACS-8
632
21292
22295
1221
426
538
PACS-9
427
540
639
22284
1213
21282
ACS_GA-1
427
539
632
21330
22157o
1211
ACS_GA-2
630
21370
22237
426
538
1211
ACS_GA-3
631
21330
22295
426
538
1211
ACS_GA-4
634
21318
22231
426
538
1211
ACS_GA-5
630
21380
22258
426
538
1211
ACS_GA-6
630
21318
22268
426
538
1211
ACS_GA-7
630
21318
22274
426
538
1211
ACS_GA-8
630
21318
22241
426
538
1211
ACS_GA-9
630
21318
22237
426
538
1211
“a” expected best value; “o” best value obtained. Values in bold mean
optimums. For the PACS and ACS_GA set of parameters PACS-i means
φ=0.i
is normally less or equal to the 20%; the number of ants are
normally between 1 and 10. But we cannot set a rule
because the set of parameters for P_kroB100 is not
following those guidelines and is performing better than the
rest. In all the instances the optimum is found but this is not
relevant because of the enormous quantity of time expended
(sometimes more than half a day) and because of the
number of times an ACS algorithm is run on a given
instance.
We have run each one of the obtained set of parameters in
all of the previous selected instances, to check the efficiency
of the parameters. As such ACS is performed with 25 trials
of 2500 iterations each.
Table III and IV show the performance, based on the
average tour length (trial-based) and the minimum tour
length found respectively, for each of the sets of parameters
on each of the instances. We expect that the parameters
related to an instance will compute the best results on this
instance. Comparing this hypothesis on Tables III and IV,
computational results show that this is in fact false. Instead,
working only in this reduced set of instances, the set of
parameters for P_kroB100 can be considered as the best
overall one which is performing optimally in most of the
6
TABLE V.
TIME COMPARISON BEST SETS OF PARAMETERS
Avg. FI
Avg. Time
P_kroB100
4,53
241,29
Best of PACS
1,35
69,53
Best of GA-ACS
3,16
162,89
“Avg. FI” is the average time (considering all refereed instances) used
for finding the best solution proposed by the algorithm, considering
the time from the trial this solution was computed. “Avg. Time” is the
average computational time of the algorithm running in all the
proposed instances. In bold are the best times. For PACS and
GA_ACS lowest times for all possible Φ values are used: are named
respectively “Best of PACS” and “Best of GA-ACS”.
REFERENCES
[1]
[2]
[3]
[4]
instances and with an efficient average value.
[5]
Comparing the sets of parameters of PACS and
ACS_GA, one can observe that PACS is performing better
in bigger instances and ACS_GA in smaller ones. Further,
ACS_GA performs better on the average.
Computational times of PACS, ACS_GA and P_kroB100
set of parameters are compared in Table V. PACS perform
best in both measures and P_kroA100 is the worst. The set
of parameters of P_kroB100 is performing better than ACS
and ACS_GA in most of the instances, but it has a greater
computational time.
[6]
[7]
[8]
[9]
[10]
[11]
V. CONCLUSIONS
Computational results seem to show that there is no
uniquely optimal set of ACS parameters yielding best
quality solutions in all the TSP instances. Nevertheless the
PSO-ACS has been able to find a set of ACS parameters
that work optimally for a majority of instances unlike others
known so far.
PSO-ACS algorithm works well across different instances
because it adapts itself to the instance characteristics. But it
has a high computational overhead. A future work will try
to modify the algorithm framework to reduce this cost.
PSO-ACS also has a fast convergence that can lead to a
bad set of parameters. This may be due to two reasons: first
is the specific PSO framework used, and in modifying it we
expect to obtain better results. Secondly the way the sets of
parameters are evaluated may have to be reviewed as a bad
set of parameters could lead to a non-desired convergence.
ACKNOWLEDGMENT
D. Gómez-Cabrero thanks the University of Colombo
School of Computing for the support given while he was in
Sri Lanka. The contribution by D. Gómez-Cabrero has been
partially supported by the AVCiT of the Generalitat
Valenciana (Ref: GRUPOS03/174).
[12]
[13]
[14]
[15]
[16]
[17]
B. Adenso-Díaz and M. Laguna, “Fine-tuning of Algorithms Using
Fractional Experimental Designs and Local Search”, to appear in
Operations Research.
R. S. Barr, B. L. Golden, J. P. Kelly, M. G. C. Resende and W. R.
Stewart, “Designing and Reporting on Computational Experiments
with Heuristic Methods”, Journal of Heuristics 1:1 9-32 1995.
Mauro Birrattari, Thomas Stützle, Luis Paquete and Klaus
Varrentrapp,
“A
Racing
Algorithm
For
Configuring
Metaheuristics”. In W. B. Langdon et al editors, GECCO 2002:
Proceedings of the Genetic and Evolutionary Computation
Conference, pp. 11--18, Morgan Kaufmann Publishers, San
Francisco, CA, USA, 2002.
S. P. Coy, B. L. Golden, G. C. Runger and E. A. Wasil , “Using
Experimental Design to Find Effective parameter settings for
heuristics,” Journal of Heuristics, vol. 7, nº 1, pp. 77-97, 2001.
M. Dorigo, Optimization, Learning, and Natural Algorithms (in
Italian). PhD Thesis, Dip. Elettronica, Politecnico di Milano, 1992.
M. Dorigo and L.M. Gambardella, “Ant Colony System: A
Cooperative Learning Aproach to the Travelling Salesman Problem,”
IEEE Transactions on Evolutionary Computation, 1(1):53-66, 1997.
M. Dorigo, V. Maniezzo, and A. Colorni. Positive Feedback as a
Search Strategy. Technical Report
91-016, Dip. Elettronica,
Politecnico di Milano, 1991.
M. Dorigo, V. Maniezzo, and A. Colorni, “The Ant System:
Optimization by a Colony of Cooperating Agents,” IEEE
Transactions on Systems, Man and Cybernetics, Part B, 26(1):29-42,
1996.
M. Dorigo and T. Stützle, Ant Colony Optimization, MIT Press, 2004.
R.C. Eberhart, P.K. Simpson and R.W. Robbins, Computational
Intelligence PC Tools, Academic Press Professional: Boston 1996.
L.M. Gambardella and M. Dorigo, “Ant-Q: A reinforcement Learning
Approach to the Symmetric and Asymmetric Travelling Salesman
Problems”, in Proceedings of the IEEE International Conference on
Evolutionary Computation (ICEC’96), pages 252-260. Morgan
Kaufmann, 1995.
D. Gies and Y. Rahmat-Samii, “Particle Swarm Optimization for
Reconfigurable Phase Differentiated Array Design,” Microwave and
Optical Technology Letters, vol. 38, nº 3, pp. 168-175, 2003.
J. Kennedy and R.C. Eberhart, “Particle Swarm Optimization,” Proc.
Of the IEEE International Conference on Neural Networks,
Piscataway, NJ, USA, pp. 1942-1948, 1995.
J. Kennedy and R.C. Eberhart, Swarm Intelligence, Morgan
Kaufmann Publishers, 2001.
E.C. Laskari, K.E. Parsopoulos and M.N. Vrahatis, “Particle Swarm
Optimization for Integer Programming,” Proceedings of the 2002
IEEE Congress on Evolutionary Computation, pp. 1582-1587, 2002.
M. L. Pilat and T. White, “Using Genetic Algorithms to optimize
ACS-TSP,” Proceedings of the Third International Workshop on Ant
Algorithms, pages 282-287, 2002.
G. Reinelt, “TSPLIB- A travelling Salesman Problem Library for TSP
Applications,” ORSA Journal of Computing, 3:376-384, 1991.
| 9cs.NE
|
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
Systematic and Integrative Analysis of Proteomic
Data using Bioinformatics Tools
Rashmi Rameshwari
Dr. T. V. Prasad
Asst. Professor, Dept. of Biotechnology,
Manav Rachna International University,
Faridabad, India
Dean (R&D), Lingaya‟s University,
Faridabad, India
Abstract— The analysis and interpretation of relationships
between biological molecules is done with the help of networks.
Networks are used ubiquitously throughout biology to represent
the relationships between genes and gene products. Network
models have facilitated a shift from the study of evolutionary
conservation between individual gene and gene products towards
the study of conservation at the level of pathways and complexes.
Recent work has revealed much about chemical reactions inside
hundreds of organisms as well as universal characteristics of
metabolic networks, which shed light on the evolution of the
networks. However, characteristics of individual metabolites
have been neglected in this network. The current paper provides
an overview of bioinformatics software used in visualization of
biological networks using proteomic data, their main functions
and limitations of the software.
Keywords- Metabolic network; protein interaction network;
visualization tools.
I.
INTRODUCTION
Molecular interaction network visualization is one of the
most user-friendly features developed for the simulation
process of biological interactions [29]. Drawing of any
molecule for example, protein may seem to be easy but
generating the same protein with all the types of conformation
that it can attain during any interactions and to simulate this
process is quite difficult. In this context one of the greatest
manually produced molecular structures of its time was done
by Kurt Kohn's 1999 map of cell cycle control. Protein
interactions network visualization deals with territory that is
very similar to that of protein-protein interaction prediction, but
differs in several key ways. proteomics data are often
associated with pathways or protein interactions, and both of
these are easily visualized as networks[22]. Even types of data
not normally viewed as networks (e.g. microarray results) are
often painted onto signaling, metabolic, or other pathways or
protein interaction networks for visualization and analysis.
Visualization and analysis tools are commonly used to
interact with proteomic data. Most, visualization tools were
developed simply for illustrating the big picture represented by
protein-protein interaction data or expression data, for
qualitative assessment, not necessarily for quantitative analysis
or prediction. Expression and interaction experiments tend to
be on such a large scale that it is difficult to analyze them, or
indeed grasp the meaning of the results of any analysis. Visual
representation of such large and scattered quantities of data
allows trends that are difficult to pinpoint numerically to stand
out and provide insight into specific avenues of molecular
functions and interactions that may be worth exploring first out
of the bunch, either through confirmation or rejection and then
later of significance or insignificance to the research problem at
hand. With a few recent exceptions, visualization tools were
not designed with the intent of being used for analysis so much
as to show the workings of a molecular system more clearly.
Visualization tools also do not actually predict molecular
interactions themselves or their characteristics. On the contrary,
visualization tools only create a graphical representation of
what is already "known" in literature and in molecular
interaction repositories such as the Gene Ontology (GO) [7]. A
side-effect of displaying interaction networks by treating some
proteins as members of more general family groupings or going
by interactions in different tissues or as analogues in other
tissues or organisms is the apparent display of a protein-protein
or other molecular interaction that is inferred but may or may
not have actually been observed and documented, something
that could be misconstrued as a predicted interaction.
The most significant difference between molecular
interaction network visualization and molecular interaction
prediction is the nature of the information it provides [11].
Protein-protein interaction prediction is characterized by its
concern with how proteins will interact, where they will
interact, under what conditions they were interact, and what
parts are necessary for their interaction. These characteristics
are governed by physical and chemical properties of the
proteins or other molecules involved, which may be actual
molecules that have been described through extensive
proteomics experiments or hypothetical, in silico generated
species that are being investigated for pharmaceutical and other
applications. Interaction network visualization tools are given
no knowledge of physical or chemical properties of the proteins
that, why they interact. As a result, the information they
inadvertently impart concerns only whether or not certain
proteins putatively interact with certain other proteins, not how
they interact, when they interact, or why they interact.
29 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
biological systems. Comprehensive data of protein interactions
is also suitable for systems level evolutionary analysis.
There are many commercial software such as Ingenuity
Pathway Analysis (Figure 2), MetaCore and Pathway studio,
designed to visualize high-throughput data in the context of
biological networks. Biological networks have a scale-free and
modular organization. In a scale free networks the degree of
distribution follows a power-law, which means that only a
small number of nodes, called hubs are highly connected [12].
Hubs usually play essential roles in biological systems [13]. On
the other hand, groups of proteins with similar functions tend to
form clusters or modules in the network architecture. Many
commercial software for network visualization follow this law.
Metacore is an integrated knowledge database and software
suite for pathway analysis of experimental data and gene lists.
It is based on manually curetted databases of human proteinprotein, Protein-DNA and protein compound interactions. This
package includes easy to use, intuitive tools for search, data
visualization, mapping and exchange, biological networks and
interactomes [32].
Figure:(A)
Figure: (B)
Figure1: Comparative visualization of Protein-Network drawn by different
tools like Pajek (Fig.A), and Cytoscape (Fig. B). Here genes are represented
as nodes and interaction as edges.
II.
BACKGROUND AND RELATED WORK
The other software tool known as Pathway studio, is based
on a mammalian database, named: ResNet 5 mammalian,
which is generated by text mining of the PubMed database and
43 full text journals [14]. The advantages of using this tool are
that it increases the depth of analysis of high-throughput data
generation experiments like microarray gene expression,
proteomics, metabolomics. Enables data sharing in a common
analysis environment. This tool simplifies keeping up to date
with the literature and brings this knowledge into an analysis
environment. This also enables visualization of gene
expression values and status in the context of protein
interaction networks and pathways. However, free software
like Pathway Voyager (Figure 5), GenMapp and Cytoscape are
also available. Pathway Voyager applies flexible approach that
uses the KEGG database [27] to pathway mapping [17].
GenMapp (Figure 6) is also designed to visualize gene
expression data on maps representing biological pathways and
gene groupings. GenMapp has more option which can modify
or design new pathways and apply complex criteria for viewing
gene expression data on pathways [14].
With the recent advances in high-throughput technologies,
software tools have been developed to visualize and analyze
large-scale data. This paper deals with various visualization
techniques for proteomic data. Major emphasis is on network
graph generated during protein-protein interaction. Many tools
are being used for this purpose which is based on different
algorithm. For example, Pajek [6] and Cytoscape [8] use force
directed layout algorithm which produces graph by computing
force between pairs of nodes in each iteration of the
optimization process. The networks can be visualized as
indicated in Fig 1A and Fig 1B.
As protein interactions data also helps in study related to
evolutionary analysis. In a new era, it is necessary to
understand how the components which involve in biological
systems from the various biological data and knowledge of
components at molecular level. It reveals the structure of
biological systems and lead to “ontological” comprehension of
Figure 2: G-Protein signaling_Ras family GTPases in kinase cascades. Image
generated by Ingenuity Pathway Analysis.
30 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
Clustered graphs are common occurrences in the biological
field. Examples of pre-clustered graphs include clustering of
proteins or genes based on some biological functionality,
structural geometry, expression pattern or other chemical
property [11].
Figure 3: A Network generated by Metacore Software [32]
III.
LIMITATIONS AND FUTURE DEVELOPMENTS
After the completion of Human Genome Project, 36,000
genes were discovered, which has potential to synthesize more
than 1,00,000 proteins [19], less than 50% of these genes can
be assigned a putative biological function on the basis of
sequence data [18]. With the advancement of technology many
software has been designed to explore biological networks, like
protein interactions network, protein-DNA interactions etc.,
that are based on databases like DIP [31], MINT [30], of
mammalian database. The tools represented in this paper are
applicable to a wide range of problems and their distinct
features make them suitable for a wide range of applications.
Figure 4: Actin Cytoskeleton Regulation, network generated by Pathway
Studio
In any “post- omic” analysis two essential concepts must be
applied to understand biological functions at a systems level.
First integrate different levels of information and second, view
cells in terms of their underlying network structure. The
information about biological entity is scattered in different
databases [28]. Hence, the information retrieval from diverse
databases is done, which is bit time consuming. Current
databases are good for the analysis of a particular protein or
small interaction networks. But they are not as useful for
integration of complex information on cellular regulation
pathways, networks, cellular roles and clinical data and they
lack coordination and the ability to exchange information
between multiple data sources [20]. There is need of software
that can integrate information from different database as well
as from diverse sources.
To analyze data, at present many software are there like
Ingenuity Pathway analysis, Metacore and Pathway Studio.
They work on owner curetted database at the same time there
high price make them unaffordable for academic institute to
use them. At the same time Cytoscape which has many
properties for visualization of high throughput data, can be
alternative for users [15].
However network constructed with Cytoscape are
sometimes liable to show errors. So there is need to improve
quality of available curetted databases and also to develop
integrative knowledge bases that are especially designed to
construct biological networks.
Figure 5: Interactive KEGG Pathway display. The screenshot illustrates
KEGG pathway mapping for the glycolysis / gluconeogenesis pathway using
the predicted ORFeome of the GAMOLA annotated L. acidophilus NCFM
genome as query template [19].
31 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
Figure 6: Myometrial Relaxation and contraction Pathway, Image generated
by GenMapp.
A metabolic network is a reliable source of information.
The reconstruction of GRNs is largely promoted by advances
in high-throughput technologies, which enable to measure the
global response of a biological system to specific interventions.
For instance, large-scale gene expression monitoring using
DNA microarrays is a popular technique for measuring the
abundance of mRNAs. However, integration of different types
of „omics‟ data from genomics, proteomics and metabolomic
studies can be undertaken. Although the metabolic network has
important features for drug discovery, its use in case of human
is very limited [18]. Further, Proteomics may yield crucial
information on the regulation of biological functions and the
mechanism of diseases. In this sense it is a highly promising
area for drug discovery. Hence, additional efforts will be
required for metabolic network reconstruction and analysis.
IV.
CONCLUSION
From the above tools it can be concluded that metabolic
pathways stored as directed acyclic graphs can be considered a
basic concept for the visualization tool for metabolic pathway.
With respect to visualization, single network views provide
little more than brief glimpses of the large datasets.
Visualization tools need to support many different types of
views, each network view at a different level of detail.
Dynamic navigation from one view to another will be a key to
showing the connection between different views. Navigating
from one time series point to another, for instance, could
involve a view showing only the differences between the two
time points. If the time points are consecutive, the number of
differences will tend to be quite small. A similar approach
could be applied to sub-cellular localization information as
well. To adequately address each of these issues, active
cooperation is required between a variety research fields
including graph drawing, information visualization, network
analysis and of course biology. Though all the mentioned tool
differs significantly in its approach for pathway
reconstructions. Hence for future a tool is needed which
describe all pathways that make up a cell and how they interact
as a system in the overall physiology of an organism is
required.
Figure 7: Pathway voyager mapping procedure [17].
REFERENCES
[1]
Lenzerini M: Data Integration: A Theoretical Perspective. PODS 2002 ,
243-246.
[2] B., Bar-Joseph, Z., Gerber, G.K., Lee, T.I., Rinaldi, N.J., Yoo, J.Y.,
Robert, F., Gordon, D.Fraenkel, E., Jaakkola, T.S., Young, R.A.,
Gifford, D.K. (2003). Computational discovery of gene module and
regulatory networks. Nat Biotechnol. 21(11):1337-42.
[3] Fields S, Song O: A novel genetic system to detect protein-protein
interactions. Nature 1989, 340:245-246.
[4] Spirin, V. and Mirny, L.A. (2003) PNAS, 100,12123-128
[5] Jeong H, Mason SP, Barabasi AL, Oltvai ZN: Lethality and centrality in
protein networks. Nature 2001, 411:41-42.
[6] Batagelj V, Mrvar A: Pajek – Program for Large Network Analysis.
Connections 1998, 21:47-57.
[7] The Gene Ontology Consortium: Gene Ontology: tool for the unification
of biology. Nat Genet 2000, 25:25-29.
[8] Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, Amin
N, Schwikowski B, Ideker T: Cytoscape: a software environment for
integrated models of biomolecular interaction networks. Genome Res
2003, 13(11):2498-2504.17
[9] Hermjakob H, Montecchi-Palazzi L, Lewington C, Mudali S, Kerrien S,
Orchard S, Vingron M, Roechert B, Roepstorff P, Valencia A, Margalit
H, Armstrong J, Bairoch A, Cesareni G, Sherman D, Apweiler R:
IntAct: an open source molecular interaction database. Nucleic Acids
Res 2004, (32 Database):D452-455
[10] Alfarano, C., Andrade, C.E., Anthony, K., Bahroos, N., Bajec, M.,
Bantoft, K., Betel, D., Bobechko, B., Boutilier, K., Burgess, E., et al.
(2005) The Biomolecular Interaction Network Database and related tools
2005 update Nucleic Acids Res, . 33, D418–D42.
[11] Rashmi Rameshwari and Dr. T. V. Prasad, “A Survey of Various
Protein-Protein Interaction Tools”, Proc. of National Conference on
32 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
Advances in Knowledge Management NCAKM ‟10, Lingaya‟s
University, Faridabad, 2010
Barabasi AL, Oltvai ZN. Network biology: understanding the cell‟s
functional organization. Nat Rev Genet 2004;5:101-13
Han JD, Bertin N, Hao T, Goldberg DS, Berriz GF, Zhang LV, et al.
Evidence for dynamically organized modularity in the yeast proteinprotein interaction network. Nature 2004;430:88-93.
Dahlquist KD, Salomonis N, Vranizan K, Lawlor SC, Conklin BR.
GenMAPP, a new tool for viewing and analyzing microarray data on
biological pathways. Nat Genet 2002;31:19-20.
Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, et al.
Cytoscape: A Software Environment for Integrated Models of
Biomolecular Interaction Networks. Genome Res 2003;13:2498-504.
Daraselia N, Yuryev A, Egorov S, Novichkova S, Nikitin A, Mazo I.
Extracting human protein interactions from MEDLINE using a fullsentence parser. Bioinformatics 2004;20:604-11.
Altermann E, Klaenhammer TR. PathwayVoyager: pathway mapping
using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.
BMC Genomics 2005;6:60.
Overington JP, Al Lazikani B, Hopkins AL. How many drug targets are
there? Nat Rev Drug Discov 2006;5:993-6.
Lander ES, Linton LM, Birren B, Nusbaum C, Zody MC, Baldwin J, et
al. Initial sequencing and analysis of the human genome. Nature
2001;409:860-921.
Tucker CL, Gera JF, Uetz P. Towards an understanding of complex
protein networks. Trends in Cell Biology 2001;11:102-6.
Funahashi,A. et al. (2003) CellDesigner: a process diagram editor for
gene-regulatory and biochemical networks. Biosilico, 1, 159–162.
Kell,D.B. (2006) Systems biology, metabolic modelling and
metabolomics in drug discovery and development. Drug Discovery
Today, 11, 1085–1092.
Kitano,H. et al. (2005) Using process diagrams for the graphical
representation of biological networks. Nat Biotechnol., 23, 961–966.
Nikitin,A. et al. (2003) Pathway studio –- the analysis and navigation of
molecular networks. Bioinformatics, 19, 2155–2157.
Altermann E, Klaenhammer TR: GAMOLA: a new local solution for
sequence annotation and analyzing draft and finished prokaryotic
genomes. OMICS 2003, 7:161-169.
[26] Goesmann A, Haubrock M, Meyer F, Kalinowski J, Giegerich R: PathFinder: reconstruction and dynamic visualization of metabolic pathways.
Bioinformatics 2002, 18:124-129.
[27] Ogata,H., Goto,S., Sato,K., Fujibuchi,W., Bono,H. andKanehisa,M.
(1999) Kegg: Kyoto encyclopedia of genes and genomes. Nucleic Acids
Res., 27, 29–34.
[28] Alfarano,C. et al. (2005) The biomolecular interaction network database
and related tools 2005 update. Nucleic Acids Res., 33, 418–424.
[29] Kohn,K.W. et al. (2006) Molecular interaction maps of bioregulatory
networks: a general rubric for systems biology. Mol. Biol. Cell, 17, 1–
13.
[30] Zanzoni,A. et al. (2002) MINT: a molecular interaction database. FEBS
Lett., 513, 135–140.
[31] Xenarios,I. et al. (2000) DIP: the database of interacting proteins.
Nucleic Acids Res., 28, 289–291.
[32] MetaCore Pathway Analysis Software, available at www.genego.com
AUTHORS PROFILE
Ms. Rashmi Rameshwari has received two master‟s degree from one from
T.M.B.U., Bhagalpur, Bihar and other from Jamia Hamdard, New Delhi in the
area of biotechnology and bioinformatics respectively. She is currently
associated with Manav Rachna International University as Assistant Professor
in Dept. of Biotechnology. Her research interests include systems biology,
proteomics, Microarray Technology, Chemoinformatics, etc.
Dr. T. V. Prasad received his master‟s degree in Computer Science from
Nagarjuna University, AP India and doctoral degree from Jamia Milia Islamia
University, New Delhi, India. With over 16 years of academic and
professional experience, he has a deep interest in planning and executing
major IT projects, pursuing research interest in CS/IT and bioinformatics. He
has authored over 65 publications in reputed journals and conferences. He has
also authored 4 books. He has also held respectable positions such as Deputy
Director with Bureau of Indian Standards, New Delhi. His areas of interest
include bioinformatics, artificial intelligence, consciousness studies, computer
organization and architecture. He is a member of reputed bodies like Indian
Society of Remote Sensing, Computer Society of India, APBioNet,
International Association of Engineers, etc.
33 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
Table 1
COMPARATIVE STATEMENT OF VARIOUS NETWORK VISUALIZATION AND ANALYSIS SOFTWARE
S.
N
o.
1
Parameters/
Features
Ingenuity
Pathway Analysis
Metacore
Pathway Studio
GenMAPP
Cytoscape
Pathway
Voyager
PathFinder
Developed
by
Ingenuity Systems
Inc.
GeneGo Inc.
Ariadne
Genomics Inc.
Gladstone
Institutes
Description
Database of
Biological
networks created
from millions of
relationships,
between proteins,
genes ,complexes,
cells, tissues,
drugs and diseases
A manually
curetted
database of
human
Proteinprotein
interaction &
Protein-DNA
interactions,
transcriptional
factors,
Signaling,
metabolism
and bioactive
molecules
Archived
Maps were
drawn based
on textbooks,
articles and
public
pathway
databases or
generated
from the
public
database
maintained by
the Gene
Ontology
Project
3
4
Availability
Based on
database
Commercial
Ingenuity
pathways
knowledge base
Commercial
Human
Database
Public
KEGG
Public
KEGG
Public
KEGG
Public
RDBMS is usec
along with
KEGG
5
6
7
Web access
Platform
Special
features
Enabled
Java 1.5 or higher
Solution is given
to Pharmaceutical,
Bio-technology,
and Academics
Enabled
Java
has graphics
tools for
constructing
and modifying
pathways.
Used for
analyzing
microarray
data include
statistical
filters and
pattern finding
algorithms
such as
hierarchical
clustering
Enabled
Java
mRNA Expression
profiles, Gene
annotations from
Gene Ontology
(GO) & KEGG.
Incorporates
statistical analysis
Enabled
Perl/TK
No dedicated
hardware or
software are
necessary to
analyze given
datasets
Enabled
Perl/TK
Aim at
comparing
pathways at
microscopic
level and
therefore it can
be used for
dynamic
visualizations of
metabolisms
from a whole
genome
perspective
8
Drawbacks
Limited to Human,
Mouse, Rat and
Canine
Enabled
Java
unique ability
to
concurrently
visualize
multiple types
of
experimental
data
such as gene
expression,
proteomic,
metabolomics,
SAGE, MPSS,
SNP, HCS,
HTS,
microRNA
and
clinical and
pre-clinical
phenotypic
data
Specific server
is required
These databases
are a collection of
eukaryotic
molecular
interactions
generated by
MedScan Text to
knowledge suit
using the entire
PubMed database
and 43 full text
journals. Also
works with public
database of
signaling and
biochemical
pathways
Commercial
Mammalian
Database: ResNet
5, ResNet plant
database, KEGG,
BIND, HPRD
Enabled
Python
Analyze
proteomic,
Metabolomics
and other high
throughput data.
Eric Altermann
and Todd R
Klaenhammer
Utilizes the
KEGG online
database for
pathway
mapping of
partial and
whole
prokaryotic
genomes
Alexander
Goesmann et.al
2
Paul Shanon,
Andrew Markiel et
al.
Software for
integrating
biomolecular
interaction
Networks with
high-throughput
expression data
and other
molecular states
into a unified
conceptual
framework
Generally
used to
explore
Microarray
data
Focuses only on
high-level
networks, lowlevel models of
components and
interactions
addressed by
ongoing projects
such as Ecell
(Tomita et al.
1999), VirtualCell;
mechanisms
for bridging highlevel interactions
with lower level,
physico-chemical
For certain
selectable
pathways (e.g.
Ribosomal
reference
pathway)
KEGG does not
yet support
organism
independent
marking. For
practical
reasons, no hits
will be displayed
for these
pathways
Uses RDBMS
based internet
application.
Being integrated
into the locally
developed
genome
annotation
system GENDB
with extended
functionality
It offers a wizard
interface for
creating very
simple network
and data queries
and only
Biological
Networks
provides a
language
interface for
expressing such
queries.
A tool for the
dynamic
visualization of
metabolic
pathways based
on annotation
data.
34 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 5, 2011
S.
N
o.
9
10
11
12
13
14
15
16
17
Parameters/
Features
Ingenuity
Pathway Analysis
GenMAPP
Cytoscape
Pathway
Voyager
PathFinder
GAMOLA
GENDB
Integrated
with
Storing
result
Data format
Only based on
Ingenuity Product
Present
Gene Go
MedScan
MAPP
Present
Present
Present
Present
Present
Present
Not Specific.
Not Specific
ResNet Exchange
XML formats.
PSI-MI,
GML/XGMML
Fasta Files
EMBL OR
Genbank
Graph
comparison
for species
Graphical
user
Interface
Visualization
Technique
Ease of use
and report
generation
Graphical
Representati
on
Classificatio
n Technique
Present
Present
Present
CSV, GPML,
WikiPathways
, MAPP
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Present
Excellent
Excellent
Excellent
Excellent
Very Good
Good
Good
Present
Present
Present
Present
Present
Present
Present
Compare affected
pathways and
phenotypes across
time, dose, or
patient population.
Disease,
Tissue,
Species, Sub
cellular
localization,
Interactions,
Metabolites
Internet
Explorer 6.0
or higher
4 GB
Interrogate
Different species,
Multiple genomes
Hierarchical
clustering
None
Hierarchical
clustering
Chunks and
Subway
Internet Explorer
6.0 or higher
Internet Explorer
6.0 or higher
512 Mb
Internet
Explorer 6.0 or
higher
512 Mb
Internet Explorer
6.0 or higher
1 GB
Internet
Explorer 6.0
or higher
512 Mb
Linux 2.1, 3.0
or Red Hat 9.0
Windows
Windows
Window
Windows
UNIX, Windows
www.genego.c
om
www.ariadnegena
mics.com
www.genmap
p.org/
www.cytoscape.or
g
www.bioinform
at
ics.ai.sri.com/pt
ools
http://bibi
serv.TechFak
.UniBielefeld
.DE/pathfind
er/
[32]
[24]
[14]
[15]
[17]
[25], [26]
Web
browser
Internet Explorer
6.0 or higher
19
Memory
20
Operating
System
supported
Reference
URL
512 Mb
(minimum), 1GB
(Recommended)
Vista, Window XP
and Macintosh
10.3,10.4, 10.5
www.ingenuity.co
m/
References
[12]
22
Pathway Studio
models of specific
biological
processes are
required
Numerous
18
21
Metacore
512 Mb
35 | P a g e
www.ijacsa.thesai.org
| 5cs.CE
|
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY
COMPACT GROUPS
arXiv:1702.07955v1 [math.GR] 25 Feb 2017
FRIEDRICH MARTIN SCHNEIDER
Abstract. We note a generalization of Whyte’s geometric solution to the von
Neumann problem for locally compact groups in terms of Borel and clopen
piecewise translations. This strengthens a result of Paterson on the existence
of Borel paradoxical decompositions for non-amenable locally compact groups.
Along the way, we study the connection between some geometric properties of
coarse spaces and certain algebraic characteristics of their wobbling groups.
1. Introduction
In his seminal article [18] von Neumann introduced the concept of amenability for groups in order to explain why the Banach-Tarski paradox occurs only for
dimension greater than two. He proved that a group containing an isomorphic
copy of the free group F2 on two generators is not amenable. The converse, i.e.,
the question whether every non-amenable group would have a subgroup being isomorphic to F2 , was first posed in print by Day [5], but became known as the von
Neumann problem (or sometimes von Neumann-Day problem). The original question has been answered in the negative by Ol’šanskiı̆ [19]. However, there are very
interesting positive solutions to variants of the von Neumann problem in different
settings: a geometric solution by Whyte [27], a measure-theoretic solution by Gaboriau and Lyons [9] and its generalization to locally compact groups by Gheysens
and Monod [14], as well as a Baire category solution by Marks and Unger [13].
Whyte’s geometric version reads as follows.
Theorem 1.1 (Theorem 6.2 in [27]). A uniformly discrete metric space of uniformly bounded geometry is non-amenable if and only if it admits a partition whose
pieces are uniformly Lipschitz embedded copies of the 4-regular tree.
In particular, the above applies to Cayley graphs of finitely generated groups
and in turn yields a geometric solution to the von Neumann problem.
The aim of the present note is to extend Whyte’s relaxed version of the von
Neumann conjecture to the realm of locally compact groups. For this purpose, we
need to view the result from a slightly different perspective. Given a uniformly
discrete metric space X, its wobbling group (or group of bounded displacement ) is
defined as
W (X) := {α ∈ Sym(X) | ∃r ≥ 0 ∀x ∈ X : d(x, α(x)) ≤ r}.
Wobbling groups have attracted growing attention in recent years [11, 12, 3]. Since
the 4-regular tree is isomorphic to the standard Cayley graph of F2 , one can easily
Date: 28th February 2017.
2010 Mathematics Subject Classification. Primary 22D05, 43A07, 20E05, 20F65.
This research has been supported by funding of the German Research Foundation (reference
no. SCHN 1431/3-1) as well as by funding of the Excellence Initiative by the German Federal and
State Governments.
1
2
FRIEDRICH MARTIN SCHNEIDER
reformulate Whyte’s in terms of semi-regular subgroups. Let us recall that a subgroup G ≤ Sym(X) is said to be semi-regular if no non-identity element of G has
a fixed point in X.
Corollary 1.2 (Theorem 6.1 in [27]). A uniformly discrete metric space X of
uniformly bounded geometry is non-amenable if and only if F2 is isomorphic to a
semi-regular subgroup of W (X).
For a finitely generated group G, the metrics generated by any two finite symmetric generating sets containing the neutral element are equivalent and hence give
rise to the very same wobbling group W (G). It is easy to see that W (G) is just the
group of piecewise translations of G, i.e., a bijection α : G → G belongs to W (G) if
and only if exists a finite partition P of G such that
∀P ∈ P ∃g ∈ G :
α|P = λg |P .
Furthermore, we note that the semi-regularity requirement in the statement above
cannot be dropped: in fact, van Douwen [6] showed that W (Z) contains an isomorphic copy of F2 , despite Z being amenable. As it turns out, F2 embeds into the
wobbling group of any coarse space of positive asymptotic dimension (see Proposition 4.3 and Remark 4.4).
We are going to present a natural counterpart of Corollary 1.2 for general locally
compact groups. Let G be a locally compact group. We call a bijection α : G → G
a clopen piecewise translation of G if there exists a finite partition P of G into
clopen subsets such that
∀P ∈ P ∃g ∈ G :
α|P = λg |P ,
i.e., on every member of P the map α agrees with a left translation of G. It is
easy to see that the set C (G) of all clopen piecewise translations of G constitutes
a subgroup of the homeomorphism group of the topological space G and that the
mapping Λ : G → C (G), g 7→ λg embeds G into C (G) as a regular, i.e., semi-regular
and transitive, subgroup. Similarly, a bijection α : G → G is called a Borel piecewise
translation of G if there exists a finite partition P of G into Borel subsets with
∀P ∈ P ∃g ∈ G :
α|P = λg |P .
Likewise, the set B(G) of all Borel piecewise translations of G is a subgroup of the
automorphism group of the Borel space of G and contains C (G) as a subgroup.
For a locally compact group G, both B(G) and C (G) are reasonable analogues
of the wobbling group. Yet, the mere existence of an embedding of F2 as a semiregular subgroup of B(G), or even C (G), does not prevent G from being amenable.
In fact, there are many examples of compact (thus amenable) groups that admit
F2 as a (non-discrete) subgroup and hence as a semi-regular subgroup of C (G).
For example, since F2 is residually finite, it embeds into the compact group formed
by the product of its finite quotients. Therefore, we have to seek for a topological
analogue of semi-regularity, which amounts to a short discussion.
Remark 1.3. Let X be a set. A subgroup G ≤ Sym(X) is semi-regular if and only
if there exists a (necessarily surjective) map ψ : X → G such that ψ(gx) = gψ(x)
for all g ∈ G and x ∈ X. Obviously, the latter implies the former. To see the
converse, let σ : X → X be any orbit cross-section for the action of G on X, i.e.,
σ(X) ∩ Gx = {σ(x)} for every x ∈ X. Since G is semi-regular, for each x ∈ X there
is a unique ψ(x) ∈ G such that ψ(x)σ(x) = x. For all g ∈ G and x ∈ X, we have
gψ(x)σ(gx) = gψ(x)σ(x) = gx = ψ(gx)σ(gx),
which readily implies that ψ(gx) = gψ(x). So, ψ : X → G is as desired.
The purpose of this note is to show the following.
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
3
Theorem 1.4. Let G be a locally compact group. The following are equivalent.
(1) G is not amenable.
(2) There exist a homomorphism ϕ : F2 → C (G) and a Borel measurable map
ψ : G → F2 such that ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 .
(3) There exist a homomorphism ϕ : F2 → B(G) and a Borel measurable map
ψ : G → F2 such that ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 .
We remark that any map ϕ as in (2) or (3) of Theorem 1.4 has to be injective.
In view of the discussion above, we also note that for finitely generated discrete
groups the statement of Theorem 1.4 reduces to Whyte’s geometric solution to the
von Neumann problem. More specifically, the existence of a map ψ as in (2) or (3)
above may be thought of as a Borel variant of the semi-regular embedding condition
in Corollary 1.2. In general, we cannot arrange for ψ to be continuous, as there
exist non-amenable connected locally compact groups
Both (2) and (3) of Theorem 1.4 may be considered relaxed versions of containing
F2 as a discrete subgroup: according to a result of Feldman and Greenleaf [7], if
H is a σ-compact metrizable closed (e.g., countable discrete) subgroup of a locally
compact group G, then the right coset projection G → H \ G, x 7→ Hx admits a
Borel measurable cross-section τ : H \G → G, and hence the H-equivariant map
ψ : G → H,
x 7→ xτ (Hx)−1
is Borel measurable, too. This particularly applies if H ∼
= F2 is discrete.
The proof of Theorem 1.4 combines a result of Rickert resolving the original
von Neumann problem for almost connected locally compact groups (Theorem 3.3)
with a slight generalization of Whyte’s result for coarse spaces (Theorem 2.2) and
in turn refines an argument of Paterson proving the existence of Borel paradoxical
decompositions for non-amenable locally compact groups [21]. In fact, Theorem 1.4
implies Paterson’s result [21].
Corollary 1.5 (Paterson [21]). A locally compact group G is non-amenable if and
only if it admits a Borel paradoxical decomposition, i.e., there exist finite partitions
P and Q of G into Borel subsets and gP , hQ ∈ G (P ∈ P, Q ∈ Q) such that
[
[
G = · gP P ∪· · hQ Q.
P ∈P
Q∈Q
This note is organized as follows. Building on some preparatory work concerning
coarse spaces done in Section 2, we prove Theorem 1.4 in Section 3. Since our
approach to proving Theorem 1.4 involves wobbling groups, and there has been
recent interest in such groups, we furthermore include some complementary remarks
about finitely generated subgroups of wobbling groups in Section 4.
2. Revisiting Whyte’s result
Our proof of Theorem 1.4 will make use of Whyte’s argument [27] – in the form
of Corollary 2.3. More precisely, we will have to slightly generalize his result from
metric spaces to arbitrary coarse spaces. However, this will just require very minor
adjustments, and we only include a proof for the sake of completeness.
For convenience, let us recall some terminology from coarse geometry as it may
be found in [24]. For a relation E ⊆ X × X on a set X and x ∈ X, A ⊆ X, let
[
E[x] := {y ∈ X | (x, y) ∈ E},
E[A] := {E[z] | z ∈ A}.
A coarse space is a pair (X, E ) consisting of a set X and a collection E of subsets
of X × X (called entourages) such that
• the diagonal ∆X = {(x, x) | x ∈ X} belongs to E ,
4
FRIEDRICH MARTIN SCHNEIDER
• if F ⊆ E ∈ E , then also F ∈ E ,
• if E, F ∈ E , then E ∪ F, E −1 , E ◦ F ∈ E .
A coarse space (X, E ) is said to have bounded geometry if
∀E ∈ E ∀x ∈ X :
E[x] is finite,
and (X, E ) has uniformly bounded geometry if
∀E ∈ E ∃m ≥ 0 ∀x ∈ X :
|E[x]| ≤ m.
Among the most important examples of coarse spaces are metric spaces: if X is
a metric space, then we obtain a coarse space (X, EX ) by setting
EX := {E ⊆ X × X | sup{d(x, y) | (x, y) ∈ E} < ∞}.
Another crucial source of examples of coarse spaces is given by group actions.
Indeed, if G is a group acting on a set X, then we obtain a coarse space (X, EG ) of
uniformly bounded geometry by
EG := {R ⊆ X × X | ∃E ⊆ G finite : R ⊆ {(x, gx) | x ∈ X, g ∈ E}}.
Note that the coarse structure induced by a finitely generated group G acting on
itself by left translations coincides with the coarse structure on G generated by the
metric associated with any finite symmetric generated subset of G containing the
neutral element.
Now we come to amenability. Adopting the notion from metric coarse geometry,
we call a coarse space (X, E ) of bounded geometry amenable if
∀θ > 1 ∀E ∈ E ∃F ⊆ X finite, F 6= ∅ :
|E[F ]| ≤ θ|F |,
which is (easily seen to be) equivalent to saying that
∃θ > 1 ∀E ∈ E ∃F ⊆ X finite, F 6= ∅ :
|E[F ]| ≤ θ|F |.
This definition is compatible with the existing notion of amenability for group
actions (Proposition 2.1). Recall that an action of a group G on a set X is amenable
if the space ℓ∞ (X) of all bounded real-valued functions on X admits a G-invariant
mean, i.e., there exists a positive linear functional µ : ℓ∞ (X) → R with µ(1) = 1
and µ(f ◦ g) = µ(f ) for all f ∈ ℓ∞ (X) and g ∈ G.
Proposition 2.1 (cf. Rosenblatt [25]). An action of a group G on a set X is
amenable if and only if the coarse space (X, EG ) is amenable.
Proof. Generalizing Følner’s work [8] on amenable groups, Rosenblatt [25] showed
that an action of a group G on a set X is amenable if and only if
∀θ > 1 ∀E ⊆ G finite ∃F ⊆ X finite, F 6= ∅ :
|EF | ≤ θ|F |,
which is easily seen to be equivalent to the amenability of (X, EG ).
Let us turn our attention towards Theorem 1.1. A straightforward adaptation of
Whyte’s original argument readily provides us with the following only very slight
generalization (Theorem 2.2). For a binary relation E ⊆ X × X, we will denote the
associated undirected graph by
Γ(E) := (X, {{x, y} | (x, y) ∈ E}).
Furthermore, let gr(f ) := {(x, f (x)) | x ∈ X} for any map f : X → Y . Our proof
of Theorem 2.2 will utilize the simple observation that, for a map f : X → X, the
graph Γ(gr(f )) is a forest, i.e., it contains no cycles, if and only if f has no periodic
points, which means that P (f ) := {x ∈ X | ∃n ≥ 1 : f n (x) = x} is empty.
Theorem 2.2. Let d ≥ 3. A coarse space (X, E ) of bounded geometry is nonamenable if and only if there is E ∈ E such that Γ(E) is a d-regular forest.
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
5
Proof. (⇐=) Due to a very standard fact about isoperimetric constants for regular
trees [2, Example 47], if E ⊆ X × X is symmetric and Γ(E) is a d-regular tree, then
|E[F ]| ≥ (d − 1)|F | for every finite subset F ⊆ X. Of course, this property passes
to d-regular forests, which readily settles the desired implication.
(=⇒) Suppose that (X, E ) is not amenable. Then there is a symmetric entourage
E ∈ E such that |E[F ]| ≥ d|F | for every finite F ⊆ X. Consider the symmetric
relation R := E \ ∆X ⊆ X × X. Since |R[x]| < ∞ for every x ∈ X and
|R[F ]| ≥ |E[F ] \ F | ≥ |E[F ]| − |F | ≥ (d − 1)|F |
for every finite subset F ⊆ X, the Hall harem theorem [1, Theorem H.4.2] asserts
that there exists a function f : X → X with gr(f ) ⊆ R and |f −1 (x)| = d − 1 for
all x ∈ X. Notice that f does not have any fixed points as R ∩ ∆X = ∅. Since
the set of f -orbits of its elements
S partitions the set P (f ), we may choose a subset
P0 ⊆ P (f ) such that P (f ) = · x∈P0 {f n (x) | n ∈ N}. Furthermore, choose functions
g, h : P0 × N → X such that, for all x ∈ P0 and n ≥ 1,
• g(x, 0) = x and h(x, 0) = f (x),
• {g(x, n), h(x, n)} ∩ P (f ) = ∅,
• f (g(x, n)) = g(x, n − 1) and f (h(x, n)) = h(x, n − 1).
It follows that g and h are injective functions with disjoint ranges. Now we define
f∗ : X → X by setting
g(z, n + 2) if x = g(z, n) for z ∈ P0 and even n ≥ 0,
g(z, n − 2) if x = g(z, n) for z ∈ P and odd n ≥ 3,
0
f∗ (x) :=
f 2 (x)
if
x
=
h(z,
n)
for
z
∈
P
0 and n ≥ 2,
f (x)
otherwise
for x ∈ X. We observe that
gr(f∗ ) ⊆ gr(f 2 )−1 ∪ gr(f 2 ) ∪ gr(f ).
In particular, gr(f∗ ) ⊆ E ◦ E and therefore gr(f∗ ) ∈ E . Moreover, it follows that
P (f∗ ) ⊆ P (f ). However, for every x ∈ P (f ), there exists a smallest m ∈ N such that
f m (x) ∈ P0 , and we conclude that f∗m+1 (x) = f∗ (f m (x)) = g(f m (x), 2) ∈
/ P (f ) and
hence f∗m+1 (x) ∈
/ P (f∗ ), which readily implies that x ∈
/ P (f∗ ). Thus, P (f∗ ) = ∅.
In particular, f∗ has no fixed points. Furthermore,
(f −1 (x) ∪ {g(z, n − 2)}) \ {g(z, n + 1)} if x = g(z, n) for z ∈ P0
and even n ≥ 2,
(f −1 (x) ∪ {g(z, n + 2)}) \ {g(z, n + 1)} if x = g(z, n) for z ∈ P0
and odd n ≥ 1,
−1
f∗ (x) =
−1
(f
(x)
∪
{h(z,
n
+
2)})
\
{h(z,
n
+
1)}
if
x = h(z, n) for z ∈ P0
and n ≥ 1,
−1
(f
(x)
∪
{h(z,
2)})
\
{z}
if
x = f (z) for z ∈ P0 ,
−1
f (x)
otherwise
and thus |f∗−1 (x)| = d−1 for each x ∈ X. Hence, Γ(gr(f∗ )) is a d-regular forest.
Just as Theorem 1.1 corresponds to Corollary 1.2, we can translate Theorem 2.2
into an equivalent statement about wobbling groups. Given a coarse space (X, E ),
we define its wobbling group (or group of bounded displacement ) as
W (X, E ) := {α ∈ Sym(X) | gr(α) ∈ E }.
Since the 4-regular tree is isomorphic to the standard Cayley graph of the free group
on two generators, we now obtain the following consequence of Theorem 2.2.
6
FRIEDRICH MARTIN SCHNEIDER
Corollary 2.3. A coarse space X of bounded geometry is non-amenable if and only
if F2 is isomorphic to a semi-regular subgroup of W (X).
We note that Corollary 2.3 for group actions has been applied already (though
without proof) in the recent work of the author and Thom [26, Corollary 5.12],
where a topological version of Whyte’s result for general (i.e., not necessarily locally
compact) topological groups in terms of perturbed translations is established. In
the present note, Corollary 2.3 will be used to prove Theorem 1.4, which generalizes
Whyte’s result to locally compact groups by means of clopen and Borel piecewise
translations and is in turn quite different to [26, Corollary 5.12].
3. Proving the main result
In this section we prove Theorem 1.4. For the sake of clarity, recall that a locally
compact group G is said to be amenable if there is a G-invariant1 mean on the space
Cb (G) of bounded continuous real-valued functions on G, i.e., a positive linear map
µ : Cb (G) → R with µ(1) = 1 and µ(f ◦ λg ) = µ(f ) for all f ∈ Cb (G) and g ∈ G.
In preparation of the proof of Theorem 1.4, we note the following standard fact,
whose straightforward proof we omit.
Lemma 3.1. Let H be a subgroup of a locally compact group G and consider the
usual action of G on the set G/H of left cosets of H in G. If µ : ℓ∞ (G/H) → R is a
G-invariant mean and ν : Cb (H) → R is an H-invariant mean, then a G-invariant
mean ξ : Cb (G) → R is given by
ξ(f ) := µ(xH 7→ ν((f ◦ λx )|H ))
(f ∈ Cb (G)).
It is a well-known fact (see Section 2 in [10]) that a locally compact group G
(considered together with a left Haar measure) is amenable if and only if there
exists a G-invariant mean on L∞ (G), i.e., a positive linear map µ : L∞ (G) → R
such that µ(1) = 1 and µ(f ◦ λg ) = µ(f ) for all f ∈ Cb (G) and g ∈ G. An easy
calculation now provides us with the following.
Lemma 3.2. Let G be a locally compact group.
(1) A mean µ : L∞ (G) → R is G-invariant if and only if µ is B(G)-invariant.
(2) Let H be a locally compact group, let ϕ : H → B(G) be a homomorphism
and ψ : G → H be Borel measurable with ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ H.
If G is amenable, then so is H.
Proof. (1) Clearly, B(G)-invariance implies G-invariance. To prove the converse,
suppose that µ is G-invariant. Let α ∈ B(G) and let P be a finite partition of G
into Borel subsets and gP ∈ G (P ∈ P) with α|P = λgP |P for each P ∈ P. Now,
X
X
µ(f ◦ α) =
µ ((f ◦ α) · 1P ) =
µ ((f ◦ λgP ) · 1P )
P ∈P
=
X
P ∈P
=
X
P ∈P
P ∈P
X
=
µ (f · (1gP P ))
µ f · 1P ◦ λg−1
P
P ∈P
µ f · 1α(P ) = µ(f )
for every f ∈ L∞ (G), as desired.
(2) Let ν : L∞ (G) → R be a G-invariant mean. Define µ : Cb (H) → R by
µ(f ) := ν(f ◦ ψ)
(f ∈ Cb (H)).
It is easy to see that µ is a mean. Furthermore, (1) asserts that
µ(f ◦ λg ) = ν(f ◦ λg ◦ ψ) = ν(f ◦ ψ ◦ ϕ(g)) = ν(f ◦ ψ) = µ(f )
1In case of ambiguity, invariance shall always mean left invariance.
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
for all f ∈ Cb (H) and g ∈ H. Hence, µ is H-invariant.
7
We note that Lemma 3.2 readily settles the implication (3)=⇒(1) of Theorem 1.4.
The remaining part of the proof of Theorem 1.4 will rely on some structure theory
for locally compact groups – most importantly the following remarkable result of
Rickert [22] building on [23]. We recall that a locally compact group G is said to
be almost connected if the quotient of G by the connected component of its neutral
element is compact.
Theorem 3.3 (Theorem 5.5 in [22]). Any almost connected, non-amenable, locally
compact group has a discrete subgroup being isomorphic to F2 .
Now everything is prepared to prove our main result.
Proof of Theorem 1.4. Evidently, (2) implies (3) as C (G) is a subgroup of B(G).
Furthermore, (3) implies (1) due to Lemma 3.2 and the non-amenability of F2 .
(1)=⇒(2). Let G be a non-amenable locally compact group. It follows by classical work of van Dantzig [4] that any locally compact group contains an almost
connected, open subgroup (see, e.g., [20, Proposition 12.2.2 (c)]). Choose any almost connected, open (and hence closed) subgroup H of G. We will distinguish
two cases depending upon whether H is amenable.
H is not amenable. According to Theorem 3.3, H contains a discrete subgroup F
being isomorphic to F2 , and so does G. By a result of Feldman and Greenleaf [7],
the right coset projection π : G → F \ G, x 7→ F x admits a Borel measurable
cross-section, i.e., there exists a Borel measurable map τ : F \ G → G such that
π ◦ τ = idF \G . Clearly, the F -equivariant map ψ : G → F, x 7→ xτ (F x)−1 is Borel
measurable. This readily settles the first case: the maps
ϕ : F2 ∼
= F → C (G),
g 7→ λg
and ψ are as desired.
H is amenable. Since G is not amenable, Lemma 3.1 implies that the action
of G on the set G/H is not amenable. By Proposition 2.1, this means that the
coarse space X := (G/H, EG ) is not amenable. Due to Corollary 2.3, there exists
an embedding ϕ : F2 = F (a, b) → W (X) such that ϕ(F2 ) is semi-regular. Thus, by
definition of W (X), there exists some finite subset E ⊆ G such that
∀x ∈ {a, b} ∀z ∈ X ∃g ∈ E :
ϕ(x)(z) = gz.
Hence, we find a finite partition P of X along with gP , hP ∈ E (P ∈ P) such that
ϕ(a)|P = λgP |P and ϕ(b)|P = λhP |P for every P ∈ P. Consider the projection
π : G → G/H, x 7→ xH. Since H is an open subgroup of G, the quotient topology on
G/H, i.e., the topology induced by π, is discrete. So, π −1 (P) = {π −1 (P ) | P ∈ P}
is a finite partition of G into clopen subsets. What is more,
[
[
[
G = · π −1 (ϕ(a)(P )) = · π −1 (gP P ) = · gP π −1 (P ),
P ∈P
P ∈P
P ∈P
[
[
[
G = · π −1 (ϕ(b)(P )) = · π −1 (hP P ) = · hP π −1 (P ).
P ∈P
P ∈P
P ∈P
Therefore, we may define ϕ : {a, b} → C (G) by setting
ϕ(a)|π−1 (P ) = λgP |π−1 (P ) ,
ϕ(b)|π−1 (P ) = λhP |π−1 (P )
∗
(P ∈ P).
∗
Consider the unique homomorphism ϕ : F2 → C (G) satisfying ϕ |{a,b} = ϕ. Since
π ◦ ϕ(x) = ϕ(x) ◦ π for each x ∈ {a, b}, it follows that π ◦ ϕ∗ (w) = ϕ(w) ◦ π for
every w ∈ F2 . Appealing to Remark 1.3, we find a mapping ψ : G/H → F2 such
that ψ(ϕ(w)(z)) = wψ(z) for all w ∈ F2 and z ∈ G/H. Since the quotient space
8
FRIEDRICH MARTIN SCHNEIDER
G/H is discrete, the map ψ ∗ := ψ ◦ π : G → F2 is continuous and therefore Borel
measurable. Finally, we note that
ψ ∗ (ϕ∗ (w)(x)) = ψ(π(ϕ∗ (w)(x))) = ψ(ϕ(w)(π(x))) = wψ(π(x)) = wψ ∗ (x)
for all w ∈ F2 and x ∈ G, as desired. This completes the proof.
Let us deduce Paterson’s result [21] from Theorem 1.4.
Proof of Corollary 1.5. (⇐=) This is clear.
(=⇒) Let G be a non-amenable locally compact group. By Theorem 1.4, there
exist a homomorphism ϕ : F2 → B(G) and a Borel measurable map ψ : G → F2
with ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 . Consider any paradoxical decomposition of
F2 given by P, Q, (gP )P ∈P , (hQ )Q∈Q . Taking a common refinement of suitable
finite Borel partitions of G corresponding to the elements ϕ(gP ), ϕ(hQ ) ∈ B(G)
(P ∈ P, Q ∈ Q), we obtain a finite Borel partition R of G along with mappings
σ : P × R → G and τ : Q × R → G such that
ϕ(gP )|R = λσ(P,R) |R
ϕ(hQ )|R = λτ (Q,R) |R
for all P ∈ P, Q ∈ Q, and R ∈ R. By ψ being Borel measurable, the refinements
ψ −1 (P) ∨ R and ψ −1 (Q) ∨ R are finite Borel partitions of G. What is more,
[
[
G = · ψ −1 (gP P ) ∪· · ψ −1 (hQ Q)
P ∈P
Q∈Q
[
[
= · ϕ(gP )(ψ −1 (P )) ∪· · ϕ(hQ )(ψ −1 (Q))
P ∈P
=
Q∈Q
[
·
ϕ(gP )(ψ
−1
(P ) ∩ R) ∪·
(P,R)∈P×R
=
[
·
[
·
ϕ(hQ )(ψ −1 (Q) ∩ R)
(Q,R)∈Q×R
σ(P, R)(ψ
−1
(P,R)∈P×R
(P ) ∩ R) ∪·
[
·
τ (Q, R)(ψ −1 (Q) ∩ R).
(Q,R)∈Q×R
Thus, the data
ψ −1 (P) ∨ R,
ψ −1 (Q) ∨ R,
(σ(P, R))(P,R)∈P×R ,
(τ (Q, R))(Q,R)∈Q×R
constitute a Borel paradoxical decomposition of G.
4. Further remarks on wobbling groups
We are going to conclude with some additional remarks about wobbling groups,
which we consider noteworthy complements of Corollary 2.3. As van Douwen’s
result [6] shows, the presence of F2 as a subgroup of the wobbling group does not
imply the non-amenability of a coarse space. As it turns out, containment of F2 is
just a witness for positive asymptotic dimension (Proposition 4.3).
Let us once again recall some terminology from [24]. The asymptotic dimension
asdim(X, E ) of a coarse space (X, E ) is defined as the infimum of all those n ∈ N
such that, for every E ∈ E , there exist C0 , . . . , Cn ⊆ P(X) with
S
S
• X = C0 ∪ . . . ∪ Cn ,
• (C
S × D) ∩ E = ∅ for all i ∈ {0, . . . , n} and C, D ∈ Ci with C 6= D,
• {C × C | C ∈ Ci , i ∈ {0, . . . , n}} ∈ E .
The concept of asymptotic dimension was first introduced for metric spaces by
Gromov [15] and later extended to coarse spaces by Roe [24]. We refer to [24] for
a thorough discussion of asymptotic dimension, related results and examples.
As we aim to describe positive asymptotic dimension in algebraic terms, we will
unravel the zero-dimensional case in the following lemma. Let us denote by [R] the
equivalence relation on a set X generated by a given binary relation R ⊆ X × X.
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
9
Lemma 4.1. Let (X, E ) be a coarse space. Then asdim(X, E ) = 0 if and only if
[E] ∈ E for every E ∈ E .
Proof. (=⇒) Let E ∈ E . Without loss of generality, assume that E contains ∆X .
As asdim(X, E ) = 0, there exists C0 ⊆ P(X) such that
S
(1) X = C0 ,
(2) S
(C × D) ∩ E = ∅ for all C, D ∈ C0 with C 6= D,
(3) {C × C | C ∈ C0 } ∈ E .
As ∆X ⊆ E, assertion (2) implies that any two distinct members of C0 are disjoint.
Hence,
S(1) gives that C0 is a partition of X. By (2), the induced equivalence relation
R := {C × C | C ∈ C0 } contains E, thus [E]. By (3), it follows that [E] ∈ E .
(⇐=) Let E ∈ E . It is straightforward to check that C0 := {[E][x] | x ∈ X} has
the desired properties. Hence, asdim(X, E ) = 0.
Our proof of Proposition 4.3 below will rely upon the following slight modification
of the standard argument for residual finiteness of free groups. For an element
w ∈ F2 = F (a, b), let us denote by |w| the length of w with respect to the generators
a and b, i.e., the smallest integer n ≥ 0 such that w can be represented as a word
of length n in the letters a, a−1 , b, b−1 .
Lemma 4.2. Let w ∈ F2 with w 6= e and let M := {0, . . . , 2|w|}. Then there exists
a homomorphism ϕ : F2 → Sym(M ) such that ϕ(w) 6= e and |ϕ(v)(i) − i| ≤ 2|v| for
all i ∈ M and v ∈ F2 .
Proof. Let (k0 , . . . , kn ) ∈ (Z \ {0})n × ZP
and (ℓ0 , . . .P
, ℓn ) ∈ Z × (Z \ {0})n such that
n
k0 ℓ0
kn ℓn
w = a b · · · a b . Of course, |w| = i=0 |ki | + ni=0 |ℓi |. Let
Xi−1
Xi−1
Xi
Xi−1
|ℓj |
|kj | +
|ℓj |,
βi :=
|kj | +
αi :=
j=0
j=0
j=0
j=0
for i ∈ {0, . . . , n} and let βn+1 := |w|. We will define a map ϕ : {a, b} → Sym(M ).
First, let us define ϕ(a) ∈ Sym(M ) by case analysis as follows: if i ∈ [2αj , 2βj+1 ]
for some j ∈ {0, . . . , n} with kj > 0, then
i + 2 if i is even and i ∈ [2αj , 2βj+1 − 2],
i − 1 if i = 2β ,
j+1
ϕ(a)(i) :=
i
−
2
if
i
is
odd
and i ∈ [2αj + 3, 2βj+1 − 1],
i − 1 if i = 2αj + 1,
if i ∈ [2αj , 2βj+1 ] for some j ∈ {0, . . . , n} with kj < 0, then
i − 2 if i is even and i ∈ [2αj + 2, 2βj+1 ],
i + 1 if i = 2α ,
j
ϕ(a)(i) :=
i
+
2
if
i
is
odd
and i ∈ [2αj + 1, 2βj+1 − 3],
i + 1 if i = 2βj+1 − 1,
S
and if i ∈
/ {[2αj , 2βj+1 ] | j ∈ {0, . . . , n}, kj 6= 0}, then ϕ(a)(i) := i. Analogously,
let us define ϕ(b) ∈ Sym(M ) by case analysis as follows: if i ∈ [2βj , 2αj ] for some
j ∈ {0, . . . , n} with ℓj > 0, then
i + 2 if i is even and i ∈ [2βj , 2αj − 2],
i − 1 if i = 2α ,
j
ϕ(b)(i) :=
i − 2 if i is odd and i ∈ [2βj + 3, 2αj − 1],
i − 1 if i = 2βj + 1,
10
FRIEDRICH MARTIN SCHNEIDER
if i ∈ [2βj , 2αj ] for some j ∈ {0, . . . , n} with ℓj < 0, then
i − 2 if i is even and i ∈ [2βj + 2, 2αj ],
i + 1 if i = 2β ,
j
ϕ(b)(i) :=
i
+
2
if
i
is
odd
and i ∈ [2βj + 1, 2αj − 3],
i + 1 if i = 2αj − 1,
S
and if i ∈
/ {[2βj , 2αj ] | j ∈ {0, . . . , n}, ℓj 6= 0}, then ϕ(b)(i) := i. It is easy to
check that ϕ(a) and ϕ(b) are well-defined permutations of M , and that moreover
|ϕ(x)(i) − i| ≤ 2 for each x ∈ {a, b} and all i ∈ M . Considering the unique
homomorphism ϕ∗ : F2 → Sym(M ) with ϕ∗ |{a,b} = ϕ, we observe that
ϕ∗ (w)(0) = ϕ(a)kn ϕ(b)ℓn · · · ϕ(a)k0 ϕ(b)ℓ0 (0) = 2|w|
and thus ϕ∗ (w) 6= e. Also, |ϕ∗ (v)(i) − i| ≤ 2|v| for all i ∈ M and v ∈ F2 .
For the sake of clarity, we recall that a group is locally finite if each of its finitely
generated subgroups is finite. For a subset S of a group G, we will denote by hSi
the subgroup of G generated by S.
Proposition 4.3. Let X be a coarse space of uniformly bounded geometry. The
following are equivalent.
(1) asdim(X) > 0.
(2) W (X) is not locally finite.
(3) F2 embeds into W (X).
Proof. We will denote by E the coarse structure of X.
(2)=⇒(1). Let us recall a general fact: for a finite group G and any set M , the
group GM is locally finite. Indeed, considering a finite subset S ⊆ GM and the
induced equivalence relation R := {(x, y) ∈ M × M | ∀α ∈ S : α(x) = α(y)} on M ,
we observe that N := {R[x] | x ∈ M } is finite, due to G and S being finite. The
map π : M → N, x 7→ R[x] induces a homomorphism ϕ : GN → GM , α 7→ α ◦ π.
Evidently, S is contained in the finite group ϕ(GN ), and so is hSi.
Suppose now that asdim(X) = 0. Consider a finite subset S ⊆ W (X). We aim
to show that H := hSi is finite. To this end, we first observe that
Y
ϕ: H →
Sym(Hx), α 7→ (α|Hx )x∈X
x∈X
S
constitutes a well-defined embedding. Since D := {gr(α) | α ∈ S} belongs to E ,
Lemma 4.1 asserts that E := [D] ∈ E , too. Note that gr(α) ∈ E for all α ∈ H.
Hence, Hx ⊆ E[x] for every x ∈ X. Due to X having uniformly bounded geometry,
there exists m ≥ 0 such that |E[x]| ≤ m and thus |Hx|
Q ≤ m for every x ∈ X. Now,
let M := {0, . . . , m − 1}. It follows that the group x∈X Sym(Hx) is isomorphic to
a subgroup of Sym(M )X , and so is H by virtue of ϕ. Since H is finitely generated
and Sym(M )X is locally finite by the remark above, this implies that H is finite.
(3)=⇒(2). This is trivial.
(1)=⇒(3). Suppose that asdim(X) > 0. By Lemma 4.1, there exists E ∈ E such
that [E] ∈
/ E .SWithout loss of generality, we may assume that ∆X ⊆ E = E −1 .
Hence, [E] = {E n | n ∈ N}. For each n ∈ N, let us define
Tn := x ∈ X n+1 |{x0 , . . . , xn }| = n + 1, ∀i ∈ {0, . . . , n − 1} : (xi , xi+1 ) ∈ E .
Claim. For every n ∈ N and every finite subset F ⊆ X, there exists x ∈ Tn such
that {x0 , . . . , xn } ∩ F = ∅.
Proof of claim. Let n ∈ N and let F ⊆ X be finite. Put ℓ := (n + 1)(|F | + 1).
Since E ∈ E and [E] ∈
/ E , we conclude that E ℓ * E ℓ−1 . Let x0 , . . . , xℓ ∈ X such
ℓ−1
that (x0 , xℓ ) ∈
/E
and (xi , xi+1 ) ∈ E for every i ∈ {0, . . . , ℓ − 1}. As ∆X ⊆ E, it
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
11
follows that |{x0 , . . . , xℓ }| = ℓ + 1. Applying the pigeonhole principle, we find some
j ∈ {0, . . . , ℓ−n} such that {xj , . . . , xj+n }∩F = ∅. Hence, y0 := xj , . . . , yn := xj+n
are as desired.
Since N := F2 \ {e} is countable, we may recursively apply the claim above and
choose a family (xw )w∈N such that
(i) xw ∈ T2|w| for every w ∈ N ,
(ii) {xw,0 , . . . , xw,2|w| } ∩ {xv,0 , . . . , xv,2|v| } = ∅ for any two distinct v, w ∈ N .
Let w ∈ N and define Dw := {xw,0 , . . . , xw,2|w| }. Due to Lemma 4.2, there exists a
homomorphism ϕw : F2 → Sym(Dw ) such that ϕw (w) 6= e and
ϕw (v)(xw,i ) ∈ {xw,j | j ∈ {0, . . . , 2|w|}, |i − j| ≤ 2|v|}
for all v ∈ F2 , i ∈ {0, . . . , 2|w|}. Since (xw,i , xw,i+1 ) ∈ E for i ∈ {0, . . . , 2|w| − 1},
it follows that gr(ϕw (v)) ⊆ E 2|v| for all v ∈ F2 . As Dw and Dv are disjoint for any
distinct v, w ∈ N , we may define a homomorphism ϕ : F2 → Sym(X) by setting
(
ϕw (v)(x) if x ∈ Dw for some w ∈ N,
ϕ(v)(x) :=
x
otherwise
for v ∈ F2 and x ∈ X. By construction, ϕ is an embedding, and furthermore
[
gr(ϕ(v)) ⊆ ∆X ∪ {gr(ϕw (v)) | w ∈ N } ⊆ E 2|v| ∈ E
for every v ∈ F2 . Hence, the image of ϕ is contained in W (X), as desired.
Remark 4.4. The assumption of uniformly bounded geometry in Theorem 1.4 is
needed only to prove that (2) implies (1). In fact, a similar argument as in the proof
of (1)=⇒(3) (not involving Lemma 4.2 though) shows that the wobbling group of
any coarse
Q space not having uniformly bounded geometry contains an isomorphic
copy of n∈N Sym(n), hence F2 .
One might wonder whether Proposition 4.3 could have been deduced readily
from van Douwen’s result [6] on F2 embedding into W (Z). However, there exist uniformly discrete metric spaces of uniformly bounded geometry and positive
asymptotic dimension whose wobbling group does not contain an isomorphic copy
of W (Z) (see Example 4.7). We clarify the situation in Proposition 4.5.
As usual, a group is called residually finite if it embeds into a product of finite
groups, and a group is called locally residually finite if each of its finitely generated
subgroups is residually finite. Let us recall from [24] that a map f : X → Y between
two coarse spaces X and Y is bornologous if, for every entourage E of X, the set
{(f (x), f (y)) | (x, y) ∈ E} is an entourage of Y .
Proposition 4.5. Let X be a coarse space. The following are equivalent.
(1) There is a bornologous injection from Z into X.
(2) W (X) is not locally residually finite.
(3) W (X) contains a subgroup being isomorphic to W (Z).
Remark 4.6. (i) For groups there is no difference between positive asymptotic
dimension and the existence of a bornologous injection of Z: a group has asymptotic
dimension 0 if and only if it is locally finite, and any group which is not locally
finite admits a bornologous injection of Z by a standard compactness argument
(see, e.g., [17, IV.A.12]). However, for arbitrary coarse spaces, even of uniformly
bounded geometry, the situation is slightly different (see Example 4.7).
(ii) One may equivalently replace Z by N in item (1) of Proposition 4.5: on the
one hand, the inclusion map constitutes a bornologous injection from N into Z; on
12
FRIEDRICH MARTIN SCHNEIDER
the other hand, there is a bornologous bijection f : Z → N given by
(
2n
if n ≥ 0,
f (n) :=
(n ∈ Z).
2|n| − 1 if n < 0
Unless explicitly stated otherwise, we always understand N as being equipped with
the coarse structure generated by the usual (i.e., Euclidean) metric.
(iii) Any bornologous injection f : X → Y between two coarse spaces X and Y
induces an embedding ϕ : W (X) → W (Y ) via
(
f (α(f −1 (y))) if y ∈ f (X),
ϕ(α)(y) :=
(α ∈ W (X), y ∈ Y ).
y
otherwise
Hence, by (ii), the groups W (N) and W (Z) mutually embed into each other, and
thus Z may equivalently be replaced by N in item (3) of Proposition 4.5.
Proof of Proposition 4.5. (1)=⇒(3). This is due to Remark 4.6(iii).
(3)=⇒(2). It suffices to show that W (Z) is not locally residually finite. A result
of Gruenberg [16] states that, for a finite group F , the restricted wreath product
F ≀ Z = F (Z) ⋊ Z (i.e., the lamplighter group over F ) is residually finite if and only
Sn−1
if F is abelian. For n ≥ 1, the action of Sym(n) ≀ Z on Z = · r=0 nZ + r given by
(α, m).(nk + r) := n(m + k) + αm+k (r)
α ∈ Sym(n)(Z) , m, k ∈ Z, 0 ≤ r < n
defines an embedding of Sym(n) ≀ Z into Sym(Z), the image of which is contained
in W (Z) as supz∈Z |z − (α, m).z| ≤ n(|m| + 1) for every (α, m) ∈ Sym(n) ≀ Z. Since
the embedded lamplighter groups are finitely generated and not residually finite for
n ≥ 3, it follows that W (Z) is not locally residually finite.
(2)=⇒(1). Let E denote the coarse structure of X. If X does not have bounded
geometry, then there exist E ∈ E and x ∈ X such that E[x] is infinite, and any
thus existing injection f : Z → X with f (Z) ⊆ E[x] is bornologous. Hence, we
may without loss of generality assume that X has bounded geometry. On the other
hand, there must exist E ∈ E and x ∈ X such that [E][x] is infinite. Otherwise,
W (X) would
S have to be locally residually finite: for any finite subset F ⊆ W (X),
since E := {gr(α) | α ∈ F } ∈ E , the homomorphism
Y
hF i →
Sym([E][x]), α 7→ α|[E][x] x∈X
x∈X
would embed hF i into a product of finite groups. So, let E ∈ E and x ∈ X such that
[E][x] is infinite. S
Without loss of generality, we may assume that ∆X ⊆ E = E −1 .
Therefore, [E] = {E n | n ∈ N}. We conclude that E n [x] 6= E n+1 [x] and thus
Rn := f ∈ X N f0 = x, |{f0 , . . . , fn }| = n + 1, ∀i ∈ N : (fi , fi+1 ) ∈ E
is non-empty for all
Q n ∈ N. As (Rn )n∈N is a chain
T of closed subsets of the compact
topological space m∈N E m [x], we have R := n∈N Rn 6= ∅. Since any member of
R is a bornologous injection from N into X, this implies (1) by Remark 4.6(ii).
Example 4.7. Let I be a partition of N into finite intervals with supI∈I |I| = ∞.
Consider the metric space X := (N, d) given by
(
|x − y|
if x, y ∈ I for some I ∈ I ,
d(x, y) :=
(x, y ∈ N).
max(x, y) otherwise
It is easy to see that X has uniformly bounded geometry. Moreover, by Lemma 4.1
and the unboundedness assumption for the interval lengths, it follows that X has
positive asymptotic dimension. On the other hand, essentially by finiteness of the
considered intervals, there is no bornologous injection from N into X. Due to
Proposition 4.5, this readily implies that W (Z) does not embed into W (X).
ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS
13
The interplay between certain geometric properties of coarse spaces on the one
hand and algebraic peculiarities of their wobbling groups on the other is a subject
of recent attention [12, 3]. It would be interesting to have further results in that
direction, e.g., to understand if (and how) specific positive values for the asymptotic
dimension may be characterized in terms of wobbling groups.
Acknowledgments
The author would like to thank Andreas Thom for interesting discussions about
Whyte’s variant of the von Neumann conjecture, as well as Warren Moors and Jens
Zumbrägel for their helpful comments on earlier versions of this note.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
Tullio Ceccherini-Silberstein and Michel Coornaert, The Hall Harem Theorem, In: Cellular
Automata and Groups, Springer Monographs in Mathematics, Springer Berlin Heidelberg
(2010), pp. 391–401.
Tullio Ceccherini-Silberstein, Rostislav I. Grigorchuk, and Pierre de la Harpe, Amenability
and paradoxical decompositions for pseudogroups and discrete metric spaces, Proc. Steklov
Inst. Math. 224 (1999), pp. 57–97.
Yves Cornulier, Irreducible lattices, invariant means, and commensurating actions, Math. Z.
279 (2015), no. 1–2, pp. 1–26.
David van Dantzig, Zur topologischen Algebra. III. Brouwersche und Cantorsche Gruppen,
Compositio Math. 3 (1936), pp. 408–426.
Mahlon M. Day, Amenable semigroups, Illinois J. Math. 1 (1957), pp. 509–544.
Eric K. van Douwen, Measures invariant under actions of F2 , Topology and its Applications
34 (1990), pp. 53–68.
Jacob Feldman and Frederick P. Greenleaf, Existence of Borel transversals in groups, Pacific
J. Math. 25 (1968), pp. 455–461.
Erling Følner, On groups with full Banach mean value, Math. Scand. 3 (1955), pp. 243–254.
Damien Gaboriau and Russell Lyons, A measurable-group-theoretic solution to von Neumann’s problem, Invent. Math. 177 (2009), no. 3, pp. 533–540.
Frederick P. Greenleaf, Invariant means on topological groups and their applications, Van
Nostrand Mathematical Studies, No. 16. Van Nostrand Reinhold Co., New York-Toronto,
Ont.-London, 1969, pp. ix+113.
Kate Juschenko and Nicolas Monod, Cantor systems, piecewise translations and simple
amenable groups, Ann. of Math. (2) 178 (2013), no. 2, pp. 775–787.
Kate Juschenko and Mikael de la Salle, Invariant means for the wobbling group, Bull. Belg.
Math. Soc. Simon Stevin 22 (2015), no. 2, pp. 281–290.
Andrew Marks and Spencer Unger, Baire measurable paradoxical decompositions via matchings, Adv. Math. 289 (2016), pp. 397–410.
Maxime Gheysens and Nicolas Monod, Fixed points for bounded orbits in Hilbert spaces,
August 2015, arXiv: 1508.00423[math.GR], to appear in Annales scientifiques de l’École
normale supérieure.
Michail Gromov, Asymptotic invariants of infinite groups. In: Geometric group theory, Vol.
2 (Sussex, 1991). London Math. Soc. Lecture Note Ser. 182, Cambridge Univ. Press, Cambridge, 1993, pp. 1–295.
Karl W. Gruenberg, Residual properties of infinite soluble groups, Proc. London Math. Soc.
(3) 7 (1957), pp. 29–62.
Pierre de la Harpe, Topics in geometric group theory, Chicago Lectures in Mathematics.
University of Chicago Press, Chicago, IL, 2000, pp. vi+310.
John von Neumann, Über die analytischen Eigenschaften von Gruppen linearer Transformationen und ihrer Darstellungen, Math. Z. 30 (1929), no. 1, pp. 3–42.
Alexander Ju. Ol’šanskiı̆, On the question of the existence of an invariant mean on a group,
Uspekhi Mat. Nauk 35 (1980), no. 4(214), pp. 199–200.
Theodore W. Palmer, Banach algebras and the general theory of ∗-algebras. Vol. 2, Encyclopedia of Mathematics and its Applications 79. Cambridge University Press, Cambridge,
2001, pp. i–xii and 795–1617.
Alan L. T. Paterson, Nonamenability and Borel paradoxical decompositions for locally compact groups, Proc. Amer. Math. Soc. 96 (1986), pp. 89–90.
Neil W. Rickert, Amenable groups and groups with the fixed point property, Trans. Amer.
Math. Soc. 127 (1967), pp. 221–232.
14
FRIEDRICH MARTIN SCHNEIDER
[23] Neil W. Rickert, Some properties of locally compact groups, J. Austral. Math. Soc. 7 (1967),
pp. 433–454.
[24] John Roe, Lectures on coarse geometry, University Lecture Series 31. American Mathematical
Society, Providence, RI, 2003, pp. viii+175.
[25] Joseph M. Rosenblatt, A generalization of Følner’s condition, Math. Scand. 33 (1973), no. 3,
pp. 153–170.
[26] Friedrich M. Schneider and Andreas B. Thom, On Følner sets in topological groups, August
2016, arXiv: 1608.08185[math.GR].
[27] Kevin Whyte, Amenability, bi-Lipschitz equivalence, and the von Neumann conjecture, Duke
Math. J. 99 (1999), no. 1, pp. 93–112.
Institute of Algebra, TU Dresden, 01062 Dresden, Germany
Current address: Department of Mathematics, The University of Auckland, Private Bag 92019,
Auckland 1142, New Zealand
E-mail address: [email protected]
| 4math.GR
|
On modal analysis of laminated glass: Usability of
simplified methods and enhanced effective thickness
arXiv:1801.01110v1 [cs.CE] 3 Jan 2018
Alena Zemanová, Jan Zeman, Tomáš Janda, Jaroslav Schmidt, Michal
Šejnoha
Department of Mechanics, Faculty of Civil Engineering, Czech Technical University in
Prague
Abstract
This paper focuses on the modal analysis of laminated glass beams. In these
multilayer elements, the stiff glass plates are connected by compliant interlayers with frequency/temperature-dependent behavior. The aim of our study
is (i) to assess whether approximate techniques can accurately predict the
behavior of laminated glass structures and (ii) to propose an easy tool for
modal analysis based on the enhanced effective thickness concept by Galuppi
and Royer-Carfagni.
To this purpose, we consider four approaches to the solution of the related nonlinear eigenvalue problem: a complex-eigenvalue solver based on
the Newton method, the modal strain energy method, and two effective
thickness concepts. A comparative study of free vibrating laminated glass
beams is performed considering different geometries of cross-sections, boundary conditions, and material parameters for interlayers under two ambient
temperatures. The viscoelastic response of polymer foils is represented by
the generalized Maxwell model.
We show that the simplified approaches predict natural frequencies with
an acceptable accuracy for most of the examples. However, there is a considerable scatter in predicted loss factors. The enhanced effective thickness
approach adjusted for modal analysis leads to lower errors in both quantities
compared to the other two simplified procedures, reducing the extreme error
in loss factors to one half compared to the modal strain energy method or to
one quarter compared to the original dynamic effective thickness method.
Keywords: Free vibrations, Laminated glass, Complex dynamic modulus,
Dynamic effective thickness, Enhanced effective thickness, Modal strain
Preprint submitted to arXiv
January 4, 2018
energy method, Newton method
1. Introduction
Laminated glass is a multilayer composite made of glass layers and plastic
interlayers, typically polymers. These foils improve the post-fracture behavior of the originally brittle glass elements [18, 25, 2], increase their damping [16, 29], and therefore they allow for applications prohibited to traditional
glass, such as load-bearing and fail-safe transparent structures. Avoiding resonance and reducing noise and vibrations of laminated glass components is
thus substantial not only in building structures but also in the car or ship
design process and others. Thus, the reliable prediction of natural frequencies and damping characteristics associated with each vibration mode is an
essential issue for the design of dynamically loaded structures. In the case of
laminated glass, free vibration analysis leads, because of the viscoelastic behavior of the foil [1], to an eigenvalue problem with complex eigenvalues and
eigenvectors, corresponding to natural angular frequencies and mode shapes.
In addition, the nonlinearity due to the frequency/temperature-sensitive response of the polymer interlayer adds to the complexity of the analysis.
Several approaches to analyzing vibrations of viscoelastically damped layered composites can be found in the literature, e.g., [34, 8, 16]. In this paper,
we broadly divide these methods into three groups: (i) numerical approaches
solving the complex eigenvalue problem directly [14], (ii) simplified numerical approximations dealing with a real eigenmode problem only [4], and
(iii) analytical methods and effective thickness methods derived from analytical models [19].
The comparison of selected non-linear solvers for complex-valued problems
in [14] shows that while most of them converge towards the same eigenvalues,
their computational time and the number of iterations differ. The computational cost can be reduced using simplified numerical methods, which deal
only with a real eigenvalue problem corresponding to the delayed elasticity or take into account only the real part of the complex stiffness of the
core [4]. Then, the damping parameters are obtained by post-processing the
real-valued eigenvalues and eigenmodes using, e.g., the modal strain energy
method [17] to be discussed later in the paper.
For three-layer structures with simple boundary conditions and geometry,
analytical solutions can be derived [22, 23, 31]. Because of the frequencydependent behavior of the polymer foil, they provide natural frequencies and
2
loss factors using an iterative algorithm. Recently, the dynamic effective
thickness approach for laminated glass beams was proposed by López-Aenlle
and Pelayo [19], using the complex flexural stiffness introduced in [31]. This
concept can be extended towards plates [20] and multilayer laminated glass
beams [26]. The validation of this dynamic effective thickness method against
the results of experimental testing in [19, 20, 26] shows that using this approach natural frequencies can be predicted with good accuracy but there is
a high scatter in loss factors.
Therefore, we want to analyze the accuracy of the response of effective
thickness and other simplified approaches, to investigate their usability for
the modal analysis of laminated glass elements, and to propose some improvements. More specifically, we
• perform a comparative study for free vibrating beams using selected
solvers representing the three groups as introduced above and
• propose an easy tool for modal analysis based on the enhanced effective
thickness concept by Galuppi and Royer-Carfagni [12].
To our best knowledge, no such comparison of complex and approximate
models has been performed for laminated glass so far.
All methods are compared for simply-supported, clamped-clamped, or
free-free beams with symmetric and asymmetric cross-sections under different
ambient temperatures. The viscoelastic behavior of polymer foil is described
with the generalized Maxwell model. Several sets of parameters of the chain
were taken from literature [1, 24, 33] and used in our case study, to evaluate
and discuss the effect of various materials used in laminated glass structures
and also of different Maxwell chain parameters for the the same type of
interlayer.
The structure of this paper is as follows. The geometry of a three-layer
laminated glass beam and material characterization of the glass and polymer
layers is outlined in Section 2. The approaches based on the finite element
methods, i.e., the Newton method and the modal strain energy method, are
introduced in Section 3. The closed-form formula for the complex-valued
natural frequencies presented in Section 4 is combined with the effective
thickness concept, using the dynamic effective thickness from [20] and the
enhanced effective thickness [12] which we adjust for modal analysis. The
results of our case study are presented and analyzed in Section 5. Finally,
we summarize our findings in Section 6.
3
2. Characterization of laminated glass structures
2.1. Configuration of laminated glass beam
In this paper, the most common three-layer configuration (with two face
glass plies and one polymer interlayer, see Figure 1) is handled for simplicity. However, the extension towards multilayer elements is possible for all
discussed approaches. No slipping on the interface of a glass ply and the
polymer foil is assumed.
y
x
h1
h2
h3
z
l
glass
z
polymer
glass
b
Figure 1: Basic three-layer composition of laminated glass sandwich.
2.2. Materials
The constitutive behavior of glass and polymer layers remains the same
for all presented methods. Glass is treated as an elastic material whereas the
polymer behavior is assumed to be linear viscoelastic. It can be supposed
that the damping of glass is small in comparison with that of the interlayer
and thus it is negligible. Therefore, the behavior of a glass layer is described
by three parameters: the Young modulus E1 , the Poisson ratio ν1 , and the
density ρ1 (resp. E3 , ν3 , and ρ3 for the face layers of different types of glass).
The variety of interlayer materials for laminated glass is broad, ranging
across the most common polyvinyl butyral (PVB), ethylene-vinyl acetate
(EVA), thermoplastic polyurethane (TPU), or the stiffest ionoplast (SGP –
SentryGlas(R) Plus). The density of the interlayer material is ρ2 and we
assume that its Poisson’s ratio ν2 is constant, see [37]. The viscoelastic behavior of polymers is commonly described by the generalized Maxwell model
or by a fractional derivative model [14]. In our study, the generalized Maxwell
model is used for two reasons: (i) for laminated glass, this description is the
most common in literature and (ii) according to the comparison by Hamdaoui
et al. [14], it requires less computational effort than the fractional derivative
model for non-linear eigenvalue problems.
4
G2,∞
G2,1
G2,2
G2,3
G2,P
µ1
µ2
µ3
µP
Figure 2: Generalized Maxwell chain consisting of P viscoelastic units and one elastic
spring.
The schematic representation of the Maxwell model in Figure 2 corresponds to the relaxation function provided by the Prony series, e.g., [6, page
32],
G2 (t) = G2,∞ +
P
X
− θt
G2,p exp
p
= G2,0 −
P
X
− θt
G2,p (1 − exp
p
),
(1)
p=1
p=1
where t stands for the current time, G2,∞ is the long-term shear modulus,
G2,p denotes the shear modulus of the p-th unit, θp = µp /G2,p its relaxation
P
time related to the viscosity µp , and G2,0 = G2,∞ + Pp=1 G2,p is the elastic
shear modulus of the whole chain.
In the frequency domain, the shear modulus is given by a complex valued
quantity
G2 (ω) = G2,0 + G2,ω (ω),
(2)
composed of the real frequency-independent part G2,0 and the complex part
dependent on the angular frequency ω according to
G2,ω (ω) = −
X
p
G2,p
1
ω 2 θp2 + 1
+i
X
p
G2,p
ωθp
.
ω 2 θp2 + 1
(3)
The real part (storage modulus) refers to the elastic behavior, whereas the
imaginary part (loss modulus) represents the energy dissipation effects [6].
This decomposition will be useful for the formulation of the free vibration
problem. Unlike in [8], the elastic shear modulus of the whole chain G2,0 was
set as the real frequency-independent part instead of the long-term shear
modulus G2,∞ to avoid computational difficulties for chains with G2,∞ = 0.
The Maxwell chain parameters for different interlayer materials can be
found, for instance, in [9, 27, 5, 15, 1, 33, 24]. We selected five representative
sets of Prony series and related parameters, which are reported in Appendix
5
102
LOSS MODULUS [MPa]
STORAGE MODULUS [MPa]
103
101
10-1
10-3
100
103
SGPM
TPUM
100
PVB M
PVB S
PVB A
10-2
10-3
FREQUENCY [Hz]
100
103
FREQUENCY [Hz]
Figure 3: The dependence of the real part (storage modulus) and the imaginary part (loss
modulus) of the shear modulus G2 (ω) of PVB, TPU, and SGP on a real-valued frequency
at 25 ◦ C; the Maxwell chain parameters are taken from [24] (M), [33] (S), and [1] (A).
A. It is evident from Figure 3 that the data differ even for the same polymer type. Different test methods may result in slightly different parameters
for interlayer properties, [13] or [1]. The different content of additives and
plasticizers used in the manufacturing also partially contributes to this discrepancy [9]. Besides, the authors mostly do not specify the frequency range
for which their Prony series were determined (except for PVBA [1]).
The effect of temperature is accounted for by using the time-temperature
superposition principle. For a given ambient temperature T , the relaxation
times θp in Eq. (3) are shifted by the factor aT derived from the WilliamsLandel-Ferry equation [36]
log aT = −
C1 (T − T0 )
,
C2 + T − T0
(4)
where the material constants C1 and C2 correspond to the reference temperature T0 .
3. Finite element models
3.1. Refined beam element for three-layer laminated glass
Because of the small thickness of the interlayer, we assume that the shear
deformation in the viscoelastic foil is responsible for all the damping and the
transverse compressive strain is negligible. Thus, we treat each layer as a
one-dimensional beam element in our numerical analysis. We assume planar
6
cross-sections of individual layers but not of the whole composite. The three
layers of a laminated glass beam are constrained together with compatibility
equations. This constraint can be taken into account using (i) the Lagrange
multipliers – additional unknown nodal forces holding the adjacent layers
together [37], or by (ii) the static condensation of the dependent generalized
displacements [30]. The second option is used in this study, because no
delamination is assumed during the modal analysis.
le
x
z
ϕ1,L
ϕ2,L
master DOFs
slave DOFs
ϕ3,L
u1,L
ϕ1,R
u2,L
ϕ2,R
u3,L
ϕ3,R
w1,L
w2,L
w3,L
u1,R
w1,R
u2,R
w2,R
u3,R
h1
glass
h2
polymer
h3
glass
w3,R
Figure 4: Master and slave degrees of freedom (DOFs) of the refined beam element:
horizontal central displacements u•,L , u•,R , deflections w•,L , w•,R , and rotations ϕ•,L ,
ϕ•,R .
According to the Timoshenko beam theory, there are nine unknowns per
cross-section of a three-layer beam – the horizontal and vertical centerline
displacements and the rotations for each layer, see Figure 4. Four unknowns
can be eliminated using four inter-layer compatibility conditions assuming the
perfect adhesion in the horizontal and vertical directions. This elimination
together with the outline of the element stiffness and mass matrices can
be found in Appendix B. Note that we use linear basis functions for the
evaluation of stiffness and mass matrices, and selective integration scheme to
avoid shear locking.
3.2. Governing equations for nonlinear eigenvalue problem
After the discretization, the free vibration problem is described by the
governing equation, e.g., [8],
K(ω) − ω 2 M U = 0.
(5)
The eigenvalues and the eigenvectors solving the problem are complex-valued
and represent the squared angular frequencies ω and the associated mode
7
shapes U . The mass matrix M is real-valued and constant, frequency independent, whereas the complex stiffness matrix K depends on the complex
frequency ω of the system. Similarly to Eq. (2), the stiffness matrix is decomposed as
K(ω) = K 0 + K ω (ω).
(6)
The elastic matrix K 0 includes the contributions of the glass faces and the
stiffness matrix of the interlayer corresponding to the instantaneous shear
modulus G0 . The frequency dependent part K ω (ω) = G2,ω (ω)K c , where K c
is a constant matrix.
Then, the free vibration problem can be rewritten as
K 0 + G2,ω (ω)K c − ω 2 M U = 0,
(7a)
U 0 T (U − U 0 ) = 0,
(7b)
where Eq. (7b) is introduced to obtain a well-posed problem, see [8]. The
mode shape U 0 is the eigenvector solving the real-eigenvalue problem
K 0 − ω02 M U 0 = 0.
(8)
No damping is thus assumed in this auxiliary problem, and the stiffness
matrix is constant (frequency-independent). To solve Eq. (8), we used a
solver based on the implicitly restarted Arnoldi method [28].
3.3. Natural frequencies and loss factors
The natural frequency fHz and the modal loss factor η associated with
the same mode shape are determined from the relevant squared frequencies
ω 2 from Eq. (7a) according to, e.g., [8],
ω 2 = (2πfHz )2 (1 + iη) ,
(9)
thus
p
Re[ω 2 ]
,
fHz =
2π
Im[ω 2 ]
η=
.
Re[ω 2 ]
(10a)
(10b)
Note that we decided to omit the index referring to the given mode shape
to avoid a profusion of notation. Later in Section 5, we compare the natural
frequencies and loss factors corresponding to the first three mode shapes.
8
3.4. Complex-eigenvalue solver using Newton method (CNM)
The Newton method is applied to the extended system (7) to obtain
the complex-valued pair of an eigenvalue and an eigenvector. Our approach
belongs to the class of methods iterating each individual eigenpair independently, starting from the initial eigenpairs solving Eq. (8). We express the
(k + 1)-th approximations of the searched frequency and mode shape from
the previous k-th approximations and their increments in the form
k+1
k+1
ω = k ω + δω,
(11a)
U = k U + δU .
(11b)
After linearization and evaluation of the Jacobian matrix, the Newton method
leads to the system of linear equations, e.g., [32, page 73, Eq. (4.8)],
#
"
∂T k k
k
0
( ω) U k+1 U
T ( ω)
=
,
(12)
∂ω
δω
U 0TU 0
U 0T
0
where the operator T and its derivative are provided by
T (ω) = K 0 + G2,ω (ω)K c − ω 2 M ,
∂T (ω)
∂G2,ω (ω)
=
K c − 2ωM .
∂ω
∂ω
(13a)
(13b)
The initial pair k=0 ω = ω0 and k=0 U = U 0 is obtained from the realeigenvalue problem in Eq. (8) and the stopping criterion is defined by
kT (k ω)k U k
< tol
kk U k
(14)
with the norm of residual weighted by the norm of the relevant mode shape
and the user defined tolerance tol .
Our proposed complex-valued eigenvalue solver was verified against the
semi-analytical model by Mead [21]; the results for this model were adopted
from [19] for a laminated glass beam with free ends. The response is in a
good agreement; the errors in frequencies are below 0.5%, and in loss factors
are less than 2%. These minor discrepancies can be caused by different
assumptions for the semi-analytical model (e.g., zero Young’s modulus for the
interlayer), by a different stopping tolerance of iterative solvers (which is not
specified in [19]), or by the ambiguity in Prony series, where the conversion
of the Young modulus to the shear modulus is unclear (the assumption of a
constant value of the Poisson ratio vs. the bulk modulus, see [27, 37]).
9
3.5. Real-eigenvalue solver and modal strain energy method (MSE)
Especially for large-scale structures, the direct evaluation of the complexvalued solution can be expensive. Therefore, the natural frequencies are
often approximated from a real-valued simplification of the complex-valued
problem, recall Eq. (7), as
2
K ap (k ωap ) − k+1 ωap
M k+1 U ap = 0,
(15a)
U 0 T (k+1 U ap − U 0 ) = 0.
(15b)
Due to the real-valued approximation of the stiffness matrix K ap (k ωap ), the
(k + 1)-th estimates of the eigenpairs are real squared angular frequencies
k+1 2
ωap and real mode shape vectors k+1 U ap . We consider that the approximate stiffness matrix1 corresponds to the real part of the whole stiffness
matrix evaluated for the relevant frequency
K ap (ωap ) = K 0 + Re[G2,ω (ωap )]K c .
(16)
Therefore, the stiffness matrix is updated iteratively until the convergence of
the angular frequency
kk+1 ωap − k ωap k
< tol .
(17)
kk+1 ωap k
To estimate the loss factors from the converged real eigenpairs (k+1 ωap ,
k+1
U ap ), the modal strain energy method was introduced by Ungar and Kerwin [35] and later popularized for finite element models by Johnson and
Kienholz [17]. This method assumes that no changes occur in the damped
mode shapes, Eq. (18), and therefore it is suitable only for lightly damped
structures [34]. The modal loss factor of each individual mode is determined
according to
k+1
ηap =
U ap T Im[G2,ω (k+1 ωap )]K c k+1 U ap
.
k+1 U T K (k+1 ω )k+1 U
ap
ap
ap
ap
(18)
The derivation of this formula from Rayleigh’s quotient is described in [17].
This iterative procedure provides more accurate values of the natural
frequencies and the loss factors than the original approximation technique
1
Different ways for determining the real-valued approximate stiffness matrix K ap can
be found in the literature, see [17, 4].
10
in [17]. The original one starts from a constant initial value of the interlayer
shear modulus and then only adjusts the computed loss factor by a correction factor taking into account the change of material properties due to the
frequency shift.
4. Effective thickness approaches
4.1. Closed-form expression for natural angular frequencies of beams
A few effective thickness formulations can be found in the literature for
laminated glass beams and plates under static loading, whereas, to the best
of our knowledge, only one effective thickness approach exists for dynamic
problems [19, 20]. In general, the effective thickness methods are based
on calculating a constant thickness of a monolithic element (with the same
width and length), which gives the same response as the laminated glass
beam under the identical loading and boundary conditions. For example,
a deflection-effective thickness and stress-effective thicknesses [3, 12] are defined for laminated glass structures under static bending to obtain the same
extreme values of deflection and stresses.
In this paper, this effective thickness concept is applied to vibrating laminated glass beams, and the dynamic effective thickness is used for the modal
parameters calculation. The analytical expressions for the natural angular
frequencies for a monolithic beam are then given by, e.g., [19, Eq. (22)],
k+1
ωef2
β 4 E1 h3ef (k ωef )
,
=
12m̄
(19)
where β is the wavenumber, E1 is the Young modulus of glass, m̄ is the mass
per unit length and width of the beam (m̄ = ρ1 h1 +ρ2 h2 +ρ3 h3 ), and hef (k ωef )
is the dynamic effective thickness. Two expressions for this effective thickness
will be introduced in the next two sections. Note that the wavenumbers are
usually expressed in the form β = β̂/l, where l is the length of the beam and
β̂ depends on boundary conditions and a mode shape [7, Chapter 18].
Due to the dependency of the dynamic effective thickness on frequency,
the search for ωef2 requires iterations. When the convergence is achieved
according to the criterion (17), the frequency fHz and the loss factor η are
determined from the complex valued ωef2 from Eqs. (9–10).
11
4.2. Dynamic effective thickness concept (DET)
The dynamic effective thickness was introduced by López-Aenlle and
Pelayo [19], and the iterative algorithm was extended to plates in [20]. This
complex-valued effective thickness was derived from the closed-form formula
for effective stiffness by Ross et al. [31] for a three-layer, simply-supported
beam with purely elastic face layers and a linearly viscoelastic core. However,
this analytical model is also used for different boundary conditions using the
known relevant wavenumbers, e.g., for free-free beams in [20].
Under the assumption that the glass parameters are the same for both
glass layers (E3 = E1 , ν3 = ν1 , and ρ3 = ρ1 ) the expression for the dynamic
effective thickness from [20] holds
v
u
−1 !
u
h
1
3
,
(20)
hef (ωef ) = t(h31 + h33 ) 1 + Y 1 +
g(ωef )(h1 + h3 )
where the geometric parameter
Y =
12h1 h3 (0.5h1 + h2 + 0.5h3 )
(h31 + h33 )(h1 + h3 )
(21)
depends on the thicknesses of individual layers, and the shear parameter
g(ωef ) =
G2 (ωef )
E1 h3 h2 β 2
(22)
additionally includes the wavenumber β, the Young modulus of glass E1 , and
the complex shear modulus of polymer G2 (ωef ); recall Eq. (2).
4.3. Enhanced effective thickness adjusted for modal analysis (EET)
One of the effective thickness approaches for layered beams under static
loading is the enhanced effective thickness method by Galuppi and RoyerCarfagni [12] or Galuppi et al. [10]. It is derived from an energy-based variational formulation. Under the same assumption as in the previous section
(E3 = E1 , ν3 = ν1 , and ρ3 = ρ1 ), the enhanced effective thickness for the
deflection is then expressed as [12, Eq. (4.14)],
v
u
1
u
,
(23)
hef = u
3
ζ
1−ζ
t
+
h31 + h33 + 12Is h31 + h33
12
with the coefficient of shear cohesion ζ, the non-dimensional ratio of the
glass and interlayer stiffnesses µ, and the shape coefficient ψ dependent on a
normalized shape of the deflection curve g(x) of a homogeneous beam. These
parameters follow from
R l 00
(g (x))2 dx
G2 b
1
0
.
(24)
,
µ=
ζ=
,
ψ = Rl
I1 + I3 A1 A3
0 (x))2 dx
E1 h2
(g
0
1+
ψ
µItot A1 + A3
The cross-section areas Ai for i = 1, 2, 3 and the second moments of area Ii ,
Itot , and Is are defined by relations, recall Figure 1,
Ii =
1 3
bh ,
12 i
Itot = I1 + I3 +
Is =
Ai = bhi ,
(25a)
A1 A3
(h2 + 0.5(h1 + h3 ))2 ,
A1 + A3
(25b)
h1 h3
(h2 + 0.5(h1 + h3 ))2 .
h1 + h3
(25c)
Two intuitive adjustments of this method were made for its use in modal
analysis. In Eq. (24), the complex-valued shear modulus for interlayer from
Eq. (2) was used for coefficient µ, and the shape function of the deflection
under static loading g was replaced by the one corresponding to the n-th
mode shape of a monolithic beam under given boundary conditions, e.g.,
[7, Chapter 18]. The resulting parameters ψ, corresponding to three basic
boundary conditions, are summarized in Table 1.
Beam
simply-supported
fixed-fixed
free-free
1
π /l2
40.7/l2
10.1/l2
2
Mode
2
(2π)2 /l2
82.6/l2
34.9/l2
3
(3π)2 /l2
148/l2
78.2/l2
Table 1: Shape coefficients ψ for three basic boundary conditions and the first three mode
shapes for a beam of the length l.
This effective thickness is also complex-valued, as in the DET approach
in the previous section, and the natural frequencies and the damping are
evaluated according to Eq. (10). Note that this procedure can be extended
to the plate structures, similarly to its static variant [11, 10], but we leave
this to future work.
13
5. Case study
In this section, the usability of the three methods introduced above is
assessed for laminated glass beams. The section is divided into two parts:
the first introducing the selected test examples and the second discussing
the results and the effect of input data on the usability of the modal strain
energy method and the two effective thickness approaches.
5.1. Examples
We examined a collection of 63 examples, which results from all combinations of input data for simply-supported (S-S), free-free (F-F), and clampedclamped (C-C) beams in Table 2.
Attribute
Options/Values
supports
S-S, F-F, C-C
l×b
1 × 0.1
h1 /h2 /h3
10/0.76/10, 15/0.76/5, 10/1.52/10
interlayer
SGPM , TPUM , PVBM , PVBS , PVBA
temperature
25, 50 (for PVBS and PVBA )
Units
m×m
mm/mm/mm
◦
C
Table 2: Input data and boundary conditions.
The length l and the width b of the laminated beam were constant,
whereas three cross-section configurations h1 /h2 /h3 were considered to assess the effect of symmetry/asymmetry of the cross-section layout and the
thickness of the interlayer, see Figure 1.
Material properties of glass and three interlayer materials from three different sources, Figure 3, are summarized in Appendix A. The calculations
were carried out for two ambient temperatures: the room temperature 25 ◦ C
for all cases and the elevated temperature 50 ◦ C for two PVB-interlayers,
where the temperature-shift parameters were specified.
Finally, for the sake of completeness, Table 3 contains the relation for
wavenumbers which we used in effective thickness approaches.
5.2. Results and discussion
The natural frequencies and loss factors for the first three modes2 were
determined according to the detailed and simplified algorithms described in
2
For F-F beam the first three modes corresponding to the rigid body motion were
skipped in all solvers and are not considered in this comparison.
14
Beam
simply-supported
fixed-fixed or free-free
1
π/l
4.7300/l
Mode
2
2π/l
7.8532/l
3
3π/l
10.996/l
Table 3: Wavenumbers β for three basic boundary conditions and first three mode shapes
for a beam of the length l.
Sections 3 and 4. We set the tolerance tol to 10−5 for all methods. For finite
element solvers, each beam layer was discretized by 200 elements. Compared
to the results obtained by using a refined discretization by 300 elements per
length, the largest errors for all tested examples is below 0.03% for natural
frequencies and below 0.8% for loss factors.
The modal response obtained by the simplified methods is compared
against the reference method based on the complex-valued solver using the
Newton method, recall Section 3.4. Each method is discussed in a separate
section.
5.2.1. MSE against CNM
The box-plots in Figure 5 visualize the errors of natural frequencies and
loss factors obtained using the MSE method against the results of the reference method (CNM). The red mark inside the box indicates the median,
the bottom and top edges of the box indicate the 25th and 75th percentiles,
respectively. The boxplot whiskers have the standard maximum length of
1.5 times the interquartile range, and the remaining data points are outliers
plotted individually.
For natural frequencies, the error for all 63 examples is less than 4%.
More specifically, the error is less than 1.5% in 75% of the tested cases and
the average error is less than 0.5%. Thus, this solver provide a reasonable
approximation of natural frequencies, which are mostly sufficiently accurate
for design purposes. There is no substantial difference in errors for the three
tested boundary conditions, but the errors slightly decrease for higher mode
shapes.
However, the loss factors differ significantly from the reference solution.
The errors stay below 15% for 75% of configurations, but they increase up
to 42% in some cases. The error decreases with an increasing number of
kinematic boundary constraints or for higher modes. The highest errors in
loss factors correspond to the first mode shape; the errors decrease to about
15
ERROR IN LOSS FACTOR [%]
ERROR IN FREQUENCY [%]
0
-1
-2
-3
S-S
F-F
C-C
-4
1
2
3
1
2
3
1
2
S-S
40
F-F
C-C
30
20
10
0
3
1
2
3
MODE SHAPE
1
2
3
1
2
3
MODE SHAPE
Figure 5: Errors in natural frequencies and loss factor by the modal strain energy method
(MSE) against the reference solution (CNM) for first three mode shapes for simplysupported (SS), free-free (F-F), and clamped-clamped (C-C) beams.
FREQUENCY MSE [Hz]
S-S BEAM
50
F-F BEAM
120
PCC = 0.999
C-C BEAM
120
PCC = 0.999
100
100
80
80
60
60
PCC = 0.999
40
30
30
40
50
60
80
100
120
60
80
100
120
20
30
LOSS FACTOR MSE [%]
FREQUENCY CNM [Hz]
30
30
PCC = 0.993
30
PCC = 0.995
PCC = 0.991
20
20
20
10
10
10
0
0
0
0
10
20
30
0
10
20
30
0
10
LOSS FACTOR CNM [%]
25°C:
SGPM
TPUM
50°C:
PVB S
PVB A
PVB M
PVB S
PVB A
Figure 6: Quantile-quantile plot for the first mode shape for natural frequencies and loss
factors. The response by the simplified method (MSE) is plotted against the reference
solution (CNM) along with the Pearson correlation coefficient (PCC). The Maxwell chain
parameters for PVB, TPU, and SGP are taken from [24] (M), [33] (S), and [1] (A).
16
25% for the second and below 15% for the third mode shape.
For further consideration, the quantile-quantile plots in Figure 6 show
the values of natural frequencies and loss factors by the simplified method
(MSE) against the reference solution (CNM) corresponding to the first mode
shape for all cases from Table 2. These values of frequencies and loss factors
are strongly influenced by the effect of temperature and the interlayer type.
Also, the highest errors for loss factors up to 42% appears only for the PVBM
foil; the errors remain below 18% for the other foils. The samples with the
PVBA foil show entirely different response than those with the other two
PVB foils because this the corresponding eigenfrequencies fall outside the
frequency range of this Prony series [1].
5.2.2. DET against CNM
The comparison of the modal response by the dynamic effective thickness
method (DET) against that by the reference method (CNM) are, similarly
to the previous section, shown in terms of errors of natural frequencies and
loss factors in Figure 7 and quantile-quantile plots in Figure 8.
It is clear that this method gives very accurate results for the simplysupported beams (SS). The errors are below 1% in frequencies, and the loss
factor predictions are quite accurate as well with errors below 10%. For the
other boundary conditions, the method does not provide satisfactory approximations, especially for loss factors, even if we used an adjusted wavenumber
k for given boundary conditions. This can be also seen in the quantilequantile plot in Figure 8. The errors in frequencies do not exceed 15% (the
75th percentile is below 10%) and the errors in loss factors 85%. The errors
are decreasing for the higher mode shapes.
5.2.3. EET against CNM
In analogy to the previous sections, the errors associated with the EET
method against the reference solution (CNM) are plotted in Figure 9 and the
quantile-quantile plots appear in Figure 10.
For the simply supported beam, the errors remain below 1% for frequencies and below the 10% for loss factors, as in the case of the DET method. For
the other boundary conditions, the errors in both, frequencies and loss factors, are lower than in the case of the DET approach. More specifically, they
are below 5% for frequencies and below 22% for loss factors. The highest errors appear for the free-free beams, the lowest again for the simply-supported
beam. This method provides good approximations of natural frequencies for
17
10
5
0
-5
-10
S-S
-15
1
2
F-F
3
1
2
C-C
3
1
2
100
ERROR IN LOSS FACTOR [%]
ERROR IN FREQUENCY [%]
15
50
0
-50
3
S-S
1
2
F-F
3
MODE SHAPE
1
2
C-C
3
1
2
3
MODE SHAPE
Figure 7: Errors in natural frequencies and loss factor by the dynamic effective thickness
method (DET) against the reference solution (CNM) for first three mode shapes for simplysupported (SS), free-free (F-F), and clamped-clamped (C-C) beams.
FREQUENCY DET [Hz]
S-S BEAM
50
F-F BEAM
120
PCC = 1
C-C BEAM
120
PCC = 0.993
100
100
80
80
60
60
PCC = 0.992
40
30
30
40
50
60
80
100
120
60
80
100
120
20
30
LOSS FACTOR DET [%]
FREQUENCY CNM [Hz]
30
30
PCC = 0.999
30
PCC = 0.931
PCC = 0.891
20
20
20
10
10
10
0
0
0
10
20
30
0
0
10
20
30
0
10
LOSS FACTOR CNM [%]
25°C:
SGPM
TPUM
50°C:
PVB S
PVB A
PVB M
PVB S
PVB A
Figure 8: Quantile-quantile plot for the first mode shape for natural frequencies and loss
factors. The response by the dynamic effective thickness method (DET) is plotted against
the reference solution (CNM) along with the Pearson correlation coefficient (PCC). The
Maxwell chain parameters for PVB, TPU, and SGP are taken from [24] (M), [33] (S), and
[1] (A).
18
ERROR IN LOSS FACTOR [%]
ERROR IN FREQUENCY [%]
4
3
2
1
0
1
2
3
1
2
3
1
2
10
5
0
-5
-10
-15
-20
3
1
2
3
MODE SHAPE
1
2
3
1
2
3
MODE SHAPE
Figure 9: Errors in natural frequencies and loss factor by the adjusted enhanced effective
thickness method (EET) against the reference solution (CNM) for first three mode shapes
for simply-supported (SS), free-free (F-F), and clamped-clamped (C-C) beams.
FREQUENCY EET [Hz]
S-S BEAM
50
F-F BEAM
120
PCC = 1
C-C BEAM
120
PCC = 1
100
100
80
80
60
60
PCC = 1
40
30
30
40
50
60
80
100
120
60
80
100
120
20
30
LOSS FACTOR EET [%]
FREQUENCY CNM [Hz]
30
30
PCC = 0.999
30
PCC = 1
PCC = 1
20
20
20
10
10
10
0
0
0
10
20
30
0
0
10
20
30
0
10
LOSS FACTOR CNM [%]
25°C:
SGPM
TPUM
50°C:
PVB S
PVB A
PVB M
PVB S
PVB A
Figure 10: Quantile-quantile plot for the first mode shape for natural frequencies and
loss factors. The response by the adjusted enhanced effective thickness method (EET) is
plotted against the reference solution (CNM) along with the Pearson correlation coefficient
(PCC). The Maxwell chain parameters for PVB, TPU, and SGP are taken from [24] (M),
[33] (S), and [1] (A).
19
all three boundary conditions and the best estimates of loss factors from all
tested simplified methods.
5.2.4. Comparison of simplified approaches
S-S BEAM
10
MSE
DET
SET
0
ERRORS AGAINST CNM [%]
-10
F-F BEAM
10
0
-10
C-C BEAM
10
0
PVB S - 25°C
PVB S - 50°C
PVB A - 25°C
PVB A - 50°C
PVB M - 25°C
TPUM - 25°C
PVB S - 50°C
SGPM - 25°C
PVB A - 50°C
10/1.52/10
PVB S - 25°C
PVB A - 25°C
TPUM - 25°C
PVB M - 25°C
SGPM - 25°C
PVB S - 25°C
15/0.76/5
PVB S - 50°C
PVB A - 25°C
PVB A - 50°C
TPUM - 25°C
SGPM - 25°C
PVB M - 25°C
10/0.76/10
-10
Figure 11: Summary of errors in natural frequencies obtained by all simplified methods against the reference solution (CNM) corresponding to the first mode shape for all
tested cases. Simplified approaches: modal strain energy (MSE), dynamic effective thickness (DET), and enhanced effective thickness (EET) methods.
Finally, the errors of both quantities predicted by all three methods are
summarized in Figures 11 and 12 for all tested examples for the first mode
shape. The present study thus shows that:
• For natural frequencies, modal strain energy and effective thickness
methods give very good approximations (below 4% for MSE and EET,
below 15% for DET).
20
S-S BEAM
100
MSE
DET
SET
50
ERRORS AGAINST CNM [%]
0
-50
F-F BEAM
100
50
0
-50
C-C BEAM
100
10/0.76/10
15/0.76/5
10/1.52/10
50
PVB S - 50°C
PVB A - 50°C
PVB S - 25°C
PVB A - 25°C
PVB M - 25°C
TPUM - 25°C
PVB S - 50°C
SGPM - 25°C
PVB S - 25°C
PVB A - 25°C
PVB A - 50°C
TPUM - 25°C
PVB M - 25°C
SGPM - 25°C
PVB S - 25°C
PVB S - 50°C
PVB A - 25°C
PVB A - 50°C
PVB M - 25°C
TPUM - 25°C
-50
SGPM - 25°C
0
Figure 12: Summary of errors in loss factors obtained by all simplified methods against
the reference solution (CNM) corresponding to the first mode shape for all tested cases.
Simplified approaches: modal strain energy (MSE), dynamic effective thickness (DET),
and enhanced effective thickness (EET) methods.
• The effective thickness approaches provide excellent results for simplysupported laminated glass beams; the error is less than 1%.
• However, all methods predict in some cases the loss factors of laminated glass beams with errors in tens of percent (up to 22% for EET,
42% for MSE, and 85% for DET). Therefore, they can provide only an
informative estimate of damping.
• For simply-supported beams, the effective thickness methods deliver
the loss factors with errors less than 5% for the room temperature and
below 10% for the evaluated temperature typical of, e.g., an exterior
21
facade panel in summer.
• The algorithm based on the EET provides the best estimates of loss
factors from all tested simplified methods.
• The symmetry or the asymmetry of the cross-section geometry does
not influence the level of errors.
It is worth mentioning that for the simply-supported beams the enhanced
effective method adjusted for dynamics (EET) gives exactly the same results
as the original dynamic effective thickness method (DET). We can see from
Tables 1 and 3 that shape coefficients are squared wavenumbers for a simplysupported beam ψ = β 2 . If we replaced the used shape coefficient derived
in Table 1 with squared wavenumbers from Table 3, we would obtain the
same results from both methods for all boundary conditions. A more detailed
study of this unexpected connection goes beyond the current work and will
be performed separately.
6. Conclusions
Four methods for modal analysis of laminated glass structures were introduced in this paper, i.e., the numerical complex-valued eigensolver based
on the Newton method, the real-valued eigensolver complemented with the
modal strain energy method, and the original and enhanced dynamic effective
thickness method. The aim of this paper was to assess the usability of the
last three practical methods, comparing their predictions with the complexvalued eigensolver. For the enhanced effective thickness method by Galuppi
and Royer-Carfagni, we proposed and presented our extension of the current
method for modal analysis of laminated glass. The following conclusions can
be made from the present study:
1. Our study underlines the importance of being careful when predicting
the damping in laminated glass. The loss factor is a sensitive quantity
which affects the errors of approximations provided by simplified methods. In particular, the natural frequencies can be predicted with errors
less than 5% using a suitable method, but the same does not hold for
the loss factors.
2. The level of errors in approximated quantities depends on the boundary
conditions for the effective thickness approaches and on the parameters
22
of the generalized Maxwell model for all methods. Material parameters for the interlayers from the literature leads to different natural
frequencies and loss factors even for the same type of polymer.
3. The enhanced effective thickness approach adjusted for modal analysis
provides approximations of both quantities, natural frequencies and
loss factors, with lower errors compared to the other two simplified
procedures.
Simplified methods and approaches reduce computational time and cost
and have an important place in the design of engineering structures. Under
the above limitations, our study confirms their suitability for modal analysis
of laminated glass structures.
Acknowledgments. This publication was supported by the Czech Science
Foundation under project No. 16-14770S.
References
[1] Andreozzi, L., Bati, S. B., Fagone, M., Ranocchiai, G., and Zulli, F.
(2014). Dynamic torsion tests to characterize the thermo-viscoelastic properties of polymeric interlayers for laminated glass. Construction and Building Materials, 65:1–13.
[2] Bedon, C., Kalamar, R., and Eliášová, M. (2017). Low velocity impact
performance investigation on square hollow glass columns via full-scale
experiments and finite element analyses. Composite Structures, 182:311–
325.
[3] Bennison, S. J., Qin, M. H., and Davies, P. S. (2008). High-performance
laminated glass for structurally efficient glazing. Innovative light-weight
structures and sustainable facades, Hong Kong.
[4] Bilasse, M., Daya, E., and Azrar, L. (2010). Linear and nonlinear vibrations analysis of viscoelastic sandwich beams. Journal of Sound and
Vibration, 329:4950–4969.
[5] Biolzi, L., Cagnacci, E., Orlando, M., Piscitelli, L., and Rosati, G. (2014).
Long term response of glass–PVB double-lap joints. Composites Part B:
Engineering, 63:41–49.
23
[6] Christensen, R. (1982). Theory of Viscoelasticity: An Introduction. Elsevier, second edition.
[7] Clough, R. W. and Penzien, J. (2003). Dynamics of structures. Computers & Structures, Inc.
[8] Daya, E. and Potier-Ferry, M. (2001). A numerical method for nonlinear
eigenvalue problems application to vibrations of viscoelastic structures.
Computers & Structures, 79:533–541.
[9] Duser, A. V., Jagota, A., and Bennison, S. J. (1999). Analysis of
glass/polyvinyl butyral laminates subjected to uniform pressure. Journal
of Engineering Mechanics, 125(4):435–442.
[10] Galuppi, L., Manara, G., and Royer-Carfagni, G. (2012). Practical expression for the design of laminated glass. Composites Part B: Engineering,
45:1677–1688.
[11] Galuppi, L. and Royer-Carfagni, G. (2012a). The effective thickness of
laminated glass plates. Journal of Mechanics of Materials and Structures,
7:375–400.
[12] Galuppi, L. and Royer-Carfagni, G. F. (2012b). Effective thickness of
laminated glass beams: New expression via a variational approach. Engineering Structures, 38:53–67.
[13] Giovanna, R., Zulli, F., Andreozzi, L., and Fagone, M. (2016). Test
methods for the determination of interlayer properties in laminated glass.
Journal of Materials in Civil Engineering, ASCE, 04016268:1–8.
[14] Hamdaoui, M., Akoussan, K., and Daya, E. (2016). Comparison of nonlinear eigensolvers for modal analysis of frequency dependent laminated
visco-elastic sandwich plates. Finite Elements in Analysis and Design,
121:75–85.
[15] Hooper, P., Blackman, B., and Dear, J. (2012). The mechanical behaviour of poly(vinyl butyral) at different strain magnitudes and strain
rates. Journal of Materials Science, 47:3564–3576.
[16] Huang, Z., Qin, Z., and Chu, F. (2016). Damping mechanism of elasticviscoelastic-elastic sandwich structures. Composite Structures, 153:96–107.
24
[17] Johnson, C. D. and Kienholz, D. A. (1982). Finite element prediction of
damping in structures with constrained viscoelastic layers. AIAA Journal,
20(9):1284–1290.
[18] Larcher, M., Solomos, G., Casadei, F., and Gebbeken, N. (2012). Experimental and numerical investigations of laminated glass subjected to
blast loading. International Journal of Impact Engineering, 39(1):42–50.
[19] López-Aenlle, M. and Pelayo, F. (2013). Frequency response of laminated glass elements: Analytical modeling and effective thickness. Applied
Mechanics Reviews, 65(2):020802–1–020802–13.
[20] López-Aenlle, M. and Pelayo, F. (2014). Dynamic effective thickness
in laminated-glass beams and plates. Composites Part B: Engineering,
67:332–347.
[21] Mead, D. J. (2007). The measurement of the loss factors of beams
and plates with constrained and unconstrained damping layers: A critical
assessment. Journal of Sound and Vibration, 300(3):744–762.
[22] Mead, D. J. and Markus, S. (1969). The forced vibration of a three-layer,
damped sandwich beam with arbitrary boundary conditions. Journal of
Sound and Vibration, 10(2):163–175.
[23] Mead, D. J. and Markus, S. (1970). Loss factors and resonant frequencies
of encastr damped sandwich beams. Journal of Sound and Vibration,
12(1):99–112.
[24] Mohagheghian, I., Wang, Y., Jiang, L., Zhang, X., Guo, X., Yan, Y.,
Kinloch, A., and Dear, J. (2017a). Quasi-static bending and low velocity
impact performance of monolithic and laminated glass windows employing
chemically strengthened glass. European Journal of Mechanics – A/Solids,
63:165–186.
[25] Mohagheghian, I., Wang, Y., Zhou, J., Yu, L., Guo, X., Yan, Y., Charalambides, M., and Dear, J. (2017b). Deformation and damage mechanisms of laminated glass windows subjected to high velocity soft impact.
International Journal of Solids and Structures, 109:46–62.
25
[26] Pelayo, F. and López-Aenlle, M. (2017). Natural frequencies and damping ratios of multi-layered laminated glass beams using a dynamic effective
thickness. Journal of Sandwich Structures and Materials, 0:1–25.
[27] Pelayo, F., López-Aenlle, M., Hermans, L., and Fraile, A. (2013). Modal
scaling of a laminated glass plate. 5th International Operational Modal
Analysis Conference, pages 1–10.
[28] Radke, R. J. (1996). A Matlab implementation of the implicitly restarted
Arnoldi method for solving large-scale eigenvalue problems. PhD thesis,
Rice University.
[29] Rao, M. D. (2003). Recent applications of viscoelastic damping for noise
control in automobiles and commercial airplanes. Journal of Sound and
Vibration, 262(3):457–474.
[30] Rikards, R., Chate, A., and Barkanov, E. (1993). Finite element analysis of damping the vibrations of laminated composites. Computers &
structures, 47(6):1005–1015.
[31] Ross, D., Ungar, E., and Kerwin, E. (1959). Damping of plate flexural vibrations by means of viscoelastic laminae. Proc. Colloq. Structural
Damping, American Society of Mechanical Engineers, pages 49–87.
[32] Schreiber, K. (2008). Nonlinear Eigenvalue Problems: Newton-type
Methods and Nonlinear Rayleigh Functionals. PhD thesis, Technischen
Universitat Berlin.
[33] Shitanoki, Y., Bennison, S., and Koike, Y. (2014). A practical, nondestructive method to determine the shear relaxation modulus behavior of
polymeric interlayers for laminated glass. Polymer Testing, 37:59–67.
[34] Treviso, A., Van Genechten, B., Mundo, D., and Tournour, M. (2015).
Damping in composite materials: Properties and models. Composites Part
B: Engineering, 78:144–152.
[35] Ungar, E. and Kerwin, E. (1962). Loss factors of viscoelastic systems
in terms of energy concepts. The Journal of the Acoustical Society of
America, 34(7):954–957.
26
[36] Williams, M., Landel, R., and Ferry, J. (1955). The temperature dependence of relaxation mechanisms in amorphous polymers and other glassforming liquids. Journal of the American Chemical Society, 77(14):3701–
3707.
[37] Zemanová, A., Zeman, J., and Šejnoha, M. (2017). Comparison of viscoelastic finite element models for laminated glass beams. International
Journal of Mechanical Sciences, 131–132:380–395.
Appendix A. Material properties of glass and interlayers
In this appendix, we summarize the used material properties of glass and
three different polymers (taken from the literature) in Tables A.4, A.5, A.6,
and A.7.
Soda-lime glass
Density
Young’s modulus of elasticity
Poisson’s ratio
ρ1 = ρ3
E1 = E3
ν1 = ν3
2,500 kg.m−3
72 GPa
0.22 –
Table A.4: Glass properties [19].
Density
Poisson’s ratio
ρ2
ν2
SGPM
950
0.49
TPUM
1070
0.49
PVBM , PVBS , PVBA
1100
0.49
Table A.5: Interlayer properties [24].
27
kg.m−3
–
SGPM
TPUM PVBM
Long-term and instantaneous shear moduli
G2,∞
1.8
1.56
0.22
MPa
G2,0
274.1
94.6
213.6
MPa
Coefficients and relaxation times
p
G2,p /G2,0
θp [s]
1
0.07767 0.42077 0.39262 10−6
2
0.03764 0.18113 0.19225 10−5
3
0.05631 0.19280 0.20957 10−4
4
0.06501 0.09969 0.12621 10−3
5
0.07409 0.04750 0.05694 10−2
6
0.09317 0.01928 0.01536 10−1
7
0.11867 0.00903 0.00325
100
8
0.20551 0.00414 0.00103
101
9
0.18131 0.00307 0.00077
102
10
0.05361 0.00230 0.00010
103
11
0.01856 0.00371 0.00029
104
12
0.01180 0.00004 0.00053
105
Table A.6: Parameters of the generalized Maxwell model for SGPM , TPUM , and
PVBM [24].
28
G2,∞
p
1
2
3
4
5
6
7
8
9
10
11
12
13
14
T0
C1
C2
PVBS
PVBA
Long-term shear moduli
0
0
Shear moduli and relaxation times
G2,p [MPa]
θp [s]
G2,p [MPa]
θp [s]
−7
51.25
4.273×10
0.514628 9.51×10−2
−6
31.75
3.546×10
0.280116 4.71×10−1
12.80
1.330×10−5
0.144282
2.72×100
32.90
4.279×10−5
0.086904
2.11×101
−4
39.90
2.984×10
0.076190
2.21×102
0.092202
2.12×103
37.80
2.170×10−3
−3
21.94
8.274×10
0.098780
1.74×104
25.01
2.937×10−2
0.085555
1.31×105
27.58
1.658×10−1
0.070251
1.05×106
−1
11.98
7.774×10
0.107653
2.99×107
6.345
3.293×100
2.692
1.698×101
8.718
2.041×102
0.6969
3.588×104
Parameters for temperature shifting
20.46 ◦ C
30 ◦ C
37.30
12.5
203.61 ◦ C
89 ◦ C
Table A.7: Parameters of the generalized Maxwell model for PVBS [33] and PVBA [1].
29
Appendix B. Numerical aspects
The finite element discretization is briefly presented in this section, following the exposition by Rikards et al. [30]. Our study deals with a laminated
glass beam made of three layers. After discretization, we have 18 unknowns
per element – 9 unknowns per cross section, recall Figure 4. Therefore, the
kinematics of the three-layer element is specified by the vector of nodal displacements and rotations
u1,L w1,L ϕ1,L u1,R w1,R ϕ1,R u2,L w2,L ϕ2,L
ufull
e =
u2,R w2,R ϕ2,R u3,L w3,L ϕ3,L u3,R w3,R ϕ3,R T .
Using the Timoshenko beam theory, we express 8 inter-layer compatibility
conditions corresponding to perfect horizontal and vertical adhesion at each
inter-layer interface as
h1
ϕ1,L
2
h2
u2,L + ϕ2,L
2
w1,L
w2,L
u1,L +
h2
ϕ2,L ,
2
h3
= u3,L − ϕ3,L ,
2
= w2,L ,
= w3,L ,
h1
ϕ1,R
2
h2
u2,R + ϕ2,R
2
w1,R
w2,R
= u2,L −
u1,R +
h2
ϕ2,R ,
2
h3
= u3,R − ϕ3,R ,
2
= w2,R ,
= w3,R .
= u2,R −
Therefore, we decompose the vector of the generalized nodal displacements
as follows
ue = u1,L w1,L ϕ1,L u3,L ϕ3,L u1,R w1,R ϕ1,R u3,R ϕ3,R T ,
T
uslave
=
u
w
ϕ
w
u
w
ϕ
w
,
2,L
2,L
2,L
3,L
2,R
2,R
2,R
3,R
e
where ue are the independent unknowns and uslave
are the dependent ones.
e
Then, the compatibility conditions can be written in the compact form
ufull
e = T e ue
30
with the transformation matrix
1
1
1
1
1
1
h1
h3
1
1
−4
2
4
2
1
−1
h1
h3
1
− 2h
− 2h
h2
2h2
2
2
Te =
h1
1
2
4
1
h1
− h12
− 2h
2
1
1
1
1
1
2
− h43
1
2h2
h3
− 2h
2
1
.
1
Using this elimination of unknowns, the element mass matrix M e , the initial stiffness matrix K 0,e and the matrix K c,e , recall Eq. (7), corresponding
to the independent generalized nodal displacements follow from
M e = T T M full
e T,
K 0,e = T T K full
0,e T ,
K c,e = T T K full
c,e T .
Therefore, they are 10×10 matrices derived from the original 18×18 one. We
use the consistent mass matrix and the stiffness matrices are derived using
the selective integration scheme to avoid shear locking.
31
| 5cs.CE
|
Online Representation Learning with Single and
Multi-layer Hebbian Networks for Image Classification
Yanis Bahroun and Andrea Soltoggio
arXiv:1702.06456v3 [cs.NE] 29 Jan 2018
Loughborough University, Computer Science Department,
LE11 3TU, Leicestershire, United Kingdom
{y.bahroun,a.soltoggio}@lboro.ac.uk
Abstract. Unsupervised learning permits the development of algorithms that are
able to adapt to a variety of different data sets using the same underlying rules
thanks to the autonomous discovery of discriminating features during training.
Recently, a new class of Hebbian-like and local unsupervised learning rules for
neural networks have been developed that minimise a similarity matching costfunction. These have been shown to perform sparse representation learning. This
study tests the effectiveness of one such learning rule for learning features from
images. The rule implemented is derived from a nonnegative classical multidimensional scaling cost-function, and is applied to both single and multi-layer
architectures. The features learned by the algorithm are then used as input to a
SVM to test their effectiveness in classification on the established CIFAR-10 image dataset. The algorithm performs well in comparison to other unsupervised
learning algorithms and multi-layer networks, thus suggesting its validity in the
design of a new class of compact, online learning networks.
Keywords: Classification; Competitive learning; Feature learning; Hebbian learning; Online algorithm; Neural networks; Sparse coding; Unsupervised learning.
1 Introduction
Biological synaptic plasticity is hypothesized to be one of the main phenomena responsible for human learning and memory. One mechanism of synaptic plasticity is inspired
by the Hebbian learning principle which states that connections between two units, e.g.,
neurons, are strengthened when they are simultaneously activated. In artificial neural
networks, implementations of Hebbian plasticity are known to learn recurring patterns
of activations. The use of extensions of this rule, such as Oja’s rule [8] or the Generalized Hebbian rule, also called Sanger’s rule [14], have permitted the development of
algorithms that have proved particularly efficient at tasks such as online dimensionality reduction. Two important properties of brain-inspired models, namely competitive
learning [13] and sparse coding [9] can be performed using Hebbian and anti-Hebbian
learning rules. Such properties can be achieved with inhibitory connections, which extend the capabilities of such learning rules beyond simple extraction of the principal
component of input data. The continuous and local update dynamics of Hebbian learning also make it suitable for learning from a continuous stream of data. Such an algorithm can take one image at a time with memory requirements that are independent of
the number of samples.
2
Online Representation Learning with Hebbian Networks
This study employs Hebbian/anti-Hebbian learning rules derived from a similarity
matching cost-function [11] and applies it to perform online unsupervised learning of
features from multiple image datasets. The rule proposed in [11] is applied here for the
first time to online features learning for image classification with single and multi-layer
architectures. The quality of the features is assessed visually and by performing classification with a linear classifier working on the learned features. The simulations show
that a simple single-layer Hebbian network can outperform more complex models such
as Sparse Autoencoders (SAE) and Restricted Boltzmann machines (RBM) for image
classifications tasks [2]. When applied to multi-layer architectures, the rule learns additional features. This study is the first of its kind to perform multi-layer sparse dictionary
learning based on the similarity matching principle developed in [11] and to apply it to
image classification.
2 Hebbian/anti-Hebbian Network Derived From a Similarity
Matching Cost-Function
The rule implemented by the Hebbian/anti-Hebbian network used in this work derives
from an adaptation of Classical MultiDimensional Scaling (CMDS). CMDS is a popular
embedding technique [3]. Unlike most dimensionality reduction techniques, e.g. PCA,
the CMDS uses as input the matrix of similarity between inputs to generate a set of
embedding coordinates. The advantage of MDS is that any kind of distance or similarity
matrix can be analyzed. However, in its simplest form, CMDS produces dense features
maps which are often unsuitable when considered for image classification. Therefore an
adaptation of the CMDS introduced recently in [11] is used to overcome this weakness.
The model implemented is a nonnegative classical multidimensional scaling that has
three properties: it takes a similarity matrix as input, it produces sparse codes, and can
be implemented using a new biologically plausible Hebbian model. The Hebbian/antiHebbian rule introduced in [11] is given as follows: for a set of inputs xt ∈ Rn for
t ∈ {1, . . . , T }, the concatenation of the inputs defines an input matrix X ∈ Rn×T .
The output matrix Y of encodings is an element of Rm×T that corresponds to a sparse
overcomplete representation of the input if m > n, or to a low-dimensional embedding
of the input if m < n. The objective function proposed by [11] is:
Y ∗ = arg min kX ′ X − Y ′ Y k2F
(1)
Y ≥0
where F is the Frobenius norm and X ′ X is the Gram matrix of the inputs which corresponds to the similarity matrix. Solving Eq.1 directly requires storing Y ∈ Rm×T
+
which increases with time T making online learning difficult. Thus instead an online
learning version of Eq.1 is expressed as:
(y T )∗ = arg min kX ′ X − Y ′ Y k2F
.
(2)
y T ≥0
The components of the solution of Eq.2, found in [11] using coordinate descent, are :
(yiT )∗ = max WiT xT − MiT (y T )∗ , 0
∀i ∈ {1, . . . , m},
(3)
Online Representation Learning with Hebbian Networks
where
WijT
=
TP
−1
t=1
TP
−1
t=1
W
T
yit xtj
;
MijT
=
(yit )2
TP
−1
t=1
TP
−1
t=1
3
yit yjt
1i6=j
.
(4)
(yit )2
T
and M can be found using the recursive formulations:
T −1
T −1 T −1
T −1 T −1
T
Wij = Wij + yi (xj
− Wij yi ) ŶiT
T −1
T −1 T −1
T −1 T −1
T
Mij6
=
M
+
y
(y
−
M
y
)
ŶiT
=i
ij
i
j
ij
i
ŶiT = ŶiT −1 + (yiT −1 )2
T
.
(5)
(6)
(7)
T
W (green arrows) and M (blue arrows) can be interpreted respectively as feedforward synaptic connections between the input and the hidden layer and lateral synaptic inhibitory connections within the hidden layer. The weight matrices are of fixed sizes
and updated sequentially, which makes the model suitable for online learning. The architecture of the Hebbian/anti-Hebbian network is represented in Figure 1.
Input
layer
Hidden
layer
Output
layer
y1T
xT1
xT2
xT3
y2T
y3T
y4T
Feed-forward
connections W T
Lateral synaptic
connections M T
Fig. 1: Hebbian/anti-Hebbian network with lateral connections derived from Eq.2
3 A Model to Learn Features From Images
In the new model presented in this study, the input data vectors (x1 , . . . , xT ) are composed of patches taken randomly from a training dataset of images. For every new input
xt presented, the model first computes a sparse post-synaptic activity y t . Second, the
synaptic weights are modified based on local Hebbian/anti-Hebbian learning rules requiring only the current pre- post-synaptic neuronal activities. The model can be seen as
a sparse encoding followed by a recursive updating scheme, which are both well suited
to solve large-scale online problems.
A multi-class SVM classifies the pictures using output vectors obtained by a simple pooling of the feature vectors, Y ∗ , obtained for the input images from the trained
network. In particular, given an input image, each neuron in the output layer produces
a new image, called a feature map, which is pooled in quadrants [2] to form 4 terms of
the input vector for the SVM.
4
Online Representation Learning with Hebbian Networks
3.1 Multi-layer Hebbian/anti-Hebbian Neural Network
In the proposed approach, layers of Hebbian/anti-Hebbian network are stacked similarly to the Convolutional DBN [4], and Hierarchical K-means. In the multi-layer
Hebbian/anti-Hebbian network, both the weights of the first layer and second layer are
continuously updated. Unlike other CNNs, the non-linearity used in each layer is not
only due to the positivity constraint, but to the combination of a rectified linear unit activation function and of interneuronal competition. This model combines the powerful
architecture of convolutional neural networks using ReLU activation with interneuronal
competition, while all synaptic weights are updated using online local learning rules. In
between layers, a 2 × 2 average pooling is used to downsample the feature maps.
3.2 Overcompleteness of the Representation and Multi-resolution
As part of the evaluation of the new model, it is important to assess its performance with
different sizes (m) of the hidden layers. If the number of neurons exceeds the size of
the input (m > n), the representation is called overcomplete. Overcompleteness may be
beneficial, but requires increased computation, particularly for deep networks in which
the number of neurons has to grow exponentially in order to keep this property. One
motivation for overcompleteness is that it may allow more flexibility in matching the
output structure with the input. However, not all learning algorithms can learn and take
advantage of overcomplete representations. The behaviour of the algorithm is analysed
in the transition between undercomplete (m < n) and overcomplete (m > n) representations.
Although the model might benefit from a large number of neurons, from a practical
perspective an increase in the number of neurons is a challenge for such models due
to the number of operations required in the coordinate descent. In order to limit the
computational cost of training a large network while still benefiting from overcomplete
representations, this study proposes to train simultaneously three single-layer neural
networks, each of them having different receptive field sizes (4 × 4, 6 × 6, and 8 × 8
pixels). Thus, a variation of the model tested here is composed of three different networks. This architecture of parallel networks with different receptive field sizes requires
less computational time and memory than a model with only one receptive field size and
the same total number of neurons, because the synaptic weights only connect neurons
within each neural network. This model will be called multi-resolution in the following.
3.3 Parameters and Preprocessing
The architecture used here has the following tunable parameters: the receptive field size
(n) of the neurons and the number of neurons (m). These parameters are standard to
CNNs but their influence on this online feed-forward model needs to be investigated.
For computer vision models, understanding the influence of input preprocessing is
of critical importance for both biological plausibility and practical applicability. Recent
findings [1], confirm partial decorrelation of the input signal in the retinal ganglion
cells. The influence of input decorrelation by applying whitening will be investigated.
Online Representation Learning with Hebbian Networks
5
4 Results
The effectiveness of the algorithm is assessed by measuring the performance on an
image classification task. We acknowledge that classification accuracy is at best an implicit measure evaluating the performance of representation learning algorithms, but
provides a standardised way of comparing them. In the following, single and multilayer Hebbian/anti-Hebbian neural networks combined with the standard multi-class
SVM are trained on the CIFAR-10 dataset [5].
4.1 Evaluation of the Single-layer Model
A first experiment tested the performance of the model with and without whitening of
the input data. Although there exist Hebbian networks that can perform online whitening [10], an offline technique based on singular value decomposition [2] is applied in
these experiments. Figure 2a and 2b show the features learned by the network from
raw input and whitened input respectively. The features learned from raw data (Fig.2a)
are neither sharp nor localised filters and just slightly capture edges. With whitened
data (Fig.2b), the features are sharp, localised, and resemble Gabor filters, which are
observed in the primary visual cortex [9].
Fig. 2: Sample of features learned from raw (2a) and whitened input (2b). Classification
accuracy with raw (2c) and whitened input (2d).
(a) Features learned from raw data
(b) Features learned from whitened data
(c) Accuracy using raw data
(d) Accuracy using whitened data
72
70
74
400 Neurons
500 Neurons
600 Neurons
800 Neurons
Accuracy in %
Accuracy in %
74
68
66
64
3
72
70
68
400 Neurons
500 Neurons
600 Neurons
800 Neurons
66
64
5
7
Receptive Field Size
9
3
5
7
Receptive Field Size
9
In a second set of experiments, the performance of the network was tested for varying receptive field sizes (Fig.2c-2d) and varying network sizes (400, 500, 600, and 800
neurons). The results show that the performance peaks at a receptive field size of 7
pixels and then begins to decline. This property is common to most unsupervised learning algorithms [2], showing the difficulty of learning spatially extended features. Fig-
6
Online Representation Learning with Hebbian Networks
ures 2c and 2d also show that for every configuration, the performance of the algorithm
is largely and uniformly improved when whitening is applied to the input.
4.2 Comparison to State-of-the-art Performances and Online Training
Various unsupervised learning algorithms have been tested on the CIFAR-10 dataset.
Spherical K-means, in particular, proved in [2] to outperform autoencoders and restricted Boltzmann machines, providing a very simple and efficient solution for dictionary learning for image classification. Thus, spherical K-means is used here as a
benchmark to evaluate the performance of the single-layer network. As with other unsupervised learning algorithms, increasing the number of output neurons to reach overcompleteness also improved classification performance (Fig.3a). Although the singlelayer neural network has a higher degree of sparsity than the K-means proposed in [2]
(results not shown here), they appear to have the same performance in their optimal
configurations (Fig.3a).
The classification accuracy of the network during training is shown in Fig.3b. The
graph (Fig.3b) suggests that the features learned by the network over time help the
system improve the classification accuracy. This is significant because it demonstrates
for the first time the effectiveness of features learned with a Hebb-like cost-function
minimisation. It is not obvious a priori that the online optimisation of a cost-function
for sparse similarity matching (Eq.2) produces features suitable for image classification.
Fig. 3: (a) Proposed model vs K-means, (b) Classification accuracy
(b) Online training
(a) Optimal setup vs K-means
80
70
60
50
40
0
200
Hebbian Network
K-Means
400
600
800
Number of Neurons
Accuracy in %
Accuracy in %
80
70
60
50
40
30
20
0
5
10
Hebbian Network
15
20
Number of input streamed in 10k
As shown in Table 1, the multi-resolution network outperforms the single resolution
network and K-means algorithm [2], reaching 80.42% accuracy on the CIFAR-10. The
multi-resolution model shows better performance, while requiring less computation and
memory than the single resolution model. It also outperforms the single layer NOMP
[6], sparse TIRBM [15], CKN-GM and CKN-PM [7], which are more complex models.
It was outperformed only by combined models or models with three layers or more.
Online Representation Learning with Hebbian Networks
Algorithm
Accuracy
Single-Layer, Single Resolution (4k neurons)
Single-Layer, Multi-Resolution (3×1.6k neurons)
79.58 %
80.42 %
Single-layer K-means [2] (4k neurons)
Multi-layer K-means [2] (3 Layers, >4k neurons)
Sparse RBM
Convolutional DBN [4]
Sparse TIRBM [15] (4k neurons)
TIOMP-1/T [15] (combined transformations, 4k neurons)
Single Layer NOMP [6] ( 5k neurons)
Multi-Layer NOMP [6] (3 Layers, >4k neurons)
Multi-Layer CKN-GM [7]
Multi-Layer CKN-PM [7]
Multi-Layer CKN-CO [7] (combining CKN-GM & CKN-PM)
79.60 %
82.60 %
72.40 %
78.90 %
80.10%
82.20 %
78.00 %
82.90 %
74.84 %
78.30 %
82.18 %
7
Table 1: Comparison of the single-layer network with unsupervised learning algorithms
on CIFAR-10.
4.3 Evaluation of the Multi-layer Model
A single resolution, double-layer neural network with different numbers of neurons in
each layer was trained similarly to the single-layer network in the previous section. In
Table 2, φ1 and φ2 correspond respectively to the features learned by the first and second
layer. The results show that φ2 alone are less discriminative than φ1 as indicated in Fig.
3a. However, when combined (φ1 + φ2 ) the model achieves better performance than
each layer considered separately. Nevertheless, the preliminary results indicate that the
sizes of the two layers unevenly affect the performance of the network. A future test may
investigate if a multi-layer architecture can outperform the largest shallow networks.
50
φ2 54.9%
100 Neurons Layer 1
φ1 +φ2 67.2%
φ2 55.8%
200 Neurons Layer 1
φ1 +φ2 69.9%
#Neurons Layer 2
100 200 400
59.7% 64.7% 68.7%
68.1% 69.9% 72.4%
60.6% 65.3% 70.3%
70.8% 71.9% 73.7%
800
71.45%
73.81%
72.7%
75.1%
Table 2: Classification accuracy for a two-layer network.
5 Conclusion
This work proposes a multi-layer neural network exploiting Hebbian/anti-Hebbian rules
to learn features for image classification. The network is trained on the CIFAR-10 image dataset prior to feeding a linear classifier. The model successfully learns online
more discriminative representations of the data when the number of neurons and the
number of layers increase. The overcompleteness of the representation is critical for
learning relevant features. The results show that a minimum unsupervised learning time
is needed to optimise the network leading to better classification accuracy. Finally, one
8
Online Representation Learning with Hebbian Networks
key factor in improving image classification is the appropriate choice of the receptive
field size used for training the network.
Such findings prove that neural networks can be trained to solve problems as complex as sparse dictionary learning with Hebbian learning rules, delivering competitive
accuracy compared to other encoder, including deep neural networks. This makes deep
Hebbian networks attractive for building large-scale image classification systems. The
competitive performances on the CIFAR-10 suggests that this model can offer an alternative to batch trained neural networks. Ultimately, thanks to its bio-inspired architecture and learning rules, it also stands as a good candidate for memristive devices [12].
Moreover, if a decaying factor is added to the proposed model that might result in an
algorithm that can deal with complex datasets with temporal variations of the distributions.
References
1. Abbasi-Asl, R., Pehlevan, C., Yu, B., Chklovskii, D.B.: Do retinal ganglion cells project
natural scenes to their principal subspace and whiten them? arXiv preprint arXiv:1612.03483
(2016)
2. Coates, A., Lee, H., Ng, A.Y.: An analysis of single-layer networks in unsupervised feature
learning. In: AISTATS 2011. vol. 1001 (2011)
3. Cox, T.F., Cox, M.A.: Multidimensional scaling. CRC press (2000)
4. Krizhevsky, A., Hinton, G.: Convolutional deep belief networks on cifar-10. Unpublished
manuscript 40 (2010)
5. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images (2009)
6. Lin, T.h., Kung, H.: Stable and efficient representation learning with nonnegativity constraints. In: Proceedings of the 31st International Conference on Machine Learning (ICML14). pp. 1323–1331 (2014)
7. Mairal, J., Koniusz, P., Harchaoui, Z., Schmid, C.: Convolutional kernel networks. In: Advances in Neural Information Processing Systems. pp. 2627–2635 (2014)
8. Oja, E.: Neural networks, principal components, and subspaces. International journal of neural systems 1(01), 61–68 (1989)
9. Olshausen, B.A., et al.: Emergence of simple-cell receptive field properties by learning a
sparse code for natural images. Nature 381(6583), 607–609 (1996)
10. Pehlevan, C., Chklovskii, D.: A normative theory of adaptive dimensionality reduction in
neural networks. In: Advances in Neural Information Processing Systems. pp. 2269–2277
(2015)
11. Pehlevan, C., Chklovskii, D.B.: A Hebbian/anti-Hebbian network derived from online nonnegative matrix factorization can cluster and discover sparse features. In: 2014 48th Asilomar
Conference on Signals, Systems and Computers. pp. 769–775. IEEE (2014)
12. Poikonen, J.H., Laiho, M.: Online linear subspace learning in an analog array computing
architecture. CNNA 2016 (2016)
13. Rumelhart, D.E., Zipser, D.: Feature discovery by competitive learning. Cognitive science
9(1), 75–112 (1985)
14. Sanger, T.D.: Optimal unsupervised learning in a single-layer linear feedforward neural network. Neural networks 2(6), 459–473 (1989)
15. Sohn, K., Lee, H.: Learning invariant representations with local transformations. In: Proceedings of the 29th International Conference on Machine Learning (ICML-12). pp. 1311–1318
(2012)
| 1cs.CV
|
ATTACHING LEAVES AND PICKING CHERRIES TO
CHARACTERISE THE HYBRIDISATION NUMBER FOR A SET OF
PHYLOGENIES∗
arXiv:1712.04131v1 [q-bio.PE] 12 Dec 2017
SIMONE LINZ† AND CHARLES SEMPLE‡
Abstract. Throughout the last decade, we have seen much progress towards characterising and
computing the minimum hybridisation number for a set P of rooted phylogenetic trees. Roughly
speaking, this minimum quantifies the number of hybridisation events needed to explain a set of
phylogenetic trees by simultaneously embedding them into a phylogenetic network. From a mathematical viewpoint, the notion of agreement forests is the underpinning concept for almost all results
that are related to calculating the minimum hybridisation number for when |P| = 2. However, despite various attempts, characterising this number in terms of agreement forests for |P| > 2 remains
elusive. In this paper, we characterise the minimum hybridisation number for when P is of arbitrary
size and consists of not necessarily binary trees. Building on our previous work on cherry-picking
sequences, we first establish a new characterisation to compute the minimum hybridisation number
in the space of tree-child networks. Subsequently, we show how this characterisation extends to the
space of all rooted phylogenetic networks. Moreover, we establish a particular hardness result that
gives new insight into some of the limitations of agreement forests.
Key words. agreement forest, cherry-picking sequence, minimum hybridisation, phylogenetic
networks, reticulation, tree-child networks
AMS subject classifications. 05C05; 92D15
1. Introduction. In our quest for faithfully describing evolutionary histories,
we are currently witnessing a shift from the representation of ancestral histories by
phylogenetic (evolutionary) trees towards phylogenetic networks. The latter not only
represent speciation events but also non-tree like events such as hybridisation and
horizontal gene transfer that have played an important role throughout the evolution
of certain groups of organisms as for example in plants and fish [10, 17, 18, 23].
In this paper, we focus on a problem that is related to the reconstruction of phylogenetic networks. Called Minimum Hybridisation and formally stated at the end
of this section, this problem was first introduced by Baroni et al. [2]. While Minimum
Hybridisation was historically motivated by attempting to quantify hybridisation
events, it is now more broadly regarded as a tool to quantify all non-tree like events to
which we collectively refer to as reticulation events. Pictorially speaking, Minimum
Hybridisation aims at the reconstruction of a phylogenetic network that simultaneously embeds a given set of phylogenetic trees while minimising the number of
reticulation events that are represented by vertices in the network whose in-degree is
at least two. More formally, the problem is based on the following underlying question. Given a collection P of rooted phylogenetic trees on the same set of taxa that
have correctly been reconstructed for different parts of the species’ genomes, what
is the smallest number of reticulation events that is needed to explain P? Over the
last ten years, we have seen significant progress in characterising and computing this
minimum number for when |P| = 2 (e.g. see [1, 3, 4, 8, 16, 25]). However, except
for some heuristic approaches [7, 26], less is known for when |P| ≥ 3. This is due to
∗ Submitted
to the editors December 13, 2017.
Funding: We thank the New Zealand Marsden Fund for their financial support.
† Department
of
Computer
Science,
University
of
Auckland,
New
([email protected]).
‡ School
of Mathematics and Statistics, University of Canterbury, New
([email protected]).
1
Zealand.
Zealand.
2
SIMONE LINZ AND CHARLES SEMPLE
the fact that the notion of agreement forests, which underlies almost all results that
are related to Minimum Hybridisation, appears to be ungeneralisable to more than
two trees.
Previously, together with Humphries, we introduced cherry-picking sequences and
characterised a restricted version of Minimum Hybridisation for P being binary and
of arbitrary size [12]. Instead of minimising the number of reticulation events needed
to explain P over the space of all rooted phylogenetic networks, this restricted version
only considers binary temporal tree-child networks. Such networks are the binary
intersection of the classes of temporal networks and tree-child networks introduced
by Moret et al. [19] and Cardona et al. [6], respectively. Disadvantageously, this
restriction is so strong that not even if |P| = 2 are we guaranteed to have a solution,
i.e. there may be no such network explaining P [11, Figure 2].
Here, we advance our work on cherry-picking sequences and establish two new
characterisations to quantify the amount of reticulation events that are needed to
explain a set of (not necessarily binary) phylogenetic trees. The first characterisation
solves the problem over the space of tree-child networks. Unlike temporal networks,
we show that every collection P of rooted phylogenetic trees has a solution, i.e. the
trees in P can simultaneously be embedded into a tree-child network. Subsequently,
we extend this characterisation to the space of all rooted phylogenetic networks and,
hence, provide the first characterisation for Minimum Hybridisation in its most
general form. Both characterisations are based on computing a cherry-picking sequence for P, while the latter characterisation makes also use of an operation that
attaches auxiliary leaves to the trees in P.
In addition to the two new characterisations, we return back to agreement forests
and investigate why they seem to be of limited use to solve Minimum Hybridisation
for an arbitrary size set P of rooted phylogenetic trees. Roughly speaking, given P,
one can compute a particular type of agreement forest F of smallest size and, if
|P| = 2, then each but one component in F contributes exactly one to the minimum
number of reticulation events that is needed to explain P. On the other hand, if
|P| > 2, the contribution of each component in F to this minimum number is much
less clear. Motivated by this drawback of agreement forests, we consider a set P of
rooted binary phylogenetic trees as well as the agreement forest F induced (formally
defined in Section 5) by a phylogenetic network that explains P and minimises the
number of reticulations events and ask whether or not, it is computationally hard
to calculate the minimum number of reticulation events that is needed to explain P.
We call the associated decision problem Scoring Optimum Forest. This problem
was first mentioned in [13], where the authors conjecture that Scoring Optimum
Forest is NP-complete. Using the machinery of cherry-picking sequences, we show
that Scoring Optimum Forest is NP-complete for when one considers the smaller
space of tree-child networks.
The paper is organised as follows. The remainder of the introduction contains
some definitions and preliminaries on phylogenetic networks. In Section 2, we state the
two new characterisations in terms of cherry-picking sequences. The first optimises
Minimum Hybridisation within the space of tree-child networks and the second
optimises Minimum Hybridisation within the space of all phylogenetic networks.
The second characterisation is an extension of the first by additionally allowing the
attachment of auxiliary leaves. We then establish proofs for both characterisations in
Section 3 as well as a formal description of the analogous algorithm. In Section 4, we
establish an upper bound on the number of auxiliary leaves that, given a collection
of phylogenetic trees, are needed to characterise Minimum Hybridisation over the
MINIMUM HYBRIDISATION
3
space of all rooted phylogenetic networks. Lastly, in Section 5, we formally state the
problem Scoring Optimum Forest and show that it is NP-complete. We finish the
paper with some concluding remarks in Section 6.
Throughout the paper, X denotes a non-empty finite set. A phylogenetic network
N on X is a rooted acyclic digraph with no parallel edges that satisfies the following
properties:
(i) the (unique) root has out-degree two,
(ii) the set X is the set of vertices of out-degree zero, each of which has in-degree
one, and
(iii) all other vertices either have in-degree one and out-degree two, or in-degree at
least two and out-degree one.
For technical reasons, if |X| = 1, we additionally allow N to consist of the single
vertex in X. The set X is the leaf set of N and the vertices in X are called leaves. We
sometimes denote the leaf set of N by L(N ). For two vertices u and v in N , we say
that u is a parent of v and v is a child of u if (u, v) is an edge in N . Furthermore, the
vertices of in-degree at most one and out-degree two are tree vertices, while the vertices
of in-degree at least two and out-degree one are reticulations. An edge directed into
a reticulation is called a reticulation edge while each non-reticulation edge is called
a tree edge. We say that N is binary if each reticulation has in-degree exactly two.
Lastly, a directed path P in N ending at a leaf is a tree path if every intermediate
vertex in P is a tree vertex.
A phylogenetic network N on X is tree child if each non-leaf vertex in N is the
parent of a tree vertex or a leaf. An example of two tree-child networks N and N 0 is
given at the bottom of Figure 1. Note that the phylogenetic network obtained from
N by deleting the leaf labelled 4 and suppressing the resulting degree-two vertex v
results in a network that is not tree child.
A rooted phylogenetic X-tree T is a rooted tree with no degree-2 vertices except
possibly the root which has degree at least two, and with leaf set X. If |X| = 1, then
T consists of the single vertex in X. As for phylogenetic networks, the set X is called
the leaf set of T and is denoted by L(T ). In addition, T is binary if |X| = 1 or, apart
from the root which has degree two, all interior vertices have degree three. Since we
are only interested in rooted phylogenetic trees and rooted binary phylogenetic trees
in this paper, we will refer to such trees simply as phylogenetic trees and binary
phylogenetic trees, respectively. For a phylogenetic X-tree T , we consider two types
of subtrees. Let X 0 be a subset of X. The minimal subtree of T that connects all the
leaves in X 0 is denoted by T (X 0 ). Moreover, the restriction of T to X 0 , denoted by
T |X 0 , is the phylogenetic X 0 -tree obtained from T (X 0 ) by suppressing all degree-two
vertices apart from the root. Lastly, for two phylogenetic X-trees T and T 0 , we say
that T 0 is a refinement of T if T can be obtained from T 0 by contracting a possibly
empty set of internal edges in T 0 . In addition, T 0 is a binary refinement of T if T 0 is
binary.
Let T be a phylogenetic X 0 -tree. A phylogenetic network N on X with X 0 ⊆ X
displays T if, up to suppressing vertices with in-degree 1 and out-degree 1, there
exists a binary refinement of T that can be obtained from N by deleting edges, leaves
not in X 0 , and any resulting vertices of out-degree zero, in which case we call the
resulting acyclic digraph an embedding of T in N . If P is a collection of phylogenetic
X-trees, then N displays P if each tree in P is displayed by N . For example, the two
phylogenetic networks at the bottom of Figure 1 both display each of the four trees
shown in the top part of the same figure.
Let N be a phylogenetic network with vertex set V and root ρ. The hybridisation
4
SIMONE LINZ AND CHARLES SEMPLE
P
3 2 1 4 5 6
1 2 3 4 5 6
3 2 1 4 5 6
3 2 1 4 5 6
v
1
2
3
4
5
6
1
2
3 4 5
6
N0
N
Fig. 1. Top: A set P of four phylogenetic X-trees with X = {1, 2, . . . , 6}. Bottom: Two
tree-child networks displaying P with h(N ) = 3 and h(N 0 ) = 4.
number of N , denoted h(N ), is the value
X
h(N ) =
d− (v) − 1 ,
v∈V −{ρ}
where d− (v) denotes the in-degree of v. For example, the phylogenetic networks N
and N 0 that are shown in Figure 1 have hybridisation number 3 and 4, respectively.
Observe that each tree vertex and each leaf contributes zero to this sum, but each
reticulation v contributes d− (v)−1. Furthermore, for a set P of phylogenetic X-trees,
we denote by htc (P) and h(P), respectively, the values
min{h(N ) : N is a tree-child network on X that displays P}
and
min{h(N ) : N is a phylogenetic network on X that displays P}.
Remark. While the above definition of a phylogenetic network is restricted to networks whose tree vertices have out-degree exactly two, we note that the results in this
paper also hold for networks with tree vertices whose out-degree is at least two. More
particularly, if a set P of phylogenetic X-trees is displayed by a phylogenetic network
N whose tree vertices have out-degree at least two, then, by “refining” such vertices,
we can obtain a phylogenetic network N 0 whose tree vertices have out-degree exactly
two, displays P, and h(N 0 ) = h(N ). Thus no generality is lost with this restriction.
We next formally state the two decision problems that this paper is centred
around.
Minimum Tree-Child Hybridisation
Instance. A set P of phylogenetic X-trees and a positive integer k.
Question. Does there exist a tree-child network N on X that displays P such that
h(N ) ≤ k?
Minimum Hybridisation
Instance. A set P of phylogenetic X-trees and a positive integer k.
MINIMUM HYBRIDISATION
5
Question. Does there exist a phylogenetic network N on X that displays P such
that h(N ) ≤ k?
We will see at the end of this section that, for any given set P of phylogenetic
X-trees, Minimum Tree-Child Hybridisation has a solution, i.e. there exists a
tree-child network that displays P.
It was shown in [3] that Minimum Hybridisation is NP-hard, even for when
P consists of two rooted binary phylogenetic X-trees. To see that Minimum TreeChild Hybridisation is also computationally hard, we again consider this restricted
version of the problem and recall the following observation that was first mentioned
in [12] and can be derived by slightly modifying the proof of [2, Theorem 2].
Observation 1.1. Let P = {T , T 0 } be a collection of two binary phylogenetic
X-trees. If there exists a phylogenetic network N that displays P with h(N ) = k,
then there also exists a tree-child network N 0 that displays P with h(N 0 ) ≤ k.
The next theorem, whose straightforward proof is omitted, follows from Observation 1.1 and the fact that, given a tree-child network N and a binary phylogenetic
tree T , it can be checked in polynomial time whether or not N displays T [14, 21].
Theorem 1.2. The decision problem Minimum-Tree-Child Hybridisation is
NP-complete.
We end this section by showing that every collection of phylogenetic trees can be
displayed by a tree-child network. For n = 2, let U2 be the unique binary phylogenetic
tree on two leaves, x1 and x2 say, whose root ρ is a vertex at the end of a pendant
edge adjoined to the original root. Now for a positive integer n > 2, obtain Un from
Un−1 by adding an edge that joins a new vertex v and a new leaf xn and, for each tree
edge e in Un−1 , subdividing e with a new vertex ue and adding the edge (ue , v). We
call Un the universal network on n leaves and note that Un is unique up to relabelling
its leaves.
Theorem 1.3. Let Un be the universal network on X = {x1 , x2 , . . . , xn } with
n ≥ 2. Then Un is tree child and displays all binary phylogenetic X-trees.
Proof. By construction of Un from Un−1 it is straightforward to check that, as U2
is tree child, Un is tree child. To see that Un displays all binary phylogenetic X-trees,
we use induction on n. Clearly, U2 displays the unique binary phylogenetic tree on two
leaves. For n ≥ 3, assume that the universal network Un−1 on X 0 = {x1 , x2 , . . . , xn−1 }
displays all binary phylogenetic X 0 -trees. Observe that Un−1 can be obtained from
Un by deleting xn , the parent of xn and all their incident edges, and suppressing
all resulting vertices with in-degree 1 and out-degree-1. Now, let Tn be a binary
phylogenetic X-tree, and let Tn−1 be Tn |X 0 . Furthermore, let C be the subset of X 0
that consists of the descendant leaves of the parent of xn in Tn . As Un−1 displays
Tn−1 , there exist an embedding E of Tn−1 in Un−1 and an edge (u, v) in E such that
the set of descendants of v in E is precisely C. If (u, v) is a tree edge in Un−1 , then it
is easily checked that Un displays Tn by construction. On the other hand, if (u, v) is a
reticulation edge in Un−1 , then v has out-degree 1 in E. Let (v, w) be the unique edge
in E that is directed out of v. Note that, as Un−1 is tree child, w is a tree vertex in
Un−1 . Then, as (v, w) is a tree edge in Un−1 that is subdivided by a new vertex in the
construction of Un from Un−1 , it again follows that Un displays Tn . This completes
the proof of the theorem.
The next corollary is an immediate consequence of Theorem 1.3 and the fact that
6
SIMONE LINZ AND CHARLES SEMPLE
every phylogenetic tree has a binary refinement on the same leaf set.
Corollary 1.4. Let P be a set of phylogenetic X-trees. There exists a tree-child
network on X that displays P.
While every collection of phylogenetic X-trees can be displayed by a tree-child
network on X, a simple counting argument shows that the analogous result is not
true for binary tree-child networks. Specifically, a binary tree-child network on X has
at most |X| − 1 reticulations [6, Proposition 1] and so displays at most 2|X|−1 distinct
binary phylogenetic X-trees. But for large enough X, there are many more distinct
binary phylogenetic X-trees than 2|X|−1 . For related results, we refer the interested
reader to [22].
2. Cherry-picking characterisations. In this section, we state the two cherrypicking characterisations whose proofs are given in the next section. Let T be a
phylogenetic X-tree with root ρ, where |X| ≥ 2. If x is a leaf of T , we denote by
T \x the operation of deleting x and its incident edge and, if the parent of x in T
has out-degree 2, suppressing the resulting degree-two vertex. Note that if the parent
of x is ρ and ρ has out-degree 2, then T \x denotes the operation of deleting x and
its incident edge, and then deleting ρ and its incident edge. Observe that T \x is a
phylogenetic tree on X − {x}. A 2-element subset {x, y} of X is a cherry of T if x
and y have the same parent. Clearly, every phylogenetic tree with at least two leaves
contains a cherry. In this paper, we typically distinguish the leaves in a cherry, in
which case we write {x, y} as the ordered pair (x, y) depending on the roles of x and
y.
Let P be a set of phylogenetic X-trees. A sequence
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −), (xs+2 , −), . . . , (xt , −)
of ordered pairs in X × (X ∪ {−}) is a cherry-picking sequence of P if the following
algorithm returns a set of phylogenetic trees each of which consists of a single vertex
in {xs+1 , xs+2 , . . . , xt }.
Algorithm. Picking Cherries
Input. A set P of phylogenetic X-trees and a cherry-picking sequence
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −), (xs+2 , −), . . . , (xt , −)
for P.
Output. A set Ps of phylogenetic trees.
Step 1. Set P0 = P and, for each tree T ∈ P, set T0 = T . Set i = 1.
Step 2. Set Pi to be the set of phylogenetic trees obtained from Pi−1 by performing
exactly one of the following two operations for each tree Ti−1 ∈ Pi−1 :
(a) If {xi , yi } is a cherry of Ti−1 , then set Ti = Ti−1 \xi .
(b) Else, set Ti = Ti−1 .
Step 3. If i < s, increment i by one and repeat Step 2; otherwise, return Ps .
For all i ∈ {1, 2, . . . , s}, we say that Pi is obtained from P by picking x1 , x2 , . . . , xi .
Furthermore, if Pi 6= Pi+1 for each i ∈ {1, 2, . . . , s − 1}, we say that each ordered pair
in σ is essential. The weight of σ, denoted w(σ), is the value t − |X|. Observe that if
σ is a cherry-picking sequence of P, then
t − |X| ≥ 0
MINIMUM HYBRIDISATION
7
as each element in X must appear as the first element in an ordered pair in σ.
A particular type of cherry-picking sequence underlies our characterisation of
htc (P). To this end, let P be a set of phylogenetic X-trees. A cherry-picking sequence
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −), (xs+2 , −), . . . , (xt , −)
for P is called a tree-child sequence if t = s + 1 and, for all i ∈ {1, 2, . . . , s}, we
have xi 6∈ {yi+1 , yi+2 , . . . , ys }. Now, let σ be a tree-child sequence for P. We call
σ a minimum-tree-child sequence of P if w(σ) is of smallest value over all tree-child
sequences of P. This smallest value is denoted by stc (P). It will follow from the
results in the next section (Lemma 3.4) that every collection P of phylogenetic trees
has a tree-child sequence and so stc (P) is well defined.
Referring to Figure 1,
σ = (3, 2), (3, 4), (5, 6), (5, 4), (1, 2), (4, 2), (4, 6), (2, 6), (6, −)
is a tree-child sequence with weight w(σ) = 9 − 6 = 3 for the four trees shown at the
top of this figure.
Remark. As noted in the introduction, cherry-picking sequences were introduced
in [12]. In the set-up of this paper, the difference is as follows. Instead of a cherrypicking sequence consisting of a set of ordered pairs, a cherry-picking sequence in [12]
consists of an ordering of the elements in X. Moreover, this ordering has the additional property that, in the step analogous to Step 2 of Picking Cherries, xi is part
of a cherry of every tree in Pi−1 . At this step, xi is deleted from each tree in Pi−1 ,
and the iterative process continues. The weighting of such a sequence is based, across
all i, on the number of different cherries of which xi is part of. It is not difficult to
see how this could be interpreted as a special type of tree-child sequence.
The first of our new characterisations is the next theorem. For a given set P of
phylogenetic X-trees, it writes htc (P) in terms of tree-child sequences for P.
Theorem 2.1. Let P be a set of phylogenetic X-trees. Then
htc (P) = stc (P).
To state the second characterisation, we require an additional concept. Let T be
a phylogenetic X-tree. Consider the operation of adjoining a new leaf z to T in one
of the following three ways.
(i) Subdivide an edge of T with a new vertex, u say, and add the edge (u, z)
(ii) View the root ρ of T as a degree-one vertex adjacent to the original root and
add the edge (ρ, z).
(iii) Add the edge (v, z), where v is an interior vertex of T .
We refer to this operation as attaching a new leaf z to T . More generally, if Z is a
finite set of elements such that X ∩ Z is empty, then attaching Z to T is the operation
of attaching, in turn, each element in Z to T to eventually obtain a phylogenetic tree
on X ∪ Z. We refer to Z as a set of auxiliary leaves. Lastly, attaching Z to a set P of
phylogenetic X-trees is the operation of attaching Z to each tree in P.
Let P be a set of phylogenetic X-trees. A tree-child sequence σ for P is leaf
added if it is a tree-child sequence of a set of phylogenetic trees obtained from P
8
SIMONE LINZ AND CHARLES SEMPLE
P
1 2 3 4 5 6 7 8
1 4 2 3 5 6 7 8
1 2 3 4 6 7 8 5
P0
1 2 3 4 z 5 6 7 8
1 4 2 3 z 5 6 7 8
1 2 3 z 4 6 7 8 5
Fig. 2. Two sets P and P 0 of phylogenetic trees, where P 0 is obtained by attaching z to P. (In
parts, adapted from [13, Figure 1].)
by attaching a set of auxiliary leaves. We denote the minimum weight amongst all
leaf-added tree-child sequences of P by s`-tc (P). Of course, s`-tc (P) ≤ stc (P), but
this inequality can also be strict. To illustrate, consider the two sets P and P 0 of
phylogenetic trees shown in Figure 2. Now
σ = (4, 5), (4, 1), (4, 3), (5, 6), (5, 3), (5, 8),
(2, 3), (3, 1), (6, 7), (7, 8), (1, 8), (8, −)
is a tree-child sequence for P of weight w(σ) = 12 − 8 = 4. In fact, it follows
from [13, 15] that htc (P) = 4 (see Section 6 for details). On the other hand,
σ 0 = (5, z), (5, 8), (4, z), (4, 1), (z, 3), (z, 6),
(2, 3), (3, 1), (6, 7), (7, 8), (1, 8), (8, −)
is a tree-child sequence for P 0 of weight w(σ 0 ) = 12 − 9 = 3. Since P 0 can be obtained
by attaching z to P, it follows that σ 0 is a leaf-added tree-child sequence for P and
s`-tc (P) ≤ 3.
For a given set P of phylogenetic X-trees, the next theorem characterises h(P)
in terms of leaf-added tree-child sequences.
Theorem 2.2. Let P be a set of phylogenetic X-trees. Then
h(P) = s`-tc (P).
It is worth noting that, for a set P of phylogenetic X-trees, it follows from Theorems 2.1 and 2.2 that htc (P) and h(P) can be determined without constructing a
phylogenetic network.
3. Proofs of Theorems 2.1 and 2.2. In this section, we prove Theorems 2.1
and 2.2. Most of the work is in proving Theorem 2.1. We begin by showing that
htc (P) ≤ stc (P).
Lemma 3.1. Let P be a set of phylogenetic X-trees. Let σ be a tree-child sequence
for P. Then there exists a tree-child network N on X that displays P with h(N ) ≤
w(σ) satisfying the following properties:
MINIMUM HYBRIDISATION
9
(i) If u is a tree vertex in N and not the parent of a reticulation, then there are
leaves `1 and `2 at the end of tree paths starting at the children v1 and v2 of u,
respectively, such that (`1 , `2 ) is an element in σ.
(ii) If u is a tree vertex in N and the parent of a reticulation v, then there are leaves
`u and `v at the end of tree paths starting at u and v, respectively, such that
(`v , `u ) ∈ σ.
Proof. Let
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −)
be a tree-child sequence for P. The proof is by induction on s. If s = 0, then
|X| = 1 and each tree in P consists of the single vertex in X. It immediately follows
that choosing N to be the phylogenetic network consisting of the single vertex in X
establishes the lemma for s = 0.
Now suppose that s ≥ 1, and that the lemma holds for all tree-child sequences
for sets of phylogenetic trees on the same leaf set whose length is at most s. Let
σ 0 = (x2 , y2 ), (x3 , y3 ), . . . , (xs , ys ), (xs+1 , −)
and let P 0 be the set of phylogenetic trees obtained from P by picking x1 .
First assume each tree in P 0 has the same leaf set, namely X 0 = X − {x1 }. Then
0
σ is a tree-child sequence for P 0 . By induction, there is a tree-child network N 0 on
X 0 that displays P 0 with h(N 0 ) ≤ w(σ 0 ) and satisfies (i) and (ii). Since each tree in P 0
has the same leaf set, {x1 , y1 } is a cherry in each tree in P. Therefore, as N 0 displays
a binary refinement of each tree in P 0 , the tree-child network obtained from N 0 by
subdividing the edge directed into y1 with a new vertex u and adding the edge (u, x1 )
displays P. Furthermore, as h(N 0 ) ≤ w(σ 0 ) and N 0 satisfies (i) and (ii) relative to σ 0 ,
we have h(N ) = h(N 0 ) ≤ w(σ 0 ) = w(σ) and it is easily seen that N satisfies (i) and
(ii) relative to σ.
Now assume that not every tree in P 0 has the same leaf set. Let P10 denote the
subset of trees in P 0 whose leaf set is X − {x1 }. Since P 0 − P10 is non-empty, there
exists an ordered pair in σ 0 whose first coordinate is x1 . Note that (x1 , −) is not in
σ; otherwise there is an ordered pair in σ whose second coordinate is x1 and so σ is
not a tree-child sequence for P. Let (x1 , yi ) be the first such ordered pair. Let T1 be
a tree in P10 and, using σ 0 , consider applying i − 2 iterations of Picking Cherries
to T1 . Let X1 denote the subset of leaves in X − {x1 } that are deleted from T1 in this
process. Observe that, as yi is the second coordinate in (x1 , yi ), we have yi 6∈ X1 .
We next add x1 to T1 to obtain a phylogenetic X-tree for which σ 0 is a tree-child
sequence. Let w be the (unique) vertex of T1 that is closest to the root with the
property that yi is a descendant leaf of w, and the child of w on the path from w to yi
has all its descendant leaves in X1 ∪ {yi }. Let T10 be the binary phylogenetic X-tree
obtained from T by adding the edge (w, x1 ). We now show that σ 0 is a tree-child
sequence for T10 . Suppose that σ 0 is not a tree-child sequence for T10 . Let u be the
parent of w in T10 . Then amongst the first i − 2 ordered pairs in σ 0 is an ordered pair
of the form (xj , yi ) that is essential when, using σ 0 , Picking Cherries is applied to
T1 , where xj is a descendant leaf of u in T10 . But then, each descendant leaf of w is in
X1 ∪ {yi }, contradicting the choice of w.
Repeating this placement of x1 for each tree in P10 , we obtain a set P100 of phylogenetic X-trees from P10 . Let P 00 = P100 ∪ (P 0 − P10 ) and observe that σ 0 is a tree-child
sequence for P 00 . Therefore, by induction, there is a tree-child network N 0 on X that
displays P 00 with h(N 0 ) ≤ w(σ 0 ) and satisfies (i) and (ii).
10
SIMONE LINZ AND CHARLES SEMPLE
Let p denote the parent of x1 in N 0 . If p is a reticulation, let N be the phylogenetic
network obtained from N 0 by subdividing the edge directed into y1 with a new vertex
u and adding the edge (u, p). Since N 0 is tree child and displays P 0 , it follows that
N is tree child and displays P. Furthermore,
h(N ) = h(N 0 ) + 1 ≤ w(σ 0 ) + 1 = w(σ).
Additionally, as (x1 , y1 ) ∈ σ, it also follows that, as N 0 satisfies (i) and (ii) relative
to σ 0 , we have N satisfies (i) and (ii) relative to σ.
Thus we may assume that p is a tree vertex. Let w denote the child of p that
is not x1 in N 0 . If w is a reticulation, then, as N 0 satisfies (ii), σ 0 contains a cherry
in which x1 is the second coordinate. But (x1 , y1 ) is the first ordered pair in σ and
so, as σ is tree child, x1 is never the second coordinate in an ordered pair in σ; a
contradiction. Therefore w is either a tree vertex or a leaf in N 0 . So, as N 0 satisfies
(i) and no ordered pair has x1 as the second coordinate, it follows that σ 0 contains an
ordered pair, (x1 , yj ) say, where yj is the leaf at the end of a tree path in N 0 starting
at w. Now let N be the phylogenetic network obtained from N 0 by subdividing the
edges directed into y1 and x1 with new vertices u and v, respectively, and adding the
edge (u, v). Since N 0 is tree child and h(N 0 ) ≤ w(σ 0 ), it is easily seen that N is tree
child and h(N ) = h(N 0 ) + 1 ≤ w(σ 0 ) + 1 = w(σ). Furthermore, N 0 displays P 0 − P10 as
well as P100 , and therefore P10 |(X − {x1 }). Thus N displays P. To see that N satisfies
(i) and (ii) relative to σ, it suffices to show that N satisfies (ii) for p and u. Indeed,
the two ordered pairs (x1 , yj ) and (x1 , y1 ) in σ verify (ii) for p and u, respectively.
This completes the proof of the lemma.
The next corollary immediately follows from Lemma 3.1.
Corollary 3.2. Let P be a set of phylogenetic X-trees. Then htc (P) ≤ stc (P).
For the proof of the converse of Corollary 3.2, we begin with an additional lemma.
Let N be a phylogenetic network, and let x and y be two leaves in N . Generalising
cherries to phylogenetic networks, we say that {x, y} is a cherry in N if x and y have
a common parent. Moreover, we call {x, y} a reticulated cherry if the parent of x, say
px , and the parent of y, say py , are joined by a reticulation edge (py , px ) in which case
we say that x is the reticulation leaf relative to {x, y}. We next define two operations
on N . First, reducing a cherry {x, y} is the operation of deleting one of the two leaves
in {x, y}, and suppressing the resulting degree-2 vertex. Second, reducing a reticulated
cherry {x, y} is the operation of deleting the reticulation edge joining the parents of
x and y and suppressing any resulting degree-2 vertices. The proof of the next lemma
is similar to the analogous result for binary tree-child networks [5, Lemma 4.1] and is
omitted.
Lemma 3.3. Let N be a tree-child network on X. Then the following hold.
(i) If |X| ≥ 2, then N contains either a cherry or a reticulated cherry.
(ii) If N 0 is obtained from N by reducing either a cherry or a reticulated cherry,
then N 0 is a tree-child network.
Lemma 3.4. Let P be a set of phylogenetic X-trees. Then htc (P) ≥ stc (P).
Proof. Let N be a tree-child network on X that displays P. By Corollary 1.4,
such a network exists. We establish the lemma by explicitly constructing a tree-child
sequence σ for P such that w(σ) ≤ h(N ).
Let ρ denote the root of N , and let v1 , v2 , . . . , vr denote the reticulations of N .
Let `ρ , `1 , `2 , . . . , `r denote the leaves at the end of tree paths Pρ , P1 , P2 , . . . , Pr in N
MINIMUM HYBRIDISATION
11
starting at ρ, v1 , v2 , . . . , vr , respectively. Observe that these paths are pairwise vertex
disjoint. We now construct a sequence of ordered pairs as follows:
Step 1. Set N = N0 and σ0 to be the empty sequence. Set i = 1.
Step 2. If Ni−1 consists of a single vertex xi , then set σi to be the concatenation of
σi−1 and (xi , −), and return σi .
Step 3. If {xi , yi } is a cherry in Ni−1 , then
(a) If one of xi and yi , say xi , equates to `j for some j ∈ {1, 2, . . . , r} and
vj is not a reticulation in Ni−1 , then set σi to be the concatenation of
σi−1 and (xi , yi ).
(b) Otherwise, set σi to be the concatenation of σi−1 and (xi , yi ), where
xi 6∈ {`ρ , `1 , `2 , . . . , `r }.
(c) Set Ni to be the tree-child network obtained from Ni−1 by deleting xi ,
thereby reducing the cherry {xi , yi }.
(d) Increase i by one and go to Step 2.
Step 4. Else, there is a reticulated cherry {xi , yi } in Ni−1 , where xi say is the reticulation leaf.
(a) Set σi to be the concatenation of σi−1 and (xi , yi ).
(b) Set Ni to be the tree-child network obtained from Ni−1 by reducing the
reticulated cherry {xi , yi }.
(c) Increase i by one and go to Step 2.
First note that it is easily checked that the construction is well defined, that is,
it returns a sequence of ordered pairs. Moreover, in each iteration i of the above
construction, it follows from Lemma 3.3 that Ni is tree child. We next show that, if
{xi , yi } is a cherry in Ni−1 , and xi and yi equate to `j and `j 0 , respectively, where `j
and `j 0 are elements in {`1 , `2 , . . . , `r }, then exactly one of vj and vj 0 is a reticulation
in Ni−1 . To see this, if vj and vj 0 are both reticulations in Ni−1 , then Pj and Pj 0
are not vertex disjoint in N ; a contradiction. On the other hand, suppose neither vj
and vj 0 are reticulations in Ni−1 . Without loss of generality, we may assume {xi , yi }
is the first such cherry for which this holds. Since N is tree child, and therefore has
no tree vertex that is the parent of two reticulations, there is an iteration i0 < i, in
which the cherry (xi0 , yi0 ) is concatenated with σi0 −1 , where yi0 ∈ {xi , yi }, and Ni0
has {xi , yi } as a cherry but Ni0 −1 does not. If xi0 = `ρ or xi0 ∈ {`1 , `2 , . . . , `r }, we
contradict the construction by the choice of {xi , yi }. Also, if xi0 6∈ {`ρ , `1 , `2 , . . . , `r },
then we again contradict the construction. Hence, we may assume for the remainder
of the proof that exactly one of vj and vj 0 is a reticulation in Ni−1 .
Let
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xi−1 , yi−1 ), (xi , −)
be the sequence returned by the construction. We prove by induction on i that σ is
a tree-child sequence for P whose weight is at most h(N ). If i = 1, then N consists
of the single vertex in X and the construction correctly returns such a sequence.
Now suppose that i ≥ 2, and consider the first iteration of the construction. Either
{x1 , y1 } is a cherry or a reticulated cherry of N0 . If {x1 , y1 } is a cherry, then {x1 , y1 }
is a cherry of each tree in P. In this instance, let P 0 denote the set of phylogenetic
X 0 -trees obtained from P by picking x1 , where X 0 = X − {x1 }. Observe that N1 is
a tree-child network on X 0 that displays P 0 .
Now assume that {x1 , y1 } is a reticulated cherry with x1 as the reticulation leaf.
Let P1 be the subset of trees in P not displayed by N1 and let P2 = P −P1 . Note that
12
SIMONE LINZ AND CHARLES SEMPLE
{x1 , y1 } is a cherry of each tree in P1 . For each tree in P1 , delete the edge incident
with x1 , suppress any resulting degree-2 vertex, and reattach x1 to the rest of the tree
containing y1 by subdividing an edge with a new vertex and adding an edge joining
this vertex and x1 so that the resulting phylogenetic X-tree is displayed by N1 . It is
easily seen that this is always possible. Let P10 denote the resulting collection of trees
obtained from P1 . For this instance, let P 0 = P10 ∪ P2 and observe that N1 displays
P 0 . To complete the induction it suffices to show that if
σ 0 = (x2 , y2 ), (x3 , y3 ), . . . , (xi−1 , yi−1 ), (xi , −)
is a tree-child sequence for P 0 whose weight w(σ 0 ) is at most h(N1 ), then σ is a
tree-child sequence for P whose weight w(σ) is at most h(N0 ), that is, at most h(N ).
First assume that {x1 , y1 } is a cherry of N0 . Then, as {x1 , y1 } is a cherry of
each tree in P, it follows by induction that σ is a cherry-picking sequence of P.
Furthermore, as σ 0 is tree child and x1 ∈
/ L(N1 ), σ is also tree child. Since x1 only
appears once as the first coordinate of an ordered pair in σ, we have
w(σ) = w(σ 0 ) ≤ h(N1 ) = h(N0 ).
Now assume that {x1 , y1 } is a reticulated cherry of N0 with x1 as the reticulation
leaf. Without loss of generality, let v1 denote the associated reticulation, so that
x1 = `1 . Since each tree in P1 has {x1 , y1 } as a cherry and σ 0 is a tree-child sequence
for P 0 , it follows that σ is a cherry-picking sequence for P.
We next show that σ is tree child. If the in-degree of v1 is at least three in
N0 , then v1 exists in N1 and so, by construction, x1 does not appear as the second
coordinate of an ordered pair in σ 0 as well as in σ. Therefore, if the in-degree of v1 is
at least three in N0 , then σ is a tree child.
Now suppose that the in-degree of v1 is two in N0 . To establish that σ is tree
child, assume to the contrary that x1 appears as the second coordinate of an ordered
pair in σ 0 . Let (z, x1 ) denote the first such ordered pair. Then, at some iteration
j, either {z, x1 } is a cherry or a reticulated cherry of Nj−1 . If {z, x1 } is a cherry
of Nj−1 , then, since x1 = `1 , we are in Step 3(a) in iteration j of the construction
and so the ordered pair should be (x1 , z); a contradiction. On the other hand, if
{z, x1 } is a reticulated cherry of Nj−1 , then z is the reticulation leaf of {z, x1 } and,
by construction of σ, one of the parents of v1 in N0 is the parent of two reticulations
in N0 , namely v1 and the reticulation for which, by construction, there is a tree path
starting at this reticulation and ending at z; a contradiction as N0 is tree child. Hence
σ is tree child. Furthermore, as w(σ) = w(σ 0 ) + 1 and h(N0 ) = h(N1 ) + 1,
w(σ) = w(σ 0 ) + 1 ≤ h(N1 ) + 1 = h(N0 ).
This completes the proof of the lemma.
Proof of Theorem 2.1. Combining Corollary 3.2 and Lemma 3.4 establishes the
theorem.
We next establish Theorem 2.2.
Proof of Theorem 2.2. We first show that h(P) ≤ s`-tc (P). Let P 0 be a set of
phylogenetic trees obtained from P by attaching a set Z such that X ∩ Z is empty
and s`-tc (P) = stc (P 0 ). It follows by Theorem 2.1 that there is a tree-child network
N 0 on X ∪ Z that displays P 0 with h(N 0 ) = stc (P 0 ). Observe that N 0 displays P. Let
N be the phylogenetic network on X obtained from N 0 by deleting every vertex that
MINIMUM HYBRIDISATION
13
is not on a directed path from the root to a leaf in X, and suppressing any resulting
non-root vertex of degree two. Noting that no deleted vertex is used to display a
phylogenetic tree in P, it is easily checked that, up to the root having out-degree one,
N displays P. Furthermore, h(N ) ≤ h(N 0 ). Therefore, by Theorem 2.1,
h(P) ≤ h(N ) ≤ h(N 0 ) = stc (P 0 ) = s`-tc (P).
In particular, h(P) ≤ s`-tc (P).
To prove the converse, h(P) ≥ s`-tc (P), let N be a phylogenetic network on X
that displays P and h(N ) = h(P). Let N 0 be the phylogenetic network obtained by
attaching a new leaf to each reticulation edge in N , i.e. for each reticulation edge e,
subdivide e with a new vertex u and add a new edge (u, ze ), where ze ∈
/ X. It is easily
checked that N 0 is tree child and h(N 0 ) = h(N ). Let Z denote the set of new leaves
attached to N . For each tree T in P, let Tr denote a binary refinement of T that is
displayed by N , and let Tr0 be a binary phylogenetic tree with leaf set X ∪ Z that
is displayed by N 0 and obtained from Tr by attaching Z. Note that Tr0 is a binary
refinement of a tree that can be obtained from T by attaching Z. Set
Pr0 = {Tr0 : T ∈ P},
and note that htc (Pr0 ) ≤ h(N 0 ) as N 0 is tree child and displays Pr0 . Since each tree in
Pr0 is a binary refinement of a tree that can be obtained from a tree in P by attaching
Z, we have s`-tc (P) ≤ stc (Pr0 ). Thus, by Theorem 2.1,
s`-tc (P) ≤ stc (Pr0 ) = htc (Pr0 ) ≤ h(N 0 ) = h(N ) = h(P),
and so s`-tc (P) ≤ h(P).
We end this section with the pseudocode of an algorithm—called Construct
Tree-Child Network—that constructs a tree-child network from a tree-child sequence. Specifically, given a tree-child sequence σ for a set P of phylogenetic X-trees,
Construct Tree-Child Network returns a tree-child network N on X that displays P and h(N ) ≤ w(σ). This is the same construction as that used to prove
Lemma 3.1 and so the proof of its correctness is not given.
Algorithm. Construct Tree-Child Network
Input. A set P of phylogenetic X-trees, and a tree-child sequence
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −)
for P.
Output. A tree-child network N on X that displays P and h(N ) ≤ w(σ).
Step 1. If |X| = 1, set Ns+1 to be the phylogenetic network consisting of the single
vertex xs+1 in X and return Ns+1 . Otherwise, set Ns+1 to be the phylogenetic network consisting of the single edge (ρ, xs+1 ) and set i = s.
Step 2. Depending on which holds, do exactly one of the following three steps.
(a) If xi ∈ L(Ni+1 ) and the parent pi of xi is a reticulation in Ni+1 , then
obtain Ni from Ni+1 by subdividing the edge directed into yi with a
new vertex u and adding a new edge (u, pi ).
(b) If xi ∈ L(Ni+1 ) and the parent pi of xi is not a reticulation in Ni+1 ,
then obtain Ni from Ni+1 by subdividing the edge directed into yi with
a new vertex u, subdividing the edge (pi , xi ) with a new vertex v, and
adding a new edge (u, v).
14
SIMONE LINZ AND CHARLES SEMPLE
(c) Else xi ∈
/ L(Ni+1 ), and obtain Ni from Ni+1 by subdividing the edge
directed into yi with a new vertex u and adding a new edge (u, xi ).
Step 3. If i = 1, then set N to be the network obtained from Ni by deleting the
unique edge incident with ρ and return N . Otherwise, decrement i by one
and go to Step 2.
Now, let σ be a leaf-added tree-child sequence for a set P of phylogenetic X-trees.
Then there exists a set P 0 of phylogenetic trees on X ∪Z obtained from P by attaching
Z such that σ is a tree-child sequence for P 0 . It is straightforward to check that the
network N on X resulting from calling Construct Tree-Child Network for P 0
and σ and, subsequently, restricting to vertices and edges on a path from the root to
leaves in X as described in the first direction of the proof of Theorem 2.2 displays P
and h(N ) ≤ w(σ).
4. Bounding the maximum number of auxiliary leaves. In light of Theorem 2.2, a natural question to ask is how many auxiliary leaves need to be attached
to a given set P of phylogenetic X-trees in order to calculate h(P). Attaching auxiliary leaves to P is necessary whenever h(P) < htc (P). Here, we provide an upper
bound on the number of auxiliary leaves in terms of htc (P). We start by introducing
two operations that, repeatedly applied, transform any phylogenetic network N that
displays P into a tree-child network without increasing h(N ) and that displays a set
of binary phylogenetic trees obtained from P by attaching auxiliary leaves.
Let N be a phylogenetic network on X, and let (u, v) be an edge in N such that
u and v are reticulations. Obtain a phylogenetic network N 0 from N by contracting
(u, v) and, for each resulting pair of parallel edges, repeatedly deleting one of the two
edges in parallel and suppressing the resulting degree-two vertex. We say that N 0 has
been obtained from N by a contraction.
Lemma 4.1. Let P be a collection of phylogenetic X-trees, and let N be a phylogenetic network on X that displays P. Let N 0 be a phylogenetic network obtained
from N by a contraction. Then N 0 displays P and h(N 0 ) ≤ h(N ).
Proof. Let (u, v) be the edge in N that is incident with two reticulations and
contracted in the process of obtaining N 0 from N . Furthermore, let w be the vertex in
N 0 that results from identifying u and v. We have d− (w) ≤ d− (u)+d− (v)−1 while all
other reticulations w0 in N 0 correspond to a reticulation u0 in N and d− (w0 ) = d− (u0 ).
It now follows that h(N 0 ) ≤ h(N ).
Now, let T be a tree in P. If v is not used to display T in N , then u is also not
used to display T in N , and it is easily seen that N 0 displays T without using w. On
the other hand, if v is used to display T in N , then exactly one parent, t say, of v
is used to display T in N . If t 6= u, it is clear that N 0 displays P. Furthermore, if
t = u, then exactly one parent of u, say s, is used to display T in N . Now, regardless
of whether or not s is also a parent of v in which case s is suppressed in obtaining N 0
from N , it again follows that N 0 displays T . Hence N 0 displays each tree in P and
the lemma follows.
We call a phylogenetic network with no edges whose end vertices are both reticulations stack free. It follows from repeated applications of Lemma 4.1 that if P is
a collection of phylogenetic X-trees, then there is a stack-free network N on X that
displays P such that h(N ) = h(P).
For the second operation, let N be a phylogenetic network on X, and let u be a
tree vertex in N whose two children are both reticulations. Furthermore, let (u, v) be
15
MINIMUM HYBRIDISATION
u
u
v
v
0
v
w
w
z
(a)
(b)
Fig. 3. The phylogenetic networks at the bottom are obtained from their respective networks at
the top by (a) a contraction and (b) a leaf-attaching operation.
a reticulation edge, and let z ∈
/ X. Obtain a phylogenetic network N 0 on X ∪{z} from
N by subdividing the edge (u, v) with a new vertex w and adding a new edge (w, z).
We say that N 0 has been obtained from N by a leaf-attaching operation. Figure 3
illustrates (a) a contraction and (b) a leaf-attaching operation.
We are now in a position to establish the main result of this section.
Theorem 4.2. Let P be a collection of phylogenetic X-trees. There exists a set
Z of auxiliary leaves with the following two properties.
(i) |Z| ≤ htc (P), and
(ii) there is a collection PZ of binary phylogenetic trees with leaf set X∪Z obtained
from P by attaching Z such that h(P) = htc (PZ ).
Proof. Let N be a stack-free network on X that displays P with h(N ) = h(P).
By Lemma 4.1, N exists. Now obtain a phylogenetic network NZ from N by a
minimum number of repeated applications of the leaf-attaching operation until each
tree vertex in the resulting network has at least one child that is a tree vertex or a
leaf. Clearly, h(N ) = h(NZ ). Moreover, since no leaf-attaching operation results in
a new edge in NZ that is incident with two reticulations, NZ is stack free. It now
follows that NZ is tree child. Let Z = L(NZ ) − X. Then, by construction, the size of
Z is equal to the number of tree vertices in N whose children are both reticulations.
Let PZ be a set of phylogenetic trees obtained from P by attaching Z to P such that
NZ displays PZ . Since N displays P, such a set PZ always exists. By construction,
h(P) ≥ htc (PZ ). Moreover, as each tree in P is a restriction of a tree in PZ , it follows
that h(P) ≤ htc (PZ ); thereby establishing part (ii) of the theorem.
Using the construction of the previous paragraph, we now establish part (i) of
the theorem. Let Er be the set of reticulation edges in N , and let Vt be the set of
tree vertices of N whose children are both reticulations. Recall that |Vt | = |Z|. We
next make two observations. First, each vertex in Vt is incident with two edges in
Er . Second, each edge in Er is incident with at most one vertex in Vt . In summary,
this implies that |Z| ≤ 21 |Er |. Furthermore, we have |Er | = h(P) + |Vr |, where Vr is
the set of reticulations in N . Therefore, as |Vr | ≤ h(P), we have |Er | ≤ 2h(P). As
h(P) ≤ htc (P), it now follows that
|Z| ≤ 12 |Er | ≤
1
2
· 2h(P) ≤
This establishes part (i) of the theorem.
1
2
· 2htc (P) = htc (P).
16
SIMONE LINZ AND CHARLES SEMPLE
5. Scoring an optimum forest. For a collection P of binary phylogenetic
X-trees, acyclic-agreement forests characterise h(P) for when P consists of exactly
two trees. Indeed, many algorithms and theoretical results that deal with Minimum
Hybridisation for two trees are deeply-anchored in the notion of acyclic-agreement
forests [1, 2, 4, 16]. In this section, we establish a particular hardness result that
contributes to an explanation of why acyclic-agreement forests appear, however, to
be of little use to solve Minimum Hybridisation for more than two trees. This result
is a particular instance of a conjecture in [13].
For the purpose of the upcoming definitions, we regard the root of a binary
phylogenetic X-tree T as a vertex labelled ρ at the end of a pendant edge adjoined
to the original root. Furthermore, we view ρ as an element of the leaf set of T ; thus
L(T ) = X ∪ {ρ}. Let T and T 0 be two binary phylogenetic X-trees. An agreement
forest F = {Lρ , L1 , L2 , . . . , Lk } for T and T 0 is a partition of X ∪{ρ} such that ρ ∈ Lρ
and the following conditions are satisfied:
(i) For all i ∈ {ρ, 1, 2, . . . , k}, we have T |Li ∼
= T 0 |Li .
(ii) The trees in
{T (Li ) : i ∈ {ρ, 1, 2, . . . , k}}
and
{T 0 (Li ) : i ∈ {ρ, 1, 2, . . . , k}}
are vertex-disjoint subtrees of T and T 0 , respectively.
Now, let F = {Lρ , L1 , L2 , . . . , Lk } be an agreement forest for T and T 0 . Let GF be
the directed graph that has vertex set F and an arc from Li to Lj precisely if i 6= j
and
(iii) the root of T (Li ) is an ancestor of the root of T (Lj ) in T ,or the root of
T 0 (Li ) is an ancestor of the root of T 0 (Lj ) in T 0 .
We call F an acyclic-agreement forest for T and T 0 if GF has no directed cycle.
Moreover, if F contains the smallest number of elements over all acyclic-agreement
forests for T and T 0 , we say that F is a maximum acyclic-agreement forest for T and
T 0 , in which case, we denote this number minus one by ma (T , T 0 ).
Baroni et al. [2] established the following characterisation for when a collection
of binary phylogenetic X-trees contains exactly two trees.
Theorem 5.1. Let P = {T , T 0 } be a collection of two binary phylogenetic Xtrees. Then h(P) = ma (T , T 0 ).
Let N be a phylogenetic network on X with root ρ that displays a set P of binary
phylogenetic X-trees. As above, we regard ρ as a vertex at the end of a pendant edge
adjoined to the original root. We obtain a forest F from N by deleting all reticulation
edges, repeatedly contracting edges where one end-vertex has degree one and is not in
X ∪ {ρ}, deleting isolated vertices not in X ∪ {ρ} and, lastly, suppressing all vertices
with in-degree one and out-degree one. We say that F is the forest induced by N .
Moreover, F is said to be optimum if N is a tree-child network with htc (P) = h(N ).
For example, up to regarding ρ as a new vertex that is adjoined to the original root
of the two phylogenetic networks N and N 0 shown in Figure 1,
F = {{ρ, 1, 2, 6}, {3}, {4}, {5}} and F 0 = {{ρ, 1, 3, 5, 6}, {2}, {4}}
is the induced forest of N and N 0 , respectively.
In [13], the authors investigate Minimum Hybridisation for three trees and
conjecture that, given a set P of three binary phylogenetic X-trees and the induced
forest of a phylogenetic network N that displays P and h(P) = h(N ), it is NP-hard
MINIMUM HYBRIDISATION
17
to determine N . For |P| ≥ 3, we affirmatively answer their conjecture in the context
of tree-child networks. More precisely, using cherry-picking sequences, we show that
the following decision problem is NP-complete.
Scoring Optimum Forest
Instance. A non-negative integer k, a collection P of binary phylogenetic X-trees,
an optimum forest F induced by a tree-child network N on X that displays P.
Question. Is htc (P) ≤ k?
If |P| = 2, then, by Observation 1.1, htc (P) = h(P) and F is a maximum acyclicagreement forest with htc (P) = |F| − 1. Hence, Scoring Optimum Forest is
polynomial time when |P| = 2. However, the general problem is NP-complete.
Theorem 5.2. The problem Scoring Optimum Forest is NP-complete.
The remainder of this section consists of the proof of Theorem 5.2. To establish
the result, we use a reduction from a particular instance of the NP-complete problem
Shortest Common Supersequence. Let Σ be a finite alphabet, and let W be a
finite subset of words in Σ∗ . A word z ∈ Σ∗ is a common supersequence of W if each
word in W is a subsequence of z.
Shortest Common Supersequence (SCS)
Instance. A non-negative integer k, a finite alphabet Σ, and a finite subset W of
words in Σ∗ .
Question. Is there a supersequence of the words in W with at most k letters?
Timkovskii [24, Theorem 2] established the next theorem. The orbit of a letter
in Σ is the set of its occurrences in the words in W . Note that if a word in W uses a
letter, b say, twice, then that word contributes two occurrences to the orbit of b.
Theorem 5.3. The decision problem SCS is NP-complete even if each word in
W has 3 letters and the size of all orbits is 2.
A consequence of Theorem 5.3 is the next corollary.
Corollary 5.4. The decision problem SCS is NP-complete even if each word in
W has 3 letters, the size of all orbits is at most 2, and no word in W contains a letter
twice.
Proof. Let k, Σ, and W be an instance of SCS, where each word W has 3 letters
and all orbits have size 2. Let Y be the subset of W that consists of those words in
W in which no letter occurs twice. Observe that if b ∈ Σ and b occurs twice in a word
in W , then no word in Y contains b. Furthermore, with regards to Y , each word has
3 letters, the size of all orbits is at most 2, and no word contains a letter twice. Let
t denote the number of distinct letters that occur in two distinct words in W − Y .
Note that the construction of Y and the computation of t can both be done in time
polynomial in |W |. The corollary will follow from Theorem 5.3 by showing that SCS
with parameters Σ and W has a supersequence of length at most
2|W − Y | + t + k
if and only if SCS with parameters Σ and Y has a supersequence of length at most
k.
Suppose that SCS with parameters Σ and Y has a supersequence z of length at
most k. Now iteratively extend z to a sequence z 0 as follows. Let w ∈ W − Y . Then
18
SIMONE LINZ AND CHARLES SEMPLE
w contains two occurrences of a letter, b say, in Σ. Let d denote the third letter in w.
Note that b occurs in no other word in Σ and d occurs in exactly one other word in
Σ. First assume that d occurs in a word in Y . Depending on whether d is the first,
second, or third letter in w, extend z by adding bb to the end of z, adding b at the
beginning and b at the end of z, or adding bb at the beginning of z, respectively. The
resulting sequence is a supersequence for Y ∪ {w}. Second assume that d does not
occur as a word in Y . Then d occurs in a word w0 in W − Y . Let c denote the letter
occurring twice in w0 . Extend z by adding w to the beginning of z and then, to the
resulting sequence, add two occurrences of c after w, add one occurrence of c before
w and one occurrence of c after w, or two occurrences of c before w depending on
whether d is the first, second, or third letter of w0 , respectively. The resulting sequence
is a supersequence for Y ∪ {w, w0 }. Taking the resulting sequence and repeating this
process for each remaining word in W − (Y ∪ {w}) or W − (Y ∪ {w, w0 }), respectively,
we eventually obtain a supersequence z 0 for W . Moreover, z 0 has length
2|W − Y | + t + k.
For the converse, suppose that there is a supersequence z of W of length
2|W − Y | + t + k.
Let z 0 be the sequence obtained from z by deleting each occurrence of a letter that
occurs twice in a word in W and deleting exactly one occurrence of a letter that
occurs in two distinct words in W − Y and, hence, does not occur in a word in Y . It is
easily checked that z 0 is a supersequence of Y . Furthermore, since there are 2|W − Y |
deletions of the first type and t deletions of the second type, it follows that z 0 has
length k. This completes the proof of the corollary.
The decision problem described in the statement of Corollary 5.4 is the one we
will use for the reduction in proving Theorem 5.2. Let k, Σ, and W be an instance of
SCS such that each word in W has 3 letters, the size of all orbits is at most 2, and no
word in W contains a letter twice. Without loss of generality, we may assume that,
for each ` ∈ Σ, there is a word in W containing `, and that |W | ≥ 3, so no letter is
contained in each word. Let
W = {w1 , w2 , . . . , wq }
and, for each i ∈ {1, 2, . . . , q}, let wi = wi1 wi2 wi3 . Also, let
o(Σ) = `1 , `2 , . . . , `|Σ|
denote a fixed ordering of the letters in Σ. For each wi in W , we denote the sequence
obtained from o(Σ) by removing each of the three letters in wi by o(Σ) − wi .
We now construct an instance of Scoring Optimum Forest. A rooted caterpillar is a binary phylogenetic tree T whose leaf set can be ordered, say x1 , x2 , . . . , xn , so
that {x1 , x2 } is a cherry and if pi denotes the parent of xi , then, for all i ∈ {3, 4, . . . , n},
we have (pi , pi−1 ) as an edge in T . Here, we denote the rooted caterpillar by
(x1 , x2 , . . . , xn ).
Now, for each i ∈ {1, 2, . . . , q}, let Ti denote the rooted caterpillar
Ti = (α, wi1 , wi2 , wi3 , β1 , β2 , . . . , βq|Σ| , o(Σ) − wi ),
and let P = {T1 , T2 , . . . , Tq }. Note that each of the trees in P has leaf set
X = {α, β1 , β2 , . . . , βq|Σ| } ∪ Σ
MINIMUM HYBRIDISATION
19
and P can be constructed in time polynomial in the size of Σ and W .
We next establish a lemma that reveals a relationship between the weight of a
tree-child sequence for P and the length of a supersequence for W . Let
σ = (x1 , y1 ), (x2 , y2 ), . . . , (xs , ys ), (xs+1 , −)
be a tree-child sequence for a set P of binary phylogenetic X-trees. For each i ∈
{1, 2, . . . , s}, we say that (xi , yi ) corresponds to n trees in P if {xi , yi } is a cherry in
exactly n trees obtained from P by picking x1 , x2 , . . . , xi−1 , where 1 ≤ n ≤ |P|.
Lemma 5.5. Let k ≤ 2|Σ| be a positive integer. Then there is a tree-child sequence
of P of weight k if and only if there is a supersequence of W of length k.
Proof. First suppose there is a common supersequence z of W of length k. Let
z = m1 m2 · · · mk ,
and let σ denote the sequence
(m1 , α), . . . , (mk , α), (β1 , α), . . . , (βq|Σ| , α), (`1 , α), . . . , (`|Σ| , α), (α, −).
Since z is a supersequence of W , it is easily seen that σ is a tree-child sequence of P.
Moreover,
w(σ) = (k + |X|) − |X| = k.
Now suppose that there is a tree-child sequence σ of P of weight k. Without loss
of generality, we may assume that each ordered pair in σ is essential. We first show
that there is a positive integer i0 such that, after i0 iterations of Picking Cherries
applied to P and σ, each of the trees in Pi0 has a cherry consisting of two elements in
{α, β1 , β2 , . . . , βq|Σ| }. If not, then there is a word wi in W such that either (wij , βq|Σ| )
or (βq|Σ| , wij ) is an ordered pair in σ, where j ∈ {1, 2, 3}. Now, by considering a word
not containing wij and its associated tree in P, it is easily seen that w(σ) ≥ q|Σ|−1 as
each of the elements in {β1 , β2 , . . . , βq|Σ|−1 } appear at least twice as the first element
of an ordered pair in σ. But then
q|Σ| − 1 > 2|Σ|
as q ≥ 3 and |Σ| ≥ 3, contradicting the assumption k ≤ 2|Σ|.
Consider the first i0 ordered pairs in σ. For each tree Ti in P, there are exactly
three ordered pairs whose first and second elements are in {α, wi1 , wi2 , wi3 } and for
which Picking Cherries picks three of these elements in Ti . Since the size of all
orbits is at most 2, such an ordered pair corresponds to at most two trees in P. We
next construct a sequence σ 0 of ordered pairs obtained from σ. We start by modifying
the first i0 ordered pairs of σ as follows:
(a) Amongst the first i0 ordered pairs, replace each ordered pair of the form (α, `)
with (`, α), where ` ∈ Σ.
(b) With the sequence obtained after (a) is completed, sequentially move along the
sequence to the i0 -th ordered pair replacing each ordered pair of the form (`, `0 ),
where `, `0 ∈ Σ in one of the following ways:
(i) If (`, `0 ) corresponds to exactly one tree in P, then replace it with (`, α)
or (`0 , α) depending on whether (`0 , α) or (`, α), respectively, is an earlier
ordered pair.
20
SIMONE LINZ AND CHARLES SEMPLE
(ii) If (`, `0 ) corresponds to two trees, Ti and Tj say, in P and the order of the
letters ` and `0 is the same in wi and wj , then replace it with (`, α) or (`0 , α)
depending on whether (`0 , α) or (`, α), respectively, is an earlier ordered pair.
(iii) If (`, `0 ) corresponds to two trees, Ti and Tj say, in P and the order of the
letters ` and `0 in wi is not the same as that in wj , then replace it with (`, α)
if (`, α) occurs as an ordered pair before (`0 , α) earlier in the sequence; otherwise, (`0 , α) occurs as an ordered pair before (`, α) earlier in the sequence
and so replace it with (`0 , α).
With this modification of σ after (b) is completed, let σ10 denote the subsequence of
the first i0 ordered pairs whose coordinates are in Σ ∪ {α}, and let σ20 denote the
subsequence σ − σ10 . Let σ 0 denote the concatenation of σ10 and σ20 .
By considering each tree Ti in P together with its corresponding ordered pairs in
σ and the associated ones in σ 0 , a routine check shows that the first |σ10 | iterations of
Picking Cherries applied to σ 0 and P sets the tree corresponding to Ti in P|σ10 | to
be the rooted caterpillar
(α, β1 , β2 , . . . , βq|Σ| , o(Σ) − wi ).
In particular, (wi1 , α), (wi2 , α), (wi3 , α) is subsequence of σ10 .
We next extend σ10 to a tree-child cherry-picking for P of weight at most k.
Consider σ and σ 0 . If σ1 denotes the subsequence of ordered pairs in σ corresponding
to σ10 , then |σ1 | = |σ10 | and
|σ| − |σ1 | ≥ q|Σ| + |Σ| + 1
as each of the elements in {β1 , β2 , . . . , βq|Σ| } ∪ Σ as well as at least one element in
{β1 , β2 , . . . , βq|Σ| }∪Σ∪{α} appears as the first coordinate of an ordered pair in σ −σ1 .
Here, an element in {β1 , β2 , . . . , βq|Σ| } ∪ Σ may be counted twice as it appears as the
first coordinate of two ordered pairs in σ − σ1 . It follows that the sequence of ordered
pairs that is the concatenation of σ10 and
(β1 , α), (β2 , α), . . . , (βq|Σ| , α), (`1 , α), (`2 , α), . . . , (`|Σ| , α), (α, −)
is a tree-child sequence of P whose weight is at most k.
Let σ10 be the sequence
(m1 , α), (m2 , α), . . . , (mk0 , α).
Since (wi1 , α), (wi2 , α), (wi3 , α) is a subsequence of σ10 for each tree Ti ,
m1 m2 · · · mk0
is a common supersequence of W . Moreover, as w(σ) = k, we have k 0 ≤ k. It follows
that there is a supersequence of W of length k.
To complete the proof of Theorem 5.2, let
F = {ρ, α, β1 , β2 , . . . , βq|Σ| } ∪ {`} : ` ∈ Σ
be a partition of X ∪ {ρ}. We next show that F is an optimum forest induced by a
tree-child network on X with root ρ and that displays P. Let z = z1 z2 · · · zk be a
common supersequence of W of minimum length, and suppose this length is k. Since
MINIMUM HYBRIDISATION
21
all orbits have size at most 2 and z is of minimum length, each letter in Σ appears at
most twice in z, and so k ≤ 2|Σ|. Let T be the ‘multi-labelled’ rooted caterpillar
T = (α, z1 , z2 , . . . , zk , β1 , β2 , . . . , βq|Σ| , o(Σ))
and let N be the tree-child network with root ρ obtained from T as follows. For each
` ∈ Σ, identify the leaves labelled ` and adjoin a new pendant edge to the identified
vertex with the leaf-end labelled `. Since z is a common supersequence of W , it
is easily checked that N displays P. Furthermore, h(N ) = k. By Theorem 2.1 and
Lemma 5.5, htc (P) = k, and so, as F is induced by N , it follows that F is an optimum
forest for P.
Now, given an arbitrary phylogenetic network, it can be verified in polynomial
time whether it is tree child, it displays P [21], its hybridisation number is at most k,
and it induces F. Hence, Scoring Optimum Forest is in NP.
Theorem 5.2 now follows by combining Corollary 5.4 with Theorem 2.1 and
Lemma 5.5.
6. Concluding remarks. In this paper, we have generalised the concept of
cherry-picking sequences as introduced in [12] and shown how this generalisation can
be used to characterise the minimum number of reticulation events that is needed to
explain any set of phylogenetic X-trees in the space of tree-child networks as well as in
the space of all phylogenetic networks. To see that these two minima can be different
for a fixed set of phylogenetic trees, consider the set P of trees presented in Figure 2.
It was shown in [13, 15] that h(P) = 3 and that there are six phylogenetic networks
each of which displays P and has a hybridisation number of three. However, none
of these six phylogenetic networks is tree child. Moreover, using tree-child sequences
a straightforward check shows that htc (P) = 4. Furthermore, we have shown that
Scoring Optimum Forest is NP-complete. Hence, given an optimum forest, it is
computationally hard to compute htc (P) for when P is a set of binary phylogenetic
X-trees, where |P| ≥ 3. This contrasts with the two-tree case for which Scoring
Optimum Forest is polynomial-time solvable and further hints at that agreement
forests are of limited use beyond the two-tree case.
Of course, restricting to collections of binary phylogenetic trees, one could generalise the definition of an acyclic-agreement forest for two binary phylogenetic trees
to more than two trees in the most obvious way. That is, one requires Conditions (i),
(ii), and (iii) in the definition of an acyclic-agreement forest to hold for each tree in an
arbitrarily large collection of binary phylogenetic X-trees. With this generalisation
in mind and observing that the number of components in a forest that is induced by
a tree-child network is equal to its number of reticulations plus one, one might conjecture that, given a set P of binary phylogenetic X-trees, the number of components
in a maximum acyclic-agreement forest for P is the same as the minimum number of
components in an optimum forest for P. To see that this is not true, we refer back to
Figure 1. Let F be the forest induced by N , and let F 0 be the forest induced by N 0 .
Since |F 0 | = 3, a maximum acyclic-agreement forest for P has at most three elements.
Moreover, since h(N )=3 and N is tree child, we have htc (P) ≤ 3. Indeed, it can be
checked that htc (P) = 3. Moreover, there is no tree-child network that displays P
and induces an optimum forest that is also a maximum acyclic-agreement forest for
P. Consequently, an approach that exploits maximum acyclic-agreement forests for a
set P of binary phylogenetic trees to compute htc (P), such as computing a maximum
acyclic-agreement forest F for P and, subsequently, scoring F in a way that reflects
the number of edges that are directed into each reticulation vertex in a network that
22
SIMONE LINZ AND CHARLES SEMPLE
induces F, is unlikely to give the desired result.
Lastly, from a computational viewpoint, the introduction of acyclic-agreement
forests [2] has triggered significant progress towards the development of ever faster
algorithms to solve Minimum Hybridisation for when the input contains exactly
two phylogenetic trees (e.g. see [1, 4, 8, 9, 20, 25]). We look forward to seeing a
similar development now for solving Minimum Hybridisation for arbitrarily many
phylogenetic trees by using cherry-picking sequences. In turn, this is likely to be of
benefit to biologists who often wish to infer evolutionary histories that are not entirely
tree-like and for data sets that usually consists of more than two phylogenetic trees.
REFERENCES
[1] B. Albrecht, C. Scornavacca, A. Cenci, and D. H. Huson (2012). Fast computation of minimum
hybridization networks. Bioinformatics, 28, 191–197.
[2] M. Baroni, S. Grünewald, V. Moulton, and C. Semple, (2005), Bounding the number of hybridization events for a consistent evolutionary history, Journal of Mathematical Biology,
51, 171–182.
[3] M. Bordewich and C. Semple (2007). Computing the minimum number of hybridization events
for a consistent evolutionary history. Discrete Applied Mathematics, 155, 914–928.
[4] M. Bordewich and C. Semple (2007). Computing the hybridization number of two phylogenetic
trees is fixed-parameter tractable. IEEE/ACM Transactions on Computational Biology
and Bioinformatics, 4, 458–466.
[5] M. Bordewich and C. Semple (2016). Determining phylogenetic networks from inter-taxa distances. Journal of Mathematical Biology, 73, 283–303.
[6] G. Cardona, F. Rosselló, and G. Valiente (2009). Comparison of tree-child phylogenetic networks. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 6, 552–
569.
[7] Z. Z. Chen and L. Wang (2012). Algorithms for reticulate networks of multiple phylogenetic
trees. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 9, 372–
384.
[8] Z. Z. Chen and L. Wang (2013). An ultrafast tool for minimum reticulate networks. Journal of
Computational Biology, 20, 38–41.
[9] J. Collins, S. Linz, and C. Semple (2011). Quantifying hybridization in realistic time. Journal
of Computational Biology, 18, 1305–1318.
[10] J.-M. Drezen, J. Gauthier, T. Josse , A. Bézier, E. Herniou, E. Huguet (2016). Foreign DNA acquisition by invertebrate genomes. Journal of Invertebrate Pathology, doi:
10.1016/j.jip.2016.09.004.
[11] P. J. Humphries, S. Linz, and C. Semple (2013). On the complexity of computing the temporal
hybridization number for two phylogenies. Discrete Applied Mathematics, 161, 871–880.
[12] P. J. Humphries, S. Linz, and C. Semple (2013). Cherry picking: a characterization of the
temporal hybridization number for a set of phylogenies. Bulletin of Mathematical Biology,
75, 1879–1890.
[13] L. van Iersel, S. Kelk, N. Lekić, C. Whidden, and N. Zeh, (2016). Hybridization number on
three rooted binary trees is EPT. SIAM Journal on Discrete Mathematics, 30, 1607–1631.
[14] L. van Iersel, C. Semple, and M. Steel (2010). Locating a tree in a phylogenetic network.
Information Processing Letters, 110, 1037–1043.
[15] S. Kelk (2012). Personal communication.
[16] S. Kelk, L. van Iersel, N. Lekić, S. Linz, C. Scornavacca, and L. Stougie (2012). Cycle killer
... qu’est-ce que c’est? On the comparative approximability of hybridization number and
directed feedback vertex set. SIAM Journal on Discrete Mathematics, 26,1635–1656.
[17] J. Mallet, N. Besansky, and M. W. Hahn (2016). How reticulated are species? BioEssays, 38,
140–149.
[18] T. Marcussen, S. R. Sandve, L. Heier, M. Spannagl, M. Pfeifer, International Wheat Genome
Sequencing Consortium, K. S. Jakobsen, B. B. Wulff, B. Steuernagel, K. F. Mayer, and
O. A. Olsen (2014). Ancient hybridizations among the ancestral genomes of bread wheat.
Science, 345, 1250092.
[19] B. M. E. Moret, L. Nakhleh, T. Warnow, C. R. Linder, A. Tholse, A. Padolina, J. Sun,
and R. Timme (2004). Phylogenetic networks: modeling, reconstructibility, and accuracy.
IEEE/ACM Transactions on Computational Biology and Bioinformatics, 1, 13–23.
MINIMUM HYBRIDISATION
23
[20] T. Piovesan and S. Kelk (2012). A simple fixed parameter tractable algorithm for computing
the hybridization number of two (not necessarily binary) trees, IEEE Transactions on
Computational Biology and Bioinformatics, 10, 18–25.
[21] C. Semple. Double-u-free networks, in preparation.
[22] J. Simpson. Tree display in tree-child networks. PhD thesis, University of Canterbury, in preparation.
[23] S. M. Soucy, J. Huang, and J. P. Gogarten (2015). Horizontal gene transfer: building the web
of life. Nature Reviews Genetics, 16, 472–482.
[24] V. G. Timkovskii (1989). Complexity of common subsequence and supersequence problems and
related problems. Cybernetics, 25, 565–580.
[25] Y. Wu and J. Wang (2010). Fast computation of the exact hybridization number of two phylogenetic trees. In: International Symposium on Bioinformatics Research and Applications,
Springer, pp. 203–214.
[26] Y. Wu (2010). Close lower and upper bounds for the minimum reticulate network of multiple
phylogenetic trees. Bioinformatics, 26, i140–i148.
| 8cs.DS
|
arXiv:0711.3628v1 [q-bio.PE] 22 Nov 2007
A Perl Package and an Alignment Tool for Phylogenetic
Networks
Francesc Rosselló
Research Institute of Health Science
University of the Balearic Islands
E-07122 Palma de Mallorca
Spain
Gabriel Cardona
Department of Mathematics
and Computer Science
University of the Balearic Islands
E-07122 Palma de Mallorca
Spain
Gabriel Valiente
Algorithms, Bioinformatics, Complexity
and Formal Methods Research Group
Technical University of Catalonia
E-08034 Barcelona
Spain
January 29, 2018
Abstract
Phylogenetic networks are a generalization of phylogenetic trees that allow for the
representation of evolutionary events acting at the population level, like recombination between genes, hybridization between lineages, and lateral gene transfer. While
most phylogenetics tools implement a wide range of algorithms on phylogenetic trees,
there exist only a few applications to work with phylogenetic networks, and there are no
open-source libraries either. In order to improve this situation, we have developed a Perl
package that relies on the BioPerl bundle and implements many algorithms on phylogenetic networks. We have also developed a Java applet that makes use of the aforementioned Perl package and allows the user to make simple experiments with phylogenetic
networks without having to develop a program or Perl script by herself. The Perl
package has been accepted as part of the BioPerl bundle. It can be downloaded from
the url http://dmi.uib.es/~gcardona/BioInfo/Bio-PhyloNetwork.tgz. The webbased application is available at the url http://dmi.uib.es/~gcardona/BioInfo/.
The Perl package includes full documentation of all its features.
Background
We briefly recall some definitions and results from [2] on phylogenetic networks.
A phylogenetic network on a set S of taxa is any rooted directed acyclic graph whose
leaves (those nodes without outgoing edges) are bijectively labeled by the set S.
Let N = (V, E) be a phylogenetic network on S. A node u ∈ V is said to be a tree node
if it has, at most, one incoming edge; otherwise it is called a hybrid node. A phylogenetic
network on S is a tree-child phylogenetic network if every node either is a leaf or has at
least one child that is a tree node.
Let S = {ℓ1 , . . . , ℓn } be the set of leaves. We define the µ-vector of a node u ∈ V as
the vector µ(u) = (m1 (u), . . . , mn (u)), where mi (u) is the number of different paths from
1
u to the leaf ℓi . The multiset µ(N ) = {µ(v) | v ∈ V } is called the µ-representation of
N and, provided that N is a tree-child phylogenetic network, it turns out to completely
characterize N , up to isomorphisms, among all tree-child phylogenetic networks on S.
This allows us to define a distance on the set of tree-child phylogenetic networks on S:
the µ-distance between two given networks N1 and N2 is the symmetric difference of their
µ-representations,
dµ (N1 , N2 ) = |µ(N1 ) △ µ(N2 )| .
This defines a true distance, and when N1 and N2 are phylogenetic trees, it coincides with
the well-known partition distance [8].
This representation also allows us to define an optimal alignment between two treechild phylogenetic networks on S, say n = |S|. Given two such networks N1 = (V1 , E1 ) and
N2 = (V2 , E2 ) (where, for the sake of simplicity, we assume |V1 | 6 |V2 |), an alignment is
just an injective mapping M : V1 → V2 . The weight of this alignment is
X
(kµ(v) − µ(M (v))k + χ(v, M (v))),
w(M ) =
v∈V1
where k · k stands for the Manhattan norm of a vector and χ(u, v) is 0 if both u and v are
tree nodes or hybrid nodes, and 1/(2n) if one of them is a tree node and the other one is a
hybrid node. An optimal alignment is, then, an alignment with minimal weight.
The Extended Newick Format
The eNewick (for “extended Newick”) string defining a phylogenetic network appeared in
the packages PhyloNet [7] and NetGen [5] related to phylogenetic networks, with some
differences between them. The former encodes a phylogenetic network with k hybrid nodes
as a series of k trees in Newick format, while the latter encodes it as a single tree in Newick
format but with k repeated nodes.
Whereas the Perl module we introduce here accepts both formats as input, a complete
standard for eNewick is implemented, based mainly on NetGen and following the suggestions of D. Huson and M. M. Morin (among others), to make it as complete as possible. The
adopted standard has the practical advantage of encoding a whole phylogenetic network as
a single string, and it also includes mandatory tags to distinguish among the various hybrid
nodes in the network.
The procedure to obtain the eNewick string representing a phylogenetic network N goes
as follows: Let {H1 , . . . , Hm } be the set of hybrid nodes of N , ordered in any fixed way. For
each hybrid node H = Hi , say with parents u1 , u2 , . . . , uk and children v1 , v2 , . . . , vℓ : split
H in k different nodes; let the first copy be a child of u1 and have all v1 , v2 , . . . , vℓ as its
children; let the other copies be children of u2 , . . . , uk (one for each) and have no children.
Label each of the copies of H as
[label]#[type]tag[:branch_length]
where the parameters are:
• label (optional) string providing a labelling for the node;
• type (optional) string indicating if the node H corresponds to a hybridization (indicated by H) or a lateral gene transfer (indicated by LGT) event; note that other types
can be considered in the future;
2
r
x
h
1
2
r
y
x
3
y
h h
1
2
3
Figure 1: A phylogenetic network N (left), and tree (right) associated to N for computing
its eNewick string.
r
r
x
x
h
y
1
2
3
1
2
3
Figure 2: Representation of a lateral gene transfer event (left) as a hybrid node in a
phylogenetic network (right).
• tag (mandatory) integer i identifying the node H = Hi .
• branch_length (optional) number giving the length of the branch from the copy of
H under consideration to its parent.
In this way, we get a tree whose set of leaves is the set of leaves of the original network
together with the set of hybrid nodes (possibly repeated). Then, the Newick string of
the obtained tree (note that some internal nodes will be labeled and some leaves will be
repeated) is the eNewick string of the phylogenetic network. The leftmost occurrence of
each hybrid node in an eNewick string corresponds to the full description of the network
rooted at that node, and although node labels are optional, all labeled occurrences of a
hybrid node in an eNewick string must carry the same label.
Consider, for example, the phylogenetic network depicted together with its decomposition in Figure 1. The eNewick string for this network would be ((1,(2)#H1),(#H1,3));
or ((1,(2)h#H1)x,(h#H1,3)y)r; if all internal nodes are labeled. The leftmost occurrence
of the hybrid node in the latter string corresponds to the full description of the network
rooted at that node: (2)h#H1.
Obviously, the procedure to recover a network from its eNewick string is as simple as
recovering the tree and identifying those nodes that are labeled as hybrid nodes with the
same identifier.
Notice that gene transfer events can be represented in a unique way as hybrid nodes.
Consider, for example, the lateral gene transfer event depicted in Figure 2, where a gene
is transferred from species 2 to species 3 after the divergence of species 1 from species
2. The eNewick string ((1,(2,(3)h#LGT1)y)x,h#LGT1)r; describes such a phylogenetic
network. A program interpreting the eNewick string can use the information on node types
in different ways; for instance, to render tree nodes circled, hybridization nodes boxed, and
lateral gene transfer nodes as arrows between edges.
3
The Perl Module
The Perl module Bio::PhyloNetwork implements all the data structures needed to work
with tree-child phylogenetic networks, as well as algorithms for:
• reconstructing a network from its eNewick string (in all its different flavours),
• reconstructing a network from its µ-representation,
• exploding a network into the set of its induced subtrees,
• computing the µ-representation of a network and the µ-distance between two networks,
• computing an optimal alignment between two networks,
• computing tripartitions [4, 3] and the tripartition error between two networks, and
• testing if a network is time consistent [1], and in such a case, computing a temporal
representation.
The underlying data structure is a Graph::Directed object, with some extra data, for
instance the µ-representation of the network. It makes use of the Perl module Bio::Phylo
Network::muVector that implements basic arithmetic operations on µ-vectors. Two extra
modules, Bio::PhyloNetwork::Factory and Bio::PhyloNetwork::RandomFactory, are
provided for the sequential and random generation (respectively) of all tree-child phylogenetic networks on a given set of taxa.
The web interface and the java applet
The web interface, available at http://dmi.uib.es/~gcardona/BioInfo/, allows the user
to input one or two phylogenetic networks, given by their eNewick strings. A Perl script
processes these strings and uses the Bio::PhyloNetwork package to compute all available
data for them, including a plot of the networks that can be downloaded in PS format; these
plots are generated through the application GraphViz and its companion Perl package.
Given two networks on the same set of leaves, their µ-distance is also computed, as well
as an optimal alignment between them. The algorithm to compute such an alignment relies
on the Hungarian algorithm [6]. If their sets of leaves are not the same, their topological
restriction on the set of common leaves is first computed followed by the µ-distance and an
optimal alignment.
A Java applet displays the networks side by side, and whenever a node is selected,
the corresponding node in the other network (with respect to the optimal alignment) is
highlighted, provided it exists. This is also extended to edges. Similarities between the
networks are thus evident at a glance and, since the weight of each matched node is also
shown, it is easy to see where the differences are.
Authors contributions
All authors conceived the method, prepared the manuscript, contributed to the discussion, and have approved the final manuscript. GC implemented the software. GV also
implemented part of the software.
4
Acknowledgements
The research described in this paper has been partially supported by the Spanish CICYT
project TIN 2004-07925-C03-01 GRAMMARS and by Spanish DGI projects MTM200607773 COMGRIO and MTM2006-15038-C02-01.
References
[1] Mihaela Baroni, Charles Semple, and Mike Steel. Hybrids in real time. Syst. Biol.,
55(1):46–56, 2006.
[2] Gabriel Cardona, Francesc Rosselló, and Gabriel Valiente. Comparison of tree-child
phylogenetic networks. http://arxiv.org/abs/0708.3499/, 2007.
[3] Gabriel Cardona, Francesc Rosselló, and Gabriel Valiente. Tripartitions do not always
discriminate phylogenetic networks. Math. Biosci., 2008. In press.
[4] Bernard M. E. Moret, Luay Nakhleh, Tandy Warnow, C. Randal Linder, Anna Tholse,
Anneke Padolina, Jerry Sun, and Ruth Timme. Phylogenetic networks: Modeling,
reconstructibility, and accuracy. IEEE T. Comput. Biol., 1(1):13–23, 2004.
[5] M. M. Morin and B. M. E. Moret. NETGEN: generating phylogenetic networks with
diploid hybrids. Bioinformatics, 22(15):1921–1923, 2006.
[6] J. Munkres. Algorithms for the assignment and transportation problems. J. SIAM,
5(1):32–38, 1957.
[7] Rice University BioInformatics Group. Phylonet: Phylogenetic networks toolkit (v. 1.4).
Available at http://bioinfo.cs.rice.edu/phylonet/, 2007.
[8] D. F. Robinson and L. R. Foulds. Comparison of phylogenetic trees. Math. Biosci.,
53(1/2):131–147, 1981.
5
| 5cs.CE
|
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
1
Deep Metric Learning for Practical Person
Re-Identification
arXiv:1407.4979v1 [cs.CV] 18 Jul 2014
Dong Yi, Zhen Lei, Member, IEEE Stan Z. Li, Fellow, IEEE
Abstract—Various hand-crafted features and metric learning
methods prevail in the field of person re-identification. Compared
to these methods, this paper proposes a more general way that
can learn a similarity metric from image pixels directly. By
using a “siamese” deep neural network, the proposed method
can jointly learn the color feature, texture feature and metric in
a unified framework. The network has a symmetry structure with
two sub-networks which are connected by Cosine function. To
deal with the big variations of person images, binomial deviance
is used to evaluate the cost between similarities and labels, which
is proved to be robust to outliers.
Compared to existing researches, a more practical setting is
studied in the experiments that is training and test on different
datasets (cross dataset person re-identification). Both in “intra
dataset” and “cross dataset” settings, the superiorities of the
proposed method are illustrated on VIPeR and PRID.
Index Terms—Person Re-Identification, Deep Metric Learning,
Convolutional Network, Cross Dataset
I. I NTRODUCTION
The task of person re-identification is to judge whether
two person images belong to the same subject or not. In
practical applications, the two images are usually captured by
two cameras with disjoint views. The performance of person
re-identification is closely related to many other applications,
such as cross camera tracking, behaviour analysis, object
retrieval and so on. The algorithms proposed in this field are
also overlapped with other fields in pattern recognition. In
recent years, the performance of person re-identification has
increased continuously and will increase further.
The essence of person re-identification is very similar to
biometric recognition problems, such as face recognition.
The core of them is to find a good representation and a
good metric to evaluate the similarities between samples.
Compared to biometric problems, person re-identification is
more challenging due to the low quality and high variety of
person images. Person re-identification usually need to match
the person images captured by surveillance cameras working in
wide-angle mode. Therefore, the resolution of person images
are very low (e.g., around 48 × 128 pixels) and the lighting
conditions are unstable too. Furthermore, the direction of
cameras and the pose of persons are arbitrary. These factors
cause the person images under surveillance scenarios have
two distinctive properties: large variations in intra class, and
ambiguities between inter classes. In summary, the challenges
of person re-identification come from the following aspects:
Dong Yi is with the National Laboratory of Pattern Recognition, Institute
of Automation, Chinese Academy of Sciences, Beijing, China, 100190 email:[email protected]
Manuscript received June 1, 2014; revised June 1, 2014.
camera view change, pose variation, non-rigid deformation,
unstable illumination and low resolution.
However, another challenge is less studied in existing work,
which is “cross dataset person re-identification”. In practical
systems, we usually collect some large datasets first and train
a model on them. Then the trained model is applied to other
datasets or videos for person re-identification. We call the
training datasets as source domain and test datasets as target
domain. The source and target datasets are totally different,
because they are usually captured by different cameras under
different environments, i.e., have different probability distribution. A practical person re-identification algorithm should
has good generalization with respect to the dataset changes.
Therefore, cross dataset person re-identification is an important
rule to evaluate the performance of algorithms in practice.
Since the pixels of person images are unstable, effective representations are important and needed for person reidentification. To this end, existing methods borrow many
sophisticated features from other fields, such as HSV histogram, Gabor, HOG and so on. Based on the features, direct
matching or discriminative learning are then used to evaluate
the similarity. Existing methods mainly focus on the second
step that is how to learn a metric to discriminate the persons.
Many good metric learning methods have been proposed in
this context, such as KISSME [1], RDC [2] and so on.
The majority of existing methods include two separate steps:
feature extraction and metric learning. The features usually
come from two separate sources: color and texture, some of
which are designed by hand, some of which are learned, and
they are finally connected or fused by simple strategies. On
the contrary, this paper proposes a new method to combine the
separate modules together, learning the color feature, texture
feature and metric in a unified framework, which is called as
“Deep Metric Learning” (DML).
The main idea of DML is inspired by a “siamese” neural network [3], which is originally proposed for signature
verification. Given two person images x and y, we want to
use a siamese deep neural network to assess their similarity
s = DM L(x, y). Being more specific than the original
work [3], our DML first abstracts the siamese network to
two sub-networks, a connection function and a cost function
(see Figure 1), and then carefully design the architecture
for person images. In this way, DML can adapt well to
person re-identification. Denoting the connection function
as S, the similarity equation of DML can be written as
s = DM L(x, y) = S(B1 (x), B2 (y)), where B1 and B2
denote the two sub-networks of DML. Depending on specific
applications, B1 and B2 need to share or not to share their
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
parameters.
Compared with existing person re-identification methods,
DML has the following advantages:
1) DML can learn a similarity metric from image pixels
directly. All layers in DML are optimized by a common
objective function, which are more effective than the
hand-crafted features in traditional methods;
2) The multi-channel filters learned in DML can capture
the color and texture information simultaneously, which
are more reasonable than the simple fusion strategies
in traditional methods, e.g., feature concatenation, sum
rule;
3) The structure of DML is flexible that can switch between
view specific and general person re-identification tasks
by whether sharing the parameters of sub-networks.
DML is tested on two popular person re-identification
datasets, VIPeR [4] and PRID 2011 [5], using the common
evaluation protocols. The results show that DML outperforms
or on a par with the state-of-the-art methods, such as [6],
and [7]. To appeal the practical requirements and evaluate
the generalization of DML, we also conduct more challenging
cross dataset experiments, which are training on i-LIDS [8],
CUHK Campus [6] and testing on VIPeR, PRID. The results
of the cross dataset experiments are significantly better than
existing methods [9] under similar experimental settings. To
our knowledge, this is the first work to conduct strict cross
dataset experiment in the field of person re-identification.
Finally, we fuse the features learned from multiple datasets
to improve the performance on VIPeR and PRID further.
II. R ELATED W ORK
This work uses deep learning to learn a metric for person
re-identification. Related works in four aspects are reviewed in
this section: feature representation, metric learning for person
re-identification, siamese convolutional neural network and
cross dataset related methods.
Early papers mainly focus on how to construct effective
feature representation. From 2005, numerous features are used
or proposed for person re-identification [10]. The features
can be divided into two categories: color based and texture
based features. The most popular features include HSV color
histogram [11], [6], LAB color histogram [12], SIFT [12], LBP
histogram [6], Gabor features [6] and their fusion. Among
these features, color has the most contribution to the final
results. The recent advance in this aspect is the color invariant
signature [13]. By combing color histogram, covariance feature and segmentation information, [13] achieved rank1=24%
on VIPeR.
On the other hand, [11] has proved that using the silhouette
and symmetry structure of person can improve the performance significantly, therefore the color and texture features
are usually extracted in a predefined grid or finely localized
parts. [14] proposed a MCMC based method to do part
localization and person re-identification simultaneously, which
obtained rank1=23% on VIPeR. The above work all prove that
the performance can be improved significantly when know
the geometric configuration of person explicitly or implicitly.
2
Compared to part based method, salience based methods [12],
[7] proposed by Zhao et al. relaxed the spatial constraint
further and could deal with larger pose variations. Similar to
the history of face recognition [15], the future direction of
feature representation must be based on precise body parts
segmentation, person alignment and pose normalization.
Based on the extracted features, naive feature matching or
unsupervised learning methods usually got moderate results,
and state-of-the-art results were achieved by supervised methods, such as Boosting [16], Rank SVM [17], PLS [17] and
Metric learning [2], [1], [6]. Among these methods, metric
learning is the main stream due to its flexibility. Compared
with standard distance measures, e.g., L1 , L2 norm, the learned
metric is more discriminative for the task on hand and more
robust to large variations of person images across view. Most
papers used a holistic metric to evaluate the similarity of two
samples, but [6] first divided the samples into several groups
according to their pose and then learned metrics for each
group. By using the pose information explicitly, [6] obtains
the highest performance on VIPeR. Recently, [18] proposed a
novel “human in the loop” style method, which illustrated that
the performance of person re-identification can be improved
drastically by human intervention. Although the results of this
paper is hard to reproduce, it supplies a benchmark to reflect
the performance of human (Rank1=71.08% on VIPeR). Closing the gap to human performance is the target of researchers.
Early in 1993, a siamese neural network [3] was proposed
to evaluate the similarity of two signature samples. In the
same year, a neural network [19] with similar structure was
proposed for fingerprint verification. Different from traditional
neural networks, the siamese architecture is composed by
two sub-networks sharing the same parameters. Each subnetwork is a convolutional neural network. Then the siamese
neural network was used for face verification [20] by the same
research group. The best property of siamese neural network is
its unified and clear objective function. Guided by the objective
function, the end-to-end neural network can learn a optimal
metric towards the target automatically. The responsibility of
the last layer of the siamese neural network is to evaluate
the similarity of the output of two sub-networks, which can
be in any form [20], such as L1 , L2 norm and Cosine.
Although good experimental results have been obtained in [3],
[19] and [20], their disadvantages are lacking implementation
details and lacking comparison with other methods. This paper
will design a siamese neural network for person image and
apply it in the person re-identification problem. In this paper,
the implementation details will be described and extensive
comparisons will be reported.
In this young field, cross dataset problem has not attracted
much attention. Majority researchers do their best to improve
the performance within single dataset, i.e., training on VIPeR
and test on VIPeR too. Not long ago, [9] started to concern
this issue. In [9], the authors proposed a transfer Rank SVM
(DTRSVM) to adapt a model trained on the source domain
(i-LIDS or PRID) to target domain (VIPeR). All image pairs
in the source domain and the negative image pairs in the target
domain were used for training. Different from DTRSVM, in
our cross dataset experiments, the proposed network is trained
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
3
Input
3@128x48
C1
64@48x48
S2
64@24x24
C3
64@24x24
S4
64@12x12
F5
500
CNN
Connection
Function
Cost
Function
Label
[-1, 1]
CNN
Fig. 1. The structure of the siamese convolutional neural network (SCNN),
which is composed by three components: CNN, connection function and cost
function.
Parameters
Sharing
Convolution
in the source domain only and its performance is tested in the
target domain.
Convolution
Normalization &
Max pooling
Full connection
Normalization &
Max pooling
Fig. 2. The structure of the 5-layer CNN used in our method.
III. D EEP M ETRIC L EARNING
Under the joint influence of resolution, illumination and
pose changes, the ideal metric for person re-identification may
be highly nonlinear. Deep learning is exact one of the most
effective tools to learn the nonlinear metric function. This
section introduces the architecture, parameters, cost function
and implementation details of the proposed convolutional
network for deep metric learning.
A. Architecture
For most of pattern recognition problems, neural network
works in a standalone mode. The input of neural network is a
sample and the output is a predicted label. This mode works
well for handwritten digit recognition, object recognition and
other classification problems when the labels of the training
set are the same as the test set. For person re-identification
problem, the subjects in the training set are generally different
from those in the test set, therefore the “sample → label” style
neural network cannot apply to it. To deal with this problem,
we construct a siamese neural network, which includes two
sub-networks working in a “sample pair → label” mode.
The flowchart of our method is shown in Figure 1. Given
two person images, they are sent to siamese convolutional
neural network (SCNN). For two images x and y, SCNN can
predict a label l = ±1 to denote whether the image pair comes
from the same subject or not. Because many applications need
rank the images in the gallery based on their similarities to a
probe image, our SCNN outputs a similarity score instead.
The structure of the SCNN is shown in Figure 1, which is
composed by two convolutional neural networks (CNN). And
the two CNNs are connected by a connection function.
Existing siamese neural networks have a constraint that their
two sub-networks share the same parameters, i.e., weights and
biases. As studied in our previous work [21], this constraint
could be removed in some conditions. Without parameters
sharing, the network can deal with the view specific matching
tasks more naturally. With parameters sharing, the network is
more appropriate for general task, e.g., cross dataset person
re-identification. We call these two modes as “General” and
“View Specific” SCNN. Because cross dataset problem is the
main concern of this paper, we focus on General SCNN.
B. Convolutional Neural Network
The CNN in this paper (see Figure 2) is composed by 2
convolutional layers, 2 max pooling layers and a full connected
layer. As shown in Figure 2, the number of channels of convolutional and pooling layers are both 64. The output of the CNN
is a vector of 500 dimensions. Every pooling layer includes a
cross-channel normalization unit. Before convolution the input
data are padded by zero values, therefore the output has the
same size with input. The filter size of C1 layer is 7 × 7 and
the filter size of C2 layer is 5 × 5. ReLU neuron [22] is used
as activation function for each layer.
To capture the different statistical properties of body parts,
we train the CNN in part based way. In our previous work [21],
person images are cropped into three overlapped parts and
three networks are trained independently. Differently, we use
a faster scheme in this paper that the three parts are trained
jointly. First, the three parts share C1 layer. Second, each part
has its own C3 layer, which can help to learn part-specific
filters. Third, the high level features of all parts are fused at
F5 layer by sum rule. Then, the similarity of fused features are
evaluated by the connection function. Driven by a common
cost function, the three parts can contribute to the training
process jointly.
Overall, there are two main differences between the proposed network in Figure 2 and the network in [21]: 1) C1 layer
are shared by three parts or not; 2) the contribution of the three
parts are fused in feature level or score level. Parameter sharing
in low level can reduce the complexity of the network. Fusion
in feature level make the three parts can train jointly, which
will improve the performance slightly. Moreover, training and
test a single network is more convenient and efficient than
using three independent networks.
C. Cost Function and Learning
Before learning the parameters of SCNN, we revisit its
structure again. As shown in Figure 1, the structure of SCNN
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
4
can be abstracted into three basic components: two subnetworks, a connection function and a cost function. Connection function is used to evaluate the relationship between two
samples and cost function is used to convert the relationship
into a cost. How to choose the connection function and cost
function is closely related to the performance of SCNN.
There are many distance, similarity, or other functions can
be used as candidates to connect two vectors, such as: Euclidian distance, Cosine similarity, absolute difference, vector
concatenate and so on. Their formulas are
X
Seuc (x, y) = −
(xi − yi )2 ,
(1)
i
P
Scos (x, y)
Sabs (x, y)
xi yi
pP i P
,
i xi xi
i yi yi
X
= −
|xi − yi |,
=
(2)
(3)
i
Scon (x, y)
X
=
wi [x; y]i .
(4)
i
In the above equations, we negate the distance functions to
make them consistent to similarity. The advantage of Euclidian
distance is that its derivation has simple form, but its output
is unbounded which could make the training process unstable.
Absolute difference is non-derivable at some points. Cosine
function is bounded to [-1, 1] and invariant to the magnitude
of samples. Because of the good property of Cosine function
and it has been used widely in many pattern recognition
problems [3], [23], [24], we choose it as the connection
function.
For the cost function, [21] has given some analysis about
Square loss, Exponential loss, and Binomial deviance [25] and
chose Binomial deviance as the final cost function. Here we
discuss another popular cost function in pattern recognition:
Fisher criterion. Given a training set X = [x1 , x2 , · · · , xn ],
its corresponding similarity matrix S and mask matrix M ,
Binomial deviance and Fisher criterion are formulated as
X
Jdev =
W ◦ ln(e−α(S−β)◦M + 1),
(5)
P
=
(
−P
i,j
2
P ◦ S)
i,j (S
− S̄)2
,
(6)
where ◦ is element-wise matrix product.
S = [Sij ]n×n ,
M = [Mij ]n×n ,
W = [Wij ]n×n ,
P = [Pij ]n×n ,
i,j
where X is the output of CNN. S is the output of connection
function.
xT xj
.
(12)
S = [Sij ]n×n , Sij = q i
xTi xi xTj xj
Differentiating the cost function with respect to X, we can get
∂Jdev
g + AD),
d
= X(AB + (AB)T ) − X ◦ (AC
∂X
where
e−α(S−β)◦M
,
e−α(S−β)◦M + 1
1
B = [Bij ]n×n , Bij = q
,
T
xi xi xTj xj
A = −αW ◦ M ◦
xTi xj
,
xTi xi
xT xj
D = [Dij ]n×n , Dij = Bij Ti ,
xj xj
T
X
g = repmat(
AC
Aij Cij , d, 1),
C = [Cij ]n×n , Cij = Bij
i,j
Jf isher
pairs. S̄ is the mean of S. α and β are hyper-parameters of
Binomial deviance. The numerator of Eqn. (6) is the between
class divergence of similarity matrix S and the denominator
is the total variance. By minimizing Eqn. (5) or Eqn. (6), we
can learn a network separating the positive and negative pairs
as far as possible.
By comparing Eqn. (5) and Eqn. (6), we can see that Binomial deviance cost more focus on the false classified samples
(or the samples near the boundary), but Fisher criterion focus
on every elements of the similarity matrix S equally. In our
intuition, Binomial deviance cost can make the network be
trained mainly on the hard samples and more likely to get a
good model, which will be verified in the experiments. In the
following sections, we fix the connection function as Cosine
and the cost function as Binomial deviance.
After the connection and cost functions are determined,
Back-Propagation (BP) [26] is used to learn the parameters
of SCNN. By plugging Eqn. (2) into Eqn. (5), we can get the
forward propagation function to calculate the cost of training
set.
X
Jdev (X) =
W ◦ ln(e−α(S−β)◦M + 1),
(11)
Sij = S(xi , xj ),
1, positive pair
−1, negative pair ,
Mij =
0, neglected pair
1
n1 , positive pair
1
, negative pair ,
Wij =
n2
0, neglected pair
1
n1 , positive pair
− 1 , negative pair .
Pij =
n2
0, neglected pair
(7)
(13)
(14)
(15)
(16)
(17)
(18)
j
(8)
X
d = repmat(
AD
Aij Dij , d, 1).
(19)
i
(9)
(10)
Sij denotes the similarity of sample xi and xj . Mij denotes
whether xi and xj come from the same subject or not. n1
is the count of positive pairs. n2 is the count of negative
where “repmat” is a function to create a matrix by tiling vector
many times, as same as the function in Matlab. One can refer
Appendix A for the detail derivation of Eqn. (13).
While [21] trains the network in a pairwise way, this paper
formulates the cost and gradient in totally matrix form. When
the network is trained by Stochastic Gradient Descent (SGD)
with mini-batch, the new formulation can process more sample
pairs in one batch. When the size of batch is set to 128, in [21]
a batch includes 64 positive and 64 negative sample pairs, but
in this paper a batch can generate 128 × (128 − 1)/2 = 8128
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
pairs, which makes the network scans the training data faster
and saves the training time.
Based on Eqn. (11) and Eqn. (13), we can learn the parameters of SCNN by SGD algorithm. For general neural network,
the error is backward propagated from top to down through
a single path. On the contrary, the error of specific SCNN is
backward propagated through two branches by Eqn. (27) and
Eqn. (33) respectively as described in [21]. In practice, we
also can assign asymmetry cost on the label Mij to positive
and negative pairs to tune the performance of network, e.g.,
1 for positive pairs and −2 for negative pairs. The effect of
asymmetry cost on the performance will be discussed in the
experiments.
IV. E XPERIMENTS
There are many popular datasets were built for person reidentification, such as VIPeR [4], PRID 2011 [5], i-LIDS [8],
CUHK Campus [6] and so on. Among these datasets, the
evaluation protocols of VIPeR and PRID (single-shot version)
are the clearest two, therefore we compare our method with
other methods on them. The experiments are done in two
settings:
1) Intra dataset experiments
• training and test both on VIPeR;
• training and test both on PRID.
2) Cross dataset experiments
• training on i-LIDS and test on VIPeR;
• training on i-LIDS and test on PRID;
• training on CUHK Campus and test on VIPeR;
• training on CUHK Campus and test on PRID.
The intra dataset experiments are conducted to illustrate the
basic performance of the proposed method. The cross dataset
experiments are to illustrate the generalization ability.
A. Intra Dataset
Except for [9] and [21], other papers all conduct experiments in this setting that is training and test on the same
dataset. VIPeR includes 632 subjects, 2 images per subject
coming from 2 different camera views (camera A and camera
B). We split VIPeR into disjoint training (316 subjects) and
testing set (316 subjects) randomly, and repeat the process 11
times. The first split (Dev. view) is used for parameter tuning,
such as the number of training epoch, learning rate, weight
decay and so on. The other 10 splits (Test view) are used for
reporting the results.
As similar to VIPeR, PRID is captured by 2 cameras too
(camera A and camera B). Camera A shows 385 subjects and
camera B shows 749 subjects. The first 200 subjects appear
in both cameras. Follow the testing protocols in [5], [27], we
randomly select 100 subjects from the first 200 subjects for
training and the remain subjects for testing. The test set is
composed by probe and gallery, and their information are as
follows.
• Probe set: the remain 100 subjects of the first 200 subjects
in camera A except for the 100 training subjects;
5
Gallery set: the remain 649 subjects of camera B except
for the 100 training subjects.
Again, the whole process is repeated 11 times. The first split
is used for parameter tuning and the other 10 splits are used
for reporting the results.
In the training stage, all training images from camera A
and camera B are merged, randomly shuffled and sent to a
generic SCNN. The corresponding mask matrix is generated
according to the label of samples. Those pairs from the same
subjects are assigned 1, and those from the different subjects
are assigned -1. Because the mask matrix is symmetric, we
set the elements in the lower triangular part to 0 to avoid
redundant computation. In the testing stage, one image of each
subject is used as gallery and the other one is used as probe.
Before evaluate on Test view, we use Dev. view to investigate the most three important factors which affect the
performance of network: 1) data augmentation, 2) asymmetric
cost for positive and negative sample pairs, and 3) cost
function. Besides, we set the parameters of the cost function
Eqn. (5) as α = 2 and β = 0.5.
1) Data Augmentation: Data augmentation is a widely used
trick [28] for training neural network. In the experiments, we
will mirror all person images to double the training and test
sets. Although this trick has been used in [21], but its effect
on the performance was not analyzed. Here, we compare the
performance on Dev. view of VIPeR with and without data
augmentation. For the training set, the number of images
of each subject is increased from 2 to 4. In the testing
stage, 2 original images and their mirrored version generate 4
similarity scores and the final score are fused by sum rule.
From the results in Table I, we can see significant improvements brought by data augmentation, especially in the
top ranks. This indicates that the scale of dataset is crucial
to train good networks. If know the geometry information of
the person in the image, i.e., the 3D pose of the person, we
can augment the datasets guided by the 3D pose and generate
more virtual images to improve the performance. In this sense,
human 3D pose estimation will be an important direction for
person re-identification in the following years.
2) Asymmetric Cost: As described in Section III-C, while
generating mask matrix from the labels of training samples, the
number of negative sample pairs is far more than positive pairs.
However, in practice we usually split the training samples
into many batches first, and then the positive and negative
sample pairs are just generated within each batch. Therefore,
the negative sample pairs between batches can not be covered
by the training process. This may cause the negative pairs
prone to under-fitting.
To balance the weight of positive and negative sample pairs,
we can assign asymmetric costs to them. While fixing the cost
of positive pair to 1, we tune the cost of negative pair c from
1 to 3. The asymmetric cost can apply easily on Eqn. (11) and
Eqn. (13) by setting
1, positive pair
−c, negative pair
Mij =
(20)
0, neglected pair
•
Table II shows the relationship between the negative cost
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
6
TABLE I
T HE RANK -1 TO RANK -50 RECOGNITION RATES ON D EV. VIEW OF VIP E R WITH OR WITHOUT DATA AUGMENTATION .
Rank
With Aug.
Without Aug.
1
34.49%
18.4%
5
60.13%
49.37%
10
74.37%
67.09%
15
80.7%
74.05%
20
84.18%
80.06%
25
88.61%
84.18%
30
91.14%
87.04%
50
96.84%
96.2%
0
0.65
Training set
Test set
Training set
Test set
0.6
−5
0.55
−10
Cost
Cost
0.5
0.45
−15
0.4
0.35
−20
0.3
0.25
0
20
40
60
80
100
Epoch
120
140
160
180
−25
0
20
40
60
80
100
Epoch
120
140
160
180
Fig. 3. The epoch-cost curves of the training and test set (Dev. view of
VIPeR) when using Binomial deviance (left) or Fisher criterion (right) as
cost function.
Fig. 5. The similarity score distribution of VIPeR computed based on the 21
human labelled attributes [29]. The 21 attributes are binary and the similarity
scores are evaluated by Hamming distance.
and the recognition rate. No matter using which cost function,
the cost of the training set always drop continually. On the
contrary, the cost of the test set drops significantly just at the
beginning of training process, and then it gradually becomes
converged after tens of epochs. From the figure we can see
that the cost gap between the training and test set is small
when using Binomial deviance. But for Fisher criterion, the
gap is obviously bigger than Binomial deviance, which reflects
that Fisher criterion is easily overfiting to the training set. By
inspecting the epoch-cost curves, we set epoch= 180 based
on our experience in the following experiments.
Fig. 4. The similarity score distributions of the training and test set (Dev.
view of VIPeR) when using Binomial deviance (left) or Fisher criterion (right)
as cost function.
c and recognition rate. On Dev. view of VIPeR, the highest
overall performance is achieved at c = 2. This illustrates
that the negative pairs should be paid more attention in each
training batch.
3) Binomial Deviance vs. Fisher Criterion: Next, we compare two cost functions discussed in Section III-C: Binomial
deviance (Eqn. (5)) and Fisher criterion (Eqn. (6)). Like the
two prior two experiments, the comparison is conducted on
Dev. view of VIPeR too. In the comparison experiments, we
mirror the images to double the dataset and set the cost for
negative pairs to c = 2. The connection function is fixed to
Cosine. The differences between Binomial deviance and Fisher
criterion are evaluated in three aspects: 1) epoch-cost curve,
2) similarity score distribution, and 3) recognition rate.
Figure 3 shows the epoch-cost curves on Dev. view of
VIPeR. Low cost reflects high performance approximately,
although there is no explicit relationship between the cost
As shown in Figure 4, the two-class similarity distributions
of the two cost functions are very different. And a distribution
generated from the attributes [29] of VIPeR is also given in
Figure 5 for reference. The experiment in [29] has shown that
only using the 21 attributes and Hamming distance can achieve
very high recognition rate, i.e., rank1≈ 76%. Because the
attributes are labelled by human, they can be seen as a baseline
of human performance, and the distribution in Figure 5 can be
seen as the ideal two-class distribution of VIPeR. For Binomial
deviance, the distribution of negative similarity scores are
significantly wider than that of positive scores, which is very
coincide with the ideal distribution generated from the 21
attributes. For Fisher criterion, the distributions of positive and
negative scores are standard Gaussian with the same variance.
Although nearly perfect results can be obtained on the
training set using any cost function, their performance on the
test set are different (see Table III). From the performance, we
can see that Binomial deviance is more suitable for person reidentification problem because it mainly focus on the samples
near the boundary and less affected by the distributions of
positive and negative samples. For Fisher criterion, the ideal
distributions are highly heteroscedastic, which is conflicted
with the assumption of Fisher. Maybe some modifications can
be made on Fisher criterion to solve this problem, but we leave
the work to future.
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
7
TABLE II
T HE RECOGNITION RATE ON D EV. VIEW OF VIP E R AT DIFFERENT NEGATIVE COSTS .
Rank
c=1
c = 1.5
c=2
c = 2.5
c=3
1
26.9%
31.01%
34.49%
29.75%
27.85%
5
53.48%
58.54%
60.13%
59.18%
56.65%
10
65.82%
70.89%
74.37%
71.84%
70.57%
15
74.05%
78.8%
80.7%
80.06%
78.8%
20
78.48%
84.49%
84.18%
85.13%
85.44%
25
85.76%
87.66%
88.61%
89.24%
87.97%
30
87.34%
90.82%
91.14%
91.77%
91.46%
50
94.94%
97.78%
96.84%
96.2%
94.94%
TABLE III
T HE RANK -1 TO RANK -50 RECOGNITION RATES ON D EV. VIEW OF VIP E R USING DIFFERENT COST FUNCTIONS .
Rank
Binomial Deviance
Fisher Criterion
1
34.49%
14.24%
5
60.13%
35.13%
10
74.37%
47.15%
4) Results: After the above analysis and parameter tuning,
we test the performance of our network on Test View of VIPeR
and PRID with the following configuration:
• Augment the training and set set by mirror the samples
and fused the score by sum;
• Set the negative cost c = 2; set the number of epoch=
180;
• Use Cosine as connection function and Binomial Deviance as cost function.
The experiments are repeated 10 times and the mean recognition rates are list in Table IV and Table V.
To distinguish the modified DML from the original one
in [21], we rename the proposed method in this paper as
“Improved DML”. Compared to the pairwise version in [21],
the improved DML has 5× training speed and significant
performance improvement. Notably, the rank-1 recognition
rate increases by 6%, from 28.23% to 34.4%. The results of
other compared methods are copied from the original papers.
If the results are unavailable, they are leaved as “-”. From
the table we can see that the proposed method outperforms
most of compared methods on VIPeR including the current
state-of-the-art methods [6], [7]. From rank-1 to rank-50, our
method outperforms [6], [7] remarkably, except that the rank50 recognition rate of [6] is higher than ours slightly. Among
all methods, our method is nearly the most simple and elegant
one, which doesn’t need any pre-processing, pose information
or segmentation. From the bottom to top layers in the network, every building blocks contribute to a common objective
function and are optimized by BP algorithm simultaneously.
On PRID, we get similar results to VIPeR that the proposed
method outperforms the state-of-the art method, RPML [27].
However, the superiority of our method on PRID (1% to 3%)
is smaller than that on VIPeR (4% to 10%), the reason is
that the scale of PRID dataset is too small to train a good
network. Just using 100 sample pairs, our network outperforms
the compared methods, which can illustrate the power of deep
metric learning. Compared to VIPeR, the quality of PRID
is poorer and the size of its gallery is bigger, therefore the
recognition rates on PRID are overall lower than VIPeR.
B. Cross Dataset
In this section, we conduct many experiments in the cross
dataset setting which is more coincide with practical applica-
15
80.7%
56.96%
20
84.18%
62.66%
25
88.61%
67.41%
30
91.14%
71.84%
50
96.84%
80.06%
TABLE VI
C ROSS DATASET EXPERIMENT: COMPARISON OF THE PROPOSED METHOD
AND DTRSVM [9] ON VIP E R.
Methods
DML [21]
DTRSVM [9]
DTRSVM [9]
Improved DML
Improved DML
Improved DML
Tr. Set
CUHK
PRID
i-LIDS
i-LIDS
CUHK
i-LIDS
+ CUHK
1
16.17%
10.9%
8.26%
11.61%
16.27%
10
45.82%
28.2%
31.39%
34.43%
46.27%
20
57.56%
37.69%
44.83%
44.08%
59.94%
30
64.24%
44.87%
53.88%
52.69%
70.13%
17.72%
48.8%
63.35%
72.85%
tions. In the following, i-LIDS, CUHK Campus are used as
training set, and VIPeR, PRID are used as test set. i-LIDS is
captured at a airport in indoor environment; CUHK Campus
is captured in a campus; VIPeR and PRID are captured on
the street. Due to totally different capture environments and
devices, cross dataset experiments are more challenging than
the previous experiments. Different from [9], we don’t use any
samples in VIPeR and PRID to adapt classifiers to the target
domains.
In i-LIDS dataset, there are 119 people with total 476
images captured by multiple non-overlapping cameras with an
average of 4 images for each person. Many of these images
have large illumination changes and are subject to occlusions.
The scale of i-LIDS is smaller than VIPeR and PRID and the
resolution of images in i-LIDS are varied. CUHK Campus is
a large scale dataset captured by two cameras in 5 sessions,
includes 1816 subjects and 7264 images. Each subject has
4 images from 2 camera views. The resolution of CUHK
Campus is 60 × 160. The cross dataset networks are trained
with the same parameters as the previous experiments.
Before training, we resize all images in i-LIDS and CUHK
Campus to 40 × 128 for convenience. For the test set, we
use the same setting with the intra dataset experiments. For
VIPeR, A half of subjects and images are randomly selected
to construct the test set, which includes 316 subjects and 632
images. For PRID, the test set includes 649 subjects and 749
images. The test process is repeated 10 times too, and the
average recognition rate is reported.
1) Results: First, we use i-LIDS and CUHK Campus as
training set respectively, and test the performance of the
trained networks on VIPeR. The recognition rates on VIPeR
are shown in Table VI and the results of DTRSVM [9] are
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
8
TABLE IV
I NTRA DATASET EXPERIMENT: COMPARISON OF THE PROPOSED METHOD AND OTHER STATE - OF - THE - ART METHODS ON VIP E R.
Rank
Method
ELF [16]
RDC [2]
PPCA [30]
Salience [12]
RPML [27]
LAFT [6]
Salience1 [7]
DML [21]
Improved DML
1
5
10
15
20
25
30
50
12%
15.66%
19.27%
26.74%
27%
29.6%
30.16%
28.23%
34.4%
31%
38.42%
48.89%
50.7%
52.3%
59.27%
62.15%
41%
53.86%
64.91%
62.37%
69%
69.31%
73.45%
75.89%
81.2%
82.56%
58%
70.09%
80.28%
76.36%
83%
86.39%
87.22%
88.7%
89.53%
89.65%
92.28%
92.28%
95%
96.8%
96.68%
96.52%
TABLE V
I NTRA DATASET EXPERIMENT: COMPARISON OF THE PROPOSED METHOD AND OTHER STATE - OF - THE - ART METHODS ON PRID 2011.
Rank
Method
Descr. Model [5]
RPML [27]
Improved DML
1
5
10
15
20
25
30
50
4%
15%
17.9%
37.5%
24%
42%
45.9%
50.7%
37%
54%
55.4%
59.3%
63.1%
56%
70%
71.4%
TABLE VII
C ROSS DATASET EXPERIMENT: COMPARISON OF THE PROPOSED METHOD
AND DTRSVM [9] ON PRID 2011.
Methods
DTRSVM [9]
DTRSVM [9]
Improved DML
Improved DML
Improved DML
Tr. Set
VIPeR
i-LIDS
i-LIDS
CUHK
i-LIDS
+ CUHK
1
4.6%
3.95%
8%
7.6%
10
17.25%
18.85%
25.5%
23.4%
20
22.9%
26.6%
38.9%
30.9%
30
28.1%
33.2%
45.6%
36.1%
13.8%
35.4%
45%
51.3%
(a) VIPeR
(b) PRID
listed for comparison. From the results we can see that our
method outperforms the original DML in [21] slightly. When
training on i-LIDS, our method is on a par with DTRSVM,
and the rank-1 and rank-10 recognition rates are better than
DTRSVM slightly. When training on CUHK Campus, our
performance improved significantly. A possible reason is that
compared to i-LIDS the quality and aspect ratio of images in
CUHK Campus is more similar to VIPeR. By combining the
similarity scores of the two networks by sum, the performance
is improved further and even approaches the performance of
some methods in intra dataset setting, such as ELF [16] and
RDC [2].
Then we test the two trained networks on PRID, and give
the results in Table VII. When using i-LIDS as training set,
our method is better than DTRSVM significantly. But different from the results on VIPeR, training on CUHK Campus
decreases the recognition rates remarkably. The performance
decline is caused by the big difference between CUHK Campus and PRID datasets. Fusing the similarity scores of i-LIDS
and CHUK Campus increases the performance too, especially
the rank-1 recognition raises from 8% to 13.8%.
Compared to the intra dataset experiments, the performance
of cross dataset experiments decline sharply. For rank-30
recognition rate, the number of VIPeR and PRID both drop
more than 20%, which indicates the trained models are hard
to generalize across datasets due to the distinctive properties
of each dataset. Figure 6 shows some filters learned from
(c) i-LIDS
(d) CUHK
Fig. 6. The filters in the first layer of CNN, from top to down, which are
learned on VIPeR, PRID, i-LIDS, CUHK Campus respectively. The size of
filters are 7 × 7, and their order are sorted by Hue component for best view.
different datasets, from which we can see the distributions of
color and texture on these datasets are very diverse. Besides
of the experimental results, the diverse filters also can prove
that the model learned on a dataset can hardly adapt to another
one. How to transfer a model to target domain or how to fully
use the multiple heterogeneous datasets at hand to improve the
performance in target domain are important research topics in
the future.
V. C ONCLUSIONS
This paper proposed a deep metric learning method by
using siamese convolutional neural network. The structure
of the network and the training process were described in
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
9
detail. Extensive intra dataset and cross dataset person reidentification experiments were conducted to illustrate the
superiorities of the proposed method. This is the first work
to apply deep learning in the person re-identification problem
and is also the first work to study the person re-identification
problem in fully cross dataset setting. The experimental results
illustrated that the network can deal with the cross view and
cross dataset person re-identification problems efficiently and
outperformed the state-of-the-art methods significantly. In the
future, we will apply DML to other applications; explore
the way to pre-train the network; and investigate how to
embed geometry information into the network to improve the
robustness to pose variations. Moreover, we will continue to
research how to train a general person matching engine with
good generalization across view and dataset.
j
∂xTj xj
= 2 0 xj 0
(25)
∂X
By substituting Eqn. (23), Eqn. (24) and Eqn. (25) into
Eqn. (22) and substituting Eqn. (22) into Eqn. (21), we can
dev
get the final formulation of ∂J∂X
.
∂Jdev X
=
Aij Bij
0
∂X
ij
−αW ◦ M ◦
ij
=
X
Aij
ij
e−α(S−β)◦M ∂Sij
+ 1 ∂X
e−α(S−β)◦M
(21)
∂Sij
,
∂X
−
∂X
0
∂xTi xi
=2
∂X
0
i
0 xi
X
Aij Cij
Aij Dij
0
xi
0
(26)
j
−
X
0
xj
0
g is defined in Eqn. (18) and AD
d is defined in
where AC
Eqn. (19).
A PPENDIX B
G RADIENTS OF V IEW S PECIFIC SCNN
Although this paper doesn’t use specific (asymmetric)
SCNN for experiment, we give the gradients of cost function
for reference. We denote the output of two sub-networks by
X = [x1 , x2 , · · · , xn ] and Y = [y1 , y2 , · · · , xm ], where n
is the number of samples in view1 and m is the number
of samples in view2. The dimensions of their corresponding
similarity matrix S and mask matrix M are n × m.
Following the derivation process in Appendix A, the gradients of Jdev with respect to X and Y are as follows.
∂Jdev
g
= Y (EF )T − X ◦ EG,
∂X
(27)
where
e−α(S−β)◦M
,
+1
1
F = [Fij ]n×n , Fij = q
,
xTi xi yjT yj
E = −αW ◦ M ◦
e−α(S−β)◦M
G = [Gij ]n×n , Gij = Fij
xTi yj
,
xTi xi
T
X
g = repmat(
Eij Gij , d, 1).
EG
(28)
(29)
(30)
(31)
(32)
And
∂Jdev
d
= X(EF ) − Y ◦ EH,
∂Y
(33)
where
j
xi
0
0
j
(22)
where Bij , Cij and Dij are defined in Eqn. (15), Eqn. (16)
and Eqn. (17) respectively. The derivatives of xTi xj , xTi xi and
xTj xj with respect to X are as follows.
=
xi
ij
∂xT xj
Cij ∂xTi xi
Dij ∂xTj xj
−
−
,
=Bij i
∂X
2 ∂X
2 ∂X
i
xj
0
g + AD),
d
=X(AB + (AB)T ) − X ◦ (AC
where Ai j can be seen as a weight for each sample pair. And
∂Sij
∂X is the derivative of the similarity Sij with respect to X.
Expand the Cosine similarity Sij by Eqn. (2), we can get
p
q
T
xi xT
x
∂ xT
i
j j
T x xT x ∂xi xj − xT x
x
i j
i i j j ∂X
∂X
∂Sij
=
∂X
xTi xi xTj xj
1
=q
xTi xi xTj xj
!
∂xTi xj
xTi xj ∂xTi xi
xTi xj ∂xTj xj
− T
− T
∂X
2xi xi ∂X
2xj xj ∂X
∂xTi xj
xj
ij
For general SCNN, the two sub-networks share their parameters, therefore they have the same output for a sample. Input a training set, we denote the output of CNN by
X = [x1 , x2 , · · · , xn ], where n is the number of samples.
The cost produced by the training set X can be calculated by
Eqn. (11). The gradient of Eqn. (11) with respect to X can be
derived as follows.
∂Jdev X ∂Jij
=
∂X
∂X
ij
X
j
i
A PPENDIX A
G RADIENTS OF G ENERAL SCNN
=
i
0
(23)
(24)
xTi yj
,
yjT yj
(34)
X
d = repmat(
EH
Eij Hij , d, 1).
(35)
H = [Hij ]n×n , Hij = Fij
i
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
ACKNOWLEDGMENT
This work was supported by the Chinese National Natural Science Foundation Projects #61105023, #61103156,
#61105037, #61203267, #61375037, National Science and
Technology Support Program Project #2013BAK02B01, Chinese Academy of Sciences Project No. KGZD-EW-102-2,
Jiangsu Science and Technology Support Program Project
#BE2012627, and AuthenMetric R&D Funds.
R EFERENCES
[1] M. Kostinger, M. Hirzer, P. Wohlhart, P. Roth, and H. Bischof, “Large
scale metric learning from equivalence constraints,” in Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE Conference on, 2012, pp.
2288–2295.
[2] W.-S. Zheng, S. Gong, and T. Xiang, “Reidentification by relative
distance comparison,” Pattern Analysis and Machine Intelligence, IEEE
Transactions on, vol. 35, no. 3, pp. 653–668, 2013.
[3] J. Bromley, I. Guyon, Y. LeCun, E. Säckinger, and R. Shah, “Signature
verification using a siamese time delay neural network,” in NIPS, 1993,
pp. 737–744.
[4] D. Gray, S. Brennan, and H. Tao, “Evaluating appearance models
for recognition, reacquisition, and tracking,” in In IEEE International
Workshop on Performance Evaluation for Tracking and Surveillance,
Rio de Janeiro, 2007.
[5] M. Hirzer, C. Beleznai, P. Roth, and H. Bischof, “Person re-identification
by descriptive and discriminative classification,” in Image Analysis, ser.
Lecture Notes in Computer Science, A. Heyden and F. Kahl, Eds.
Springer Berlin Heidelberg, 2011, vol. 6688, pp. 91–102.
[6] W. Li and X. Wang, “Locally aligned feature transforms across views,”
in Computer Vision and Pattern Recognition (CVPR), 2013 IEEE
Conference on, 2013, pp. 3594–3601.
[7] R. Zhao, W. Ouyang, and X. Wang, “Person re-identification by salience
matching,” in Proceedings of IEEE International Conference on Computer Vision, 2013, pp. 2528–2535.
[8] W.-S. Zheng, S. Gong, and T. Xiang, “Person re-identification by
probabilistic relative distance comparison,” in Computer Vision and
Pattern Recognition (CVPR), 2011 IEEE Conference on, 2011, pp. 649–
656.
[9] A. J. Ma, P. C. Yuen, and J. Li, “Domain transfer support vector ranking
for person re-identification without target camera label information,”
in Proceedings of IEEE International Conference on Computer Vision,
2013, pp. 3567–3574.
[10] O. Javed, K. Shafique, and M. Shah, “Appearance modeling for tracking
in multiple non-overlapping cameras,” in Computer Vision and Pattern
Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on,
vol. 2, 2005, pp. 26–33 vol. 2.
[11] M. Farenzena, L. Bazzani, A. Perina, V. Murino, and M. Cristani,
“Person re-identification by symmetry-driven accumulation of local
features,” in Computer Vision and Pattern Recognition (CVPR), 2010
IEEE Conference on, 2010, pp. 2360–2367.
[12] R. Zhao, W. Ouyang, and X. Wang, “Unsupervised salience learning for
person re-identification,” in Computer Vision and Pattern Recognition
(CVPR), 2013 IEEE Conference on, 2013, pp. 3586–3593.
[13] I. Kviatkovsky, A. Adam, and E. Rivlin, “Color invariants for person reidentification,” Pattern Analysis and Machine Intelligence, IEEE
Transactions on, vol. 35, no. 7, pp. 1622–1634, 2013.
[14] Y. Xu, L. Lin, W.-S. Zheng, and X. Liu, “Human re-identification by
matching compositional template with cluster sampling.” ICCV, 2013.
[15] X. Zhang and Y. Gao, ““Face recognition across pose: A review”,”
Pattern Recognition, vol. 42, no. 11, pp. 2876–2896, 2009.
[16] D. Gray and H. Tao, “Viewpoint invariant pedestrian recognition with
an ensemble of localized features,” in Computer Vision C ECCV 2008,
ser. Lecture Notes in Computer Science, D. Forsyth, P. Torr, and
A. Zisserman, Eds. Springer Berlin Heidelberg, 2008, vol. 5302, pp.
262–275.
[17] B. Prosser, W.-S. Zheng, S. Gong, and T. Xiang, “Person re-identification
by support vector ranking,” in BMVC, 2010, pp. 1–11.
[18] C. Liu, C. C. Loy, S. Gong, and G. Wang, “Pop: Person re-identification
post-rank optimisation,” in International Conference on Computer Vision, 2013.
[19] P. Baldi and Y. Chauvin, “Neural networks for fingerprint recognition,”
Neural Computation, vol. 5, no. 3, pp. 402–418, 1993.
10
[20] S. Chopra, R. Hadsell, and Y. LeCun, “Learning a similarity metric
discriminatively, with application to face verification,” in CVPR (1),
2005, pp. 539–546.
[21] D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Deep metric learning for person
re-identification,” in Proceedings of International Conference on Pattern
Recognition, 2014, pp. 2666–2672.
[22] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in NIPS, 2012, pp. 1106–
1114.
[23] T. Qin, X.-D. Zhang, M.-F. Tsai, D.-S. Wang, T.-Y. Liu, and H. Li,
“Query-level loss functions for information retrieval,” Information Processing & Management, vol. 44, no. 2, pp. 838–855, 2008.
[24] H. Nguyen and L. Bai, “Cosine similarity metric learning for face
verification,” in Computer Vision C ACCV 2010, ser. Lecture Notes
in Computer Science, R. Kimmel, R. Klette, and A. Sugimoto, Eds.
Springer Berlin Heidelberg, 2011, vol. 6493, pp. 709–720.
[25] J. Friedman, R. Tibshirani, and T. Hastie, The Elements of Statistical
Learning : Data Mining, Inference, and Prediction, ser. Springer Series
in Statistics. New York: Springer-Verlag, 2009.
[26] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[27] M. Hirzer, P. Roth, M. Köstinger, and H. Bischof, “Relaxed pairwise
learned metric for person re-identification,” in Computer Vision C
ECCV 2012, ser. Lecture Notes in Computer Science, A. Fitzgibbon,
S. Lazebnik, P. Perona, Y. Sato, and C. Schmid, Eds. Springer Berlin
Heidelberg, 2012, vol. 7577, pp. 780–793.
[28] D. C. Cireşan, U. Meier, J. Masci, L. M. Gambardella, and J. Schmidhuber, “Flexible, high performance convolutional neural networks for
image classification,” in Proceedings of the Twenty-Second international
joint conference on Artificial Intelligence-Volume Volume Two. AAAI
Press, 2011, pp. 1237–1242.
[29] R. Layne, T. M. Hospedales, and S. Gong, “Attributes-based reidentification,” in Person Re-Identification. Springer, 2014, pp. 93–117.
[30] A. Mignon and F. Jurie, “Pcca: A new approach for distance learning
from sparse pairwise constraints,” in Computer Vision and Pattern
Recognition (CVPR), 2012 IEEE Conference on, 2012, pp. 2666–2672.
Dong Yi received the B.S. degree in electronic
engineering in 2003, the M.S. degree in communication and information system from Wuhan University, Wuhan, China, in 2006, and received the
Ph.D. degree in pattern recognition and intelligent
systems from CASIA, Beijing, China, in 2009. His
research areas are unconstrained face recognition,
heterogeneous face recognition, and deep learning.
He has authored and acted as a reviewer for tens of
articles in international conferences and journals. He
developed the face biometric algorithms and systems
for the Shenzhen-Hongkong immigration control project and 2008 Beijing
Olympic Games.
PLACE
PHOTO
HERE
Zhen Lei received the B.S. degree in automation
from the University of Science and Technology of
China (USTC), in 2005, and the Ph.D. degree from
the Institute of Automation, Chinese Academy of
Sciences (CASIA), in 2010, where he is now an
assistant professor. His research interests are in computer vision, pattern recognition, image processing,
and face recognition in particular. He has published
over 40 papers in international journals and conferences.
JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012
Stan Z. Li received the B.Eng. degree from Hunan
University, Changsha, China, the M.Eng. degree
from the National University of Defense Technology,
China, and the Ph.D. degree from Surrey UniverPLACE
sity, Surrey, U.K. He is currently a Professor and
PHOTO
the Director of Center for Biometrics and Security
HERE
Research (CBSR), Institute of Automation, Chinese
Academy of Sciences (CASIA). He worked at Microsoft Research Asia as a researcher from 2000 to
2004. Prior to that, he was an Associate Professor at
Nanyang Technological University, Singapore. His
research interest includes pattern recognition and machine learning, image
and vision processing, face recognition, biometrics, and intelligent video
surveillance. He has published over 200 papers in international journals and
conferences, and authored and edited eight books.
11
| 9cs.NE
|
Low noise sensitivity analysis of ℓq -minimization in oversampled
systems ∗
arXiv:1705.03533v2 [math.ST] 18 Feb 2018
Haolei Weng and Arian Maleki
Abstract
The class of ℓq -regularized least squares (LQLS) are considered for estimating β ∈ Rp from its n noisy
linear observations y = Xβ + w. The performance of these schemes are studied under the high-dimensional
asymptotic setting in which the dimension of the signal grows linearly with the number of measurements.
In this asymptotic setting, phase transition diagrams (PT) are often used for comparing the performance
of different estimators. PT specifies the minimum number of observations required by a certain estimator
to recover a structured signal, e.g. a sparse one, from its noiseless linear observations. Although phase
transition analysis is shown to provide useful information for compressed sensing, the fact that it ignores
the measurement noise not only limits its applicability in many application areas, but also may lead to
misunderstandings. For instance, consider a linear regression problem in which n > p and the signal is
not exactly sparse. If the measurement noise is ignored in such systems, regularization techniques, such
as LQLS, seem to be irrelevant since even the ordinary least squares (OLS) returns the exact solution.
However, it is well-known that if n is not much larger than p then the regularization techniques improve the
performance of OLS.
In response to this limitation of PT analysis, we consider the low-noise sensitivity analysis. We show that
this analysis framework (i) reveals the advantage of LQLS over OLS, (ii) captures the difference between
different LQLS estimators even when n > p, and (iii) provides a fair comparison among different estimators
in high signal-to-noise ratios. As an application of this framework, we will show that under mild conditions
LASSO outperforms other LQLS even when the signal is dense. Finally, by a simple transformation we
connect our low-noise sensitivity framework to the classical asymptotic regime in which n/p → ∞ and
characterize how and when regularization techniques offer improvements over ordinary least squares, and
which regularizer gives the most improvement when the sample size is large.
Key Words: High-dimensional linear model, ℓq -regularized least squares, ordinary least squares,
LASSO, phase transition, asymptotic mean square error, second-order expansion, classical asymptotics.
1 Introduction
1.1 Problem Statement
In modern data analysis, one of the fundamental models that has been extensively studied is the high-dimensional
linear model y = Xβ + w with y ∈ Rn , X ∈ Rn×p , where the length of the signal p is at the same order or
in some cases larger than the number of observations n. Since the ordinary least squares (OLS) estimate is not
accurate in the high-dimensional regime, researchers have proposed a wide range of regularization techniques
∗
1
and recovery algorithms to go beyond OLS. The existence of a variety of algorithms and regularizers has in
turn called for platforms that can provide fair comparisons among them. One of the most popular platforms
is the phase transitions analysis (PT). Intuitively speaking, phase transition diagram measures the minimum
number of observations an algorithm or an estimator requires to recover β. To understand the limitations of
PT we study the performance of the following ℓq -regularized least squares (LQLS), a.k.a. bridge estimators
Frank and Friedman (1993),
1
β̂(λ, q) ∈ arg min ky − Xβk22 + λkβkqq ,
2
β
1 ≤ q ≤ 2,
(1)
where k · kq is the usual ℓq norm and λ ≥ 0 is a tuning parameter. This family covers LASSO (Tibshirani, 1996)
and Ridge (Hoerl and Kennard, 1970), two well known estimates in high-dimensional statistics and compressed
sensing.
Phase transition analysis studies the asymptotic mean square error (AMSE) kβ̂(λ, q) − βk22 /p under the
asymptotic setting p → ∞ and n/p → δ. Then, it considers w = 0 and calculates the smallest δ for which
inf λ limp→∞ kβ̂(λ, q) − βk22 /p = 0. In this paper, we consider situations in which β is not exactly sparse. As is
intuitively expected and will be discussed later in the paper, the phase transition analysis implies that for every
1 ≤ q ≤ 2, if δ > 1, then inf λ limp→∞ kβ̂(λ, q) − βk22 /p = 0 and if δ < 1, then inf λ limp→∞ kβ̂(λ, q) −
βk22 /p 6= 0. This simple application of PT reveals some of the limitations of the phase transition analysis:
1. Phase transition analysis is concerned with w = 0, and when β is not sparse, LQLS with different values
of q have the same phase transition at δ = 1. Hence, it is not clear whether regularization can improve
the performance of ordinary least squares (OLS) and if it does, which regularizer is the best. We expect
the choice of regularizer to matter when we add some noise to the measurements.
2. Phase transition diagram is not sensitive to the magnitudes of the elements of β. Again, intuitively
speaking, this seems to have a major impact on the performance of different estimators when the noise is
present in the system.
This paper aims to present a generalization of the phase transition, called low-noise sensitivity analysis, in
which it is assumed that a small amount of noise is present in the measurements. Then it calculates AMSE of
each estimate. This framework has the following two main advantages over the phase transition analysis:
1. It reveals certain phenomena that are important in applications and are not captured by PT analysis. For
instance, one immediately sees the impact of the regularizer and the magnitudes of the elements of β on
the AMSE. Furthermore, these relations are expressed explicitly and can be interpreted easily.
2. It provides a bridge between the phase transition analysis proposed in compressed sensing, and the classical large sample-size asymptotic (n/p → ∞). We will discuss some of the implications of this connection
for the classical asymptotics in Section 3.3.
As a consequence, low noise sensitivity analysis enables us to present a fair comparison among different LQLS,
and reveal different factors that affect their performance.
Here is the organization of the rest of this paper: In Section 1.2, we discuss some related works and highlight the differences from ours. In section 2, we formally introduce the asymptotic framework adopted in our
analyses. In Section 3, we present and discuss our main contributions in details. In Section 4, we prove all the
main results.
2
1.2 Related work
The phase transition analysis for compressed sensing evolved in a series of papers by Donoho and Tanner
Donoho (2004a,b, 2006a); Donoho and Tanner (2005b,a). Donoho and Tanner characterized the phase transition curve for LASSO and some of its variants. Inspired by Donoho and Tanner’s breakthrough, many researchers explored the performance of different algorithms under the asymptotic settings n/p → δ and p → ∞
Reeves and Donoho (2013); Stojnic (2009b,a, 2013); Amelunxen et al. (2014); Thrampoulidis et al. (2016);
El Karoui et al. (2013); Karoui (2013); Donoho and Montanari (2013); Donoho et al. (2013); Donoho and Montanari
(2015); Bradic and Chen (2015); Donoho et al. (2011a,b); Foygel and Mackey (2014); Zheng et al. (2017); Rangan et al.
(2009); Krzakala et al. (2012); Bayati and Montanri (2011); Bayati and Montanari (2012); Reeves and Gastpar
(2008); Reeves and Pfister (2016). In this paper, we use the message passing analysis that was developed in a series of papers Donoho et al. (2011b, 2009); Maleki (2010); Bayati and Montanri (2011); Bayati and Montanari
(2012); Maleki et al. (2013) to characterize the asymptotic mean square errors of LQLS. As a result of this
calculation we obtain some equations whose solution can specify AMSE of an estimate. Unfortunately, the
complexity of these equations does not allow us to interpret the results and obtain useful information. Hence,
we develop a machinery to simplify the AMSE formulas and turn them to explicit and informative quantities.
Note that the AMSE formulas we derive in Theorem 1 can be calculated by the framework developed in
Thrampoulidis et al. (2016) too. Furthermore, the phase transition formulas we derive in Theorem 2 can be
derived from the framework of Amelunxen et al. (2014) as well. The low noise sensitivity analysis that we use
in this paper has been also used in Weng et al. (2016). But the analysis of Weng et al. (2016) is only concerned
with sparse signals. This paper avoids the sparsity assumption. Perhaps surprisingly, the proof techniques developed in Weng et al. (2016) for the sparse signals come short of characterizing higher order terms in dense
cases. In response to that limitation, we propose a delicate chaining argument in this paper that offers a much
more accurate characterization of the higher order terms. The study of dense signals provides much more complete understanding of the the bridge estimators. In particular, under the sparse settings, Weng et al. (2016)
reveals the monotonicity of LQLS’s performance: LASSO is optimal, and the closer q is to 1, the better LQLS
performs. However, in this paper we show that the comparison and optimality characterization of LQLS becomes much more subtle for non-sparse signals. A general and detailed treatment for different types of dense
signals are given in Section 3. Finally, we should also emphasize that the machinery required for the low-noise
sensitivity analysis of dense signals is different from (and much more complicated) the one developed for sparse
signals.
Our paper performs the asymptotic analysis of LQLS. Many researchers have used non-asymptotic frameworks for this purpose. Among all the LQLS, LASSO is the best studied one. The past decade has witnessed dramatic progress towards understanding the performance of LASSO in the tasks of parameter estimation (Donoho, 2006b; Donoho and Tanner, 2005a; Candès et al., 2006; Donoho et al., 2009; Bickel et al., 2009;
Raskutti et al., 2011), variable selection (Zhao and Yu, 2006; Wainwright, 2009; Meinshausen and Bühlmann,
2006; Reeves and Gastpar, 2008), and prediction (Greenshtein et al., 2004). We refer the reader to Bühlmann and Van De Geer
(2011); Eldar and Kutyniok (2012) for a complete list of references. The ones most related to our work are
Donoho (2006b); Candes and Tao (2006); Candes and Plan (2011). In the first two papers, the authors considered non-sparse β with the constraint that kβkq ≤ R or the ith largest component |β|(i) decays as i−α (α > 0).
The papers derived optimal (up to logarithmic factor) upper bounds on the mean square error of LASSO. However, in this paper we characterize the performance of LASSO for a generic β and derive conditions under
which LASSO outperforms other bridge estimators. Also, we should emphasize that thanks to our asymptotic
settings, unlike these two papers we are able to derive exact expressions of AMSE with sharp constants. Finally,
Candes and Plan (2011) studied a fixed signal β and obtained an oracle inequality for kβ̂(1, λ) − βk2 , with the
3
tuning λ chosen as an explicit function of p. While their results are more general than ours, the bounds suffer
from loose constants and are not sufficient to provide sharp comparison of LASSO with other LQLS. Moreover,
the tuning parameter λ in our case is set to the optimal one that minimizes the AMSE for every LQLS, which
further paves our way for accurate comparison between different LQLS.
Finally, the performance of LQLS with q ≥ 0 under classical asymptotic setting where p is fixed and
√
n → ∞ is studied in Knight and Fu (2000). The author obtained the n convergence of LQLS estimates and
derived the asymptotic distributions. His results can be used to calculate the AMSE for LQLS with optimal
tuning and show that they are all equal for q ∈ [1, 2]. However, we demonstrate in Section 3.3 that by a
second-order analysis, a more accurate comparison between the performances of different LQLS is possible. In
particular, LASSO will be shown to outperform others for certain type of dense signals. We should also mention
that the idea of obtaining higher order terms for large sample scenarios was first introduced in Wang et al.
(2017). However, all the results in Wang et al. (2017) are only concerned with the sparse signals. In this paper,
we consider dense signals. We discussed the main challenges of switching from sparse signals to dense signals
earlier in this section.
2 The asymptotic framework
The main goal of this section is to formally introduce the asymptotic framework under which we study LQLS.
In the current section only, we write vectors and matrices as X(p), β(p), y(p), and w(p) to emphasize their
dependence on the dimension of β ∈ Rp . Similarly, we may use β̂(λ, q, p) as a substitute for β̂(λ, q). We
first define a specific type of a sequence known as a converging sequence. Our definition is borrowed from
other papers Donoho et al. (2011b); Bayati and Montanri (2011); Bayati and Montanari (2012) with some minor modifications. Recall that we have the linear model y(p) = X(p)β(p) + w(p).
Definition 1. A sequence of instances {X(p), β(p), w(p)} is called a converging sequence if the following
conditions hold:
- n/p → δ ∈ (0, ∞), as n → ∞.
- The empirical distribution of β(p) ∈ Rp converges weakly to a probability measure pβ with bounded
second moment. Further, 1p kβ(p)k22 converges to the second moment of pβ .
2.
- The empirical distribution of w(p) ∈ Rn converges weakly to a zero mean distribution with variance σw
2.
Furthermore, n1 kw(p)k22 → σw
- The elements of X(p) are iid with distribution N (0, 1/n).
For each of the problem instances in the converging sequence, we solve the LQLS problem (1) and obtain
β̂(λ, q, p) as the estimate. The goal is to evaluate the accuracy of this estimate. Below we define the asymptotic
mean square error as the asymptotic measures of performance.
Definition 2. Let β̂(λ, q, p) be the sequence of solutions of LQLS for the converging sequence of instances
{β(p), X(p), w(p)}. The asymptotic mean square error of β̂(λ, q, p) is defined as the following almost sure
limit
p
1X
(β̂i (λ, q, p) − βi (p))2 .
(2)
AMSE(λ, q, σw ) , lim
p→∞ p
i=1
4
Note that AMSE depends on other factors like pβ . We have suppressed the notations for simplicity. The
performance of LQLS as defined above is affected by the tuning parameter λ. In this paper, we consider the
value of λ that gives the minimum AMSE. Let λ∗,q denote the value of λ that minimizes AMSE given in (2):1
λ∗,q = arg min AMSE(λ, q, σw )
(3)
λ∈[0,∞)
Then LQLS is solved with this specific value of λ, i.e.,
1
β̂(λ∗,q , q) ∈ arg min ky − Xβk22 + λ∗,q kβkqq .
β 2
(4)
This is the best performance that LQLS with each value of q can achieve in terms of AMSE. We can use
Corollary 1 in Weng et al. (2016) to obtain the precise formula of this optimal AMSE.
Theorem 1. For a given q ∈ [1, 2]. Consider a converging sequence {β(p), X(p), w(p)}. Suppose that
β̂(λ∗,q , q) is the solution of LQLS with optimal tuning λ∗,q defined in (4) and AMSE(λ∗,q , q, σw ) is its AMSE
defined in (2). Then
AMSE(λ∗,q , q, σw ) = min EB,Z (ηq (B + σ̄Z; χ) − B)2 ,
(5)
χ≥0
where B and Z are two independent random variables with distributions pβ and N (0, 1), respectively; ηq is
the proximal operator for the function k · kqq ; 2 and σ̄ is the unique solution of the following equation:
σ̄ 2 = σω2 +
1
min EB,Z [(ηq (B + σ̄Z; χ) − B)2 ].
δ χ≥0
(6)
Theorem 1 provides the first step in our analysis of LQLS. We first calculate σ̄ from (6). Then incorporating
the solution σ̄ into (5) gives us the result of AMSE(λ∗,q , q, σw ). Given the distribution of B, the variance of
2 , the number of observations (normalized by the number of predictors) δ, and q ∈ [1, 2], it is
the error σw
straightforward to write a computer program to (numerically) find the solution of (6) and calculate the value
of AMSE(λ∗,q , q, σw ). However, it is needless to say that this approach does not shed much light on the
performance of different LQLS estimates, since there are many factors involved in the computation and each
affects the result in a non-trivial fashion. In this paper, we perform an analytical study on the solution of (6)
and obtain an explicit characterization of AMSE in the high signal-to-noise ratio regime. The expressions
we derived will offer us an accurate view of LQLS and quantify the impact of the distribution of β on the
performance of different LQLS.
As is clear from the definition of a converging sequence in 1 and Theorem 1, the main property of β that
affects AMSE is the probability measure pβ . In the rest of the paper, we will assume pβ does not have any point
mass at zero. We use the notation B to denote a one dimensional random variable distributed according to pβ .
We present our findings for the high signal-to-noise ratio regime where the noise level σw is either zero or small
in Sections 3.1 and 3.2, respectively. We then discuss the implications of our analysis framework for classical
asymptotics in Section 3.3.
1
It turns out the optimal value of λ can be estimated accurately under the asymptotic settings discussed in this paper. See
Mousavi et al. (2017) for more information.
2
The proximal operator of k · kqq is defined as ηq (u; χ) , arg minz 21 (u − z)2 + χ|z|q . For further information on these functions,
please refer to Lemma 6 in Section 4.1.
5
3 Our main contributions
3.1 Phase transition
Suppose that there is no noise in the linear model, i.e., σw = 0. Our first goal in the phase transition analysis
is to find the minimum value of δ for which AMSE(λ∗,q , q, 0) = 0. Our next theorem characterizes the phase
transition.
Theorem 2. Let q ∈ [1, 2]. If E|B|2 < ∞, then we have
AMSE(λ∗,q , q, 0) =
(
>0
if δ < 1,
=0
if δ > 1.
The result can also be derived from several different frameworks including the statistical dimension framework in Amelunxen et al. (2014). But we derive it as a simple byproduct of our results in Section 3.2. So we do
not discuss its proof here. This result is not surprising. Since, none of the coefficients is zero, the exact recovery
is impossible if n < p. Also, note that when δ > 1 even the ordinary least squares is capable of recovering β.
Hence, the result of phase transition analysis does not provide any additional information on the performance
of different regularizers. It is not even capable of showing the advantage of regularization techniques over the
standard least squares algorithm. This is due to the fact that the result of Theorem 2 holds only in the noiseless
case. Intuitively speaking, in the practical settings where the existence of the measurement noise is inevitable,
we expect different LQLS to behave differently. For instance, even though the signal under study is not sparse,
when pβ has a large mass around zero (it is approximately sparse), we expect the sparsity promoting LASSO
to offer better performance than the other LQLS. However, the distribution pβ does not have any effect on the
phase transition diagram. Motivated by these concerns, in the next section, we investigate the performance of
LQLS in the noisy setting, and study their noise sensitivity when the noise level σw is small. The new analysis
will offer more informative answers.
3.2 Second-order noise sensitivity analysis of AMSE
As an immediate generalization of the phase transition analysis, we can study the performance of different
estimators in the presence of a small amount of noise. More formally, we derive the asymptotic expansion of
AMSE(λ∗,q , q, σw ) for every q ∈ [1, 2], when σw → 0. As will be discussed later, this generalization of phase
transitions presents a more delicate analysis of LQLS. We start with the study of AMSE for the ordinary least
squares (OLS). The result of OLS will be later used for comparison purposes.
Lemma 1. Consider the region δ > 1. For the OLS estimate β̂(0, q), we have
AMSE(0, q, σw ) =
2
σw
.
1 − 1/δ
We prove the above lemma in Section 4.2. Note that the proof we presented there, has not used the independence of the noise elements that is often assumed in the analysis of OLS. Now we can discuss LQLS with the
optimal choice of λ defined in (3). It turns out that the distribution pβ impacts AMSE(λ∗,q , q, σw ) in a subtle
way. For analysis purposes we first study the signals whose elements are bounded away from zero in Theorem
3 and then study other distributions in Theorem 4.
6
Theorem 3. Consider the region δ > 1. Suppose that P(|B| > µ) = 1 with µ being a positive constant and
E|B|2 < ∞. Then, for q ∈ (1, 2], as σw → 0
AMSE(λ∗,q , q, σw ) =
and for q = 1, as σw → 0
2
δ3 (q − 1)2 (E|B|q−2 )2 4
σw
4
−
σw + o(σw
),
1 − 1/δ
(δ − 1)3 E|B|2q−2
2
AMSE(λ∗,q , q, σw ) =
where µ̃ is any positive number smaller than µ.
µ̃ (δ−1)
2
σw
−
− |o(e δσw2 )|,
1 − 1/δ
The proof can be found in Section 4.3. We observe that the first dominant term in the expansion of
2 /(1 − 1/δ).
AMSE(λ∗,q , q, σw ) is exactly the same for all values of q, including q = 1 and is equal to σw
This is also the same as the AMSE of the OLS. We may consider this term as the ‘phase transition’ term, since
it will go to zero only when δ > 1. In a nutshell, the first term in the expansion provides the phase transition
information. However, we are able to derive the second order term for AMSE(λ∗,q , q, σw ). This term gives us
what is beyond phase transition analysis. The impact of the signal distribution pβ and the regularizer ℓq , that is
omitted in PT diagram, is revealed in the second order term. As a result, to compare the performance of LQLS
with different values of q in the low noise regime, we can compare their second order terms.
First note that all the regularizers that are studied in Theorem 3 improve the performance of OLS. When the
distribution of the coefficients is bounded away from 0, no significant gain is obtained from LASSO since the
second dominant term in the expansion of AMSE is exponentially small. However, the rate of the second order
term exhibits an interesting transition from exponential to a polynomial decay when q increases from 1. In
fact, it seems that bridge regularizers with q > 1 offer more substantial improvements over OLS. Even though
LASSO is suboptimal, it is not clear which value of q provides the best performance here. Among other LQLS
with q ∈ (1, 2], the optimality is determined by the constant involved in the second order term (they all have
the same orders). To simplify our discussions, define
Cq =
(q − 1)2 (E|B|q−2 )2
,
E|B|2q−2
q ∗ = arg max Cq .
1<q≤2
Then LQLS with q = q ∗ will perform the best. To provide some insights on q ∗ , we focus on a special family of
distributions.
Lemma 2. Consider the two-point mixture |B| ∼ α∆µ1 + (1 − α)∆µ2 , where ∆µ denotes the probability
measure putting mass 1 at µ, 0 < µ1 ≤ µ2 , α ∈ (0, 1). Then q ∗ = 2 when µ1 = µ2 , and q ∗ → 1 as
µ2 /µ1 → ∞.
∗
Proof. When µ1 = µ2 , it is clear that Cq = (q − 1)2 µ−2
1 and thus q = 2. We now consider 0 < µ1 < µ2 .
Denote κ = µ2 /µ1 . We can then write Cq as:
Cq =
(q − 1)2 (αµ1q−2 + (1 − α)µ2q−2 )2
αµ12q−2 + (1 − α)µ22q−2
=
(q − 1)2 (α + (1 − α)κq−2 )2
.
µ21 (α + (1 − α)κ2q−2 )
Define q̄ = 1 + log1 κ . We would like to show that for any ǫ > 0, Cq̄ > max1+ǫ≤q≤2 Cq for κ large enough. That
will give us q ∗ ∈ (1, 1 + ǫ) and hence finishes the proof. To show that, note that for any q ∈ [1 + ǫ, 2], κ ≥ 1,
Cq̄
Cq
(α + (1 − α)κq̄−2 )2
(α + (1 − α)κq−2 )2
α + (1 − α)κ2ǫ
≥ α2 ·
α + (1 − α)e2
=
α + (1 − α)κ2q−2 (q̄ − 1)2
·
α + (1 − α)κ2q̄−2 (q − 1)2
α2 (1 − α)
· (q̄ − 1)2 ≥
· κ2ǫ (log κ)−2 → ∞.
α + (1 − α)e2
·
7
Therefore, Cq̄ > max1+ǫ≤q≤2 Cq when κ is sufficiently large.
Lemma 2 implies that ridge (q = 2) regularizer is optimal when the two point mixture components coincide,
and the optimal value of q will shift towards 1 as the ratio of the two points goes off to infinity. Intuitively
speaking, one would expect ridge to penalize large signals more aggressively than q < 2. Hence, in cases the
signal has a large dynamic range, ridge penalizes the large signal values more and is not expected to outperform
other values of q. Note that for the two-point mixture signals, the optimal value of q can be arbitrarily close to
1, however LASSO can never be optimal because its second order term is exponentially small.
Theorem 3 only studied distributions pβ that are bounded away from zero. We next study a more informative
and practical case where the distribution of β has more mass around zero.
Theorem 4. Consider the region δ > 1 and assume E|B|2 < ∞. For any given q ∈ (1, 2), suppose that
P(|B| ≤ t) = O(t2−q+ǫ ) (as t → 0) with ǫ being any positive constant, then as σw → 0
AMSE(λ∗,q , q, σw ) =
2
δ3 (q − 1)2 (E|B|q−2 )2 4
σw
4
−
σw + o(σw
).
1 − 1/δ
(δ − 1)3 E|B|2q−2
For q = 2, as σw → 0
AMSE(λ∗,2 , 2, σw ) =
2
σw
δ3
4
−
σ 4 + o(σw
).
1 − 1/δ (δ − 1)3 E|B|2 w
For q = 1, suppose that P(|B| ≤ t) = Θ(tℓ ) with ℓ > 0, then as σw → 0
2ℓ+2
−|Θ(σw
)|
·
1
log log . . . log
{z
} σw
|
m times
ℓ
. AMSE(λ∗,1 , 1, σw ) −
2
σw
2ℓ+2
. −|Θ(σw
)|,
1 − 1/δ
where m can be any natural number.
The proof is presented in Section 4.4. Before discussing the implications of this result, let us mention a few
points about the conditions that are imposed in this theorem. Note that the condition P(|B| ≤ t) = O(t2−q+ǫ )
for q ∈ (1, 2) is necessary otherwise E|B|q−2 appearing in the second order term will be unbounded. Intuitively
speaking, for every q ∈ (1, 2) even though P(|B| ≤ t) = O(t2−q+ǫ ), the probability density function (pdf) of
|B| can still go to infinity at zero. However, the condition requires that it should not go to infinity too fast. Now,
we would like to explain some of the interesting implications of this theorem.
1. Compared to the results of Theorem 3, we see that the expansion of AMSE(λ∗,q , q, σw ) for q ∈ (1, 2]
in Theorem 4 remains the same for more general B, while the rate of the second order term for LASSO
changes to polynomial from exponential. That means LASSO is more sensitive to the distribution of β
than other LQLS.
2. The second order term of LASSO becomes smaller as ℓ decreases. Note that when ℓ decreases the mass
of the distribution around zero increases. Hence, Theorem 4 implies that LASSO performs better when
the probability mass of the signal concentrates more around zero. This can be well explained by the
sparsity promoting feature of LASSO.
3. As in the case P(|B| > µ) = 1, the first dominant term is the same for all q ∈ [1, 2]. Hence we have
to compare their second order term. For any given q ∈ (1, 2], suppose P(|B| ≤ t) = Θ(t2−q+ǫ ), then
4 , while that of LASSO is Θ(σ 6−2q+2ǫ ) (ignore the
the second term of AMSE(λ∗,q , q, σw ) is of order σw
w
8
logarithmic factor). Since both terms are negative, we can conclude LASSO performs better than LQLS
with that value of q when ǫ ∈ (0, q − 1), and performs worse when ǫ ∈ (q − 1, ∞). This observation has
an important implication. The behavior of the distribution of |B| around zero is the most important factor
in the comparison between LASSO and other LQLS. If the pdf of |B| is zero at zero, then we should not
use LASSO and when it goes to infinity LASSO performs better than LQLS with q > 1 (at least for those
values of q for which our theorem is applicable).
4. Regrading the case where the probability density function of |B| is finite and positive at zero, our calculations of LASSO are not sharp enough to give an accurate comparison between LASSO and other LQLS.
However, the comparison of LQLS for different values of q > 1 will shed more light on the performance
of different regularizers in this case. In the following we consider two of the most popular families of
distributions and present an accurate comparison among q ∈ (1, 2].
Lemma 3. Consider |B| with density function f (b) = ζ(τ, q0 )e−τ b 0 1(0 ≤ b < ∞), where q0 ∈ (0, 2], τ > 0
and ζ(τ, q0 ) is the normalization constant. Then the best bridge estimator is the one that uses q ∗ = max(1, q0 ).
q
Proof. A simple integration by parts yields, for q ∈ (1, 2]
Z ∞
Z ∞
τ q0
τ q0 E|B|q+q0−2
q0
q−2 −τ bq0
q−2
db =
ζ(τ, q0 )b e
E|B|
=
.
ζ(τ, q0 )bq+q0 −2 e−τ b db =
q−1 0
q−1
0
q+q0 −2 )2
Hence Cq = τ 2 q02 (E|B|
E|B|2q−2
. We first consider q0 ∈ [1, 2], then
Cq = τ 2 q02
[E(|B|q−1 · |B|q0 −1 )]2 (a) 2 2
≤ τ q0 E|B|2q0 −2 = Cq0 ,
E|B|2q−2
where (a) is due to Cauchy-Schwarz inequality. So we obtain q ∗ = q0 . Regrading the case q0 ∈ (0, 1), let B ′
be an independent copy of B. Then for any q ∈ [1, 2]
E|B|q+q0 −2 − E|B|q−1 E|B|q0 −1 =
1
E(|B|q−1 − |B ′ |q−1 )(|B|q0 −1 − |B ′ |q0 −1 ) ≤ 0.
2
As a result, we can derive
Cq ≤ τ 2 q02
(E|B|q+q0 −2 )2
≤ τ 2 q02 E(|B|q0 −1 )2 = C1 .
(E|B|q−1 )2
We can then conclude q ∗ = 1 = max(1, q0 ).
Note that Lemma 3 studies a family of distributions whose probability density function exists and is nonzero at zero, but exhbit very different tail behaviors. As confirmed by this lemma, in this case the tail behavior
has an influence on the performance of LQLS. In particular, LQLS with q = q0 ∈ (1, 2] is optimal for distriq
butions with the exponential decay tail e−τ b 0 . Since β̂(λ, q0 ) can be considered as the maximum a posterior
estimate (MAP), our result suggests that MAP offers the best performance in the low noise regime (among the
bridge estimators). This is in general not true. See Zheng et al. (2017) for a counterexample in large noise
cases. It is also interesting to observe that as the tail becomes heavier than that of Laplacian distribution, the
optimal q ∗ approaches 1. Again, this observation is consistent with the fact that ridge often penalizes large
signal values more aggressively than the other estimators. Hence, if the tail of the distribution is light (like
Gaussian distributions), then ridge offers the best performance, otherwise, other values of q offer better results.
In the next lemma, we further support this claim by considering a special family of distributions with light tails.
9
Lemma 4. Consider |B| follows a uniform distribution with density function f (b) =
θ > 0 is the location parameter. Then q ∗ = 2.
1
θ
1(0 ≤ b ≤ θ), where
Proof. It is clear that
q−2
E|B|
1
=
θ
Hence, q ∗ = arg max1<q≤2
Z
θ
b
0
q−2
θ q−2
,
db =
q−1
(q−1)2 (E|B|q−2 )2
E|B|2q−2
2q−2
E|B|
= arg max1<q≤2
1
=
θ
Z
2q−1
θ2
= 2.
θ
b2q−2 db =
0
θ 2q−2
.
2q − 1
3.3 Implications for classical asymptotics
In this section, we would like to show that the results we have derived in the previous sections, has, perhaps
surprisingly, some interesting connections to the classical asymptotic setting in which n → ∞, while p is fixed.
Our analysis so far has been focused on the high-dimensional setting in which n/p → δ ∈ (0, ∞). Furthermore,
we assumed that the noise variance is small. In the classical asymptotics, it is assumed that the signal-to-noise
ratio of each observation is fixed and n/p → ∞. Note that having more measurements is at the intuitive level
equivalent to less noise. Hence, we expect our low-noise sensitivity to have some implications for the classical
asymptotics too. Our goal below is to formalize this connection and explain the implications of our low-noise
analysis framework for the classical asymptotics.
Towards that goal, we will consider the scenarios where the sample size n is much larger than the dimension
p. Analytically, we let δ go to infinity and calculate the expansions for AMSE in terms of large δ (similar to
what we did in Section 3.2 for low noise). In this section, we write AMSE(λ∗,q , q, δ) for AMSE(λ∗,q , q, σw )
to make it clear that the expansion is derived in terms of δ. Before getting to the results, we should clarify an
important issue. Recall the definition of a converging sequence in Definition 1. It is straightforward to confirm
2
that the signal-to-noise ratio of each measurement is SNR ∝ E|B|
2 . Hence if we take δ → ∞, SNR of each
δσw
measurement will go to zero and this is inconsistent with the classical asymptotic setting where the SNR is in
general assumed to be fixed. To fix this inconsistency, we will scale the noise term and consider a scaled linear
model as follows,
w
(7)
y = Xβ + √ ,
δ
where {X, β, w} is the converging sequence specified in Definition 1. With the SNR remained a positive
constant, this model is well aligned with the classical setting. Again for comparison purposes we start with the
ordinary least squares estimate.
Lemma 5. Consider the model (7) and OLS estimate β̂(0, q). Then as δ → ∞,
AMSE(0, q, δ) =
2
σ2
σw
+ w
+ o(δ−2 ).
δ
δ2
Proof. This lemma is a simple application of Lemma 1. Under model (7), Lemma 1 shows that AMSE(0, q, δ) =
2
σw
δ−1 . As δ → ∞, the expansion can be easily verified.
We now discuss the bridge estimators with q ∈ [1, 2].
Theorem 5. Consider the model (7). Suppose that P(|B| > µ) = 1 with µ being a positive constant and
E|B|2 < ∞. Then for q ∈ [1, 2], as δ → ∞,
AMSE(λ∗,q , q, δ) =
2
2
σw
σ 2 E|B|2q−2 − (q − 1)2 (E|B|q−2 )2 σw
+ w
·
+ o(δ−2 ).
δ
δ2
E|B|2q−2
10
The proof can be found in Section 4.5. Since both Theorems 3 and 5 are concerned with signals that are
bounded away from zero, we can compare their results. Again all of the LQLS have the same first dominant
term. However, in the large sample regime, the second order term of LASSO is at the same order as that of
other LQLS. Interestingly, the comparison of the constant in the second order term is consistent with that in
the low noise case. Hence we obtain the same conclusions for two-point mixture distributions. For instance,
bridge with q ∈ (1, 2] outperforms OLS and q = 2 is optimal when all the mass is concentrated at one point.
See Lemma 2 for more information on the comparison of Cq .
We now discuss the implications of Theorem 5 for classical asymptotics. In the classical setting where
n → ∞ and p is fixed, the performance of LQLS has been studied in Knight and Fu (2000). In particular the
√
LQLS estimates were shown to have the regular n convergence. In our setting, we first let n/p → δ and then
δ → ∞. If we apply Theorem 2 in Knight and Fu (2000) to (7), a straightforward calculation for the asymptotic
variance will give us the first dominant term in AMSE(λ∗,q , q, δ). In other words, the classical asymptotic result
for LQLS in Knight and Fu (2000) only provides the “first-order” information regarding mean square error, and
it is the same for all the values of q ∈ [1, 2] under optimal tuning. The virtue of our asymptotic framework is
to offer the second order term that can be used to evaluate and compare LQLS more accurately. Similar results
can be derived when signals have mass around zero, as presented in the next theorem.
Theorem 6. Consider the model introduced in (7) and assume E|B|2 < ∞. For any given q ∈ (1, 2), suppose
that P(|B| ≤ t) = O(t2−q+ǫ ) (as t → 0) with ǫ being any positive constant, then as δ → ∞,
AMSE(λ∗,q , q, δ) =
2
2
σ 2 E|B|2q−2 − (q − 1)2 (E|B|q−2 )2 σw
σw
+ w
·
+ o(δ−2 ),
δ
δ2
E|B|2q−2
for q = 2, as δ → ∞,
AMSE(λ∗,q , q, δ) =
2
2
σ 2 E|B|2 − σw
σw
+ w
·
+ o(δ−2 ),
2
2
δ
δ
E|B|
and for q = 1, suppose P(|B| ≤ t) = Θ(tℓ ) with 0 < ℓ < 1, then as δ → ∞,
√ ℓ
σ2
−|Θ(δ−ℓ−1 )| · log log . . . log( δ) . AMSE(λ∗,q , q, δ) − w . −|Θ(δ−ℓ−1 )|,
|
{z
}
δ
m times
where m can be any natural number.
The proof is presented in Section 4.5. Theorem 6 can be compared with Theorem 4. Again we see that the
expansion for q ∈ (1, 2] remains the same for more general signals, while the second order term of LASSO
becomes order-wise smaller when signals put more mass around zero. For a given q ∈ (1, 2], it is clear that
LASSO outperforms this LQLS when P(|B| ≤ t) = Θ(t2−q+ǫ ) with ǫ ∈ (0, q−1). This implies that even in the
case when n is much larger than p, if the underlying signal has many elements of small values, ℓ1 regularization
will improve the performance, which is characterized by a second order analysis that is not available from
√
q
the n convergence result. Regrading the distributions with tail e−τ b 0 , we see that the comparison among
q ∈ (1, 2] in the low noise regime carries over.
The fact that regularization can improve the performance of the maximum likelihood estimate (i.e., OLS in
the context of linear regression with Gaussian noise), seems to be contradictory with the classical results that
imply MLE is asymptotically optimal under mild regularity conditions. However, note that the optimality of
MLE is concerned with the asymptotic variance (equivalently the first order term) of the estimate. Our results
show that many estimators share that first order term, while their actual performance might be different. Second
dominant terms provide much more accurate information in these cases.
11
4 Proofs of our main results
4.1 Notations and Preliminaries
Throughout the proofs, B will be a random variable having the probability measure pβ that appears in the
definition of the converging sequence, and Z will refer to a standard normal random variable. We will also use
φ(·) to denote the density function of Z and F (b) to represent the cumulative distribution function of |B|. We
further define the following useful notations:
Rq (χ, σ) = E(ηq (B/σ + Z; χ) − B/σ)2 ,
χ∗q (σ) = arg min Rq (χ, σ),
χ≥0
(8)
where B and Z are independent. Recall the proximal operator function ηq (u; χ). Since we will be using
ηq (u; χ) extensively in the later proofs, we present some useful properties of ηq (u; χ) in the next lemma.
Because ηq (u; χ) has explicit forms when q = 1, 2, we focus on the case 1 < q < 2. For notational simplicity
we may use ∂i f (x1 , x2 , . . .) to represent the partial derivative of f with respect to its ith argument.
Lemma 6. For q ∈ (1, 2), the function ηq (u; χ) satisfies the following properties.
(i) −ηq (u; χ) = ηq (−u; χ).
(ii) u = ηq (u; χ) + χq(q − 1)ηq (u; χ)sign(u).
(iii) αηq (u; χ) = ηq (αu; α2−q χ), for α > 0.
(iv)
∂ηq (u;χ)
∂u
=
1
1+χq(q−1)|ηq (u;χ)|q−2
(v)
∂ηq (u;χ)
∂χ
=
−q|ηq (u;χ)|q−1 sign(u)
1+χq(q−1)|ηq (u;χ)|q−2
(vi) The function ∂2 ηq (u; χ) is differentiable with respect to u.
Proof. Please refer to Lemmas 7, 8 and 10 in Weng et al. (2016) for the proofs.
We next write down the Stein’s lemma (Stein, 1981) that we will apply several times in the proofs.
Stein’s lemma. Suppose the function f : R → R is weakly differentiable and E|f ′ (Z)| < ∞, then
E(Zf (Z)) = Ef ′ (Z).
4.2 Proof of Lemma 1
Since δ > 1, β̂(0, q) = (X ′ X)−1 Xy is well defined with probability 1 for sufficiently large n. We first derive
AMSE(λ, 2, σw ) for the Ridge estimate β̂(λ, 2) = (X ′ X + λI)−1 X ′ y, and then obtain the AMSE for OLS by
letting λ → 0. According to Theorem 2.1 in Weng et al. (2016), it is known that for given λ > 0,
2
AMSE(λ, 2, σw ) = δ(σ 2 − σw
),
where σ is the solution to the following equation:
2
σ 2 = σw
+
4χ2 E|B|2 + σ 2
,
δ(1 + 2χ)2
λ=χ−
χ
.
δ(1 + 2χ)
After a few calculations we can obtain
AMSE(λ, 2, σw ) =
2)
δ(4χ2 E|B|2 + σw
,
δ(1 + 2χ)2 − 1
12
(9)
√
1−δ+2λδ+ (δ−1−2λδ)2 +8λδ2
with χ =
. Clearly AMSE(λ, 2, σw ) →
4δ
to derive AMSE for OLS. According to the identity below
2
σw
1−1/δ
as λ → 0. We now utilize that result
(X ′ X + λI)−1 = (X ′ X)−1 − λ (X ′ X)−1 (I + λ(X ′ X)−1 )−1 (X ′ X)−1 ,
{z
}
|
H
we have
2
2
1
σw
1
σw
2
1
kβ̂(0, q) − βk22 −
= kβ̂(λ, 2) − βk22 −
+ kλHX ′ yk22 + hβ̂(λ, 2) − β, λHX ′ yi (10)
p
1 − 1/δ
p
1 − 1/δ p
p
{z
} |
{z
}
|
{z
} |
J2
J1
J3
Let σmin (X) be the smallest non-zero singular values of X. It is not hard to confirm that
kHX ′ Y k2 ≤ kHX ′ Xβk2 + kHX ′ wk2 ≤
a.s.
Since σmin → 1 −
√1
δ
kβk2
kwk2
+
2
2
2 (X) .
λ + σmin (X) (λ + σmin (X))σmin
> 0 Bai and Yin (1993) and β, w belong to the converging sequence defined in 1, we
can conclude that J2 = O(λ2 ), a.s.. Moreover, we obtain from (9) that almost surely
J1 =
2
2)
σw
δ(4χ2 E|B|2 + σw
−
δ(1 + 2χ)2 − 1
1 − 1/δ
The results on J1 , J2 imply that J3 = O(λ), a.s.. Further note that the term on the left hand side of (10) does
not depend on λ. Therefore by letting n → ∞ and then λ → 0 on both sides of (10) finishes the proof.
4.3 Proof of Theorem 3
4.3.1
Roadmap
Since the proof has several long steps, we lay out the roadmap to help readers navigate through the details.
According to Lemma 6 part (iii) and Theorem 1, we know
AMSE(λ∗,q , q, σw ) = σ̄ 2 Rq (χ∗q (σ̄), σ̄),
(11)
where σ̄ is the unique solution of
2
σ̄ 2 = σw
+
σ̄ 2
Rq (χ∗q (σ̄), σ̄).
δ
(12)
Note from the above equation that σ̄ is a function of σw . In the regime σw → 0, we will show σ̄ → 0. This
fact combined with (11) tells us that in order to derive the second-order expansion of AMSE(λ∗,q , q, σw ) as
a function of σw , it is sufficient to characterize the convergence rate of σ̄ as σw → 0 and Rq (χ∗q (σ), σ) as
σ → 0. For that purpose, we will first study the convergence rate of χ∗q (σ) as σ → 0, which will then enables
us to obtain the convergence rate of Rq (χ∗q (σ), σ). We then utilize that result and (12) to derive the rate of σ̄ as
σw → 0. We give the proof for 1 < q ≤ 2 and q = 1 in Sections 4.3.2 and 4.3.3, respectively.
4.3.2
Proof for the case 1 < q ≤ 2
Due to the explicit form of η2 (u; χ) =
thus focus the proof on 1 < q < 2.
u
1+2χ ,
all the results for q = 2 in this section can be easily verified. We
13
Lemma 7. Let χ∗q (σ) be the optimal threshold value as defined in (8). Then χ∗q (σ) → 0 as σ → 0.
Proof. The proof is essentially the same as the one for Lemma 17 in Weng et al. (2016). Hence we do not
repeat the arguments here.
Lemma 8. For q ∈ (1, 2], suppose that P(|B| > µ) = 1 with µ being a positive constant and E|B|2 < ∞.
Then as σ → 0
Rq (Cσ q , σ) = 1 + (C 2 q 2 E|B|2q−2 − 2Cq(q − 1)E|B|q−2 )σ 2 + o(σ 2 ),
where C is any fixed positive constant.
Proof. We aim to derive the convergence rate of Rq (χ, σ) when χ = Cσ q . In this proof, we may write χ
to denote Cσ q for notational simplicity. According to Lemma 6 parts (ii)(iv) and Stein’s lemma, we have the
following formula for Rq (χ, σ):
Rq (χ, σ) − 1 = E(ηq (B/σ + Z; χ) − B/σ − Z)2 + 2EZ(ηq (B/σ + Z; χ) − B/σ − Z)
|ηq (B/σ + Z; χ)|q−2
.
= χ2 q 2 E|ηq (B/σ + Z; χ)|2q−2 −2χq(q − 1)E
{z
}
1 + χq(q − 1)|ηq (B/σ + Z; χ)|q−2
|
{z
}
|
S1
(13)
S2
It is straightforward to confirm the following
χ2 q 2 E|ηq (B/σ + Z; χ)|2q−2
σ→0
σ2
2 2
= C q lim E|ηq (B + σZ; χσ 2−q )|2q−2 = C 2 q 2 E|B|2q−2 .
S1
σ→0 σ 2
lim
=
lim
σ→0
(14)
The last equality is obtained by Dominated Convergence Theorem (DCT). The condition of DCT holds due to
Lemma 6 part (ii). We now focus on analyzing S2 . We obtain
−S2
σ2
|ηq (B/σ + Z; χ)|q−2
1 + χq(q − 1)|ηq (B/σ + Z; χ)|q−2
1
= 2Cσ q−2 q(q − 1)E
|ηq (|B|/σ + Z; χ)|2−q + χq(q − 1)
Z ∞ Z −b/σ+µ/(2σ)
1
= 2Cσ q−2 q(q − 1)
φ(z)dzdF (b)
2−q + χq(q − 1)
µ
−b/σ−µ/(2σ) |ηq (b/σ + z; χ)|
|
{z
}
= 2Cσ q−2 q(q − 1)E
T1
+ 2Cσ q−2 q(q − 1)
|
Z
∞Z
µ
1
φ(z)dzdF (b) .
2−q + χq(q − 1)
R\[−b/σ−µ/(2σ),−b/σ+µ/(2σ)] |ηq (b/σ + z; χ)|
{z
}
T2
We then consider T1 and T2 separately. For T1 , we have
T1 ≤ 2Cσ
≤ 2σ
q−2
−3
q(q − 1)
Z
∞ Z −µ/σ+µ/(2σ)
µ
−µ/σ−µ/(2σ)
µφ(µ/(2σ)) → 0, as σ → 0.
14
1
φ(z)dzdF (b)
χq(q − 1)
(15)
Regarding T2 , DCT enables us to conclude
lim T2 =
lim 2Cσ q−2 q(q − 1)E
1(Z ∈/ [−|B|/σ − µ/(2σ), −|B|/σ + µ/(2σ)])
|ηq (|B|/σ + Z; χ)|2−q + χq(q − 1)
1(Z ∈/ [−|B|/σ − µ/(2σ), −|B|/σ + µ/(2σ)])
= lim 2Cq(q − 1)E
σ→0
|ηq (|B| + σZ; χσ 2−q )|2−q + Cσ 2 q(q − 1)
σ→0
σ→0
= 2Cq(q − 1)E|B|q−2 .
(16)
Note that DCT works here because for small enough σ, Lemma 6 parts (iv)(v) implies
1(Z ∈/ [−|B|/σ − µ/(2σ), −|B|/σ + µ/(2σ)])
|ηq (|B| +
σZ; χσ 2−q )|2−q
+
Cσ 2 q(q
− 1)
≤
1
1
≤
.
2−q
2−q
|ηq (µ/2; χσ )|
|ηq (µ/2; 1)|2−q
Combining (13), (14), (15) and (16) together completes the proof.
Lemma 8 shows that by choosing an appropriate χ for σ small enough, R(χ, σ) is less than 1. This result
will be used to show that χ∗q (σ) cannot converge to zero too fast. We then utilize this fact to derive the exact
convergence rate of χ∗q (σ). This is done in the next lemma.
Lemma 9. Suppose that P(|B| > µ) = 1 with µ being a positive constant and E|B|2 < ∞, then for q ∈ (1, 2]
we have as σ → 0
(q − 1)E|B|q−2 q
σ + o(σ q ),
qE|B|2q−2
(q − 1)2 (E|B|q−2 )2 2
Rq (χ∗q (σ), σ) = 1 −
σ + o(σ 2 ).
E|B|2q−2
χ∗q (σ) =
Proof. Choosing χ =
(q−1)E|B|q−2
qE|B|2q−2
· σ q in Lemma 8, we have
Rq (χ, σ) − 1
(q − 1)2 (E|B|q−2 )2
=
−
< 0.
σ→0
σ2
E|B|2q−2
lim
(17)
That means for sufficiently small σ
Rq (χ∗q (σ), σ) ≤ Rq (χ, σ) < 1 = Rq (0, σ).
Hence we can conclude that χ∗q (σ) > 0 when σ is small enough. Moreover, by a slight change of arguments in
the proof of Lemma 8 summarized below:
1.
2.
3.
4.
the fact χσ 2−q = o(1) used several times in Lemma 8 still holds here
χσ 2−q = o(1) and χ = o(σ q ) are sufficient to have S1 = o(σ 2 )
bounding the term T1 in (15) does not depend on χ
χσ 2−q = o(1) and χ = o(σ q ) are sufficient to obtain T2 = o(1)
we can show
lim
σ→0
Rq (χ, σ) − 1
= 0,
σ2
(18)
for χ = O(exp(−c/σ)) with any fixed positive constant c. This implies that limσ→0 χ∗q (σ) · ec/σ = +∞ for
any c > 0. Otherwise there exists a sequence σn → 0 such that χq (σn )ec/σn = O(1). This result combined
with (17) and (18) contradicts with the fact that χ = χ∗q (σ) is the minimizer of Rq (χ, σ). We will use the two
15
aforementioned properties of χ∗q (σ) we have showed so far in the following proof. For notational simplicity, in
the rest of the proof we may use χ to denote χ∗q (σ) whenever no confusion is caused. Firstly since χ∗q (σ) is a
non-zero finite value, it is a solution of the first order optimality condition
written out as
0
=
(a)
=
(b)
=
∂Rq (χ,σ)
∂χ
= 0, which can be further
E((ηq (B/σ + Z; χ) − B/σ)∂2 ηq (B/σ + Z; χ))
−(ηq (B/σ + Z; χ) − B/σ − Z)q|ηq (B/σ + Z; χ)|q−1 sign(B/σ + Z)
E
1 + χq(q − 1)|ηq (B/σ + Z; χ)|q−2
+E(Z∂2 ηq (B/σ + Z; χ))
q 2 |ηq (B/σ + Z; χ)|2q−2
q(q − 1)|ηq (B/σ + Z; χ)|4−2q
χE
−
E
1 + χq(q − 1)|ηq (B/σ + Z; χ)|q−2
(|ηq (B/σ + Z; χ)|2−q + χq(q − 1))3
|
{z
} |
{z
}
U1
U2
q 2 (q
Z; χ)|2−q
− 1)|ηq (B/σ +
.
−χ E
(|ηq (B/σ + Z; χ)|2−q + χq(q − 1))3
{z
}
|
(19)
U3
We have used Lemma 6 part (v) to derive (a). To obtain (b), we have used the following steps:
1. We used Lemma 6 part (ii) to conclude that
ηq (B/σ + Z; χ) − B/σ − Z = −χq|ηq (B/σ + Z; χ)|q−1 sign(B/σ + Z).
2. We used the expression we derived in Lemma 6 part (v) for ∂2 ηq (B/σ + Z; χ) and then employed Stein’s
lemma to simplify E(Z∂2 ηq (B/σ + Z; χ)). Note that according to Lemma 6 part (vi), ∂2 ηq (B/σ + Z; χ)
is differentiable with respect to its first argument and hence Stein’s lemma can be applied.
We now evaluate the three terms U1 , U2 and U3 individually. Our goal is to show the following:
(i) limσ→0 σ 2q−2 U1 = q 2 E|B|2q−2 .
(ii) limσ→0 σ q−2 U2 = q(q − 1)E|B|q−2 .
(iii) limσ→0 σ 2q−4 U3 = q 2 (q − 1)E|B|2q−4 .
For the term U1 , we can apply Dominated Convergence Theorem (DCT)
q 2 |ηq (B + σZ; χσ 2−q )|2q−2
= q 2 E|B|2q−2 .
σ→0 1 + χσ 2−q q(q − 1)|ηq (B + σZ; χσ 2−q )|q−2
lim σ 2q−2 U1 = E lim
σ→0
We now derive the convergence rate of U2 . We have
Z ∞Z ∞
q(q − 1)|ηq (b/σ + z; χ)|4−2q
U2 =
φ(z)dzdF (b)
2−q + χq(q − 1))3
µ
−∞ (|ηq (b/σ + z; χ)|
Z ∞ Z −b+ µ
σ
2σ
q(q − 1)|ηq (b/σ + z; χ)|4−2q
φ(z)dzdF (b)
=
µ
(|ηq (b/σ + z; χ)|2−q + χq(q − 1))3
− σb − 2σ
µ
|
{z
}
U21
+
Z
∞Z
µ
|
µ
µ
,− σb + 2σ
]
z ∈[−
/ σb − 2σ
q(q − 1)|ηq (b/σ + z; χ)|4−2q
φ(z)dzdF (b) .
(|ηq (b/σ + z; χ)|2−q + χq(q − 1))3
{z
}
U22
16
(20)
First note that
σ
q−2
≤
U21 ≤ σ
q−2
Z
µ
∞ Z − σb + 2σ
µ
− σb − 2σ
µ
q(q − 1)(µ/(2σ))4−2q
φ(z)dzdF (b)
(χq(q − 1))3
µ5−2q φ(µ/(2σ))
→ 0, as σ → 0,
− 1)2
σ 7−3q 24−2q χ3 q 2 (q
(21)
where the last step is due to the fact that limσ→0 χec/σ = +∞. To evaluate U22 we first derive the following
bounds for small enough σ
µ
µ
1(z ∈/ [− σb − 2σ
, − σb + 2σ
]) · q(q − 1)|ηq (b + σz; χσ 2−q )|4−2q
(|ηq (b + σz; χσ 2−q )|2−q + χσ 2−q q(q − 1))3
q(q − 1)
q(q − 1)
≤
≤
.
2−q
2−q
|ηq (µ/2; χσ )|
|ηq (µ/2; 1)|2−q
Hence we are able to apply DCT to obtain
lim σ q−2 U22 = q(q − 1)E|B|q−2 .
σ→0
(22)
Combining (20), (21), and (22) proves the result (ii). We can use similar arguments to show result (iii). Finally,
we utilize the convergence results for U1 , U2 , U3 and Equation (19) to derive
χ
limσ→0 σ q−2 U2
(q − 1)E|B|q−2
=
lim
=
.
σ→0 σ q
σ→0 limσ→0 σ 2q−2 U2 − limσ→0 σ 2q−2 U3
qE|B|2q−2
lim
Now since we know the exact convergence order of χ∗q (σ), (17) shows the exact order of Rq (χ∗q (σ), σ).
We are in position to derive the second-order expansion of AMSE(λ∗,q , q, σw ) as σw → 0 for q ∈ (1, 2].
2) ≤
According to Equation (12) and the fact that χ = χ∗q (σ̄) minimizes Rq (χ, χ∗q (σ̄)), it is clear that δ(σ̄ 2 − σw
2
2
σ̄ Rq (0, σ̄) = σ̄ , which combined with the condition δ > 1 implies σ̄ → 0 as σw → 0. This result further
enables us to conclude from (12):
δ
σ̄ 2
,
=
2
σw →0 σw
δ−1
lim
(23)
where we have used Rq (χ∗q (σ̄), σ̄) → 1 from Lemma 9. We finally utilize Lemma 9, Equations (11), (12) and
(23) to derive the expansion of AMSE(λ∗,q , q, σw ) in the following way:
2
δ
1
σw
−4
σ̄ 2 Rq (χ∗q (σ̄), σ̄) −
= σw
(σ̄ 2 − σ̄ 2 Rq (χ∗q (σ̄), σ̄))
1 − 1/δ
δ−1
δ
∗
4
3
2
q−2
2
δ
σ̄ Rq (χq (σ̄), σ̄) − 1
−δ (q − 1) (E|B| )
=
· 4 ·
→
.
2
δ − 1 σw
σ̄
(δ − 1)3 E|B|2q−2
−4
AMSE(λ∗,q , q, σw ) −
σw
This completes the proof of Theorem 3 for q ∈ (1, 2].
4.3.3
Proof for the case q = 1
Lemma 10. Suppose that P(|B| > µ) = 1 with µ being a positive constant and E|B|2 < ∞, then for q = 1 as
σ→0
χ∗q (σ) = O(φ(µ/σ)),
Rq (χ∗q (σ), σ) − 1 = O(φ2 (µ/σ)).
17
Proof. We first claim that χ∗q (σ) → 0 as σ → 0. Otherwise, there exists a sequence σn → 0 such that
χ∗q (σn ) → C > 0 as n → ∞. And the limit C is finite. Suppose this is not true, then since η1 (u; χ) =
sign(u)(|u| − χ)+ we can apply Fatou’s lemma to conclude
lim inf Rq (χ∗q (σn ), σn ) ≥ E lim inf (η1 (B/σn + Z; χ∗q (σn )) − B/σn )2 = +∞,
n→∞
n→∞
contradicting with the fact
Rq (χ∗q (σn ), σn )
≤ Rq (0, σn ) = 1. We now calculate the following limit:
lim Rq (χ∗q (σn ), σn ) = lim E(η1 (B/σn + Z; χ∗q (σn )) − B/σn − Z)2
n→∞
n→∞
+2 lim EZ(η1 (B/σn + Z; χ∗q (σn )) − B/σn − Z) + 1 = C 2 + 1.
n→∞
The last step is due to Dominated Convergence Theorem (DCT). The condition of DCT can be verified based on
the fact |u − η1 (u; χ)| ≤ χ. We can also choose a positive constant C̃ smaller than C and use similar argument
to obtain limn→∞ Rq (C̃, σn ) = C̃ 2 + 1. That means Rq (C̃, σn ) < Rq (χ∗q (σn ), σn ) when n is large enough.
This is contradicting with the fact χ = χ∗q (σn ) minimizes Rq (χ, σn ).
We next derive the following bounds:
Rq (χ, σ) − 1
=
(a)
=
(b)
≤
E(η1 (B/σ + Z; χ) − B/σ − Z)2 + 2E(Z(η1 (B/σ + Z; χ) − B/σ − Z))
E(η1 (B/σ + Z; χ) − B/σ − Z)2 + 2E(∂1 η1 (B/σ + Z; χ) − 1)
Z −B/σ+χ
(c)
2
φ(z)dz = χ2 − 4χEφ(−B/σ + αχ).
χ − 2E
−B/σ−χ
To obtain (a) we used Stein’s lemma; note that η1 (u; χ) is a weakly differentiable function of u. Inequality (b)
holds since |η1 (u; χ) − u| ≤ χ. Equality (c) is the result of the mean value theorem and hence |α| ≤ 1 is depen√
dent on B. From the above inequality, it is straightforward to verify that if we choose χ = 3e−1 Eφ( 2B/σ),
then
Rq (χ∗q (σ), σ) ≤ Rq (χ, σ) < 1 = Rq (0, σ),
(24)
for small enough σ. This means the optimal threshold χ∗q (σ) is a non-zero finite value. Hence it is a solution to
∂Rq (χ∗q (σ),σ)
∂χ
= 0, which further implies (from now on we use χ∗ to represent χ∗q (σ) for simplicity):
χ∗ =
2Eφ(χ∗ − |B|/σ)
2φ(χ∗ − µ/σ)
Eφ(χ∗ − B/σ) + Eφ(χ∗ + B/σ)
≤
≤
,
E1(|Z + B/σ| ≥ χ∗ )
E1(|Z + B/σ| ≥ χ∗ )
E1(|Z + B/σ| ≥ χ∗ )
(25)
where the last inequality holds for small values of σ due to the condition P(|B| > µ) = 1. Since E1(|Z +
√
∗ 2
B/σ| ≥ χ∗ ) → 1, as σ → 0 and φ(χ∗ − µ/σ) ≤ φ(µ/( 2σ))e(χ ) /2 , from (25) we can first conclude
χ∗ = o(σ), which in turn (use (25) again) implies χ∗ = O(φ(µ/σ)).
We now turn to analyzing Rq (χ∗ , σ):
Rq (χ∗ , σ) − 1
=
≥
(d)
≥
E(η1 (B/σ + Z; χ∗ ) − B/σ − Z)2 + 2E(∂1 η1 (B/σ + Z; χ∗ ) − 1)
Z −µ/σ+χ∗
∗
−2E1(|B/σ + Z| ≤ χ ) ≥ −2
φ(z)dz ≥ −4χ∗ φ(µ/σ − χ∗ )
−µ/σ−χ∗
−8φ2 (χ∗ − µ/σ) (e)
∼ −8φ2 (µ/σ),
E1(|Z + B/σ| ≥ χ∗ )
where (d) is due to (25) and (e) holds because E1(|Z + B/σ| ≥ χ∗ ) → 1 and χ∗ = o(σ). This result combined
with Rq (χ∗ , σ) − 1 < 0 from (24) finishes the proof.
18
We are in position to derive the expansion of AMSE(λ∗,1 , 1, σw ). Similarly as in the proof for q ∈ (1, 2],
we can use Lemma 10 to derive (23) for q = 1. Then we apply Lemma 10 again to obtain
δ
δ
σ 2 = σ̄ 2 Rq (χ∗q (σ̄), σ̄) −
(σ̄ 2 − σ̄ 2 Rq (σq∗ (σ̄), σ̄)/δ)
δ−1 w
δ−1
δσ̄ 2 (Rq (χ∗q (σ̄), σ̄) − 1)
2
= o(exp(−µ̄2 /σ̄ 2 )) = o(exp(−µ̃2 (δ − 1)/(δσw
))),
=
δ−1
AMSE(λ∗,1 , 1, σw ) −
where 0 < µ̃ < µ̄ < µ. This closes the proof.
4.4 Proof of Theorem 4
Similar to the proof of Theorem 3, we consider two cases, i.e. 1 < q ≤ 2 and q = 1, and prove them separately.
We will follow closely the roadmap illustrated in Section 4.3.1.
4.4.1
Proof for the case 1 < q < 2
Again all the results in this section can be proved easily for q = 2. We will only consider 1 < q < 2. Before
we start the proof of our main result, we mention a simple lemma that will be used multiple times in our proof.
Lemma 11. Let T (σ) and χ(σ) be two nonnegative sequences with the property: χ(σ)T q−2 (σ) → 0, as
σ → 0. Then,
ηq (T (σ), χ(σ))
= 1.
lim
σ→0
T (σ)
Proof. The proof is a simple application of scale invariance property of ηq , i.e, Lemma 6 part (iii). We have
ηq (T (σ), χ(σ))
= lim ηq (1; χ(σ)T q−2 (σ)) = 1,
σ→0
σ→0
T (σ)
lim
where the last step is the result of Lemma 6 part (ii).
Our first goal is to show that when χ = Cσ q , then limσ→0 R(χ,σ)−1
is a negative constant by choosing an
σ2
appropriate C. However, since this proof is long, we break it to several steps. These steps are summarized in
Lemmas 12, 13, and 14. Then in Lemma 15 we employ these three results to show that if χ = Cσ q , then
R(χ, σ) − 1
= C 2 q 2 E|B|2q−2 − 2Cq(q − 1)E|B|q−2 .
σ→0
σ2
lim
Lemma 12. For any given q ∈ (1, 2), suppose that P(|B| < t) = O(t2−q+ǫ ) (as t → 0) with ǫ being any
positive constant, E|B|2 < ∞ and χ = Cσ q , where C > 0 is a fixed number. Then we have
σ q−2
Z
0
∞Z
−b
+α
σ
−b
−α
σ
1
φ(z)dzdF (b) → 0,
|ηq (b/σ + z; χ)|2−q + χq(q − 1)
as σ → 0. Note that α is an arbitrary positive constant.
Proof. The main idea of the proof is to break this integral into several pieces and prove that each piece converges
to zero. Throughout the proof, we will choose ǫ small enough to be in (0, q − 1). Based on the value of q, we
consider the following intervals. First find the unique non-negative integer of m∗ such that
1
1
q ∈ [2 − (ǫ/(ǫ + q − 1)) m∗ +1 , 2 − (ǫ/(ǫ + q − 1)) m∗ ).
19
n (l) = lm + lm+1 + · · · + ln (m ≤ n). Now we define the following intervals:
Denote Sm
#
"
b
σ q−ǫ
σ q−ǫ
b
,
,− +
I−1 = − −
σ log( σ1 ) σ log( σ1 )
"
ǫ+q−1
ǫ+q−1
ǫ
ǫ #
(2−q)i − q−1
(2−q)i − q−1
b
b
σ q−1
σ q−1
Ii = − −
,− +
, 0 ≤ i ≤ m∗ ,
σ (log(1/σ))S0i (2−q) σ (log(1/σ))S0i (2−q)
#
"
1
1
b
b
,
,− +
Im∗ +1 = − −
∗
∗
σ (log(1/σ))S0m +1 (2−q) σ (log(1/σ))S0m +1 (2−q)
−b
b
Im∗ +2 =
− α, − + α .
σ
σ
(26)
We see that for small enough σ, these intervals are nested: I−1 ⊂ I0 ⊂ I1 ⊂ . . . ⊂ Im∗ +2 . Further define
Z ∞Z
1
q−2
φ(z)dzdF (b),
P−1 = σ
2−q + χq(q − 1)
0
I−1 |ηq (b/σ + z; χ)|
Z ∞Z
1
φ(z)dzdF (b), 0 ≤ i ≤ m∗ + 2.
Pi = σ q−2
2−q + χq(q − 1)
|η
(b/σ
+
z;
χ)|
q
0
Ii \Ii−1
Using these notations we have
σ
q−2
Z
0
∞Z
−b
+α
σ
−b
−α
σ
∗
m
+2
X
1
Pi .
φ(z)dzdF
(b)
=
|ηq (b/σ + z; χ)|2−q + χq(q − 1)
(27)
i=−1
Our goal is to show that Pi → 0 as σ → 0. Since these intervals have different forms, we consider five different
cases (i) i = −1, (ii) i = 0, (iii) 1 ≤ i ≤ m∗ , (iv) i = m∗ + 1, and (v) i = m∗ + 2 and for each case we show
that Pi → 0. Let |I| denote the Lebesgue measure of an interval I. For the first term, we have for a positive
constant C̃−1 ,
Z C̃−1 σ√log(1/σ) Z
Z ∞Z
1
1
q−2
q−2
φ(z)dzdF (b) ≤ σ
φ(z)dzdF (b)
P−1 ≤ σ
χq(q
−
1)
χq(q
− 1)
0
I−1
0
I−1
Z ∞
Z
1
q−2
+σ
φ(z)dzdF (b)
√
C̃−1 σ log(1/σ) I−1 χq(q − 1)
p
p
σq−ǫ
q−2
σ q−2 φ(0)|I−1 |P(|B| ≤ C̃−1 σ log(1/σ)) σ φ(C̃−1 log(1/σ) − log(1/σ) )|I−1 |
+
≤
χq(q − 1)
χq(q − 1)
p
p
σ q−ǫ−2 φ( C̃2−1 log(1/σ))
σ q−ǫ−2 P(|B| ≤ C̃−1 σ log(1/σ))
≤ O(1)
+ O(1)
log(1/σ)
log(1/σ)
2
≤ O(1)(log(1/σ))
−q+ǫ
2
σ q−ǫ−2+C̃−1 /8
+ O(1)
→ 0,
log(1/σ)
(28)
where we have used the condition P(|B| < t) = O(t2−q+ǫ ) to obtain the last inequality and the last statement
20
holds by choosing C̃−1 large enough. We next analyze the term P0 . For a constant C̃0 > 0 we have
Z ∞Z
1
q−2
φ(z)dzdF (b)
P0 ≤ σ
|η
(b/σ
+
z; χ)|2−q
q
0
I0 \I−1
Z C̃0 σ√log(1/σ) Z
1
q−2
φ(z)dzdF (b)
= σ
|η
(b/σ
+
z; χ)|2−q
q
0
I0 \I−1
Z ∞
Z
1
φ(z)dzdF (b)
+σ q−2
√
2−q
C̃0 σ log(1/σ) I0 \I−1 |ηq (b/σ + z; χ)|
p
p
σ
q−2 φ C̃
q−2
σ
log(1/σ)
−
0
σ φ(0)|I0 |P(|B| < C̃0 σ log(1/σ))
log(1/σ) |I0 |
≤
+
.
q−ǫ
q−ǫ
2−q
2−q
σ
σ
;
χ
;
χ
ηq
η
q
log(1/σ)
log(1/σ)
We have used the fact that |b/σ + z| ≥
since
σq−ǫ
( log(1/σ)
)q−2 χ
∝
σq−ǫ
log(1/σ)
(29)
for z ∈
/ I−1 in the last step. Note that according to Lemma 11,
2
σ q −(1+ǫ)q+2ǫ (log(1/σ))2−q
→ 0, we obtain
σq−ǫ
log(1/σ)
lim
q−ǫ
σ→0 ηq ( σ
log(1/σ) ; χ)
= 1.
With the above result, it is clear that the second term of the upper bound in (29) vanishes if choosing sufficiently
large C̃0 . Regarding the first term we know
p
ǫ+4−3q
σ q−2 |I0 |P(|B| < C̃0 σ log(1/σ))
q 2 −(ǫ+2)q+3ǫ+1
2
= o(1).
∝
σ
(log(1/σ))
q−ǫ
2−q
σ
;
χ
ηq
log(1/σ)
Now we consider an arbitrary 1 ≤ i ≤ m∗ and show that Pi → 0. Similarly as bounding P0 we can have
Z C̃i σ√log(1/σ) Z
1
Pi ≤ σ q−2
φ(z)dzdF (b)
2−q
0
Ii \Ii−1 |ηq (b/σ + z; χ)|
Z ∞
Z
1
q−2
+σ
φ(z)dzdF (b)
√
2−q
C̃i σ log(1/σ) Ii \Ii−1 |ηq (b/σ + z; χ)|
ǫ+q−1
p
ǫ
(2−q)i − q−1
σ q−1
q−2 φ C̃
p
|Ii |
log(1/σ)
−
σ
i
i
σ q−2 φ(0)|Ii |P(|B| < C̃i σ log(1/σ))
(log(1/σ))S0 (2−q)
ǫ+q−1
ǫ+q−1
+
.
(30)
≤
i−1 − ǫ
i−1 − ǫ
q−1 (2−q)
q−1
q−1
2−q σ q−1 (2−q)
ηq2−q σ
η
;
χ
;
χ
q
S i−1 (2−q)
S i−1 (2−q)
(log(1/σ))
(log(1/σ))
0
0
We then use Lemma 11 to conclude for i ≥ 1
σ
lim
σ→0
ηq2−q
ǫ+q−1
i ǫ(2−q)
q−1 (2−q) − q−1
i
(log(1/σ))S1 (2−q)
ǫ+q−1
i−1 − ǫ
q−1 (2−q)
q−1
i−1
(log(1/σ))S0 (2−q)
σ
;χ
= 1.
(31)
The condition of Lemma 11 can be verified in the following:
(σ
ǫ+q−1
ǫ
(2−q)i−1 − q−1
q−1
i−1
(log(1/σ))−S0
(2−q) q−2
)
χ ∝ σ−
ǫ+q−1
(2−q)i + 2−q
ǫ+q
q−1
q−1
i
(log(1/σ))S1 (2−q) = o(1),
2−q
ǫ+q−1
2−q
i
where the last step is due to the fact that − ǫ+q−1
q−1 (2−q) + q−1 ǫ+q ≥ − q−1 (2−q)+ q−1 ǫ+q = 2q −2 > 0.
Using the result (31), it is straightforward to confirm that if C̃i is chosen large enough, the second term in (30)
21
goes to zero. For the first term,
p
σ q−2 φ(0)|Ii |P(|B| < C̃i σ log(1/σ))
ǫ+q−1
lim
i−1 − ǫ
σ→0
q−1
2−q σ q−1 (2−q)
ηq
;χ
S i−1 (2−q)
(log(1/σ))
(a)
=
σ
O(1) · lim
0
ǫ+q−1
ǫ
(2−q)i − q−1
q−2 σ q−1
i
(log(1/σ))S0 (2−q)
σ 2−q+ǫ (log(1/σ))
= O(1) · lim (log(1/σ))
ǫ(2−q)
ǫ+q−1
(2−q)i − q−1
σ q−1
i
(log(1/σ))S1 (2−q)
σ→0
2−q+ǫ
2
−q+ǫ
2
σ→0
= 0,
P ∗
where we have used (31) to obtain (a). So far we have showed limσ→0 m
i=−1 Pi = 0. Our next step is to
prove that Pm∗ +1 → 0.
Z C̃m∗ +1 σ√log(1/σ) Z
φ(z)
dzdF (b)
Pm∗ +1 ≤ σ q−2
2−q
0
Im∗ +1 \Im∗ |ηq (b/σ + z; χ)|
Z ∞
Z
φ(z)
q−2
+σ
dzdF (b)
√
2−q
C̃m∗ +1 σ log(1/σ) Im∗ +1 \Im∗ |ηq (b/σ + z; χ)|
p
σ q−2 φ(0)|Im∗ +1 |P(|B| < C̃m∗ +1 σ log(1/σ))
ǫ+q−1
≤
ǫ
m∗
2−q σ q−1 (2−q) − q−1
ηq
;χ
m∗
(log(1/σ))S0 (2−q)
p
1
q−2
∗
|Im∗ +1 |
σ φ C̃m +1 log(1/σ) −
m∗ +1 (2−q)
log(1/σ)S0
ǫ+q−1
.
(32)
+
ǫ
m∗
2−q σ q−1 (2−q) − q−1
;χ
ηq
∗
S m (2−q)
(log(1/σ))
0
Again based on (31) It is clear that if C̃m∗ +1 is large enough, then the second term in (32) goes to zero. We
now show the first term goes to zero as well:
p
σ q−2 φ(0)|Im∗ +1 |P(|B| < C̃m∗ +1 σ log(1/σ))
ǫ+q−1
lim
ǫ
m∗
σ→0
2−q σ q−1 (2−q) − q−1
;χ
ηq
∗
S m (2−q)
(log(1/σ))
σ q−2
(b)
= O(1) · lim
1
0
m∗ +1 (2−q)
log(1/σ)S0
σ→0
σ
= O(1) · lim σ
σ 2−q+ǫ (log(1/σ))
2−q+ǫ
2
ǫ+q−1
m∗ +1 − ǫ(2−q)
q−1 (2−q)
q−1
m∗ +1 (2−q)
(log(1/σ))S1
∗
ǫ−(ǫ+q−1)(2−q)m +1
q−1
σ→0
(log(1/σ))
−q+ǫ
2
(c)
= 0,
∗ +1
where (b) holds from Lemma 11 and (c) is due to the condition we imposed on m∗ that ensures (2 − q)m
ǫ
∗
∗
ǫ+q−1 . The last remaining term of (27) is Pm +2 . To prove Pm +2 → 0, we have
P
m∗ +2
≤ σ
q−2
Z
C̃m∗ +2 σ
√
log(1/σ)
0
+σ q−2
Z
∞
C̃m∗ +2 σ
√
Z
log(1/σ)
I
∗ +2 \Im∗ +1
Zm
Im∗ +2 \Im∗ +1
22
1
φ(z)dzdF (b)
|ηq (b/σ + z; χ)|2−q
1
φ(z)dzdF (b).
|ηq (b/σ + z; χ)|2−q
≤
By using the same strategy as we did for bounding Pi (0 ≤ i ≤ m∗ + 1), the second integral above will go to
zero as σ → 0, when C̃m∗ +2 is chosen large enough. And the first integral can be bounded by
p
m∗ +2
σ q−2 φ(0)2αP(|B| ≤ C̃m∗ +2 σ log(1/σ)) (d)
(2−q)
= O(1)σ ǫ log(1/σ)(2−q+ǫ)/2+S1
→ 0,
2−q
1
ηq
;χ
∗
S m +1 (2−q)
log(1/σ)
0
where (d) holds by Lemma 11 and the condition of Lemma 11 can be easily checked. This completes the
proof.
Define
α
b
α
b
I , − − 1−γ , − + 1−γ .
σ σ
σ σ
γ
In Lemma 12 we proved that:
Z ∞Z
σ q−2
0
−b
+α
σ
−b
−α
σ
(33)
1
φ(z)dzdF (b) → 0.
|ηq (b/σ + z; χ)|2−q + χq(q − 1)
In the next lemma, we would like to extend this result and show that in fact,
Z ∞Z
1
q−2
φ(z)dzdF (b) → 0.
σ
2−q + χq(q − 1)
|η
(b/σ
+
z;
χ)|
γ
q
0
I
Lemma 13. For any given q ∈ (1, 2), suppose the conditions in Lemma 12 hold. Then for any fixed 0 < γ < 1,
Z ∞Z
1
q−2
σ
φ(z)dzdF (b) → 0,
2−q + χq(q − 1)
|η
(b/σ
+
z;
χ)|
γ
q
0
I \Im∗ +2
as σ → 0. Note that Im∗ +2 is defined in (26).
Proof. As in the proof of Lemma 12, we break the integral into smaller subintervals and prove each one goes
to zero. Consider the following intervals:
b
b
α
α
Ji = − −
,− +
,
ǫ
ǫ
S0i (1−ǫ)
S0i (1−ǫ)
σ σ 1+θ
σ σ 1+θ
where θ > 0 is an arbitrarily small number and i is an arbitrary natural number. Note that {Ji } is a sequence
of nested intervals and Im∗ +2 ⊂ J0 . Our goal is to show that the following integrals go to zero as σ → 0:
Z ∞Z
1
Q−1 , σ q−2
φ(z)dzdF (b) → 0,
2−q + χq(q − 1)
0
J0 \Im∗ +2 |ηq (b/σ + z; χ)|
Z ∞Z
1
φ(z)dzdF (b) → 0, i ≥ 0.
Qi , σ q−2
2−q + χq(q − 1)
0
Ji+1 \Ji |ηq (b/σ + z; χ)|
Define σ̃i ,
ǫ
1 1+θ S0i (1−ǫ)
.
ασ
Q−1
Since |b/σ + z| ≥ α for z ∈
/ Im∗ +2 we obtain
Z ∞Z
1
≤ σ q−2
φ(z)dzdF (b)
2−q
0
J0 \Im∗ +2 |ηq (α; χ)|
Z σ log(1/σ) Z
σ̃0
1
φ(z)dzdF (b)
= σ q−2
2−q
0
J0 \Im∗ +2 |ηq (α; χ)|
Z ∞
Z
1
q−2
+σ
φ(z)dzdF (b)
2−q
σ
log(1/σ) J0 \Im∗ +2 |ηq (α; χ)|
σ̃0
≤ σ
q−2
Z
σ
σ̃0
0
log(1/σ)
− σ̃10 )|J0 |
φ( log(1/σ)
1
σ̃0
q−2
dF
(b)
+
σ
.
|ηq (α; χ)|2−q
|ηq (α; χ)|2−q
23
It is straightforward to notice that the second term above converges to zero. For the first term, by the condition
P(|B| < t) = O(t2−q+ǫ ) we derive the following bounds
σ
q−2
Z
σ
σ̃0
log(1/σ)
0
= O(1)σ
1
dF (b) ≤ O(1)σ q−2 (σσ̃0−1 log(1/σ))2−q+ǫ
|ηq (α; χ)|2−q
ǫ(q−1−ǫ+θ)
1+θ
(log(1/σ))2−q+ǫ → 0.
Now we discuss the term Qi for i ≥ 0. Similarly as we bounded Q−1 we have
Qi ≤ σ
q−2
Z
σ
σ̃i+1
log(1/σ)
0
+σ
q−2
Z
∞
σ
σ̃i+1
Z
log(1/σ)
1
φ(z)dzdF (b)
1
2−q
Ji+1 \Ji |ηq ( σ̃i ; χ)|
Z
1
φ(z)dzdF (b).
1
2−q
Ji+1 \Ji |ηq ( σ̃i ; χ)|
(34)
The second integral in (34) can be easily shown convergent to zero as σ → 0. We now focus on the first integral.
σ q−2
≤
≤
Z
σ
σ̃i+1
log(1/σ)
0
σ q−2
Z
Ji+1 \Ji
1
φ(z)dzdF (b)
|ηq ( σ̃1i ; χ)|2−q
σ
log(1/σ)
P |B| ≤
σ̃i+1
|ηq ( σ̃1i ; χ)|2−q
σ ǫ (log(1/σ))2−q+ǫ σ̃i2−q
σ ǫ (log(1/σ))2−q+ǫ (a)
O(1)
2−q+ǫ = O(1)
2−q+ǫ
|ηq ( σ̃1i ; χ)|2−q σ̃i+1
σ̃i+1
= O(1)σ
ǫ(θ+(q−1−ǫ)(1−ǫ)i+1 )
1+θ
(log(1/σ))2−q+ǫ = o(1).
We have used Lemma 11 to obtain (a). Above all we have showed that for any given natural number i ≥ 0,
Z ∞Z
1
q−2
φ(z)dzdF (b) = 0.
lim σ
2−q + χq(q − 1)
σ→0
0
Ji \Im∗ +2 |ηq (b/σ + z; χ)|
ǫ
Now note that as i goes to infinity, the exponent of σ in the interval Ji goes to 1+θ
(1+(1−ǫ)+(1−ǫ)2 +. . .) =
1
γ
1+θ . So, by choosing small enough θ and sufficiently large i we can make I ⊂ Ji , hence completing the
proof.
In the last two lemmas, we have been able to prove that for χ = Cσ q ,
Z ∞Z
1
q−2
σ
φ(z)dzdF (b) → 0.
2−q + χq(q − 1)
0
I γ |ηq (b/σ + z; χ)|
This result will be used to characterize the following limit
Z ∞Z
1
q−2
lim σ
φ(z)dzdF (b).
2−q + χq(q − 1)
σ→0
|η
(b/σ
+
z;
χ)|
0
R q
Before that we mention a simple lemma that will be applied several times in our proofs.
Lemma 14. For 1 < q < 2 we have
|ηq
2
1
≤
.
2−q
+ χq(q − 1)
|u| (q − 1)
(u; χ)|2−q
24
Proof. It is sufficient to consider u > 0. We analyze two different cases:
1
1. χ ≤ u2−q 2q
: According to Lemma 6 part (ii), since we know ηq (u; χ) ≤ u, we have
ηq (u; χ) = u − χqηqq−1 (u; χ) ≥ u − χquq−1 ≥ u − u2−q
Hence,
1 q−1 u
qα
= .
2q
2
1
22−q
2
1
≤
≤
≤
.
2−q
2−q
2−q
|ηq (u; χ)|
+ χq(q − 1)
|ηq (u; χ)|
u
(q − 1)u2−q
1
2. χ ≥ u2−q 2q
:
1
2
1
≤
≤
.
|ηq (u; χ)|2−q + χq(q − 1)
χq(q − 1)
(q − 1)u2−q
This completes our proof.
Now we can consider one of the main results of this section.
Lemma 15. For any given q ∈ (1, 2), suppose the conditions in Lemma 12 hold. Then for χ = Cσ q we have
lim
σ→0
Rq (χ, σ) − 1
= C 2 q 2 E|B|2q−2 − 2Cq(q − 1)E|B|q−2 .
σ2
Proof. We follow the same roadmap as in the proof of Lemma 8. Recall that
|ηq (B/σ + Z; χ)|q−2
.
Rq (χ, σ) − 1 = χ2 q 2 E|ηq (B/σ + Z; χ)|2q−2 −2χq(q − 1)E
{z
}
1 + χq(q − 1)|ηq (B/σ + Z; χ)|q−2
|
{z
}
|
S1
(35)
S2
The first term S1 can be calculated in the same way as in the proof of Lemma 8.
lim σ −2 S1 = C 2 q 2 E|B|2q−2 .
(36)
σ→0
We now focus on analyzing S2 . First note that restricting |B| to be bounded away from 0 makes it possible to
follow the same arguments used in the proof of Lemma 8 to obtain,
lim E
σ→0
1(|B| > 1)
= E|B|q−2 1(|B| > 1).
|ηq (|B| + σZ; Cσ 2 )|2−q + Cσ 2 q(q − 1)
(37)
Hence we next consider the event |B| ≤ 1.
1(|B| ≤ 1)
E
|ηq (|B| + σZ; Cσ 2 )|2−q + Cσ 2 q(q − 1)
1 Z −b/σ+bc /(2σ)
1
φ(z)dzdF (b)
=
2
2−q
+ Cσ 2 q(q − 1)
0
−b/σ−bc /(2σ) |ηq (b + σz; Cσ )|
|
{z
}
Z
T1
+
Z
0
|
1Z
1
R\[−b/σ−bc /(2σ),−b/σ+bc /(2σ)]
σz; Cσ 2 )|2−q
|ηq (b +
{z
T2
25
+ Cσ 2 q(q − 1)
φ(z)dzdF (b),
}
where c > 1 is a constant that we will specify later. We first analyze T2 . Note that,
T2 = E
1(|B + σZ| ≥ |B|c /2, |B| ≤ 1)
|ηq (B + σZ; Cσ 2 )|2−q + Cσ 2 q(q − 1)
,
and
1(|B + σZ| ≥ |B|c /2, |B| ≤ 1)
|ηq (B +
σZ; Cσ 2 )|2−q
+
Cσ 2 q(q
− 1)
(a)
≤
|B|c(q−2)
21(|B + σZ| ≥ |B|c /2)
,
≤ q−3
2−q
(q − 1)|B + σZ|
2 (q − 1)
where (a) is due to Lemma 14. For any 1 < q < 2, it is straightforward to verify that E|B|c(q−2) < ∞ if c is
chosen close enough to 1. We can then apply Dominated Convergence Theorem (DCT) to obtain
lim T2 = E1(|B| ≥ |B|c /2, |B| ≤ 1)|B|q−2 = E|B|q−2 1(|B| ≤ 1).
(38)
σ→0
We now turn to bounding T1 . According to Lemmas 12 and 13, we know
Z ∞Z
1
φ(z)dzdF (b) → 0,
σ q−2
2−q + χq(q − 1)
|η
(b/σ
+
z;
χ)|
γ
q
0
I
c
c
c
c
α
α
b
b
b
b
, − σb + σ1−γ
]. Define Icγ = [− σb − σ1−γ
, − σb + σ1−γ
] and Ĩ c = [− σb − 2σ
, − σb + 2σ
].
where I γ = [− σb − σ1−γ
γ
γ
For 0 ≤ b ≤ 1, we get Ic ⊆ I for any given α > 1. Therefore,
Z 1Z
1
q−2
T3 , σ
φ(z)dzdF (b) → 0.
γ |η (b/σ + z; χ)|2−q + χq(q − 1)
q
0
Ic
Hence to bound T1 , it is sufficient to bound T1 − T3 :
Z 1Z
1
q−2
φ(z)dzdF (b)
T1 − T3 = σ
2−q + χq(q − 1)
γ
0
Ĩ c \Ic |ηq (b/σ + z; χ)|
Z 1Z
1
q−2
φ(z)dzdF (b)
≤ σ
c
1−γ
γ
; χ)|2−q + χq(q − 1)
0
Ĩ c \Ic |ηq (b /σ
Z 1Z
(b)
2bc(q−2)
≤ σ q−2+(1−γ)(2−q)
φ(z)dzdF (b)
0
Ĩ c \Icγ q − 1
Z C̃σ√log(1/σ) Z
2bc(q−2)
φ(z)dzdF (b)
= σ q−2+(1−γ)(2−q)
0
Ĩ c \Icγ q − 1
|
{z
}
T4
+σ
|
q−2+(1−γ)(2−q)
Z
1
C̃σ
√
log(1/σ)
Z
{z
T5
Ĩ c \Icγ
2bc(q−2)
φ(z)dzdF (b),
q−1
}
where (b) is the result of Lemma 14 and C̃ is a positive constant. We first bound T5 in the following:
Z
p
2σ q−2+(1−γ)(2−q) 1
b
bc(q−1)
2σ q−3+(1−γ)(2−q)
T5 ≤
φ
φ(C̃ log(1/σ)/2).
dF (b) ≤
√
q−1
σ
2σ
q−1
C̃σ log(1/σ)
It is then easily seen that T5 goes to zero by choosing large enough C̃. For the remaining term T4 ,
√
Z
b
2σ q−3+(1−γ)(2−q) C̃σ log(1/σ) c(q−1)
b
φ
dF (b)
T4 ≤
q−1
2σ
0
p
p
2σ q−3+(1−γ)(2−q)
(C̃σ log(1/σ))c(q−1) φ(0)P(|B| ≤ C̃σ log(1/σ))
≤
q−1
≤ O(1)σ c(q−1)−γ(2−q)+1−q+ǫ (log(1/σ))(c(q−1)+2−q+ǫ)/2 → 0.
26
To obtain the last statement, we can choose γ close enough to zero and c close to 1. Hence we can conclude
T1 → 0 as σ → 0. This combined with the results in (37) and (38) gives us
lim −σ −2 S2 = 2Cq(q − 1)E
σ→0
1
|ηq (|B| +
σZ; Cσ 2 )|2−q
+ Cσ 2 q(q − 1)
= 2Cq(q − 1)E|B|q−2 .
The above result together with (36) finishes the proof.
As stated in the roadmap of the proof, our first goal is to characterize the convergence rate of χ∗q (σ).
Towards this goal, we first show that χ∗q (σ) cannot be either too large or too small. In particular, in Lemmas
16 and 17, we show that χ∗q (σ) = O(σ q−1 ) and χ∗q (σ) = Ω(σ q ). We then utilize such result in Lemma 18 to
conclude that χ∗q (σ) = Θ(σ q ).
Lemma 16. Suppose E|B|2 < ∞, if χσ 1−q = ∞ and χ = o(1), then Rq (χ, σ) → ∞, as σ → 0.
Proof. Consider the formula of Rq (χ, σ) in (35). Since χ = o(1), it is straightforward to apply Dominated
Convergence Theorem to obtain
lim χ−2 σ 2q−2 S1 = q 2 E|B|2q−2 .
σ→0
Because χ2 σ 2−2q → ∞, we know S1 → ∞. Also note
|S2 | ≤ 2χq(q − 1) ·
1
= 2.
χq(q − 1)
Hence, Rq (χ, σ) → ∞.
Lemma 17. Suppose that the same conditions for B in Lemma 15 hold, if χ = o(σ q ), then
Rq (χ, σ) − 1
→ 0, as σ → 0.
σ2
Proof. Consider the expression of Rq (χ, σ) − 1 in (35). First note that
χ2 σ 2−2q q 2 E|ηq (B + σZ; χσ 2−q )|2q−2
S1
=
lim
= 0.
σ→0
σ→0 σ 2
σ2
h
i
σq−ǫ
σq−ǫ
b
Now we study the behavior of S2 . Recall that we defined I−1 = − σb − log(
and I γ =
1 , −σ +
1
log( σ )
)
σ
b
α
α
in (26) and (33), respectively. It is straightforward to use the same argument as for
, − σb + σ1−γ
− σ − σ1−γ
bounding P−1 in the proof of Lemma 12 (see the derivations in (28)) to have
Z Z
Z Z
χ ∞
φ(z)
φ(z)
χ ∞
dzdF (b) ≤ 2
dzdF (b) → 0.
2−q + χq(q − 1)
σ2 0
|η
(b/σ
+
z;
χ)|
σ
χq(q
− 1)
q
I−1
0
I−1
lim
Moreover, since χ < Cσ q for small enough σ, Lemma 6 part (v) implies
|ηq (b/σ + z; χ)| ≥ |ηq (b/σ + z; Cσ q )|.
Therefore, as σ → 0
Z
∞Z
φ(z)
dzdF (b)
2−q
0
I γ \I−1 |ηq (b/σ + z; χ)|
Z ∞Z
χ
φ(z)
1
≤ q · 2−q
dzdF (b) → 0,
q 2−q
σ σ
0
I γ \I−1 |ηq (b/σ + z; Cσ )|
χ
σ2
27
R∞R
φ(z)
1
where the last statement holds because of σ2−q
0
I γ \I−1 |ηq (b/σ+z;Cσq )|2−q dzdF (b) → 0 that has already
been shown in the proof of Lemmas 12 and 13. Above all we have proved
Z Z
χ ∞
φ(z)
dzdF (b) → 0.
2
2−q + χq(q − 1)
σ 0
I γ |ηq (b/σ + z; χ)|
Based on the above result, we can easily follow the same derivations of bounding the term T1 in the proof of
Lemma 15 to conclude
Z Z
c
φ(z)
χ 1 −b/σ+b /(2σ)
dzdF (b) = 0.
(39)
lim 2
2−q + χq(q − 1)
σ→0 σ
|η
(b/σ
+
z;
χ)|
c
q
0
−b/σ−b /(2σ)
Furthermore, because χ = o(σ q ), the analyses to derive Equation (37) and bound T2 in the proof of Lemma 15
can be adapted here and yield
Z Z
φ(z)
χ ∞ +∞
dzdF (b) = 0,
(40)
lim 2
2−q + χq(q − 1)
σ→0 σ
1
−∞ |ηq (b/σ + z; χ)|
Z Z
φ(z)
χ 1
dzdF (b) = 0.
(41)
lim 2
2−q + χq(q − 1)
σ→0 σ
0
R\[−b/σ−bc /(2σ),−b/σ+bc /(2σ)] |ηq (b/σ + z; χ)|
Putting results (39), (40) and (41) together gives us
Z Z
2χq(q − 1) ∞ ∞
φ(z)
−S2
dzdF (b) = 0.
lim 2 = lim
2
2−q + χq(q − 1)
σ→0
σ→0 σ
σ
0
−∞ |ηq (b/σ + z; χ)|
This finishes the proof.
Collecting the results from Lemmas 15, 16 and 17, we can upper and lower bound the optimal threshold
value χ∗q (σ) as shown in the following corollary.
Corollary 7. Suppose the conditions for B in Lemma 15 hold. Then as σ → 0, we have
χ∗q (σ) = Ω(σ q ),
χ∗q (σ) = O(σ q−1 ).
Proof. Since χ = χ∗q (σ) minimizes Rq (χ, σ), we know
Rq (χ∗q (σ), σ) ≤ Rq (0, σ) = 1, for any σ > 0,
σ
−1
(Rq (χ∗q (σ), σ)
− 1) ≤ σ
−2
q
(Rq (Cσ , σ) − 1) < −c, for small enough σ,
(42)
(43)
where the last inequality is due to Lemma 15 with an appropriate choice of C, and c is a positive constant.
Note that we already know χ∗q (σ) = o(1). If χ∗q (σ) 6= O(σ q−1 ), Lemma 16 will contradict with (42). If
χ∗q (σ) 6= Ω(σ q ), Lemma 17 will contradict with (43).
We are now able to derive the exact convergence rate of χ∗q (σ) and Rq (χ∗q (σ), σ).
Lemma 18. For any given q ∈ (1, 2), suppose the conditions in Lemma 12 for B hold. Then we have
(q − 1)E|B|q−2 q
σ + o(σ q ),
qE|B|2q−2
(q − 1)2 (E|B|q−2 )2 2
Rq (χ∗q (σ), σ) = 1 −
σ + o(σ 2 ).
E|B|2q−2
χ∗q (σ) =
28
Proof. In this proof, we use χ∗ to denote χ∗q (σ) for notational simplicity. Using the notations in Equation (19),
we know that χ∗ satisfies the following equation:
0 = χ∗ U1 − U2 − χ∗ U3 .
Our first goal is to show that σ q−2 U2 → q(q − 1)E|B|q−2 as σ → 0. Define the interval
K = [−b/σ − (χ∗ )1/(2−q) , −b/σ + (χ∗ )1/(2−q) ].
(44)
Then we have,
U2
σ 2−q
Z Z
|ηq (b/σ + z; χ∗ )|4−2q
q(q − 1) ∞
=
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q
0
K (|ηq (b/σ + z; χ )|
Z Z
|ηq (b/σ + z; χ∗ )|4−2q
q(q − 1) ∞
φ(z)dzdF (b).
+ 2−q
∗ 2−q + χ∗ q(q − 1))3
σ
0
R\K (|ηq (b/σ + z; χ )|
(45)
We first show that the first term in (45) goes to zero. Note that ηq ((χ∗ )1/(2−q) ; χ∗ ) = (χ∗ )1/(2−q) ηq (1; 1) by
Lemma 6 part (iii), we thus have
Z Z
|ηq (b/σ + z; χ∗ )|4−2q
q(q − 1) ∞
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q
0
K (|ηq (b/σ + z; χ )|
Z Z
|ηq ((χ∗ )1/(2−q) ; χ∗ )|4−2q
q(q − 1) ∞
φ(z)dzdF (b)
≤
σ 2−q
(χ∗ q(q − 1))3
0
K
Z ∞Z
1
ηq4−2q (1; 1)
≤
φ(z)dzdF (b)
∗
2
σ 2−q 0
K χ (q(q − 1))
Z C1 σ√log(1/σ) Z
ηq4−2q (1; 1)
1
≤
φ(z)dzdF (b)
∗
2
σ 2−q 0
K χ (q(q − 1))
p
ηq4−2q (1; 1)
|K|φ(C
log(1/σ) − (χ∗ )1/(2−q) ).
+ 2
1
q (q − 1)2 σ 2−q χ∗
Since we have already shown χ∗ = Ω(σ q ) in Corollary 7, it is straightforward to see that the second integral in
the above bound is negligible for large enough C1 . For the first term, we know
Z C1 σ√log(1/σ) Z
2−q+ǫ
1
φ(z)dzdF (b) ≤ O(1)(χ∗ )(q−1)/(2−q) σ ǫ (log(1/σ)) 2 = o(1).
∗
2−q
χ σ
0
K
Our next goal is to find the limit of the second term in (45). In order to do that, we again break the integral into
several pieces. Recall the intervals I γ , I−1 , I0 , I1 , . . . , J0 , J1 , . . . that we introduced in Lemmas 12 and 13.
We consider two different cases:
q−ǫ
σ
).
1. In this case, we assume that (χ∗ )1/(2−q) = o( log(1/σ)
29
Hence K ⊆ I−1 . We have
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
I−1 \K (|ηq (b/σ + z; χ )|
Z ∞Z
1
1
φ(z)dzdF (b)
≤
σ 2−q 0
|η
(b/σ
+
z; χ∗ )|2−q
q
I−1 \K
Z ∞Z
1
1
≤
φ(z)dzdF (b)
2−q
2−q
σ
(1; 1)
0
I−1 \K χ∗ ηq
Z C2 σ√log(1/σ) Z
1
1
≤
φ(z)dzdF (b)
2−q
∗
σ 2−q 0
(1; 1)
I−1 \K χ ηq
Z ∞
Z
1
1
φ(z)dzdF (b).
+ 2−q
√
2−q
σ
C2 σ log(1/σ) I−1 \K χ∗ ηq (1; 1)
The fact that χ∗ (σ) = Ω(σ q ) enables us to conclude that the second integral above goes to zero by
choosing large enough C2 . Regarding the first term we know
Z C2 σ√log(1/σ) Z
1
φ(z)dzdF (b)
χ∗ σ 2−q 0
I−1 \K
p
−q+ǫ (a)
φ(0)|I−1 |P(|B| ≤ C2 σ log(1/σ))
σq
2
=
O(1)
·
·
(log(1/σ))
= o(1),
≤
σ 2−q χ∗
χ∗
where (a) is due to χ∗ = Ω(σ q ). We now consider another integral.
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
I γ \I−1 (|ηq (b/σ + z; χ )|
Z ∞Z
1
1
≤
φ(z)dzdF (b).
2−q
∗ 2−q
σ
0
I γ \I−1 |ηq (b/σ + z; χ )|
Our goal is to show that this integral goes to zero as well. We use the following calculations:
Z ∞Z
1
1
φ(z)dzdF (b)
2−q
∗ 2−q
σ
0
I γ \I−1 |ηq (b/σ + z; χ )|
m∗ +2 Z ∞ Z
1 X
1
≤
φ(z)dzdF (b)
2−q
∗ 2−q
σ
0
Ii \Ii−1 |ηq (b/σ + z; χ )|
i=0
Z
ℓ Z
1
1 X ∞
+ 2−q
φ(z)dzdF (b)
∗ 2−q
σ
Ji \Ji−1 |ηq (b/σ + z; χ )|
i=1 0
Z ∞Z
1
1
φ(z)dzdF (b),
+ 2−q
∗ 2−q
σ
0
J0 \Im∗ +2 |ηq (b/σ + z; χ )|
where ℓ is chosen in a way such that I γ ⊆ Jℓ . Define mi = |Ii | and m̃i = |Ji |. Note that we
did similar calculations for the case χ = Cσ q in Lemmas 12 and 13. The key argument regarding χ
that we used there to show each term above converges to zero was that ηq (mi ; Cσ q ) = Θ(mi ) and
ηq (m̃i ; Cσ q ) = Θ(m̃i ). Hence, if we can show that ηq (mi ; χ∗ ) = Θ(mi ) and ηq (m̃i ; χ∗ ) = Θ(m̃i ) in
the current case, then those proofs will carry over and we will have
Z ∞Z
1
1
φ(z)dzdF (b) → 0.
2−q
∗ 2−q
σ
0
I γ \I−1 |ηq (b/σ + z; χ )|
30
For this purpose, we make use of Lemma 11. Note that since m−1 < m0 < m1 < . . . < mm∗+2 <
m̃0 < m̃1 < m̃2 . . . < m̃ℓ , we only need to confirm the condition of Lemma 11 for m−1 . We have
q−2
χ∗ m−1
= χ∗ σ (q−ǫ)(q−2) log(1/σ)2−q = o(1),
by the assumption of Case 1. Hence in the current case we have obtained
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
φ(z)dzdF (b) → 0.
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
I γ (|ηq (b/σ + z; χ )|
Furthermore, it is clear that
σ q−2 |ηq (b/σ + z; χ∗ )|4−2q
1
.
≤
∗
2−q
∗
3
∗
2−q
2−q
(|ηq (b/σ + z; χ )|
+ χ q(q − 1))
|ηq (b + σz; χ σ )|
+ χ∗ σ 2−q q(q − 1)
We can then follow the same line of arguments for deriving limσ→0 −S2 /σ 2 in the proof of Lemma 15
to obtain limσ→0 σ q−2 U2 = q(q − 1)E|B|q−2 .
1
1
q−ǫ
q−ǫ
σ
σ
). Because (χ∗ ) 2−q = Ω( log(1/σ)
) and χ∗ = O(σ q−1 ), there
2. The other case is (χ∗ ) 2−q = Ω( log(1/σ)
1
1
exists a value of 0 ≤ m̄ ≤ m∗ + 1 such that for σ small enough, (χ∗ ) 2−q = o(|Im̄ |) and (χ∗ ) 2−q =
Ω(|Im̄−1 |). We then break the integral into:
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
I γ \K (|ηq (b/σ + z; χ )|
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
=
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
Im̄ \K (|ηq (b/σ + z; χ )|
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
+ 2−q
φ(z)dzdF (b)
(46)
∗ 2−q + χ∗ q(q − 1))3
σ
0
I γ \Im̄ (|ηq (b/σ + z; χ )|
Once we show that each of the two integrals above goes to zero as σ → 0, then the subsequent arguments
will be exactly the same as the ones in Case 1. Regarding the first integral,
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
Im̄ \K (|ηq (b/σ + z; χ )|
Z C3 σ√log(1/σ) Z
1
1
φ(z)dzdF (b)
≤
2−q
∗
σ 2−q 0
(1; 1)
Im̄ \K χ ηq
Z ∞
Z
1
1
+ 2−q
φ(z)dzdF (b)
√
2−q
∗
σ
(1; 1)
C3 σ log(1/σ) Im̄ \K χ ηq
p
p
φ(0)|Im̄ |P(|B| ≤ C3 σ log(1/σ)) φ(C3 log(1/σ)/2)
+
.
≤
σ 2−q ηq2−q (1; 1)χ∗
σ 2−q ηq2−q (1; 1)χ∗
Since χ∗ = Ω(σ q ) from Corollary 7, it is clear that the second term in the above upper bound goes to
zero by choosing large enough C3 . Regarding the first term we have
p
|Im̄ |P(|B| ≤ C3 σ log(1/σ))
σ 2−q χ∗
(a)
≤
(b)
≤
2−q+ǫ
ǫ+q−1
m̄ −
σ ǫ (log(1/σ)) 2 σ q−1 (2−q)
O(1)
m̄
χ∗ log(1/σ)S0 (2−q)
(
−q+ǫ
O(1)(log(1/σ)) 2 = o(1)
O(1)(log(1/σ))
31
ǫ+4−3q
2
ǫ
q−1
2
σ q −(2+ǫ)q+3ǫ+1
if m̄ > 0,
= o(1)
if m̄ = 0,
ǫ
(2−q)m̄ − q−1
≤ 0 by the definition of m∗ ; and (b) is
where (a) holds even when m̄ = m∗ +1 since ǫ+q−1
ǫ+q−1 q−1
m̄−1 − ǫ
q−1 (2−q)
q−1
σq−ǫ
due to the fact that (χ∗ )1/(2−q) = Ω σ
when m̄ > 0 and (χ∗ )1/(2−q) = Ω( log(1/σ)
)
m̄−1
S
(2−q)
(log(1/σ))
0
1
when m̄ = 0, according to the choice of m̄. For the second integral in (46), note that (χ∗ ) 2−q = o(|Im̄ |),
hence χ∗ |Im̄ |q−2 → 0. It implies that the arguments in calculating the second integral in Case 1 hold
here as well.
So far we have been able to derive the limit of σ q−2 U2 . We next analyze the term σ q−2 χ∗ U3 and show
that it goes to zero as σ → 0. We have
Z ∞Z
|ηq (b/σ + z; χ∗ )|2−q
χ∗
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
K (|ηq (b/σ + z; χ )|
Z ∞Z
Z ∞Z
χ∗
χ∗ ηq2−q (1; 1)
ηq2−q (1; 1)
1
≤
φ(z)dzdF
(b)
=
φ(z)dzdF (b)
∗
3
∗
3
σ 2−q 0
σ 2−q 0
K (χ q(q − 1))
K χ (q(q − 1))
The upper bound above has been shown to be zero in the preceding calculations regarding the first term
/ K,
in (45). Furthermore, note that when z ∈
|ηq (b/σ + z; χ∗ )|2−q ≥ χ∗ ηq2−q (1; 1).
We can then obtain
Z ∞Z
|ηq (b/σ + z; χ∗ )|2−q
χ∗
φ(z)dzdF (b)
∗ 2−q + χ∗ q(q − 1))3
σ 2−q 0
I γ \K (|ηq (b/σ + z; χ )|
Z ∞Z
|ηq (b/σ + z; χ∗ )|4−2q
1
≤ O(1) · 2−q
φ(z)dzdF (b).
∗ 2−q + χ∗ q(q − 1))3
σ
0
I γ \K (|ηq (b/σ + z; χ )|
The last term has been shown to converge to zero in the analysis of σ q−2 U2 . Above all we have derived
that
Z ∞Z
|ηq (b/σ + z; χ∗ )|2−q
χ∗
lim 2−q
φ(z)dzdF (b) = 0.
∗ 2−q + χ∗ q(q − 1))3
σ→0 σ
0
I γ (|ηq (b/σ + z; χ )|
This together with the fact
χ∗ |ηq (b/σ + z; χ∗ )|2−q
q −1 (q − 1)−1
≤
,
σ 2−q (|ηq (b/σ + z; χ∗ )|2−q + χ∗ q(q − 1))3
|ηq (b + σz; σ 2−q χ∗ )|2−q + σ 2−q χ∗ q(q − 1)
we can again follow the line of arguments for −σ −2 S2 in the proof of Lemma 15 to get
Z ∞Z
|ηq (b/σ + z; χ∗ )|2−q
χ∗
φ(z)dzdF (b) = 0.
lim 2−q
∗ 2−q + χ∗ q(q − 1))3
σ→0 σ
0
R (|ηq (b/σ + z; χ )|
Finally a direct application of Dominated Convergence Theorem gives us σ 2q−2 U1 → q 2 E|B|2q−2 .
Hence we are able to derive the following
σ q−2 U2 + σ q−2 χ∗ U3
(q − 1)E|B|q−2
χ∗
= lim
=
.
q
2q−2
σ→0
σ→0 σ
σ
U1
qE|B|2q−2
lim
Now that we have derived the convergence rate of χ∗ , according to Lemma 15, we can immediately obtain the
order of Rq (χ∗ , σ).
Having the convergence rate of Rq (χ∗q (σ), σ) as σ → 0 in Lemma 18, the derivation for the expansion of
AMSE(λ∗,q , q, σw ) will be the same as the one in the proof of Theorem 3.
32
4.4.2
Proof for the case q = 1
Lemma 19. Suppose that P (|B| ≤ t) = Θ(σ ℓ ) (as t → 0) and E|B|2 < ∞, then for q = 1
αm σ ℓ ≤ χ∗q (σ) ≤ βm σ ℓ (logm (1/σ))ℓ/2 ,
α̃m σ 2ℓ ≤ 1 − Rq (χ∗q (σ), σ) ≤ β̃m σ 2ℓ (logm (1/σ))ℓ ,
for small enough σ, where logm (1/σ) = log log . . . log
{z
}
|
m times
αm , βm , α̃m , β̃m > 0 are four constants depending on m.
1
σ
; m > 0 is an arbitrary integer number; and
Proof. Since the proof steps are similar to those in Lemma 10, we do not repeat every detail and instead
highlight the differences. We write χ∗ for χ∗q (σ) for notational simplicity. Using the same proof steps in
Lemma 10, we can obtain χ∗ → 0, as σ → 0 and
χ∗ =
Eφ(χ∗ − B/σ) + Eφ(χ∗ + B/σ)
.
E1(|Z + B/σ| ≥ χ∗ )
Following the same arguments from the proof of Lemma 21 in Weng et al. (2016), we can show
Θ(σ ℓ ) ≤ Eφ(χ∗ − B/σ) + Eφ(χ∗ + B/σ) ≤ Θ(σ ℓ (logm (1/σ))ℓ/2 ),
√
Θ(σ ℓ ) ≤ Eφ( 2B/σ), Eφ(−B/σ + αχ∗ ) ≤ Θ(σ ℓ (logm (1/σ))ℓ/2 ),
(47)
(48)
where α is any number between 0 and 1. Since E1(|Z + B/σ| ≥ χ∗ ) → 1 , the bounds for χ∗ is proved by
using the result (47). Furthermore, we know
Rq (χ∗ , σ) − 1 ≤ Rq (χ, σ) − 1 = E(η1 (B/σ + Z; χ) − B/σ − Z)2 + 2E(∂1 η1 (B/σ + Z; χ) − 1)
Z −B/σ+χ
φ(z)dz = χ2 − 4χEφ(−B/σ + αχ),
≤ χ2 − 2E
−B/σ−χ
√
where |α| ≤ 1 is dependent on B. If we choose χ = 3e−1 Eφ( 2B/σ) in the above inequality, it is straightforward to see that
√
Rq (χ∗ , σ) − 1 ≤ −Θ((Eφ( 2B/σ))2 ) ≤ −Θ(σ 2ℓ ),
where the last step is due to (48). For the other bound, note that
Rq (χ∗ , σ) − 1 = E(η1 (B/σ + Z; χ∗ ) − B/σ − Z)2 + 2E(∂1 η1 (B/σ + Z; χ∗ ) − 1)
Z −B/σ+χ∗
≥ −2E
φ(z)dz = −4χ∗ Eφ(−B/σ + αχ∗ ) ≥ −Θ(σ 2ℓ (logm (1/σ))ℓ ).
−B/σ−χ∗
The last inequality holds because of the upper bound on χ∗ and (48).
Based on the results of Lemma 19, deriving the expansion of AMSE(λ∗,1 , 1, σw ) can be done in a similar
way as in the proof of Theorem 3. We do not repeat it here.
4.5 Proof of Theorem 5
The idea of this proof is similar to those for Theorems 3 and 4. We make use of the result in Theorem 1:
2
AMSE(λ∗,q , q, δ) = σ̄ 2 Rq (χ∗q (σ̄), σ̄) = δσ̄ 2 − σw
.
33
(49)
Since we are in the large sample regime where δ → ∞, σ̄ is a function of δ. It is clear from (49) that 0 ≤
2 ≤ σ̄ 2 . Hence σ̄ 2 ≤ σ 2 /(δ − 1) → 0, which further leads to
δσ̄ 2 − σw
w
σ̄ 2 =
2
σw
+ o(1/δ).
δ
(50)
Due to the fact that σ̄ → 0 as δ → ∞, we will be able to use the convergence rate results of Rq (χ∗q (σ), σ) (as
σ → 0) we have proved in Lemmas 9 and 10. For 1 < q ≤ 2, Equations (49), (50) and Lemma 9 together yield
2
δ2 (AMSE(λ∗,q , q, δ) − σw
/δ) = δ2 (σ̄ 2 Rq (χ∗q (σ̄), σ̄) − (σ̄ 2 − σ̄ 2 Rq (χ∗q (σ̄), σ̄)/δ))
Rq (χ∗q (σ̄), σ̄) − 1
+ (δσ̄ 2 ) · Rq (χ∗q (σ̄), σ̄)
σ̄ 2
−(q − 1)2 (E|B|q−2 )2 4
2
→
σw + σw
.
E|B|2q−2
= (σ̄ 4 δ2 ) ·
(51)
For the case q = 1, from Lemma 10 we know Rq (χ∗q (σ̄), σ̄) − 1 is exponentially small. So the firs term in (51)
vanishes and the second term remains the same.
4.6 Proof of Theorem 6
Theorem 6 can be proved in a similar fashion as for Theorem 5. Equation (50) still holds. Equations (49), (50)
and Lemma 19 together give us for q = 1,
2
δℓ+1 (AMSE(λ∗,q , q, δ) − σw
/δ) = (σ̄ 2ℓ+2 δℓ+1 ) ·
Rq (χ∗q (σ̄), σ̄) − 1
+ (δℓ σ̄ 2 ) · Rq (χ∗q (σ̄), σ̄),
σ̄ 2ℓ
where the first term above is Θ(1) and the second one is o(1) when ℓ < 1. The case 1 < q ≤ 2 can be proved
exactly the same way as in Theorem 5 by using Lemma 18.
Acknowledgment
Arian Maleki is supported by NSF grant CCF1420328.
References
A MELUNXEN , D., L OTZ , M., M C C OY, M. B. and T ROPP, J. A. (2014). Living on the edge: phase
transitions in convex programs with random data. Information and Inference: A Journal of the IMA, 3 224.
/oup/backfile/Content_public/Journal/imaiai/3/3/10.1093/imaiai/iau005/2/iau005.pdf
URL +http://dx.doi.org/10.1093/imaiai/iau005.
BAI , Z. and Y IN , Y. (1993). Limit of the smallest eigenvalue of a large dimensional sample covariance matrix.
The annals of Probability 1275–1294.
BAYATI , M. and M ONTANARI , A. (2012). The LASSO risk for Gaussian matrices. IEEE Trans Inform Theory,
58 1997–2017.
BAYATI , M. and M ONTANRI , A. (2011). The dynamics of message passing on dense graphs, with applications
to compressed sensing. IEEE Trans. Inform. Theory, 57 764–785.
34
B ICKEL , P. J., R ITOV, Y. and T SYBAKOV, A. B. (2009). Simultaneous analysis of lasso and dantzig selector.
The Annals of Statistics 1705–1732.
B RADIC , J. and C HEN , J. (2015). Robustness in sparse linear models: relative efficiency based on robust
approximate message passing. arXiv preprint arXiv:1507.08726.
B ÜHLMANN , P. and VAN D E G EER , S. (2011). Statistics for high-dimensional data: methods, theory and
applications. Springer Science & Business Media.
C ANDES , E. J. and P LAN , Y. (2011). A probabilistic and ripless theory of compressed sensing. IEEE Transactions on Information Theory, 57 7235–7254.
C AND ÈS , E. J., ROMBERG , J. and TAO , T. (2006). Robust uncertainty principles: Exact signal reconstruction
from highly incomplete frequency information. IEEE Transactions on information theory, 52 489–509.
C ANDES , E. J. and TAO , T. (2006). Near-optimal signal recovery from random projections: Universal encoding
strategies? IEEE transactions on information theory, 52 5406–5425.
D ONOHO , D. (2004a). For most underdetermined systems of linear equations, the minimal ℓ1 -norm approximates the sparsest solution. Comm. Pure and Appl. Math.
D ONOHO , D. (2004b). For most underdetermined systems of linear equations, the minimal ℓ1 -norm nearsolution approximates the sparsest near-solution. Manuscript, submitted for publication, URL: http://wwwstat. stanford. edu/˜ donoho/Reports.
D ONOHO , D. (2006a). High-dimensional centrally symmetric polytopes with neighborliness proportional to
dimension. Discrete & Computational Geometry, 35 617–652.
D ONOHO , D., G AVISH , M. and M ONTANARI , A. (2013). The phase transition of matrix recovery from
gaussian measurements matches the minimax mse of matrix denoising. Proceedings of the National Academy
of Sciences, 110 8405–8410.
D ONOHO , D., M ALEKI , A. and M ONTANARI , A. (2011a). The noise-sensitivity phase transition in compressed sensing. IEEE Transactions Information Theory, 57 6920–6941.
D ONOHO , D. and M ONTANARI , A. (2013). High dimensional robust m-estimation: Asymptotic variance via
approximate message passing. Probability Theory and Related Fields 1–35.
D ONOHO , D. and M ONTANARI , A. (2015). Variance breakdown of huber (m)-estimators: n/p → m. arXiv
preprint arXiv:1503.02106.
D ONOHO , D. and TANNER , J. (2005a). Neighborliness of randomly projected simplices in high dimensions.
Proceedings of the National Academy of Sciences, 102 9452–9457.
D ONOHO , D. and TANNER , J. (2005b). Sparse nonnegative solution of underdetermined linear equations by
linear programming. Proceedings of the National Academy of Sciences, 102 9446–9451.
D ONOHO , D. L. (2006b). Compressed sensing. IEEE Transactions on information theory, 52 1289–1306.
D ONOHO , D. L., M ALEKI , A. and M ONTANARI , A. (2009). Message-passing algorithms for compressed
sensing. Proceedings of the National Academy of Sciences, 106 18914–18919.
35
D ONOHO , D. L., M ALEKI , A. and M ONTANARI , A. (2011b). Noise sensitivity phase transition. IEEE Trans.
Inform. Theory, 57 6920–6941.
E L K AROUI , N., B EAN , D., B ICKEL , P., L IM , C. and Y U , B. (2013). On robust regression with highdimensional predictors. Proceedings of the National Academy of Sciences, 110 14557–14562.
E LDAR , Y. C. and K UTYNIOK , G. (2012). Compressed sensing: theory and applications. Cambridge University Press.
F OYGEL , R. and M ACKEY, L. (2014). Corrupted sensing: Novel guarantees for separating structured signals.
IEEE Transactions Information Theory, 60 1223–1247.
F RANK , L. E. and F RIEDMAN , J. H. (1993). A statistical view of some chemometrics regression tools.
Technometrics, 35 109–135.
G REENSHTEIN , E., R ITOV, Y. ET AL . (2004). Persistence in high-dimensional linear predictor selection and
the virtue of overparametrization. Bernoulli, 10 971–988.
H OERL , A. E. and K ENNARD , R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems.
Technometrics, 12 55–67.
K AROUI , N. (2013). Asymptotic behavior of unregularized and ridge-regularized high-dimensional robust
regression estimators: rigorous results. arXiv preprint arXiv:1311.2445.
K NIGHT, K. and F U , W. (2000). Asymptotics for lasso-type estimators. Annals of statistics 1356–1378.
K RZAKALA , F., M ÉZARD , M., S AUSSET, F., S UN , Y. and Z DEBOROV Á , L. (2012). Statistical-physics-based
reconstruction in compressed sensing. Physical Review X, 2 021005.
M ALEKI , A. (2010). Approximate message passing algorithms for compressed sensing. Ph.D. thesis, Stanford
University.
M ALEKI , A., A NITORI , L., YANG , Z. and BARANIUK , R. (2013). Asymptotic analysis of complex lasso via
complex approximate message passing (camp). IEEE Transactions Information Theory, 59 4290–4308.
M EINSHAUSEN , N. and B ÜHLMANN , P. (2006). High-dimensional graphs and variable selection with the
lasso. The annals of statistics 1436–1462.
M OUSAVI , A., M ALEKI , A. and BARANIUK , R. G. (2017). Consistent parameter estimation for lasso and
approximate message passing. To appear in Annals of Statistics.
R ANGAN , S., G OYAL , V. and F LETCHER , A. (2009). Asymptotic analysis of map estimation via the replica
method and compressed sensing. In Advances in Neural Information Processing Systems. 1545–1553.
R ASKUTTI , G., WAINWRIGHT, M. J. and Y U , B. (2011). Minimax rates of estimation for high-dimensional
linear regression over ℓq -balls. IEEE transactions on information theory, 57 6976–6994.
R EEVES , G. and D ONOHO , D. (2013). The minimax noise sensitivity in compressed sensing. In Information
Theory Proceedings (ISIT), 2013 IEEE International Symposium on. IEEE, 116–120.
R EEVES , G. and G ASTPAR , M. (2008). Sampling bounds for sparse support recovery in the presence of noise.
In Information Theory, 2008. ISIT 2008. IEEE International Symposium on. IEEE, 2187–2191.
36
R EEVES , G. and P FISTER , H. D. (2016). The replica-symmetric prediction for compressed sensing with
gaussian matrices is exact. In Information Theory (ISIT), 2016 IEEE International Symposium on. IEEE,
665–669.
S TEIN , C. M. (1981). Estimation of the mean of a multivariate normal distribution. The annals of Statistics
1135–1151.
S TOJNIC , M. (2009a). Block-length dependent thresholds in block-sparse compressed sensing. arXiv preprint
arXiv:0907.3679.
S TOJNIC , M. (2009b).
arXiv:0907.3666.
Various thresholds for ℓ1 -optimization in compressed sensing.
arXiv preprint
S TOJNIC , M. (2013).
arXiv:1306.3774.
Under-determined linear systems and ℓq -optimization thresholds.
arXiv preprint
T HRAMPOULIDIS , C., A BBASI , E. and H ASSIBI , B. (2016). Precise error analysis of regularized m-estimators
in high-dimensions. arXiv preprint arXiv:1601.06233.
T IBSHIRANI , R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical
Society. Series B 267–288.
WAINWRIGHT, M. J. (2009). Sharp thresholds for high-dimensional and noisy sparsity recovery using ℓ1 constrained quadratic programming (lasso). IEEE transactions on information theory, 55 2183–2202.
WANG , S., W ENG , H. and M ALEKI , A. (2017). Which bridge estimator is optimal for variable selection?
arXiv preprint arXiv:1705.08617.
W ENG , H., M ALEKI , A. and Z HENG , L. (2016). Overcoming the limitations of phase transition by higher
order analysis of regularization techniques. arXiv preprint arXiv:1603.07377.
Z HAO , P. and Y U , B. (2006). On model selection consistency of lasso. Journal of Machine learning research,
7 2541–2563.
Z HENG , L., M ALEKI , A., W ENG , H., WANG , X. and L ONG , T. (2017). Does ℓp -minimization outperform
ℓ1 -minimization? IEEE Transactions on Information Theory, accepted.
37
| 10math.ST
|
1
Traffic Models of Periodic Event-Triggered Control
Systems
arXiv:1711.03599v1 [cs.SY] 9 Nov 2017
Anqi Fu, and Manuel Mazo, Jr.
Abstract—Periodic event-triggered control (PETC) [13] is a
version of event-triggered control (ETC) that only requires to
measure the plant output periodically instead of continuously. In
this work, we present a construction of timing models for these
PETC implementations to capture the dynamics of the traffic they
generate. In the construction, we employ a two-step approach.
We first partition the state space into a finite number of regions.
Then in each region, the event-triggering behavior is analyzed
with the help of LMIs. The state transitions among different
regions result from computing the reachable state set starting
from each region within the computed event time intervals.
Index Terms—systems abstractions; periodic event-triggered
control; LMI; formal methods; reachability analysis.
I. I NTRODUCTION
Wireless networked control systems (WNCS) are control
systems that employ wireless networks as feedback channels.
In such systems, the physically distributed components are
co-located with their own wireless nodes and communicate
via a wireless network. These components can be designed
with great mobility once the nodes are supported by batteries.
Besides, each component can be established and updated
easily. Therefore, WNCS have great adaptability on obtaining
different control objectives and have been attracting much
attention. However, there are two major issues that must be
considered while designing such a system: limited bandwidth
and energy supply.
Most often, control tasks are designed to be executed
periodically. This periodic strategy, also named time-triggered
control (TTC), does not regard the system’s current state and
thus may waste bandwidth and energy. Alternatively, eventtriggered control (ETC) strategies are proposed to reduce
bandwidth occupation, see e.g. [5], [17], [19], [22], [25], [26],
and references therein. In ETC, the control tasks only execute
when necessary, e.g. when some pre-designed performance indicator is about to be violated. Thus the system is tightfisted in
communication. However, to validate the pre-designed eventtriggering conditions, sensors are required to sample the plant
output continuously. This continuous monitoring can consume
large amounts of energy. To reduce this energy consumption,
naturally one may want to replace the continuously sampling
by a discrete time sampling.
When applying discrete time sampling, to compensate the
delay caused by the discretization, one can either design a
The authors are with the Delft Center for Systems and Control, Delft University of Technology, Delft 2628 CD, The Netherlands e-mail: A.Fu-1,
[email protected].
This work is partly funded by China Scholarship Council (CSC).
stricter event-triggering condition based on the system dynamics, as e.g. [18]; or modify the Lyapunov function, as e.g.
[13]. In [13], Heemels et. al. present a periodic event-triggered
control (PETC) mechanism. In a PETC implementation, the
sensors are only required to measure the plant output and
validate the event conditions periodically. Only when some
pre-designed conditions are satisfied, fresh measurements are
employed to recompute the controller output. Therefore, PETC
enjoys the benefits of both cautious communication and discrete time measurement. Compared to [18], the event conditions can be less conservative to further reduce communications. Thus the energy consumed and bandwidth occupied are
reduced. Furthermore, the transmissions of the control input
from the controller to the plant are also included in the PETC
mechanism.
To further reduce the resource consumption and to fully
extract the potential gains from ETC, one can also consider
scheduling approaches. By efficiently scheduling listening
times on wireless communications and medium access time in
general, the energy consumption in a WNCS can be reduced
and bandwidth can be more efficiently reused. To enable
such scheduling, a model for the traffic generated by ETC
is required. In [16], Kolarijani and Mazo propose a type of
approximate power quotient system, to derive models that
capture the timing behaviors of ETC systems applying the
triggering mechanism from [22]. They first partition the state
space into finite cones. In each cone, they analyze the timing
behavior by over-approximation methods (see e.g. [3], [4], [6],
[11], [14], [20], [21]), linear matrix inequality (LMI) methods,
and reachability analysis (see e.g. [1] and [2]).
Similarly, in order to fully extract the potential gains from
PETC with scheduling approaches, a model for the traffic
generated by PETC is necessary. In this work, we present a
construction of the timing models of the PETC implementations from [13]. First of all, we modify the PETC mechanism
by giving an upper bound time such that if no event happens
within that interval, the system will be forced to generate an
event by the end of it. When constructing the models, the
approach has two steps. We first divide the state space into
a finite number of partitions. For a 2-dimensional system, the
partition looks like a dartboard. Then we construct a set of
LMIs to compute the output map. Transition relations among
different regions are derived by computing the reachable state
set starting from each region. Compared with the work from
[9], we do not require that the perturbation should vanish as
the state converges. Instead, we only assume the perturbation
to be both L2 and L∞ .
This paper is organized as follows. Section II presents the
2
necessary notation and definitions. The problem to be solved
is defined in Section III. Section IV shows all the details to
construct a power quotient system to model the traffic of a
centralized PETC implementation. A numerical example is
shown in Section V. Section VI summarizes the contributions
of this paper and discusses future work. To ease the readability,
the proofs are collected in the Appendix.
II. N OTATION AND PRELIMINARIES
We denote the n-dimensional Euclidean space by Rn , the
+
positive real numbers by R+ , by R+
0 = R ∪ {0}. The natural
numbers including zero is denoted by N. When zero is not
included, we denote the natural numbers as N+ . IN+ is the
set of all closed intervals [a, b] such that a, b ∈ N+ and a ≤ b.
For any set S, 2S denotes the set of all subsets of S, i.e. the
power set of S. Mm×n and Mn are the set of all m × n real
valued matrices and the set of all n × n real-valued symmetric
matrices respectively. A symmetric matrix M ∈ Rn×n is said
to be positive (negative) definite, denoted by M 0 (M ≺ 0),
whenever xT M x > 0 (xT M x < 0) for all x 6= 0, x ∈ Rn .
M 0 (M 0) means M is a positive (negative) semidefinite matrix. When Q ⊆ Z × Z is an equivalence relation
on a set Z, [z] denotes the equivalence class of z ∈ Z and
Z/Q denotes the set of all equivalence classes. For a locally
integrable
signal w: R+ → Rn , we denote by kwkL2 =
qR
∞
|w(t)|2 dt its L2 -norm, kwkL∞ = supt≥0 kw(t)k < ∞
0
its L∞ -norm. Furthermore, we define the space of all locally
integrable signals with a finite L2 -norm as L2 , the space of
all signals with a finite L∞ -norm as L∞ .
Now we review some notions from the field of system
theory.
Definition 2.1: (Metric)[7] Consider a set T , d : T × T →
R ∪ {+∞} is a metric (or a distance function) if the following
three conditions are satisfied ∀x, y, z ∈ T :
• d(x, y) = d(y, x);
• d(x, y) = 0 ↔ x = y;
• d(x, y) ≤ d(x, z) + d(y, z).
The ordered pair (T, d) is said to be a metric space.
Definition 2.2: (Hausdorff distance)[7] Assume X and Y
are two non-empty subsets of a metric space (T, d). The
Hausdoorff distance dH (X, Y ) is given by:
max sup inf d(x, y), sup inf d(x, y) .
(1)
x∈X y∈Y
y∈Y x∈X
Definition 2.3: (System)[23] A system is a sextuple
(X, X0 , U, −→, Y, H) consisting of:
• a set of states X;
• a set of initial states X0 ⊆ X;
• a set of inputs U ;
• a transition relation −→⊆ X × U × X;
• a set of outputs Y ;
• an output map H : X → Y .
The term finite-state (infinite-state) system indicates X is a
finite (an infinite) set. For a system, if the cardinality of U is
smaller than or equal to one, then this system is said to be
autonomous.
Definition 2.4: (Metric system)[23] A system S is said to
be a metric system if the set of outputs Y is equipped with a
metric d : Y × Y → R+
0.
Definition 2.5: (Approximate simulation relation)[23] Consider two metric systems Sa and Sb with Ya = Yb , and let
∈ R+
0 . A relation R ⊆ Xa × Xb is an -approximate
simulation relation from Sa to Sb if the following three
conditions are satisfied:
• ∀xa0 ∈ Xa0 , ∃xb0 ∈ Xb0 such that (xa0 , xb0 ) ∈ R;
• ∀(xa , xb ) ∈ R we have d (Ha (xa ), Hb (xb )) ≤ ;
0
• ∀(xa , xb ) ∈ R such that (xa , ua , xa ) ∈−→ in Sa implies
a
0
∃(xb , ub , xb ) ∈−→ in Sb satisfying (x0a , x0b ) ∈ R.
b
We denote the existence of an -approximate simulation
relation from Sa to Sb by Sa S Sb , and say that
Sb -approximately simulates Sa or Sa is -approximately
simulated by Sb . Whenever = 0, the inequality
d(Ha (xa ), Hb (xb )) ≤ implies Ha (xa ) = Hb (xb ) and the
resulting relation is called an (exact) simulation relation.
We introduce a notion of power quotient system and corresponding lemma for later analysis.
Definition 2.6: (Power quotient system)[16] Let S =
(X, X0 , U, −→, Y, H) be a system and R be an equivalence
relation on X.
The power quotient of S by R, denoted
by S/R ,
is the system
X/R , X/R,0 , U/R , −→, Y/R , H/R
/R
consisting
of:
•
•
•
•
X/R = X/R;
X/R,0 = x/R ∈ X/R |x/R ∩ X0 6= ∅ ;
U
/R = U ;
x/R , u, x0/R
∈−→ if ∃(x, u, x0 ) ∈−→ in S with x ∈
/R
x/R and x0 ∈ x0/R ;
Y
• Y/R ⊂ 2 ;
S
• H/R x/R =
x∈x/R H(x).
Lemma 2.7: [Lemma 1 in [16]] Let S be a metric system,
R be an equivalence relation on X, and let the metric system
S/R be the power quotient system of S by R. For any
(2)
≥
max
d H(x), H/R x/R ,
x ∈ x/R
x/R ∈ X/R
with d the Hausdorff distance over the set 2Y , S/R approximately simulates S, i.e. S S S/R .
The definition of Minkowski addition is introduced here for
the computation of the reachable sets.
Definition 2.8: (Minkowski addition) The Minkowski addition of two sets of vectors A and B in Euclidean space is
formed by adding each vector in A to each vector in B:
A ⊕ B = {a + b|a ∈ A, b ∈ B},
where ⊕ denotes the Minkowski addition.
III. P ROBLEM DEFINITION
The centralized PETC presented in [13] is reviewed here.
Consider a continuous linear time-invariant (LTI) plant of the
form:
(
ξ˙p (t) = Ap ξp (t) + Bp v̂(t) + Ew(t)
(3)
y(t) = Cp ξp (t),
3
where ξp (t) ∈ Rnp denotes the state vector
y(t) ∈ Rny denotes the plant output vector,
denotes the input applied to the plant, w(t) ∈
the perturbation. The plant is controlled by a
controller, given by:
(
ξc (tk+1 ) = Ac ξc (tk ) + Bc ŷ(tk )
of the plant,
v̂(t) ∈ Rnv
Rnw denotes
discrete-time
v(tk ) = Cc ξc (tk ) + Dc ŷ(tk ),
(4)
where ξc (tk ) ∈ Rnc denotes the state vector of the controller,
v(tk ) ∈ Rnv denotes the controller output vector, and ŷ(tk ) ∈
Rny denotes the input applied to the controller. A periodic
sampling sequence is given by:
Ts := {tk |tk := kh, k ∈ N},
(5)
where h > 0 is the sampling interval. Define two vectors:
T
u(t) : = y T (t) v T (t) ∈ Rnu
(6)
T
û(tk ) : = ŷ T (tk ) v̂ T (tk ) ∈ Rnu ,
with nu := ny + nv . u(t) is the output of the implementation,
û(t) is the input of the implementation. A zero-order hold
mechanism is applied between samplings to the input. At each
sampling time tk , the input applied to the implementation
û(tk ) is updated ∀tk ∈ Ts :
(
u(tk ),
if ku(tk ) − û(tk )k > σku(tk )k
û(tk ) =
(7)
û(tk−1 ), if ku(tk ) − û(tk )k ≤ σku(tk )k,
where σ > 0 is a given constant. Reformulating the event
condition as a quadratic form, the event sequence can be
defined by:
Te := tb |b ∈ N, tb ∈ Ts , ξ T (tb )Qξ(tb ) > 0 .
(8)
T
T
where ξ(t) := ξp (t) ξcT (t) ŷ T (t) v̂ T (t) ∈ Rnξ , with
nξ := np + nc + ny + nv . And:
Q1 Q2
Q=
QT
Q4
2
in which:
0
(1 − σ)CpT Cp
Q
=
1
0
(1 − σ)CcT Cc
−CpT
0
Q2 =
(1 − σ)CcT Dc −CcT
I + (1 − σ)DcT Dc −DcT
Q4 =
−Dc
I
0 is a zero matrix with proper dimension, I is an identity
matrix with appropriate dimension.
It is obvious that Te ⊆ Ts . According to Theorem V.2 in
[13], if the hypothesis therein are satisfied, then the system
(3-8):
1) is globally exponential stable (GES), i.e. ∃c > 0 and ρ >
0 s.t. ∀ξ(0) ∈ Rnξ with w = 0, kξ(t)k ≤ ce−ρt kξ(0)k
for all t ∈ R+ .
2) has L2 -gain from w to z smaller than or equal to γ, i.e.
∃σ : Rnξ → R+ s.t. ∀w ∈ L2 , ξ(0) ∈ Rnξ , the corresponding solution to the system and z(t) := g(ξ(t), w(t))
satisfies kzkL2 ≤ σ(ξ(0)) + γkwkL2 .
To model the timing behaviour of a PETC system, we aim at
constructing a power quotient system for this implementation.
Remark 3.1: Because of the uncertainty brought by the
perturbation, it may happen that the perturbation compensates
the effect of sampling, helping the state of the implementation
to converge. Therefore the event condition in (8) may not be
satisfied along the timeline. As a result, there may not be an
upper bound for the event intervals. However an upper bound
is necessary for constructing a useful power quotient system.
Remark 3.2: To apply scheduling approaches, an online
scheduler is required. The model we are going to construct
is non-deterministic, meaning that after an event the system
may end up in several possible regions, but those regions are
defined in terms of ξp , which means that from a measurement
is not always clear in which region the system is. That means
this online scheduler cannot figure out where the system
is from simple output measurements. Therefore, the online
scheduler should be able to access in which region the system
is.
Assumption 3.3: The current state region at each eventtriggered time tb can be obtained in real time.
Because of the observation in Remark 3.1, we use instead
the following event condition:
tb+1 = inf {tk |tk ∈ Ts , tk > tb ,
o
_
ξ T (tk )Qξ(tk ) > 0 tk ≥ tb + τ̄R(ξ(tb )) ,
(9)
where R(ξ(tb )) is the state region on state-space Rnξ at last
sampling time tb , τ̄R(ξ(tb )) is a regional maximum allowable
event interval (MAEI), which is dependent on R(ξ(tb )).
According to Assumption 3.3, R(ξ(tb )) is obtainable. If
this value is not possible to be accessed by the triggering
mechanisms, one can always employ a global upper bound
τ̄ :≥ τ̄R(ξ(tb )) . We will discuss the computation of τ̄R(tb )
in later sections. Note that, if the PETC implementation
employing (8) can guarantee some pre-designed stability and
performance, then the PETC implementation employing (9)
can guarantee the same stability and performance.
Consider a period:
τ (x) := tb+1 − tb = kx h.
(10)
By definition û(t) is constant ∀t ∈ [tb , tb+1 [ and dependent
on ξp (tb ) and ξc (tb ). The input û(t) can be expressed as:
Cp
0
û(t) = CE x, CE :=
,
Dc Cp Cc
where
T
x := ξpT (tb ) ξcT (tb ) .
T
Let ξx (k) := ξpT (tb + kh) ξcT (tb + kh)
be the state
T
T
T
evolution with initial state x = ξp (tb ) ξc (tb ) , and k ∈ N.
Now ξx (k) can be computed as:
ξx (k) = M (k)x + Θ(k),
where
M (k) :=
M1 (k)
Θ1 (k)
,
, Θ(k) :=
M2 (k)
0
(11)
4
Z kh
eAp s ds Ap I 0
M1 (k) := I 0 +
0
+Bp Dc Cp Cc ,
X
k−1
k
0
I
M
(k)
:=A
+
Ak−1−i
Bc Cp 0 ,
2
c
c
i=0
Z kh
Θ1 (k) :=
eAp (kh−s) Ew(s)ds.
Figure 1. An example of the state space partition, into (a) finite polyhedral
cones, (b) finite homocentric spheres, and (c) finite state-space partition.
0
Thus from the event condition in (9), kx in (10) can be
computed by:
(12)
kx = min k x , k x ,
τ̄
and
where k x := R(x)
h
k x := inf k ∈ N+
))
T
M (k)x + Θ(k)
M (k)x + Θ(k)
Q
>0
.
CE x
CE x
(13)
Now we present the main problem to be solved in this paper.
Consider the system:
S = (X, X0 , U, −→, Y, H),
(14)
where
n
• X = R x , nx = np + nc ;
n
• X0 ⊆ R x ;
• U = ∅;
0
0
• −→⊆ X × U × X such that ∀x, x ∈ X : (x, x ) ∈−→
0
iff ξx (H(x)) = x ;
+
• Y ⊂N ;
nx
• H :R
→ N+ where H(x) = kx .
S is an infinite-state system. The output set Y of system S
contains all the possible amount of sampling steps tb+1h−tb ∈
N, b ∈ N that the system (3-7), and (9) may exhibit. Once
the sampling time h is chosen, the event interval then can be
computed by kx h.
Problem 3.4: Construct a finite abstraction of system S
capturing enough information for scheduling.
Inspired by [16], we solve this problem by constructing a
power quotient systems S/P based on an adequately designed
equivalence relation P defined over the state set X of S. The
constructed systems S/P are semantically equivalent to timed
automata, which can be used for automatic scheduler design
[15].
In particular, the system S/P to be constructed is as follows:
S/P = X/P , X/P,0 , U/P , −→, Y/P , H/P ,
(15)
/P
•
•
•
X/P = Rn/Px := {R1 , · · · , Rq };
X
= Rn/Px ;
/P,0
x/P , x0/P ∈−→ if ∃x ∈ x/P , ∃x0 ∈ x0/P such that
Compared with the power quotient system constructed in
[16], a main difference is that since we focus on PETC, there
is no timing uncertainty.
IV. C ONSTRUCTION OF THE QUOTIENT SYSTEM
A. State set
From the results in [8], we remark the following fact:
Remark 4.1: When w = 0, excluding the origin, all the
states that lie on a line going through the origin have an
identical triggering behavior.
We also call the following assumption:
Assumption 4.2: The perturbation w satisfies w ∈ L2 and
w ∈ L∞ . Besides, assume an upper bound W > 0 for kwkL∞ ,
i.e. kwkL∞ ≤ W, is known.
Base on Remark 4.1 and Assumption 4.2, we propose statespace partition as follows:
(
n^
x −1
xT Ξs1 ,(i,i+1) x ≥ 0
Rs1 ,s2 = x ∈ Rnx
(16)
i=1
o
^
Ws2 −1 ≤ |x| < Ws2 ,
where s1 ∈ {1, · · · , q1 }, s2 ∈ {1, · · · , q2 }, q1 , q2 ∈ N
are pre-designed scalars. Ξs1 ,(i,j) is a constructed matrix;
{Wi |i ∈ {0, · · · , q2 }} is a sequence of scalars. Note that
W0 = 0, Wq2 = +∞, and the rest Ws2 are bounded
and
It is obvious that
S somewhere in between 0 and +∞.
nx
R
=
R
.
s1 ,s2
s1 ∈{1,··· ,q1 },s2 ∈{1,··· ,q2 }
This state-space partition combines partitioning the statespace into finite polyhedral cones (named as isotropic covering
[8]) and finite homocentric spheres. From (16), we can see
that, the isotropic covering describes the relation between
entries of the state vector, while the transverse isotropic
covering is used to capture the relation between the norm of
the state vector and the L∞ norm of the perturbations, which
will be shown later in Theorem 4.4. If w = 0, the homocentric
spheres can be omitted.
Details on the isotropic covering can be found in the
Appendix. Figure 1 shows a 2-dimensional example.
/P
•
•
ξx (H(x)) = x0 ;
Y/P ⊂ 2Y ⊂ IN+;
H
h /P x/P i= minx∈x/P H(x), maxx∈x/P H(x) :=
k x/P , k̄x/P .
S/P is a finite state system.
B. Output map
We first free the system dynamics from the uncertainty
brought by the perturbation.
Lemma 4.3: Consider the system (3-7) and (9), and that
Assumptions 4.2 hold. If there exist a scalar µ ≥ 0 and a
5
symmetric matrix Ψ such that (Q1 + Ψ)1 µI, then k x
generated by (13) is lower bounded by:
kx0 := inf{k ∈ N+ |Φ(k) 0},
where
Q1 + Ψ =
(Q1 + Ψ)1
(Q1 + Ψ)3
(Q1 + Ψ)2
(Q1 + Ψ)4
(17)
(Q1 + Ψ)1 ∈ Rnp ×np ,
Φ1 (k) Φ2 (k)
0
−Ψ
0 ,
Φ(k) := ΦT
2 (k)
0
0
Φ3 (k)
T
T
Φ1 (k) =M (k)Q1 M (k) + M (k)Q2 CE
T
T
+ CE
Q3 M (k) + CE
Q4 CE
T
T
Φ2 (k) =M (k)Q1 + CE Q3
Φ (k) =khµλ
T
3
max E E dAp (k),
(18)
(19)
and
kλmax (Ap +AT
p) − 1
e
, if λmax Ap + AT
p 6= 0,
T
dAp (k) =
λmax Ap + Ap
kh,
if λmax Ap + AT
p = 0.
Next we construct LMIs that bridge Lemma 4.3 and the
state-space partition.
Theorem 4.4: (Regional lower bound) Consider a scalar
k s1 ,s2 ∈ N and regions with s2 > 1. If all the hypothesis in
Lemma 4.3 hold and there exist scalars εk,(s1 ,s2 ),(i,i+1) ≥ 0
where i ∈ {1, · · · , nx −1} such that for all k ∈ {0, · · · , k s1 ,s2 }
the following LMIs hold:
H
Φ2 (k)
0,
(20)
ΦT
−Ψ
2 (k)
according to Remark 4.1, only applying isotropic covering is
enough.
We define Rs1 ,• to represent Rs1 ,s2 , ∀s2 ∈ {1, · · · , q2 }.
Corollary 4.7: (Regional lower bound when w = 0) Consider a scalar k s1 ,• ∈ N. If there exist scalars εk,s1 ,(i,i+1) ≥ 0
where i ∈ {1, · · · , nx −1} such that for all k ∈ {0, · · · , k s1 ,• }
the following LMIs hold:
X
Φ1 (k) +
εk,s1 ,(i,i+1) Ξs1 ,(i,i+1) 0,
(22)
i∈{1,··· ,nx −1}
with Φ1 (k) defined in (19), then the inter event time (8) of the
system (3-7) with w = 0 are regionally bounded from below
by (k s1 ,• + 1)h.
Corollary 4.8: (Regional upper bound when w = 0) Let
¯l ∈ N be a large enough scalar. Consider a scalar k̄s ,• ∈
1
k s1 ,• , · · · , ¯l . If there exist scalars ε̄k,s1 ,(i,i+1)
≥ 0 where
i ∈ {1, · · · , nx − 1} such that for all k ∈ k̄s1 ,• , · · · , ¯l the
following LMIs hold:
X
Φ1 (k) −
ε̄k,s1 ,(i,i+1) Ξs1 ,(i,i+1) 0,
(23)
i∈{1,··· ,nx −1}
with Φ1 (k) defined in (19), then the inter event time (8) of the
system (3-7) with w = 0 are regionally bounded from above
by k̄s1 ,• h.
Remark 4.9: For the choice of ¯l, we follow Remark 2 in [16],
and apply a line search approach: increasing ¯l until Φ1 (¯l) 0.
This results in ¯l being a global upper bound for the inter event
time (8) of the system (3-7) with w = 0.
It is obvious that ¯l ≥ k̄s1 ,• > k s1 ,• ≥ k s1 ,s2 , ∀s2 . We can
now set the regional MAEI τ̄R(ξ(tb )) in (9) as: τ̄R(ξ(tb )) :=
k̄s1 ,• h, ∀x ∈ Rs1 ,• .
where
H =Φ1 (k) + Φ3 (k)W 2 Ws−2
I
2 −1
X
+
εk,(s1 ,s2 ),(i,i+1) Ξs1 ,(i,i+1) ,
i∈{1,··· ,nx −1}
with Φ1 (k), Φ2 (k), and Φ3 (k) defined in (19), and Ψ from
Lemma 4.3, then the inter event times (9) for system (3-7)
are regionally bounded from below by (k s1 ,s2 + 1)h. For the
regions with s2 = 1, the regional lower bound is h.
Remark 4.5: In Theorem 4.4, we discuss the situations when
s2 > 1 and s2 = 1, since for all regions with s2 > 1, it holds
that Ws2 −1 6= 0; while for all regions with s2 = 1, Ws2 −1 = 0
holds. When Ws2 −1 6= 0, one can easily validate the feasibility
of the LMI (20); while when Ws2 −1 = 0, H will be diagonal
infinity, making the LMI (20) infeasible when k > 0. However,
according to the property of PETC, i.e. tb+1 ∈ Ts and tb+1 >
tb , the regional lower bound exists and is equal to h.
Following similar ideas of Theorem 4.4, we present next
lower and upper bounds starting from each state partition when
w = 0. Consider the following event condition:
(
)
T
M
(k)x
M
(k)x
kx = inf k ∈ N+
Q
>0 .
(21)
CE x
CE x
Remark 4.6: Since (21) does not consider perturbations,
when computing the lower and upper bound for each region,
C. Transition relation
In this subsection, we discuss the construction of the transition relation and the reachable state set. Denote the initial
state set as X0,(s1 ,s2 ) , after k-th samplings without an update,
the reachable state set is denoted as Xk,(s1 ,s2 ) . According to
(11), a relation can be obtained as:
Xk,(s1 ,s2 ) = M (k)X0,(s1 ,s2 ) + Θ(k).
(24)
It is obvious that, Xk,(s1 ,s2 ) cannot be computed directly,
because the perturbation is uncertain and the state region may
not be convex. Therefore, we aim to find sets X̂k,(s1 ,s2 ) such
that:
Xk,(s1 ,s2 ) ⊆ X̂k,(s1 ,s2 ) .
To compute X̂k,(s1 ,s2 ) , we take the following steps:
1) Partition the dynamics: According to (24), X̂k,(s1 ,s2 )
can be computed by:
1
2
X̂k,(s1 ,s2 ) = X̂k,(s
⊕ X̂k,(s
,
1 ,s2 )
1 ,s2 )
1
2
where ⊕ is the Minkowski addition, X̂k,(s
and X̂k,(s
1 ,s2 )
1 ,s2 )
are sets to be computed.
6
1
1
2) Compute X̂k,(s
: One can compute X̂k,(s
by:
1 ,s2 )
1 ,s2 )
1
X̂k,(s
= M (k)X̂0,(s1 ,s2 ) ,
1 ,s2 )
where X̂0,(s1 ,s2 ) is a polytope that over approximates
X0,(s1 ,s2 ) , i.e. X0,(s1 ,s2 ) ⊆ X̂0,(s1 ,s2 ) . X̂0,(s1 ,s2 ) can be
computed as in the optimization problem (1) in [2].
2
2
3) Compute X̂k,(s
: For the computation of X̂k,(s
,
1 ,s2 )
1 ,s2 )
it follows that:
2
= {x ∈ Rnx ||x| ≤ |Θ(k)|},
X̂k,(s
1 ,s2 )
where
kh
Z
eAp (kh−s) Ew(s)ds
|Θ(k)| =
0
kh
Z
eAp (kh−s) Ew(s) ds
≤
0
kh
Z
eAp (kh−s) ds|E|kwkL∞
≤
0
kh λ
max
Z
≤
e
AT
p +Ap
2
Figure 2. State-space partition and the labeling of each region.
(kh−s)
ds|E|W.
0
In which the last inequation holds according to (2.2) in [24].
of the system
Thus the reachable set X{k
s1 ,s2 ,k s1 ,• },(s1 ,s2 )
(3-7), and (9) starting from X0,(s1 ,s2 ) can be computed by:
X{k
⊆ X̂{k
s1 ,s2 ,k s1 ,• },(s1 ,s2 )
s1 ,s2 ,k s1 ,• },(s1 ,s2 )
[
=
X̂k,(s1 ,s2 ) .
k∈{ks ,s ,··· ,ks1 ,• }
1 2
To compute the transitions in S/P , one can check the
intersection between the over approximation of reachable state
set and all the state regions Rs01 ,s02 , ∀s01 ∈ {1, · · · , q1 }, s02 ∈
{1, · · · , q2 }. More specifically, one can check if the following
feasibility problem for each state region holds:
Rs01 ,s02 ∩ X̂{k
s1 ,s2 ,k s1 ,•
},(s1 ,s2 ) 6= ∅,
Figure 3. Computed result of the regional lower bound with W = 2.
in which case
Rs1 ,s2 , Rs01 ,s02 ∈−→ .
/P
D. Main result
Now we summarize the main result of the paper in the
following theorem.
=
Theorem 4.10: The metric
system S/P
X/P , X/P,0 , U/P , −→, Y/P , H/P ,
-approximately
/P
simulates S, where = max dH (y, y 0 ), y = H(x) ∈ Y ,
y 0 = H/P (x0 ) ∈ Y/P , ∀ (x, x0 ) ∈ P , and dH (·, ·) is the
Hausdorff distance.
V. N UMERICAL EXAMPLE
In this section, we consider a system employed in [13] and
[22]. The plant is given by:
˙ = 0 1 ξ(t) + 0 v(t) + 1 w(t),
ξ(t)
−2 3
1
0
and the controller is given by:
K= 1
−4 .
This plant is chosen since it is easy to show the feasibility of
the presented theory in 2-dimensional plots. The state-space
partition is presented in Figure 2.
We set W = 2, the convergence rate ρ = 0.01, L2 gain
γ = 2, sampling time h = 0.005s, event condition σ = 0.1.
By checking the LMI presented in [13], we can see there
exists a feasible solution, thus the stability and performance
can be guaranteed. The result of the computed lower bound by
Theorem 4.4 is shown in Figure 3. Figure 4 shows a zoomed-in
version. The computed upper bound by Corollary 4.8 is shown
in Figure 5. The resulting abstraction precision is = 0.15s.
The simulation results of the system evolution and event
intervals with perturbations are shown in Figure 6. The upper
bound triggered 6 events during the 10s simulation. Note
that, increasing the number of subdivisions can lead to less
7
Figure 4. Zoomed-in result of the regional lower bound with W = 2.
Figure 6. System evolution and event intervals when w = 2 sin(πt), t ∈
[3, 8]: state evaluation and perturbance, event intervals with the bounds.
Figure 5. Computed result of the regional upper bound with w = 0.
conserved lower and upper bounds of the inter event time.
The conservativeness can also be reduced by decreasing W.
The reachable state regions starting from each region is
shown in Figure 7. As an example, the reachable state region
of the initial region (s1 , s2 ) = (4, 6) is shown in Figure 8.
We also present a simulation when w = 0. The lower
bound is shown in Figure 9. The evolution of the system is
shown in Figure 10, which shows that, the inter event intervals
are within the computed bounds. The reachable state regions
starting from each region are shown in Figure 11.
VI. C ONCLUSION
In this paper, we present a construction of a power quotient
system for the traffic model of the PETC implementations
from [13]. The constructed models can be used to estimate the
next event time and the state set when the next event occurs.
These models allow to design scheduling to improve listening
time of wireless communications and medium access time to
increase the energy consumption and bandwidth occupation
efficiency.
Figure 7. Reachable regions starting from each state region, with labeling
from Figure 2.
In this paper, we consider an output feedback system with
a dynamic controller. However, the state partition is still
based on the states of the system and controller. The system
state may not always be obtainable. Therefore, to estimate
the system state in an ETC implementation from output
measurements is a very important extension to make this work
more practical. The periodic asynchronous event-triggered
control (PAETC) presented in [10] is an extension of PETC
considering quantization. One can either treat the quantization
error as part of the perturbations, or analyze this part separately
to increase the abstraction precision, since the dynamics of the
quantization error is dependent on the states. This is also an
interesting future investigation. Another interesting extension
is reconstruction of traffic models for each sensor node to
capture the local timing behavior in a decentralized PETC
8
Figure 8. Flow pipe of (s1 , s2 ) = (4, 6): indicating initial state set (red),
reachable state set (blue), and reachable regions (cyan).
Figure 10. System evolution and event intervals when w = 0: state evaluation
and event intervals vs computed bounds.
Figure 9. Computed result of the regional lower bound with w = 0.
implementation, by either global information or even only
local information.
A PPENDIX
Isotropic
covering:
Consider
x
=
T
x1 x2 · · · xn
∈ Rn . We first present a case
when x ∈ R2 . Let Θ = [− π2 , π2 [ be an interval. Splitting this
interval into q sub-intervals and Θs = [θs , θs [ be the s-th
sub-interval. Then for each sub-interval, one can construct a
cone pointing at the origin:
n
o
Rs = x ∈ R2 |xT Ξ̃s x ≥ 0 ,
where
Ξ̃s =
− sin θs sin θs 12 sin(θs + θs )
.
1
2 sin(θ s + θ s ) − cos θ s cos θ s
Remark 4.1 shows that x and −x have the same behaviours,
therefore it is sufficient to only consider half of the state-space.
Figure 11. Reachable regions starting from each conic region, with labeling
from Figure 2.
Now we derive the case when x ∈ Rn , n > 2. Define
(x)i,j = (xi , xj ) as the projection of this point on its i − j
coordinate. Now a polyhedral cone Rs can be defined as:
(
)
n−1
^
n
T
Rs = x ∈ R
(x)(i,i+1) Ξ̃s,(i,i+1) (x)(i,i+1) ≥ 0 ,
i=1
where Ξ̃s,(i,i+1) is a constructed matrix. A relation between
Ξ̃s,(i,i+1) and Ξs,(i,i+1) from (16) is given by:
[Ξs,(i,i+1) ](i,i)
=[Ξ̃s,(i,i+1) ](1,1)
=[Ξ̃s,(i,i+1) ](1,2)
[Ξs,(i,i+1) ](i,i+1)
[Ξs,(i,i+1) ](i+1,i)
[Ξs,(i,i+1) ](i+1,i+1)
[Ξs,(i,i+1) ](k,l)
=[Ξ̃s,(i,i+1) ](2,1)
=[Ξ̃s,(i,i+1) ](2,2)
=0,
9
where [M ](i,j) is the i-th row, j-th column entry of the matrix
M , k and l satisfy (k, l) 6= (i, i + 1).
Proof of Lemma 4.3: We decouple the event triggering
mechanism in (13) first:
T
M (k)x + Θ(k)
M (k)x + Θ(k)
Q
CE x
CE x
T
T
=x Φ1 (k)x + x Φ2 (k)Θ(k) + Θ
T
(k)ΦT
2 (k)x
+ ΘT (k)Q1 Θ(k)
T
≤xT (Φ1 (k) + Φ2 (k)Ψ−1 ΦT
2 (k))x + Θ (k)(Q1 + Ψ)Θ(k),
(25)
where the last inequality comes from Lemma 6.2 in [12]. Now
for the uncertainty part, we have:
ΘT (k)(Q1 + Ψ)Θ(k)
T
Θ1 (k)
(Q1 + Ψ)1
=
0
(Q1 + Ψ)3
(Q1 + Ψ)2
(Q1 + Ψ)4
Proof of Theorem 4.4: We first consider the regions with
s2 > 1. If all the hypothesis of the theorem hold, by applying
the Schur complement to (20), one has:
(29)
xT H + Φ2 (k)Ψ−1 ΦT
2 (k) x ≤ 0.
From (16), and applying the S-procedure, it holds that:
xT Φ1 (k) + Φ3 (k)W 2 Ws−2
I + Φ2 (k)Ψ−1 ΦT
2 (k) x ≤ 0.
2 −1
(30)
From (16) we also have:
xT x ≥ Ws22 −1 .
(31)
Since Φ3 (k), W, and Ws2 −1 are non-negative scalars and
Ws2 −1 > 0, we have the following inequality:
xT Φ3 (k)W 2 Ws−2
Ix = Φ3 (k)W 2 Ws−2
xT x
2 −1
2 −1
Θ1 (k)
0
=ΘT
1 (k)(Q1 + Ψ)1 Θ1 (k).
From the hypothesis of the theorem that there exists µ such
that (Q1 + Ψ)1 µI, together with Jensen’s inequality [12],
inequality (2.2) in [24], and Assumption 4.2, i.e. w ∈ L∞ ,
ΘT (k)(Q1 + Ψ)Θ(k) can be bounded from above by:
≥Φ3 (k)W 2 Ws−2
Ws22 −1 = Φ3 (k)W 2 ≥ Φ3 (k)kwk2L∞ ,
2 −1
(32)
in which the last inequality comes form the definition of W.
Now inserting (32) into (30) results in:
2
xT Φ1 (k) + Φ2 (k)Ψ−1 ΦT
2 (k) x + Φ3 (k)kwkL∞ ≤ 0,
which together with applying the Schur complement to (18)
provides the regional lower bound.
ΘT (k)(Q1 + Ψ)Θ(k) = ΘT
1 (k)(Q1 + Ψ)1 Θ1 (k)
!
!
When s2 = 1, k > 0, H will be diagonal infinity. Thus the
T Z
Z kh
kh
LMI (20) will be infeasible. According to the event-triggered
≤µ
eAp (kh−s) Ew(s)ds
eAp (kh−s) Ew(s)ds
condition (9), which indicates that tb+1 ∈ Ts and tb+1 > tb ,
0
0
the regional lower bound for those regions with s2 = 1 is h.
(by (Q1 + Ψ) µI)
This finishes the proof.
Z kh
T
Proof of Corollary 4.7: The result can be easily obtained
≤khµ
eAp (kh−s) Ew(s)
eAp (kh−s) Ew(s) ds
0
from Theorem 4.4 considering E = 0.
(by Jensen’s equality)
Proof of Corollary 4.8: The result can be easily obtained
Z kh
analogously
to Theorem 4.4 considering E = 0: if all the
T
≤khµ
e(kh−s)λmax (Ap +Ap ) wT (s)E T Ew(s)ds
hypothesis of this
Corollary hold, then according to (23),
0
Φ
(k)
0,
k
∈
k̄s1 ,• , · · · , ¯l . According to the definition of
1
(by (2.2) in [24])
Φ1 (k) in (19), for all k ≥ k̄s1 ,• , it holds that:
Z
kh (kh−s)λmax (Ap +AT )
p dskwk2
T
≤khµλmax E T E
e
L∞
M (k)x
M (k)x
0
Q
> 0,
CE x
CE x
(by w ∈ L∞ )
=khµλmax E T E dAp (k)kwk2L∞ .
which together with event condition (21) provides the regional
(26) upper bound.
With (26), (25) can be further bounded as:
Proof of Theorem 4.10: The result follows from Lemma
T
2.7
and
the construction described in this section.
M (k)x + Θ(k)
M (k)x + Θ(k)
Q
CE x
CE x
(27)
R EFERENCES
T
−1 T
2
≤x Φ1 (k) + Φ2 (k)Ψ Φ2 (k) x + Φ3 (k)kwkL∞ .
From the hypothesis of the theorem, if Φ(k) 0 holds,
then by applying the Schur complement to (18), the following
inequality holds:
2
xT Φ1 (k) + Φ2 (k)Ψ−1 ΦT
2 (k) x + Φ3 (k)kwkL∞ ≤ 0,
which indicates:
T
M (k)x + Θ(k)
M (k)x + Θ(k)
Q
≤ 0.
CE x
CE x
(28)
Therefore, k x generated by (13) is lower bounded by kx0
generated by (17). This ends the proof.
[1] Alongkrit Chutinan and Bruce H Krogh. Computing polyhedral approximations to flow pipes for dynamic systems. In Decision and Control,
1998. Proceedings of the 37th IEEE Conference on, volume 2, pages
2089–2094. IEEE, 1998.
[2] Alongkrit Chutinan and Bruce H Krogh. Computational techniques for
hybrid system verification. IEEE transactions on automatic control,
48(1):64–75, 2003.
[3] Marieke BG Cloosterman, Laurentiu Hetel, Nathan Van de Wouw,
WPMH Heemels, Jamal Daafouz, and Henk Nijmeijer. Controller
synthesis for networked control systems. Automatica, 46(10):1584–
1594, 2010.
[4] Marieke BG Cloosterman, Nathan Van de Wouw, WPMH Heemels,
and Hendrik Nijmeijer. Stability of networked control systems with
uncertain time-varying delays. IEEE Transactions on Automatic Control,
54(7):1575–1580, 2009.
10
[5] MCF Donkers and WPMH Heemels. Output-based event-triggered
control with guaranteed-gain and improved and decentralized eventtriggering. Automatic Control, IEEE Transactions on, 57(6):1362–1376,
2012.
[6] MCF Donkers, WPMH Heemels, Nathan Van de Wouw, and Laurentiu
Hetel. Stability analysis of networked control systems using a switched
linear systems approach. IEEE Transactions on Automatic control,
56(9):2101–2115, 2011.
[7] Günter Ewald. Combinatorial convexity and algebraic geometry, volume
168. Springer Science & Business Media, 2012.
[8] Christophe Fiter, Laurentiu Hetel, Wilfrid Perruquetti, and Jean-Pierre
Richard. A state dependent sampling for linear state feedback. Automatica, 48(8):1860–1867, 2012.
[9] Christophe Fiter, Laurentiu Hetel, Wilfrid Perruquetti, and Jean-Pierre
Richard. A robust stability framework for lti systems with time-varying
sampling. Automatica, 54:56–64, 2015.
[10] Anqi Fu and Manuel Mazo Jr. Periodic asynchronous event-triggered
control. CoRR, abs/1703.10073, 2017.
[11] Rob H Gielen, Sorin Olaru, Mircea Lazar, WPMH Heemels, Nathan
van de Wouw, and S-I Niculescu. On polytopic inclusions as a
modeling framework for systems with time-varying delays. Automatica,
46(3):615–619, 2010.
[12] Keqin Gu, Jie Chen, and Vladimir L Kharitonov. Stability of time-delay
systems. Springer Science & Business Media, 2003.
[13] WPMH Heemels, MCF Donkers, and Andrew R Teel. Periodic eventtriggered control for linear systems. Automatic Control, IEEE Transactions on, 58(4):847–861, 2013.
[14] Laurentiu Hetel, Jamal Daafouz, and Claude Iung. Stabilization of
arbitrary switched linear systems with unknown time-varying delays.
IEEE Transactions on Automatic Control, 51(10):1668–1674, 2006.
[15] Arman Sharifi Kolarijani, Dieky Adzkiya, and Manuel Mazo. Symbolic
abstractions for the scheduling of event-triggered control systems. In
Decision and Control (CDC), 2015 IEEE 54th Annual Conference on,
pages 6153–6158. IEEE, 2015.
[16] Arman Sharifi Kolarijani and Manuel Mazo Jr. A formal traffic characterization of lti event-triggered control systems. IEEE Transactions on
Control of Network Systems, 2016.
[17] Manuel Mazo Jr. and Ming Cao. Asynchronous decentralized eventtriggered control. Automatica, 50(12):3197–3203, 2014.
[18] Manuel Mazo Jr and Anqi Fu. Decentralized event-triggered controller
implementations. Event-Based Control and Signal Processing, page 121,
2015.
[19] Manuel Mazo Jr. and Paulo Tabuada. Decentralized event-triggered
control over wireless sensor/actuator networks. Automatic Control, IEEE
Transactions on, 56(10):2456–2461, 2011.
[20] Joëlle Skaf and Stephen Boyd. Analysis and synthesis of state-feedback
controllers with timing jitter. IEEE Transactions on Automatic Control,
54(3):652–657, 2009.
[21] Young Soo Suh. Stability and stabilization of nonuniform sampling
systems. Automatica, 44(12):3222–3226, 2008.
[22] Paulo Tabuada. Event-triggered real-time scheduling of stabilizing
control tasks. Automatic Control, IEEE Transactions on, 52(9):1680–
1685, 2007.
[23] Paulo Tabuada. Verification and control of hybrid systems: a symbolic
approach. Springer Science & Business Media, 2009.
[24] Charles Van Loan. The sensitivity of the matrix exponential. SIAM
Journal on Numerical Analysis, 14(6):971–981, 1977.
[25] Xiaofeng Wang and Michael D Lemmon. Event-triggering in distributed
networked control systems. Automatic Control, IEEE Transactions on,
56(3):586–601, 2011.
[26] Xiaofeng Wang and Michael D Lemmon. On event design in eventtriggered feedback systems. Automatica, 47(10):2319–2322, 2011.
| 3cs.SY
|
Derandomization for k-submodular
maximization
arXiv:1610.07729v2 [cs.DS] 15 Feb 2017
Hiroki Oshima
Department of Mathematical Informatics,
Graduate School of Information Science and Technology, The University of Tokyo
Abstract. Submodularity is one of the most important property of combinatorial optimization, and k-submodularity is a generalization of submodularity. Maximization of a k-submodular function is NP-hard, and
approximation algorithm has been studied. For monotone k-submodular
functions, [Iwata, Tanigawa, and Yoshida 2016] gave k/(2k−1)-approximation
algorithm. In this paper, we give a deterministic algorithm by derandomizing that algorithm. Our algorithm is k/(2k−1)-approximation and runs
in polynomial time.
1
Introduction
A set function f : 2V → R is submodular if, for any A, B ⊆ V , f (A) + f (B) ≥
f (A ∪ B) + f (A ∩ B). Submodularity is one of the most important properties
of combinatorial optimization. The rank functions of matroids and cut capacity
functions of networks are submodular. Submodular functions can be seen as
discrete version of convex functions.
For submodular function minimization, [4] showed the first polynomial-time
algorithm. The combinatorial strongly polynomial algorithms were shown by [5]
and [8]. On the other hand, submodular function maximization is NP-hard and
we can only use approximation algorithms. Let an input function for maximization be f , a maximizer of f be S ∗ , and an output of an algorithm be S. The
approximation ratio of the algorithm is defined as f (S)/f (S ∗ ) for deterministic
algorithms and E[f (S)]/f (S ∗ ) for randomized algorithms. A randomized version of Double Greedy algorithms in [2] achieves 1/2-approximation. [3] showed
(1/2 + ǫ)-approximation requires exponential time value oracle queries. This implies that, Double Greedy algorithm is one of the best algorithms in terms of the
approximation ratio. [1] showed a derandomized version of randomized Double
Greedy algorithm, and their algorithm achieves 1/2-approximation.
k-submodularity is an extension of submodularity. k-submodular function is
defined as below.
Definition 1. Let (k + 1)V := {(X1 , ..., Xk ) | Xi ⊆ V (i = 1, ..., k), Xi ∩ Xj =
∅ (i 6= j)}. A function f : (k + 1)V → R is called k-submodular if we have
f (x) + f (y) ≥ f (x ⊓ y) + f (x ⊔ y)
for any x = (X1 , ..., Xk ), y = (Y1 , ..., Yk ) ∈ (k + 1)V . Note that
x ⊓ y = (X1 ∩ Y1 , ..., Xk ∩ Yk ) and
[
[
x ⊔ y = (X1 ∪ Y1 \ (Xi ∪ Yi ), ..., Xk ∪ Yk \ (Xi ∪ Yi )).
i6=1
i6=k
It is a submodular function if k = 1. It is called a bisubmodular function if k = 2.
Maximization for k-submodular functions is also NP-hard and approximation algorithm have been studied. An input of the problem is a nonnegative
k-submodular function. Note that, for any k-submodular function f and any
c ∈ R, a function f ′ (x) := f (x) + c is k-submodular. An output of the problem is x = (X1 , ..., Xk ) ∈ (k + 1)V . Let an input k-submodular function be f , a
maximizer of f be o, and an output of an algorithm be s. Then we define the approximation ratio of the algorithm as f (s)/f (o) for deterministic algorithms, and
E[f (s)]/f (o) for randomized algorithms. For bisubmodular functions, [6] and [9]
showed that the algorithm for submodular functions in [2] can be extended. [9]
analyzed an extension for k-submodular functions. They
p showed a randomized
1/(1 + a)-approximation algorithm with a = max{1, (k − 1)/4} and a deterministic 1/3-approximation algorithm. Now we have a 1/2-approximation algorithm shown in [7]. In particular, for monotone k-submodular functions, [7] gave
k
-approximation algorithm. They also showed any ( k+1
a 2k−1
2k +ǫ)-approximation
algorithm requires exponential time value oracle queries.
In this paper, we give a deterministic approximation algorithm for monok
-approximation and runs in
tone k-submodular maximization. It satisfies 2k−1
polynomial-time. Our algorithm is a derandomized version of algorithm for
monotone functions in [7]. We also note the derandomization scheme is from
[1], used for Double Greedy algorithm.
2
Preliminary
Define a partial order on (k + 1)V for x = (X1 , ..., Xk ) and y = (Y1 , ..., Yk ) as
follows:
def
x y ⇐⇒ Xi ⊆ Yi (i = 1, ..., k).
Sk
Also, for x = (X1 , ..., Xk ) ∈ (k + 1)V , e ∈
/ l=1 Xl , and i ∈ {1, ..., k}, define
∆e,i f (x) = f (X1 , ..., Xi−1 , Xi ∪ {e}, Xi+1 , ..., Xk ) − f (X1 , ..., Xk ).
A monotone k-submodular function is k-submodular and satisfies f (x) ≤
f (y) for any x = (X1 , ..., Xk ) and y = (Y1 , ..., Yk ) in (k + 1)V with x y.
The property of k-submodularity can be written as another form.
Theorem 1. ([9] THEOREM 7) A function f : (k + 1)V → R is k-submodular
if and only if f is orthant submodular and pairwise monotone.
Note that orthant submodularity is to satisfy
∆e,i f (x) ≥ ∆e,i f (y) (x, y ∈ (k + 1)V , x y, e ∈
/
k
[
l=1
Yl , i ∈ {1, ..., k}),
and pairwise monotonicity is to satisfy
∆e,i f (x) + ∆e,j f (x) ≥ 0 (x ∈ (k + 1)V , e ∈
/
k
[
Xl , i, j ∈ {1, ..., k} (i 6= j)).
l=1
To analyze k-submodular functions, it is often convenient to identify (k + 1)V
as {0, 1, ..., k}V . A |V |-dimensional vector x ∈ {0, 1, ..., k}V is associated with
(X1 , ..., Xn ) ∈ (k + 1)V by Xi = {e ∈ V | x(e) = i}.
3
Existing randomized algorithms
3.1
Algorithm framework
In this section, we see the framework to maximize k-submodular functions (Algorithm 1 [7]). [6] and [9] used it with specific distributions.
Algorithm 1 ([7] Algorithm 1)
Input: A nonnegative k-submodular function f : {0, 1, ..., k}V → R+ .
Output: A vector s ∈ {0, 1, ..., k}V .
s←0.
Denote the elements of V by e(1) , ..., e(n) (|V | = n).
for j = 1, ..., n do
Set a probability distribution p(j) over {1, ..., k}.
(j)
Let s(e(j) ) ∈ {1, ..., k} be chosen randomly with Pr[s(e(j) ) = i] = pi .
end for
return s
Algorithm 1 is not only used for monotone functions. However, in this paper,
we only use it for monotone functions.
Now we define some variables to see Algorithm 1. Let o be an optimal solution, and we write s(j) as s at j-th iteration. Let other variables be as follows:
(
o(j) (e) (e 6= e(j) )
(j)
(j)
(j)
(j−1)
o = (o ⊔ s ) ⊔ s , t
(e) =
0
(e = e(j) )
(j)
yi
(j)
= ∆e(j) ,i f (s(j−1) ) , ai
= ∆e(j) ,i f (t(j−1) )
Algorithm 1 satisfies following lemma.
Lemma 1. ([7] LEMMA 2.1.)
Let c ∈ R+ . Conditioning on s(j−1) , suppose that
k
X
(j)
(j)
(j)
(ai∗ − ai )pi
≤c
k
X
(j) (j)
(yi pi )
i=1
i=1
holds for each j with 1 ≤ j ≤ n, where i = o(e(j) ). Then E[f (s)] ≥
∗
1
1+c f (o).
3.2
A randomized algorithm for monotone functions
k
In [7], a randomized 2k−1
-approximation algorithm for monotone k-submodular
functions (Algorithm 2) is shown.
Algorithm 2 ([7] Algorithm 3)
Input: A monotone k-submodular function f : {0, 1, ..., k}V → R+ .
Output: A vector s ∈ {0, 1, ..., k}V .
s ← 0, t ← k − 1.
Denote the elements of V by e(1) , ..., e(n) (|V | = n).
for j = 1, ..., n do
(j)
yi ← ∆e(j) ,i f (s) (1 ≤ i ≤ k).
P
(j)
β ← ki=1 (yi )t .
(j)
(j)
if β 6= 0 then pi ← (yi )t /β (1 ≤ i ≤ k).
else
(j)
(j)
p1 = 1, pi = 0 (i = 2, ..., k).
end if
(j)
Lets(e(j) ) ∈ {1, ..., k} be chosen randomly with Pr[s(e(j) ) = i] = pi .
end for
return s
Algorithm 2 runs in polynomial time. The approximation ratio of Algorithm
2 satisfies the theorem below.
Theorem 2. ([7] THEOREM 2.2.) Let o be a maximizer of a monotone ksubmodular function f and let s be the output of Algorithm 2. Then E[f (s)] ≥
k
2k−1 f (o).
In the proof of this theorem (see [7]), the inequality of Lemma 1 is proved
with c = 1 − k1 . We get ai ≥ 0 (∀i ∈ {1, ..., k}) from monotonicity, and ai ≤
yi (∀i ∈ {1, ..., k}) from orthant submodularity. Hence, the inequality
X
i6=i∗
(j) (j)
(yi∗ pi )
k
1 X (j) (j)
(y p )
≤ 1−
k i=1 i i
(1)
is used. The inequality of Lemma 1 is satisfied when the inequality (1) is valid.
4
Deterministic algorithm
In this section, we give a polynomial-time deterministic algorithm for maximizing
monotone k-submodular functions. Our algorithm is Algorithm 3. Algorithm 3
is a derandomized version of Algorithm 2. We note the derandomization scheme
of this algorithm is from [1].
Algorithm 3 A deterministic algorithm
Input: A monotone k-submodular function f : {0, 1, ..., k}V → R+ .
Output: A vector s ∈ {0, 1, ..., k}V .
P
D0 ← (1, 0), (D = {(p, s) | s ∈ (k + 1)V , 0 ≤ p ≤ 1} ( s∈D p = 1)).
(1)
(n)
Denote the elements of V by e , ..., e
(|V | = n).
for j = 1, ..., n do
yi (s) ← ∆e(j) ,i f (s) (∀s ∈ supp(Dj−1 ), i ∈ {1, ..., k}).
Find an extreme point solution (pi,s )i=1,...,k, s∈supp(Dj−1 ) of the following linear
formulation:
" k
#
X
1
1−
Es∼Dj−1
(2)
pi,s yi (s) ≥ Es∼Dj−1 [(1 − pl,s )yl (s)]
k
i=1
(l ∈ {1, ..., k})
k
X
pi,s = 1 (∀s ∈ supp(Dj−1 ))
(3)
pi,s ≥ 0 (∀s ∈ supp(Dj−1 ), i ∈ {1, ..., k}).
(4)
i=1
Construct a new distribution Dj :
Dj ←
k
[
{(pi,s · PrDj−1 [s], se(j) ,i ) | s ∈ supp(Dj−1 ), pi,s > 0}
i=1
se(j) ,i (e) =
end for
return arg maxs∈supp(Dn ) {f (s)}
(
s(e)
i
(e 6= e(j) )
(e = e(j) )
!
.
(5)
In the algorithm, we construct a distribution D which satisfies Es∼D [f (s)] ≥
Then the algorithm outputs the best solution in supp(D) := {s |
(p, s) ∈ D}. We can see the right hand side of (2) in Algorithm 3 is the expected value of the left hand side of (1) for s ∼ Dj−1 with i∗ = l. it is because
P
i6=l pi,s yl (s) = (1 − pl,s )yl (s). Also the left hand side of (2) is the expected
value of the right hand side of (1) with c = 1 − 1/k. From (3) and (4), Dj in (5)
is constructed as a distribution.
k
2k−1 f (o).
Algorithm 3 achieves the same approximation ratio as Algorithm 2.
Theorem 3. Let o be a maximizer of a monotone nonnegative k-submodular
k
f (o).
function f and let z be the output of Algorithm 3. Then f (z) ≥ 2k−1
Proof. We consider the j-th iteration. From (5), we get
Es∼Dj−1
" k
X
i=1
#
pi,s yi (s) = Es∼Dj−1
= Es∼Dj−1
"
"
k
X
#
pi,s (f (se(j) ,i ) − f (s))
i=1
k
X
#
pi,s f (se(j) ,i ) − f (s)
i=1
′
= Es′ ∼Dj [f (s )] − Es∼Dj−1 [f (s)] .
(6)
Now, we consider o[s] := (o ⊔ s) ⊔ s. Define the variables as follows:
r(e) =
(
o[s](e) (e 6= e(j) )
0
(e = e(j) )
ai (s) = ∆e(j) ,i f (r)
Then we have
f (o[s]) − f (o[se(j) ,i ]) = ai∗ (s) − ai (s) (i∗ = o(e(j) ))
(7)
From monotonicity and orthant submodularity of f , we have
ai∗ (s) − ai (s) ≤ yi∗ (s).
(8)
From (7) and (8), we get
Es∼Dj−1 [f (o[s])] − Es′ ∼Dj [f (o[s′ ])] = Es∼Dj−1
= Es∼Dj−1
"
"
k
X
#
pi,s f (o[s]) − f (o[se(j) ,i ])
i=1
k
X
pi,s
i=1
= Es∼Dj−1
≤ Es∼Dj−1
#
f (o[s]) − f (o[se(j) ,i ])
X
pi,s (ai∗ (s) − ai (s))
X
pi,s (yi∗ (s))
i6=i∗
i6=i∗
= Es∼Dj−1 [(1 − pi∗ ,s ) (yi∗ (s))] .
(9)
pi,s satisfies (2) for all l ∈ {1, 2, ..., k}. Hence we obtain
1
Es′ ∼Dj [f (s′ )] − Es∼Dj−1 [f (s)] ≥ Es∼Dj−1 [f (o[s])]−Es′ ∼Dj [f (o[s′ ])]
1−
k
(10)
from (6) and (9). By the summation of (10), we get
1
(Es′ ∼Dn [f (s′ )] − Es∼D0 [f (s)]) ≥ Es∼D0 [f (o[s])] − Es′ ∼Dn [f (o[s′ ])] .
1−
k
(11)
Note that o[s′ ] = s′ for s′ ∈ supp(Dn ), and o[s] = o for s ∈ supp(D0 ). Now we
have
1
1
′
′
Es ∼Dn [f (s )] − 1 −
f (0)
f (o) ≤ 2 −
k
k
1
≤ 2−
Es′ ∼Dn [f (s′ )]
k
1
≤ 2−
max {f (s′ )}
k s′ ∈supp(Dn )
The algorithm performs a polynomial number of value oracle queries.
Theorem 4. Algorithm 3 returns a solution after O(n2 k 2 ) value oracle queries.
Proof. Algorithm 3 uses the value oracle to caluculate yi (s). At j-th iteration,
the number of yi (s) is k|Dj−1 |. From (5), |Dj | equals the number of pi,s 6= 0.
Then we have to consider pi,s 6= 0 at j-th iteration.
By the definition, (pi,s )i=1,...,k, s∈supp(Dj−1 ) is an extreme point solution of
(2), (3), and (4). Note that, we can get a solution by setting (pi,s ) as the distribution of Algorithm 2 for each s ∈ supp(Dj−1 ). We can also see the feasible
region of (2), (3), and (4) is bounded. Then some extreme point solution exists.
Let |Dj−1 | = m. By (pi,s )i=1,...,k, s∈supp(Dj−1 ) ∈ Rkm and k equalities of
(3), km − k inequalities are tight at any extreme point solution. (2) have m
inequalities and (4) have km inequalities. Then, at least km − k − m inequalities
of (4) are tight. Hence, the number of pi,s 6= 0 is at most m + k.
Now we have |Dj | ≤ |Dj−1 | + k. We can also see |Dj | ≤ jk + 1. Then the
number of value oracle queries is
n
X
j=1
k|Dj−1 | ≤
n
X
k(jk + 1).
j=1
In our algorithm, we have to search for an extreme point solution. We can do
it by solving LP for some objective function. If we use LP for our algorithm, it
is polynomial-time not only for the number of queries but also for the number of
operations. The simplex method is not proved to be a polynomial-time method.
However, it is practical. Our algorithm needs only an extreme point solution,
then if we get a basic solution, it is enough. So we can use the first phase of
two-phase simplex method to find an extreme point solution.
5
Conclusion
We showed a derandomized algorithm for monotone k-submodular maximizak
tion. It is 2k−1
-approximation and polynomial-time algorithm.
One of open problems is a faster method for finding an extreme point solution
of the linear formulation. For submodular functions, [1] showed greedy methods
are effective. It is because their formulation is the form of fractional knapsack
problem. Our formulation is similar to theirs, and ours can be seen as the form of
an LP relaxation of multidimensional knapsack problem. However, faster methods are not given than general LP solutions. The number of constraints in our
formulation depends on k and the number of iterations. It is therefore difficult
to find an extreme point faster.
Constructing a deterministic algorithm for nonmonotone functions is also an
important open problem. For nonmonotone functions, we have pairwise monotonicity instead of monotonicity. In such a situation, for some i, ai can be negative. However, if yj > 0 for all j, we can’t find such i. Then, if we try to
use the same derandomizing method, the number of constraints in the linear
formulation and the size of D will be exponential. So algorithm can’t finish in
polynomial-time.
References
1. N. Buchbinder and M. Feldman. Deterministic algorithms for submodular maximization problems. In Proceedings of the Twenty-Seventh Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 392–403. SIAM, 2016.
2. N. Buchbinder, M. Feldman, J. Seffi, and R. Schwartz. A tight linear time (1/2)approximation for unconstrained submodular maximization. SIAM Journal on
Computing, 44(5):1384–1402, 2015.
3. U. Feige, V. S. Mirrokni, and J. Vondrak. Maximizing non-monotone submodular
functions. SIAM Journal on Computing, 40(4):1133–1153, 2011.
4. M. Grötschel, L. Lovász, and A. Schrijver. The ellipsoid method and its consequences
in combinatorial optimization. Combinatorica, 1(2):169–197, 1981.
5. S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial algorithm for minimizing submodular functions. Journal of the ACM, 48(4):761–777,
2001.
6. S. Iwata, S. Tanigawa, and Y. Yoshida. Bisubmodular function maximization and
extensions. Technical report, Technical Report METR 2013-16, The University of
Tokyo, 2013.
7. S. Iwata, S. Tanigawa, and Y. Yoshida. Improved approximation algorithms for ksubmodular function maximization. In Proceedings of the Twenty-Seventh Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 404–413. SIAM, 2016.
8. A. Schrijver. A combinatorial algorithm minimizing submodular functions in
strongly polynomial time. Journal of Combinatorial Theory, Series B, 80(2):346–
355, 2000.
9. J. Ward and S. Živný. Maximizing k-submodular functions and beyond. ACM
Trans. Algorithms, 12(4):47:1–47:26, August 2016.
| 8cs.DS
|
1
Graph Distances and Controllability of Networks
arXiv:1608.04309v1 [cs.SY] 15 Aug 2016
A. Yasin Yazıcıoğlu, Waseem Abbas, and Magnus Egerstedt
Abstract—In this technical note, we study the controllability of
diffusively coupled networks from a graph theoretic perspective.
We consider leader-follower networks, where the external control
inputs are injected to only some of the agents, namely the leaders.
Our main result relates the controllability of such systems to the
graph distances between the agents. More specifically, we present
a graph topological lower bound on the rank of the controllability
matrix. This lower bound is tight, and it is applicable to systems
with arbitrary network topologies, coupling weights, and number
of leaders. An algorithm for computing the lower bound is also
provided. Furthermore, as a prominent application, we present
how the proposed bound can be utilized to select a minimal set
of leaders for achieving controllability, even when the coupling
weights are unknown.
I. I NTRODUCTION
Networks of diffusively coupled agents appear in numerous
systems such as sensor networks (e.g., [1]), distributed robotics
(e.g., [2]), power grids (e.g., [3]), social networks (e.g., [4]),
and biological systems (e.g., [5]). A central question regarding
such networks is whether a desired global behavior can be
induced by directly manipulating only a small subset of the
agents, referred to as the “leaders” in the network. This
question has motivated numerous studies on the controllability
of networks. In particular, there has been a large interest in
relating the network controllability to the structure of the interaction graph. In this technical note, we relate the controllability
of diffusively coupled agents with single integrator dynamics
to the distances between the nodes on the interaction graph.
Various graph theoretic tools have recently been utilized
to provide some topology based characterizations of network
controllability. Some of the graph theoretic constructs that are
widely employed for this purpose include equitable partitions
(e.g., [6], [7]), maximum matching (e.g., [8], [11]), centrality
based measures (e.g., [13], [14]), and dominating sets (e.g.,
[15]). Recently, the graph distances have been used to acquire further insight on how the network structure and the
locations of the leaders influence the network controllability
[16], [17], [18]. The graph distances, which rely purely on the
shortest paths between the nodes, provide a computationally
tractable and perceptible characterization of the graph structure. Thus, the distance-based relationships reveal some innate
connections between the network topology and the network
controllability.
The main contribution of this technical note is a distancebased tight lower bound on the dimension of the controllable
A. Yasin Yazıcıoğlu is with the Laboratory for Information and Decision
Systems, Massachusetts Institute of Technology, [email protected].
Waseem Abbas is with the Institute for Software Integrated Systems,
Vanderbilt University, [email protected]
Magnus Egerstedt is with the School of Electrical and Computer Engineering, Georgia Institute of Technology, [email protected].
The paper was presented in part at the 51st IEEE Conference on Decision
and Control, Maui, HI, December 10-13, 2012 (see [18]).
subspace (Theroem 3.2). The bound is generic in the sense
that it is applicable to systems with arbitrary network topologies, coupling weights, and number of leaders. Based on the
distances between the leaders and the followers, we first define
the distance-to-leaders (DL) vectors. Then, we define a certain
ordering rule and derive the lower bound as the maximum
length of the sequences of the DL vectors that satisfy the rule
(Section III). An algorithm to compute the proposed bound is
also presented in Section IV.
A prominent attribute of the proposed bound is that, unlike
the dimension of the controllable subspace, it does not depend
on the coupling weights. Thus, the bound is useful in many
applications, particularly when the information about the network is incomplete. As an example, in Section V, we present
how the bound can be used in leader selection for achieving the
controllability of any given network, even when the coupling
weights are unknown. Finally, some conclusions are provided
in Section VI.
II. P RELIMINARIES
A. Graph Theory
A graph G = (V, E) consists of a node set
V = {1, 2, . . . , n} and an edge set E ⊆ V × V . For an
undirected graph, each edge is represented as an unordered
pair of nodes. For each i ∈ V , let Ni denote the neighborhood
of i, i.e.,
Ni = {j ∈ V | (i, j) ∈ E}
(1)
A path between a pair of nodes i, j ∈ V is a sequence of
nodes {i, . . . , j} such that each pair of consecutive nodes are
linked by an edge. The distance between the nodes, dist(i, j),
is equal to the number of the edges that belong to the shortest
path between the nodes. A graph is connected if there exists
a path between every pair of nodes. A graph is weighted if
there is a corresponding weighting function w : E 7→ R+ . The
adjacency matrix, A, of a weighted graph is defined as
w(i, j) if (i, j) ∈ E
Aij =
(2)
0
otherwise.
For any adjacency matrix A, the corresponding degree matrix,
∆, is defined as
P
if i = j
k∈Ni Aik
∆ij =
(3)
0
otherwise,
The graph Laplacian, L, is defined as the difference of the
degree matrix and the adjacency matrix, i.e.,
L = ∆ − A.
(4)
2
B. Leader-Follower Networks
A network of diffusively coupled agents can be represented
as a graph, where the nodes correspond to the agents, and the
weighted edges exist between the coupled agents. For such a
network G = (V, E), let the dynamics of each agent i ∈ V be
X
w(i, j)(xj − xi ),
(5)
ẋi =
j∈Ni
where xi denotes the state of i, and w(i, j) ∈ R+ represents
the strength of the coupling between i and j.
In a leader-follower setting, the objective is to drive the
overall system by injecting external control inputs to some of
the nodes, which are called the leaders. The set of leaders
can be represented as L = {l1 , . . . , lm } ⊆ V , where, without
loss of generality, the leaders are labeled such that lj < lj+1 .
For any leader-follower network, a global state vector x can
be obtained by stacking the states of all the nodes. Without
loss of generality, let x ∈ Rn , and let u ∈ Rm be the control
input injected to the leaders. Then, the overall dynamics of
the system can be expressed as
ẋ = −Lx + Bu,
where B is an n × m matrix with the following entries
1 if i = lj .
Bij =
0 otherwise.
(6)
(7)
For the system in (6), the controllable subspace consists of
the states that can be reached from x(0) = 0 in any finite time
via an appropriate choice of u(t). The controllable subspace
is the range space of the controllability matrix, i.e.,
Γ = B (−L)B (−L)2 B . . . (−L)n−1 B . (8)
III. L EADER -F OLLOWER D ISTANCES AND
C ONTROLLABILITY
In this section, we present a connection between the controllability of networks and the distances of the nodes to the
leaders on the interaction graph. More specifically, we utilize
such distances to define a tight lower bound on the dimension
of the controllable subspace, i.e. the rank of the controllability
matrix. First, we provide some definitions prior to our analysis.
Definition 3.1 (Distance-to-Leaders (DL) Vector): For each
node i in a network with m leaders, the DL vector di ∈ Rm
is defined as
di,j = dist(i, lj ),
(9)
where di,j denotes the j th entry of di , and lj denotes the j th
leader for j ∈ {1, 2, . . . , m}.
In our analysis, we utilize a specific sequence, which we define
as a pseudo-monotonically increasing sequence, of the DL
vectors. For any vector sequence D, let Di be the ith vector
in the sequence, and let Di,j denote the j th entry of Di .
Definition 3.2 (Pseudo-Monotonically Increasing (PMI) Sequence): A sequence D of vectors, where each vector is in
Rm , is a PMI sequence if for every Di there exists some
α(i) ∈ {1, 2, · · · , m} such that
Di,α(i) < Dj,α(i) , ∀i < j.
(10)
Condition (10) simply means that if Di is the ith vector in
the sequence with Di,α(i) being its α(i)th entry, then the
corresponding (i.e. α(i)th ) entries of all the subsequent vectors
in the sequence should be greater than Di,α(i) .
Example – Consider the network shown in Fig. 1. For this
network, one can build a PMI sequence of five DL vectors as
D = (D1 , D2 , D3 , D4 , D5 ) = (d1 , d6 , d5 , d3 , d4 )
2
3
0
1
2
,
,
,
,
,
=
0
3
1
2
2
where, for each Di , the α(i)th entry that satisfies (10) is
encircled. For instance, consider D1 , for which α(1) = 1
and D1,1 = 0. Note that the first entries of all the subsequent
vectors are greater than 0. Similarly, for the second vector D2 ,
α(2) = 2 and D2,2 = 0. Note that Dj,2 > 0 for j > 2, and so
on. For this example, another PMI sequence of five DL vectors
could be build as (d1 , d3 , d6 , d5 , d4 ), where α(1) = α(2) = 1
and α(3) = α(4) = α(5) = 2.
h
h
h
0
3
i
1
1
3
i
2
2
2
i
4
1
2
3
h
i
2
1
5
h
i
6
h
3
0
i
Fig. 1. A leader-follower network with two leaders (shown in gray), l1 = 1
and l2 = 6. Each node i has its DL vector di given next to itself.
For any connected G = (V, E) and any set of leaders
L ⊆ V , let DL denote the set of all PMI sequences of the
corresponding DL vectors. Furthermore, let δL denote the
length of the longest sequence in DL , i.e.
δL = max |D|.
(11)
D∈DL
In the following analysis, we show that, for any weighting
function w : E 7→ R+ , the rank of the resulting controllability
matrix is lower bounded by δL .
Lemma 3.1. Let G = (V, E) be a connected graph. Then, for
any weighting function w : E 7→ R+ ,
= 0 if 0 ≤ r < dist(i, j),
r
[(−L) ]ij
(12)
6= 0 if r = dist(i, j),
where dist(i, j) is the distance between i and j on G.
Proof: Using (4), (−L)r can be expanded as
(−L)r = (A − ∆)r = Ar +
r−1
X
(−1)r−m Sm ,
(13)
m=0
where Sm denotes the sum of all matrices that can be
expressed as a multiplication of m copies of A and r − m
copies of ∆. For instance, if r = 2, then S0 = ∆2 and
S1 = A∆ + ∆A. Note that, for w : E 7→ R+ , any matrix
3
that can be represented as such a multiplication has only nonnegative entries since ∆ and A have only non-negative entries.
Moreover, ∆ is a diagonal matrix and, for any connected
graph, it has only positive entries on the main diagonal. As
such, it does not alter the signs of entries when multiplied by
a matrix. Hence,
[Sm ]ij = 0 ⇔ [Am ]ij = 0.
(14)
Using (13),
[(−L)r ]ij = [Ar ]ij +
r−1
X
(−1)r−m [Sm ]ij .
(15)
m=0
Since A is the adjacency matrix of a weighted graph with
positive edge weights, [Ak ]ij is equal to a positive scalar times
the number of walks of length k from node i to node j. Hence,
[Ak ]ij = 0 for all 0 ≤ k < dist(i, j), and [Adist(i,j) ]ij 6= 0
for any connected graph. Consequently, (14) and (15) together
imply (12).
Theorem 3.2. Let G = (V, E) be a connected graph, and let
L ⊆ V be the set of leaders. Then, for any weighting function
w : E 7→ R+ , the controllability matrix Γ satisfies
rank(Γ) ≥ δL ,
(16)
where δL is defined in (11).
Proof: For any connected G = (V, E) and any set
of leaders L, let {d1 , d2 , · · · , dn } be the corresponding DL
vectors. Let D = (D1 , D2 , · · · , DδL ) be a PMI sequence of
maximum length. Now, consider vectors of the form
(−L)rp bα(p) ,
(17)
where α(p) is the index of Dp as per the definition of a PMI
sequence (Definition 3.2), rp = Dp,α(p) , and bα(p) denotes
the α(p)th column of the input matrix B. If Dp corresponds
to the DL vector of node i, i.e. Dp = di , then rp = di,α(p) .
As a result of Lemma 3.1, ith entry of the vector in (17) is
non-zero. Also, for any node j with dj,α(i) > di,α(i) , the j th
entry of the vector in (17) equal to zero. Using this along with
the definition of PMI sequences, we conclude that the n × δL
matrix
(−L)r1 bα(1) (−L)r2 bα(2) . . . (−L)rδL bα(δL ) ,
(18)
has a full column rank since each column contains the leftmost non-zero entry in some rows. Note that for every
p ∈ {1, 2, . . . , δL }, we have rp = Dp,α(p) ≤ n − 1 since
the distance between any two nodes is always smaller than or
equal to n − 1. Hence, each column of the matrix in (18) is
also a column of Γ. Consequently, rank(Γ) ≥ δL .
Since (16) holds for any weighting function w : E 7→ R+ ,
the proposed lower bound is closely related to the notion
of strong structural controllability. A network is said to be
structurally controllable if and only if there exists w : E 7→ R+
such that the resulting controllability matrix is of full rank
[9]. Furthermore, a network is said to be strongly structurally
controllable if and only if for any w : E 7→ R+ , the resulting
controllability matrix is of full rank [10]. In this regard, the
dimension of controllable subspace in the sense of strong
structural controllability can be defined as the minimum of the
ranks of all controllability matrices obtained by using arbitrary
weighting functions w : E 7→ R+ [12]. Accordingly, δL is
essentially a lower bound on the dimension of controllable
subspace in the sense of strong structural controllability.
We also emphasize that the lower bound in Theorem 3.2 is
tight, i.e., there exist G = (V, E), L ⊆ V , and w : E 7→ R+
such that rank(Γ) = δL . Any cycle graph with two adjacent
nodes being the leaders, or any path graph with a leaf node
being the leader are some of the examples that satisfy (16) with
equality. In these two examples, rank(Γ) = δL follows from
the fact that both cases lead to δL = n, where n is the total
number of nodes in the graph. In the remainder of this section,
we present the connections between the proposed lower bound
and some closely related distance-based measures, namely the
maximum distance to the leaders and the number of unique
DL vectors.
For networks with a single leader, any PMI sequence of
the DL vectors consists of one dimensional vectors with
monotonically increasing entries. Hence, for such networks,
δL is equal to one plus the maximum distance to the leader,
which was indeed proposed in [16] as a lower bound of the
controllability matrix for single-leader networks. An extension
to the case of multiple leaders was later presented in [17] by
taking the maximum of this value among all the leaders, i.e.
µL = max dist(i, j) + 1.
i∈V,j∈L
(19)
The relationship between µL and δL can be seen through
the following observation: If one considers only the PMI
sequences that satisfy (10) for some fixed entry, i.e. α(1) =
. . . = α(|D|), then the longest PMI sequence in this con∗
⊆ DL , has length µL . Hence, one can
strained set, DL
conclude that the following inequality holds for any leaderfollower network:
rank(Γ) ≥ δL ≥ µL .
(20)
In light of (20), while the two quantities are equal for singleleader networks, the proposed lower bound δL is richer than
µL in capturing the controllability of networks with multiple
leaders. In fact, since the maximum distance between any two
nodes in a graph is the diameter of the graph by definition,
µL is always less than or equal to one plus the diameter of
the graph, even when every node is a leader. In general, the
difference between δL and µL depends on the graph topology
and the leader assignment. For instance, the example in Fig. 1
yields δL = 5 and µL = 3. Numerical comparisons of the
bounds for Erdös-Renyi and Barabási-Albert graphs with two
leaders are illustrated in Fig. 2. In this figure, each point on
the plot corresponds to the average value for 50 randomly
generated cases (each case is a randomly generated graph and a
pair of randomly assigned leaders). The results indicate that δL
provides a significantly better utilization of the graph distances
in the controllability analysis of multi-leader networks, even
when the network has only a pair of leaders.
The proposed lower bound is also closely related to the
number of unique DL vectors. For any connected G = (V, E)
4
BA graphs
ER graphs
10
µL
δL
the leaders. If such a pair is assigned as the leaders, then
rank(Γ) < υL .
We would like to conclude this section with a remark
regarding the application of the presented results to directed
networks.
µL
δL
10
Lower bound
Lower bound
8
6
8
6
4
4
2
0.1
0.2
p
0.3
0.4
1
2
3
m
4
5
Fig. 2. Comparison of the lower bounds for two randomly selected leaders
on Erdös-Renyi random graphs with 50 nodes in which any two nodes are
adjacent with the probability p; and Barabási-Albert graph with 50 nodes in
which each new node is connected to m existing nodes through a preferential
attachment strategy.
and any leader set L ⊆ V , let υL be the number of unique DL
vectors. Note that, due to (10), each vector in a PMI sequence
has an entry that is strictly smaller than the corresponding
entries of all the following vectors in that sequence. Hence,
a PMI sequence cannot contain two identical vectors. Consequently, the proposed lower bound is always less than or equal
to the number of unique DL vectors, i.e.
υL ≥ δL .
(21)
The relationship in (21) facilitates a deeper understanding
for potential applications of the proposed lower bound. For
instance, for any k ∈ {1, . . . , n}, having υL = k is a necessary
condition to have δL = k. Hence, it is possible to conclude that
a network is completely controllable as per the proposed lower
bound, only if each follower has a distinctive DL vector. The
relationships in (20) and (21) may naturally yield the question
of whether υL can capture rank(Γ) better than δL and µL .
In the following result, we show that there is no such universal
relationship between rank(Γ) and υL .
Proposition 3.3. For leader-follower networks, the number of
unique DL vectors, υL , is not a universal bound of rank(Γ).
Proof: We prove this statement by providing some examples both for rank(Γ) > υL and for rank(Γ) < υL .
1) A path graph with uniform edge weights is controllable
from any single node if and only if the the number of nodes
is a power of two, i.e. n = 2k for some k ∈ N [19]. Note that,
for a path graph with a single leader, δL = n if and only if
the leader is a leaf node. Hence, for any n = 2k , if a non-leaf
node is the only leader of a path graph with uniform edge
weights, then rank(Γ) > υL .
2) A cycle graph with uniform edge weights is controllable
from any pair of nodes if and only if the the number of nodes
is a prime number [19]. Consider any cycle graph of n nodes
such that n is an odd composite number. For such a graph
with two leaders, the clockwise and counterclockwise paths
between the leaders have different lengths. Hence, any pair
of nodes that have equal distances from one of the leaders
have different distances from the other leader, i.e. υL = n.
Furthermore, in light of [19], there exists a pair of nodes
that render the system uncontrollable if they are assigned as
Remark 3.1 (Directed Networks) The formulation in (1)-(8)
is applicable to directed networks with the interaction graph
G = (V, E), where each (i, j) ∈ E denotes that i is influenced
by j as in (5). For such a network, powers of the adjacency
matrix A have the property that [Ar ]ij = 0 for all 0 ≤ r <
dist(i, j) and [Adist(i,j) ]ij 6= 0, where dist(i, j) is the length
of the shortest directed path from i to j. Hence, by using the
corresponding DL vectors as in (9), the results in both Lemma
3.1 and Theorem 3.2 can be extended to strongly connected
networks (i.e., there exists a directed path from every node to
every other node).
IV. A N A LGORITHM FOR C OMPUTING THE LOWER BOUND
In this section, we present an algorithm to compute the
proposed lower bound, δL . Note that the main contribution
of this work is the lower bound itself, and the algorithm in
this section is provided to facilitate some practical use of our
result.
Let S = {d1 , d2 , . . . , dn } be the set of all DL vectors for a
given leader-follower network. Given these vectors, we present
an iterative way of generating the longest PMI sequences. Let
Cp be the set of all DL vectors that can be assigned as the pth
element of such a sequence D. According to these definitions,
C1 = S. Once a vector from Cp is assigned as the pth element
of the sequence, Dp , and an index α(p) satisfying (10) is
chosen, the resulting Cp+1 can be obtained from Cp as
Cp+1 = {di ∈ Cp | di,α(p) > Dp,α(p) }.
(22)
In order to obtain longer sequences, this iteration must
be continued until Cp = ∅. However, in general there are
too many possible sequences that can be obtained this way,
and it is not feasible to find the maximum length for PMI
sequences by searching among all these possibilities. Instead,
we present a necessary condition for a PMI sequence to
have the maximum possible length. This necessary condition
significantly lowers the number of sequences to consider.
Lemma 4.1. Let D be a PMI sequence of DL vectors with
the maximum possible length, then its pth entry, Dp , satisfies
Dp,α(p) = min di,α(p) .
di ∈Cp
(23)
Proof: For the sake of contradiction, assume that this is
not true for a PMI sequence D with the maximum length.
Then, there exists a DL vector dj ∈ Cp such that dj,α(p) <
Dp,α(p) . By the construction of a PMI sequence, dj can not
be added to this sequence after Dp . However, dj can be
added right before Dp while keeping all the other parts of
D the same since dj,α(p) can be selected to satisfy (10) in
the resulting sequence. Hence, it is possible to obtain a longer
PMI sequence, which leads to the contradiction that D does
not have the maximum possible length.
5
In light of (22), as far as the sequence length is concerned,
the only important decision at each step p in building a PMI
sequence satisfying (23) is the choice of α(p). Based on this
observation, we propose Algorithm I for computing the lower
bound.
Algorithm I
1: initialize: C1 = {{d1 , d2 , . . . , dn }}; p = 1
2: while Cp,y =
6 ∅ for some y ∈ {1, · · · , |Cp |} do
3:
q=1
4:
for i = 1 : |Cp | do
5:
if Cp,i =
6 ∅ then
6:
for j = 1 : m do
7:
Cp+1,q = {dt ∈ Cp,i | dt,j > min ds,j }
to an element of Cp that is computed in the (p − 1)st iteration
of the while loop. The left child of a node Cp,i is obtained
from Cp,i by deleting all the DL vectors whose first entries
are equal to the minimum value of the first entries among all
the DL vectors in Cp,i , i.e., obtained from Cp,i as per (23) and
(22) for α(p) = 1. Similarly, the right child of a node Cp,i
is obtained by following the same procedure for α(p) = 2.
Accordingly, the elements in the first three levels are
C2,1
ds ∈Cp,i
8:
q =q+1
9:
end for
10:
end if
11:
end for
12:
p=p+1
13: end while
14: return p − 1
C2,2
C3,1
Proposition 4.2. Given the DL vectors for any connected
leader-follower network, Algorithm I returns δL .
Proof: By combining (22) and (23) (line 7 of Algorithm
I), at each step p, Algorithm I builds all the possible Cp+1 that
correspond to the PMI sequences of p vectors satisfying (23).
Hence, when Cp,y = ∅ for every y ∈ {1, · · · , |Cp |}, there is not
a longer PMI sequence that satisfies the necessary condition
in Lemma 4.1. Consequently, Algorithm I always returns the
length of the longest PMI sequence, δL .
Note that, in light of Remark 3.1, Algorithm I can also be
used to compute the proposed lower bound for any directed
network with a strongly connected interaction graph.
For a sample run of Algorithm I, consider the network in
Fig. 1. For this example, the algorithm terminates after the
fifth iteration of the while loop, and δL = 5. The resulting
flow of Algorithm I can be represented via a tree diagram as
illustrated in Fig. 3.
C1,1
C1 :
C3 :
Fig. 3.
C2,2
C2,1
C3,1
0
C3,2
C3,3
C3,4
An illustration of the flow of Algorithm I for the network in Fig. 1.
In Fig. 3., each node Cp,i at a given level p > 1 corresponds
C3,4 =
,
3
1
2
0
3
2
1
C3,2 = C3,3 =
In Algorithm I, the variable Cp is the multiset, where
each element Cp,i is the Cp resulting from (22) for specific
choices of α(1), . . . , α(p − 1), subject to the corresponding
PMI sequences satisfying (23). The main while loop iterates as
long as there exists a longer PMI sequence that satisfies (23).
Note that for each Cp,i 6= ∅, there are m (number of leaders)
different Cp+1,q , each corresponding to a particular choice of
α(p). As such, Algorithm I computes δL by generating no
more than mδL elements Cp,i .
C2 :
=
=
=
C1,1 =
0
3
1
,
2
1
,
3
2
,
1
2
,
2
;
0
3
;
3
1
2
0
1
1
2
2
,
,
,
,
;
2
3
1
2
2
3
,
,
;
2
0
1
1
2
2
;
,
,
,
2
3
1
2
1
1
2
,
,
,
.
2
3
2
,
1
,
2
,
2
,
3
V. D ISTANCE -BASED L EADER S ELECTION FOR
C ONTROLLABILITY
One of the main attributes of the proposed lower bound
δL is its independence of the edge weights. In contrast,
rank(Γ) depends on the values of the edge weights unless
there is some constraint such as all the weights being equal.
Consequently, computing δL typically requires significantly
less information about the overall system. This minimality of
the required information makes the lower bound attractive in
many applications such as leader selection for controllability.
Leader selection problems typically require finding a leader
set L that optimizes a system objective such as robustness,
mixing time, or controllability (e.g., [20], [21], [22], [23]).
For instance, consider the problem of finding a minimum
number of leaders that render a given network controllable
under the resulting leader-follower dynamics. If the edge
weights are known (or if they are known to be identical),
then the rank of the controllability matrix can be computed
for any set of leaders. Hence, a possible, yet not scalable,
way to find a minimal set of leaders for controllability is
to execute an exhaustive search. Note that, aside from the
complexity issues, the rank computation is not applicable if
the edge weights are unknown and arbitrary. In such cases,
the leader selection problem needs to be solved by leveraging
the structural properties of the interaction graph.
One approach to achieving controllability under arbitrary
coupling weights is to choose the minimal L that achieves
structural controllability (e.g., [8]). Structural controllability
implies that the selected leaders provide complete controllability for some, not all, weighting functions w : E 7→ R+ .
Hence, this approach may fall short in some applications,
especially when there are constraints on the admissible edge
weights. For instance, if all the edge weights in a network are
equal by design, then it is known that a complete graph is not
controllable by any single leader [6] whereas a single leader is
enough to achieve structural controllability [8]. Alternatively,
6
the notion of strong structural controllability can be employed
in the leader selection (e.g., [11]). In this regard, the proposed
bound δL can be used to ensure that the dimension of the
controllable subspace is not smaller than some desired value,
k ∈ {1, 2, . . . , n}, for any w : E 7→ R+ by formulating the
leader selection problem as
|L|
minimize
L∈2V
(24)
subject to δL ≥ k,
In light of Theorem 3.2, any element in the feasible set of
the problem in (24) renders rank(Γ) ≥ k for any weighting
function w : E 7→ R+ . Note that the problem in (24)
is always feasible, i.e., for any given G = (V, E) and
k ∈ {1, 2, . . . , n}, there always exists some L ⊆ V such that
δL ≥ k. Indeed, the feasibility can be shown by considering
the trivial case, L = V . In that case, for each DL vector, the
entry that contains the distance of the corresponding node
from itself satisfies (10) for any sequence of the corresponding
DL vectors. Consequently, any sequence of the DL vectors
is a PMI sequence if L = V , and δV = n. Note that the
feasibility would not be guaranteed if the problem was posed
by using µL instead of δL since µL is always upper bounded
by one plus the graph diameter. An example, where the leaders
were assigned by solving (24) for k = n, is illustrated in
Fig. 4. For this example, the leaders were selected through an
exhaustive search by first looking for a single-leader solution
and incrementing the number of leaders until a solution exists.
Algorithm I was used to compute δL for each candidate L.
1
2
15
3
12
4
5
6
14
13
11
8
10
7
9
Fig. 4. A graph of 15 nodes, G = (V, E), and a minimal selection of leaders
(shown in gray), L = {2, 6, 9, 12, 14, 15}, such that δL = 15. The network
is completely controllable via L for any weighting function w : E 7→ R+ .
VI. C ONCLUSION
In this technical note, we presented that the distances
between the leaders and the followers on the interaction graph
contain some fundamental information about the controllability of the leader-follower networks. In particular, we used the
distance-to-leaders (DL) vectors to derive a tight lower bound
on the dimension of the controllable subspace. The proposed
bound is applicable to networks with arbitrary interaction
graphs and weighting functions w : E 7→ R+ . We also
provided some connections between the proposed lower bound
and a pair of closely related distance-based measures, namely
the maximum distance from the leaders and the number
of distinct DL vectors. While the results were presented
for undirected networks, we also showed how they can be
extended to directed networks. Furthermore, we presented
an algorithm for computing the lower bound. The proposed
bound may find its applications in various networked control
problems, especially when the edge weights are unknown. As
a prominent application, we presented how it can be utilized
to find a minimal set of leaders that ensure the controllability
of a leader-follower network with a given interaction graph
under any weighting function w : E 7→ R+ .
R EFERENCES
[1] A. Speranzon, C. Fischione, and K. H. Johansson, “Distributed and collaborative estimation over wireless sensor networks”, IEEE Conference
Decision and Control, pp. 1025–1030, 2006.
[2] A. Jadbabaie, J. Lin and A. S. Morse, “Coordination of groups of mobile
autonomous agents using nearest neighbor rules”, IEEE Transactions on
Automatic Control, 48(6): 988–1001, 2003.
[3] F. Dorfler and F. Bullo, “Synchronization and transient stability in
power networks and nonuniform Kuramoto oscillators”, SIAM Journal
on Control and Optimization, 50(3): 1616–1642, 2012.
[4] J. Ghaderi and R. Srikant, “Opinion dynamics in social networks: A local
interaction game with stubborn agents,” American Control Conference,
pp. 1982–1987, 2013.
[5] S. Gu et al., “Controllability of structural brain networks,” Nature
Communications, 6, 2015.
[6] A. Rahmani, M. Ji and M. Egerstedt, “Controllability of multi-agent
systems from a graph theoretic perspective”, SIAM J. Control Optim.,
48(1): 162–186, 2009.
[7] M. Egerstedt, S. Martini, M. Cao, K. Camlibel, and A. Bicchi, “Interacting with networks: How does structure relate to controllability in
single-leader consensus networks?” IEEE Control Systems Magazine,
32(4): 66–73, 2012.
[8] Y.-Y. Liu, J.-J Slotine, and A.-L.Barabási, “Controllability of complex
networks,” Nature, 473(7346): 167–173, 2011.
[9] C.-T Lin, “Structural controllability,” IEEE Transactions on Automatic
Control, 19(3): 201–208, 1974.
[10] H. Mayeda and T. Yamada, “Strong structural controllability,” SIAM
Journal on Control and Optimization, 17(1): 123–138, 1979.
[11] A. Chapman and M. Mesbahi, “On strong structural controllability
of networked systems: a constrained matching approach,” American
Control Conference, pp. 6126–6131, 2013.
[12] J. Jarczyk, F. Svaricek, and B. Alt, “Strong structural controllability of
linear systems revisited,” IEEE Conference on Decision and Control,
pp. 1213–1218, 2013.
[13] Y.-Y.Liu, J.-J Slotine, and A.-L.Barabási, “Control centrality and hierarchical structure in complex networks,” Plos One, 7(9), 2012.
[14] Y. Pan and X. Li, “Structural controllability and controlling centrality
of temporal networks,” PloS One, 9(4), 2014.
[15] J. C. Nacher and T. Akutsu, “Analysis of critical and redundant nodes in
controlling directed and undirected complex networks using dominating
sets,” Journal of Complex Networks, 2(4): 394–412, 2014.
[16] S. Zhang, M. K. Camlibel, and M. Cao, “Controllability of diffusivelycoupled multi-agent systems with general and distance regular coupling
topologies”, IEEE Conference on Decision and Control, 2011.
[17] S. Zhang, M. Cao, and M. K. Camlibel, “Upper and lower bounds for
controllable subspaces of networks of diffusively coupled agents,” IEEE
Transactions on Automatic Control, 59(3): 745–750, 2014.
[18] A. Y. Yazıcıoğlu, W. Abbas, and M. Egerstedt, “A tight lower bound on
the controllability of networks with multiple leaders,” IEEE Conference
on Decision and Control, pp. 1978–1983, 2012.
[19] G. Parlangeli and G. Notarstefano, “On the reachability and observability
of path and cycle graphs,” IEEE Transactions on Automatic Control,
57(3): 743–748, 2012.
[20] F. Lin, M. Fardad, and M. R. Jovanovic, “Algorithms for leader selection
in stochastically forced consensus networks,” IEEE Transactions on
Automatic Control, 59(7): 1789–1802, 2014.
[21] A. Clark, B. Alomair, L. Bushnell, and R. Poovendran, “Leader selection
in multi-agent systems for smooth convergence via fast mixing,” IEEE
Conference on Decision and Control, pp. 818–824, 2012.
[22] A. Y. Yazıcıoğlu and M. Egerstedt, “Leader selection and network
assembly for controllability of leader-follower networks,” American
Control Conference, pp. 3802–3807, 2013.
[23] C. O. Aguilar and B. Gharesifard, “Graph controllability classes for the
Laplacian leader-follower dynamics,” IEEE Transactions on Automatic
Control, 60(6): 1611–1623, 2015.
| 3cs.SY
|
Continuous Hands-off Control by
CLOT Norm Minimization
Niharika Challapalli ∗ Masaaki Nagahara ∗∗
Mathukumalli Vidyasagar ∗∗∗
arXiv:1611.02071v1 [cs.SY] 7 Nov 2016
∗
Department of Electrical Engineering University of Texas at Dallas
Richardson, TX 75080, USA (e-mail: [email protected])
∗∗
Institute of Environmental Science and Technology, The University
of Kitakyushu, Hibikino 1-1, Wakamatsu-ku, Kitakyushu, Fukuoka
808-0135, JAPAN (e-mail: [email protected]).
∗∗∗
Department of Systems Engineering, University of Texas at Dallas,
Richardson, TX 75080, USA, and Department of Electrical
Engineering, Indian Institute of Technology Hyderabad, Kandi,
Telangana, India 502285 (e-mail: [email protected],
[email protected])
Abstract: In this paper, we consider hands-off control via minimization of the CLOT
(Combined L-One and Two) norm. The maximum hands-off control is the L0 -optimal (or the
sparsest) control among all feasible controls that are bounded by a specified value and transfer
the state from a given initial state to the origin within a fixed time duration. In general, the
maximum hands-off control is a bang-off-bang control taking values of ±1 and 0. For many real
applications, such discontinuity in the control is not desirable. To obtain a continuous but still
relatively sparse control, we propose to use the CLOT norm, a convex combination of L1 and
L2 norms. We show by numerical simulation that the CLOT control is continuous and much
sparser (i.e. has longer time duration on which the control takes 0) than the conventional EN
(elastic net) control, which is a convex combination of L1 and squared L2 norms.
Keywords: Optimal control, convex optimization, sparsity, maximum hands-off control,
bang-off-bang control
1. INTRODUCTION
Sparsity has recently emerged as an important topic in
signal/image processing, machine learning, statistics, etc.
If y ∈ Rm and A ∈ Rm×n are specified with m < n, then
the equation y = Ax is underdetermined and has infinitely
many solutions for x if A has rank m. Finding the sparsest
solution (that is, the solution with the fewest number of
nonzero elements) can be formulated as
min kzk0 subject to Az = b.
z
However, this problem is NP hard, as shown in (Natarajan,
1995). Therefore other approaches have been proposed for
this purpose. This area of research is known as “sparse regression.” One of the most popular is LASSO (Tibshirani,
1996), also referred to as basis pursuit (Chen et al., 1999),
in which the ℓ0 -norm is replaced by the ℓ1 -norm. Thus the
problem becomes
min kzk1 subject to Az = b.
z
The advantage of LASSO is that it is a convex optimization problem and therefore very large problems can be
solved efficiently, for example by using the Matlab-based
package cvx (Grant and Boyd, 2014). Moreover, under
mild technical assumptions, the LASSO-optimal solution
has no more than m nonzero components (Osborne et al.,
2000). However, the exact location of the nonzero components is very sensitive to the vector y. To overcome this
deficiency, another approach known as the Elastic Net was
proposed in (Zou and Hastie, 2005), where the ℓ1 norm in
LASSO is replaced by a weighted sum of ℓ1 and squared
ℓ2 norms. This leads to the optimization problem
min λ1 kzk1 + λ2 kzk22 subject to Az = b,
z
where λ1 and λ2 are positive weights such that λ1 +λ2 = 1.
It is shown in (Zou and Hastie, 2005, Theorem 1) that the
EN formulation gives the grouping effect ; If two columns of
the matrix A are highly correlated, then the corresponding
components of the solution for x have nearly equal values.
This ensures that the solution for x is not overly sensitive
to small changes in y. The name “elastic net” is meant
to suggest a stretchable fishing net that retains all the big
fish.
During the past decade and a half, another research area
known as “compressed sensing” has witnessed a great deal
of interest. In compressed sensing, the matrix A is not
specified; rather, the user gets to choose the integer m
(known as the number of measurements), as well as the
matrix A. The objective is to choose the matrix A as well
as a corresponding “decooder” map ∆ : Rm → Rn such
that, the unknown vector x is sparse and the measurement
vector y equals Ax, then ∆(Ax) = x for all sufficiently
sparse vectors x. More generally, if measurement vector
y = Ax + η where η is the measurement noise, and the
vector x is nearly sparse (but not exactly sparse), then
the recovered vector ∆(Ax + η) should be sufficiently close
to the true but unknown vector x. This is referred to as
“robust sparse recovery.” Minimizing the ℓ1 -norm is among
the more popular decoders. See the books by (Elad, 2010),
(Eldar and Kutyniok, 2012), and (Foucart and Rauhut,
2013) for the theory and some applications. Due to its
similarity to the LASSO formulation of (Tibshirani, 1996),
this approach to compressed sensing is also referred to as
LASSO.
Until recently the situation was that LASSO achieves
robust sparse recovery in compressed sensing, but did not
achieve the grouping effect in sparse regression. On the flip
side, EN achieves the grouping effect, but it was not known
whether it achieves robust sparse recovery. A recent paper
(Ahsen et al., 2016) sheds some light on this problem. It
is shown in (Ahsen et al., 2016) that EN does not achieve
robust sparse recovery. To achieve both the grouping effect
in sparse regression as well as robust sparse recovery in
compressed sensing, (Ahsen et al., 2016) has proposed the
CLOT (Combined L-One and Two) formulation:
min λ1 kzk1 + λ2 kzk2 subject to Az = b and λ1 + λ2 = 1.
z
The difference between EN and CLOT is the ℓ2 norm term;
EN has the squared ℓ2 norm while CLOT has the pure
ℓ2 norm. This slight change leads to both the grouping
effect and robust sparse recovery, as shown in (Ahsen et al.,
2016).
In parallel with these advances in sparse regression and recovery of unknown sparse vectors, sparsity techniques have
also been applied to control. Sparsity-promoting optimization has been applied to networked control in (Nagahara
et al., 2014), where quantization errors and data rate can
be reduced at the same time by sparse representation of
control packets. Other examples of control applications include optimal controller placement by (Casas et al., 2012;
Clason and Kunisch, 2012; Fardad et al., 2011), design of
feedback gains by (Lin et al., 2013; Polyak et al., 2013),
state estimation by (Charles et al., 2011), to name a few.
More recently, a novel control called the maximum handsoff control has been proposed in (Nagahara et al., 2016) for
continuous-time systems. The maximum hands-off control
is the L0 -optimal control (the control that has the minimum support length) among all feasible controls that are
bounded by a fixed value and transfer the state from a
given initial state to the origin within a fixed time duration. Such a control is effective for reduction of electricity or fuel consumption; an electric/hybrid vehicle shuts
off the internal combustion engine (i.e. hands-off control)
when the vehicle is stopped or the speed is lower than a
preset threshold; see (Chan, 2007) for example. Railway
vehicles also utilize hands-off control, often called coasting control, to cut electricity consumption; see (Liu and
Golovitcher, 2003) for details. In (Nagahara et al., 2016),
the authors have proved the theoretical relation between
the maximum hands-off control and the L1 optimal control under the assumption of normality. Also, important
properties of the maximum hands-off control have been
proved in (Ikeda and Nagahara, 2016) for the convexity of
the value function, and in (Chatterjee et al., 2016) for the
existence and the discreteness.
In general, the maximum hands-off control is a bang-offbang control taking values of ±1 and 0. For many real
applications, such a discontinuity property is not desirable.
To obtain a continuous but still sparse control, (Nagahara
et al., 2016) has proposed to use a combined L1 and
squared L2 minimization, like EN mentioned above. Let
us call this control an EN control. As in the case of EN
in the vector optimization, the EN control often shows
much less sparse (i.e. has a larger L0 norm) than the
maximum hands-off control. Then, we proposed to use the
CLOT norm, a convex combination of L1 and non-squared
L2 norms. We call the minimum CLOT-norm control the
CLOT control. We show by numerical simulation that the
CLOT control is continuous and much sparser (i.e. has
longer time duration on which the control takes 0) than
the conventional EN control.
The remainder of this article is organized as follows. In
Section 2, we formulate the control problem considered in
this paper. In Section 3, we give a discretization method
to numerically compute the optimal control. Results of the
numerical computations on a variety of problems are presented in Section 4. These examples illustrate the advantages of the CLOT control compared with the maximum
hands-off control and the EN control. We present some
conclusions in Section 5.
Notation
Let T > 0 and m ∈ N. For a continuous-time signal
u(t) ∈ R over a time interval [0, T ], we define its Lp (p ≥ 1)
and L∞ norms respectively by
Z T
1/p
p
kukp ,
|u(t)| dt
, kuk∞ , sup |u(t)|.
t∈[0,T ]
0
We denote the set of all signals with kukp < ∞ by Lp [0, T ]
for p ≥ 1 or p = ∞. We define the L0 norm of a signal u(t)
on the interval [0, T ] as
Z T
φ0 (u(t))dt,
kuk0 ,
0
where φ0 is the L0 kernel function defined by
1, if α 6= 0,
φ0 (α) ,
0, if α = 0
(1)
for a scalar α ∈ R. The L0 norm can be represented by
kuk0 = µL supp(u) ,
where supp(u) is the support of the signal u, and µL is the
Lebesgue measure on R.
2. PROBLEM FORMULATION
Let us consider a linear time-invariant system described
by
dx
(t) = Ax(t) + Bu(t), t ≥ 0, x(0) = ξ.
(2)
dt
Here we assume that x(t) ∈ Rn , u(t) ∈ R, and the initial
state x(0) = ξ is fixed and given. The control objective is
to drive the state x(t) from x(0) = ξ to the origin at time
T > 0, that is
x(T ) = 0.
(3)
We limit the control u(t) to satisfy
kuk∞ ≤ Umax
(4)
for fixed Umax > 0.
If the system (2) is controllable and the final time T is
larger than the optimal time T ∗ (the minimal time in
which there exist a control u(t) that drives x(t) from
x(0) = ξ to the origin; see (Hermes and Lasalle, 1969)),
then there exists at least one u(t) ∈ L∞ [0, T ] that satisfies
equations (2), (3), and (4). Let us call such a control a
feasible control. From (2) and (3), any feasible control u(t)
on [0, T ] satisfies
Z T
eA(T −t) Bu(t)dt,
0 = x(T ) = eAT ξ +
0
or
Z
First, we divide the time interval [0, T ] into N subintervals,
[0, T ] = [0, h) ∪ · · · ∪ [(N − 1)h, N h], where h is the
discretization step (or the sampling period) such that
T = N h. We assume that the state x(t) and the control
u(t) in (2) are constant over each subinterval. On the
discretization grid, t = 0, h, . . . , N h, the continuous-time
system (2) is described as
xd [m + 1] = Ad xd [m] + Bd ud [m], m = 0, 1, . . . , N − 1,
(11)
where xd [m] , x(mh), ud [m] , u(mh), and
Z h
Ad , eAh , Bd ,
eAt Bdt.
(12)
0
T
e
−At
Bu(t)dt + ξ = 0.
(5)
0
Define a linear operator Φ : L∞ [0, T ] → Rn by
Z T
Φu ,
e−At Bu(t)dt, u ∈ L∞ [0, T ].
0
By this, we define the set U of the feasible controls by
U , {u ∈ L∞ : Φu + ξ = 0, kuk∞ ≤ 1} .
Define the control vector
ud , [ud [0], ud[1], . . . , ud [N − 1]]⊤ .
Note that the final state x(T ) can be described as
x(T ) = xd [N ] = AN
d ξ + ΦN ud ,
(14)
−1
−2
ΦN , AN
Bd , AN
Bd , . . . , B d .
d
d
(15)
where
(6)
The problem of the maximum hands-off control is then
described by
min kuk0 subject to u ∈ U.
(7)
u
The L0 problem (7) is very hard to solve since the L0
cost function is non-convex and discontinuous. For this
problem, (Nagahara et al., 2016) has shown that the L0
optimal control in (7) is equivalent to the following L1
optimal control:
min kuk1 subject to u ∈ U,
(8)
Then the set U in (6) is approximately represented by
UN , ud ∈ RN : AN
d ξ + ΦN ud = 0, kud k∞ ≤ 1 . (16)
Next, we approximate the L1 norm of u by
Z T
kuk1 =
|u(t)|dt
=
≈
u
if the plant is normal, that is, if the (2) is controllable and
the matrix A is nonsingular. Let us call the L1 optimal
control as the LASSO control. If the plant is normal, then
the LASSO control is in general a bang-off-bang control
that is piecewise constant taking values in {0, ±1}. The
discontinuity of the LASSO solution is not desirable in real
applications, and a smoothed solution is also proposed in
(Nagahara et al., 2016) as
min kuk1 + λkuk22 subject to u ∈ U,
u
(13)
(9)
where λ > 0 is a design parameter for smoothness. Let us
call this control the EN (elastic net) control. In (Nagahara
et al., 2016), it is proved that the solution of (9) is a
continuous function on [0, T ].
While the EN control is continuous, it is shown by numerical experiments that the EN control is not sometimes
sparse. This is an analogy of the EN for finite-dimensional
vectors that EN does not achieve robust sparse recovery.
Borrowing the idea of CLOT in (Ahsen et al., 2016), we
define the CLOT optimal control problem by
min kuk1 + λkuk2 subject to u ∈ U.
(10)
u
=
0
N
−1 Z (m+1)h
X
m=0 mh
N
−1 Z (m+1)h
X
m=0
N
−1
X
m=0
mh
|u(t)|dt
|ud [m]|dt
(17)
|ud [m]|h
= kud k1 h.
In the same way, we obtain approximation of the L2 norm
of u as
Z
kuk22 =
T
0
|u(t)|2 dt ≈ kud k22 h.
(18)
Finally, the optimal control problems (8), (9) and (10) can
be approximated by
min hkuk1 subject to ud ∈ UN
(19)
ud ∈RN
min hkud k1 + hλkud k22 subject to ud ∈ UN
√
min hkud k1 + hλkud k2 subject to ud ∈ UN
ud ∈RN
ud ∈RN
(20)
(21)
The optimization problems are convex and can be efficiently solved by numerical software packages such as cvx
with Matlab; see (Grant and Boyd, 2014) for details.
We call this optimal control the CLOT control.
4. NUMERICAL EXAMPLES
3. DISCRETIZATION
Since the problems (8)–(10) are infinite dimensional, we
should approximate it to finite dimensional problems. For
this, we adopt the time discretization.
In this section we present numerical results from applying
the CLOT norm minimization approach to seven different
plants, and compare the results with those from applying
LASSO and EN.
Plant
P1 (s)
P1 (s)
P2 (s)
P2 (s)
P3 (s)
P4 (s)
7
P5 (s)
8
9
P6 (s)
P7 (s)
Poles
(0,0,0,0)
(0,0,0,0)
−0.025 ± j
−0.025 ± j
−1 ± 0.2j, ±j
−1 ± 0.2j,
−0.3 ± j
−5 ± j,
−0.3 ±√
2j ,
−1 ± 2 2j
0, 0, 0, 0, ±j
0, 0, 0, 0, ±j
T
20
20
20
20
20
20
x(0)
e4
e4
e2
(10, 1)⊤
e4
e4
λ
1
0.1
0.1
0.1
0.1
0.1
Figs.
1,2
3,4
5,6
7,8
9, 10
11, 12
20
e6
0.1
13, 14
40
40
e6
e6
0.1
0.1
L2 norm of state
25
EN
LASSO
CLOT
20
||x(t)|| 2
No.
1
2
3
4
5
6
15, 16
17, 18
15
10
5
Table 1. Details of various plants studied.
0
4.1 Details of Various Plants Studied
To save space in the table, the plant zeros are not shown;
P3 (s) has a zero at s = −2, P6 (s) has a zero at s = 2,
while P7 (s) has zeros at s = 1, 2. The remaining plants
do not have any zeros, so that the plant numerator equals
one.
Once the plant zeros and poles are specified, the plant numerator and denominator polynomials n, d were computed
using the Matlab command poly. Then the transfer function was computed as P = tf(n,d), and the state space realization was computed as [A,B,C,D] = ssdata(P). The
maximum control amplitude is taken 1, so that the control
must satisfy |u(t)| ≤ 1 for t ∈ [0, T ]. To save space, we
use the notation el to denote an l-column vector whose
elements all equal one. Note that in all but one case, the
initial condition equals en where n is the order of the plant.
10
15
20
Fig. 1. State trajectory for the plant P1 (s) with the initial
state (1, 1, 1, 1)⊤ and λ = 1.
control
1
EN
LASSO
CLOT
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
0
5
10
15
20
t
Fig. 2. Control trajectory for the plant P1 (s) with the
initial state (1, 1, 1, 1)⊤ and λ = 1.
L2 norm of state
25
EN
LASSO
CLOT
Note that, with T = 20, the problems with plants P6 (s)
and P7 (s) are not feasible (meaning that T is smaller than
the minimum time needed to reach the origin); this is why
we took T = 40.
20
||x(t)|| 2
All optimization problems were solved after discretizing
the interval [0, T ] into both 2,000 as well as 4,000 samples,
to examine whether the sampling time affects the sparsity
density of the computed optimal control.
5
t
u(t)
For the reader’s convenience, the details of the various
plants are given in Table 1. The figure numbers show where
the corresponding computational results can be found.
Some conventions are adopted to reduce the clutter in the
table, as described next. All plants are of the form
np
nz
Y
Y
n(s)
P (s) =
(s − pi ).
(s − zi ), d(s) =
, n(s) =
d(s)
i=1
i=1
0
15
10
5
4.2 Plots of Optimal State and Control Trajectories
0
The plots of the ℓ2 -norm (or Euclidean norm) of the
state vector trajectory and the control signal for all these
examples are shown in the next several plots.
We begin with the plant P1 (s), the fourth-order integrator.
Figures 1 and 2 show the state and control trajectories
when λ = 1. The same system is analyzed using a smaller
value of λ = 0.1. One would expect that the resulting
control signals would be more sparse with a smaller λ, and
this is indeed the case. The results are shown in Figures
3 and 4. Based on the observation that the control signal
0
5
10
15
20
t
Fig. 3. State trajectory for the plant P1 (s) with the initial
state (1, 1, 1, 1)⊤ and λ = 0.1.
becomes more sparse with λ = 0.1 than with λ = 1, all
the other plants are analyzed with λ = 0.1.
Figures 5 and 6 display the state trajectory and the
control trajectories of the plant P2 (s) (damped harmonic
oscillator) when the initial state is (1, 1)⊤ . Figures 7 and
L2 norm of state
control
1
12
EN
LASSO
CLOT
0.8
0.6
10
0.4
8
0
||x(t)|| 2
u(t)
0.2
-0.2
-0.4
6
-0.6
-0.8
-1
4
EN
LASSO
CLOT
0
2
4
6
8
10
12
14
16
18
20
t
2
Fig. 4. Control trajectory for the plant P1 (s) with the
initial state (1, 1, 1, 1)⊤ and λ = 0.1.
L2 norm of state
1.5
0
0
5
10
15
20
t
Fig. 7. State trajectory for the plant P2 (s) with the initial
state (10, 1)⊤ and λ = 0.1.
EN
LASSO
CLOT
control
1
1
||x(t)|| 2
0.8
0.6
0.4
0.5
u(t)
0.2
0
-0.2
0
-0.4
0
5
10
15
20
t
-0.6
EN
LASSO
CLOT
-0.8
Fig. 5. State trajectory for the the plant P2 (s) with the
initial state (1, 1)⊤ and λ = 0.1.
control
1
-1
0
5
u(t)
0.6
0.4
5
0.2
4.5
0
4
-0.2
3.5
||x(t)|| 2
-0.4
-0.6
-1
EN
LASSO
CLOT
0
15
20
t
Fig. 8. Control trajectory for the plant P2 (s) with the
initial state (10, 1)⊤ and λ = 0.1.
0.8
-0.8
10
L2 norm of state
EN
LASSO
CLOT
3
2.5
2
1.5
5
10
15
20
t
1
0.5
Fig. 6. Control trajectory for the the plant P2 (s) with the
initial state (1, 1)⊤ and λ = 0.1.
0
0
5
10
15
20
t
8 show the state and control trajectories with the initial
state (10, 1)⊤ . It can be seen that, with this intial state,
the control signal changes sign more frequently.
Fig. 9. State trajectory for the plant P3 (s) with the initial
state (1, 1, 1, 1)⊤ and λ = 0.1.
To compare the sparsity densities of the three control
signals, we compute the fraction of time that each signal
is nonzero. In this connection, it should be noted that
the LASSO control signal is the solution of a linear programming problem; consequently its components exactly
equal zero at many time instants. In contrast, the EN
and CLOT control signals are the solutions of convex
optimization problems. Consequently, there are many time
instants when the control signal is “small” without being
smaller than the machine zero. Therefore, to compute
the sparsity density, we applied a threshold of 10−4 , and
treated a component of a control signal as being zero if
control
1
L2 norm of state
12
EN
LASSO
CLOT
0.8
10
0.6
0.4
8
||x(t)|| 2
u(t)
0.2
0
-0.2
4
-0.4
-0.6
2
EN
LASSO
CLOT
-0.8
-1
6
0
5
10
15
0
20
0
5
t
control
1
EN
LASSO
CLOT
0.8
0.6
3
0.4
2.5
0.2
2
u(t)
||x(t)|| 2
20
Fig. 13. State trajectory for the plant P5 (s) with the initial
state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
L2 norm of state
3.5
0
1.5
-0.2
1
-0.4
-0.6
0.5
EN
LASSO
CLOT
-0.8
0
0
5
10
15
-1
20
t
0
5
20
L2 norm of state
200
0.8
EN
LASSO
CLOT
180
160
0.4
140
||x(t)|| 2
0.6
0.2
0
-0.2
120
100
80
60
40
EN
LASSO
CLOT
-0.4
-0.6
15
Fig. 14. Control trajectory for the plant P5 (s) with the
initial state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
control
1
10
t
Fig. 11. State trajectory for the plant P4 (s) with the initial
state (1, 1, 1, 1)⊤ and λ = 0.1.
u(t)
15
t
Fig. 10. Control trajectory for the plant P3 (s) with the
initial state (1, 1, 1, 1)⊤ and λ = 0.1.
4
10
0
20
5
10
15
20
t
Fig. 12. Control trajectory for the plant P4 (s) with the
initial state (1, 1, 1, 1)⊤ and λ = 0.1.
its magnitude is smaller than this threshold. With this
convention, the sparsity densities of the various control
signals are as shown in Table 2. From this table it can
be seen that the control signal generated using CLOT
norm minimization has significantly lower sparsity density
compared to that of EN, and is not much higher than
0
0
10
20
30
40
t
Fig. 15. State trajectory for the plant P6 (s) with the initial
state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
that of LASSO. Also, as expected, the sparsity density
of LASSO does not change with λ, whereas the sparsity
densities of both EN and CLOT decrease as λ is decreased.
For this reason, in other examples we present only the
results for λ = 0.1.
No.
1
2
3
4
5
6
7
8
9
control
1
0.8
0.6
0.4
u(t)
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
EN
LASSO
CLOT
0
5
10
15
20
25
30
35
40
t
Fig. 16. Control trajectory for the plant P6 (s) with the
initial state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
L2 norm of state
200
EN
LASSO
CLOT
180
160
||x(t)|| 2
140
EN
0.5915
0.3270
0.1155
0.5555
0.3050
0.0395
0.1100
0.1438
0.1438
CLOT
0.4475
0.2480
0.0830
0.4225
0.2180
0.0805
0.0845
0.1125
0.1125
Table 3. Sparsity densities for optimal controllers produced by various methods
three methods LASSO, EN, and CLOT. The advantage
of using the sparsity density instead of the sparsity count
(the absolute number of nonzero entries) is that when the
sample time is reduced, the sparsity count would increase,
whereas we would expect the sparsity density to remain
the same. As explained above, we have applied a threshold
of 10−4 in computing the sparsity densities of various
control signals.
Table 3 shows the sparsity densities for the nine examples
studied in Table 1, in the same order. From this table
it can be seen that the CLOT norm-based control signal
is always more sparse than the EN-based control signal.
Indeed, in some cases the sparsity density of the CLOT
control is comparable to that of the LASSO control.
120
100
80
60
40
20
0
LASSO
0.1690
0.1690
0.0480
0.4055
0.1655
0.0040
0.0595
0.0568
0.0568
0
10
20
30
40
t
Fig. 17. State trajectory for the plant P7 (s) with the initial
state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
We also increased the number of samples from 2,000
to 4,000, and the optimal values changed only in the
third significant figure in almost all examples for all three
methods. Therefore the figures in Table 3 are essentially
equal to the Lebesgue measure of the support set divided
by T .
5. CONCLUSIONS
control
1
0.8
0.6
0.4
u(t)
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
EN
LASSO
CLOT
0
5
10
15
20
25
30
35
40
t
Fig. 18. Control trajectory for the plant P7 (s) with the
initial state (1, 1, 1, 1, 1, 1)⊤ and λ = 0.1.
λ
λ=1
λ = 0.1
LASSO
0.1725
0.1725
EN
0.6050
0.3795
CLOT
0.5900
0.2665
Table 2. Sparsity indices of the control signals
from various algorithms for the plant P1 (s)
(fourth-order integrator) with the initial state
(1, 1, 1, 1).
4.3 Comparison of Sparsity Densities
In this subsection we analyze the sparsity densities, that
is, the fraction of samples that are nonzero, using the
In this article, we propose the CLOT norm-based control
that minimizes the weighted sum of L1 and L2 norms
among feasible controls, to obtain a continuous control
signal that is sparser than the EN control introduced in
(Nagahara et al., 2016). We have shown a discretization
method, by which the CLOT optimal control problem
can be solved via finite-dimensional convex optimization.
Numerical experiments have shown the advantage of the
CLOT control compared with the LASSO and EN controls. Future work includes the analysis of the sparsity
and continuity of the CLOT control as a continuous-time
signal.
ACKNOWLEDGEMENTS
The research of MN was supported in part by JSPS
KAKENHI Grant Numbers 15H02668, 15K14006, and
16H01546. The research of MV and NC was supported
by the US National Science Foundation under Award No.
ECCS-1306630, the Cancer Prevention and Research Institute of Texas (CPRIT) under award No. RP140517, and
a grant from the Department of Science and Technology,
Government of India.
REFERENCES
Ahsen, M.E., Challapalli, N., and Vidyasagar, M. (2016).
Two new approaches to compressed sensing exhibiting
both robust sparse recovery and the gourping effect.
Journal of Machine Learning Research. (submitted;
preprint at arXiv 1410:8229).
Casas, E., Clason, C., and Kunisch, K. (2012). Approximation of elliptic control problems in measure spaces with
sparse solutions. SIAM J. Control Optim., 50, 1735–
1752.
Chan, C. (2007). The state of the art of electric, hybrid,
and fuel cell vehicles. Proc. IEEE, 95(4), 704–718.
Charles, A., Asif, M., Romberg, J., and Rozell, C. (2011).
Sparsity penalties in dynamical system estimation. In
45th Annual Conference on Information Sciences and
Systems (CISS), 1–6.
Chatterjee, D., Nagahara, M., Quevedo, D.E., and Rao,
K.M. (2016). Characterization of maximum hands-off
control. Systems & Control Letters, 94, 31–36.
Chen, S.S., Donoho, D.L., and Saunders, M.A. (1999).
Atomic decomposition by basis pursuit.
SIAM
J. Sci. Comput., 20(1), 33–61.
Clason, C. and Kunisch, K. (2012). A measure space
approach to optimal source placement. Comput. Optim.
Appl,, 53, 155–171.
Elad, M. (2010). Sparse and Redundant Representations.
Springer.
Eldar, Y.C. and Kutyniok, G. (2012). Compressed Sensing: Theory and Applications. Cambridge University
Press.
Fardad, M., Lin, F., and Jovanović, M. (2011). Sparsitypromoting optimal control for a class of distributed
systems. In American Control Conference (ACC), 2011,
2050–2055.
Foucart, S. and Rauhut, H. (2013). A Mathematical
Introduction to Compressive Sensing. Birkhäuser.
Grant, M. and Boyd, S. (2014). CVX: Matlab software for disciplined convex programming, version 2.1.
http://cvxr.com/cvx.
Hermes, H. and Lasalle, J.P. (1969). Functional Analysis
and Time Optimal Control. Academic Press.
Ikeda, T. and Nagahara, M. (2016). Value function in maximum hands-off control for linear systems. Automatica,
64, 190–195.
Lin, F., Fardad, M., and Jovanovic, M.R. (2013). Design
of optimal sparse feedback gains via the alternating
direction method of multipliers. IEEE Trans. Autom.
Control, 58(9), 2426–2431.
Liu, R. and Golovitcher, I.M. (2003). Energy-efficient
operation of rail vehicles. Transportation Research Part
A: Policy and Practice, 37(10), 917–932.
Nagahara, M., Quevedo, D.E., and Nevsić, D. (2016).
Maximum hands-off control: a paradigm of control effort
minimization. IEEE Trans. Autom. Control, 61(3), 735–
747.
Nagahara, M., Quevedo, D., and Østergaard, J. (2014).
Sparse packetized predictive control for networked control over erasure channels. IEEE Trans. Autom. Control,
59(7), 1899–1905.
Natarajan, B.K. (1995). Sparse approximate solutions to
linear systems. SIAM J. Comput., 24(2), 227–234.
Osborne, M.R., Presnell, B., and Turlach, B.A. (2000). On
the LASSO and its dual. Journal of Computational and
Graphical Statistics, 9, 319–337.
Polyak, B., Khlebnikov, M., and Shcherbakov, P. (2013).
An lmi approach to structured sparse feedback design in
linear control systems. In European Control Conference
(ECC), 833–838.
Tibshirani, R. (1996). Regression shrinkage and selection
via the LASSO. J. R. Statist. Soc. Ser. B, 58(1), 267–
288.
Zou, H. and Hastie, T. (2005). Regularization and variable
selection via the Elastic Net. Journal of the Royal
Statistical Society, Series B, 67, 301–320.
| 3cs.SY
|
ZERO DIVISOR GRAPHS OF QUOTIENT RINGS
arXiv:1508.02432v1 [math.AC] 10 Aug 2015
RACHAEL ALVIR
Abstract. The compressed zero-divisor graph ΓC (R) associated with a commutative ring R has
vertex set equal to the set of equivalence classes {[r] | r ∈ Z(R), r 6= 0} where r ∼ s whenever
ann(r) = ann(s). Distinct classes [r], [s] are adjacent in ΓC (R) if and only if xy = 0 for all
x ∈ [r], y ∈ [s]. In this paper, we explore the compressed zero-divisor graph associated with quotient
rings of unique factorization domains. Specifically, we prove several theorems which exhibit a method
of constructing Γ(R) for when one quotients out by a principal ideal, and prove sufficient conditions
for when two such compressed graphs are graph-isomorphic. We show these conditions are not
necessary unless one alters the definition of the compressed graph to admit looped vertices, and
conjecture necessary and sufficient conditions for two compressed graphs with loops to be isomorphic
when considering any quotient ring of a unique factorization domain.
1. Introduction
In Coloring of Commutative Rings [3], I. Beck introduces the zero-divisor graph Γ(R) associated with
the zero-divisor set of a commutative ring, whose vertex set is the set of zero-divisors. Two distinct
zero-divisors x, y are adjacent in Γ(R) if and only if xy = 0. The zero-divisor graph establishes a
"connection between graph theory and commutative ring theory which hopefully will turn out to
mutually beneficial for those two branches of mathematics." M. Axtell and J. Stickles [2] remark
that "in general, the set of zero-divisors lacks algebraic structure," suggesting that turning to the
zero-divisor graph may both reveal both ring-theoretical properties and impose a graph-theoretical
structure. Beck’s hopes have certainly been met; his now classical paper motivated an explosion
of research in this and similar associated graphs in the past decade. His definition has since been
modified to emphasize the fundamental structure of the zero-divisor set: Anderson’s definition given
in [1], which excludes the vertex 0 from the graph, is now considered standard.
In 2009, S. Spiroff and C. Wickham formalized the compressed zero-divisor graph ΓC (R), which has
vertex set equal to the set of equivalence classes {[r] | r ∈ Z(R), r 6= 0} where r ∼ s whenever
ann(r) = ann(s). Distinct classes [r], [s] are adjacent in ΓC (R) if and only if xy = 0 for all
x ∈ [r], y ∈ [s]. It is proved in [7] that xy = 0 for all x ∈ [r], y ∈ [s] if and only if rs = 0. In
other words, multiplication of equivalence classes is well-defined. This graph "represents a more
succinct description of the zero-divisor activity" [9]. Spiroff’s compressed graph was inspired by S.
Mulay’s work in [7] and has also been called the Mulay graph or the graph of equivalence classes of
zero-divisors. The compressed graph has many advantages over the traditional zero-divisor graph,
as it is often finite in cases where the zero-divisor set is infinite, and reveals associated primes.
1
The purpose of this paper is to study the compressed zero-divisor graph, and discover in what way
it better describes the core behavior of the zero-divisor set. This question becomes particularly
tractable in unique factorization domains, building off intuition built in Z/hni. Our main theorem
reveals that the overall structure of the zero-divisor activity for quotient rings of unique factorization
domains is determined by a finite set of key elements, which we call the zero-divisor basis for the
ring. As a result of this observation, we are able to state sufficient conditions for two compressed
graphs of such rings to be graph-isomorphic, and conjecture necessity when the compressed graph
admits looped vertices. In the final section, we conjecture that these results may be extended to
any quotient ring of a unique factorization domain.
2. Background
One motivation for the use of the compressed zero-divisor graph is the observation that new zerodivisors may be trivially found; every multiple of a zero-divisor is also a zero-divisor. Moreover,
these multiples often behave in exactly the same way as their parent zero-divisor: they annihilate
the same elements. This phenomena is expressed in the following lemma.
Lemma 1. Let R be a commutative ring, z ∈ Z(R), n 6∈ Z(R). Then [nz] = [z].
Proof. (⊆) Let x ∈ ann(nz). Then (nz)x = 0 ⇒ n(zx) = 0. Since n 6∈ Z(R), it follows by negating
the definition of zero-divisor that zx = 0. Therefore x ∈ ann(z).
(⊇) Let x ∈ ann(z). Then zx = 0 ⇒ (nz)x = 0 ⇒ x ∈ ann(nz).
The above is useful when considering the zero-divisor graph and thus appears in the literature,
although often in its weaker form where n is a unit. This observation leads us to consider that
some zero-divisors do not determine the behavior of the zero-divisor set, and have their behavior
determined by other "important" zero-divisors. Naturally, one wishes to identify which zero-divisors
actually determine the zero-divisor graph’s structure. To this end, we define two zero-divisors r, s
to be equivalent if ann(r) = ann(s), and denote the equivalence class of r with [r]. Because copycat
zero-divisors add unnecessary visual information to Γ(R), we wish to associate a graph with this
equivalence relation. The compressed zero-divisor graph then becomes the new object of study.
Are there any other situations, other than that expressed in lemma 1, in which the annihilators of
two elements are equal? This is addressed by the following lemma:
Lemma 2. Let S be a commutative ring, c ∈ S, I an ideal of S. Then c + I ∈ Z(S/I) if and only
if ∃d ∈ S − I such that cd + I = I.
Proof.
c + I ∈ Z(S/I) ⇐⇒ ∃d ∈ S s.t. cd + I = I and d + I 6= I
⇐⇒ ∃d ∈ S s.t. cd + I = I and d 6∈ I
2
We may restate this lemma so that it explicitly relates to the zero-divisor graph. We follow standard
notation in graph theory: given a graph G, its vertex set is denoted V (G), and edge set E(G).
Corollary 3. Let S be a commutative ring, I an ideal of S. Then ∃c, d ∈ S −I such that cd+I = I,
c + I 6= d + I if and only if
(1) c + I ∈ V (Γ(S/I)), and
(2) e = {c + I, d + I} ∈ E(Γ(S/I)).
Proof. The extra condition that c ∈ S − I ensures that 0 is not a vertex of Γ(S/I). The condition
that c + I 6= d + I ensures that edges exist only between distinct vertices.
Using this simple intuition as a springboard, we proceeed to consider the consequences of this
relationship between the ideal and the zero-divisor set of the quotient ring.
3. Results
Corollary 3 indicates that in order to investigate the zero-divisor set of a quotient ring, one must
look at the factors of the elements in the kernel the homomorphism. Not every factor, however, will
do. If I is the principal ideal hni, we will wish to restrict our attention only to the factors of n. We
therefore utilize the following definition.
Definition 1. Let D be a unique factorization domain and hni a principal ideal of D. A ZeroDivisor Basis B for D/hni is a set of all nontrivial divisors of n such that no two distinct elements
in the basis are associates. In symbols,
• d ∈ B ⇒ d | n, d is not a unit, d is not an associate of n.
• d1 , d2 ∈ B ⇒ d1 = d2 or d1 , d2 are not associates.
It is our claim that such factors are the only ones which determine the behavior of the zero-divisor
set, i.e., every zero-divisor has the same annihilator as the image of some irreducible divisor of n.
Lemma 4. Let D be a UFD, I an ideal of D, and φ the canonical homomorphism from D onto
D/I. Let a, n ∈ D. Then [φ(gcd(a, n))] is well-defined.
Proof. If I is trivial, then there is only one annihilator class in D/I and we are done. If not, then
since φ is onto, the image of a unit under φ is a unit. But a unit is never a zero-divisor. By Lemma
1, images of associates therefore share equivalence classes. However, every gcd(a, n) is unique up to
associates, so [φ(gcd(a, n))] is well-defined.
Theorem 5. Let D be a UFD, hni a principal ideal of D, and φ the canonical homomorphism from
D onto D/hni. Then ∀a ∈ D, [φ(a)] = [φ(gcd(a, n))].
3
Proof. If hni is maximal or trivial, there is only Q
one annihilator class in D/hni and we are done. If
not, then n is nonzero and not a unit. Let n = psi i be a factorization of n into irreducibles. Fix
a ∈ D.
If a = 0, then [φ(gcd(0, n))] = [φ(n)] = [0] and we are done.
Q
Suppose a 6= 0. We may write a = y pki i where
Q kigcd(y, n) =Q1. ki Then by Lemma 2 and since
y 6= 0, φ(y) is
not
a
zero-divisor,
so
[φ(a)]
=
[φ(
pi )]. But [φ( pi )] = [φ(gcd(a, n))] if and only
Q ki
if ∀x ∈ D, x Qpi ∈ hni ⇐⇒ x gcd(a, n) ∈ hni. It therefore suffices to show that the equivalence
∀x ∈ D, n | x pki i ⇐⇒ n | x gcd(a, n) holds.
(⇐) Let x ∈ D, and suppose Q
n | x gcd(a, n). Since gcd(y, n)Q= 1, then gcd(a, n) = gcd(
i
which implies that gcd(a, n) | pki Q
. We conclude that n | x pki i .
(⇒) Let x ∈ D, and suppose n | x pki i . Then
Y min(k ,s ) Y s −min(k ,s ) Y min(k ,s ) Y k −min(k ,s )
i i
i i
i i
i i
.
x
pi
pi i
pi
pi i
Q
pki i , n),
Since the cancellation property holds in UFD’s,
Y s −min(k ,s ) Y k −min(k ,s )
i i
i i
.
x
pi i
pi i
However, by definition of minimum, if ki − min(ki , si ) 6= 0, then si − min(ki , si ) = 0. In other
Q k −min(ki ,si ) si −min(ki ,si )
) = 1. By Euclid’s Lemma, we conclude that
, pi
words, gcd( pi i
Y s −min(k ,s )
i i
x,
pi i
which implies that
Y
Therefore n | x
Q
min(ki ,si )
pi
min(ki ,si )
pi
Y
s −min(ki ,si )
pi i
x
Y
min(ki ,si )
pi
, which is what we desired to show.
.
Theorem 6. Let D be a unique factorization domain, and hni a nontrivial principal ideal of D.
Let φ be the canonical homomorphism from D onto D/hni.
Each vertex of ΓC (D/hni) is the equivalence class [φ(d)] of exactly one d in the zero-divisor basis
B for D/hni. An edge exists between two distinct nodes [φ(d1 )], [φ(d2 )] if and only if d1 d2 ∈ hni.
Proof. If hni is maximal, then B is empty andQso also is ΓC (D/hni). Suppose hni is not maximal.
As hni is neither maximal nor trivial, let n = psi i .
It has already been shown that [φ(a)] = [φ(gcd(a, n))] for all a ∈ D. Thus every equivalence class
is the equivalence class of a divisor of n. Since the vertices of ΓC (D/hni) do not contain the classes
[0] or [1], every vertex of ΓC (D/hni) is the equivalence class of φ(d) for some d in B.
We show next that this is true for exactly one d ∈ B. Suppose that d1 , d2 are distinct elements in the
zero-divisor
Because the zero-divisor basis is unique up to associates, we may write
Q ti basis forQD/hni.
ki
d1 = u pi , d2 = v pi where u, v are units. Since d1 6= d2 and by definition of zero-divisor basis
4
Q
they are not associates, some tj 6= kj . Without loss of generality, let kj > tj . Consider d3 = psi i −ki .
Q t +(s −k )
Q
Then clearly φ(d3 ) annihilates φ(d2 ). However, d1 d3 = u pii i i = u pisi +ti −ki . But kj > tj
implies that sj + tj − kj < sj , so φ(d1 )φ(d3 ) 6= 0. We conclude that ann(φ(d1 )) 6= ann(φ(d2 )).
That an edge exists between distinct vertices [φ(d1 )], [φ(d2 )] if and only if d1 , d2 ∈ hni follows
immediately from the definition of compressed zero-divisor graph.
Corollary 7. The compressed graph of D/hni is always finite.
Proof. If hni is maximal or trivial, we are done because the compressed graph is empty. If not, then
because n may be expressed as a finite product of irreducibles, the zero-divisor basis of D/hni is
always finite.
Theorem 8. Let D1 , D2 be two unique
and let n1 ∈ D1 and n2 ∈ D2 be
QB tdomains,
QA sifactorization
i
non-zero and not units. Let n1 = i=1 pi , n2 = i=1 qi be factorizations of n1 , n2 into products of
irreducibles. If A = B and each si = ti , then ΓC (D1 /hn1 i) ∼
= ΓC (D2 /hn2 i).
Proof. A zero-divisor basis of D1 /hn1 i is
B1 = {d =
A
Y
pvi i | d 6= n, d 6= 1, 0 ≤ vi ≤ si },
i=1
and a basis of D1 /hn2 i is given by
B2 = {d =
B
Y
i
pw
i | d 6= n, d 6= 1, 0 ≤ wi ≤ si }.
i=1
Consider the map ψ (whose domain does not contain the equivalence classes of 0 or 1) defined by
hY
i hY
i
ψ
pzi i + hn1 i =
qizi + hn2 i .
By Theorem 6, ψ is a map between the vertex sets of ΓC (D1 /hn1 i) and ΓC (D2 /hn2 i). Since A = B,
each si = ti , and again by Theorem 6 each vertex corresponds to exactly one element in the zerodivisor basis, we conclude that ψ is a one-to-one correspondence.
5
Q
Q
Suppose [ pfi i ], [ pgi i ] are adjacent in ΓC (D1 /hn1 i). This happens if and only if
Y f Y g
pi i
pi i + hn1 i = hn1 i
Y f +g
⇐⇒
pi i i ∈ hn1 i
⇐⇒ ∀i, fi + gi ≥ si = ti
Y f +g
⇐⇒
qi i i ∈ hn2 i
Y f Y g
⇐⇒
qi i
qi i + hn2 i = hn2 i
hY i hY i
⇐⇒ ψ
pfi i , ψ
pgi i
are adjacent in ΓC (D2 /hn2 i).
Therefore ψ is an isomorphism.
Corollary 9. Let D be a UFD and hni a principal ideal of D. Then ΓC (D/hni) is isomorphic to
ΓC (Z/hmi) for some m ∈ Z.
It is perhaps appropriate we mention that requiring Γ(R) (and consequently ΓC (R)) be simple seems
to have caused some disagreement among the zero-divisor graph community. Axtell’s definition of
zero-divisor graph in [2] does not require that edges exist only between distinct vertices and refers
explicitly to "loops" in the graph, despite the fact that he claims Anderson’s definition is standard.
Anderson himself must remind his reader that "the zero-divisor graph does not detect the nilpotent
elements" [1]. For Redmond, Anderson’s definition posed potential problems, requiring "some
conditions that make the nilpotent elements easier to locate" [8].
In this paper also, we advocate changing the standard definitions of Γ(R) and ΓC (R) so that looped
vertices are allowed. First, only a graph with loops models all zero-divisor activity in the ring.
Additionally, although the former theorem holds for either definition of ΓC (R), its conditions are
not necessary unless we stipulate that ΓC (R) contain loops. As a counterexample, note that in the
unlooped case, ΓC (Z/hp3 i) ∼
= ΓC (Z/hpqi). We believe that sufficiency is held when ΓC (R) admits
loops, and that this sufficiency need only be shown for Z/hmi by the above corollary.
For certain quotient rings, the next theorem gives necessary and sufficient conditions for compressed
graph isomorphism regardless of whether or not we admit loops in ΓC (R).
Theorem 10. Let D1 , D2 be two UFD’s. Let p, p1 , . . . , pA ∈ D1 , q, q1 , . . . qB ∈ D2 be irreducibles,
where no two of the pi , qi ’s are associates respectively. Then ΓC (D1 /hps i) ∼
= ΓC (D2 /hq t i) if and only
QA
Q
B
if s = t, and ΓC (D1 /h i=1 pi i) ∼
= ΓC (D2 /h i=1 qi i) if and only if A = B.
Proof. Sufficiency has already been shown. We prove necessity by contraposition.
In the first case, suppose s 6= t. Then it is clear that ps , q t have different numbers of non-trivial
divisors up to associates, so the zero-divisor bases for the two rings will have different cardinalities.
Therefore the vertex sets of ΓC (D1 /hps i), ΓC (D2 /hq t i) cannot be put into a one-to-one correspondence, and the graphs cannot be isomorphic.
6
In the second case, suppose A 6= B. Similarly, it is clear that the vertex sets of
A
B
Y
Y
ΓC (D1 /h pi i), ΓC (D2 /h qi i)
have different cardinalities.
These graph-isomorphism results generalize those in [4], in which is is proved that Γ(Zp [x]/hxn i) ∼
=
Γ(Zpn ).
4. Conjectures
We would like to discuss when our results may be extended to Γ(R) itself. The following conjectures
sugges when this might be possible.
Conjecture 11. Let R1 , R2 be two commutative rings. Then Γ(R1 ) ∼
= Γ(R2 ) if and only if
(1) ΓC (R1 ) ∼
= ΓC (R2 ), and
(2) |R1 − Z(R1 )| = |R2 − Z(R2 )|.
We would also like to extend our results to when one quotients a UFD by any ideal I. Note first that
every ideal of a ring may be written as a union of principal ideals. The trivial union I = ∪x∈I hxi
suffices to show that this is true; equality holds because ideals are closed under multipliciation by
the elements of a ring by definition. A more general definition of zero-divisor basis follows.
Definition 2. Let D be a unique factorization domain, and let I = ∪α∈Λ hnα i be an expression of I
as a union of principal ideals. A zero-divisior basis B of D/I is the set of all divisors d of some
nα , d 6∈ I, such that no two distinct elements of B are associates.
If I is expressed minimally, we may be able to replace the stipulation that d 6∈ I with the requirement
that the divisor d merely be nontrivial.
Conjecture 12. Let D be a UFD, I = ∪α∈Λ hnα i be an ideal of D expressed as a minimal union, and
φ the canonical homomorphism from D onto D/I. Then ∀a ∈ D, [φ(a)] = [φ(gcd{a, nα for all α ∈
Λ}].
Conjecture 13. Let D be a unique factorization domain, and I an ideal of D where
I = ∪α∈Λ h
Qα
Y
sα
pαkk i
k=1
is a minimal expression of I as a union of principal ideals. The distinct nodes of the Q
compressed
zero-divisor graph for D/I are given by the equivalence classes of elements of the form
φ(
diQ
) where
Q
each di is an element
Q of the zero-divisor basis. An edge exists between two nodes φ( di ), φ( bi ) if
and only if nα | di bi for some α ∈ Λ.
7
Conjecture 14. Let D1 , D2 be unique factorization domains and I1 , I2 be any two ideals of D1 , D2
respectively which are neither maximal nor trivial. Suppose
I1 = ∪α∈Λ h
Qα
Y
sα
pαkk i.
k=1
Then ΓC (D1 /I1 ) ∼
= ΓC (D2 /I2 ) if we may write
I2 = ∪α∈Λ h
Qα
Y
sα
qαkk i.
k=1
References
[1] D. Anderson and P. Livingston, The Zero-Divisor Graph of a Commutative Ring. Journal of Algebra, Vol. 217,
pp. 434-447, 1999.
[2] M. Axtell and P. Livingston, Graphs and Zero-Divisors. The College Mathematics Journal, Vol. 41, No. 5, pp.
396-399, 2010.
[3] I. Beck, Coloring of Commutative Rings Journal of Algebra, Vol. 116, pp. 208-226, 1986
[4] D. Endean, K. Henry, and E. Manlove, Zero-Divisor Graphs of Zn and Polynomial Quotient Rings over Zn .
Rose-Hulman Undergraduate Mathematics Journal, Vol. 8, Nov. 12, 2007.
[5] D. Burton, Elementary Number Theory. The College Mathematics Journal, Mc-Graw Hill, 6th Edition, 2007.
[6] J. Gallian, Contemporary Abstract Algebra. Brooks/Cole, 8th Edition, 2013.
[7] S. Mulay, Cycles and Symmetries of Zero-Divisors. Commutative Algebra, Vol. 30, pp. 3533-3558, 2002.
[8] S. Redmond Generalizations of the Zero-Divisor Graph of a Ring Doctoral Dissertations, The University of
Tenessee, Knoxville, 2001.
[9] S. Spiroff and C. Wickham, A Zero-Divisor Graph Determined by Equivalence Classes of Zero-Divisors. Commutative Algebra, Vol. 39, pp. 2338-2348, 2011.
8
| 0math.AC
|
Control Strategy for Anaesthetic Drug Dosage with
Interaction Among Human Physiological Organs
Using Optimal Fractional Order PID Controller
Saptarshi Das1,2
Sourav Das2 and Koushik Maharatna1
1. School of Electronics and Computer Science, University
of Southampton, Southampton SO17 1BJ, UK.
Email: [email protected], [email protected]
2. Department of Power Engineering, Jadavpur University,
LB-8, Sector 3, Kolkata-700098, India.
Email: [email protected], [email protected]
Abstract—In this paper, an efficient control strategy for
physiological interaction based anaesthetic drug infusion model is
explored using the fractional order (FO) proportional integral
derivative (PID) controllers. The dynamic model is composed of
several human organs by considering the brain response to the
anaesthetic drug as output and the drug infusion rate as the
control input. Particle Swarm Optimisation (PSO) is employed to
obtain the optimal set of parameters for PID/FOPID controller
structures. With the proposed FOPID control scheme much less
amount of drug-infusion system can be designed to attain a
specific anaesthetic target and also shows high robustness for
±50% parametric uncertainty in the patient’s brain model.
Keywords—anaesthetic drug; dosage control; fractional order
PID controller; physiological organs; PSO
I.
INTRODUCTION
Control strategy formulation for anaesthetic drug dosage is
very crucial in clinical surgery and falls in a particular category
of biomedical system design known as pharmacology [1-3]. In
pharmacology, there are two major steps involved known as
pharmacokinetics and pharmacodynamics [4]. The anaesthetic
drug injected in a patient’s body gets infused in the arterial
blood flow and then the arterial blood carrying the drug reaches
to different physiological organs. Determining the drug
concentration in arterial blood flow and in different tissues i.e.
dosage to concentration is known as the pharmacokinetics.
The interaction of the drug with different physiological organs
and the overall effect of the drug i.e. concentration to effect is
known as pharmacodynamics [4]. Amongst many others
anaesthetic drugs, fentanyl is widely used in relief of acute pain
like cancer [5] and in different surgeries [6]. The drug dosage
in clinical surgery is generally controlled by the
Electroencephalogram (EEG) recording during the anaesthesia
[7] until a predefined unconsciousness in not reached which
can be characterized by observing slow oscillations in EEG
signals. Since different physiological organs have different
time constants to absorb, react and finally release the drug
again in the blood stream, their interaction and contribution on
the overall physiological dynamics of anaesthetic study is
highly complex and may lead clinicians to misinterpret the
observed event. As an example, some organs may store the
drug in a larger extent or reacts to the drug slowly than the
others and its immediate effect on the reduction of fast
oscillations in EEG waves may not be significant. This type of
phenomena may confuse the clinicians to enhance the drug
dosage to attend a pre-specified EEG activity indicative to an
anaesthetic state in a typical case which might be lifethreatening for the patient. In this scenario a physiological
model based simulation study is necessary to device a control
strategy with possible variation in patient’s mathematical
model from the nominal case, since over dosage of the drug by
an open-loop type EEG observation based control may
endanger the patient in the process of anaesthesia. A realistic
(data-based) model for control strategy formulation can thus be
implemented for automated anaesthesia in clinical surgery
using brain activity (EEG) monitoring as a sensory feedback to
the comparator to generate the tracking error. This results in a
control action going to an actuator to pump the drug into
patient’s body [8], using a prior knowledge of the set-point or
reference as quantitative measure of consciousness [4, 8].
Therefore the task of the controller design in the present
scenario can be summarised as minimising the tracking error of
the brain response given by the Hill equation, modified Hill
equation (in frequency domain) and bispectral index (BIS) of
EEG by delicately manipulating the drug input to the patient
[4]. A realistic mathematical model for fentanyl drug was
developed in [9-10] and is known as Mapleson-Higgins model.
The Mapleson model was composed of a few set of algebraic
equations derived from biochemistry. In this model, each organ
of human body like lungs, peripheral shunt, kidney, gut and
spleen, liver, other viscera, muscle, fat, sample brain is
considered as a separate compartment. It is also assumed that
each of the organs gets equal arterial blood flow. Mahfouf et al.
[11] translated the static algebraic equation based Mapleson
model into a dynamic model which considers the temporal
variation in the drug concentration at the inlet and outlet of
each physiological organ. They also carried out model
reduction of the large dynamical system for generalized
predictive control (GPC) design. In this paper, we use the
original (unreduced) higher order dynamic model in order to
design efficient control scheme. It has been shown by Das et
al. [12]-[13] that fractional order (FO) controllers are very
effective in handling higher order dynamics due to their
inherent infinite-dimensional nature over integer order
controllers which formulates the scope of the present study.
Control strategies for anaesthetic dosage have been formulated
on much simpler models using PID controllers [14]-[16], fuzzy
[17] and neuro-fuzzy [18] controllers in earlier research. Wada
et al. [19] developed a more detailed physiological system
level pharmacokinetic model without any control scheme. A
three state nonlinear compartmental model for clinical
pharmacology has been described for different control studies
including adaptive control [20], neural network control [21-22],
nonnegative dynamical systems [23], disturbance rejection
control [24]. Results of clinical trials of anaesthesia control
scheme, based on three-compartmental model with noise EEG
measurements has been reported in Haddad et al. [25] for 10
patients. The focus of the present study is to formulate an
optimal FO control strategy [26-28] with the dynamical model
reported in Mahfouf et al. [11] considering interaction amongst
physiological organs.
The rest of the paper is organised as follow: section II
describes the overall system model for different physiological
organs and their interaction with the drug. The optimal
controller design task is described in section III and the system
simulation in section IV. The paper ends with the conclusion in
section V, followed by the references.
II.
DYNAMIC MODEL OF FENTANYL INTERACTION WITH
PHYSIOLOGICAL ORGANS
A. Overall System Description for Automated Anaesthesia
controller to generate a control action u ( t ) , giving a command
to the actuator or mechanical pump to pump the drug (Fig. 1).
B. Dynamical Model for Each Human Organs
Mahfouf et al. [11] developed individual system models by
applying system identification techniques on original Mapleson
model with the consideration of 70 kg body weight and
6.481/min cardiac output which represent a base-case clinical
scenario. In order to study the generalising capability of the
model outside the nominal range, an interpolation based model
identification has been reported in [11] for realistic variation
either in body-weight, cardiac output or simultaneously both of
them. From the generalized interpolated scheme, a patient
model was developed with 93 kg body weight, 5.41/min body
weight while 100μg of fentanyl has been injected over a period
of 60 sec. The corresponding interpolated transfer function
models [11] of various human organs are described in (1)-(9).
G Fat =
− 1.437 × 10 − 8 s + 1.722 × 10 −6
s 2 + 0.4126 s + 0.0003241
G Lungs = 64.78 ( s 2 + 4.17 s + 6.97 )
G Gut-spleen =
As described earlier, the first static model of fentanyl
interaction from pharmacokinetics and pharmacodynamics
point of view was known as Mapleson model [9]-[10].
Mahfouf et al. [11] extended the concept for dynamic models
by representing each organs dynamics with ordinary
differential equations or single-input-single-output (SISO)
continuous time transfer function models. The model mainly
captures the drug flow from one organ to other and not the drug
concentration in a specific organ. The drug is added through
intravenous (IV) injection which gets infused in the arterial
blood and then is perfused in all organs with a fraction of the
total arterial blood flow. The brain reaction to the drug y ( t ) ,
manifested in the form of EEG is evaluated using the Hillequation or some index like BIS and then fed-back for
comparison with the target anaesthetic level to generate an
error e ( t ) which will be minimised by the FOPID or PI λ D μ
(2)
(3)
G Kidneys = 0.0132 ( s + 0.7436 )
(4)
G Liver = 0.006243 ( s + 0.04257 )
(5)
G Other-viscera = 0.001725 ( s + 0.09009 )
(6)
− 2.764 × 10 − 7 s + 3.312 × 10 − 5
s 2 + 0.4152 s + 0.001867
(7)
G Muscle =
Fig. 1. Overall fentanyl dose-effect model with proposed control strategy.
− 1.34 × 10 − 5 s + 0.001604
s 2 + 0.9059 s + 0.09356
(1)
G Brain = 1.614 × 10 − 5
( s + 0.1533 )
(8)
G Nasal = 5.459 × 10 −6
( s + 0.08507 )
(9)
The peripheral shunt has been modelled as only as a gain
without any time constants i.e. K peripheral-shunt = 0.0241 . The
above models explain the dynamic relationship between the
outgoing drug concentration in the blood flow from each tissue
and the incoming arterial pool drug amount whereas it cannot
model the drug concentration in each issue. If the drug
concentration in the outgoing flow from brain is represented as
C b , then the drug effect on the brain be calculated by the Hill
equation (10) using the effect-site concentration at 50% of drug
effect ( EC 50 = 7.8 ng/ml) and steepness or slope factor ( γ ).
Effect ( C b ) = C bγ
III.
( EC
γ
50
+ C bγ ) ,
γ = 4.3
(10)
OPTIMAL CONTROL SCHEME OF DRUG INFUSION WITH
FRACTIONAL ORDER PID CONTROLLER
A. Choice of the Control Objective
In clinical practices of anaesthesia, computer controlled
drug infusion is generally adopted to meet a target
concentration infusion (TCI) rather than manually controlling
the infusion rate [8, 14, 15, 22, 25]. Most of these techniques
rely on open loop control which assumes that the population
model is a good representative of any patient which may not be
valid in many cases. A more sensible scheme should be to
continuously monitor brain response (EEG) to the drug-dosage
and use it as a feedback mechanism to minimise the set-point
tracking error by an efficient controller structure. The control
signal essentially modifies the intravenous drug infusion rate
and should not be violently manipulated to reach a faster
anaesthetic effect. So apart from minimizing the tracking
error e ( t ) , it is also an important task to minimize the variation
(or derivative) of the drug infusion rate to prevent any sudden
shock in the automated IV injection pump and chance of
infusing large amount of drug in small time i.e. an increased
control effort u ( t ) . The objective function for the optimal
controller design has been formulated as (11) as a weighted
sum of Integral of Time multiplied Squared Error (ITSE) and
Integral of Squared Deviation Control Output (ISDCO).
J =
∫
∞
0
(w ⋅t ⋅e
1
2
( t ) + w2 ( Δ u ( t ) )
2
) dt
(11)
B. Controller Structure and its Tuning Using PSO Optimiser
Fractional order PI λ D μ controller design with various
other time domain performance criteria has been explored in
Das et al. [12] and the results shows that squared error term in
(11) puts more penalties on the tracking error and the time
multiplication term makes the overall response faster and
reduces the chance of loop oscillations in later stages. It is
evident that such a tracking criterion will definitely increase the
required control effort whose variation (temporal derivative) is
thus added as a minimising criteria in (11). The above control
objective is to be met by an integer and fractional order PID
controller structure (12) where the controller gains
{K p , K i , K d } and the integro-differential orders {λ , μ } are to
individual particle or pbest ( pi ) and that of the global swarm
or gbest ( pg ), while the velocity and positions are manipulated
over successive iterations using (14), until all particles
converges to the global best solution.
vi ( t + 1) = ωvi ( t ) + c1ϕ1 ( pi ( t ) − xi ( t ) ) + c2ϕ2 ( pg ( t ) − xi ( t ) )
xi ( t + 1) = xi ( t ) + vi ( t + 1)
(14)
Here, {c1 , c2 } = {0.5,1} are known as the inertia factor, cognitive
learning rate and social learning rate and {ϕ1 , ϕ2 } are uniformly
distributed random variables within the interval [ 0,1] . The
parameter ω , known as the inertia factor for the swarm is
linearly varied from 0.9 to 0.1. In the present study, the
unconstrained version of PSO is employed with only bound on
the controller parameters as { K p , K i , K d } ∈ [ 0.001,10 ] and
{λ , μ } ∈ [ 0, 2 ] . Due the implementation issues of ORA for FO
operators for {λ , μ } > 1 , we explored four classes of FOPID
structures (15) and tested the tracking performance and control
effort of each controller structure.
FOPID 1 →
({λ , μ } < 1) ;
FOPID 3 → ( λ > 1, μ < 1) ;
IV.
FOPID 2 → ( λ < 1, μ > 1) ;
FOPID 4 →
(15)
({λ , μ } > 1)
SIMULATION, RESULTS AND DISCUSSION
be tuned with a global optimisation algorithm.
C (s ) = K p + (Ki sλ ) + Kd sμ
(12)
For PID controller only the gains are to be optimized by
considering the orders as unity. In (12), each FO operator is
continuously rationalised within the optimization process using
a 5th order Oustaloup’s recursive approximation or ORA (13)
for a chosen frequency band of ω ∈ {ωl , ωb } = 10−2 ,102
{
}
rad/sec [29] and the rationalised pole-zero and gains are given
by { K , ωk , ωk′ } .
N
s K ∏ ( s + ωk′ ) ( s + ωk )
γ
(13)
k =− N
ωk = ωb (ωh ωb )
k + N + (1+γ ) 2
2 N +1
, ωk′ = ωb ( ωh ωb )
k + N + (1−γ ) 2
2 N +1
γ
, K = ωh
Here, the PID/FOPID controller parameters are optimized
using the PSO algorithm which is a widely used global
optimiser. The swarm starts with particles having velocity vi
and position xi and in each time step they try to move towards
the global best with latest value of the best found solution for
Fig. 2. Convergence characterteristics of different controller structures.
TABLE I.
OPTIMUM VALUES OF THE CONTROLLER PARAMETERS
Controller
Jmin
Controller parameters
Kp
Ki
Kd
λ
μ
PID
499.2007
3.8243
8.6647
0.001
-
-
FOPID1
58.3018
0.0212
2.3014
0.0783
0.8301
0.1013
FOPID2
792.1196
5.5757
3.3381
0.001
0.798
0.1488
FOPID3
68.721
0.2106
1.5393
0.001
0.001
0.021
FOPID4
728.6634
0.4213
1.6329
0.001
0.5237
0.1033
The PSO based optimal parameter selection is carried out
for the 5 controller structure with a goal of minimizing the
objective function (11) by considering a step input target at
t = 1 min and has been reported in Table I. The located minima
( J min ) in Table I show that the FOPID1 structure with
{λ , μ } < 1 gives the most optimal result than other FOPID
variants and PID controller which is counter-proved by the
PSO convergence characteristic, corresponding to FOPID1. In
most of the case low derivative gain implies that a PI/PIλ type
controller is more suitable for this particular system.
A. Performance on Nominal Patient Model
With the optimal parameters of the PID/FOPID controller
in Table 1, the nominal anaesthetic drug delivery system,
described in (1)-(10) and Fig. 1 is now simulated and the
tracking performance and required control efforts are compared
in Fig. 3 and Fig. 4 respectively. It is evident that although the
with PID controller the tracking is much faster, it needs more
amount of drug infusion within a short period of time (5 mins).
Whereas the FOPID1 structure meets the same anaesthetic
target while pushing much less amount of drug into patient’s
body over a longer period of time (20 mins).
Testing vulnerability of control loops with dc gain variation is
a widely used robustness measure and has been described in
Das et al. [12]. The dc gain of the nominal brain model is
1.0528×10-4. The rest of the physiological organs i.e.
peripheral-shunt, kidney, liver, other-viscera, muscle, fat, nasal
receive the same arterial blood flow in Mapleson model
described in Fig. 1, but does not directly affect the output y ( t ) .
Hence, they will have much less influence if the respective
parameters are perturbed. Fig. 5-6 reports simulations with the
best found cases of PID and FOPID controller respectively
under a ±50% dc gain perturbation in the brain model (8).
Fig. 5. Tracking performance and control effort with PID controller for
±50% dc-gain perturbation of the brain model.
Fig. 3. Tracking performance for different controller structures.
Fig. 6. Tracking performance and control effort with FOPID controller for
±50% dc-gain perturbation of the brain model.
Fig. 4. Control effort for different controller structures.
B. Robustness of Control Scheme in Perturbed Condition
In the previous subsection, the simulations have been
reported for the nominal model whereas it is highly likely that
the model differs significantly for different patients and for
different conditions of the same patient. Since the brain model
directly affects the output of the system as shown in Fig. 1, i.e.
the brain response to fentanyl drug directly feedback to the
controller, we here report simulation studies of the brain
response for the drug in a ±50% dc gain perturbation scenario.
It is evident from Fig. 5 that for a decrease in the dc-gain of
nominal brain model, the control effort rapidly increases,
signifying that more drug is injected into the patient’s body. It
is evident that the rise time of the PID controlled system is
much faster (10 mins), but the amount of drug-injected is
significantly high (17-35 units). On contrary, the FOPID
controller in Fig. 6 provides a relatively slower tracking
performance (50 mins). But there is a significant amount of
saving in the drug injection level which is bounded within 7-18
units even in the perturbed condition. In clinical practices,
meeting the specified unconsciousness level is not the sole
criterion because this may create a sudden shock in different
human physiological systems which may be harmful for the
patient [8]. Also, in order to attain same level of anaesthetic
effect but within a shorter time- period, the drug concentration
in brain needs to be raised rapidly which has both economical
constraint and physiological ill-effects along with faster
mechanical pumping needed for automated drug-delivery.
Whereas a smoother and bounded control action with less
amount of drug infusion into the patient can attain the same
level of anaesthesthetic fentanyl drug, if equipped with an
efficient control strategy i.e. using proposed optimal FOPID
controller.
[14]
[15]
[16]
V.
CONCLUSION
The paper devises a new fractional order control strategy
for automatic fentanyl drug dosage control for anaesthesia in
clinical practices. The proposed FOPID control scheme
requires less amount of dug to be injected than with a PID
controller to meet same anaesthetic target. The smoother
control action provided by FOPID controller outperforms
classical PID controller in a sense of restricting the chance of
feeling a sudden shock in the brain response due to rapid
increase in the amount of anaesthetic drug concentration in
arterial blood flow within a very short time interval. Future
work may be directed towards validation of the control scheme
applied on Mapleson’s dynamic model, with real clinical data.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
M.C.K. Khoo, “Physiological control systems: analysis, simulation and
estimation”, IEEE Press Series on Biomedical Engineering, 2000.
G.D. Baura, “System theory and practical applications of biomedical
systems”, IEEE Press Series on Biomedical Engineering, 2002.
R.B. Northorp, “Signals and systems analysis in biomedical
engineering”, CRC Press, 2010.
J.M. Bailey and W.M. Haddad, “Drug dosing control in clinical
pharmacology”, IEEE Control Systems Magazine, vol. 25, no. 2, pp. 3551, Apr. 2005.
R.K. Portenoy et al., “Oral transmucosal fentanyl citrate (OTFC) for the
treatment of breakthrough pain in cancer patients: a controlled dose
titration study”, Pain, vol. 79, no. 2–3, pp. 303-312, Feb. 1999.
K.J.S. Anand, W.G. Sippell, and A.Aynsley Green, “Randomised trial
on fentanyl anaesthesia in preterm babies undergoing surgery: effects on
the stress response”, The Lancet, vol. 329, no. 8527, pp. 243-248, Jan.
1987.
N. Ty Smith et al. “EEGs during high-dosage fentanyl-, sufentanil-, or
morphine-oxygen anesthesia”, Anesthesia & Analgesia, vol. 63, no. 4,
pp. 386-393, 1984.
H. Schwilden and H. Stoeckel, “Effective therapeutic infusions produced
by closed-loop feedback control of methohexital administration during
total intrvanous anesthesia with fentanyl”, Anesthesiology, vol. 73, no. 2,
pp. 225-229, 1990.
N.R. Davis and W.W. Mapleson, “A physiological model for the
distribution of injected agents, with special reference to pethidine”,
British Journal of Anaesthesia, vol. 70, no. 3, pp. 248-258, 1993.
W.W. Mapleson, “Circulation-time models of the uptake of inhaled
anaesthetics and data for quantifying them”, British Journal of
Anaesthesia, vol. 45, no. 4, pp. 319-334, 1973.
M. Mahfouf, D.A. Linkens, and D. Xue, “A new generic approach to
model reduction for complex physiologically based drug models”,
Control Engineering Practice, vol. 10, no. 1, pp. 67-81, Jan. 2002.
S. Das, S. Saha, S. Das, and A. Gupta, “On the selection of tuning
methodology for FOPID controllers for the control of higher order
processes”, ISA Transactions, vol. 5, no. 3, pp. 376-388, July 2011.
S. Das, A. Gupta, and S. Das, “Generalized frequency domain robust
tuning of a family of fractional order PI/PID controllers to handle higher
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
order process dynamics”, Advanced Material Research, vol. 403-408
(MEMS, NANO and Smart Systems), pp. 4859-4866, 2012.
D.A. O’Hara, D.K. Bogen, and A. Noordergraaf, “The use of computer
controlling the delivery of anesthesia”, Anesthesiology, vol. 77, no. 3,
pp. 563-581, 1992.
T. Sakai, A. Matsuki, P.F. White, and A.H. Giesecke, “Use of an EEGbiseptral closed-loop delivery system for administering propofol”, Acta
Anaesthesia Scandinavica, vol. 44, no. 8, pp. 1007-1010, Sep. 2000.
R.R. Jaklitsch and D.R. Westenskow, “A model-based self-adjusting
two-phase controller for vacuronium-induced muscle relaxation during
anesthesia”, IEEE Transactions on Biomedical Engineering, vol. BME34, no. 8, pp. 583-594, Aug. 1987.
J.S. Shieh, D.A. Linkens, and J.E. Peacock, “Hierarchical rule-based and
self-organizing fuzzy logic control for depth of anaesthesia”, IEEE
Transactions on Systems, Man, and Cybernetics-Part C: Applications
and Reviews, vol. 29, no. 1, pp. 98-109, Feb. 1998.
M. Mahfouf, C.S. Nunes, D.A. Linkens, and J.E. Peacock, “Modelling
and multivariable control in anaesthesia using neural-fuzzy paradigms:
part II. closed loop control of simultaneous administration of propofol
and remifentanil”, Artificial Intelligence in Medicine, vol. 35, no. 3, pp.
207-213, Nov. 2005.
D.R. Wada, D.R. Stanski, and W.F. Ebling, “A PC-based graphical
simulator for physiological pharmacokinetic models”, Computer
Methods and Programs in Biomedicine, vol. 46, no. 3, pp. 245-255, Apr.
1995.
W.M. Haddad, T. Hayakawa, and J.M. Bailey, “Adaptive control for
nonlinear compartmental dynamical systems with applications to clinical
pharmacology”, Systems & Control Letters, vol. 55, no. 1, pp. 62-70,
Jan. 2006.
W.M. Haddad, J.M. Bailey, T. Hayakawa, and N. Hovakimyan, “Neural
network adaptive output feedback control for intensive care unit sedation
and intraoperative anesthesia”, IEEE Transactions on Neural Networks,
vol. 18, no. 4, pp. 1049-1066, Jul. 2007.
W.M. Haddad and J.M. Bailey, “Closed-loop control for intensive care
unit sedation”, Best Practice & Research Clinical Anaesthesiology, vol.
23, no. 1, pp. 95-114, Mar. 2009.
W.M. Haddad, T. Hayakawa, and J.M. Bailey, “Adaptive control for
non-negative and compartmental dynamical systems with applications to
general anesthesia”, International Journal of Adaptive Control and
Signal Processing, vol. 17, no. 3, pp. 209-235, Apr. 2003.
K. Volyanskyy, W.M. Haddad, and J.M. Bailey, “Adaptive disturbance
rejection control for compartmental systems with application to
intraoperative anesthesia influence by hemorrhage and hemodilution
effects”, International Journal of Adaptive Control and Signal
Processing, vol. 23, no. 1, pp. 1-29, Jan. 2009.
W.M. Haddad, K.Y. Volyanskyy, J.M. Bailey, and J.J. Im,
“Neuroadaptive output feedback control for automated anesthesia with
noisy EEG measurements”, IEEE Transactions on Control Systems
Technology, vol. 19, no. 2, pp. 311-326, Mar. 2011.
S. Das, I. Pan, S. Das, and A. Gupta, “Master-slave chaos
synchronization via optimal fractional order PIλDμ controller with
bacterial foraging algorithm”, Nonlinear Dynamics, vol. 69, no. 4, pp.
2193-2206, Sep. 2012.
S. Das, I. Pan, K. Halder, S. Das and A. Gupta, “LQR based improved
discrete PID controller design via optimum selection of weighting
matrices using fractional order integral performance index”, Applied
Mathematical Modelling, vol. 37, no. 6, pp. 4253-4268, Mar. 2013.
I. Pan and S. Das, “Intelligent fractional order systems and control: an
introduction”, Series: Studies in Computational Intelligence, vol. 438,
2013, Springer-Heidelberg.
F. Merrikh-Bayat, “Rules for selecting the parameters of Oustaloup
recursive approximation for the simulation of linear feedback systems
containing PIλDμ controller”, Communications in Nonlinear Science and
Numerical Simulation, vol. 17, no. 4, pp. 1852-1861, Apr. 2012.
| 3cs.SY
|
SYSTEMS OF PARAMETERS AND HOLONOMICITY OF
A-HYPERGEOMETRIC SYSTEMS
arXiv:1302.0048v1 [math.AG] 1 Feb 2013
CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER
Abstract. The main result is an elementary proof of holonomicity for A-hypergeometric
systems, with no requirements on the behavior of their singularities, originally due to
Adolphson [Ado94] after the regular singular case by Gelfand and Gelfand [GG86]. Our
method yields a direct de novo proof that A-hypergeometric systems form holonomic families over their parameter spaces, as shown by Matusevich, Miller, and Walther [MMW05].
Introduction
An A-hypergeometric system is the D-module counterpart of a toric ideal. Solutions
to A-hypergeometric systems are functions, with a fixed infinitesimal homogeneity, on an
affine toric variety. The solution space of an A-hypergeometric system behaves well in
part because the system is holonomic, which in particular implies that the vector space
of germs of analytic solutions at any nonsingular point has finite dimension.
This note provides an elementary proof of holonomicity for arbitrary A-hypergeometric
systems, relying only on the statement that a module over the Weyl algebra in n variables
is holonomic precisely when its characteristic variety has dimension at most n (see [Gab81]
or [BGK+ 87, Theorem 1.12]), along with standard facts about transversality of subvarieties and about Krull dimension. In particular, our proof requires no assumption about
the singularities of the A-hypergeometric system; equivalently, the associated toric ideal
need not be standard graded. Holonomicity was proved in the regular singular case by
Gelfand and Gelfand [GG86], and later by Adolphson [Ado94, §3] regardless of the behavior of the singularities of the system. Adolphson’s proof relies on careful algebraic
analysis of the coordinate rings of a collection of varieties whose union is the characteristic variety of the system. Another proof of the holonomicity of an A-hypergeometric
system, by Schulze and Walther [SW08], yields a more general result: for a weight vector L from a large family of possibilities, the L-characteristic variety for the L-filtration
is a union of conormal varieties and hence has dimension n; holonomicity follows when
L = (0, . . . , 0, 1, . . . , 1) induces the order filtration on the Weyl algebra. The L-filtration
method uses an explicit combinatorial interpretation of initial ideals of toric ideals, which
requires a series of technical lemmas.
Holonomicity of A-hypergeometric systems forms part of the statement and proof, by
Matusevich, Miller, and Walther [MMW05], that A-hypergeometric systems determine
Date: 22 January 2013.
2010 Mathematics Subject Classification. Primary: 33C70; Secondary: 13N10, 14M25, 16S32, 33C99.
EM had support from NSF grant DMS-1001437. SG acknowledges the financial support of Fondecyt
Proyecto Regular 1110072.
1
2
CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER
holonomic families over their parameter spaces. The new proof of that statement here
serves as a model suitable for generalization to hypergeometric systems for reductive
groups, in the sense of Kapranov [Kap98].
The main step (Theorem 1.2) in our proof is an easy geometric argument showing that
the Euler operators corresponding to the rows of an integer matrix A form part of a system
of parameters on the product kn × XA , where k is any algebraically closed field and XA is
the toric variety over k determined by A. This observation leads quickly in Section 2 to
the conclusion that the characteristic variety of the associated A-hypergeometric system
has dimension at most n, and hence that the system is holonomic. Since the algebraic
part of the proof holds when the entries of β are considered as independent variables that
commute with all other variables, the desired stronger consequence is immediate: the Ahypergeometric system forms a holonomic family over its parameter space (Theorem 2.1).
1. Systems of parameters via transversality
Fix a field k. Let x = x1 , . . . , xn and ξ = ξ1 , . . . , ξn be sets of coordinates on kn and let
xξ denote the column vector with entries x1 ξ1 , . . . , xn ξn . Given a rectangular matrix L
with n columns, write Lxξ for the vector of bilinear forms given by multiplying L times xξ.
Lemma 1.1. Let k2n = knx × knξ have coordinates (x, ξ) and let X ⊆ knξ be a subvariety.
If L is an ℓ × n matrix with entries in k, then the variety Var(Lxξ) of Lxξ in k2n is
transverse to kn × X at any smooth point of kn × X whose ξ-coordinates are all nonzero.
Proof. It suffices to prove the statement after passing to the algebraic closure of k, so
assume k is algebraically closed. Let (p, q) be a smooth point of kn × X that lies in
Var(Lxξ) and has all coordinates of q nonzero. The tangent space to kn × X at (p, q)
contains kn × {0}. The tangent space T(p,q) to Var(Lxξ) is the kernel of the ℓ × 2n matrix
[L(q) L(p)], where L(p) (respectively, L(q)) is the ℓ×n matrix that results after multiplying
each column of L by the corresponding coordinate of p (respectively, q). Since the q
coordinates are all non-zero, T(p,q) projects surjectively onto the last n coordinates; indeed,
if η ∈ knξ is given, then taking yi = −pi ηi /qi yields y ∈ knx with L(q)y + L(p)η = 0. Thus
the tangent spaces at (p, q) sum to the ambient space, so the intersection is transverse.
The next result applies the lemma to an affine toric variety X. A fixed d × n integer
matrix A = [a1 a2 · · · an−1 an ] defines an action of the algebraic torus T = (k∗ )d on knξ by
t · ξ = (ta1 ξ1 , . . . , tan ξn ).
The orbit Orb(A) of the point 1 = (1, . . . , 1) ∈ kn is the image of an algebraic map
T → kn that sends t → t · 1. The closure of Orb(A) in kn is the affine toric variety
XA = Var(IA ) cut out by the toric ideal
IA = hξ u − ξ v | Au = Avi ⊆ k[ξ],
of A in the polynomial ring k[ξ] = k[ξ1 , . . . , ξn ]. The T -action induces an A-grading on k[ξ]
via deg(ξi ) = ai , and the semigroup ring SA = k[ξ]/IA is A-graded [MS, Chapters 7–8].
SYSTEMS OF PARAMETERS AND HOLONOMICITY OF A-HYPERGEOMETRIC SYSTEMS
3
For any face τ of the real cone R≥0 A generated by the columns of A, write τ A and let
1 ∈ {0, 1}n ⊂ kn be the vector with nonzero entry 1τi = 1 precisely when A has
F a nonzero
column ai ∈ τ . The variety XA decomposes as a finite disjoint union XA = τ A Orb(τ )
of orbits, where Orb(τ ) = T ·1τ . Each orbit has dimension dim Orb(τ ) = rank(Aτ ), where
Aτ is the submatrix of A consisting of those columns lying in τ , and dim XA = rank(A).
Theorem 1.2. The ring k[x, ξ]/ IA + hAxξi has Krull dimension n. In particular, if A
has rank d then the forms Axξ are part of a system of parameters for k[x] ⊗k SA .
τ
Proof. Let kτ ⊆ kn be the subspace consisting of vectors with 0 in coordinate i if
ai ∈
/ τ , and let |τ | be its dimension. Since k[x, ξ]/IA = k[x] ⊗k SA has dimension
n + rank(A) and the number of k-linearly independent generators of hAxξi is at most
rank(A), the Krull dimension in question is at least n. Hence it suffices to prove that
kn × Orb(τ ) ∩ Var(Axξ) ⊆ kn × kτ has dimension at most n. Let xτ and ξτ denote the
subsets corresponding to τ in the variable sets x and ξ, respectively. The projection of
the intersection onto the subspace kτ × kτ has image contained in
kτ × Orb(τ ) ∩ Var(Aτ xτ ξτ ) ⊆ kτ × kτ .
It therefore suffices to show that the dimension of this latter intersection is at most |τ |.
By Lemma 1.1, the intersection is transverse in kτ × kτ . But the dimension of Orb(τ ) is
the codimension of Var(Aτ xτ ξτ ) in kτ × kτ , which completes the proof.
2. Hypergeometric holonomicity
In this section, the matrix A is a d × n integer matrix of full rank d. Let
D = Chx, ∂ [∂i , xj ] = δij and [xi , xj ] = 0 = [∂i , ∂j ]i
denote the Weyl algebra over the complex numbers C, where x = x1 , . . . , xn and ∂i
corresponds to ∂x∂ i . This is the ring of C-linear differential operators on C[x].
For β ∈ Cd , the A-hypergeometric system with parameter β is the left D-module
MA (β) = D/D · (IA∂ , {Ei − βi }di=1 ),
where IA∂ = h∂ u − ∂ v | Au = Avi ⊆ C[∂] is the toric ideal associated to A and
n
X
Ei − βi =
aij xj ∂j − βi
j=1
are Euler operators associated to A.
The order filtration F filters D by order of differential operators. The symbol of an
operator P is its image in(P ) ∈ grF D. Writing ξi = in(∂i ), this means grF D is the
commutative polynomial ring C[x, ξ]. The characteristic variety of a left D-module M is
the variety in A2n of the associated graded ideal grF ann(M) of the annihilator of M. A
nonzero D-module is holonomic if its characteristic variety has dimension n; this is equivalent to requiring that the dimension be at most n; see [Gab81] or [BGK+ 87, Theorem 1.12].
4
CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER
The rank of a holonomic D-module M is the (always finite) dimension of C(x) ⊗C[x] M as
a vector space over C(x); this number equals the dimension of the vector space of germs
of analytic solutions of M at any nonsingular point in Cn [SST00, Theorem 1.4.9].
Viewing the A-hypergeometric system MA (β) as having a varying parameter β ∈ Cd ,
the rank of MA (β) is upper semicontinuous as a function of β [MMW05, Theorem 2.6].
This follows by viewing MA (β) as a holonomic family [MMW05, Definition 2.1] parametrized by β ∈ Cd . By definition, this means not only that MA (β) is holonomic for
each β, but also that it satisfies a coherence condition over Cd , namely: after replacing
β with variables b = b1 , . . . , bd , the module C(x) ⊗C[x] MA (b) is finitely generated over
C(x)[b]. (The definition of holonomic family in [MMW05] allows sheaves of D-modules
over arbitrary complex base schemes, but that generality is not needed here.)
The derivation of the holonomic family property for MA (b) from the holonomicity of the
A-hypergeometric system is more or less the same as in [MMW05, Theorem 7.5], which was
phrased in the generality of Euler–Koszul homology of toric modules. The brief deduction
here isolates the steps necessary for A-hypergeometric systems; its brevity stems from the
special status of affine semigroup rings among all toric modules [MMW05, Definition 4.5].
Theorem 2.1. The module MA (b) forms a holonomic family over Cd with coordinates b.
In more detail, as a D[b]-module the parametric A-hypergeometric system MA (b) satisfies:
1. the fiber MA (β) = MA (b) ⊗C[b] C[b]/hb − βi is holonomic for all β; and
2. the module C(x) ⊗C[x] MA (b) is finitely generated over C(x)[b].
Proof. Since R = C[x, ξ]/hin(IA ), Axξi surjects onto grF MA (β), it is enough to show that
the ring R has dimension n. If MA (β) is standard Z-graded (equivalently, the rowspan
of A over the rational numbers contains the row [1 1 · · · 1 1] of length n), then in(IA ) =
IA ⊆ C[ξ], and the result follows from Theorem 1.2.
When MA (β) is not standard Z-graded, let  be the (d + 1) × (n + 1) matrix obtained
by adding a row of 1’s across the top of A and then adding as the leftmost column
(1, 0, . . . , 0). If ξ0 denotes a new variable corresponding to the leftmost column of Â, and
ˆ
ξˆ = {ξ0 } ∪ ξ, then C[ξ]/ in(IA ) ∼
= C[ξ]/hI
 , ξ0 i. In particular,
ˆ
C[x̂, ξ]
C[x̂, ξ]
∼
,
=
ˆ
hin(IA ), Axξi
hIÂ , ξ0 , Âx̂ξi
where x̂ = {x0 } ∪ x. Since hI , ξ0i is Â-graded and  has a row [1 1 · · · 1 1], we have
reduced to the case where MA (β) is Z-graded, completing part 1.
With R as in part 1, the ring R[b] surjects onto grF MA (b), so it suffices for part 2
to show that R[b] becomes finitely generated over C(x)[b] upon inverting all nonzero
polynomials in x. Since the ideal hin(IA ), Axξi has no generators involving b variables, it
suffices to show that R(x) itself has finite dimension over C(x). The desired result from
the statement proved for part 1: any scheme of dimension n has finite degree over Cnx .
SYSTEMS OF PARAMETERS AND HOLONOMICITY OF A-HYPERGEOMETRIC SYSTEMS
5
References
[Ado94]
Alan Adolphson, Hypergeometric functions and rings generated by monomials, Duke Math.
J. 73 (1994), 269–290.
[BGK+ 87] A. Borel, P.-P. Grivel, B. Kaup, A. Haefliger, B. Malgrange, and F. Ehlers, Algebraic Dmodules, Perspectives in Mathematics 2, Academic Press, Inc., Boston, MA, 1987.
[Gab81]
Ofer Gabber, The integrability of the characteristic variety, Amer. J. Math. 103 (1981), no.
3, 445–468.
[GG86]
I. M. Gelfand and S. I. Gelfand, Generalized hypergeometric equations, (Russian) Dokl. Akad.
Nauk SSSR 288 (1986), no. 2, 279–283. English Translation: Soviet Math. Dokl. 33 (1986),
no. 3, 643–646.
[GGZ87] I. M. Gelfand, M. I. Graev, and A. V. Zelevinsky, Holonomic systems of equations and series
of hypergeometric type, Dokl. Akad. Nauk SSSR 295 (1987), no. 1, 14–19.
[GKZ90] I. M. Gelfand, M. Kapranov, and A. V. Zelevinsky, Generalized Euler integrals and Ahypergeometric functions, Adv. Math. 84 (1990), no. 2, 255–271.
[GKZ93] I. M. Gelfand, A. V. Zelevinsky, and M. M. Kapranov, Hypergeometric functions and toric
varieties, Funktsional. Anal. i Prilozhen. 23 (1989), no. 2, 12–26. Correction in ibid, 27
(1993), no. 4, 91.
[GKZ94] I. M. Gelfand, M. Kapranov, and A. V. Zelevinsky, Discriminants, resultants and multidimensional determinants, Mathematics: Theory & Applications. Birkhäuser Boston, Inc., Boston,
MA, 1994.
[Kap98]
Mikhail Kapranov, Hypergeometric functions on reductive groups, Integrable systems and algebraic geometry (Kobe/Kyoto, 1997), World Sci. Publishing, River Edge, NJ, 1998, pp. 236–
281.
[MMW05] Laura Felicia Matusevich, Ezra Miller, and Uli Walther, Homological methods for hypergeometric families, J. Amer. Math. Soc. 18 (2005), no. 4, 919–941.
[MS]
Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graduate Texts in
Mathematics, 227. Springer-Verlag, New York, 2005.
[SST00]
Mutsumi Saito, Bernd Sturmfels, and Nobuki Takayama, Gröbner Deformations of Hypergeometric Differential Equations, Springer–Verlag, Berlin, 2000.
[SW08]
Mathias Schulze and Uli Walther, Irregularity of hypergeometric systems via slopes along
coordinate subspaces, Duke Math. J. 142 (2008), no. 3, 465–509.
Department of Mathematics, Duke University, Box 90320, Durham, NC 27708
E-mail address: [email protected]
Instituto de Mathemática y Fı́sica, Universidad de Talca, Camino Lircay, Talca, Chile
E-mail address: [email protected]
Department of Mathematics, Duke University, Box 90320, Durham, NC 27708
E-mail address: [email protected]
| 0math.AC
|
Top-Down Synthesis of Multi-Agent Formation Control:
An Eigenstructure Assignment based Approach
arXiv:1708.04851v1 [cs.SY] 16 Aug 2017
Takatoshi Motoyama and Kai Cai
Abstract— We propose a top-down approach for formation
control of heterogeneous multi-agent systems, based on the
method of eigenstructure assignment. Given the problem of
achieving scalable formations on the plane, our approach
globally computes a state feedback control that assigns desired closed-loop eigenvalues/eigenvectors. We characterize the
relation between the eigenvalues/eigenvectors and the resulting
inter-agent communication topology, and design special (sparse)
topologies such that the synthesized control may be implemented locally by the individual agents. Moreover, we present
a hierarchical synthesis procedure that significantly improves
computational efficiency. Finally, we extend the proposed approach to achieve rigid formation and circular motion, and
illustrate these results by simulation examples.
I. I NTRODUCTION
Cooperative control of multi-agent systems has been an
active research area in the systems control community [1],
[2], [3], [4], [5]. Among many problems, formation control
has received much attention [6] owing to its wide applications such as satellite formation flying, search and rescue,
terrain exploration, and foraging. A main problem studied
is stabilization to a rigid formation, where the goal is to
steer the agents to achieve a formation with a specified
size and only freedoms of translation and rotation. Several
control strategies have been proposed: affine feedback laws
[7], [8], nonlinear gradient-based control [9], [10], and anglebased algorithms [11]. Achieving a scalable formation with
unspecified size has also been studied [12], [13]; a scalable
formation may allow the group to adapt to unknown environment with obstacles or targets. In addition, [14], [15] have
presented methods of controlling formations in motion.
These different methods for formation control have a
common feature in design: namely bottom-up. Specifically,
the inter-agent communication topology is given a priori,
which defines the neighbors for each agent. Then based only
on the neighborhood information, local control strategies
are designed for the individual agents. The properties of
the designed local strategies are finally analyzed at the
systemic (i.e. global) level, and correctness is proved under
certain graphical conditions on the communication topology.
This bottom-up design is indeed the mainstream approach
for cooperative control of multi-agent systems that places
emphasis on distributed control.
In this paper, we propose a distinct, top-down approach
for formation control, based on a known method called
eigenstructure assignment [16], [17], [18], [19]. Different
The authors are with Dept. of Electrical and Information Engineering,
Osaka City University, Japan. Emails: [email protected],
[email protected].
from the bottom-up approach, here there need not be any
communication topology imposed a priori (in fact the agents
are typically assumed independent, i.e. uncoupled), and
no design will be done at the local level. Indeed, given
a multi-agent formation control problem characterized by
specific eigenvalues and eigenvectors (precisely defined in
Section II), our approach constructs on the global level a
feedback matrix (if it exists) that renders the closed-loop
system to possess those desired eigenvalues/eigenvectors,
thereby achieving desired formations. Moreover, the synthesized feedback matrix (its off-diagonal entries being zero or
nonzero) defines the communication topology, and accordingly the computed feedback control may be implemented
by individual agents. Thus our approach features “compute
globally, implement locally”.
The inter-agent communication topology is a result of
control synthesis, rather than given a priori. We characterize the relation between the resulting topology and the
eigenstructure chosen for the synthesis. Further, we show
that by appropriately choosing desired eigenvalues and the
corresponding eigenvectors, special topologies (star, cyclic,
line) can be designed, and the computed feedback control
may be implemented locally over these (sparse) topologies.
Although our method requires centralized computation of
control gain matrices, we show that a straightforward extension of the approach to a hierarchical synthesis procedure
significantly reduces computation time. Empirical evidence is
provided to show the efficiency of the proposed hierarchical
synthesis procedure; in particular, computation of a feedback
control for a group of 1000 agents needs merely a fraction
of a second, which is likely to suffice for many practical
purposes.
The main advantage of our top-down approach is that
it is systematic, in the sense that it treats heterogeneous
agent dynamics and different cooperative control specifications (characterizable by desired eigenstructure) by the
same synthesis procedure. We show that scalable formation,
rigid formation, and cooperative circular motion can all be
addressed using the same method. Additionally, we show that
this method is amenable to deal with more general cases
where some agents are not self-stabilizable and the initial
inter-agent connections are arbitrary.
We first proposed this eigenstructure assignment based
approach in [20], where we applied it to solve the consensus
problem. Then the conference precursor [21] of this paper
extended the approach to solve scalable and rigid formations, and proposed a hierarchical synthesis procedure to
significantly shorten the computation time. This paper differs
from [21] in the following aspects. (i) A precise relation
between eigenstructure and topology is characterized (Section III). (ii) A method for imposing topological constraints
on eigenstructure assignment is presented (Section III.B).
(iii) More general cases where the initial inter-agent topology is arbitrary and/or there exist non-stabilizable agents
are addressed (Section IV). (iv) The problem of achieving
cooperative circular motion is solved. (v) All the proofs are
provided.
We note that [22] also proposed an eigenstructure assignment method and applied it to the multi-agent consensus
problem. Their approach is bottom-up: first a communication
topology is imposed among the agents, then local control
strategies are designed based on eigenstructure assignment
respecting the topology, and finally the correctness of the
proposed strategies is verified at the global level. By contrast, our approach is top-down: no topology is imposed
a priori, and topology is a result of control synthesis.
Moreover, we characterize the relation between topology and
eigenstructure, and design special topologies by selecting
special eigenstructures. In addition, the problems addressed
in the paper are distinct, namely scalable/rigid formation
and circular motion on the plane, which involve complex
eigenvalues and eigenvectors.
The rest of the paper is organized as follows. In Section II
we review the basics of eigenstructure assignment and formulate the multi-agent formation control problem. In Section III
we solve the problem by eigenstructure assignment, and
discuss the relations between eigenvalues/eigenvectors and
topologies. In Section IV we study the more general cases
where the initial inter-agent topology is arbitrary and/or there
exist non-stabilizable agents. In Section V we present a
hierarchical synthesis procedure to reduce computation time,
and in Section VI extend the method to achieve rigid formation and circular motion. Simulation examples are given in
Section VII and our conclusions stated in Section VIII.
N1 (λ)
N (λ) :=
∈ C(n+m)×m
N2 (λ)
First we review the basics of eigenstructure assignment [16]. Consider a linear time-invariant finite-dimensional
system modeled by
(1)
where x ∈ Cn is the state vector, u ∈ Cm the input vector,
and A ∈ Rn×n , B ∈ Rn×m .
Suppose we modify (1) by state feedback u = F x. It
is well-known (e.g. [23]) that F may be chosen to assign
any (self-conjugate) set of closed-loop eigenvalues for ẋ =
(A + BF )x if and only if (A, B) is controllable. Unless
m = 1 (single input), however, F is not uniquely determined
by a set of closed-loop eigenvalues. Indeed, with state
feedback F one has additional freedom to assign certain sets
of closed-loop eigenvectors. Simultaneously assigning both
eigenvalues and eigenvectors is referred to as eigenstructure
assignment.
(2)
with linearly independent columns such that
N1 (λ)
λI − A B
= 0.
N2 (λ)
(3)
Thus the columns of N (λ) form a basis of Ker[λI − A B];
Ker denotes kernel. Also we will use Im to denote image.
Lemma 1: ([16]) Consider the system (1) and suppose that
(A, B) is controllable and KerB = 0. Let {λ1 , . . . , λn } be
a set of distinct complex numbers, and {v1 , . . . , vn } a set of
linearly independent vectors in Cn . Then there is a unique
F such that for every i ∈ [1, n], (A + BF )vi = λi vi if and
only if
(∀i ∈ [1, n])vi ∈ ImN1 (λi )
(4)
where N1 (·) is in (2).
Lemma 1 provides a necessary and sufficient condition
of eigenstructure assignment. When the condition holds and
thus F exists for assigning distinct complex eigenvalues λi
and the corresponding eigenvectors vi (i ∈ [1, n]), F may be
constructed by the following procedure [16].
(i) For each λi compute a basis of Ker[λi I − A B]. Stack
the basis vectors to form N (λi ) in (2); partition N (λi )
properly to get N1 (λi ) and N2 (λi ).
(ii) Find wi = −N2 (λi )ki , where ki ∈ C is such that
N1 (λi )ki = vi (the condition KerB = 0 in Lemma 1
ensures that N1 (λi ) has independent columns; thus ki may
be uniquely determined).
(iii) Compute F by
F = [w1 · · · wn ][v1 · · · vn ]−1 .
II. PRELIMINARIES AND PROBLEM
FORMULATION
A. Preliminaries on Eigenstructure Assignment
ẋ = Ax + Bu
Let λ ∈ C. It is shown in [16] that if (A, B) is controllable,
then there exists
(5)
Note that the entries of F may include complex numbers
in general. If {λ1 , . . . , λn } is a self-conjugate set of distinct
complex numbers, and vi = v̄j wherever λi = λ̄j (¯· denotes
complex conjugate), then all entries of F are real numbers.
The procedure (i)-(iii) of computing F has complexity
O(n3 ), inasmuch as the calculations involved are solving
systems of linear equations, matrix inverse and multiplication
(e.g. [24]).
We note that the above eigenstructure assignment result
may be extended to the case of repeated eigenvalues with
generalized eigenvectors. For details refer to [17] or Appendix.
B. Problem Formulation
Consider a heterogeneous multi-agent system where each
agent is modeled by a first-order ODE:
ẋi = ai xi + bi ui ,
i = 1, . . . , n.
(6)
Here xi ∈ C is the state variable, ui ∈ C the control variable,
ai ∈ R and bi (6= 0) ∈ R are constant parameters. Thus each
agent is a point mass moving on the complex plane, with
possibly stable (ai < 0), semistable (ai = 0) or unstable
(ai > 0) dynamics. The requirement bi 6= 0 is to ensure
stabilizability/controllability of (ai , bi ); thus each agent is
stabilizable/controllable. Note that represented by (6), the
agents are independent (i.e. uncoupled) and no inter-agent
topology is imposed at this stage.
In vector-matrix form, the system of n independent agents
is
ẋ = Ax + Bu
(7)
where x := [x1 · · · xn ]⊤ ∈ Cn , u := [u1 · · · un ]⊤ ∈ Cn ,
A := diag(a1 , . . . , an ) and B := diag(b1 , . . . , bn ); here
diag(·) denotes a diagonal matrix with the specified diagonal
entries. Consider modifying (7) by a state feedback u = F x
and thus the closed-loop system is
ẋ = (A + BF )x.
(8)
Straightforward calculation shows that the diagonal entries
of A + BF are ai + bi Fii , and the off-diagonal entries bi Fij .
Since bi 6= 0, the off-diagonal entries (A + BF )ij 6= 0 if
and only if Fij 6= 0 (i 6= j).
In view of the structure of A + BF , we can define a
corresponding directed graph G = (V, E) as follows: the
node set V := {1, . . . , n} with node i ∈ V standing for
agent i (or state xi ); the edge set E ⊆ V × V with edge
(j, i) ∈ V if and only if F ’s off-diagonal entry Fij 6= 0 .
Since Fij 6= 0 implies that xi uses xj in its state update,
we say for this case that agent j communicates its state
xj to agent i, or j is a neighbor of i. The graph G is
therefore called a communication network among agents,
whose topology is decided by the off-diagonal entries of F .
Thus the communication topology is not imposed a priori,
but emerges as the result of applying the state feedback
control u = F x.
Now we define the formation control problem of the multiagent system (7).
Problem 1: Consider the multi-agent system (7) and specify a vector f ∈ Cn (f 6= 0). Design a state feedback
control u = F x such that for every initial condition x(0),
limt→∞ x(t) = cf for some constant c ∈ C.
In Problem 1, the specified vector f represents a desired formation configuration in the plane. By formation
configuration we mean that the geometric information of
the formation remains when scaling and rotational effects
are discarded. Indeed, by writing the constant √
c ∈ C in
the polar coordinate form (i.e. c = ρejθ , j = −1), the
final formation cf is the configuration f scaled by ρ and
rotated by θ. The constant c is unknown a priori and in
general depends on the initial condition x(0). Note also that
Problem 1 includes the consensus problem as a special case
when f = 1 := [1 · · · 1]⊤ .
To solve Problem 1, we note the following fact.
Proposition 1: Consider the multi-agent system (7) and
state feedback u = F x. If A+BF has a simple eigenvalue 0,
with the corresponding eigenvector f , and other eigenvalues
have negative real parts, then for every initial condition x(0),
limt→∞ x(t) = cf for some c ∈ C.
Proof: The solution of the closed-loop system (8)
is x(t) = e(A+BF )t x(0). Since A + BF has a simple
eigenvalue 0, with the corresponding eigenvector f , and other
eigenvalues have negative real parts, it follows from the
standard linear systems analysis that x(t) → (w⊤ x(0))f as
t → ∞. Here w ∈ Cn is the left-eigenvector of A+BF with
respect to the eigenvalue 0. Therefore limt→∞ x(t) = cf ,
where c := w⊤ x(0).
In view of Proposition 1, if the specified eigenvalues and
the corresponding eigenvectors may be assigned by state
feedback u = F x, then Problem 1 is solved. To this end,
we resort to eigenstructure assignment.
III. M AIN R ESULTS
In this section, we solve Problem 1, the formation control
problem of multi-agent systems, by the method of eigenstructure assignment. The following is our first main result.
Theorem 1: Consider the multi-agent system (7) and let
f be a desired formation configuration. Then there always
exists a state feedback control u = F x that solves Problem 1,
i.e.
(∀x(0) ∈ Cn )(∃c ∈ C) lim x(t) = cf.
t→∞
Proof: By Proposition 1, the multi-agent system (7)
with u = F x achieves a formation configuration f ∈
Cn if the closed-loop matrix A + BF has the following
eigenstructure: (i) its eigenvalues {λ1 , λ2 , . . . , λn } satisfy
0 = λ1 < |λ2 | ≤ · · · ≤ |λn |, and (∀i ∈ [2, n])Re(λi ) < 0
(9)
(ii) the corresponding eigenvectors
{v1 , v2 , . . . , vn } are linearly independent, and v1 = f
(10)
So we must verify that the above eigenstructure is assignable
by state feedback u = F x for the multi-agent system (7).
Note that except for (λ1 , v1 ) which is fixed, we have freedom
to choose (λi , vi ), i ∈ [2, n]. For simplicity we let λi be all
distinct, and thus Lemma 1 can be applied.
In (7), we have A = diag(a1 , . . . , an ), B =
diag(b1 , . . . , bn ), and bi 6= 0 for all i ∈ [1, n]. Thus it is easily
checked that the pair (A, B) is controllable and KerB = 0.
To show that there exists F such that (A + BF )vi = λi vi ,
for each i ∈ [1, n] (with λi , vi specified above), it suffices to
verify the condition (4) of Lemma 1.
First, for λ1 = 0, we find a basis for
Ker[λ1 I − A B] = Ker[−A B]
and derive N1 (λ1 ) = B and N2 (λ1 ) = A (N1 (·), N2 (·) in
(2)). Thus ImN1 (λ1 ) = Cn , and hence v1 = f ∈ ImN1 (λ1 ),
i.e. the condition (4) of Lemma 1 holds.
Next, let i ∈ [2, n]; we find a basis for Ker[λi I − A B]
and derive N1 (λi ) = B and N2 (λi ) = A − λi I. So again
ImN1 (λi ) = Cn , and vi ∈ ImN1 (λi ), i.e. the condition (4) of
control gain). As an example, for (ii) in Example 1 assign
the second largest eigenvalue λ2 = −2 (originally −1), and
change v4 = [0 0 21 − 1]. This results in a new feedback
matrix
−4 1 2 1
−1 −2 2 1
F2′ =
−2 0 0 2
0
1 2 −3
Fig. 1.
Example 1
Lemma 1 holds. Therefore, we conclude that there always
exists a state feedback u = F x such that the multi-agent
system (7) achieves the formation configuration f .
In the proof we considered distinct eigenvalues λi , i ∈
[1, n]; hence the control gain matrix F may be computed
by (5). The computed F in turn gives rise to the agents’
communication graph G. The following is an illustrative
example.
Example 1: Consider the multi-agent system (7) of 4
single integrators (that is, ai = 0 and bi = 1, i = 1, ..., 4).
(i) Square
formation with f = [1 j − 1 − j]⊤
√
(j = −1). Let the desired closed-loop eigenvalues be
λ1 = 0, λ2 = −1, λ3 = −2, λ4 = −3 and the corresponding
eigenvectors be v1 = f , v2 = [1 1 0 0]⊤ , v3 =
[0 1 1 0]⊤ , v4 = [0 0 1 0]⊤ . By (5) one computes the
control gain matrix F1 , which determines the corresponding
communication graph G1 (see Fig. 1).
Observe that F1 contains complex entries, which may be
viewed as control gains for the real and imaginary axes,
respectively, or scaling and rotating gains on the complex
plane. Also note that G1 has a spanning tree with node 4
the root, and the computed feedback control u = F x can be
implemented by the four agents individually.
(ii) Consensus with f = [1 1 1 1]⊤ . Let the desired
eigenvalues be λ1 = 0, λ2 = −1, λ3 = −3, λ4 = −4 and
the corresponding eigenvectors be v1 = f , v2 = [1 1 0 1]⊤ ,
v3 = [1 0 0 1]⊤ , v4 = [0 0 1 − 1]⊤ . Again by (5) one
computes the control gain matrix F2 and the corresponding
graph G2 (see Fig. 1).
Note that in this case F2 is real and G2 strongly connected.
But unlike the usual consensus algorithm (e.g. [2]), −F2
is not a graph Laplacian matrix for the entries (2, 1) and
(3, 1) are positive. Thus our eigenstructure assignment based
approach may generate a larger class of consensus algorithms
with negative weights.
We remark that in our approach, the convergence speed
to the desired formation configuration is assignable. This is
because the convergence speed is dominated by the eigenvalue λ2 , with the second largest real part, of the closedloop system ẋ = (A + BF )x; and in our approach λ2
is freely assignable. The smaller the Re(λ2 ) is, the faster
the convergence to formation occurs (at the cost of higher
which has zero entries at the same locations as F2 . Thus with
the same topology, F2′ achieves faster convergence speed.
As we have seen in Example 1, the feedback matrix
F ’s off-diagonal entries, which determine the topology of
G, are dependent on the choice of eigenvalues as well
as eigenvectors. Namely different sets of eigenvalues and
eigenvectors result in different inter-agent communication
topologies. Our next result characterizes a precise relation
between the eigenvalues/eigenvectors and the topologies.
Theorem 2: Consider the multi-agent system (7) and f
a desired formation configuration. Let the eigenstructure λi
and vi (i = 1, ..., n) be as in (9) and (10), and denote the
rows of [v1 · · · vn ]−1 by vi∗ . Then the communication graph
G = (V, E) of the (closed-loop) multi-agent system is such
that
(i1 , j1 ), . . . , (iK , jK ) ∈
/E
(K ≥ 1)
if and only if the vector
[λ2 · · · λn ]⊤
is orthogonal to the subspace spanned by the K vectors:
[v2 i1 v2∗ j1 · · · vn i1 vn∗ j1 ]⊤ , · · · , [v2 iK v2∗ jK · · · vn iK vn∗ jK ]⊤ .
Proof: For each λi , i ∈ [1, n], we derive from (3) that
(λi I − A)N1 (λi ) + BN2 (λi ) = 0
Choose N1 (λi ) = B and N2 (λi ) = −(λi I − A) to satisfy
the above equation. Then ki = N1−1 (λi )vi = B −1 vi and
wi = −N2 (λi )ki = (λi I − A)B −1 vi . By (5) we have
F = [w1 · · · wn ][v1 · · · vn ]−1
= [(λ1 I − A)B −1 v1 · · · (λn I − A)B −1 vn ][v1 · · · vn ]−1
= B −1 (−A[v1 · · · vn ] + [λ1 v1 · · · λn vn ])[v1 · · · vn ]−1
= −B −1 A + B −1 [v1 · · · vn ]diag(λ1 , . . . , λn )[v1∗ · · · vn∗ ]⊤ .
Thus the closed-loop matrix
A + BF = [v1 · · · vn ]diag(λ1 , . . . , λn )[v1∗ · · · vn∗ ]⊤ .
(11)
The (i, j)-entry of A + BF is
∗
∗
∗
(A + BF )ij = λ1 v1i v1j
+ λ2 v2i v2j
+ · · · + λn vni vnj
∗
∗
= λ2 v2i v2j
+ · · · + λn vni vnj
= [λ2 · · ·
∗
λn ][v2i v2j
···
(λ1 =
∗ ⊤
vni vnj ] .
0)
Therefore (A + BF )i1 j1 = · · · = (A + BF )iK jK = 0, i.e. in
the communication graph (i1 , j1 ), . . . , (iK , jK ) ∈
/ E, if and
only if the vector [λ2 · · · λn ]⊤ is orthogonal to each of the
following K vectors:
[v2 i1 v2∗ j1 · · · vn i1 vn∗ j1 ]⊤ , · · · , [v2 iK v2∗ jK · · · vn iK vn∗ jK ]⊤ .
Namely [λ2 · · · λn ]⊤ is orthogonal to the subspace spanned
by these K vectors.
Once the desired eigenvalues and eigenvectors are chosen,
Theorem 2 provides a necessary and sufficient condition
to check the interconnection topology among the agents,
without actually computing the feedback matrix F . On
the other hand, the problem of choosing an appropriate
eigenstructure to match a given topology is more difficult,
inasmuch as there are many free variables to be determined in
the eigenvalues and eigenvectors. While we shall investigate
the general problem of eigenstructure design for imposing
particular topologies in our future work, in the next subsection, nevertheless, we show that choosing certain appropriate
eigenstructures results in certain special (sparse) topologies.
With these topologies the synthesized control u = F x may
be implemented in a distributed fashion.
A. Special Topologies
We show how to derive the following three types of special
topologies by choosing appropriate eigenstructures.
1) Star Topology: A directed graph G = (V, E) is a star
topology if there is a single root node, say node 1, and
E = {(1, i)|i ∈ [2, n]}. Thus all the other nodes receive
information from, and only from, the root node 1. In terms
of the total number of edges, a star topology is one of the
sparsest topologies, with the least number (n − 1) of edges,
that contain a spanning tree. Now consider the following
eigenstructure.
eigenvalues: λ1 = 0, λ2 , . . . , λn distinct
and Re(λ2 ), . . . , Re(λn ) < 0
f1 0 · · ·
f2 1 · · ·
eigenvectors: [v1 v2 · · · vn ] = .
.. . .
..
.
.
(independent)
fn 0 · · ·
we derive
0
0
λ2
..
.
0
− λf2 f2
1
A + BF =
..
.
− λnf1fn
···
···
..
.
···
0
0
..
.
λn
.
Therefore the corresponding graph G is a star topology with
node 1 the root.
2) Cyclic Topology: A directed graph G = (V, E) is a
cyclic topology if E = {(1, 2), (2, 3), ..., (n − 1, n), (n, 1)}.
Consider the following eigenstructure.
eigenvalues: {λ1 , λ2 , . . . , λn } = {0, ω − 1, . . . , ω n−1 − 1}
eigenvectors: [v1 v2 · · · vn ] =
(independent)
where ω := e2πj/n (j =
√
f1
f2
f3
..
.
f1
f2 ω
f3 ω 2
..
.
···
···
···
..
.
f1
f2 ω n−1
f3 ω 2(n−1)
..
.
fn
fn ω n−1
···
fn ω (n−1)(n−1)
(13)
−1).
Proposition 3: Consider the multi-agent system (7). If the
eigenstructure (13) is used in the synthesis of feedback
control u = F x, then Problem 1 is solved and the resulting
G is a cyclic topology.
Proof: Following the same lines as the proof of
Proposition 2, and using the eigenstructure in (13) and
1
1
1
···
f1
f2
fn
1
ω̄
ω̄ n−1
···
f1
f2
fn
11
ω̄ 2
ω̄ 2(n−1)
−1
···
[v1 v2 · · · vn ] = f1
f2
fn
n .
..
..
..
.
.
(n−1)
(n−1)(n−1)
ω̄
ω̄
1
·
·
·
f1
f2
fn
we derive
0
0
..
.
1
(12)
Proposition 2: Consider the multi-agent system (7). If the
eigenstructure (12) is used in the synthesis of feedback
control u = F x, then Problem 1 is solved and the resulting
graph G is a star topology.
Proof: First, it follows from the proof of Theorem 1
that the eigenstructure (12) can be assigned to the closedloop matrix A + BF . Then by Proposition 1, Problem 1 is
solved.
Now proceed analogously to the proof of Theorem 2
and derive A + BF as in (11). Substituting into (11) the
eigenvalues and eigenvectors in (12), as well as
1
0 ··· 0
f1
− f2 1 · · · 0
f1
−1
[v1 v2 · · · vn ] =
.. . .
..
..
. .
.
.
− ffn1 0 · · · 1
−1 ff21
0 −1
..
..
.
.
A + BF =
0
fn
f1
0
0
0
f2
f3
..
.
0
0
···
···
..
.
..
.
0
0
..
.
fn−1
fn
···
−1
.
By inspection, we conclude that the corresponding graph G
is a cyclic topology.
3) Line Topology: A directed graph G = (V, E) is a
(directed) line topology if there is a single root node, say
node 1, and E = {(1, 2), (2, 3), ..., (n − 1, n)}. A line
topology is also one of the sparsest topologies containing
a spanning tree. Now consider the following eigenstructure.
eigenvalues: λ1 = 0, λ2 = · · · = λn = −1
f1
f2
..
.
eigenvectors: [v1 v2 · · · vn ] =
(independent)
fn−1
fn
0
0
..
.
0
−fn
···
···
.
..
···
0
−f2
..
.
−fn−1
−fn
(14)
Proposition 4: Consider the multi-agent system (7). If the
eigenstructure (14) is used in the synthesis of feedback
control u = F x, then Problem 1 is solved and the resulting
G is a line topology.
Note that in (14) we have repeated eigenvalues (λ2 )
and the corresponding generalized eigenvectors. As a result,
Lemma 1 and (5) for computing the control gain matrix
F cannot be applied to this case. Instead, we resort to
the generalized method of [17], and provide a proof of
Proposition 4 in Appendix.
By (16) we set the new feedback matrix
B. Topology Constrained Eigenstructure Assignment
namely F̂ is the same as the originally computed F except
for the ith row replaced by F̂i⊤ computed in (16) and F̂ij =
0.1
While the new feedback matrix F̂ satisfies the imposed
topological constraint, the eigenstructure of A + B F̂ is
generally different from that of A + BF . Therefore, we
must verify if the new eigenvalues/eigenvectors still satisfy
(9) and (10), i.e. achieve formation control. This verification
need not always be successful, but in case it does turn
out successful, we are guaranteed to achieve the desired
formation with a feedback matrix satisfying the imposed
topological constraint. We illustrate the above method by the
following example.
Example 2: Consider the multi-agent system (7) of 5 single integrators (that is, ai = 0 and bi = 1, i = 1, ..., 5), and
the desired formation is simply consensus (f = [1 1 1 1 1]⊤ ).
Choose the following eigenvalues and eigenvectors
We end this section by presenting an alternative approach
to imposing topological constraints on eigenstructure assignment.
Suppose that we have computed by (5) a feedback matrix
F to achieve a desired formation, i.e. the closed-loop matrix
(A+BF )’s eigenvalues λ1 , ..., λn and eigenvectors v1 , ..., vn
satisfy (9) and (10). Now assume that a topological constraint
is imposed such that agent i cannot receive information from
agent j (for reasons such as cost or physical impossibility).
But unfortunately, in the computed F the (i, j)-entry Fij 6=
0. Thus our goal is to derive a new feedback matrix F̂ , by
suitably modifying F , such that F̂ij = 0. In doing so, the
new closed-loop matrix A+B F̂ will generally have different
eigenvalues λ̂1 , ..., λ̂n and eigenvectors v̂1 , ..., v̂n . Hence we
must check if these new eigenvalues and eigenvectors still
satisfy (9) and (10).
Our approach proceeds as follows, which is inspired by
the “constrained feedback” method in [18]. Writing V :=
[v1 · · · vn ] and J := diag(λ1 , ..., λn ), we have from (11):
(A + BF )V = V J
⇒BF V = V J − AV
⇒F V = B −1 (V J − AV ).
F1
..
.
F̂ :=
F̂i1
···
F̂i(j−1)
0
..
.
F̂i(j+1)
···
Fn
F̂in
(17)
eigenvalues: {λ1 , λ2 , λ3 , λ4 , λ5 } = {0, −1, −2, −3, −4}
1
1
eigenvectors: [v1 v2 v3 v4 v5 ] =
1
1
1
0
1
0
0
0
0
0
1
0
1
0
0
1
1
0
0
1
−1
0
1
the feedback matrix F to achieve
Let ψ := B −1 (V J − AV ) and denote by ψi the ith row
of ψ, also Fi the ith row of F . Then the above equation is
rewritten in terms of Kronecker product and row stacking as
follows:
⊤ ⊤
⊤
ψ1
F1
V
0
.
.
.
..
.. = .. .
and compute by (5)
consensus:
0
2.5
F =
2
3
3
To constrain the (i, j)-entry of F to be zero, we focus on
the equation V ⊤ Fi⊤ = ψi⊤ . Deleting Fij from Fi⊤ as well
as the jth column of V ⊤ , we obtain the reduced equation:
Suppose that the topological constraint is that agent 2
cannot receive information from agent 4, i.e. the (2, 4)-entry
of F (F24 = −1.5) must be set to be zero. For this we first
derive equation V ⊤ Fi⊤ = ψi⊤ as follows:
0
V⊤
Fn⊤
V̂ ⊤ F̂i⊤ = ψi⊤
ψn⊤
(15)
where V̂ ⊤ ∈ Cn×(n−1) is the matrix V ⊤ with jth column deleted and F̂i⊤ ∈ Cn−1 the vector Fi⊤ with jth
element deleted. Now view the entries of F̂i⊤ as the unknowns (i.e. (15) contains n equations with n − 1 unknowns
F̂i1 , ..., F̂i(j−1) F̂i(j+1) , ..., F̂in ). Using the pseudoinverse of
V̂ ⊤ , denoted by (V̂ ⊤ )† , we derive
F̂i⊤ = (V̂ ⊤ )† ψi⊤ .
(16)
1
0
0
0
0
1
1
0
0
1
1
0
1
1
−1
0
0
0
0
−1 1.5 −1.5 −1.5
0 −3
0
1
.
0
0
−3
0
0
1
−1
−3
1
0
0
1
0
1 F21
0
F22 −1
0
1
F23 = 0 .
0 F24 0
1 F25
−4
1 The above derivation may be readily extended to deal with more than
one topological constraint.
Deleting F24 from Fi⊤ and the 4th column of V ⊤ , we obtain
the reduced equation (15):
0
1 1 1 1 F̂
−1
0 1 0 0 21
0 0 1 1 F̂22 = 0 .
0
0 0 1 0 F̂23
−4
0 1 −1 1 F̂25
Solve this equation for the unknowns F̂21 , F̂22 , F̂23 , F̂25 , we
compute by (16):
F̂21
1.8571
F̂22 −1.4286
F̂2⊤ =
F̂23 = 0.8571 .
−1.2857
F̂25
Finally set as (17) the new feedback matrix
0
0
0
0
0
1.8571 −1.4286 0.8571 0 −1.2857
.
0
−3
0
1
F̂ =
2
3
0
0
−3
0
3
0
1
−1
−3
A. Arbitrary Inter-Agent Connections
First we consider the case where the agents have arbitrary
initial interconnection, while keeping the assumption that
they are individually stabilizable. That is, we consider the
following multi-agent system
ẋ = Ax + Bu
n
n
(18)
n×n
where x ∈ C , u ∈ C , A ∈ R
and B =
diag(b1 , . . . , bn ) (bi 6= 0). The matrix A is now an arbitrary
real matrix, modeling an arbitrary (initial) communication
topology among the agents.
It turns out, despite the general A matrix, that the same
conclusion as Theorem 1 holds.
Theorem 3: Consider the multi-agent system (18) and let
f be a desired formation configuration. Then there always
exists a state feedback control u = F x that achieves
formation control, i.e.
(∀x(0) ∈ Cn )(∃c ∈ C) lim x(t) = cf.
t→∞
Proof: The proof proceeds similarly to that of Theorem 1. First, for the diagonal matrix B in (18), we have
(regardless of A) that the pair (A, B) is controllable and
Observe that (only) the second row of F̂ has entries all KerB = 0.
It is left to verify if the (distinct) eigenvalues λ1 , ..., λn
different from that of the original F , and the (2, 4)-entry
F̂24 = 0. Moreover the eigenstructure of the new closed- and eigenvectors v1 , ..., vn as in (9) and (10) satisfy the
condition (4) of Lemma 1. Let i ∈ [1, n]. Since
loop matrix A + B F̂ (= F̂ ) is
B
λi I − A B
=0
eigenvalues: {λ̂1 , λ̂2 , λ̂3 , λ̂4 , λ̂5 } = {0, −1.4286, −2, −3, −4}
B −1 (A − λi I)B
eigenvectors: [v̂1 v̂2 v̂3 v̂4 v̂5 ]
we find a basis for Ker[λi I − A B] and derive
1
1
=
1
1
1
0
0
0
0
2.2361 −1.0477 −0.8047 −1.1352
0
−1.3969 1.4753
1.3623
.
0
0
1.4753
0
0
−1.3969
0
−1.3623
Hence these new eigenvalues/eigenvectors still satisfy (9)
and (10), and therefore consensus is achieved despite of the
imposed topological constraint. (In fact, since λ̂2 < λ2 , we
have faster convergence with the new F̂ .)
IV. G ENERAL M ULTI -AGENT S YSTEMS
So far we have considered the multi-agent system in
(7), where the agents are uncoupled and each is (self-)
stabilizable (the matrices A, B are diagonal and B’s diagonal
entries nonzero). For (7) we have shown in Theorem 1 that a
state feedback control, based on eigenstructure assignment,
always exists to drive the agents to a desired formation.
More generally, however, the agents may be initially
interconnected (owing to physical coupling or existence of
communication channels), and/or some agents might not be
capable of stabilizing themselves (though they can receive
information from others). It is thus of interest to inquire,
based on the eigenstructure assignment approach, what conclusions we can draw for formation control in these more
general cases.
N1 (λi ) = B
N2 (λi ) = B −1 (A − λi I)B.
Thus ImN1 (λi ) = Cn , and vi ∈ ImN1 (λi ), i.e. the condition (4) of Lemma 1 holds. Therefore, there always exists
a state feedback u = F x such that the multi-agent system
(18) achieves the formation configuration f .
In the proof of Theorem 1, we derived N1 (λi ) = B and
N2 (λi ) = A − λi I. Since A was diagonal and diagonal
matrices commute, there held
B
λi I − A B
= 0.
A − λi I
For a general A as in Theorem 3, we have found instead
N1 (λi ) = B and N2 (λi ) = B −1 (A − λi I)B that deal
with arbitrary A without depending on the commutativity
of matrices.
Theorem 3 asserts that, as long as the agents are individually stabilizable, formation control is achievable by
eigenstructure assignment regardless of how the agents are
initially interconnected. The final topology, on the other
hand, is in general determined by the initial connections
‘plus’ additional ones resulted from the chosen eigenvalues/eigenvectors (as has been discussed in Section III). It
may also be possible, however, that the initial connections are
‘decoupled’ by the corresponding entries of the synthesized
feedback matrix. This is illustrated by the following example.
Consider again Example 1(i), but change A from the zero
matrix to the following
0
0.5 0
0
0 0
A=
−0.5 0 0
0
0 2
0
0
;
0
0
bi 6= 0, i ∈ [1, m].
that is, agents 1 and 2, 3 and 1, 4 and 3 are initially interconnected. Assigning the same eigenstructure as in Example 1(i),
we obtain the feedback matrix
−1 −0.5 0
j
1
−2
0 −2 − j
.
F =
−0.5
1
−3 1 − 2j
0
0
−2
0
Then the closed-loop matrix A+BF (where B is the identity
matrix) is
−1 0
1 −2
A + BF =
−1 1
0
0
Theorem 4: Consider the multi-agent system (19) and let
f be a desired formation configuration. Also let λ1 , ..., λn
and v1 , ..., vn be the desired eigenvalues and eigenvectors
satisfying (9) and (10). If
(i) the pair (A, B) is controllable, and
(ii) ImB ⊆ Im(A − λi I) for all i ∈ [1, n], and
(iii) vi ∈ ImN1 (λi ) for all i ∈ [1, n], where N1 (λi )
satisfies (A − λi I)N1 (λi ) = B,
then there exists a state feedback control u = F x that
achieves formation control, i.e.
(∀x(0) ∈ Cn )(∃c ∈ C) lim x(t) = cf.
t→∞
Proof: First observe from (19) that KerB = 0, since B’s
columns are linearly independent. Now let i ∈ [1, n]. Since
(A, B) is controllable (condition (i)), there exist N1 (λi ) and
N2 (λi ) such that (3) holds. Setting N2 (λi ) = I, we derive
from (3) the following matrix equation
0
j
0 −2 − j
−3 1 − 2j
0
0
which is the same as the feedback matrix F1 (as well as
the closed-loop matrix) in Example 1(i). Thus despite the
initial coupling, the final topology turns out to be the same
as that of Example 1(i). In particular, in the final topology
agents 1 and 2, 4 and 3 are uncoupled – their initial couplings
are ‘canceled’ by the corresponding entries of the feedback
matrix F .
B. Existence of Non-Stabilizable Agents
Continuing to consider arbitrary initial topology (i.e. general A), we further assume that some agents cannot stabilize
themselves (i.e. the corresponding diagonal entries of B
in (18) are zero). Equivalently, the non-stabilizable agents
have no control inputs. In this case, achieving a desired
formation is possible only if those non-stabilizable agents
may take advantage of information received from others (via
connections specified by A). This is a problem of global
formation stabilization with locally unstabilizable agents,
which has rarely been studied in the literature. We aim
to provide an answer using our top-down eigenstructure
assignment based approach.
Without loss of generality, assume that only the first m
(< n) agents are stabilizable. Thus the multi-agent system
we consider in this subsection is
ẋ = Ax + Bu
where x ∈ Cn , u ∈ Cm , A ∈ Rn×n and
b1
..
.
bm
B=
∈ Rn×m (m < n)
0
(19)
(A − λi I) N1 (λi ) = B.
(20)
Since ImB ⊆ Im(A − λi I) (condition (ii)), this equation has
a solution N1 (λi ) (which is determined by A, B, λi ). Finally,
since vi ∈ ImN1 (λi ) (condition (iii)), the condition (4) of
Lemma 1 is satisfied. Therefore the desired eigenvalues and
eigenvectors satisfying (9) and (10) may be assigned by a
state feedback control u = F x, i.e. formation control is
achieved.
Theorem 4 provides sufficient conditions to ensure solvability of the formation control problem for multi-agent
systems with non-stabilizable agents. In the following we
illustrate this result by working out a concrete example,
where A represents a directed line topology and there is only
one agent that is stabilizable (i.e. B is simply a vector).
Example 3: Consider the multi-agent system ẋ = Ax +
Bu with
a1 0
0
b1
â2 a2
0
A=
, B = ..
.. ..
.
.
.
0
ân
an
0
where a1 , ..., an , â2 , ..., ân , b1 are nonzero. Namely A represents a directed line topology with agent 1 the root, and B
means that only agent 1 is stabilizable. Thus this is a singleinput multi-agent system – by controlling only the root of a
directed line.
First, it is verified that (A, B) is controllable, i.e. condition (i) of Theorem 4 is satisfied. To ensure condition (ii),
ImB ⊆ Im(A − λi I), it suffices to choose each desired
eigenvalue λi (i ∈ [1, n]) such that λi 6= aj for j ∈ [1, n] (i.e.
every eigenvalue is distinct from the nonzero diagonal entries
of A). At the same time, these eigenvalues must satisfy (9).
Having condition (ii) hold, equation (20) has a solution
N1 (λi ). Let us solve (20)
b1
a1 − λi
0
0
â2
a
−
λ
2
i
0
N1 (λi ) = .
.
.
..
..
..
0
0
ân an − λi
and obtain an explicit solution
b1
a1 −λi
â2 b1
− (a1 −λi )(a2 −λi )
.
N1 (λi ) =
..
.
â2 ···ân b1
n−1
(−1)
(a1 −λi )(a2 −λi )···(an −λi )
Hence to ensure condition (iii) of Theorem 4, we must
choose each desired eigenvector vi (i ∈ [1, n]) such that
vi ∈ ImN1 (λi ) and (10) is satisfied. In particular, for i = 1
we have λi = 0 and v1 = f ; thus v1 ∈ ImN1 (λ1 ) means
that the formation vector f must be such that
i⊤
h
n b1
(21)
f = c ab11 − aâ12ab12 · · · (−1)n−1 aâ12a···â
2 ···an
where c ∈ C (c 6= 0). This characterizes the set of
all achievable formation configurations for the single-input
multi-agent system under consideration.
We conclude that, by controlling only one agent, indeed
the root agent of a directed line topology, it is not possible
to achieve arbitrary formation configurations but those determined by the nonzeros entries of the matrices A and B in
the specific manner as given in (21).
V. H IERARCHICAL E IGENSTRUCTURE A SSIGNMENT
In the previous sections, we have shown that a control gain
matrix F can always be computed (as long as every agent is
stabilizable) such that the multi-agent formation Problem 1
is solved. Computing such F by (5) (see the eigenstructure
assignment procedure in Section II) has complexity O(n3 ),
where n is the number of agents. Consequently the computation cost becomes expensive as the number of agents
increases.
To address this issue of centralized computation, we
propose in this section a hierarchical synthesis procedure.
We shall show that the control gain matrix F computed by
this hierarchical procedure again solves Problem 1, which
moreover significantly improves computational efficiency
(empirical evidence provided in Section VII).
For clarity of presentation, let us return to consider the
multi-agent system (7), and Problem 1 with the desired
formation configuration f ∈ Cn (f 6= 0). Partition the agents
into l (≥ 1) pairwise disjoint groups. Let group k (∈ [1, l])
have nk (≥ 1) agents; nk may be different and Σlk=1 nk = n.
Now for the configuration f and x, u, A, B in (7), write
in accordance with the partition (possibly with reordering)
w1
y1
g1
..
..
..
f = . ,x = . ,u = . ,
gl
A1
A=
..
.
wl
B1
,
B
=
yl
Al
..
.
Bl
where gk , yk , wk ∈ Cnk and Ak , Bk ∈ Cnk ×nk , k ∈ [1, l].
Thus for each group k, the dynamics is
ẏk = Ak yk + Bk wk .
(22)
For later use, also write gk1 , yk1 , wk1 (resp. Ak1 , Bk1 )
for the first component of gk , yk , wk (resp. (1,1)-entry
of Ak , Bk ), and g0 := [g11 · · · gl1 ]⊤ , y0 := [y11 · · · yl1 ]⊤ ,
w0 := [w11 · · · wl1 ]⊤ , A0 := diag(A11 , · · · , Al1 ), B0 :=
diag(B11 , · · · , Bl1 ).
The vector gk (k ∈ [1, l]) is the local formation configuration for group k, while g0 is the formation configuration
for the set of the first component agent from each group. We
assume that these configurations are all nonzero, i.e. gk 6= 0
for k ∈ [1, l] and g0 6= 0. Now we present the hierarchical
synthesis procedure.
(i) For each group k ∈ [1, l] and its dynamics (22),
compute Fk by (5) such that Ak + Bk Fk has a simple
eigenvalue 0 with the corresponding eigenvector gk , and
other eigenvalues have negative real parts; moreover the
topology defined by Fk has a unique root node yk1 (e.g.
star or line by the method given in Section III.A).
(ii) Treat {yk1 |k ∈ [1, l]} (the group leaders) as a higherlevel group, with the dynamics
ẏ0 = A0 y0 + B0 w0 .
(23)
Compute F0 ∈ Cl×l by (5) such that A0 + B0 F0 has a
simple eigenvalue 0 with the corresponding eigenvector g0 ,
and other eigenvalues have negative real parts.
(iii) Set the control gain matrix F := F low + F high,
where
F1
..
F low :=
.
Fl
and F high is partitioned according to F low , with each block
(i, j), i, j ∈ [1, l]
1 0 ··· 0
0 0 · · · 0
(F high )ij = (F0 )ij · . . .
..
.
.
.
. .
. .
(F0 )ij
0
= .
..
0
0 0 ··· 0
0 ··· 0
0 · · · 0
.. . .
.. .
.
.
.
0 ··· 0
The computational complexity of Step (i) is O(n̂3 ), where
n̂ := max{n1 , ..., nl }; and Step (ii) is O(l3 ). Let ñ :=
max{n̂, l}. Then the complexity of the entire hierarchical
synthesis procedure is O(ñ3 ). With proper group partition,
this hierarchical procedure can significantly reduce computation time, as demonstrated by an empirical study in
Section VII.
Note that in Step (i) of the above procedure, requiring
the topology defined by each Fk to have a unique root, i.e.
a single leader, is for simplicity of presentation. It can be
extended to the case of multiple leaders, and then in Step (ii)
treat all the leaders at the higher-level. On the other hand, the
number of leaders should be kept small such that the highlevel control synthesis in Step (ii) can be done efficiently.
The correctness of the hierarchical synthesis procedure is
asserted in the following.
Theorem 5: Consider the multi-agent system (7) and let f
be a desired formation configuration. Then the state feedback
control u = F x synthesized by the hierarchical synthesis
procedure solves Problem 1, i.e.
(∀x(0) ∈ Cn )(∃c ∈ C) lim x(t) = cf.
t→∞
Proof: For each k ∈ [1, l] let yk′ := [yk2 · · · yknk ]⊤ and
:= [gk2 · · · gknk ]⊤ ∈ Cnk −1 . Thus yk′ and gk′ are yk , gk
with the first element removed. By Step (i) of the hierarchical
synthesis procedure, since yk1 is the unique root node, we
can write ẏk = (Ak + Bk Fk )yk as follows:
yk1
0 |
0
ẏk1
′ .
′ =
y
ẏ Hk
Gk
gk′
k
k
Then by the eigenstructure of Ak +Bk Fk , all the eigenvalues
of Gk have negative real parts and
gk1
Hk
Gk
(24)
gk′ = 0.
⊤
⊤
Reorder x = [y1⊤ · · · yl⊤ ]⊤ to get x̂ := [y0⊤ y1′ · · · yn′ ]⊤ .
Then there is a permutation matrix that similarly transforms
the control gain matrix F in Step (iii) to F̂ , and x̂˙ = F̂ x̂ is
ẏ0
y0
A0 + B0 F0
0
ẏ1′ H1
y1′
G1
=
..
.. .
..
..
.
.
.
.
′
Hl
Gl
ẏl
yl′
It then follows from the eigenstructure of A0 + B0 F0
assigned in Step (ii) and (24) above that the matrix F̂ has
a simple eigenvalue 0 with the corresponding eigenvector
′
′
fˆ := [g0⊤ g1⊤ · · · gl ⊤ ]⊤ , and other eigenvalues have negative
real parts. Hence (cf. Proposition 1),
(∀x̂(0) ∈ Cn )(∃ĉ ∈ C) lim x̂(t) = ĉfˆ.
t→∞
Since x̂ (resp. fˆ) is just a reordering of x (resp. f ), the
conclusion follows and the proof is complete.
VI. R IGID F ORMATION
AND
C IRCULAR M OTION
In this section we show that our method of eigenstructure
assignment may be easily extended to address problems of
rigid formation and circular motion.
A. Rigid Formation
First, we extend our method to study the problem of
achieving a rigid formation, one that has translational and
rotational freedom but fixed size.
Problem 2: Consider the multi-agent system (7) and specify f ∈ Cn (f 6= 0) and d > 0. Design a control u such that
for every initial condition x(0), limt→∞ x(t) = c1 + df ejθ
for some c ∈ C and θ ∈ [0, 2π).
In Problem 2, the goal of the multi-agent system (7) is to
achieve a rigid formation df , with translational freedom in
c, rotational freedom in θ, and fixed size d.
We now present the rigid-formation synthesis procedure.
(i) Compute F by (5) such that A + BF has two eigenvalues 0 with the corresponding (non-generalized) eigenvectors
1 and f , and other eigenvalues have negative real parts;2
moreover the topology defined by F is 2-rooted3 with exactly
2 roots (say nodes 1 and 2). This topology may be achieved
by assigning appropriate eigenstructures, e.g.
eigenvalues: λ1 = λ2 = 0, λ3 , . . . , λn distinct
and Re(λ3 ), . . . , Re(λn ) < 0
1 f1 0
1 f2 0
eigenvectors: [v1 v2 v3 · · · vn ] = 1 f3 1
..
..
..
(independent)
.
.
.
1 fn
0
···
···
···
..
.
0
0
0
..
.
···
1
(25)
(ii) Let f1 , f2 be the first two components of f , and set
ẋ1
(x2 − x1 )(||x2 − x1 ||2 − d2 |f2 − f1 |2 )
=
=: r(x1 , x2 ).
2
2
2
ẋ2
(x1 − x2 )(||x1 − x2 || − d |f1 − f2 | )
(iii) Set the control
u := F x + B −1
r(x1 , x2 )
0
.
(26)
The idea of the above synthesis procedure is to first use
eigenstructure assignment to achieve a desired formation
configuration with two leaders, and then control the size of
the formation by stabilizing the distance between the two
leaders to the prescribed d. The latter is inspired by [13].
Our result is the following.
Proposition 5: Consider the multi-agent system (7) and
let f ∈ Cn , d > 0. Then the control u in (26) synthesized
by the rigid-formation synthesis procedure solves Problem 2
for all initial conditions x(0) with x1 (0) 6= x2 (0).
2 For repeated eigenvalues with non-generalized eigenvectors, the eigenstructure assignment result Lemma 1 and the computation of control gain
matrix F in (5) remain the same as for the case of distinct eigenvalues.
3 A 2-rooted topology is one where there exist 2 nodes from which every
other node v can be reached by a directed path after removing an arbitrary
node other than v [13].
(∀x(0) ∈ Cn )(∃c, c′ ∈ C) lim x(t) = c1 + c′ f.
t→∞
Moreover, choosing the eigenstructure in (25) and following
similarly to the proof of Proposition 2, we can show that
the resulting topology defined by F is 2-rooted with nodes
1 and 2 the only two roots.
With the 2-rooted topology and the design in Step (ii),
it follows from [13, Theorem 4.4] that for all x(0) with
x1 (0) 6= x2 (0), we have c′ = df ejθ for some θ ∈ [0, 2π).
An illustrative example of achieving rigid formations is
provided in Section VII below.
B. Circular Motion
We apply the eigenstructure assignment approach to solve
a cooperative circular motion problem, in which the agents
all circle around the same center while keeping a desired
formation configuration. This cooperative task may find
useful applications in target tracking and encircling (e.g. [25],
[26]).
Problem 3: Consider the multi-agent system (7) and specify f ∈ Cn (f 6= 0) and b ∈ R (b 6= 0). Design a state
feedback control u = F x such that for every initial condition
x(0),√limt→∞ x(t) = c1 + c′ f ebjt for some c, c′ ∈ C and
j = −1.
In Problem 3, the goal is that the agents of (7) all
circle around the same center c at rate b, while keeping the
formation configuration f scaled by |c′ |.
Our result is the following.
Proposition 6: Consider the multi-agent system (7) and
let f ∈ Cn , b ∈ R. Then there always exists a state feedback
control u = F x that solves Problem 3.
Proof: By a similar argument to that in the proof of
Theorem 1, we can show that for (7) there always exists F
such that (A + BF ) has the following eigenstructure:
eigenvalues: λ1 = 0, λ2 = bj, λ3 , . . . , λn distinct
and Re(λ3 ), . . . , Re(λn ) < 0
eigenvectors: v1 = 1, v2 = f, {v1 , v2 , ..., vn } independent
Then (cf. Proposition 1),
(∀x(0) ∈ Cn )(∃c, c′ ∈ C) lim x(t) = c1 + c′ f ebjt .
1
position(Im)
Proof: First, by a similar argument to that in the
proof of Theorem 1, we can show that the desired eigenvalues/eigenvectors (two eigenvalues at 0 with eigenvectors
1 and f ; all other eigenvalues with negative real parts) may
always be assigned for the multi-agent system (7). As a result
(cf. Proposition 1),
0.5
0
-0.5
-1
-1
0
1
position(Re)
Fig. 2. Scalable regular pentagon formation (x: initial positions, ◦: steadystate positions)
system (7) can be made to achieve circular motion while
keeping a rigid formation with some specified size d > 0.
Circular motion may be applied to the task of target
encircling, which is illustrated by an example in the next
section.
VII. SIMULATIONS
We illustrate the eigenstructure assignment based approach
by several simulation examples. For all the examples, we
consider the multi-agent system (7) with 5 heterogeneous
agents, where
A = diag(1.6, 4.7, 3.0, −0.7, −4.2)
B = diag(0.2, 1.5, −0.5, −3.3, −3.7).
Thus the first 3 agents are unstable while the latter 2 are
stable; all agents are stabilizable.
First, to achieve a scalable (regular) pentagon formation,
assign the following eigenstructure:
eigenvalues: {λ1 , . . . , λ5 } = {0, −1, −2, −3, −4}
2πj×1
e 5
−1 0
0
e 2πj×2
5
1
0
0
2πj×3
eigenvectors: [v1 · · · v5 ] = e 5
−2 −1 1
2πj×4
e 5
−2 0 −1
2πj×5
e 5
−2 0 −2
0
0
0 .
0
1
By (5) we compute the control gain matrix
t→∞
That is, Problem 3 is solved.
The key point to achieving circular motion is to assign
one, and only one, pure imaginary eigenvalue bj, associated
with the formation vector f . The circular motion is counterclockwise if b > 0, and clockwise if b < 0. One may easily
speed up or slow down the circular motion by specifying the
value |b|.
Also note that, by a similar synthesis procedure to that for
rigid formation in the previous subsection, the multi-agent
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
F =
−10.5 − 1.8164j
0.3333 + 0.2422j
−0.4721 − 2.3511j
−0.0442 − 0.4403j
−0.3979 + 0.4391j
2.5 − 1.8164j
−3.4667 + 0.2422j
−0.4721 − 2.3511j
1.1679 − 0.4403j
0.1427 + 0.4391j
0
0
0
0
0
0
.
10
−2
0
0
0.697
0
0 −0.5405 −0.0541
Simulating the closed-loop system with initial condition
x(0) = [1+j 1−0.5j 1 j −1+j]⊤ , the result is displayed
in Fig. 2. Observe that a regular pentagon is formed, and the
topology determined by F contains a spanning tree.
Next, to achieve a rigid pentagon formation, we follow the
method presented in Section VI.A: first assign the following
agent1
agent2
agent3
agent4
agent5
2.5
2
position(Im)
position(Im)
10
5
0
-5
1.5
1
target
0.5
0
agent1
agent2
agent3
agent4
agent5
-0.5
-10
0
5
10
15
20
25
-1
-2
-1
0
1
2
position(Re)
position(Re)
Fig. 3. Rigid regular pentagon formation, with size d = 5, 10, 15 (x:
initial positions, ◦: steady-state positions)
Fig. 4. Target encircling by circular motion (x: initial positions, ◦: steadystate positions)
Moreover, choose the following eigenstructure
eigenstructure:
eigenvalues: {λ1 , . . . , λ6 } = {0, j, −1, −2, −3, −4}
eigenvalues: {λ1 , . . . , λ5 } = {0, 0, −1, −2, −3}
2πj×1
0 0 0
1 e 5
1 e 2πj×2
5
0 0 0
2πj×3
eigenvectors: [v1 · · · v5 ] = 1 e 5
1 0 0
2πj×4
1 e 5
0 1 0
2πj×5
0 0 1
1 e 5
and by (5) compute the control gain matrix
F =
−8
0
0.618 + 1.9021j
−0.303 + 0.9326j
−1.0614 + 0.7711j
0
−3.1333
−2.618 − 1.9021j
−0.303 − 0.9326j
0.2506 − 0.7711j
0
0
0
0
0
0
.
8
0
0
0 0.3939
0
0
0
−0.3243
Thus the topology determined by F is 2-rooted with nodes
1 and 2 the only two roots. Then for different sizes (d =
5, 10, 15), we obtain by (26) the control u. Simulating the
closed-loop system with the same initial condition x(0) as
above, the result is displayed in Fig. 3, where pentagons with
specified sizes are formed.
Further, we consider the task of target encircling and solve
it by circular motion introduced in Section VI.B. Suppose
that there is a static target, say x6 with ẋ6 = u6 (u6 is
constantly zero), and the goal is make the same 5 agents as
above circle around x6 . For this we treat the target x6 as
part of the multi-agent system; hence the augmented A′ and
B ′ are
A′ = diag(1.6, 4.7, 3.0, −0.7, −4.2, 0)
B ′ = diag(0.2, 1.5, −0.5, −3.3, −3.7, 1).
1
1
1
eigenvectors: [v1 · · · v6 ] =
1
1
1
2πj×1
e 5
2πj×2
e 5
2πj×3
e 5
2πj×4
e 5
2πj×5
e 5
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
.
0
1
0
Thus the desired formation is (again) a regular pentagon and
the target x6 is at the center of this pentagon. Moreover, by
the eigenvectors the resulting topology will contain a spanning tree with x6 (the target) being the root. Corresponding
to the formation vector is the eigenvalue j; hence the agents
will perform circular motion at rate 1. Since the center will
not move (x6 at the center is the root), this makes the first
5 agents encircle the target x6 .
By (5) we compute the control gain matrix
F =
−8 + 5j
−0.428 + 0.84j
4.4116 − 0.7331j
0.5574 + 0.7795j
−0.5911 + 0.9447j
0
0
0
0
0
−5j
−3.8 0
0
0
1.0947 − 0.84j
0
10
0
0
−8.4116 + 0.7331j
0
0 0.697
0
−1.4664 − 0.7795j
0
0
0
−0.0541 −0.49 − 0.9447j
0
0
0
0
0
to assign the above eigenstructure. Indeed, the corresponding
topology has a spanning tree whose root is x6 and ẋ6 = 0.
Simulating the closed-loop system with the initial condition
x(0) = [1−0.5j −2+2j −2+j −1+j −j 1+j]⊤ , the
result is displayed in Fig. 4. Observe that the target stays put
at its initial position 1 + j, while other agents circle around
it.
Finally, we present an empirical study on the computation
time of synthesizing feedback matrix F . In particular, we
compare the centralized synthesis by (5) and the hierarchical
synthesis in Section V; the result is listed in Table I for different numbers of agents.4 Here for the hierarchical synthesis,
we partition the agents in such a way that the number of
4 Computation is done by Matlab R2014b on a laptop with Intel(R)
Core(TM) i7-4510U [email protected] 2.60GHz and 8.00GB memory.
TABLE I
C OMPARISON OF COMPUTATION TIME ( UNIT: SECONDS )
agent #
100
500
900
1000
centralized method by (5)
0.398
57.308
552.8419
1068.729
hierarchical method in Sec. V
0.027
0.179
0.394
0.525
groups and the number of agents in each group are ‘balanced’
(to make ñ small): e.g. 100 agents are partitioned into 10
groups of 10 agents each; 500 agents are partitioned into 16
groups of 23 agents each plus 6 groups of 22 each. Observe
that the hierarchical synthesis is significantly more efficient
than the centralized one, and the efficiency increases as the
number of agents increases. In particular, for 1000 agents
only 0.525 seconds needed, the hierarchical approach might
well be sufficient for many practical purposes.
VIII. C ONCLUDING R EMARKS
We have proposed a top-down, eigenstructure assignment
based approach to synthesize state feedback control for
solving multi-agent formation problems. The relation between the eigenstructures used in control synthesis and the
resulting topologies among agents has been characterized,
and special topologies have been designed by choosing
appropriate eigenstructures. More general cases where the
initial inter-agent coupling is arbitrary and/or there exist
non-stabilizable agents have been studied, and a hierarchical
synthesis procedure has been presented that improves computational efficiency. Further, the approach has been extended
to achieve rigid formation and circular motion.
In our view, the proposed top-down approach to multiagent formation control is complimentary to the existing
(mainstream) bottom-up approach (rather than opposed to).
Indeed, the bottom-up approach, if successful, can produce
scalable control strategies effective for possibly time-varying
topologies, nonlinear agent dynamics, and robustness issues
like communication failures, which are the cases very difficult to be dealt with by the top-down approach. On the other
hand, bottom-up design is generally challenging, requiring
significant insight into the problem at hand and possibly
many trial-and-errors in the design process; by contrast, topdown design is straightforward and can be automated by
algorithms. Hence we suggest the following. When a control
researcher or engineer faces a distributed control design
problem for achieving some new cooperative tasks, one can
start with a linear time-invariant version of the problem and
try the top-down approach to derive a solution. With the ideas
and insights gained from such a solution, one may then try
the bottom-up design possibly for time-varying and nonlinear
cases.
In future work, we aim to apply the top-down, eigenstructure assignment based approach to solve more complex
cooperative control problems of multi-agent systems. In
particular, our immediate goals are to achieve formations in
three dimensions with obstacle avoidance abilities, as well
as to deal with agents with higher-order (heterogeneous,
possibly non-stabilizable) dynamics.
IX. A PPENDIX
We provide the proof of Proposition 4. For this we
first briefly review the eigenstructure assignment method in
[17] for dealing with repeated eigenvalues and generalized
eigenvectors.
Lemma 2: ([17]) Consider the system (1) and suppose that (A, B) is controllable and KerB = 0. Let
{d1 , . . . , dk } (k ≤ n) be a set of positive integers satisfying
P
k
i=1 di = n, {λ1 , . . . , λk } be a set of complex numbers,
and {v11 , ..., v1d1 , . . . , vk1 , ..., vkdk } a set of linearly independent vectors in Cn . Then there is a feedback matrix F
such that for every i ∈ [1, k],
(λi I − A − BF )vi1 = 0
(λi I − A − BF )vij = vi(j−1) ,
j = 2, . . . , di
if and only if
(∃wij ∈ Cm )
vi1 (λi )
λi I − A − B
=0
wi1 (λi )
vij (λi )
λi I − A − B
= vi(j−1) ,
wij (λi )
j = 2, . . . , di .
(27)
Lemma 2 provides a necessary and sufficient condition
for assigning repeated eigenvalues λi (i ∈ [1, k]) with
eigenvector vi1 and generalized eigenvectors vi2 , ..., vidi
(corresponding to a Jordan block of the closed-loop matrix
A + BF ). When the condition holds and thus F exists, F
may be constructed by the following procedure [17].
(i) Compute the following (maximal-rank) matrices
N1i
S
Ni =
, Si = 1i
N2i
S2i
satisfying
[λi I − A B]
N1i
= 0,
N2i
[λi I − A B]
S1i
= −I.
S2i
(Si needs to be computed only when di > 1.)
(ii) From the following vector chain
vi1 = N1i pi1
vi2 = N1i pi2 − S1i vi1
..
.
vidi = N1i pidi − S1i vi(di −1)
find the vectors pi1 , ..., pidi . Then generate a new vector
chain as follows:
wi1 = −N2i pi1
wi2 = −N2i pi2 + S2i vi1
..
.
widi = −N2i pidi + S2i vi(di −1) .
(iii) Compute F satisfying F vij = wij for all i ∈
[1, k], j ∈ [1, di ]. (If no solution exists, alter one or more
of the vectors pij in Step (ii).)
Now we are ready to prove Proposition 4.
Proof of Proposition 4: Consider the multi-agent system (7) and assign the following eigenstructure:
eigenvalues: λ1 = 0, λ2 = · · · = λn = −1
eigenvectors: [v11 v21 · · · v2(n−1) ] =
f1
f2
..
.
fn−1
fn
0
0
..
.
0
−fn
···
···
.
..
···
0
−f2
..
.
−fn−1
−fn
First, for λ1 = 0 (with d1 = 1), in Step (i) we only need
to compute N1 and obtain N11 = B, N21 = A. Then in
Step (ii), from
v11 = N11 p11
we have p11 = B −1 v11 . Hence
w11 = −N21 p11 = −AB −1 v11 = [−
an f n ⊤
a1 f 1
···−
] .
b1
bn
It is verified that w11 , together with v11 , λ1 , satisfies the first
equation of (27).
Second, for λ2 = −1 (with d2 = n − 1), in Step (i)
we derive N12 = B, N22 = (A + I), S12 = B, S22 =
(A + I) − B −1 . Then in Step (ii), from the chain
v21 = N12 p21
v22 = N12 p22 − S12 v21
..
.
v2(n−1) = N12 p2(n−1) − S12 v2(n−2)
we find
fn ⊤
]
bn
fn
fn−1
−
− fn ]⊤
= [0 · · · 0 −
bn−1
bn
..
.
f2
f3
fn
= [0 −
−
− f3 · · · −
− fn ]⊤ .
b2
b3
bn
p21 = [0 · · · 0 −
p22
p2(n−1)
Hence we obtain
w21 = N12 p21
(an + 1)fn ⊤
]
bn
w22 = N12 p22 − S12 v21
(an−1 + 1)fn−1 an fn ⊤
]
= [0 · · · 0
bn−1
bn
..
.
w2(n−1) = N12 p2(n−1) − S12 v2(n−2)
an fn ⊤
(a2 + 1)f2 a3 f3
···
] .
= [0
b2
b3
bn
= [0 · · · 0
The chain {w21 , w22 , ..., w2(n−1) }, together with
{v21 , v22 , ..., v2(n−1) } and λ2 (= −1), is verified to
satisfy (27). Hence it follows from Lemma 2 that the above
eigenstructure can be assigned to the closed-loop matrix
A + BF . Then by Proposition 1, Problem 1 is solved.
Finally we compute the feedback matrix F . Let W =
[w11 w21 · · · w2(n−1) ]. Since
1
0
···
0
0
f1
1
0
0
· · · fn−1
− f1n
.
.
−1
.
.
.
.
V
=
.
1
.
.
0
f2
1
1
−
0
f1
f2
we obtain by F = W V −1 that
−a1
0
b1
−(1+a2 )
f2
b2 f1
b2
F =
..
.
0
···
..
0
.
fn
bn fn−1
Therefore the closed-loop matrix is
0
0
···
f2 −1
f1
A + BF =
..
..
.
.
fn
0
fn−1
−(1+an )
bn
0
.
−1
and the corresponding graph G is a line topology.
R EFERENCES
[1] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of
mobile autonomous agents sing nearest neighbor rules,” IEEE Trans.
Autom. Control, vol. 48, no. 6, pp. 988–1001, 2003.
[2] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and
cooperation in networked multi-agent systems,” Proc. IEEE, vol. 95,
no. 1, pp. 215–233, 2007.
[3] W. Ren and R. W. Beard, Distributed Consensus in Multi-vehicle
Cooperative Control: Theory and Applications.
Springer Verlag,
2008.
[4] F. Bullo, J. Cortés, and S. Martı́nez, Distributed Control of Robotic
Networks. Princeton University Press, 2009.
[5] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multiagent
Networks. Prenceton University Press, 2010.
[6] B. Anderson, C. Yu, B. Fidan, and J. M. Hendrickx, “Rigid graph
control architectures for autonomous formations,” IEEE Control Syst.
Mag., vol. 28, no. 6, pp. 48–63, 2008.
[7] J. A. Fax and R. M. Murray, “Information flow and cooperative control
of vehicle formations,” IEEE Trans. Autom. Control, vol. 49, no. 9,
pp. 1465–1476, 2004.
[8] J. Cortés, “Global and robust formation-shape stabilizatioin of relative
sensing networks,” Automatica, vol. 45, no. 10, pp. 2754–2762, 2009.
[9] L. Krick, M. Broucke, and B. Francis, “Stabilisation of infinitesimally
rigid formations of multi-robot networks,” Int. J. Control, vol. 82,
no. 3, pp. 423–439, 2009.
[10] M. Cao, A. S. Morse, C. Yu, and B. Anderson, “Maintaining a directed,
triangular formation of mobile autonomous agents,” Commun. Inform.
and Syst., vol. 11, no. 1, pp. 1–16, 2011.
[11] M. Basiri, A. Bishop, and P. Jensfelt, “Distributed control of triangular
formations with angle-only constraints,” Syst. & Control Lett., vol. 59,
no. 2, pp. 147–154, 2010.
[12] S. Coogan and M. Arcak, “Scaling the size of a formation using
relative position feedback,” Automatica, vol. 48, no. 10, pp. 2677–
2685, 2012.
[13] Z. Lin, L. Wang, Z. Han, and M. Fu, “Distributed formation control
of multi-agent systems using complex laplacian,” IEEE Trans. Autom.
Control, vol. 59, no. 7, pp. 1765–1777, 2014.
[14] H. Bai, M. Arcak, and J. T. Wen, “Adaptive design for reference
velocity recovery in motion coordination,” Syst. & Control Lett.,
vol. 57, no. 8, pp. 602–610, 2008.
[15] W. Ding, G. Yan, and Z. Lin, “Collective motions and formations under
pursuit strategies on directed acyclic graphs,” Automatica, vol. 46,
no. 1, pp. 174–181, 2010.
[16] B. C. Moore, “On the flexibility offered by state feedback in multivariable systems beyond closed loop eigenvalue assignment,” IEEE
Trans. Autom. Control, vol. 21, no. 5, pp. 689–692, 1976.
[17] G. Klein and B. C. Moore, “Eigenvalue-generalized eigenvector assignment with state feedback,” IEEE Trans. Autom. Control, vol. 22,
no. 1, pp. 140–141, 1977.
[18] A. N. Andry, E. Y. Shapiro, and J. C. Chung, “Eigenstructure assignment for linear systems,” IEEE Trans. Aerospace and Electronic
Systems, vol. AES-19, no. 5, pp. 711–729, 1983.
[19] G. P. Liu and R. J. Patton, Eigenstructure Assignment for Control
System Design. Wiley, 1998.
[20] K. Cai and T. Motoyama, “Eigenstructure assignment for synthesis of
multi-agent consensus algorithms,” in Proc. 58th Japan Joint Autom.
Control Conf., Kobe, Japan, 2015, pp. 2B1–2.
[21] T. Motoyama and K. Cai, “Top-down synthesis of multi-agent formation control: an eigenstructure assignment based approach,” in Proc.
of American Control Conf., Seattle, WA, 2017, pp. 259–264.
[22] A. Wu and T. Iwasaki, “Eigenstructure assignment with applicaiton
to consensus of linear heterogeneous agents,” in Proc. 54th Conf.
Decision and Control, Osaka, Japan, 2015, pp. 2067–2072.
[23] W. M. Wonham, “On pole assignment in multi-input controllable linear
systems,” IEEE Trans. Autom. Control, vol. 12, no. 6, pp. 660–665,
1967.
[24] G. H. Golub and C. F. van Loan, Matrix Computations. The Johns
Hopkins University Press, 3rd ed, 1996.
[25] T. Kim and T. Sugie, “Cooperative control for target-capturing task
based on a cyclic pursuit strategy,” Automatica, vol. 43, no. 8, pp.
1426–1431, 2007.
[26] Y. Lan, G. Yan, and Z. Lin, “Distributed control of cooperative target
enclosing based on reachability and invariance analysis,” Syst. &
Control Lett., vol. 59, no. 7, pp. 381–389, 2010.
| 3cs.SY
|
arXiv:1606.07264v2 [math.GT] 13 Sep 2016
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET
INTERSECTION THEOREM
PRANAB SARDAR
Abstract. We define the notion of limit set intersection property for a collection of subgroups of a hyperbolic group; namely, for a hyperbolic group G and
a collection of subgroups S we say that S satisfies the limit set intersection
property if for all H, K ∈ S we have Λ(H) ∩ Λ(K) = Λ(H ∩ K). Given a hyperbolic group admitting a decomposition into a finite graph of hyperbolic groups
structure with QI embedded condition, we show that the set of conjugates of
all the vertex and edge groups satisfy the limit set intersection property.
1. Introduction
Limit set intersection theorems first appear in the work of Susskind and Swarup
([SS92]) in the context of geometrically finite Kleinian groups. Later on Anderson
([Anda], [Andb]) undertook a detailed study of this for general Kleinian groups.
In the context of (Gromov) hyperbolic groups this is true for quasiconvex subgroups(see [GMRS97], Lemma 2.6). (Recently W. Yang has looked at the case
of relatively quasiconvex subgroups of relatively hyperbolic groups. See [Yan12].)
However, this theorem is false for general subgroups of hyperbolic groups and no
characterizations other than quasi-convexity are known for a pair of subgroups H, K
of a hyperbolic group G which guarantee that Λ(H) ∩ Λ(K) = Λ(H ∩ K). This
motivates us to look for subgroups other than quasiconvex subgroups which satisfy
the limit set intersection property. Our starting point is the following celebrated
theorem of Bestvina and Feighn.
Theorem 1.1. ([BF92]) Suppose (G, Y ) is a graph of hyperbolic groups with QI
embedded condition and the hallways flare condition. Then the fundamental group,
say G, of this graph of groups is hyperbolic.
Graphs of groups are briefly recalled in section 3. There are many examples of
hyperbolic groups admitting such a decomposition into graphs of groups where the
vertex or edge groups are not quasiconvex. Nevertheless, with the terminologies of
the above theorem, we have:
Theorem. The set of all conjugates of the vertex and edge groups of G satisfy
the limit set intersection property.
We note that a special case of our theorem was already known by results of Ilya
Kapovich ([Kap01]). There the author showed that given a k-acylindrical graph
of hyperbolic groups (G, Y ) with quasi-isometrically embedded condition and with
fundamental group G (which turnns out to be hyperbolic by Theorem 1.1,) the
Date: March 30, 2018.
2010 Mathematics Subject Classification. Primary 20F67.
Key words and phrases. Hyperbolic groups, limit sets, Bass-Serre theory.
1
2
PRANAB SARDAR
vertex groups are quasiconvex subgroups of G. Hence, the conjugates of all the
vertex groups satisfy the limit set intersection property in this case.
Acknowledgments: This work started during the author’s post doctoral term
in the University of California, Davis. The author would like to thank Michael
Kapovich for many helpful discussions in this regard. The author would also like to
thank Mahan Mj for helpful discussions and Ilya Kapovich for useful email correspondence. This research was supported by the University of California, Davis. It is
partially supported by the DST INSPIRE faculty award DST /IN SP IRE/04/2014/002236.
2. Boundary of Gromov hyperbolic spaces and Limit sets of subspaces
We assume that the reader is familiar with the basics of (Gromov) hyperbolic
metric spaces and the coarse language. We shall however recall some basic definitions and results that will be explicitly used in the sections to follow. For details
one is referred to [Gro85] or [BH99].
Notation and convention. In this section we shall assume that all the hyperbolic metric spaces are proper geodesic metric spaces. We use QI to mean both
quasi-isometry and quasi-isometric depending on the context. Hausdorff distance
of two subsets A, B of a metric space Z is denoted by Hd(A, B). For any subset A
of a metric space Y and any D ≥ 0, ND (A) will denote the D-neighborhood of A
in Y . We assume that all our groups are finitely generated. For any graph Y we
shall denote by V (Y ) and E(Y ) the vertex and the edge sets of Y respectively.
Definition 2.1. 1. Suppose G is a group generated by a finite set S ⊂ G and let γ ⊂
Γ(G, S) be a path joining two vertices u, v ∈ Γ(G, S). Let u0 = u, u1 , u2 , ..., un = v
be the consecutive vertices on γ. Let ui+1 = ui xi , xi ∈ S ∪ S −1 for 0 ≤ i ≤ n − 1.
Then we shall say that the word w = x0 x1 ...xn−1 labels the path γ.
2. Also, given w ∈ F(S)- the free group on S, its image in G under the natural
map F(S) → G will be called the element of G represented by w.
Definition 2.2. (See [BH99]) 1. Let X be a hyperbolic metric space and x ∈ X
be a base point. Then the (Gromov) boundary ∂X of X is the equivalence classes
of geodesic rays α such that α(0) = x where two geodesic rays α, β are said to be
equivalent if Hd(α, β) < ∞.
The equivalence class of a geodesic ray α is denoted by α(∞).
2. If {xn } is an unbounded sequence of points in X, we say that {xn } converges
to some boundary point ξ ∈ ∂X if the following holds: Let αn be any geodesic
joining x to xn . Then any subsequence of {αn } contains a subsequence uniformly
converging on compact sets to a geodesic ray α such that α(∞) = ξ. In this case,
we say that ξ is the limit of {xn } and write limn→∞ xn = ξ.
3. The limit set of a subset Y of X is the set {ξ ∈ ∂X : ∃ {yn } ⊂ Y with limn→∞ yn =
ξ}. We denote this set by Λ(Y ).
The following lemma is a basic exercise in hyperbolic geometry and so we mention
it without proof. It basically uses the thin triangle property of hyperbolic metric
spaces. (See [BH99], Chapter III.H, Exercise 3.11.)
Lemma 2.3. Suppose {xn }, {yn } are two sequences in a hyperbolic metric space X
both converging to some points of ∂X. If {d(xn , yn )} is bounded then limn→∞ xn =
limn→∞ yn .
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET INTERSECTION THEOREM
3
Lemma 2.4. 1. There is a natural topology on the boundary ∂X of any proper
hyperbolic metric space X with respect to which ∂X becomes a compact space.
2. If f : X → Y is a quasi-isometric embdedding of proper hyperbolic metric
spaces then f induces a topological embedding ∂f : ∂Y → ∂X.
If f is a quasi-isometry then ∂f is a homeomorphism.
We refer the reader to Proposition 3.7 and Theorem 3.9 in Chapter III.H of
[BH99] for a proof of Lemma 2.4.
Definition 2.5. 1. A map f : Y → X between two metric spaces is said to be a
proper embedding if for all M > 0 there is N > 0 such that dX (f (x), f (y)) ≤ M
implies dY (x, y) ≤ N for all x, y ∈ Y .
A family of proper embeddings between metric spaces fi : Xi → Yi , i ∈ I- where
I is an indexing set, is said to be uniformly proper if for all M > 0 there is an
N > 0 such that for all i ∈ I and x, y ∈ Xi , dYi (fi (x), fi (y)) ≤ N implies that
dXi (x, y) ≤ M .
2. If f : Y → X is a proper embeddings of proper hyperbolic metric spaces then
we say that Cannon-Thurston(CT) map exists for f if f gives rise to a continuous
map ∂f : ∂Y → ∂X.
This means that given a sequence of points {yn } in Y converging to ξ ∈ ∂Y , the
sequence {f (yn )} converges to a point of ∂X and the resulting map is continuous.
Note that our terminology is slightly different from Mitra([Mit98]). The following
lemma is immediate.
Lemma 2.6. Suppose X, Y are hyperbolic metric spaces and f : Y → X is a proper
embedding. If the CT map exists for f then we have Λ(f (Y )) = ∂f (∂Y ).
We mention the following lemma with brief remarks about proofs since it states
some standard facts from hyperbolic geometry.
Lemma 2.7. Suppose Z is proper δ-hyperbolic metric space and {xn } and {yn }
are two sequences in Z such that limn→∞ xn = limn→∞ yn = ξ ∈ ∂X. For each n
let αn , βn be two geodesics in X joining x1 , xn and y1 , yn respectively.
(1) Then there is subsequences {nk } of natural numbers such that the sequences
of geodesics {αnk } and {βnk } converge uniformly on compact sets to two geodesics
α, β joining x1 , y1 respectively to ξ.
(2) Moreover, there is a constant D depending only on δ and d(x1 , y1 ) and
sequences of points pnk ∈ αnk and qnk ∈ βnk such that d(pnk , qnk ) ≤ D, and
limk→∞ pnk = limk→∞ qnk = ξ.
(3) The conclusion (2) remains valid if we replace αn , βn by K-quasi-geodesics
for some K ≥ 1; in other words if xn , yn are joined to x1 , y1 by K-quasi-geodesics
αn , βn respectively then there is a constant D depending on δ, d(x1 , y1 ) and K, a
subsequence {nk } of natural numbers and sequences of points pnk ∈ αnk , qnk ∈ βnk
such that d(pnk , qnk ) ≤ D, and limk→∞ pnk = limk→∞ qnk = ξ.
For a proof of (1), (2) see Lemma 3.3, and Lemma 3.13 and for (3) Theorem
1.7(stability of quasi-geodesics) in Chapter III.H of [BH99]. More precisely, for
′
′
proving (3) we may choose geodesic segments αn ’s, βn ’s connecting the endpoints of
the quasi-geodesics αn ’s and βn ’s respectively and then apply (1) for these geodesics
′
′
′
′
to extract subsequences {αnk } and {βnk } of {αn } and {βn } respectively both converging uniformly on compact sets. Then we can find two sequences of points
4
PRANAB SARDAR
′
′
′
′
pnk ∈ αnk , qnk ∈ βnk satisfying (2). Finally, by stability of quasi-geodesics for all k
′
′
there are pnk ∈ αnk , qnk ∈ βnk such that d(pnk , pnk ) and d(qnk , qnk ) are uniformly
small. That will prove (3).
Definition 2.8. Suppose G is a Gromov hyperbolic group. Let S be any collection
of subgroups of G. We say that S has the limit set intersection property if for all
H, K ∈ S we have Λ(H) ∩ Λ(K) = Λ(H ∩ K).
We state two elementary results on limit sets for future use.
Lemma 2.9. Suppose G is a hyperbolic group and H is any subset of G. Then for
all x ∈ G we have
(1) Λ(xH) = Λ(xHx−1 ).
(2) Λ(xH) = xΛ(H).
P roof : (1) follows from Lemma 2.3. For (2) one notes that G acts naturally on
a Cayley graph X of G by isometries and thus by homeomorphisms on ∂X = ∂G
by Lemma 2.4. ✷
3. Graphs of groups
We presume that the reader is familiar with the Bass-Serre theory. However, we
briefly recall all the concepts that we shall need. For details one is referred to section
5.3 of J.P. Serre’s book Trees ([Ser00]). Although we always work with nonoriented
metric graphs like Cayley graphs, we need oriented graphs possibly with multiple
edges between adjacent vertices and loops to describe graphs of groups. Hence the
following definition is quoted from [Ser00].
Definition 3.1. A graph Y is a pair (V, E) together with two maps
E →V ×V,
e 7→ (o(e), t(e)) and
E → E,
e 7→ ē
such that o(ē) = t(e), t(ē) = o(e) and ē¯ = e for all e ∈ E.
For an edge e we refer to o(e) as the origin and t(e) as the terminus of e; the edge
ē is the same edge e with opposite orientation. We write V (Y ) for V and E(Y ) for
E. We refer to V (Y ) as the set of vertices of Y and E(Y ) as the set of edges of Y .
We shall denote by |e| the edge e without any orientation.
Definition 3.2. A graph of groups (G, Y ) consists of the following data:
(1) a (finite) graph Y as defined above,
(2) for all v ∈ V (Y ) (and edge e ∈ E(Y )) there is group Gv (respectively Ge )
together with two injective homomorphisms φe,0(e) : Ge → Go(e) and φe,t(e) : Ge →
Gt(e) for all e ∈ E(Y ) such that the following conditions hold:
(i) Ge = Gē ,
(ii) φe,o(e) = φē,t(ē) and φe,t(e) = φē,o(ē) .
We shall refer to the maps φe,v ’s as the canonical maps of the graph of groups.
We shall refer to the groups Gv and Ge , v ∈ V (Y ) and e ∈ E(Y ) as vertex groups
and edge groups respectively. For topological motivations of graph of groups and
the following definition of the fundamental group of a graph of groups one is referred
to [SW79] or [Hat01].
Definition 3.3. Fundamental group of a graph of groups Suppose (G, Y ) is
a graph of groups where Y is a (finite) connected oriented graph. Let T ⊂ Y be a
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET INTERSECTION THEOREM
5
maximal tree. Then the fundamental group G = π1 (G, Y, T ) of (G, Y ) is defined in
terms of generators and relators as follows:
The generators of G are the elements of the disjoint union of the generating sets
of the vertex groups Gv , v ∈ V (Y ) and the set E(Y ) of oriented edges of Y .
The relators are of four of types: (1) Those coming from the vertex groups; (2)
ē = e−1 for all edge e, (3) e = 1 for |e| ∈ T and (4) eφe,t(e) (a)e−1 = φe,o(e) (a) for
all oriented edge e and a ∈ Ge .
Bass Serre tree of a graph of groups
Suppose (G, Y ) is a graph of groups and let T be a maximal tree in Y as in the
above definition. Let G = π1 (G, Y, T ) be the fundamental group
F of the graph of
groups. The Bass-Serre tree, say T , is the tree with vertex set v∈V (Y ) G/Gv and
F
edge set e∈E(Y ) G/Gee where Gee = φe,t(e) (Ge ) < Gt(e) . The edge relations are
given by
t(gGee ) = geGt(e) , o(gGee ) = gGo(e) ....(∗)
Note that when |e| ∈ T then we have e = 1 in G.
Tree of metric spaces from a graph of groups
Given a graph of groups (G, Y ) and a maximal tree T ⊂ Y one can form in a
natural way a graph, say X, on which the fundamental group G = π1 (G, Y, T ) acts
by isometries properly and cocompactly and which admits a simplicial Lipschitz
G-equivariant map X → T . The construction of X can be described as follows.
We assume that Y is a finite connected graph and all the vertex groups and
the edge groups are finitely generated. We fix a finite generating set Sv for each
one of the vertex groups Gv ; similarly for each edge groups Ge we fix a finite
generating set Se and assume that φe,t(e) (Se ) ⊂ St(e) for all e ∈ E(Y ). Let S =
∪v∈V (Y ) Sv ∪ (E(Y ) \ E(T )) be a generating set of G where in E(Y ) \ E(T ) we shall
include only nonoriented edges of Y not in T . We define X from the disjoint union
of the following graphs by introducing some extra edges as follows:
(1) Vertex spaces: For all ṽ = gGv ∈ V (T ), where v ∈ Y and g ∈ G we
let Xṽ denote the subgraph of Γ(G, S) with vertex set the coset gGv ; two vertices
gx, gy ∈ Xṽ are connected by an edge iff x−1 y ∈ Sv . We shall refer to these
subspaces of X as vertex spaces.
(2) Edge spaces: Similarly for any edge ẽ = gGee of T , let Xẽ denote the
subgraph of Γ(G, S) with vertex set geGee where where two vertices gex, gey are
connected by an edge iff x−1 y ∈ φe,t(e) (Se ). We shall refer to these subspaces of X
as edge spaces.
(3) The extra edges connect the edge spaces with the vertex spaces as follows:
For all edge ẽ = gGee of T connecting the vertices ũ = gGo(e) and ṽ = geGt(e) of
T , and x ∈ Gee join gex ∈ Xẽ = geGee to gex ∈ Xṽ = geGt(e) and gexe−1 ∈ Xũ =
gGo(e) by edges of length 1/2 each. We define fẽ,ṽ : Xẽ → Xṽ and fẽ,ũ : Xẽ → Xũ
by setting fẽ,ṽ (gex) = gex and fẽ,ũ (gex) = gexe−1 .
We have a natural simplicial map π : X → T (more precisely to the first barycentric subdivision of T ). This map is the coarse analog of the tree of metric spaces
introduced by [BF92] (see also [Mit98]). By abuse of terminology we shall refer
to this also as a tree of metric spaces or graphs. We recall some notations and
definitions from [Mit98] and collect some basic properties.
(1) We note that Xu = π −1 (u) and Xẽ = π −1 (e) for all u ∈ V (T ) and e ∈ E(T ).
For all u ∈ V (T ) the intrinsic path metric of Xu will be denoted by du . Similarly,
6
PRANAB SARDAR
we use de for the intrinsic path metric on Xe . It follows that with these intrinsic
metrics the metric spaces Xe , Xu are isometric to the Cayley graphs Γ(Ge , Se ) and
Γ(Gu , Su ) respectively. Therefore, if all the vertex and edge groups are Gromov
hyperbolic then the vertex and edge spaces of X are uniformly hyperbolic metric
spaces.
(2) Quasi-isometric lifts of geodesics: Suppose u, v ∈ T and let [u, v] denote
the geodesic in T joining them. A K-QI section of π over [u, v] or a K-QI lift of
[u, v] (in X) is a set theoretic section s : [u, v] → X of π which is also a K-QI
embedding. In general, we are only interested in defining these sections over the
vertices in [u, v].
(3) Hallways flare condition: We will say that π : X → T satisfies the
hallways flare condition if for all K ≥ 1 there are numbers λK > 1, MK ≥ 1, nK ≥ 1
such that given a geodesic α : [−nK , nK ] → T and two K-QI lifts α1 , α2 of α, if
dα(0) (α1 (0), α2 (0)) ≥ MK then
max{dα(nK ) (α1 (nK ), α2 (nK )), dα(−nK ) (α1 (−nK ), α2 (−nK ))} ≥ λK dα(0) (α1 (0), α2 (0)).
(4) Graphs of groups with QI embedded conditions. Suppose (G, Y ) is
a graph of groups such that each vertex and edge group is finitely generated. We
say that it satisfies the QI embedded condition if all the inclusion maps of the edge
groups into the vertex groups are quasi-isometric embeddings with respect any choice
of finite generating sets for the vertex and edge groups.
It is clear that if (G, Y ) is a graph of groups with QI embedded condition then
all the maps fe,u : Xe → Xu are uniform QI embeddings.
Lemma 3.4. There is a naturally defined proper and cocompact action of G on X
such that the map π : X → T is G-equivariant.
Proof: We note that X is obtained from the disjoint union of the cosets of the
vertex and edge groups of (G, Y ). The group G has a natural action on this disjoint
union. It is also easy to check that under this action adjacent vertices of X go to
adjacent vertices. Thus we have a simplicial G-action on X. Clearly the natural
map π : X → T is G-equivariant. To show that the action is proper it is enough to
show that the vertex stabilizers are uniformly finite. However, if a point x ∈ gGv
is fixed by an element h ∈ G then h fixes gGv ∈ V (T ). However, stabilizers of gGv
is simply gGv g −1 and the action of gGv g −1 on gGv ⊂ X is fixed point free. Hence,
the G-action on X is fixed point free.
That the G-action is cocompact on X follows from the fact that the G-actions
on V (T ) and E(T ) are cofinite. ✷
Fix a vertex v0 ∈ Y and the vertex Gv0 ∈ V (T ). Look at the corresponding
vertex space Gv0 ⊂ X and let x0 denote 1 ∈ Gv0 . Let Θ : G → X denote the orbit
map g 7→ gx0 . By Milnor-Schwarz lemma this orbit map is a quasi-isometry since
the G-action is proper and cocompact by the above lemma.
Lemma 3.5. There is a constant D0 such that for all vertex space gGv ⊂ X we
have Hd(Θ(gGv ), gGv ) ≤ D0 .
It follows that for any g.x ∈ gGv ⊂ X we have g.x ∈ Θ−1 (B(gx, D0 )).
P roof : For proving the lemma let γv be a geodesic in X joining x0 to the identity
element of Gv . Then for all x ∈ Gv , gxγv is a path joining gxx0 and gx ∈ gGv .
Hence one can choose D0 to be the maximum of the lengths of γv ’s, v ∈ V (Y ). ✷
The following corollary is an immediate consequence of the above two lemmas.
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET INTERSECTION THEOREM
7
Corollary 3.6. The vertex spaces and edge spaces of X are uniformly properly
embedded in X.
Notation: We shall use iw : Xw → X denote the canonical inclusion of the
vertex and edge spaces of X into X. Let ṽ = gGv ∈ V (T ). It follows from
the above corollary that Θ induces a coarsely well-defined quasi-isometry from
gGv ⊂ G to Xṽ . Namely, we can send any x ∈ gGv to a point y of Xṽ such that
dX (Θ(x), y) ≤ D0 . where D0 is as in the above corollary. We shall denote this by
Θg,v : gGv → Xṽ .
4. The main theorem
For the rest of the paper we shall assume that G is a hyperbolic group which
admits a graph of groups decomposition (G, Y ) with the QI embedded condition
where all the vertex and edge groups are hyperbolic. Let T be the Bass-Serre tree
of this graph of groups.
We aim to show that in G the family of subgroups {Gv : v ∈ V (T )} satisfies the
limit set intersection property:
Theorem 4.1. Suppose a hyperbolic group G admits a decomposition into a graph
of hyperbolic groups (G, Y ) with quasi-isometrically embedded condition and suppose
T is the correspoding Bass-Serre tree. Then for all w1 , w2 ∈ V (T ) we have Λ(Gw1 )∩
Λ(Gw2 ) = Λ(Gw1 ∩ Gw2 ).
The idea of the proof is to pass to the tree of space π : X → T using the orbit
map Θ : G → X defined in the previous section and then use the techniques of
[Mit98]. The following theorem is an important ingredient of the proof.
Theorem 4.2. ([Mit98]) The inclusion maps iw : Xw → X admit CT maps ∂iw :
∂Xw → ∂X for all w ∈ V (T ).
Recall that if u, v ∈ V (T ) are connected by an edge e there are natural maps
fe,u : Xe → Xu and fe,v : Xe → Xv . We know that these maps are uniform QI
embeddings. We assume that they are all K-QI embeddings for some K > 1. They
induce embeddings ∂fe,u : ∂Xe → ∂Xu and ∂fe,v : ∂Xe → ∂Xv by Lemma 2.4.
Therefore, we get partially defined maps from ∂Xu to ∂Xv with domain Im(∂fe,u ).
Let us denote this by ψu,v : ∂Xu → ∂Xv . By definition for all x ∈ Xe we have
ψu,v (fe,u (x)) = fe,v (x).
Definition 4.3. (1) If ξ ∈ ∂Xu is in the domain of ψu,v and ψu,v (ξ) = η then we
say that η is a flow of ξ and that ξ can be flowed to ∂Xv .
(2) Suppose w0 6= wn ∈ V (T ) and w0 , w1 , · · · , wn are consecutive vertices of the
geodesic [w0 , wn ] ⊂ T . We say that a point ξ ∈ ∂Xw0 can be flowed to ∂Xwn if there
are ξi ∈ ∂Xwi , 0 ≤ i ≤ n where ξ0 = ξ such that ξi+1 = ψwi ,wi+1 (ξi ), 0 ≤ i ≤ n − 1.
In this case, ξn is called the flow of ξ0 in Xwn .
Since the maps ψu,v are injective on the their domains for all u 6= v ∈ V (T ) and
ξ ∈ ∂Xu the flow of ξ in ∂Xv is unique if it exists.
Lemma 4.4. Suppose w1 , w2 ∈ V (T ) and ξi ∈ ∂Xwi , i = 1, 2 such that ξ2 is a flow
of ξ1 . Let αi be a geodesic in the vertex space Xwi such that αi (∞) = ξi , i = 1, 2.
Then Hd(α1 , α2 ) < ∞.
8
PRANAB SARDAR
P roof : It is enough to check it when w1 , w2 are adjacent vertices. Suppose e is
the edge connecting w1 , w2 . The lemma follows from the stability of quasi-geodesics
in the hyperbolic space Xwi ’s and the fact that every point of Xe is at distance 1/2
from Xwi , i = 1, 2. ✷
Corollary 4.5. Under the CT maps ∂iwj : ∂Xwj → ∂X, j = 1, 2 the points ξ1 , ξ2
go the same point of ∂X, i.e. ∂iw1 (ξ1 ) = ∂iw2 (ξ2 ).
Lemma 4.6. Let w1 , w2 ∈ T and suppose they are joined by an edge e. Suppose
ξ1 ∈ ∂Xw1 can not be flowed to ∂Xw2 . Let α ⊂ Xw1 be a geodesic ray such that
α(∞) = ξ1 . Then for all D > 0 the set ND (α) ∩ fe,w1 (Xe ) is bounded.
P roof : If ND (α) ∩ fe,w1 (Xe ) is not bounded for some D > 0 then ξ1 is in
the limit set of fe,w1 (Xe ) and so ξ1 can be flowed to ∂Xw2 by Lemma 2.6. This
contradiction proves the lemma. ✷
Now we briefly recall the ladder construction of Mitra which was crucial for the
proof of the main theorem of [Mit98]. We shall need it for the proof of Theorem
4.1.
Mitra’s Ladder B(λ).
Fix D0 , D1 > 0. Let v ∈ V (T ) and λ be a finite geodesic segment of Xv . We
shall define the set B(λ) to be a union of vertex space geodesics λw ⊂ Xw where w
is in a subtree T1 of T containing v. The construction is inductive. Inductively one
constructs the n-sphere ST1 (v, n) of T1 centred at v and the corresponding λw ’s,
w ∈ S(v, n).
ST1 (v, 1): There are only finitely many edges e incident on v such that ND0 (λ) ∩
fe,v (Xe ) 6= ∅. Then S(v, 1) is the set of terminal points of all the edges e that
start at v such that the diameter of ND0 (λ) ∩ fe,v (Xe ) is at least D1 . In this
case, for each edge e connecting v to say v1 ∈ S(v, 1), we choose two points,
say x, y ∈ ND0 (λ) ∩ fe,v (Xe ) such that dv (x, y) is maximum. Then we choose
x1 , y1 ∈ Xv1 such that d(x, x1 ) = 1 and d(y, y1 ) = 1 and define λv1 to be a geodesic
in Xv1 joining x1 , y1 .
ST1 (v, n + 1) from ST1 (v, n): Suppose w1 ∈ S(v, n). Then a vertex w2 adjacent
to w1 with dT (v, w2 ) = n + 1 belongs to S(v, n + 1) if the diameter of ND0 (λw1 ) ∩
fe,w1 (Xe ) is at least D1 , where e is the edge connecting w1 , w2 , has diameter at
least D1 in Xw1 . To define λw2 one chooses two points x, y ∈ ND0 (λw1 ) ∩ fe,w1 (Xe )
such that dw1 (x, y) is maximum, then let x1 , y1 ∈ Xw2 be such that d(x, x1 ) = 1
and d(y, y1 ) = 1 and define λw2 to be a geodesic in Xw2 joining x1 , y1 .
Theorem 4.7. (Mitra [Mit98]) There are constants D0 > 0, D1 > 0 and C > 0
depending on the defining parameters of the tree of metric spaces π : X → T such
that the following holds:
For any v ∈ V (T ) and a geodesic segment λ ⊂ Xv the corresponding ladder B(λ)
is a C-quasi-convex subset of X.
To prove this theorem, Mitra defines a coarse Lipschitz retraction map P : X →
B(λ) which we now recall. For the proof of how this works one is referred to [Mit98].
However, we shall subsequently assume that appropriate choices of D0 , D1 are made
in our context so that all the ladders are uniformly quasiconvex subsets of X.
Coarsely Lipschitz retraction on the ladders
Suppose λ ⊂ Xv is a geodesic. Let T1 = π(B(λ)). For each w ∈ T1 , λw =
Xw ∩ B(λ) is a geodesic in Xw . We know that there is a coarsely well defined
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET INTERSECTION THEOREM
9
nearest point projection Pw : Xw → λw . (See Proposition 3.11 in Chapter III.Γ
of [BH99].) Now for each x ∈ Xw , w ∈ T1 define P (x) = Pw (x). If x ∈ Xw and
w 6∈ T1 then connect w to T1 by a geodesic in T . Since T is a tree there is a unique
such geodesic. Let w1 ∈ T1 be the end point of this geodesic and let e be the edge
on this geodesic incident on w1 going out of T1 . Mitra proved that in this case the
projection of fe,w1 (Xe ) on λw1 is uniformly small. It follows by careful choice of
D0 , D1 . (See Lemma 3.1 in [Mit98].) Choose a point xw1 on this projection. Define
P (x) = xw1 .
Theorem 4.8. ([Mit98]) The map P : X → B(λ) is a coarsely Lipschitz retraction.
In other words, it is a retraction and there are constants A, B such that d(P (x), P (y)) ≤
Ad(x, y) + B for all x, y ∈ X.
Using the above theorems of Mitra now we shall prove the converse of Corollary
4.5. This is the last ingredient for the proof of Theorem 4.1.
Proposition 4.9. Suppose v 6= w ∈ T and there are points ξv ∈ ∂Xv and ξw ∈
∂Xw which map to the same point ξ ∈ ∂X under the CT maps ∂Xv → ∂X and
∂Xw → ∂X respectively. Then ξv can be flowed to ∂Xw .
P roof : Using Lemma 4.4 we can assume that the point v is such that ξv can
not further be flowed along vw and similarly ξw can not be flowed in the direction
of wv where v 6= w. Let α : [0, ∞) → Xv and β : [0, ∞) → Xw be geodesic rays in
Xv , Xw respectively such that α(∞) = ξv and β(∞) = ξw . Let ev , ew be the first
edges from the points v, w along the direction of vw and wv respectively. Then
ND0 (α) ∩ fev ,v (Xev ) ⊂ Xv and ND0 (β) ∩ few ,w (Xew ) ⊂ Xw are both bunded sets
by Lemma 4.6 where D0 is as in Theorem 4.7.
For all n ∈ N let αn := α|[0,n] and βn := β|[0,n] respectively. The ladders
B(αn ), B(βn ) are uniformly quasi-convex subsets of X by Theorem 4.8. Hence there
are uniform ambient quasi-geodesics of X in these ladders joining α(0), α(n) and
β(0), β(n) respectively. Choose one such for each one of them and let us call them γn
′
and γn respectively. Now, since α and β limit on the same point ξ ∈ ∂X, by Lemma
2.7(3) there is a uniform constant D such that for a subsequence {nk } of natural
′
numbers there are points xnk ∈ γnk and ynk ∈ γnk such that d(xnk , ynk ) ≤ D and
limk→∞ xnk = limk→∞ ynk = ξ ∈ ∂X.
Let v1 , w1 be the vertices on the geodesic [v, w] ⊂ T adjacent to v, w respectively.
Let Avk := B(αnk ) ∩ Xv1 and Awk := B(βnk ) ∩ Xw1 respectively.
If we remove the edge space Xev from X then the remaining space has two
components- one containing Xv and the other containing Xw . Call them Y1 , Y2
respectively. We note that since the diameter of Avk is uniformly bounded, if at all
nonempty, the portion of γnk contained in Y2 , if at all it travels into Y2 , is uniformly
bounded. This implies that the portion of γnk joining α(0) and Avk is uniformly
small if Avk 6= ∅.
Hence, there are infinitely many k ∈ N such that xnk ∈ Y1 and ynk ∈ Y2 . Since
we are dealing with a tree of spaces and d(xnk , ynk ) ≤ D for all k ∈ N, this implies
there are points zk ∈ fev (Xev ) such that d(xnk , zk ) ≤ D for all k ∈ N. Thus ξv can
be flowed to ∂Xv1 by Lemma 2.6. This contradiction proves the proposition. ✷
Proof of Theorem 4.1:
Suppose wi = gi Gvi , i = 1, 2, for some g1 , g2 ∈ G and v1 , v2 ∈ V (Y ). This
implies Gwi = gi Gvi gi−1 , i = 1, 2. Also, Λ(gi Gvi gi−1 ) = Λ(gi Gvi ) by Lemma 2.9(1).
10
PRANAB SARDAR
Hence we need to show that Λ(g1 Gv1 )∩Λ(g2 Gv2 ) = Λ(g1 Gv1 g1−1 ∩g2 Gv2 g2−1 ). Using
Lemma 2.9(2), therefore, it is enough to show that
Λ(Gv1 ) ∩ Λ(gGv2 ) = Λ(Gv1 ∩ gGv2 g −1 ) for all v1 , v2 ∈ V (Y ), g ∈ G.
Clearly, Λ(Gv1 ∩ gGv2 g −1 ) ⊂ Λ(Gv1 ) ∩ Λ(gGv2 ). Thus we need to show that
Λ(Gv1 ) ∩ Λ(gGv2 ) ⊂ Λ(Gv1 ∩ gGv2 g −1 ).
Given an element ξ ∈ Λ(Gv1 ) ∩ Λ(gGv2 ) there are ξ1 ∈ ∂Gv1 and ξ2 ∈ ∂(gGv2 )
both of which map to ξ under the CT maps ∂Gv1 → ∂G and ∂(gGv2 ) → ∂G by
Lemma 2.6. Now, we have a quasi-isometry Θ : Γ(G, S) → X. By Lemma 3.5 each
coset of any vertex group in G is mapped uniformly Hausdorff close to the same
coset in X. Hence Θ induces uniform quasi-isometries Θg,v from gGv ⊂ Γ(G, S)
to gGv ⊂ X for all g ∈ G, v ∈ Y . For avoiding confusion let us denote the subset
Gv1 ⊂ X by Xw1 and gGv2 ⊂ X by Xw2 . It follows that ∂Θ1,v1 (ξ1 ) ∈ ∂Xw1 and
∂Θg,v2 (ξ2 ) ∈ ∂Xw2 are mapped to the same element of ∂X under the CT maps
∂Xwi → ∂X, i = 1, 2. Hence by Proposition 4.9 ∂Θ1,v1 (ξ1 ) can be flowed to,
′
′
say ξ2 ∈ ∂Xw2 . By Lemma 4.4 the image of ξ2 and Θg,v2 (ξ2 ) under the CT map
∂Xw2 → ∂X are the same. Hence, we can replace ξ2 by (∂Θg,v2 )−1 (ξ2 ) and assume
that Θ1,v1 (ξ1 ) flows to Θg,v2 (ξ2 ).
Then by Lemma 4.4 for any geodesic rays αi ⊂ Xwi with αi (∞) = ∂Θxi ,vi (ξi )
for i = 1, 2 where x1 = 1 and x2 = g we have Hd(α1 , α2 ) < ∞. Pulling back these
geodesics by Θ1,v1 and Θg,v2 we get uniform quasi-geodesic rays, say β1 ⊂ Gv1 and
β2 ⊂ gGv2 such that βi (∞) = ξi , i = 1, 2 and Hd(β1 , β2 ) < ∞.
Now let pi = β1 (i) and qi ∈ β2 , i ∈ N be such that d(pi , qi ) ≤ D where
Hd(β1 , β2 ) = D. Join pi to qi by a geodesic in Γ(G, S). Suppose wi is the word labeling this geodesic. Since there are only finitely many possibilities for such words,
′
−1
there is a constant subsequence {wnk } of {wn }. Let hk = p−1
n1 pnk and hk = qn1 qnk .
′
Let x be the group element represented by wnk . Then we have pn1 .hk .x = pn1 .x.hk
′
′
or hk = xhk x−1 . Since hk connects two elements of gGv2 , it is in Gv2 . Hence hk ∈
−1
−1
= Gv1 ∩ gGv2 g −1 .
Gv1 ∩ xGv2 x−1 . Thus pn1 hk p−1
n1 ∈ pn1 Gv1 pn1 ∩ pn1 xGv2 (pn1 x)
−1
−1
Finally, since d(pn1 hk pn1 , pn1 hk ) = d(pn1 hk pn1 , pnk ) = d(1, pn1 ) for all k ∈ N,
limn→∞ pn1 hk p−1
n1 = limn→∞ pnk = ξ1 . This completes the proof. ✷
The following corollary has been pointed out by Mahan Mj. We use the same
notations as in the main theorem.
Corollary 4.10. If Hi ⊂ Gwi , i = 1, 2 are two quasiconvex subgroups then Λ(H1 )∩
Λ(H2 ) = Λ(H1 ∩ H2 ).
P roof : Assume that wi = gi Gvi , i = 1, 2. Then Gwi = gi Gvi gi−1 . Let Ki =
< Gvi . We may construct a new finite graph starting from Y by adding
two vertices u1 , u2 where ui is connected to vi by an edge ei , i = 1, 2. Let us call
this graph Y1 . Define a new graph of groups (G1 , Y1 ) by keeping the definition same
on Y and setting Gui = Gei = Ki , i = 1, 2 and defining φei ,ui = 1Ki and φei ,vi
to be the inclusion map Ki ⊂ Gvi . This produces a new graph of groups with QI
embedded condition and with fundamental group isomorphic to G. Suppose the
Bass-Serre tree of the new graph of groups is T1 .
′
Now we can apply Theorem 4.1 to Gw′ , i = 1, 2 where wi = gi Ki ∈ T1 , i = 1, 2
i
to finish the proof. ✷
gi−1 Hi gi
Example 4.11. We now give an example where intersection of limit sets is not
equal to the limit set of the intersection. Suppose G is a hyperbolic group with an
GRAPHS OF HYPERBOLIC GROUPS AND A LIMIT SET INTERSECTION THEOREM 11
infinite normal subgroup H such that G/H is not torsion. Let g ∈ G be such that its
image in G/H is an element of infinite order. Let K =< g >. Then H ∩ K = (1)
whence Λ(H ∩ K) = ∅. However, H being an infinite normal subgroup of G we have
Λ(H) = ∂G. Thus Λ(H) ∩ Λ(K) = Λ(K) 6= ∅.
We end with a question.
Question 4.12. If a hyperbolic group G admits a decomposition into a graph of hyperbolic groups with qi embedded condition and Gv is a vertex group how to describe
Λ(Gv ) ⊂ ∂G?
It has been pointed out to me by Prof. Ilya Kapovich that the first interesting
case where this question should be considered is a hyperbolic strictly ascending
HNN extension of a finitely generated nonabelian free group F
G =< F, t|t−1 wt = φ(w), ∀w ∈ F >
where φ : F → F is an injective but not surjective endomorphism of F . One would
also like to describe ∂G in this case.
References
[Anda]
J. W. Anderson, The limit set intersection theorems for finitely generated kleinian
groups, Math. Res. Lett, vol 3(1996), no 5, 675–692.
, Limit set intersection theorems for kleinian groups and a conjecture of
[Andb]
susskind, Comput. Methods Funct. Theory, October 2014, vol 14, issue 2, 453–464.
[BF92]
M. Bestvina and M. Feighn, A Combination theorem for Negatively Curved Groups,
J. Differential Geom., vol 35 (1992), 85–101.
[BH99]
M. Bridson and A Haefliger, Metric spaces of nonpositive curvature, Grundlehren der
mathematischen Wissenchaften, Vol 319, Springer-Verlag (1999).
[GMRS97] R. Gitik, M. Mitra, E. Rips, and M. Sageev, Widths of Subgroups, Trans. AMS (Jan.
’97), 321–329.
[Gro85]
M. Gromov, Hyperbolic Groups, in Essays in Group Theory, ed. Gersten, MSRI
Publ.,vol.8, Springer Verlag (1985), 75–263.
[Hat01]
A. Hatcher, Algebraic topology, Cambridge University Press (2001).
[Kap01]
I. Kapovich, The combination theorem and quasiconvexity, Int. J. Algebra Comput.
11, Issue 02 (2001), 185–216.
[Mit98]
M. Mitra, Cannon-Thurston Maps for Trees of Hyperbolic Metric Spaces, J. Differential Geom. 48 (1998), 135–164.
[Mit04]
Mahan Mitra, Height in Splittings of Hyperbolic Groups, Proc. Indian Acad. of Sciences, v. 114, no.1 (Feb. 2004), 39–54.
[Ser00]
J. P. Serre, Trees, Springer-Verlag, ISBN 3-540-44237-5 (2000).
[SS92]
P. Susskind and G. A. Swarup, Limit sets of geometrically finite hyperbolic groups,
Amer. J. Math., Vol. 114, No. 2 (Apr., 1992) (1992), 233–250.
[SW79]
P. Scott and C. T. C. Wall, Topological Methods in Group Theory, Homological group
theory (C. T. C. Wall, ed.), London Math. Soc. Lecture Notes Series, vol. 36, Cambridge Univ. Press (1979).
[Yan12]
Wen-Yuan Yang, Limit sets of relatively hyperbolic groups, Geom. Dedicata, vol 156,
issue 1 (2012), 1–12.
Indian Institute of Science Education and Research Mohali
| 4math.GR
|
1
A Statistical Characterization of Localization
Performance in Wireless Networks
Christopher E. O’Lone, Graduate Student Member, IEEE,
arXiv:1710.07716v1 [cs.NI] 20 Oct 2017
Harpreet S. Dhillon, Member, IEEE, and R. M. Buehrer, Fellow, IEEE
Abstract
Localization performance in wireless networks has been traditionally benchmarked using the CramérRao lower bound (CRLB), given a fixed geometry of anchor nodes and a target. However, by endowing
the target and anchor locations with distributions, this paper recasts this traditional, scalar benchmark
as a random variable. The goal of this work is to derive an analytical expression for the distribution of
this now random CRLB, in the context of Time-of-Arrival-based positioning.
To derive this distribution, this work first analyzes how the CRLB is affected by the order statistics
of the angles between consecutive participating anchors (i.e., internodal angles). This analysis reveals
an intimate connection between the second largest internodal angle and the CRLB, which leads to
an accurate approximation of the CRLB. Using this approximation, a closed-form expression for the
distribution of the CRLB, conditioned on the number of participating anchors, is obtained.
Next, this conditioning is eliminated to derive an analytical expression for the marginal CRLB
distribution. Since this marginal distribution accounts for all target and anchor positions, across all
numbers of participating anchors, it therefore statistically characterizes localization error throughout an
entire wireless network. This paper concludes with a comprehensive analysis of this new network-wideCRLB paradigm.
Index Terms
Cramér-Rao lower bound, localization, order statistics, Poisson point process, stochastic geometry,
Time of Arrival (TOA), mutual information, wireless networks.
The authors are with Wireless@VT, Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg,
VA, USA. Email: {olone, hdhillon, buehrer}@vt.edu.
This paper was presented in part at the IEEE ICC 2017 Workshop on Advances in Network Localization and Navigation
(ANLN), Paris, France [1].
2
I. I NTRODUCTION
The Global Positioning System (GPS) has for decades been the standard mechanism for position location anywhere in the world. However, the deployment locations of recent and emerging
wireless networks have begun to put a strain on the effectiveness of GPS as a localization solution.
For example, as populations increase, precipitating the expansion of urban environments, cell
phone use in urban canyons, as well as indoors, is continually increasing. The rise of these GPSconstrained environments highlight the need to fall back on the existing network infrastructure
for localization purposes.
Additionally, with the emergence of Wireless Sensor Networks (WSNs) and their increased
emphasis on energy efficiency and cost-effectiveness, [2], [3], the possibility of equipping each
potential target node with a GPS chip quickly becomes impractical. Furthermore, the deployment of these networks in GPS-constrained environments further necessitates a reliance on the
terrestrial network for a localization solution. Thus, localization within a network, performed by
the network itself in the absence of GPS, has begun to garner attention, [4], [5].
Benchmarking localization performance in wireless networks has traditionally been done using
the Cramér-Rao lower bound, which provides a lower bound on the position error of any unbiased
estimator [6]. Common practice has been to analyze the CRLB in fixed scenarios of anchor nodes
and a target, e.g., [7], [8], [9], [10]. This strategy produces a scalar/fixed value for the CRLB and
is specific to the scenario being analyzed. While this idea does provide insight into fundamental
limits of localization performance, it is rather limited in that it does not take into account all
possible setups of anchor nodes and target positions within a network.
In order to account for all possible setups, it is useful to appeal to the field of stochastic
geometry. Whereas in the past stochastic geometry has been applied towards the study of
“connectivity, capacity, outage probability, and other fundamental limits of wireless networks,”
[11], [12], we now however apply it towards the study localization performance in wireless
networks. Modeling anchor node and target placements with point processes opens the possibility
of characterizing the CRLB over all setups of anchor nodes and target positions. Thus, the CRLB
is no longer a fixed value, but rather a random variable (RV) conditioned on the number of
participating anchors, where the randomness is induced by the inherent randomness of the anchor
positions. Upon marginalizing out the number of participating anchors, the resulting marginal
distribution of the CRLB will characterize localization performance throughout an entire wireless
3
network.
A. Related Work
The quest for a network-wide distribution of localization performance comprises two main
steps. The first step involves finding the distribution of the CRLB conditioned on the number
of participating anchor nodes, and the second step involves finding the probability that a given
number of anchors can participate in a localization procedure.
With regards to the first step, there have been several attempts in the literature to obtain this
conditional distribution. An excellent first attempt can be found in the series of papers, [13],
[14], [15], in which approximations of this conditional distribution were presented for ReceivedSignal-Strength (RSS), Time-of-Arrival (TOA), and Angle-of-Arrival (AOA) based localization,
respectively. These approximations were obtained through asymptotic arguments by driving the
number of participating anchor nodes to infinity. While these approximate distributions are very
accurate for larger numbers of participating anchors, they are less than ideal for lower numbers.
However, it is desirable that this conditional distribution be accurate for lower numbers of
participating anchors, since this is the dominate case in terrestrial networks, e.g., cellular.
This conditional distribution of the CRLB was also explored in [16]. Here, the authors were
able to derive the true expression for this conditional distribution through a clever re-writing of
the CRLB using complex exponentials. This distribution was then used to derive and analyze the
so-called “localization outage probability” in scenarios with a fixed number of randomly placed
anchor nodes. While this expression represents the true conditional distribution, its complexity
puts it at a disadvantage over simpler approximations (discussed further in Section III-D).
The second step, which involves finding the participation probability of a given number of
anchor nodes, was explored in [17]. In this work, the authors modeled a cellular network with a
homogeneous Poisson point process (PPP), which consequently allowed them to derive bounds
on L-localizability, i.e., the probability that a mobile device can hear at least L base stations for
participation in a localization procedure.1 Further, by employing a “dominant interferer analysis,”
they were able to derive an accurate expression for the probability of L-localizability, which
can easily be extended to give the probability of hearing a given number of anchor nodes for
participation in a localization procedure.
1 The
term hearability (or hearable) is used to describe anchor nodes that are able to participate in a localization procedure,
i.e., their received SINR at the target is above some threshold.
4
B. Contributions
This paper proposes a novel statistical characterization of a wireless network’s ability to
perform TOA-based localization. Using stochastic geometry to model target and anchor node
placements throughout a network, and using the CRLB as the localization performance benchmark, this paper presents an analytical derivation of the CRLB’s distribution.2 This distribution
offers many insights into localization performance within wireless networks that previously were
only attainable through lengthy, parameter-specific, network simulations. Thus, this distribution
1) offers a means for comparing networks in terms of their localization performance, by
enabling the calculation of network-wide localization statistics, e.g., avg. localization error,
2) unlocks insight into how changing network parameters, such as SINR thresholds, processing
gain, frequency reuse, etc., affect localization performance throughout a network, and
3) provides network designers with an analytical tool for determining whether a network meets,
for example, the FCC E911 standards [18].
In pursuit of this distribution, this paper makes four key contributions. First, this work presents
an analysis of how the CRLB is affected by the order statistics of the internodal angles. This
analysis reveals an intimate connection between the second largest internodal angle and the
CRLB, which leads to an accurate approximation of the CRLB. Second, this approximation is
then used to obtain the distribution of the CRLB conditioned on the number of hearable anchors.
Although this is a distribution of the CRLB approximation, its simplicity and accuracy offer clear
advantages over the true distribution presented in [16] and the approximate distribution presented
in [14]. These advantages are discussed further in Section III-D.
Third, this work then takes a major step by combining this conditional distribution of the CRLB
with the distribution of the number of participating anchors. This eliminates the conditioning
on a given number of anchors, allowing for the analytical expression of the marginal CRLB
distribution to be obtained. Since this marginal distribution now simultaneously accounts for
all possible target and anchor node positions, across all numbers of participating anchor nodes,
it therefore statistically characterizes localization error throughout an entire wireless network,
and thus signals a departure from the existing literature. Additionally, since the two component
distributions are parameterized by various network parameters, our resulting marginal distribution
2 We
will be using the square root of the CRLB as our performance benchmark, however, we just state the CRLB here as to
not unnecessarily clutter the discussion. This will be described further in Section III-A.
5
of the CRLB will also be parameterized by these network parameters. Consequently, our final
contribution involves a comprehensive analysis of this new network-wide CRLB paradigm, where
we examine how varying network parameters affects the distribution of the CRLB — thereby
revealing how these network parameters affect localization performance throughout the network.
II. P ROBLEM S ETUP
This section details the assumptions we propose in determining the network layout as well
as the localization procedure. Additionally, we describe important notation and definitions used
throughout the paper and conclude with how the assumptions impact the network setup.
A. Network Setup and Localization Assumptions
Assumption 1. We assume a ubiquitous, two-dimensional wireless network with anchor nodes
distributed according to a homogeneous PPP over R2 . Further, we assume potential targets to be
distributed likewise, where the anchor and target point processes are assumed to be independent.
Remark. This assumption for modeling wireless networks is common in the literature, e.g., [19],
[20], [21], [22], [23].
Assumption 2. We assume a 2-Way TOA-based positioning technique is used within the twodimensional network.
Remark. Although Time-Difference-of-Arrival (TDOA) is also commonly implemented, 2-Way
TOA represents a viable approach, and additionally offers a lower bound on TDOA performance.
Furthermore, the 2-Way assumption eliminates the need for clock synchronization between the
target and anchor nodes.
Assumption 3. Range measurements are independent and exhibit zero-mean, normally distributed range error.
Remark. A reader familiar with wireless positioning will recognize this as a classic Line-of-Sight
(LOS) assumption. However, those familiar with localization in terrestrial networks will realize
that Non-Line-of-Sight (NLOS) measurements are more common. Thus, while we move forward
under this LOS assumption in order to make progress under this new paradigm, we will see in
Section IV that we can adapt our model to accommodate NLOS measurements by selecting
ranging errors consistent with NLOS propagation.
6
TABLE I
S UMMARY OF N OTATION
Symbol
fX (·)
h(X)
u(x)
k·k
1[·]
tr(X)
L
S
Θk
∠k
X(k)
α
β
K
Description
Probability distribution function (pdf) of X
Differential entropy of X
Unit step function, 0 when x < 0, 1 when x ≥ 0
`2 -norm
Indicator function; 1[A] = 1 if A true, 0 otherwise
Trace of the matrix X
Number of participating anchors, L ∈ {0, 1, 2, . . . }
The localization performance benchmark
Anchor node angle k, a random variable
Internodal angle k, a random variable
k th order statistic of a sequence {Xi } of RVs
Path loss exponent, α > 2
Post-proc. SINR threshold
Frequency reuse factor
Symbol
FX (·)
I(X; Y )
N (µ, σ 2 )
P[A]
b·c
XT
N
M
θk
ϕk
λ
γ
q
σr2
Description
Cumulative distribution function (cdf) of X
Mutual Information between X and Y
Normal distribution, mean µ, variance σ 2
Probability of event A
Floor function
Transpose of the matrix X
Max anchors tasked to perform localization
Fixed localization error, in meters, M ∈ R+
A realization of Θk
A realization of ∠k
Density of PPP of anchor locations
Processing gain
Average network load (i.e., network traffic)
Common range error variance
Assumption 4. The range error variance, σr2 , is common among measurements from participating
anchor nodes and is considered a known quantity.
Remark. This assumption is often made in the literature for range-based localization, e.g., [10],
[25], and although not realistic in every scenario, it allows us to gain initial insight into the
problem and will be relaxed in future work.
B. Notation
The notation used throughout this paper can be found in Table I.
C. Localizability
We now define the terms localizable and unlocalizable, which were introduced in [17].
Definition 1. We say that a target is localizable if it detects localization signals from a sufficient
number of anchor nodes such that its position can be determined without ambiguity.
Remark. Under Assumption 2, this implies that L ≥ 3. We also define unlocalizable to be the
negation of Definition 1.
For the purposes of this setup and subsequent derivations, we will initially only consider
scenarios in which the target is localizable, to avoid unnecessary complication. Later in Section
7
III-E and those which follow, we will account for scenarios in which the target is unlocalizable,
and will modify our results accordingly.
D. Impact of Assumptions
With these assumptions in place, we now describe how they impact the network setup.
From Assumption 1, since the anchors and potential targets are distributed by independent,
homogeneous PPPs over R2 (i.e., stationary), then without loss of generality, we may perform
our analysis for a typical target placed at the origin of the xy-plane [21]. This is due to the
fact that the independence and stationarity assumptions imply that no matter where the target is
placed in the network, the distribution of anchors relative to the target appears the same.
Next, we assume that the number of hearable anchors is some fixed value, L, and begin by
numbering these anchors in terms of their increasing distance from the origin (target position).
This is depicted in Fig. 1 for a particular realization of a homogeneous PPP in which there
are four hearable anchor nodes. Fig. 1 also depicts how their corresponding angles, measured
counterclockwise from the +x-axis, are labeled accordingly. Assumption 1 further implies that
these angles of the hearable anchors are i.i.d. random variables that come from a uniform
distribution on [0, 2π).
Definition 2. If the target is placed at the origin of an x y-plane, then the term anchor node
angle, Θ k , is defined to be the angle corresponding to hearable anchor node k, measured
i.i.d.
counterclockwise from the +x-axis. Note that Θ k ∼ unif[0, 2π), ∀k ∈ {1, . . . , L}.
In later sections we will see that the distances between the anchor nodes and the target are
important for determining how many anchors are able to participate in a given localization procedure. However, under Assumption 4, once particular anchor nodes are identified as participating
in a localization procedure, then they are endowed with the common range error variance. As
we will see in the following section, this assumption, along with Assumptions 2 and 3 will
lead to the CRLB (with L fixed) as being dependent on only the angles between participating
anchor nodes (i.e., internodal angles). Thus, since the CRLB expression is only dependent on
the internodal angles, the distances between participating anchor nodes and the target need not
be considered. Hence, we may view the participating anchors as being placed on a circle about
the origin. This is depicted in Fig. 2, under the same PPP realization as in Fig 1.
8
y
1
y
θ1
θ3
2
θ2
θ (3)
4
θ4
x
ϕ(2)
θ (2)
ϕ(1) θ (1)
ϕ(3)
ϕ(4)
x
θ (4)
3
Fig. 1. I NITIAL L ABELING S CHEME . The dots represent
Fig. 2. E QUIVALENT S ETUP. This is the realization from
a particular realization of anchors placed according to a
Fig. 1, where only participating anchor nodes and the in-
homogeneous PPP. The origin represents the location of the
ternodal angles they trace out are considered, whereas their
target. The anchors participating in the localization procedure
distances from the target are not. The realizations of the RVs
are labeled in increasing order w.r.t. their distance from the
are given by Θ(k) = θ (k) and ∠(k) = ϕ(k) . Note, the anchor
origin. Their corresponding anchor node angles are labeled
node angle order stats “renumber” the participating anchors
accordingly. Note the realization of the RV Θk is Θk = θ k .
in terms of increasing angle c.c.w. starting from the +x-axis.
Next, we formally define the term internodal angle. Since the anchor node angles are such
i.i.d.
that Θ k ∼ unif[0, 2π), for k ∈ {1, . . . , L}, we may examine their corresponding order statistics,
Θ(1), Θ(2), . . . , Θ(L) , where 0 ≤ Θ(1) ≤ Θ(2) ≤ · · · ≤ Θ(L) ≤ 2π by definition. Thus, the order
statistics of the participating anchor node angles effectively “renumber” the nodes in terms of
increasing angle, starting counterclockwise from the +x-axis. This is also depicted in Fig. 2.
Definition 3. If participating anchor nodes are considered according to their anchor node angle
order statistics, then an internodal angle, ∠ k , is defined to be the angle between two consecutive
participating anchor nodes. That is, ∠ k = Θ(k+1) − Θ(k) for 1 ≤ k < L and ∠ L = 2π − (Θ(L) − Θ(1) ).
Remark. Since the internodal angles are functions of RVs, they themselves are RVs. Thus, we
may also consider their order statistics, ∠(1), ∠(2), . . . , ∠(L) . These order statistics of the internodal
angles are depicted for the particular PPP realization in Fig. 2.
In summary, Fig 2 depicts an example of a typical setup realization, given L = 4, once all of
the assumptions are taken into consideration.
9
III. D ERIVATION OF THE N ETWORK -W IDE CRLB D ISTRIBUTION
In this section, we first formally define our localization performance benchmark: the square
root of the CRLB. Using this definition and assuming a random placement of anchors, as well
as a random L, we then describe how this work generalizes localization performance results
currently in the literature. In what follows, we present the steps necessary to derive the marginal
distribution of our localization performance benchmark.
A. The Localization Performance Benchmark
Consider the traditional localization scenario, where the number of participating anchor nodes
(L) and their positions, as well as the target position, are all fixed. We represent the set of
coordinates of these anchors by
n
o
T
ΨL = ψi ∈ R2 ψi = xi, yi , i ∈ {1, 2, 3, ..., L} .
T
The coordinates of the target are denoted by ψt = xt, yt .
Next, under Assumptions 2, 3, and 4, the 2-Way range measurements between the target and
the L participating anchors are given by
ri = di + ni ,
where ri is the measured distance divided by 2 (i.e., the measured 1-Way distance),
q
2
2
xi − xt + yi − yt
di = kψi − ψt k =
i.i.d.
is the true 1-Way distance, and ni ∼ N (0, σr2 ) for all i ∈ {1, 2, . . . , L}.
Remark. Note that under Assumption 4, σr2 is common among the range measurements. For
2
2
2-Way TOA we set this to be σr2 = σ2-Way
/4. If 1-Way TOA is considered, then σr2 = σ1-Way
.
Thus, moving forward we may consider the range measurements to always be 1-Way, regardless
of whether 1-Way or 2-Way TOA is used.
Continuing, Assumption 3 enables the likelihood function to be easily written as a product.
Denoting the vector of range measurements as r = [r1, . . . , rL ]T , the likelihood function is
!
L
Ö
2
(r
−
d
)
1
i
i
exp −
.
L ψt r, ΨL, σr2 =
√
2
2σ
2πσ
r
r
i=1
From this likelihood function, we obtain the following Fisher Information Matrix (FIM)
ÍL
ÍL
2
i=1 cos θ i
i=1 cos θ i sin θ i
1
J L (φt ) = 2
,
ÍL
σr Í L cos θ sin θ
2
sin θi
i
i
i=1
i=1
10
where cos θi =
xi −xt
di
and sin θi =
yi −yt
di .
Remark. Note that if the target is placed at the origin, then the angles here, i.e., the θi ’s, are a
particular realization of the anchor node angles from Definition 2.
Taking the inverse of the FIM above, then the CRLB for any unbiased estimator, ψ̂t = [ x̂t, ŷt ]T ,
of the target position, ψt = [xt, yt ]T , is given by
CRLB = tr J−1
(ψ
)
t ,
L
(1)
as defined in Ch. 2.4.1 of [24].
Definition 4. We define our localization performance benchmark, S, to be the square root of
the CRLB in (1):
r
S , tr J−1
(ψ
)
t .
L
(2)
Remark. This benchmark is often referred to as the position error bound (PEB) in the literature,
[25], [26].
To conclude, notice that from (2), a closed-form expression for S can be obtained:
√
L
S = σr v
u
!2 ,
t L
L
L
Õ
Õ
Õ
cos2 θi
sin2 θi −
cos θi sin θi
i=1
i=1
(3)
i=1
which is a function of the anchor node angles.
B. Departure From the Traditional Localization Setup
In the previous section, we assumed a traditional setup, in which the number of participating
anchor nodes (L) and their positions (ΨL ), as well as the target position (ψt ), were all fixed. In
this section however, we now assume that both ΨL and L are random and briefly describe how
the random placement of anchors impacts our localization performance benchmark, S, and how
the randomness of L signals a departure from the existing literature.
We begin by describing how the random placement of anchors affects our localization performance benchmark. This is accomplished by now invoking Assumption 1 and examining
the impact that this has on the expression for S given by (3). Recall from Section II-D that
Assumption 1 implies that ψt = [0, 0]T and that the anchor node angles are i.i.d. on [0, 2π).
Thus, the θi realizations in (3) are now replaced with the random variables, Θi , from Definition
11
2. Since S is now a function of random variables, it itself becomes a random variable, and we
may now seek its distribution.
While work in the past has sought this distribution for S, e.g., [14], [16], there always
remained one implicit assumption: a fixed L. Therefore, the results presented thus far in the
literature have only applied to localization setups with a fixed number of anchor nodes, and hence
were not applicable network-wide. To address this issue, we now consider L to be a random
variable, whose distribution statistically quantifies the number of anchor nodes participating in
a localization procedure. This new interpretation of L will consequently allow us to decouple S
from L, thereby enabling the marginal distribution of S to account for all possible positioning
scenarios within a network. In addition to the contributions outlined in Section I-B, taking
advantage of this new interpretation of L, to subsequently obtain the marginal distribution of S,
is the main contribution setting this work apart from the existing literature.
C. Approximation of the CRLB
Before deriving the conditional distribution of S given L, we use this section to acquire an
approximation of our expression for S in (3). This will consequently allow for an accurate,
tractable, closed-form expression for the conditional distribution of S to be obtained.
1) Approximation Preliminaries and Goals: To facilitate the search for an accurate approximation, we recall that (3) is now in terms of the random variables Θi , and thus can be rewritten
using the internodal angles from Definition 3. This given by
√
L
S = σr v
u
! ,
tÕ
j−1
L−1 Õ
L
Õ
∠k
sin2
i=1 j=i+1
(4)
k=i
Definition 5. We define the terms underneath the square-root in the denominator in (4) by the
random variable D. That is,
D=
L−1 Õ
L
Õ
i=1 j=i+1
sin2
j−1
Õ
!
∠k .
(5)
k=i
Thus, we would like to find an approximation for D which comprises two key traits: (1) it
allows for a straightforward transformation of random variables, i.e., the number of sin2 (·) terms
does not change with L (and would ideally only involve a single term), and (2) it simultaneously
does not sacrifice accuracy, i.e., the approximation should preserve as much “information”
12
as possible about the setup of anchors, implying that the approximation should dominate (or
contribute the most to) the total value of D.
2) Initial Approach and Intuition: In trying to find an approximation that satisfies both
traits, we consider the following possibilities. First, consider approximating D with the sine
squared of an arbitrary internodal angle, sin2 (∠ k ), or with a sum of consecutive internodal angles,
sin2 (∠ k +∠ k+1 +...), where the starting angle, ∠ k , is arbitrary. While these possible approximations
may seem like reasonable candidates for satisfying the first trait, they unfortunately fall short
of satisfying the second trait. To see why, it is illustrative to examine Fig. 2 under different
realizations/placements of anchor nodes. By only looking at the same unordered internodal angle
on every realization, little knowledge is gained about the total setup of anchors. For example,
on one realization, the arbitrary internodal angle being examined might be large and would
therefore give a strong indication of how the rest of the anchors are placed, however, on another
realization, this same internodal angle might be small, thus giving little information about the
placement of the remaining anchors. Hence, in general, arbitrary internodal angles do not provide
accurate approximations due to their inconsistency in describing the anchor node setup, which
consequently leads to their (sine squared terms’) inability to capture the total value of D across
all realizations of anchors for a given L.
3) A Quantitative Approach Using Mutual Information: Taking advantage of the intuition
gained above, it should now be clear that we would like an approximation that utilizes angles
which tend to consistently dominate any given setup. Therefore, it makes sense to examine
larger internodal angles, and thus, the use of internodal angle order statistics follows naturally.
Since we ideally desire a single-term approximation, we examine the possibility of using the
sine-squared of the largest, second largest, and third largest internodal angles, i.e.,3
W(L) = sin2 (∠(L) ),
W(L−1) = sin2 (∠(L−1) ),
and
W(L−2) = sin2 (∠(L−2) ).
(6)
Note that these larger internodal angles should intuitively contain more “information” about the
setup of anchors, and consequently D, since they greatly restrict the placement of the remaining
anchors. Since each of the order statistic approximations above might seem viable under this
3 We
will not examine sums of the larger internodal angle order statistics, e.g., sin2 (∠(L) + ∠(L−1) ) or sin2 (∠(L) ) + sin2 (∠(L−1) ),
for the following two reasons: (1) they would lead to a more complex expression for the conditional distribution of S than
desired, and (2) they offer no gain in accuracy over using the sine squared of just a single internodal angle order statistic, as
evidenced through simulation.
13
1.6
I(D; W(L) |L)
I(D; W(L−1) |L)
I(D; W(L−2) |L)
Mutual Information (Bits)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
3
4
5
6
7
Number of Participating Anchor Nodes (L)
Fig. 3.
J USTIFYING A PPROXIMATIONS T HROUGH M UTUAL I NFORMATION. The mutual informations were calculated
numerically by computing (8) and (9), where the necessary distributions were generated using a Monte Carlo simulation of
10 million anchor node realizations. The bin width of these distributions was chosen to be 0.01, Matlab’s ‘spline’ option was
used to interpolate the integrands in (8) and (9), and the supports of D and W(i) are given by D = [0, L 2 /4] and W(i) = [0, 1],
where i ∈ {L, L − 1, L − 2}. Furthermore, we adopt the convention: 0 log2 0 = 0 “based on continuity arguments” [29].
qualitative notion of information, we thus turn towards a quantitative notion, in order to justify
the use of one of these approximations.
Towards this end, we utilize the concept of mutual information. The reason behind this choice,
over that of correlation for example, is because mutual information captures both linear and
nonlinear dependencies between two random variables, since it is zero if and only if the two
random variables are independent [28]. Hence, we examine the mutual information between D
and the random variables W(L) , W(L−1) , and W(L−2) , so that we may quantify which approximation
carries the most information about D. Thus, we condition on L equaling some integer `(≥ 3)
and calculate:
I(D; W(i) |L = `) = h(D|L = `) − h(D|W(i), L = `),
where i ∈ {L, L − 1, L − 2}, and the differential entropies are given by
∫
1
h(D|L = `) = fD (d|`) log2
dd,
and
fD (d|`)
D
∫ ∫
1
h(D|W(i), L = `) =
fDW(i) (d, w|`) log2
dd dw,
fD (d|w, `)
(7)
(8)
(9)
W(i) D
where W(i) and D are the supports of W(i) and D, respectively [29]. The mutual information
between the approximations in (6) and D are given in Fig. 3, versus L. From Fig. 3, it is evident
14
that the mutual information between D and the approximation W(L−1) is the highest across all
numbers of participating anchors shown.
4) Investigating the High Mutual Information of D and W(L−1) : To explore the reasoning
behind this result, we examine the effect that W(L−1) has on the total value of D. We begin by
rewriting D as follows:
Proposition 1. The random variable D from Definition 5 can be equivalently expressed as
!
j+i−1
L−2 Õ
L−i
L
Õ
Õ
Õ
sin2 ∠(k) +
sin2
∠k .
D=
i=2 j=1
k=1
k= j
Proof. See Appendix A.
By separating the sine-squared terms of the internodal angle order statistics from the total
sum, Proposition 1 makes it clearer as to how our approximations from (6) may affect the total
value of D. To reveal the effects of W(L−1) in particular, we present the following lemma along
with its corollaries:
Lemma 2. The cdf of the second largest order statistic of the internodal angles, ∠(L−1) , conditioned on L, is given by
F∠(L−1)
X
Õ
nϕ
n−1 L
ϕ|L =
(−1)
(n − 1) 1 −
2π
n
n=0
! L−1
,
(10)
where X = min L, b2π/ϕc and the support is 0 ≤ ∠(L−1) ≤ π. (Note L ≥ 2, since if L < 2,
∠(L−1) would not exist.)
Proof. We refer the reader to the conference version of this paper, i.e., Appendix B of [1].
Corollary 2.1. Given a finite L, the expected value of the second largest order statistic of the
internodal angles, ∠(L−1) , conditioned on L, is given by
!
L
Õ
L
2π(n
−
1)
E[∠(L−1) |L] =
(−1)n
n
nL
n=2
Proof. See Appendix B.
Corollary 2.2. Given a finite L, the variance of ∠(L−1) , conditioned on L, is given by
!"
#
L
2
Õ
4π
L n−1
2
c
VAR[∠(L−1) |L] =
(−1)n
+ ,
L n=2
n
n
n(L − 1) L
(11)
(12)
15
where c =
m+1 L m−1
m=2 (−1)
m
m
ÍL
.
2
2
2
Proof. Note: VAR[∠(L−1) |L] = E[∠(L−1)
|L] − E[∠(L−1) |L] , where the derivation of E[∠(L−1)
|L]
is analogous to that of E[∠(L−1) |L] in the proof of Corollary 2.1.
Next, we plot Corollary 2.1, plus/minus one and two standard deviations of ∠(L−1) . This is
given in Fig. 4, versus L. Here, we can see that the second largest internodal angle is centered
and concentrated around π/2, suggesting that W(L−1) = sin2 (∠(L−1) ) will be concentrated about
its maximum of one. This implies that, for the majority of anchor node placements, sin2 (∠(L−1) )
will be a dominant term in the expression for D in Prop. 1. Thus, the sin2 (∠(L−1) ) term will tend
to contribute the most, that a given sin2 (·) term could contribute, to the total value of D. This
is especially true for small values of L, which is our focus.
Also for low L, ∠(L−1) is intuitively the dominant angle in that it places the greatest constraints
on the remaining angles. That is, once ∠(L−1) is determined, it restricts both ∠(L) and ∠(L−2) ,
and thus gives the greatest sense of the total setup of anchors. Note, when considering order
statistics other than ∠(L−1) , the constraints placed on the remaining angles are not as pronounced.
Furthermore, by examining different realizations of anchors (Fig. 2 as an example), along with
Prop. 1, one can see that when W(L−1) is small (∠(L−1) ≈ 0 or π), than so is D, and when W(L−1)
is large (∠(L−1) ≈ π/2), a large value of D follows. Thus, W(L−1) ’s consistency as a dominant
term in Prop. 1, along with its intuitive correlation with D, offer supporting evidence as to why
I(D; W(L−1) |L = `) is higher than both I(D; W(L) |L = `) and I(D; W(L−2) |L = `) for low L.
In summary, mutual information has proved its utility by revealing that W(L−1) is perhaps the
best approximation of D, for the desirable lower values of L. Since W(L−1) possesses the two
desirable traits for an approximation, discussed at the beginning of this section, we henceforth
use W(L−1) in our approximation of D.
5) Completing the Approximation: To complete the approximation of D, and consequently
S, all that now remains is to ensure that D and W(L−1) have the same range of possible values,
i.e., the same support. This will ensure that our ultimate approximation for S will produce the
same range of values as the true S. In order to accomplish this, we approximate D with a scaled
version of W(L−1) , i.e., D ≈ k W(L−1) , and thus search for the value of the constant k so that
kW(L−1) yields the desired support. Since D = [0, dmax ] and W(L−1) = [0, 1] (which follows from
the support of ∠(L−1) , Lemma 2), then in order to have the support of kW(L−1) equal that of D,
we simply need to set k = dmax . The value of dmax is presented in the following lemma:
16
1
π
E[6
E[6
E[6
7 π /8
(L−1)
(L−1)
(L−1)
| L]
| L] ± σ
| L] ± 2σ
0.9
0.8
P[S ≤ abscissa | L]
0.7
5 π /8
π /2
3 π /8
E[6
(L−1) |L]
(Radians)
6 π /8
Increasing L
(L = 3, 5, 7, 9)
0.6
0.5
0.4
0.3
2 π /8
0.2
π /8
Solid: True
Dashed: Theorem 4
0.1
0
3
4
5
6
7
8
9
0
10
20
30
Number of Participating Anchor Nodes (L)
40
50
60
70
Meters
Fig. 4. T HE S ECOND L ARGEST I NTERNODAL A NGLE O R -
Fig. 5. ACCURACY OF T HEOREM 4. The true conditional cdf
S TATISTIC . This figure gives a sense of the concentration
of S given L was generated using a Monte Carlo simulation of
of the distribution of ∠(L−1) around π/2. E[∠(L−1) |L] is given
q
by Cor. 2.1 and σ = VAR[∠(L−1) |L] is from Cor. 2.2.
(4) over 1 million random setup realizations of the internodal
DER
angles. Note, σr = 20 m.
Lemma 3. Let L be finite. If D is given as in (5), then its maximum value is dmax = L 2 /4.
Proof. See Appendix C.
Thus, we now have D ≈ (L 2 /4) · W(L−1) , which completes our approximation of D. Lastly,
substituting this approximation for D into the expression for S in (4) finally yields:
Approximation 1. The localization performance benchmark, S, can be approximated by
r
4
1
S ≈ σr ·
·
,
L sin(∠(L−1) )
(13)
where 0 ≤ ∠(L−1) ≤ π, as stated in Lemma 2.
D. The Conditional CRLB Distribution
Theorem 4. If the localization performance benchmark, S, is given by Approximation 1, then
the cdf of S conditioned on L is
! L−1 X
! L−1
X2
1
Õ
Õ
L
nϕ
L
nϕ
2
1
FS (s | L, σr ) =
(−1)n−1
(n − 1) 1 −
−
(−1)n−1
(n − 1) 1 −
, (14)
n
2π
n
2π
n=0
n=0
p
where ϕ1 = sin−1 (a/s), ϕ2 = π − sin−1 (a/s), a = σr · 4/L, X1 = min L, b2π/ϕ1 c , X2 =
min L, b2π/ϕ2 c , and the support is S ∈ [a, ∞).
17
Proof. See Appendix D.
Remark. Although Theorem 4 is the conditional distribution of our approximation of S, it
provides two clear advantages over the true conditional distribution presented in [16] and over
the approximate conditional distribution presented in [14]. First, Theorem 4 offers a simple,
closed-form, algebraic expression involving only finite sums, as opposed to the rather complex
expression in [16] involving an improper integral of products of scaled Bessel functions. Second,
Theorem 4 is remarkably accurate for lower numbers of participating anchor nodes, see Fig. 5.
This comes in contrast to the approximate distribution presented in [14], which was derived
asymptotically and therefore only accurate for higher numbers of participating anchors. This
selective accuracy of Theorem 4 is desirable since a device is more likely to hear lower numbers
of participating anchors, especially in infrastructure-based, terrestrial wireless networks.4
E. The Distribution of the Number of Participating Anchors
The next step needed to achieve our goal is to find the distribution of the number of participating anchors, fL (`). In this section, we build upon the localizability results from [17] in order
to obtain this distribution. Towards this end, we present the relevant theorems from this work
and modify them for our use here. Finally, we conclude with a discussion on the applicability
of these results.
1) Overview of Localizability Work: Recall from Section I-A that the goal of [17] was to
derive an expression for the probability that a mobile can hear at least ` base stations for
participation in a localization procedure in a cellular network, i.e., P[L ≥ `]. To derive this
expression, the authors assumed that base stations were placed according to a homogeneous
PPP, and then examined the SIRs of the base station signals received at a “typical user” placed
at the origin.5 Specifically, they examined the SIR of the ` th base station (denoted SIR` ), since
this was used directly to determine P[L ≥ `].
Since SIR` depends on the locations of interfering base stations, then the base stations’
placement according to a PPP implies that SIR` becomes a random variable. Consequently, its
distribution also becomes a function of the PPP density, λ. Additionally, the authors incorporate
4 By
infrastructure-based wireless networks, we mean any wireless network setup with mobile devices, fixed access points,
and separate uplink/downlink channels.
5 SIR
= Signal to Interference Ratio. Noise is ignored since [17] assumes an interference-limited network.
18
a network loading parameter, q, into SIR` , where 0 ≤ q ≤ 1. This means that any given base
station can be considered active (i.e., interfering with base station `’s signal) with probability
q. Furthermore, SIR` is also a function of pathloss, α, where α > 2, and the distances of base
stations to the target.
With SIR` statistically characterized, the authors were able to determine P[L ≥ `] by noting
that P[L ≥ `] = P[SIR` ≥ β/γ], where β/γ is the pre-processing SIR threshold for detection
of a signal.6 Here, γ is the processing gain at the mobile (assumed to also average out the
effect of small scale fading), and β is the post-processing threshold. Thus, since P[L ≥ `] =
P[SIR` ≥ β/γ], then P[L ≥ `] must also depend on all of the network parameters described
above. We denote this dependency by P[L ≥ ` | α, λ, q, γ, β].
Before continuing to the localizability results, we mention one last caveat regarding the PPP
density. That is, when shadowing is present, it can easily be incorporated into the PPP network
model through small displacements of the base station locations. This results in a new PPP
density, which accounts for this effect of shadowing. This new shadowing-transformed density
is given by λ̃ = λ E[Sz2/α ], where Sz is assumed to be a log-normal random variable representing
the effect of shadowing on the signal from base station z to the origin [31].7 Thus, by using λ̃,
we incorporate shadowing under the log-normal model presented in Section II of [31].
2) The Localizability Results: In this section, we present the main theorem which will enable
us to obtain fL (`).
Lemma 5. (Theorem 2, [17]) The probability that a mobile device can hear at least ` base
stations for participation in a localization procedure is given by
P[L ≥ ` | α, λ̃, q, γ, β] =
`−1
© Õ
α−2
1− e− 2qβ/γ
n=0
«
α−2
2qβ/γ
n!
n
∫ ∞∫ r`
`−1
−α
` Õ
ª
r
β
4(
λ̃π)
`
® fΩ (0) +
1
≥
f
(ω)
Ω
2−α −r 2−α
®
r
(` − 1)! ω=1
γ
2πq
λ̃
2(ω−1)
2−α
`
1
0 0 r −α +
· r 2 −r 2 + α−2 r`
2−α
1
¬
`
1
× r1 (r`2 − r12 )ω−1r`2(`−ω)−1 ωe−λ̃πr` dr1 dr`,
2
where ` ∈ {1, 2, . . . }, and Ω is a random variable denoting the number of active participating
6 Note
that for q < 1, this equality holds for all but a few rare corner cases. However, the probability of these cases occurring
is vanishingly small and thus has little to no impact on the accuracy of the subsequent localizability results [17].
7 The
Sz are assumed to be i.i.d. ∀z. Sz ’s log-normal behavior implies it is distributed normally when expressed in dB [31].
19
base stations interfering with the ` th . Note, Ω ∼ Binomial(` − 1, q) = fΩ (ω). Additionally, for the
trivial case of ` = 0, we define P[L ≥ 0] = 1. (Note: γ and β are in linear terms, not dB.)
Remark. This theorem was derived under the following assumptions: (1) a dominant interferer
and (2) interference-limited networks. We refer the reader to Section III-D of [17] for further
details regarding these assumptions and the consequent derivation of this theorem.
3) The Distribution of L: With these localizability results, we now finally present the distribution of the number of participating anchors.
Theorem 6. The pdf of L is given by
fL (` | α, λ̃, q, γ, β) = P[L ≥ ` | α, λ̃, q, γ, β] − P[L ≥ ` + 1 | α, λ̃, q, γ, β],
where the support is L ∈ {0, 1, 2, . . . } and the probabilities are given by Lemma 5.
4) The Distribution of L with Frequency Reuse: Using Theorem 6, we may obtain another
expression for fL (`) which incorporates a frequency reuse parameter, K. This parameter models
the ability of base stations to transmit on K separate frequency bands, thereby limiting interference to a per-band basis. This can easily be incorporated into the model by considering K
independent PPPs whose densities are that of the original PPP divided by K. Thus, if nk is
the number of participating base stations in band k, then the total number of participating base
Í
stations is given by L = K
k=1 n k . Thus, to find P[L = `] under frequency reuse, we simply need
to account for all of the per-band combinations of participating base stations such that their sum
equals `. This is given in the following corollary, which is a modification of Theorem 3 of [17].
Corollary 6.1. The pdf of L, given a frequency reuse factor of K, is
!
K
Õ Ö
λ̃
fL (` | α, λ̃, q, γ, β, K) =
fL nk α, , q, γ, β ,
K
k=1
{nÍ1,...,n K }
i ni =`
where the multiplicands are given by Thm. 6, K ∈ {1, 2, . . . }, and the support is L ∈ {0, 1, 2, . . . }.
Remark. When K = 1, this corollary reduces to Theorem 6. Further, this corollary may be
evaluated numerically through the use of a recursive function.
5) Applicability of the Results: Now that we have obtained the pdf of L, we conclude
with a brief discussion regarding its applicability. We begin by taking note of the support of
fL (`). Whereas up until this section we have proceeded under the assumption that the target is
20
localizable, the support of fL (`) now allows us to consider cases where the target is unlocalizable,
i.e., L = 0, 1, 2. Thus, as will be addressed in the following section, we may use these cases to
determine the percentage of the network where a target is unlocalizable.
Lastly, we note that while the localizability results in [17] (Lemma 5) were presented in
the context of cellular networks, these results are actually applicable to any infrastructure-based
wireless network using downlink measurements, so long as the distribution parameters are altered
accordingly. This implies that the distribution for L (Corollary 6.1) also has this applicability,
since it was derived using Lemma 5. Thus, since we use Corollary 6.1, along with a modified
Theorem 4, to derive the marginal distribution of S, then this final distribution will also be
applicable to any infrastructure-based wireless network employing a TOA localization strategy.
F. The Marginal CRLB Distribution
In this section, we modify Theorem 4 and combine this with Corollary 6.1 to obtain the
marginal distribution of S. First, we state one last network assumption often used in practice:
Assumption 5. For a given localization procedure, only a finite number of anchor nodes, N
(≥ 3), are ever tasked to transmit localization signals.
Remark. Just because N anchors are tasked, does not mean that all N signals are necessarily
heard.
Under this assumption, and further considering scenarios where the target is unlocalizable, we
modify Theorem 4 as follows:
FS (s | L = N, σr )
0
FS (s | L, σr ) = FS (s | L, σr )
u(s − M)
L≥N
L = 3, . . . , N − 1
(15)
L = 0, 1, 2
where FS0 is the new modified conditional distribution of S, FS is the previous conditional
distribution given in Theorem 4, and M ∈ R+ is a predetermined localization error value used
to account for unlocalizable scenarios (described in more detail below).
Remark. While Theorem 4 only accounted for scenarios in which the target was localizable,
this modified form, however, now accounts for unlocalizable scenarios. For these scenarios, this
modified conditional distribution yields a step function, which is a valid cdf and corresponds to
a deterministic value for the localization error, i.e., P[S = M | L] = 1. Thus, we account for
21
cases where the target is unlocalizable by assigning an arbitrary localization error value for M,
which is chosen to represent cases where there is ambiguity in a target’s position estimate.8
Remark. It is possible that a mobile may hear more anchors than are tasked to perform the
localization procedure, i.e., L > N. In this case, the N participating anchors are likely those with
the highest received SIR at the target, if connectivity information is known a priori. Therefore,
in this scenario, localization performance will only be based on the N anchors tasked. This is
clearly reflected in the modified conditional distribution of S in (15).
Using this modified Theorem 4 in (15), along with Corollary 6.1, we may now obtain the
distribution of localization error for an entire wireless network:
Theorem 7. The marginal cdf of the localization performance benchmark, S, is
h
i
FS (s | σr , α, λ̃, q, γ, β, K, M, N) = FS (s | ` = N, σr ) 1 − P[L ≤ N − 1|α, λ̃, q, γ, β, K]
+
N−1
Õ
FS (s | `, σr ) fL (` | α, λ̃, q, γ, β, K)
`=3
+ u(s − M) P[L ≤ 2|α, λ̃, q, γ, β, K],
where FS (s | L = `, σr ) is given by Theorem 4, fL (` | α, λ̃, q, γ, β, K) is given by Corollary 6.1,
Íx
and P[L ≤ x|α, λ̃, q, γ, β, K] = `=0
fL (` | α, λ̃, q, γ, β, K).
Proof. Multiplying the modified conditional distribution of S, given in (15), by the marginal
distribution of L from Corollary 6.1 gives the joint distribution of S and L. Then, setting L
equal to a particular realization, `, and summing over all realizations, gives us the marginal cdf
of S, as desired.
Remark. First, recall that the conditional distribution of our approximation of S, FS (s | L = `, σr )
given by Theorem 4, is accurate for lower ` values. Next, note that: (1) fL (`) declines rapidly
as ` increases (an intuitive result, since the probability of hearing many anchor nodes should
be small in infrastructure-based wireless networks), and (2) only a maximum of N nodes are
tasked to perform a localization procedure (Assumption 5). Thus, these two facts validate the
use of our approximation, since the cases where our approximation is less than ideal (i.e., for
8 In
this paper, we choose a value for M that applies for all L < 3. Thus, a large enough M allows for a clear distinction
between the localizable and unlocalizable portions of the network through a quick examination of the cdf of S. Note however,
that one could account for the L = 0, 1, 2 scenarios separately. For example, if L = 1 then in a cellular network one may want
to choose M to be the cell radius, since the user equipment typically knows in which cell it is located [32].
22
large L) will now either be multiplied by fL (`) ≈ 0, or considered invalid in a realistic network
under Assumption 5. As a consequence, Theorem 7 will also retain this accuracy.
Remark. We conclude by noting that this distribution accounts for localization error over all
setups of anchor nodes, numbers of participating anchors, and placements of a target anywhere in
the network. Hence, this distribution completely characterizes localization performance throughout an entire wireless network and represents the main contribution of this work.
IV. N UMERICAL A NALYSIS
In this section, we examine the accuracy of Theorem 7 and investigate how changing network
parameters affects localization performance throughout the network.
A. Description of Simulation Setup
Here, we discuss the parameters that are fixed across all simulations and include a description
of how the simulations were conducted.
1) Fixed Parameter Choices and their Effect on Model Assumptions: For these simulations,
we consider the case of a cellular network, and thus, we place anchor nodes such that the
PPP density matches that of a ubiquitous hexagonal grid with 500m intersite distances (i.e.,
√
λ = 2/ 3 · 5002 m2 ) [17]. Furthermore, we choose a shadowing standard deviation of 8 dB,
which defines our shadowing-transformed density parameter, λ̃.
Next, we set our pathloss exponent, α ≈ 4, which was chosen to represent a pathloss similar to
that seen in a typical cellular network. Note that this pathloss value is indicative of NLOS range
measurements, which are inherently a part of localization in cellular networks. Recall however,
that Assumption 3 implied the use of Line-of-Sight (LOS) measurements. Thus, we attempt to
mimic NLOS in our simulations by selecting a range error to account for a reasonable delay
spread under NLOS conditions. This is described further in the following section. We note that
a subject of future work will be to seek a refinement of this model by incorporating NLOS
directly into the range measurements.
The last parameter that remains fixed across simulations is M. This parameter was chosen to
be large enough such that an examination of the cdf of S will reveal which percentage of the
network the target is unlocalizable. Towards this end, it was sufficient to choose M ≈ 200 m for
our simulations here. Again, the choice of this value is arbitrary and left up to network designers
and how they would like to treat the unlocalizable cases.
23
1
1
Increasing
Frequency Reuse
(K = 1, 2, 3, 6)
0.9
0.8
0.9
0.8
0.7
P[S ≤ abscissa]
P[S ≤ abscissa]
0.7
0.6
0.5
0.4
0.3
0.6
Decreasing q
(q = 1, 0.75, 0.50, 0.25)
0.5
0.4
0.3
0.2
0.2
Solid: True
Dashed: Theorem 7
0.1
Solid: True
Dashed: Theorem 7
0.1
0
0
0
5
10
15
20
25
30
35
40
45
50
0
5
10
15
20
Meters
Fig. 6.
T HE E FFECT
OF
F REQUENCY R EUSE. This result
25
30
35
40
45
50
Meters
Fig. 7.
T HE I MPACT
OF
D ECREASING N ETWORK L OAD.
exposes the large impact that frequency reuse has on local-
This plot demonstrates the improvement in localization perfor-
ization performance throughout the network. The parameters
mance due to a decrease in network loading. The parameters
are: N = 10, β = 10 dB, γ = 20 dB, q = 1, and σr = 20 m.
were chosen as follows: N = 10, β = 10 dB, γ = 20 dB,
The plots’ “piece-wise” appearance is due to L being discrete.
K = 2, and σr = 20 m.
2) Conducting the Simulations: The true marginal cdf of S was generated through a simulation
over 100,000 positioning scenarios. Each scenario consisted of an average placement of 1,000
anchor nodes, placed according to a homogeneous PPP, with the target located at the origin. Next,
the anchor nodes whose SIRs surpassed the detection threshold were deemed to participate in
the localization procedure, and their corresponding coordinates were used to calculate the true
value of S given by Definition 4. If more than N anchor nodes had signals above the threshold,
then the N anchors with highest SIRs were used to calculate S.
B. The Effect of Frequency Reuse on the Network-Wide Distribution of Localization Error
In this section, we explore how frequency reuse impacts localization performance throughout
the network. This simulation, as well as all subsequent simulations, compares the true (simulated)
distribution of S with our analytical model from Theorem 7. All parameters were fixed at the
levels stated in Fig. 6, while the only parameter varied was the frequency reuse factor, K. The
range error standard deviation, σr , was chosen according to the detection threshold, β, and
the CRLB of a range estimate (see [6], equ. 3.40), assuming a 10MHz channel bandwidth.
Approximately 10m was added to account for a reasonable delay spread in NLOS conditions.
From Fig. 6, the most notable impact that frequency reuse has on localization performance is
24
that of localizability. That is, with just a small increase in frequency reuse from K = 1 to K = 2,
the portion of the network with which a target is localizable increases from only ≈ 25% to an
astonishing ≈ 85%. Furthermore, localization error is also reduced, although the improvement
is not as drastic as the increase in localizability. Additionally, as frequency reuse increases, the
gains in localizability stop after K = 3, with the gains in localization error also declining after
K = 3 as well. Thus, we can conclude that an increase in frequency reuse is strongly advisable if
once desires an increase in localization performance within a network, a result which coincides
with what has been seen in practice, viz. 3GPP. Lastly, we note the excellent match between
the true simulated distribution and our analytically derived distribution given by Theorem 7. We
will see that this accuracy of Theorem 7 is retained across all of our results in this section.
C. Examining the Effects of Network Loading
Here we examine the effect that network loading has on localization performance throughout
the network. This is accomplished by varying the percentage of the network, q, actively transmitting (interfering) during a localization procedure. All parameter values other than q were
fixed and σr was chosen in the same manner as in the frequency reuse case.
From the distributions plotted in Fig. 7, we can see that a decrease in network load leads to
an improvement in localizability, as well as an improvement in localization error. However, the
improvement is not as pronounced as in the frequency reuse case. Further, examining the 80th
percentile for example, it is evident that the rate of improvement in localization error declines
as the network load declines as well. Thus, since low network traffic is usually never desirable,
a network designer looking to optimize localization performance may find solace in the fact that
gains in performance begin to decline as network loading decreases also.
D. The Impact of Processing Gain
In this section, we examine the effects of changing the processing gain, since it is perhaps the
easiest parameter for a network designer to change in practice. (Note that we choose σr here as in
the previous simulations.) From Fig. 8, it is evident that as the processing gain increases, there is
a corresponding improvement in localizability across the network, as well as an improvement in
localization error. As a consequence, there exists a clear trade-off between sacrificing processing
time for gains in localization performance. However, it appears that these improvements begin
to level off at a processing gain of ≈25 dB. This is promising, as processing gains higher than
1
1
0.9
0.9
0.8
0.8
0.7
0.7
P[S ≤ abscissa]
P[S ≤ abscissa]
25
0.6
Increasing Processing Gain
(γ = 15 dB, 20 dB, 25 dB)
0.5
0.4
0.3
Increasing σr
(σr = 20m, 40m, 60m, 80m)
0.6
0.5
0.4
0.3
0.2
0.2
Solid: True
Dashed: Theorem 7
0.1
0
Solid: True
Dashed: Theorem 7
0.1
0
0
10
20
30
40
50
60
70
0
Meters
Fig. 8. T HE E FFECT
OF I NCREASING
20
40
60
80
100
120
140
160
Meters
P ROCESSING G AIN.
Fig. 9. T HE I MPACT OF I NCREASING R ANGE E RROR. Inject-
This figure highlights the trade-off that exists between pro-
ing range error results in a predictable effect on localization
cessing time and localization performance. The distribution
performance throughout the network, yet has no effect on
parameter values are: N = 10, β = 5 dB, q = 1, K = 2, and
localizability. The parameter values are: N = 10, β = 10 dB,
σr = 30 m.
γ = 20 dB, K = 3, and q = 1.
this can quickly become impractical. Examining the 80th percentile, we can see that a 10 dB
increase in processing gain can lead to about a 30m improvement in localization error throughout
the network. Thus, increasing the processing gain can be an easily implementable solution for
achieving moderate gains in localization performance.
E. Range Error and its Impact on Localization Performance within the Network
With this last result, we attempt to mimic the effect of an increasing NLOS bias by injecting
additional range error into the measurements. In examining Fig. 9, we note that the first σr value,
i.e., σr = 20 m, was chosen as in the frequency reuse simulation. Therefore, the subsequent
choices represent that of mimicking the impact of increasing NLOS bias. From Fig. 9, we can
see that increasing range error has no effect on localizability within the network. This should
be clear from the analytical model, since σr does not appear as a parameter in Corollary 6.1.
Additionally, injecting more range error into the measurements results in a predictable effect
on the distribution of localization performance. This is also evidenced by examining Theorem
4, where σr appears as a scale parameter. Thus, mimicking the effects of NLOS measurements
results in a scaling of the distribution of S, implying a predictable reduction in localization
performance.
26
V. C ONCLUSION
This paper presents a novel parameterized distribution of localization error, applicable throughout an entire wireless network. Invoking a PPP network model, as well as common assumptions
for TOA localization, has enabled this distribution to simultaneously account for all possible positioning scenarios within a network. Deriving this result involved two main steps: (1) a derivation
of an approximate distribution of our localization performance benchmark, S, conditioned on
the number of hearable anchor nodes, L, which yielded a conditional distribution with desirable
accuracy and tractability properties, and (2) a modification of results from [17] to attain the
distribution of L. Thus, using these two distributions, we arrived at the final, marginal distribution
of S, which also retained the same parameterizations as its two component distributions.
What followed was a numerical analysis of this network-wide distribution of localization
performance. This analysis revealed that our distribution can offer an accurate, baseline tool for
network designers, in that it can be used to get a sense of localization performance within a
wireless network, while also providing insight into which network parameters to change in order
to meet localization requirements.
Since this marginal distribution of S is a distribution of the TOA-CRLB throughout the
network, it consequently provides a benchmark for describing localization performance in any
network employing an unbiased, efficient, TOA-based localization algorithm. The amount of
insights that this network-wide distribution of localization error reveals are numerous, and the
results presented in this paper have only begun to explore this new paradigm. It is our hope
that this work spawns additional research into this new concept, and that future work, such
as incorporating NLOS measurements, accounting for collaboration, adding other localization
strategies (TDOA/AOA), etc., can further refine the model presented here.
In closing, this work presents an initial attempt to provide network designers with a tool for
analyzing localization performance throughout a network, freeing them from lengthly simulations
by offering an accurate, analytical solution.
27
A PPENDIX A
P ROOF OF P ROPOSITION 1
It first helps to visualize the sin2 (·) terms of the sum in (5) on a 2-D grid, where the i’s
represent the rows and the j’s represent the columns. As an example, the case of L = 4 gives
j=1 j=2
i=1
j=3
j=4
∠1 + ∠2 ∠1 + ∠2 + ∠3
∠1
i=2
∠2 + ∠3
∠2
i=3
∠3
where just the arguments of the sin2 (·) terms are represented in the grid for clarity. From this
arrangement, it is evident that the sum in (5) represents the process of summing each row
sequentially, starting at i = 1.
Now, however, we choose to sum the terms diagonally, starting with the lowest diagonal and
working our way upward. This yields
D=
L−1 Õ
L−i
Õ
i=1 j=1
sin2
j+i−1
Õ
!
∠k .
(16)
k= j
Considering the cases i = 1 and i = L − 1 separately, we may rewrite (16) as
!
!
j+i−1
L−1
L−1
L−2 Õ
L−i
Õ
Õ
Õ
Õ
2
2
2
∠ k + sin
∠k .
D=
sin ∠ k +
sin
i=2 j=1
k=1
k= j
k=1
Next, note that
2
sin
L−1
Õ
!
∠ k = sin
2
2π −
L−1
Õ
!
∠ k = sin2 2π − (Θ(L) − Θ(1) ) = sin2 ∠ L ,
(17)
k=1
k=1
where the last two equalities follow from Definition 3. Hence,
D=
L
Õ
k=1
sin2 ∠ k +
L−2 Õ
L−i
Õ
sin2
i=2 j=1
j+i−1
Õ
!
∠k .
k= j
Lastly, we complete the proof by noting that the first sum may be equivalently expressed by
replacing the internodal angles with their order statistics.
A PPENDIX B
P ROOF OF C OROLLARY 2.1
Using Lemma 2 and our assumption of a finite L, the pdf of ∠(L−1) conditioned on L is
!
! L−2
X
Õ
d
L
n(n
−
1)(L
−
1)
nϕ
f∠(L−1) (ϕ|L) =
F∠ (ϕ|L) =
(−1)n
1−
,
dϕ (L−1)
n
2π
2π
n=0
28
where X = min L, b2π/ϕc and the support is 0 ≤ ∠(L−1) ≤ π, as in Lemma 2.
Next, note that the lower summation limit may be rewritten as n = 2 and that the upper
summation limit, X, can be simplified to just L by appending an indicator function to the
summand. This gives the logically equivalent expression:
!
! L−2 "
#
L
Õ
n(n − 1)(L − 1)
nϕ
2π
n L
f∠(L−1) (ϕ|L) =
(−1)
1−
1 ϕ≤
.
n
2π
2π
n
n=2
Using the above expression for the pdf, the expectation is derived as follows:
∫ π
E[∠(L−1) |L] =
ϕ f∠(L−1) (ϕ|L) dϕ
0
!
! L−2
L n(n − 1)(L − 1)
nϕ
ϕ
(−1)
1−
dϕ
=
n
2π
2π
0
n=2
! ∫ 2π
! L−2
L
Õ
n
nϕ
n(n − 1)(L − 1)
n L
=
(−1)
ϕ 1−
dϕ
n
2π
2π
0
n=2
!
L
Õ
L
2π(n
−
1)
=
(−1)n
,
nL
n
n=2
∫
2π
n
L
Õ
n
(18)
(19)
(20)
where (18) follows from absorbing the indicator function into the integration limits since n ≥ 2,
(19) follows from our assumption of a finite L, and (20) is derived through integration by parts.
A PPENDIX C
P ROOF OF L EMMA 3
This is a straightforward application of the lowest Geometric Dilution Of Precision (GDOP)
presented in [30]. First, since L is finite, then D is a continuous real function defined on a
compact subset of R L . Thus, its maximum must exist. Call it dmax .
Next, under our Assumptions 2, 3, and 4, GDOP, presented in [27], can be written as GDOP =
p
L/D, where D follows from the lemma assumption. Further, under our assumptions, [30] asserts
√
that the lowest GDOP is given by GDOPmin = 2/ L. Since GDOPmin must occur when D is at
p
√
its maximum, we have that 2/ L = L/dmax , and the lemma follows.
A PPENDIX D
P ROOF OF T HEOREM 4
Under Approximation 1, we have S = a/sin(∠(L−1) ), where a is defined in Theorem 4. To
determine the support of S conditioned on L, we know from Lemma 2 that if 0 ≤ ∠(L−1) ≤ π,
29
then the support for the RV sin(∠(L−1) ) must be 0 ≤ sin(∠(L−1) ) ≤ 1. From here, we see that
0 ≤ sin(∠(L−1) ) ≤ 1
⇒
1≤
1
sin(∠(L−1) )
⇒
a≤
a
sin(∠(L−1) )
⇒
a ≤ S,
and hence, S ∈ [a, ∞).
Next, to find the cdf of S conditioned on L, consider the following
FS (s | L, σr ) = P[S ≤ s | L, σr ]
"
a
≤s
=P
sin(∠(L−1) )
"
= P ∠(L−1) ≤ π−sin−1
#
(21)
L
a
s
!
#
"
L − P ∠(L−1) ≤ sin−1
a
s
!
#
L
= P[∠(L−1) ≤ ϕ2 | L] − P[∠(L−1) ≤ ϕ1 | L]
= F∠(L−1) ϕ2 | L − F∠(L−1) ϕ1 | L ,
(22)
(23)
where (21) follows from Approximation 1 and the fact that we may drop the parameter, σr , as
a condition since the dependency is now explicit. Further, ϕ1 and ϕ2 from (22) are defined in
Theorem 4, and (23), through use of Lemma 2, gives (14) from Theorem 4, as desired.
R EFERENCES
[1] C. E. O’Lone and R. M. Buehrer, “Towards a Characterization of Localization Performance in Networks with Random
Geometries,” in IEEE International Conference on Communications (ICC), Paris, France, May 2017.
[2] P. Baronti, P. Pillai, V. W. C. Chook, S. Chessa, A. Gotta, and Y. F. Hu, “Wireless sensor networks: A survey on the state
of the art and the 802.15.4 and ZigBee standards,” in Comput. Commun. 30, pp. 1655-1695, May 2007.
[3] M. A. M. Vieira, C. N. Coelho, D. C. da Silva, and J. M. da Mata, “Survey on wireless sensor network devices,” in Proc.
of IEEE Conference on Emerging Technologies and Factory Automation., vol. 1, pp. 537-544, 2003.
[4] N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero, R. L. Moses, and N. S. Correal, “Locating the nodes: cooperative
localization in wireless sensor networks,” in IEEE Signal Processing Magazine, vol. 22, no. 4, pp. 54-69, July 2005.
[5] N. A. Alrajeh, M. Bashir, and B. Shams, “Localization Techniques in Wireless Sensor Networks,” in International Journal
of Distributed Sensor Networks, vol. 9, no. 6, 2013.
[6] S. M. Kay, Fundamentals of Statistical Signal Processing. Upper Saddle River, NJ: Prentice-Hall, Inc., vol. 1, 1993.
[7] C. Chang and A. Sahai, “Estimation bounds for localization,” in First Annual IEEE Communications Society Conference
on Sensor and Ad Hoc Communications and Networks. IEEE SECON, pp. 415-424., 2004.
[8] H. Wang, L. Yip, K. Yao, and D. Estrin, “Lower bounds of localization uncertainty in sensor networks,” in IEEE International
Conference on Acoustics, Speech, and Signal Processing, vol. 3, pp. iii-917-20, 2004.
[9] I. Guvenc and C. C. Chong, “A Survey on TOA Based Wireless Localization and NLOS Mitigation Techniques,” in IEEE
Communications Surveys & Tutorials, vol. 11, no. 3, pp. 107-124, 3rd Quarter 2009.
[10] A. Savvides et al., “On The Error Characteristics of Multihop Node Localization In Ad-Hoc Sensor Networks,” in ISPN’03
Proceedings of the 2nd international conference on Information Processing in Sensor Networks, pp. 317-332, 2003.
30
[11] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic Geometry and Random Graphs for
the Analysis and Design of Wireless Networks,” in IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1029-1046, Sept., 2009.
[12] M. Haenggi, Stochastic Geometry for Wireless Networks. New York: Cambridge University Press, 2013
[13] B. Huang, T. Li, B. D. O. Anderson, and C. Yu, “On the performance limit of sensor localization,” in 50th IEEE Conference
on Decision and Control and European Control Conference, Orlando, FL, pp. 7870-7875, 2011.
[14] B. Huang et al., “On the performance limit of single-hop TOA localization,” in 12th International Conference on Control
Automation Robotics & Vision (ICARCV), Guangzhou, pp. 42-46, 2012.
[15] B. Huang et al., “Performance Limits in Sensor Localization,” in Automatica, vol. 49, no. 2, pp. 503-509, Feb. 2013.
[16] F. Zhou and Y. Shen, “On the Outage Probability of Localization in Randomly Deployed Wireless Networks,” in IEEE
Communications Letters, vol. 21, no. 4, pp. 901-904, April 2017.
[17] J. Schloemann, H. S. Dhillon, and R. M. Buehrer, “Towards a Tractable Analysis of Localization Fundamentals in Cellular
Networks,” in IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 1768-1782, March 2016.
[18] Code of Federal Regulations, “911 Service,” 47 C.F.R. 20.18 (h), 2015
[19] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A Tractable Approach to Coverage and Rate in Cellular Networks,” in IEEE
Trans. Commun., vol. 59, no. 11, pp. 3122-3134, Nov. 2011.
[20] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and Analysis of K-Tier Downlink Heterogeneous
Cellular Networks,” in IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 550-560, Apr. 2012.
[21] B. Blaszczyszyn, M. K. Karray, and H. P. Keeler, “Using Poisson processes to model lattice cellular networks,” in Proc.
IEEE INFOCOM, pp. 773-781, Apr. 2013.
[22] H. P. Keeler, N. Ross, and A. Xia, “When do wireless network signals appear Poisson?,” in arXiv:1411.3757, 2014.
[23] B. Blaszczyszyn, M. K. Karray, and H. P. Keeler, “Wireless networks appear Poissonian due to strong shadowing,” in
IEEE Trans. Wireless Commun., vol. 14, no. 8, Aug. 2015.
[24] R. Zekavat and R. M. Buehrer, Handbook of Position Location: Theory, Practice, and Advances. Wiley, 2012.
[25] D. B. Jourdan, D. Dardari, and M. Z. Win, “Position Error Bound for UWB Localization in Dense Cluttered Environments,”
IEEE International Conference on Communications, Istanbul, 2006, pp. 3705-3710.
[26] D. B. Jourdan and N. Roy, “Optimal Sensor Placement for Agent Localization,” IEEE/ION Position, Location, And
Navigation Symposium, 2006, pp. 128-139.
[27] J. Schloemann, H. S. Dhillon, and R. M. Buehrer, “A Tractable Metric for Evaluating Base Station Geometries in Cellular
Network Localization”, IEEE Wireless Commun. Lett., vol. 5, no. 2, pp. 140-143, Apr. 2016.
[28] J. Nounagnon, “Using Kullback-Leibler Divergence to Analyze the Performance of Collaborative Positioning,” Ph.D.
Dissertation, Bradley Dept. of Electrical and Computer Eng., Virgina Tech, Blacksburg, VA, 2016.
[29] T. M. Cover and J. A. Thomas, Elements of Information Theory. New York, NY: Wiley-Interscience, 1991, chs. 2 and 9.
[30] X. Lv, K. Liu and P. Hu, “Geometry Influence on GDOP in TOA and AOA Positioning Systems,” Second International
Conference on Networks Security, Wireless Communications and Trusted Computing, Wuhan, Hubei, 2010, pp. 58-61.
[31] H. S. Dhillon and J. G. Andrews, “Downlink Rate Distribution in Heterogeneous Cellular Networks under Generalized
Cell Selection,” in IEEE Wireless Communications Letters, vol. 3, no. 1, pp. 42-45, February 2014.
[32] T. Bhandari, H. S. Dhillon, and R. M. Buehrer, “The Impact of Proximate Base Station Measurements on Localizability
in Cellular Systems, in Proc. IEEE SPAWC, Edinburgh, UK, July 2016. (Invited paper)
| 7cs.IT
|
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE∗
arXiv:1602.08298v2 [cs.DS] 22 Jan 2018
JOHN AUGUSTINE† , WILLIAM K. MOSES JR.‡ , AMANDA REDLICH§ , AND ELI UPFAL¶
Abstract.
Load balancing is a well-studied problem, with balls-in-bins being the primary framework. The
greedy algorithm Greedy[d] of Azar et al. places each ball by probing d > 1 random bins and placing
the ball in the least loaded of them. With high probability, the maximum load under Greedy[d]
is exponentially lower than the result when balls are placed uniformly randomly. Vöcking showed
that a slightly asymmetric variant, Left[d], provides a further significant improvement. However, this
improvement comes at an additional computational cost of imposing structure on the bins.
Here, we present a fully decentralized and easy-to-implement algorithm called FirstDiff[d] that
combines the simplicity of Greedy[d] and the improved balance of Left[d]. The key idea in FirstDiff[d]
is to probe until a different bin size from the first observation is located, then place the ball. Although
the number of probes could be quite large for some of the balls, we show that FirstDiff[d] requires
only at most d probes on average per ball (in both the standard and the heavily-loaded settings).
Thus the number of probes is no greater than either that of Greedy[d] or Left[d]. More importantly,
we show that FirstDiff[d] closely matches the improved maximum load ensured by Left[d] in both the
standard and heavily-loaded settings. We further provide a tight lower bound on the maximum load
up to O(log log log n) terms. We additionally give experimental data that FirstDiff[d] is indeed as
good as Left[d], if not better, in practice.
Key words.
allocation
Load balancing, FirstDiff, Balanced allocation, Randomized algorithms, Task
AMS subject classifications. 60C05, 60J10, 68R05
1. Introduction. Load balancing is the study of distributing loads across multiple entities such that the load is minimized across all the entities. This problem
arises naturally in many settings, including the distribution of requests across multiple servers, in peer-to-peer networks when requests need to be spread out amongst
the participating nodes, and in hashing. Much research has focused on practical
implementations of solutions to these problems [12, 6, 13].
Our work builds on several classic algorithms in the theoretical balls-in-bins
model. In this model, m balls are to be placed sequentially into n bins and each
ball probes the load in random bins in order to make its choice. Here we give a new
algorithm, FirstDiff[d], which performs as well as the best known algorithm, Left[d],
while being significantly easier to implement.
The allocation time for a ball is the number of probes made to different bins before
placement. The challenge is to balance the allocation time versus the maximum bin
load. For example, using one probe per ball, i.e. placing each ball uniformly at
∗ An earlier version of this paper appeared in [1]. The previous version had only expectation
upper bounds for average number of probes and no lower bounds. This version has lower bounds on
maximum load and high probability upper bounds for average number of probes, as well as cleaner
proofs for the same.
† Department of Computer Science & Engineering, Indian Institute of Technology Madras, Chennai, India. ([email protected]). Supported by the IIT Madras New Faculty Seed Grant, the
IIT Madras Exploratory Research Project, and the Indo-German Max Planck Center for Computer
Science (IMPECS).
‡ Department of Computer Science & Engineering, Indian Institute of Technology Madras, Chennai, India. ([email protected]).
§ Department of Mathematics, Bowdoin College, ME, USA ([email protected]). This material
is based upon work supported by the National Science Foundation under Grant No. DMS-0931908
while the author was in residence at the Institute for Computational and Experimental Research in
Mathematics in Providence, RI, during the Spring 2014 semester.
¶ Department of Computer Science, Brown University, RI, USA ([email protected]).
1
2
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
random, the maximum load of any bin when m = n will be lnlnlnnn (1 + o(1)) (with high
probability1 ) and total allocation time of n probes [11]. On the other hand, using d
probes per ball and placing the ball in the lightest bin, i.e. Greedy[d], first studied by
Azar et. al [2], decreases the maximum load to lnlnlndn + O(1) with allocation time of
nd. In other words, using d ≥ 2 choices improves the maximum load exponentially,
at a linear allocation cost.
Vöcking [15] introduced a slightly asymmetric algorithm, Left[d], which quite
surprisingly guaranteed (w.h.p.) a maximum load of dlnlnlnφnd + O(1) (where φd is a
constant between 1.61 and 2 when d ≥ 2) using the same allocation time of nd probes
as Greedy[d] when m = n. This analysis of maximum load for Greedy[d] and Left[d]
was extended to the heavily-loaded case (when m ≫ n) by Berenbrink et al. [3].
However, Left[d] in [15] and [3] utilizes additional processing. Bins are initially sorted
into groups and treated differently according to group membership. Thus practical
implementation, especially in distributed settings, requires significant computational
effort in addition to the probes themselves.
Our Contribution. We present a new algorithm, FirstDiff[d]. This algorithm requires no pre-sorting of bins; instead FirstDiff[d] uses real-time feedback to adjust its
number of probes for each ball2 .
The natural comparison is with the classic Greedy[d] algorithm; FirstDiff[d] uses
the same number of probes, on average, as Greedy[d] but produces a significantly
smaller maximum load. In fact, we show that the maximum load is as small as that of
Left[d] when m = n. Furthermore, it is comparable to Left[d] when heavily loaded. For
both the m = n and heavily loaded cases, FirstDiff[d] has much lower computational
overhead than Left[d].
This simpler implementation makes FirstDiff[d] especially suitable for practical
applications; it is amenable to parallelization, for example, and requires no central
control or underlying structure. Some applications have a target maximum load and
aim to minimize the necessary number of probes. From this perspective, our algorithm
again improves on Greedy[d]: the maximum load of FirstDiff[ln d] is comparable to that
of Greedy[d], and uses exponentially fewer probes per ball.
Theorem 3.1 Use FirstDiff[d], where maximum number of probes allowed per ball is
22d/3 , to allocate n balls into n bins. The average number of probes required per ball
is at most d on expectation and w.h.p. Furthermore, the maximum load of any bin is
log n
at most log
n ≥ max(2, n0 ), where
0.66d + O(1) with high probability when d ≥ 4 and
4
log n
≤ n12 .
n0 is the smallest value of n such that for all n > n0 , 36 log n 72e5n
Theorem 4.1 Use FirstDiff[d], where maximum number of probes allowed per ball
is 2d/2.17 , to allocate m balls into n bins. When m ≥ 72(nλ log n + n) where λ is
taken from Lemma 4.5, n ≥ n0 where n0 is the smallest value of n that satisfies
0.00332n(λ log n + 1)d/2d/2.17 ≥ log n, and d ≥ 6, it takes at most d probes on average
to place every ball on expectation and with high probability. Furthermore, for an
absolute constant c,
Pr Max. load of any bin >
m
n
+
log log n
0.46d
+ c log log log n ≤ c(log log n)−4 .
Our technique for proving that the average number of probes is bounded is novel
1 We use the phrase “with high probability” (or w.h.p. in short) to denote probability of the form
1 − O(n−c ) for some suitable c > 0. Furthermore, every log in this paper is to base 2 unless otherwise
mentioned.
2 Thus we are concerned with the average number of probes per ball throughout this paper.
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
3
to the best of our knowledge. As the number of probes required by each ball is
dependent on the configuration of the balls-in-bins at the time the ball is placed,
the naive approach to computing its expected value quickly becomes too conditional.
Instead, we show that this conditioning can be eliminated by carefully overcounting
the number of probes required for each configuration, leading to a proof that is then
quite simple. The heavily-loaded case is significantly more complex than the m = n
case; however the basic ideas remain the same.
The upper bound on the maximum load is proved using the layered induction
technique. However, because FirstDiff[d] is a dynamic algorithm, the standard recursion used in layered induction must be altered. We use coupling and some more
complex analysis to adjust the standard layered induction to this context.
We furthermore provide a tight lower bound on the maximum load for a broad
class of algorithms which use variable probing.
Theorem 5.1 Let Alg[k] be any algorithm that places m balls into n bins, where
m ≥ n, sequentially one by one and satisfies the following conditions:
1. At most k probes are used to place each ball.
2. For each ball, each probe is made uniformly at random to one of the n bins.
3. For each ball, each probe is independent of every other probe.
The maximum load of any bin after placing all m balls using Alg[k] is at least m
n +
ln ln n
ln k − Θ(1) with high probability.
We use the above theorem to provide a lower bound on the maximum load of
FirstDiff[d], which is tight up to O(log log log n) terms.
Theorem 5.2 The maximum load of any bin after placing m balls into n bins using
FirstDiff[d], where maximum number of probes allowed per ball is 2Θ(d) , is at least
ln ln n
m
n + Θ(d) − Θ(1) with high probability.
Related Work. Several other algorithms in which the number of probes performed
by each ball is adaptive in nature have emerged in the past, such as work done by
Czumaj and Stemann [5] and by Berenbrink et al. [4].
Czumaj and Stemann [5] present an interesting “threshold” algorithm. First they
define a process Adaptive-Allocation-Process, where each load value has an associated
threshold and a ball is placed when the number of probes that were made to find a bin
of a particular load exceeded the associated threshold. Then, by carefully selecting the
thresholds for the load values, they develop M-Threshold, where each ball probes bins
until it finds one whose load is within some predetermined threshold. The bounds
on maximum load and on average allocation time are better than our algorithm’s,
but the trade-off is that computing the required threshold value often depends on the
knowledge (typically unavailable in practical applications) of the total number of balls
that will ever be placed. Furthermore their proofs are for m = n and don’t extend
easily to when m > n.
More recently, Berenbrink et al. [4] develop a new threshold algorithm, Adaptive,
which is similar to M-Threshold but where the threshold value used for a given ball
being placed depends on the number of balls placed thus far. They analyze this algorithm when m ≥ n and also extend the analysis of M-Threshold from [5] to the
m ≥ n case. They show that both algorithms have good bounds on maximum load
and average allocation time, but again this comes at the trade-off of requiring some
sort of global knowledge when placing the balls. In the case of Adaptive, each ball
must know what its order in the global placement of balls is, and in the case of M-
4
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Threshold, each ball must know the total number of balls that will ever be placed.
Our algorithm is unique in that it requires no such global knowledge at all; it is able
to make decisions based on the probed bins’ load values alone.
Definitions. In the course of this paper we will use several terms from probability
theory, which we define below for convenience.
Consider two Markov chains At and Bt over time t ≥ 0 with state spaces S1 and
S2 respectively. A coupling (cf. [8]) of At and Bt is a Markov chain (At , Bt ) over time
t ≥ 0 with state space S1 × S2 such that At and Bt maintain their original transition
probabilities.
Consider two vectors u, v ∈ Zn . Let u′ and v ′ be permutations of u and v
′
respectively such that u′i ≥ u′i+1 and vi′ ≥ vi+1
for all 1 ≤ i ≤ n − 1. We say u
majorizes v (or v is majorized by u) when
i
i
P
P
u′j ≥
vj′ , ∀1 ≤ i ≤ n.
j=1
j=1
For a given allocation algorithm C which places balls into n bins, we define the load
vector ut ∈ (Z∗ )n of that process after t balls have been placed as follows: the ith
index of ut denotes the load of the ith bin (we can assume a total order on the bins
according to their IDs). Note that ut , t ≥ 0, is a Markov chain.
Consider two allocation algorithms C and D that allocate m balls. Let the load
vectors for C and D after t balls have been placed using the respective algorithms be
ut and v t respectively. We say that C majorizes D (or D is majorized by C) if there
is a coupling between C and D such that ut majorizes v t for all 0 ≤ t ≤ m.
Berenbrink et al. [3] provide an illustration of the above ideas being applied in
the load balancing context.
We also use Theorem 2.1 from Janson [7] in order to achieve high probability
concentration bounds on geometric random variables. We first set up the terms in
the theorem and then restate it below. Let X1 , . . . , Xn be n ≥ 1 geometric random
n
P
Xi ,
variables with parameters p1 , . . . , pn respectively. Define p∗ = mini pi , X =
and µ = E[X] =
n
P
i=1
i=1
1
pi .
Now we have the following lemma.
Lemma 1.1 (Theorem 2.1 in [7]).
P r(X ≥ Λµ) ≤ e−p∗ µ(Λ−1−ln Λ) .
For any p1 , . . . , pn ∈ (0, 1] and any Λ ≥ 1,
Organization of Paper. The structure of this paper is as follows. In Section 2, we
define the model formally and present the FirstDiff[d] algorithm. We then analyze the
algorithm when m = n in Section 3 and give a proof that the total number of probes
used by FirstDiff[d] to place n balls is nd with high probability, while the maximum
log n
bin load is still upper bounded by log
0.66d + O(1) with high probability. We provide
the analysis of the algorithm when m > n in Section 4, namely that the number
of probes is on average d per ball with high probability and the maximum bin load
log log n
is upper bounded by m
n + 0.46d + O(log log log n) with probability close to 1. We
provide a matching lower bound for maximum bin load tight up to the O(log log log n)
term for algorithms with variable number of probes and FirstDiff[d] in particular in
Section 5. In Section 6, we give experimental evidence that our FirstDiff[d] algorithm
indeed results in a maximum load that is comparable to Left[d] when m = n. Finally,
we provide some concluding remarks and scope for future work in Section 7.
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
5
2. The FirstDiff[d] Algorithm. The idea behind this algorithm is to use probes
more efficiently. In the standard d-choice model, effort is wasted in some phases. For
example, early on in the distribution, most bins have size 0 and there is no need to
search before placing a ball. On the other hand, more effort in other phases would
lead to significant improvement. For example, if .9n balls have been distributed, most
bins already have size at least 1 and thus it is harder to avoid creating a bin of size
2. FirstDiff[d] takes this variation into account by probing until it finds a difference,
then making its decision.
This algorithm uses probes more efficiently than other, fixed-choice algorithms,
while still having a balanced outcome. Each ball probes at most 2Θ(d) bins (where
d ≥ 6 and by extension 2Θ(d) > 2) uniformly at random until it has found two bins
with different loads (or a bin with zero load) and places the ball in the least loaded of
the probed bins (or the zero loaded bin). If all 2Θ(d) probed bins are equally loaded,
the ball is placed (without loss of generality) in the last probed bin. The pseudocode
for FirstDiff[d] is below. Note that we use the Θ() to hide a constant value. The exact
values are different for m = n and m ≫ n and are 2/3 and 1/2.17 respectively.
Algorithm 1 FirstDiff[d]
(Assume 2Θ(d) > 2. The following algorithm is executed for each ball.)
1: Repeat 2Θ(d) times
2:
Probe a new bin chosen uniformly at random
3:
if The probed bin has zero load then
4:
Place the ball in the probed bin and exit
5:
if The probed bin has load that is different from those probed before then
6:
Place the ball in the least loaded bin (breaking ties arbitrarily) and exit
7: Place the ball in the last probed bin
As we can see, the manner in which a ball can be placed using FirstDiff[d] can be
classified as follows:
1. The first probe was made to a bin with load zero.
2. All probes were made to bins of the same load.
3. One or more probes were made to bins of larger load followed by a probe to
a bin of lesser load.
4. One or more probes were made to bins of lesser load followed by a probe to
a bin of larger load.
3. Analysis of FirstDiff[d] when m = n.
Theorem 3.1. Use FirstDiff[d], where maximum number of probes allowed per
ball is 22d/3 , to allocate n balls into n bins. The average number of probes required per
ball is at most d on expectation and w.h.p. Furthermore, the maximum load of any
log n
bin is at most log
n0 ),
0.66d + O(1) with high probability when d ≥ 4 and n ≥ max(2,
4
72e log n
≤ n12 .
where n0 is the smallest value of n such that for all n > n0 , 36 log n
5n
Proof. First, we show that an upper bound on the average number of probes per
ball is d on expectation and w.h.p. Subsequently, we show that the maximum load at
the end of placing all n balls is as desired w.h.p.
6
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
3.1. Proof of Number of Probes.
Lemma 3.2. The number of probes required to place m = n balls into n bins using
FirstDiff[d], where maximum number of probes allowed per ball is 22d/3 , is at most nd
on expectation and with high probability when d ≥ 4.
Proof. Let k be the maximum number of probes allowed to be used by FirstDiff[d]
per ball, i.e. k = 22d/3 . We show that the total number of probes required to place
all balls does not exceed 1.5n log k w.h.p. and thus nd probes are required to place
all balls.
Let the balls be indexed from 1 to n in the order in which they are placed. Our
analysis proceeds in two phases. For a value of T that will be fixed subsequently, the
first T + 1 balls are analyzed in the first phase and remaining balls are analyzed in
the second. Consider the ball indexed by t, 1 ≤ t ≤ n. Let Xt be the random variable
denoting the number of probes it takes for FirstDiff[d] to place ball t.
Phase One: t ≤ T + 1. We couple FirstDiff[d] with the related process that probes
until it finds a difference in bin loads or runs out of probes, without treating empty
bins as special; in other words, the FirstDiff[d] algorithm without lines 3 and 4. One
additional rule for the related process is that if an empty bin is probed first, then after
the process finishes probing, the ball will be placed in that first probed bin, i.e. the
empty bin. Note that this is a valid coupling; if an empty bin is probed then under
both FirstDiff[d] and this process the ball is placed in an empty bin, and if no empty
bin is probed the two processes are exactly the same. Let Yt be the number of probes
required by this related process to place ball t in the configuration where there are
t − 1 bins of load 1 and n − t + 1 bins of load 0. Notice that for any configuration
of balls in bins, Xt ≤ Yt ; furthermore, the configuration after placement under both
FirstDiff[d] and this new process is the same. You can see this by a simple sequence
of couplings.
First, choose some arbitrary configuration with nαi bins of size i for i = 0, 1, 2, . . ..
That configuration will be probed until bins of two different sizes are discovered,
i.e. until the set probed
Pn intersects two distinct αi and αj . Couple this with the
configuration that has i=1 nαi bins of size 1 and the rest of size 0. This configuration
requires more probes than the original configuration; it continues until the set probed
intersects α0 and α6=0 . Second, note that the configuration with t − 1 bins of size 1
and n − t + 1 bins of size 0 requires even more probes than this one. This is because
restricting the bins to size 1 can only decrease the number of empty bins. Finally,
note that the ball’s placement under either FirstDiff[d] or this new process leads to
the same configuration at time t + 1 (up to isomorphism); if FirstDiff[d] places a ball
in an empty bin, so does this process.
We first derive the expected value of Yt . The expected number of probes used
by FirstDiff[d] is upper bounded by the expected number of probes until a size-0 bin
appears, i.e. the expected number of probes used by the FirstDiff[d] algorithm without
line 1. This is of course n/(n − t + 1). The overall expected number of probes for the
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
7
first T + 1 steps is
E[Y ] = E[
T
+1
X
Yt ]
t=1
=
T
+1
X
E[Yt ]
t=1
≤
T
+1
X
t=1
=n
n
n−t+1
n
X
1
i
i=n−T
∼ n(log n − log(n − T ))
n
= n log
.
n−T
Now we will find T such that the expected number of probes in phase one, E[Y ],
is n log k, i.e.
n
= n log k.
n log
n−T
Solving, we get T = n(1 − 1/k). Now, recall that we want a high probability bound
on the number of probes required to place each ball in Phase One when running
Pn(1−1/k)
FirstDiff[d], i.e.
Xt . Recall that Xt ≤ Yt , and as such a high probability
t=0
Pn(1−1/k)
Yt suffices. We can now use Lemma 1.1 with Λ = 1.01, µ =
bound on t=0
n log k, and p∗ = k1 .
n(1−1/k)
n(1−1/k)
X
X
Pr
Xt ≥ 1.01n log k ≤ P r
Yt ≥ 1.01n log k
t=0
t=0
≤e
1
·(n log k)·(1.01−1−ln 1.01)
−k
1
since k << n
≤O
n
Phase Two: t > T + 1. Rather than analyzing in detail, we use the fact that the
number of probes for each ball is bounded by k, i.e. Xt ≤ k, ∀t > T + 1. So the
number of probes overall in this phase is at most
k(n − T − 1) = k(n − n(1 − 1/k) − 1) = n − k.
So the total number of probes w.h.p. is 1.01n log k + n − k ≤ 1.5n log k (when
k = 22d/3 and d ≥ 4). When k = 22d/3 , an upper bound on the number of probes to
place all n balls is nd probes on expectation and w.h.p., as desired.
3.2. Proof of Maximum Load.
Lemma 3.3. The maximum load in any bin after using FirstDiff[d], where maximum number of probes allowed per ball is 22d/3 , to allocate n balls into n bins is at
log n
n ≥ max(2, n0 ), where n0
most log
0.66d + O(1) with high probability when d ≥ 4 and
4
log n
is the smallest value of n such that for all n > n0 , 36 log n 72e5n
≤ n12 .
8
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Proof. While the proof follows along the lines of the standard layered induction
argument [8, 14], we have to make a few non-trivial adaptations to fit our context
where the number of probes is not fixed.
Let k be the maximum number of probes allowed to be used by FirstDiff[d] per
ball, i.e. k = 22d/3 . Define vi as the fraction of bins of load at least i after n balls are
placed. Define ui as the number of balls of height at least i after n balls are placed.
It is clear that vi ∗ n ≤ ui .
We wish to show that the Pr(Max. load ≥ logloglogk n + γ) ≤ n1c for some constants
γ ≥ 1 and c ≥ 1. Set i∗ = logloglogk n + 11 and γ = 15. Equivalently, we wish to show
that Pr(vi∗ +4 > 0) ≤ n1c for some constant c ≥ 1.
In order to aid us in this proof, let us define a non-increasing series of numbers
1
.
β11 , β12 , . . . , βi∗ as upper bounds on v11 , v12 , . . . vi∗ . Let us set β11 = 11
Now,
Pr(vi∗ +4 > 0) = Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ )Pr(vi∗ ≤ βi∗ )
+ Pr(vi∗ +4 > 0|vi∗ > βi∗ )Pr(vi∗ > βi∗ )
≤ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ ) + Pr(vi∗ > βi∗ )
= Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ )
+ Pr(vi∗ > βi∗ |vi∗ −1 ≤ βi∗ −1 )Pr(vi∗ −1 ≤ βi∗ −1 )
+ Pr(vi∗ > βi∗ |vi∗ −1 > βi∗ −1 )Pr(vi∗ −1 > βi∗ −1 )
≤ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ )
∗
(1)
+
i
X
Pr(vi > βi |vi−1 ≤ βi−1 ) + Pr(v11 > β11 )
i=12
Here, Pr(v11 > β11 ) = 0. It remains to find upper bounds for the remaining two
terms in the above equation.
We now derive a recursive relationship between the βi ’s for i ≥ 11. βi+1 acts as
an upper bound for the fraction of bins of height at least i + 1 after n balls are placed.
In order for a ball placed to land up at height at least i + 1, one of 3 conditions must
occur:
• All k probes are made to bins of height at least i.
• Several probes are made to bins of height at least i and one is made to a bin
of height at least i + 1.
• One probe is made to a bin of height at least i and several probes are made
to bins of height at least i + 1.
Thus the probability that a ball is placed at height at least i + 1, conditioning on
vj ≤ βj for j ≤ i + 1 at that time, is
≤ βik + βi βi+1 1 + βi + βi2 + . . . + βik−2
k−2
2
+ . . . + βi+1
+ βi βi+1 1 + βi+1 + βi+1
!
k−1
1 − βi+1
1 − βik−1
k
+
≤ βi + βi βi+1
1 − βi
1 − βi+1
1
≤ βik + β11 βi+1 2 ∗
1 − β11
2β
i+1
≤ βik +
10
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
9
Let vi+1 (t) be the fraction of bins with load at least i + 1 after the 1 ≤ t ≤ n ball
is placed in a bin.
Let t∗ = min[arg mint vi+1 (t) > βi+1 , n], i.e. t∗ is the first t such that vi+1 (t) >
βi+1 or n if there is no such t. The probability that t∗ < n is bounded by the
i+1
) is greater than βi+1 n.
probability that a binomial random variable B(n, βik + 2β10
Fix βi+1 =
10 k
3 βi
≥
2n(βik +
n
2βi+1
10
)
. Then using a Chernoff bound, we can say that
with high probability, t∗ = n or vi+1 ≤ βi+1 , so long as e
some constant c ≥ 1.
n
, e−
Now, so long as βi+1 ≥ 18 log
n
18 log n
. This
the value of βi dips below
n
log β11 = − log 11 and log βi+1 = log( 10
3 )
log βi∗ = log
10
3
βik +
2βi+1
βik +
10
−
3
= O( n1c ) for
2βi+1
10
3
= O( n1c ). Notice that at i = i∗ ,
can be seen by solving the recurrence with
+ k log βi .
(1 + k + k 2 + . . . + k logk log n−1 )
+ k logk logn (− log 11)
log log n
10
k k
−1
= log
− (log n)(log 11)
3
k−1
10
− log 11
≤ (log n) log
3
≤ −1.7 log n
Therefore, as it is, βi∗ ≤
n
, we set
βi at least at 18 log
n
(2)
1
n1.7
≤
18 log n
n
βi+1 = max
when n ≥ 2. In order to keep the value of
10 k 18 log n
β ,
3 i
n
With the values of βi defined, we proceed to bound Pr(vi > βi |vi−1 ≤ βi−1 ), ∀12 ≤
i ≤ i∗ . For a given i,
Pr(vi > βi |vi−1 ≤ βi−1 ) = Pr(nvi > nβi |vi−1 ≤ βi−1 )
≤ Pr(ui > nβi |vi−1 ≤ βi−1 )
We upper bound the above inequality using the following idea. Let Yr be an
indicator variable set to 1 when the following 2 conditions are met: (i) the rth ball
placed is of height at least i and (ii) vi−1 ≤ βi−1 . Yr is set to 0 otherwise. Now
2
k
βi ≤
for all 1 ≤ r ≤ n, the probability that Yr = 1 is upper bounded by βi−1
+ 10
βi
3
2
β
+
β
≤
.
Therefore,
the
probability
that
the
number
of
balls
of
height
at
10 i
10 i
2
βi
least i exceeds βi is upper bounded by Pr(B(n, 2 ) > nβi ), where B(·, ·) is a binomial
random variable with given parameters.
µδ2
Recall the Chernoff bound, for 0 < δ ≤ 1, Pr(X ≥ (1 + δ)µ) ≤ e− 3 , where X is
the sum of independent Poisson trials and µ is the expectation of X. If we set δ = 1,
then we have
10
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
βi
) > nβi )
2
n · ( β2i )
−
3
≤e
Pr(vi > βi |vi−1 ≤ βi−1 ) ≤ Pr(B(n,
n
n · ( 18 log
)
n
18 log n
6
, ∀i ≤ i∗ )
≤e
(since βi ≥
n
1
≤ 3
n
−
Thus we have
∗
i
X
Pr(vj > βj |vj−1 ≤ βj−1 ) ≤
j=12
log log n
n3
∗
(3)
=⇒
i
X
j=l+1
Pr(vj > βj |vj−1 ≤ βj−1 ) ≤
1
(since n ≥ 2)
2n2
Finally, we need to upper bound Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ ). Consider a particular bin
of load at least i∗ . Now the probability that a ball will fall into that bin is
!!
1
1
+ 2βi∗ +1
≤ · βik−1
∗
n
1 − βik∗ +1
1
1
≤ · βik−1
+ 2βi∗ +1
∗
n
1 − β11
22
1
+ βi∗ (since βi is a non-increasing function)
≤ · βik−1
∗
n
10
1 32
≤ ·
· βi∗ (since k ≥ 2 and βi∗ ≤ 1)
n 10
Now, we upper bound the probability that 4 balls fall into a given bin of load at
least i∗ and then use a union bound over all the bins of load at least i∗ to show that
the probability that the fraction of bins of load at least βi∗ +4 exceeds 0 is negligible.
First, the probability that 4 balls fall into a given bin of load at least βi∗ is
1 32
≤ Pr(B(n, ( ·
∗ βi∗ )) ≥ 4)
n 10
4
n
1 32
·
· βi∗
≤
n 10
4
4
1
1 32
·
· βi∗ ·
≤ e·n·
n 10
4
4
32 eβi∗
·
≤
10
4
Taking the union bound across all possible βi∗ n bins, we have the following inequality
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
Pr(vi∗ +4 > 0|vi∗
=⇒ Pr(vi∗ +4 > 0|vi∗
=⇒ Pr(vi∗ +4 > 0|vi∗
(4)
11
4
32 eβi∗
≤ βi∗ ) ≤ (βi∗ n) ·
·
10
4
4
32 18e log n
·
≤ βi∗ ) ≤ (18 log n) ·
10
4n
1
≤ βi∗ ) ≤
(since n ≥ n0 )
2n2
Putting together equations 1, 3, and 4, we get
1
1
+ 2
2n2
2n
1
≤ 2
n
Pr(vi∗ +4 > 0) ≤
Thus
log log n
Pr Max. Load ≥
+ 15 = Pr(vi∗ +4 > 0)
log k
1
≤ 2
n
From Lemma 3.2 and Lemma 3.3, we immediately arrive at Theorem 3.1.
4. Analysis of FirstDiff[d] when m ≫ n.
Theorem 4.1. Use FirstDiff[d], where maximum number of probes allowed per
ball is 2d/2.17 , to allocate m balls into n bins. When m ≥ 72(nλ log n + n) where λ
is taken from Lemma 4.5, n ≥ n0 where n0 is the smallest value of n that satisfies
0.00332n(λ log n + 1)d/2d/2.17 ≥ log n, and d ≥ 6, it takes at most d probes on average
to place every ball on expectation and with high probability. Furthermore, for an
absolute constant c,
Pr Max. load of any bin >
m
n
+
log log n
0.46d
+ c log log log n ≤ c(log log n)−4 .
Proof. First we show that the average number of probes per ball is at most d on
expectation and w.h.p. We then show the maximum load bound holds w.h.p.
4.1. Proof of Number of Probes.
Remark: The earlier version of this paper [1] had a different proof in this subsection.
The overall idea of overcounting the number of probes remains the same but the
specific argument of how to justify and go about such an overcounting is changed and
cleaner now. More specifically, we have replaced Lemmas 4.2, 4.3, and 4.4 with the
argument that follows the header “Overcounting method” and concludes at the header
“Expectation bound”. Also note that Lemma 4.6 from our earlier version is no longer
required due to the way we’ve constructed our argument.
The main difficulty with analyzing the number of probes comes from the fact that
the number of probes needed for each ball depends on where each of the previous balls
were placed. Intuitively, if all the previous balls were placed such that each bin has
the same number of balls, the number of probes will be 2d/2.17 . On the other hand, if a
significant number of bins are at different load levels, then, the ball will be placed with
very few probes. One might hope to prove that the system always displays a variety of
12
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
loads, but unfortunately, the system (as we verified experimentally) oscillates between
being very evenly loaded and otherwise. Therefore, we have to take a slightly more
nuanced approach that takes into account that the number of probes cycles between
high (i.e. as high as 2d/2.17 ) when the loads are even and as low as 2 when there is
more variety in the load.
Lemma 4.2. When m ≥ 72(nλ log n + n) where λ is taken from Lemma 4.5, n ≥
n0 where n0 is the smallest value of n that satisfies 0.00332n(λ log n + 1)d/2d/2.17 ≥
log n, and d ≥ 6, using FirstDiff[d], where maximum number of probes allowed per ball
is 2d/2.17 , takes at most md probes on expectation and with high probability to place
the m balls in n bins.
Proof. Let the maximum number of probes allowed per ball using FirstDiff[d] be
k, i.e. k = 2d/2.17 . Throughout this proof, we will assume that the maximum load
of any bin is at most m/n + λ log n, which holds with high probability owing to
Lemma 4.5. The low probability event that the maximum load exceeds m/n + λ log n
will contribute very little to the overall number of probes because the probability
that any ball exceeds a height of m/n + λ log n will be an arbitrarily small inverse
polynomial in n. Therefore, such a ball will contribute o(1) probes to the overall
number of probes even when we liberally account k probes for each such ball (as long
as k ≪ n). Let m balls be placed into n bins; we assume m ≥ 72(nλ log n + n).
In order to prove the lemma, we proceed in three stages. In the first stage, we
consider an arbitrary sequence of placing m balls into the n bins. We develop a
method that allows us to overcount the number of probes required to place the ball at
each step of this placement. In the second stage, we proceed to calculate the expected
number of probes required to place the balls. Finally in the third stage, we show how
to get a high probability bound on the number of probes required to place each ball.
Overcounting method
First, we couple the process of FirstDiff[d] with a similar process where the zero bin
condition to place balls is not used. In this similar process, if a zero bin is probed first,
more probes are made until either a bin with a different load is probed or until all k
probes are made. Then, the ball is placed in the first bin probed (the zero bin). In
case the first bin probed is not a zero bin, then the process acts exactly as FirstDiff[d].
It is clear that this process will take more probes than FirstDiff[d] while still making
the same placements as FirstDiff[d]. Thus, any upper bounds on number of probes
obtained for this process apply to FirstDiff[d]. For the remainder of this proof, we
analyze this process.
Now we describe the method we use to overcount the number of probes. Consider
an arbitrary sequence of placing m balls into n bins. We will describe a method
to associate each configuration that arises from such a placement with a “canonical”
configuration, which we define later in this proof. Each such canonical configuration
requires more probes to place the ball than the actual configuration. We ensure
that the mapping of actual configurations to canonical configurations is a one-to-one
mapping. Thus, by counting the number of probes required to place a ball in every
possible canonical configuration, we overcount the number of probes required to place
every ball.
Imagine coupling according to probe sequences. That is, consider all [n]k × [n]k ×
. . . × [n]k = [n]km possible sequences of probes. Each timestep corresponds to a
particular length-k sequence of bin labels [n], which direct the bins to be probed.
Note that the probe sequences are each equally likely, and that they fully determine
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
13
the placement.
For each probe sequence S, let X S be the sequence of configurations generated by
S. Let xS be the sequence of numbers-of-probes-used at each timestep of X S . (Note
that S gives k potential probes at each timestep, but the entries of xS are often less
than k, as often the ball is placed without using the maximum number of probes.)
For convenience, we drop the superscript S subsequently and use X and x to refer to
the vectors.
Now we give a few definitions. Consider a particular bin with balls placed in
it, one on top of another. If exactly ℓ − 1 balls are placed below a given ball, we
say that ball is at height ℓ or level ℓ. Balls of the same height are said to be of the
same level. We say a level contains b balls if there are b balls at that level. A level
containing n balls is said to be complete. A level containing at least one ball but less
than n balls is said to be incomplete. For a given configuration, consider the highest
complete level ℓ. For the given configuration, we define a plateau as any level ≥ ℓ
with at least one ball at that level. Intuitively, the plateaus for the configuration are
the highest complete level and any higher incomplete levels. Notice that it is possible
that a given configuration may only have one plateau when there are no incomplete
levels. Further, notice that for a given configuration if two plateaus exist at levels ℓ
and ℓ + 2 with number of balls b1 and b2 , it implies that there exists a plateau at level
ℓ + 1 with number of balls in [b2 , b1 ].
Consider a particular configuration Xi in the sequence X. Call its number of
plateaus p. Consider the plateaus in increasing order of their levels and call them
ℓ1 , ℓ2 , . . . ℓp . Call the number of balls at each level ℓi , bi .
We now define the canonical configuration Cℓ,b . This is the configuration with no
balls at level greater than ℓ, b balls at level ℓ, and n balls at every level less than ℓ.
We will associate each configuration Xi with a set of canonical configurations Ci .
For each plateau of Xi at level ℓj , include the canonical configuration Cℓj ,bj in the
set.
Note that for any probe sequence (not just the specific sequence S), the number of
probes utilized by Xi (e.g. xi for the sequence S) is less than or equal to the number
of probes utilized by any of the configurations in Ci . Therefore the expected number
of probes used to place a ball in configuration Xi is less than or equal to the expected
number of probes used to place a ball in any of the configurations in Ci . We now
describe a way to select one particular configuration out of Ci to associate with each
configuration Xi . We choose this configuration such that the mapping between all
configurations Xi , 1 ≤ i ≤ m, and the selected canonical configurations will be a oneto-one mapping. Furthermore, we show that every selected canonical configurations
will be unique from the others and thus the set of all selected canonical configurations
will not be a multiset. Thus, by counting the number of probes required to place balls
in every possible canonical configuration, we overcount the number of probes required
to place balls in any sequence S.
Look at the canonical configurations associated with configurations Xi , 1 ≤ i ≤ m
over some entire sequence. Let C0 refer to the set of canonical configurations before
any ball is placed and let it be the empty set. The set of canonical configurations Ci−1
differs from Ci in at most three configurations. Consider that i − 1 balls have been
placed and there are now p plateaus with levels ℓi , 1 ≤ i ≤ p and corresponding bi ,
1 ≤ i ≤ p values.
• If the ith ball is placed at level ℓj , level ℓj+1 exists (≥ 1 balls are present at
level ℓj+1 ), and bj+1 6= n − 1, then Ci = (Ci−1 \{Cℓj+1 ,bj+1 }) ∪ {Cℓj+1 ,bj+1 +1 }.
• If level ℓj+1 exists and bj+1 = n − 1, then Ci = (Ci−1 \{Cℓj ,bj , Cℓj+1 ,bj+1 }) ∪
14
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
{Cℓj+1 ,n }.
• If level ℓj+1 does not exist, then Ci = Ci−1 ∪ {Cℓj +1,1 }.
Notice that in every scenario, there is exactly one configuration that is added to
Ci−1 to get Ci . We denote the newly-added configuration as the selected canonical
configuration of Ci .
Now, for a given sequence of configurations Xi , 1 ≤ i ≤ m it is clear that each
selected canonical configuration is uniquely chosen. We show in the following lemma,
Lemma 4.3, that every selected canonical configuration in the sequence is unique and
different from the other selected canonical configurations. In other words, the set of
selected canonical configurations will not be a multiset. Furthermore, it has earlier
been established that for any configuration and one of its canonical configurations, it
takes at least as many probes to place a ball in the latter as it does in the former.
Thus, by calculating the S
number of probes it would take to place balls in all possible
m
canonical configurations i=1 Ci , we can overcount the number of probes required to
place all balls into bins.
Lemma 4.3. The set of selected canonical configurations for any sequence of configurations Xi , 1 ≤ i ≤ m will not be a multiset.
Proof. Consider an arbitrary sequence and within it an arbitrary configuration Xi
for some 1 ≤ i ≤ m. To reach this configuration, a ball was placed previously in some
level ℓ − 1 and extended level ℓ from b balls to b + 1 balls, 0 ≤ b ≤ n − 1. The selected
canonical configuration for Xi will be Cℓ,b+1 . Since balls can only be added and never
deleted, once a level is extended to some b + 1 number of balls, placing another ball
can never extend that same level to b + 1 balls. That level can only be henceforth
extended to a larger number of balls up to n balls. Thus a given configuration Cℓ,b+1
will never appear twice in the set of selected canonical configurations.
Expectation bound
As mentioned earlier, we assume the maximum load of any bin is m/n+ λ log n. Thus,
for any given sequence of placements, the final configuration never has any balls at
level m/n + λ log n + 1. Thus, when calculating the number of probes taken to place
all balls, we need only consider the number of probes required to place a ball in every
canonical configuration Cℓ,b , 0 ≤ ℓ ≤ m/n + λ log n, 0 ≤ b ≤ n − 1.
For a given canonical configuration Cℓ,b let Yℓ,b be a random variable denoting
the number of probes required to place the ball using FirstDiff[d] without the zero
bin condition. Let Y be a random variable denoting the total number of probes
required to place a ball in each of the possible canonical configurations. Thus Y =
Pm/n+λ log n Pn−1
ℓ=1
b=0 Yℓ,b .
For a given configuration Cℓ,b , a ball is placed when either it first hits bins of level
ℓ several times and then a bin of level ℓ − 1, it hits bins of level ℓ − 1 several times and
then a bin of level ℓ, or it makes k probes. Thus, using geometric random variables,
we see that E[Yℓ,b ] = min(b/(n − b) + (n − b)/b, k). For the first nk and last nk − 1
canonical configurations for a given level, let us give away the maximum number of
probes, i.e. Yℓ,b = k for any ℓ and for 0 ≤ b ≤ nk − 1 and n − nk + 1 ≤ b ≤ n − 1.
We now want to calculate the expected number of probes for the middle canonical
configurations.
Therefore
15
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
m
n +λ log n
n− n
k
X
X
b= n
k
ℓ=0
n
X n − b n−
Xk b
E[Yℓ,b ] =
+ λ log n + 1
+
n
b
n
n n − b
b= k
b= k
n
n
n−
n− n
n− n
n−
m
Xk 1
Xk b
Xk 1
Xk n − b
+ λ log n + 1 n
−
+n
−
=
n
n b
n b
n n − b
n n − b
b= k
b= k
b= k
b= k
n
n
n−
k
m
Xk 1
X
4n
1
+ λ log n + 1 n
+n
− 2n +
=
n
b
y
k
b= n
y=n− n
k
k
n
n− k
m
X 1
2
=
+ λ log n + 1 2n
− 1 +
n
b
k
n
b= k
n
2
n
− log
−1+
≈ 2 (m + nλ log n + n) log n −
k
k
k
2
= 2 (m + nλ log n + n) log(k − 1) − 1 +
k
m
n− n
k
High probability bound
Now, we may apply Lemma 1.1 with Λ = 1.01, µ taken from above, and p∗ = k1 .
m
n
Pr
≤e
+λ log n n− n
k
X
ℓ=0
X
b= n
k
1
·2(m+nλ log
−k
≤O
Yℓ,b > 2.02(m + nλ log n + n)(log(k − 1) − 1 +
2
n+n)(log(k−1)−1+ k
)·(1.01−1−ln 1.01)
2
)
k
1
(since k > 2 and n ≥ n0 )
n
Therefore, with high probability, the total number of probes
m
n +λ log n
Y =
X
n−1
X
X
X
ℓ=0
Yℓ,b
b=0
m
n
n +λ log n k −1
≤
ℓ=0
b=0
m
n +λ log
Yℓ,b +
X
ℓ=0
n n− n
k
X
m
n +λ log n
(Yℓ,b ) +
b= n
k
X
ℓ=0
n−1
X
Yℓ,b
b=n− n
k +1
n
n
m
2
+
+ λ log n + 1 k + 2.02 (m + nλ log n + n) log(k − 1) − 1 +
+ λ log n + 1 k
−1
n
k
k
n
k
≤ 2.14(m + nλ log n + n) log k
≤
m
≤ 2.17m log k (since m ≥ 72(nλ log n + n))
Thus when m balls are placed into n bins, an upper bound on both the expected
total probes and the total probes with high probability is 2.17m log k. Therefore on
expectation and with high probability, the number of probes per ball is at most d
since k = 2d/2.17 .
16
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
4.2. Proof of Maximum Load.
Lemma 4.4. Use FirstDiff[d], where maximum number of probes allowed per ball
is 2d/2.17 , to allocate m balls into n bins. For any m, for an absolute constant c,
Pr Max. load of any bin >
m
n
+
log log n
0.46d
+ c log log log n ≤ c(log log n)−4 .
Proof. This proof follows along the lines of that of Theorem 2 from [14]. In order
to prove Lemma 4.4, we make use of a theorem from [10] which gives us an initial,
loose, bound on the gap Gt between the maximum load and average load for an
arbitrary m. We then use a lemma to tighten this gap. We use one final lemma to
show that if this bound on the gap holds after all m balls are placed, then it will hold
at any time prior to that.
First, we establish some notation. Let k be the maximum number of probes
permitted to be made per ball by FirstDiff[d], i.e. k = 2d/2.17 . After placing nt
balls, let us define the load vector X t as representing the difference between the load
of each bin and the average load (as in [3, 10]). Without loss of generality we order the individual values of the vector in non-increasing order of load difference, i.e.
X1t ≥ X2t ≥ . . . ≥ Xnt . So Xit is the load in the ith most loaded bin minus t. For
convenience, denote X1t (i.e. the gap between the heaviest load and the average) as Gt .
Initial bound on gap
We now give an upper bound for the gap between the maximum loaded bin and the
average load after placing some arbitrary number of balls nt. In other words, we show
Pr(Gt ≥ x) is negligible for some x. This x will be our initial bound on the gap Gt .
Lemma 4.5. For arbitrary constant c, after placing an arbitrary nt balls into bins
n
bn
under FirstDiff[d], there exist some constants a and b such that Pr(Gt ≥ c log
a ) ≤ nc .
1
t
Thus there exists a constant λ that gives Pr(G ≥ λ log n) ≤ nc for a desired c value.
In order to prove Lemma 4.5, we need two additional facts. The first is the following
basic observation:
Lemma 4.6. FirstDiff[d] is majorized by Greedy[2] when d ≥ 2.
Proof. Let the load vectors for FirstDiff[d] and Greedy[2] after t balls have been
placed using the respective algorithms be ut and v t respectively. Now we follow
the standard coupling argument (refer to Section 5 in [3] for an example). Couple
FirstDiff[d] with Greedy[2] by letting the bins probed by Greedy[2] be the first 2 bins
probed by FirstDiff[d]. We know that FirstDiff[d] makes at least 2 probes when d ≥ 2.
It is clear that FirstDiff[d] will always place a ball in a bin with load less than or equal
to that of the bin chosen by Greedy[2]. This ensures that if majorization was preserved
prior to the placement of the ball, then the new load vectors will continue to preserve
majorization; again, see [3] for a detailed example. Initially, u0 is majorized by v 0
since both vectors are the same. Using induction, it can be seen that if ut is majorized
by v t at the time the tth ball was placed, it would continue to be majorized at time
t + 1, 0 ≤ t ≤ m − 1. Therefore, FirstDiff[d] is majorized by Greedy[2] when d ≥ 2.
The other fact is the following theorem about Greedy[d] taken from [10] (used
similarly in [14] as Theorem 3).
Theorem 4.7. [10] Let Y t be the load vector generated by Greedy[d]. Then for
every d > 1 there exist positive constants
a and
b such that for all n and all t,
P a|Y t |
E
e i
≤ bn.
i
We are now ready to prove Lemma 4.5.
17
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
Proof of Lemma 4.5.. Combining Lemma 4.6 with Theorem 4.7 tells us that, if
X t is the load vector generated by FirstDiff[d],
!
X
a|Xit |
≤ bn.
e
E
i
Clearly, Pr(Gt ≥
c log n
a )
Pr(Gt ≥
t
= Pr(eaG ≥ nc ). Observe that
P
t
t
ea|Xi | ≥ eaG . Then
i
t
c log n
) = Pr(eaG ≥ nc )
a
t
E[eaG ]
≤
(by Markov’s inequality)
nc
bn
≤ c (by Theorem 4.7 and Lemma 4.6)
n
and the theorem is proved.
Reducing the gap
Lemma 4.5 gives an initial bound on Gt of order log n. The next step is to reduce it
to our desired gap value. For this reduction, we use a modified version of Lemma 2
from [14], with a similar but more involved proof. We now give the modified lemma
and prove it.
Lemma 4.8. For every k, there exists a universal constant γ such that the fol1
lowing holds: for any t, ℓ, L such that 1 ≤ ℓ ≤ L ≤ n 4 , L = Ω(log log n) and
1
t
Pr(G ≥ L) ≤ 2 ,
3
+ n12 ,
Pr(Gt+L ≥ logloglogk n + ℓ + γ) ≤ Pr(Gt ≥ L) + 16bL
eaℓ
where a and b are the constants from Theorem 4.7.
Proof. This proof consists of many steps. We first observe that Lemma 4.8 follows directly from Lemma 4.5 for sufficiently small n. We then use layered induction
to bound the proportion of bins of each size for larger n. This in turn allows us to
compute our desired bound on the probability of a large gap occurring.
Proof of
Define n1
Define n2
Lemma 4.8 for smaller values of n
to be the minimum value of n such that L ≥ 2 (recall L = Ω(log log n)).
tobe the minimum
value of n such that
1
(18 log n) ∗
4
18en 4 log n
n
≤
1
2n2 .
Define n3 to be the minimum value of n such that
n ≥ 54 log n. Define absolute constant n0 = max(n1 , n2 , n3 ).
Notice that, when n ≤ n0 , Lemma 4.5 implies that Lemma 4.8 holds with γ =
O(log n0 ). If n ≤ n0 , then
Pr(Gt+ℓ ≥ log log n + ℓ + γ) ≤ Pr(Gt+L ≥ γ).
Consider the right hand side of Lemma 4.8.
Pr(Gt ≥ L) +
1
16bL3
+ 2 ≥ n−2 ,
exp(aℓ) n
so it will be sufficient to prove the inequality
Pr(Gt+L ≥ γ) ≤ n−2 .
18
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Since there are no conditions on t in Lemma 4.5, we may rewrite it as
Pr(Gt+L ≥ λ log n) ≤ n−c .
Let c = 2 and compute the constant λ accordingly. Set γ = λ log n0 ≥ λ log n. Then
n−2 ≥ Pr(Gt+L ≥ λ log n) ≥ Pr(Gt+L ≥ γ),
and we are done.
Rewriting initial probability inequality
We now prove Lemma 4.8 assuming n > n0 . Start by rewriting the probability in
terms of Pr(Gt ≥ L).
Pr(Gt+L ≥
log log n
log log n
+ ℓ + γ) = Pr(Gt+L ≥
+ ℓ + γ|Gt ≥ L)Pr(Gt ≥ L)
log k
log k
log log n
+ Pr(Gt+L ≥
+ ℓ + γ|Gt < L)Pr(Gt < L)
log k
log log n
≤ Pr(Gt ≥ L) + Pr(Gt+L ≥
+ ℓ + γ|Gt < L)
log k
To prove the theorem, then, it is enough to to show that
3
1
Pr(Gt+L ≥ logloglogk n + ℓ + γ|Gt < L) ≤ 16bL
eaℓ + n2 .
Bins’ loads
Define vi to be the fraction of bins of load at least t + L + i after (t + L)n balls are
placed. Let us set i∗ = logloglogk n + ℓ and set γ = 4. Using this new notation, we want to
show that Pr(Gt+L ≥ i∗ + 4|Gt < L) is negligible. This can be thought of as showing
that the probability of the fraction of bins of load at least t + L + i∗ + 4 exceeding 0
after (t + L)n balls are placed, conditioned on the event that Gt < L, is negligible.
Suppose we have a non-increasing series of numbers β0 , β1 , . . . , βi , . . . that are
upper bounds for v0 , v1 , . . . , vi , . . .. Then we know that
Pr(vi∗ +4 > 0) = Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ )Pr(vi∗ ≤ βi∗ ) + Pr(vi∗ +4 > 0|vi∗ > βi∗ )Pr(vi∗ > βi∗ )
≤ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ ) + Pr(vi∗ > βi∗ )
∗
i
X
≤ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ ) +
Pr(vj > βj |vj−1 ≤ βj−1 ) + Pr(vℓ > βℓ ) (successively expanding
j=ℓ+1
and bounding the Pr(vi∗ > βi∗ ) term and its derivatives)
Conditioning both sides on Gt < L, we have
Pr(vi∗ +4 > 0|Gt < L) ≤ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ , Gt < L)
∗
+
i
X
Pr(vi > βi |vi−1 ≤ βi−1 , Gt < L)
i=ℓ+1
(5)
+ Pr(vℓ > βℓ |Gt < L)
It remains to find appropriate βi values. We use a layered induction approach to show
that vi ’s don’t exceed the corresponding βi ’s with high probability. This then allows
19
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
us to upper bound each of the 3 components of equation 5.
Base case of layered induction
In order to use layered induction, we need a base case. Let us set βℓ =
in the statement of the theorem. Now,
1
8L3 ,
for the ℓ
T
Pr((vℓ > 8L1 3 ) (Gt < L))
Pr(Gt < L)
1
1
≤ 2 · Pr(vℓ >
) (since, by the statement of the theorem Pr(Gt < L) ≥ )
3
8L
2
8bL3
≤ 2 · aℓ (applying Markov’s inequality and using Theorem 4.7)
e
16bL3
≤ aℓ
e
Pr(vℓ > βℓ |Gt < L) =
Therefore we have the third term of Equation 5 bounded:
(6)
Pr(vℓ > βℓ |Gt < L) ≤
16bL3
eaℓ
Recurrence relation for layered induction
We now define the remaining βi values recursively. Note that for all i ≥ ℓ, βi ≤ βℓ .
Let ui be defined as the number of balls of height at least t + L + i after (L + t)n balls
are placed.
Initially there were nt balls in the system. Then we threw another nL balls into
the system. Remember that t+L is the average load of a bin after nL balls are further
placed. Because we condition on Gt < L, we have it that any ball of height i, i ≥ 1,
must have been one of the nL balls placed.
Therefore the number of bins of load t + L + i + 1 after (t + L)n balls are placed is
upper bounded by the number of balls of height at least t+ L + i + 1. So vi+1 n ≤ ui+1 .
In order to upper bound vi+1 , we can upper bound ui+1 .
Recall the algorithm places a ball in a bin of load t + L + i if it probes k times
and sees a bin of load t + L + i each time; or if it probes j < k times and sees a bin
of load t + L + i each time, then probes a bin of load ≥ t + L + i + 1; or if it probes
j < k times and sees a bin of load at least t + L + i + 1 each time (where the load of
the bin probed each time is the same), then probes a bin of load t + L + i. Thus the
probability that a ball will end up at height at least t + L + i + 1 is
k−2
2
≤ βik + βi βi+1 1 + βi + βi2 + . . . + βik−2 + βi βi+1 1 + βi+1 + βi+1
+ . . . + βi+1
!
k−1
1 − βi+1
1 − βik−1
k
≤ βi + βi βi+1
+
1 − βi
1 − βi+1
1
≤ βik + βl βi+1 2 ∗
1 − βl
2βi+1
≤ βik +
8L3 − 1
Let vi+1 (f ) be the fraction of bins with load at least t + i + 1 after the tn + f th ,
1 ≤ f ≤ nL, ball is placed in a bin.
20
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Let f ∗ = min[arg minf vi+1 (f ) > βi+1 , nL], i.e. f ∗ is the first f such that
vi+1 (f ) > βi+1 or nL if there is no such f . By our preceding argument, the probability that f ∗ < nL is bounded by the probability that a binomial random variable
2βi+1
B(nL, ≤ βik + 8L
3 −1 ) is greater than βi+1 nL.
Fix
2βi+1
2nL(βik + 8L
3 −1 )
8L3 − 1
k
βi+1 = 2L 3
β ≥
.
8L − 4L − 1 i
n
Then using a Chernoff bound,
we can
say that with high probability, f ∗ = nL or
vi+1 ≤ βi+1 , so long as e−
βik +
2βi+1
8L3 −1
3
= O( n1c ) for some constant c ≥ 1.
2βi+1
βik +
8L3 −1
−
3
n
Now, so long as βi+1 ≥ 18 log
, e
= O( n1c ). In other words, this
n
upper bound holds for the placement of all nt + nL balls.
We now show that according to the previous recurrence relation, βi∗ dips below
We later propose a modified recurrence relation which sets the value of βi to
n
. This ensures
the maximum of the value of obtained from the recurrence and 18 log
n
18 log n
. This upper bound will be used later in the argument. We have,
that βi∗ =
n
from the value of βℓ and the above discussion,
log βℓ = −3 log(2L) and
3
−1
log βi+1 = k log βi + log(2L) + log( 8L8L
3 −4L−1 )
Solving the recursion for log βℓ+log log n , we get
2L(8L3 − 1)
k log log n − 1
− 3k log log n log(2L)
log
log βℓ+log log n =
k−1
8L3 − 4L − 1
8L3 − 1
≤ k log log n (−3k + 4) log(2L) + log
8L3 − 4L − 1
8L3 − 1
≤ k log log n (−6 + 4) log(2L) + log
8L3 − 4L − 1
18 log n
.
n
≤ k log log n ((−1.5) log(2L)) (when L ≥ 2)
≤ 2log log n ((−1.5) log(2L))
≤ (−1.5)(log n)
n
,
Therefore, βi∗ ≤ n−1.5 , when L ≥ 2. Since n ≥ n1 , we have L ≥ 2. Thus βi∗ < 18 log
n
as desired.
Now, we need to bound Pr(vi > βi |vi−1 ≤ βi−1 , Gt < L) for all i’s from ℓ + 1 to
3
−1
k 18 log n
i∗ . Let us set βi+1 = max(2L 8L8L
).
3 −4L−1 βi ,
n
Using the values of βi generated above, we prove that for all i such that ℓ + 1 ≤
i ≤ i∗ , Pr(vi > βi |vi−1 ≤ βi−1 , Gt < L) ≤ n13 .
For a given i,
Pr(vi > βi |vi−1 ≤ βi−1 , Gt < L) = Pr(nvi > nβi |vi−1 ≤ βi−1 , Gt < L)
≤ Pr(ui > nβi |vi−1 ≤ βi−1 , Gt < L)
We now upper bound the above inequality using the following idea. Let Yr be an
indicator variable set to 1 when all three of the following conditions are met: (i) the
21
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
nt + rth ball placed is of height at least t + L + i, (ii) vi−1 ≤ βi−1 and (iii) Gt < L.
Yr is set to 0 otherwise. Now for all 1 ≤ r ≤ nL, the probability that Yr = 1 is upper
2βi+1
βi
8L3 −1
k
t
bounded by βik + 8L
3 −1 ≤ 8L3 −4L−1 βi−1 ≤ 2L . Since we condition on G < L, the
number of balls of height at least t + L or more come only from the nL balls placed.
Therefore, the probability that the number of balls of height at least n + L + i exceeds
βi
βi is upper bounded by Pr(B(nL, 2L
) > βi ), where B(., .) is a binomial random
variable with given parameters.
µδ2
According to Chernoff’s bound, for 0 < δ ≤ 1, Pr(X ≥ (1 + δ)µ) ≤ e− 3 , where
X is the sum of independent Poisson trials and µ is the expectation of X. If we set
δ = 1, then we have
Pr(vi > βi |vi−1 ≤ βi−1 , Gt < L)
βi
≤ Pr(B(nL,
) > βi )
2L
n ∗ ( β2i )
−
3
≤e
n
n ∗ ( 18 log
)
n
18 log n
6
≤e
(since βi ≥
, ∀i ≤ i∗ )
n
1
≤ 3
n
−
Thus we bound the middle term in Equation 5
∗
i
X
Pr(vj > βj |vj−1 ≤ βj−1 , Gt < L) ≤
log log n
n3
Pr(vj > βj |vj−1 ≤ βj−1 , Gt < L) ≤
1
(since n ≥ n1 )
2n2
j=ℓ+1
∗
(7)
=⇒
i
X
j=ℓ+1
Top layers of layered induction
Finally, we need to upper bound the first term in Equation 5,
Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ , Gt < L). Consider a bin of load at least i∗ . We will upper
bound the probability that a ball falls into this specific bin. Regardless of how the
probes are made for that ball, one of them must be made to that specific bin. Thus
we have a formula similar to our original recursion, but with a factor of 1/n.
22
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Therefore the probability that a ball will fall into that bin is
≤
≤
≤
≤
≤
≤
1
1 k−1 1
+ βi∗ +1 1 + βi∗ +1 + βi2∗ +1 + . . . + βik−2
βi
+ βi∗ +1 1 + βi∗ + βi∗ + . . . + βik−2
∗
∗ +1
n
n
n
1
1
k−1
· βi∗ + 2βi∗ +1
n
1 − βi∗ +1
1
1
· βik−1
+ 2βi∗
∗
n
1 − βi∗
1
2n
· βik−1
βi∗
+
∗
n
n − 18 log n
1 3n − 18 log n
·
· βi∗ (since k ≥ 2 and βi∗ ≤ 1)
n n − 18 log n
4
· βi∗ (since n > n3 )
n
Now, we upper bound the probability that 4 balls fall into a given bin of load
at least βi∗ and then use a union bound over all the bins of height at least βi∗ to
show that the probability that the fraction of bins of load at least βi∗ +4 exceeds 0 is
negligible.
First, the probability that 4 balls fall into a given bin of load at least βi∗ is
4
≤ Pr(B(nL, ( · βi∗ )) ≥ 4)
n
4
nL
4
· βi∗
≤
n
4
4
4
1
≤ e · nL · ( · βi∗ ) ·
n
4
4
≤ (eLβi∗ )
Taking the union bound across all possible βi∗ n bins, we have the following inequality
4
Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ , Gt < L) ≤ (βi∗ n) · (eLβi∗ )
4
18eL log n
t
∗
∗
∗
=⇒ Pr(vi +4 > 0|vi ≤ βi , G < L) ≤ (18 log n) ·
n
1
(since n ≥ n2 )
=⇒ Pr(vi∗ +4 > 0|vi∗ ≤ βi∗ , Gt < L) ≤
2n2
(8)
Putting together equations 5, 6, 7, and 8, we get
16bL3
1
1
+ 2+ 2
eaℓ
2n
2n
1
16bL3
+ 2
≤
aℓ
e
n
Pr(vi∗ +4 > 0|Gt < L) ≤
Thus
23
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
Pr(Gt+L ≥
log log n
+ ℓ + 4|Gt < L) = Pr(vi∗ +4 > 0|Gt < L)
log k
1
16bL3
+ 2
≤
eaℓ
n
Finally
Pr(Gt+L ≥
log log n
log log n
+ ℓ + 4) ≤ Pr(Gt ≥ L) + Pr(Gt+L ≥
+ ℓ + 4|Gt < L)
log k
log k
1
16bL3
≤ Pr(Gt ≥ L) + aℓ + 2
e
n
Hence Lemma 4.8 is proved.
By Lemma 4.5, we know that at some arbitrary time t, the gap will be O(log n)
with high probability. Now applying Lemma 4.8 once with L = O(log n) and ℓ =
O(log log n) with appropriately chosen constants, we get Pr(Gt+L ≥ logloglogk n +O(log log n)+
γ) ≤ O((log log n)−4 ). Applying the lemma again with L = O(log log n) and ℓ =
O(log log log n) with appropriately chosen constants, we get
c
Pr(Gt > logloglogk n + c log log log n) ≤ (log log
n)4 when time t = ω(log n).
We now show that as more balls are placed, the probability that the gap exceeds
a particular value increases. This is true by Lemma 4 from [14]:
′
′
Lemma 4.9. [14] For t ≥ t′ , Gt is stochastically dominated by Gt . Thus E[Gt ] ≤
′
E[Gt ] and for every z, Pr(Gt ≥ z) ≤ Pr(Gt ≥ z).
Although the setting is different in [14], their proof of Lemma 4.9 applies here
as well. Thus knowing the gap is large when time t = ω(log n) with probability
O((log log n)−4 ), implies that for all values of t′ < t, the gap exceeds the desired value
with at most the same probability. Substituting k = 2d/2.17 in Pr(Gt > logloglogk n +
c
c log log log n) ≤ (log log
n)4 and modifying the inequality to talk about max. load after
m balls have been thrown results in the lemma statement.
Thus concludes the proof of Lemma 4.4.
Putting together Lemma 4.2 and Lemma 4.4, we get Theorem 4.1.
5. Lower Bound on Maximum Bin Load. We now provide a lower bound
to the maximum load of any bin after using FirstDiff[d] as well as other types of algorithms which use a variable number of probes for Class 1 type algorithms as defined
by Vöcking [15]. Class 1 algorithms are those where for each ball, the locations are
chosen uniformly and independently at random from the bins available. We first give
a general theorem for this type of algorithm and then apply it to FirstDiff[d].
Theorem 5.1. Let Alg[k] be any algorithm that places m balls into n bins, where
m ≥ n, sequentially one by one and satisfies the following conditions:
1. At most k probes are used to place each ball.
2. For each ball, each probe is made uniformly at random to one of the n bins.
3. For each ball, each probe is independent of every other probe.
The maximum load of any bin after placing all m balls using Alg[k] is at least m
n +
ln ln n
−
Θ(1)
with
high
probability.
ln k
24
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
Proof. We show that Greedy[k] is majorized by Alg[k], i.e. Greedy[k] always performs better than Alg[k] in terms of load balancing. Thus any lower bound that
applies to the max. load of any bin after using Greedy[k] must also apply to Alg[k].
Let the load vectors for Greedy[k] and Alg[k] after t balls have been placed using
the respective algorithms be ut and v t respectively. We use induction on the number
of balls placed to prove our claim of majorization. Initially, no ball is placed and by
default u0 is majorized by v 0 . Assume that ut−1 is majorized by v t−1 . We now use the
standard coupling argument to prove the induction hypothesis. For the placement of
the tth ball, let Alg[k] use wt probes. Couple Greedy[k] with Alg[k] by letting the first
wt bins probed by Greedy[k] be the same bins probed by Alg[k]. Greedy[k] will always
make at least wt probes and thus possibly makes probes to lesser loaded bins than
those probed by Alg[k]. Since Greedy[k] places a ball into the least loaded bin it finds,
it will place a ball into a bin with load at most the same as the one chosen by Alg[k].
Therefore ut is majorized by v t . Thus by induction, we see that ut is majorized by v t
for all 0 ≤ t ≤ m. Therefore Greedy[k] is majorized by Alg[k].
It is known that the max. load of any bin after the placement of m balls into n
ln ln n
bins (m ≥ n) using Greedy[k] is at least m
n + ln k − Θ(1) with high probability [3].
Therefore, the same lower bound also applies to Alg[k].
Now we are ready to prove our lower bound on the max. load of any bin after
using FirstDiff[d].
Theorem 5.2. The maximum load of any bin after placing m balls into n bins
using FirstDiff[d], where maximum number of probes allowed per ball is 2Θ(d) , is at
ln ln n
least m
n + Θ(d) − Θ(1) with high probability.
Proof. We see that FirstDiff[d] uses at most 2Θ(d) probes and satisfies the requirements of Theorem 5.1. Thus by substituting k = 2Θ(d) , we get the desired bound.
Table 1
Experimental results for the maximum load for n balls and n bins based on 100 experiments for
each configuration. Note that the maximum number of probes per ball in FirstDiff[d], denoted as k,
is chosen such that the average number of probes per ball is fewer than d.
n
28
Greedy[d]
2...11%
3...87%
4... 2%
d = 2, k = 3
Left[d]
FirstDiff[d]
2...43%
2...81%
3...57%
3...19%
Greedy[d]
2...88%
3...12%
d = 3, k = 10
Left[d]
FirstDiff[d]
2...100%
2...100%
Greedy[d]
2...100%
d = 4, k = 30
Left[d]
FirstDiff[d]
2...100%
2...100%
2...10%
3...90%
2...12%
3...88%
2...96%
3... 4%
2...100%
2...93%
3... 7%
2...100%
2...100%
2...100%
3...100
2...31%
3...69%
2...100%
3...100%
2...49%
3...51%
2...100%
3...98%
4... 2%
2...100%
2...100%
3...100%
3...100%
3...100%
4...100%
3...96%
4... 4%
2...100%
2...100%
3...100%
3...100%
3...100%
4...100%
3...37%
4...63%
212
3...99%
4... 1%
3...100%
216
3...63%
4...37%
2...100%
220
3...100%
2...100%
224
3...100%
6. Experimental Results. We experimentally compare the performance of
FirstDiff[d] with Left[d] and Greedy[d] in Table 1. Similar to the experimental results in [15], we perform all 3 algorithms in different configurations of bins and d
BALANCED ALLOCATION: PATIENCE IS NOT A VIRTUE
25
values. Let k be the maximum number of probes allowed to be used by FirstDiff[d]
per ball. For each value of d ∈ [2, 4], we choose a corresponding value of k such
that the average number of probes required by each ball in FirstDiff[d] is at most d.
For each configuration, we run each algorithm 100 times and note the percentage of
times the maximum loaded bin had a particular value. It is of interest to note that
FirstDiff[d], despite using on average less than d probes per ball, appears to perform
better than both Greedy[d] and FirstDiff[d] in terms of maximum load.
7. Conclusions and Future Work. In this paper, we have introduced a novel
algorithm called FirstDiff[d] for the well-studied load balancing problem. This algorithm combines the benefits of two prominent algorithms, namely, Greedy[d] and
Left[d]. FirstDiff[d] generates a maximum load comparable to that of Left[d], while
being as fully decentralized as Greedy[d]. From another perspective, we observe that
FirstDiff[log d] and Greedy[d] result in a comparable maximum load, while the number
of probes used by FirstDiff[log d] is exponentially smaller than that of Greedy[d]. In
other words, we exhibit an algorithm that performs as well as an optimal algorithm,
with significantly less computational requirements. We believe that our work has
opened up a new family of algorithms that could prove to be quite useful in a variety
of contexts spanning both theory and practice.
A number of questions arise out of our work. From a theoretical perspective, we
are interested in developing a finer-grained analysis of the number of probes; experimental results suggest the number of probes used to place the ith ball depends on
the congruence class of i modulo n. From an applied perspective, we are interested in
understanding how FirstDiff[d] would play out in real world load balancing scenarios
like cloud computing, where the environment (i.e. the servers, their interconnections,
etc.) and the workload (jobs, applications, users, etc.) are likely to be a lot more
heterogeneous and dynamic.
Acknowledgements. We are thankful to Anant Nag for useful discussions and
developing a balls-in-bins library [9] that was helpful for our experiments. We are
also grateful to Thomas Sauerwald for his helpful thoughts when he visited Institute
for Computational and Experimental Research in Mathematics (ICERM) at Brown
University. Finally, John Augustine and Amanda Redlich are thankful to ICERM for
having hosted them as part of a semester long program.
REFERENCES
[1] J. Augustine, W. K. Moses Jr., A. Redlich, and E. Upfal, Balanced allocation: patience
is not a virtue, in Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on
Discrete Algorithms, Society for Industrial and Applied Mathematics, 2016, pp. 655–671.
[2] Y. Azar, A. Z. Broder, A. R. Karlin, and E. Upfal, Balanced allocations, SIAM journal
on computing, 29 (1999), pp. 180–200.
[3] P. Berenbrink, A. Czumaj, A. Steger, and B. Vöcking, Balanced allocations: The heavily
loaded case, SIAM Journal on Computing, 35 (2006), pp. 1350–1385.
[4] P. Berenbrink, K. Khodamoradi, T. Sauerwald, and A. Stauffer, Balls-into-bins with
nearly optimal load distribution, in Proceedings of the 25th ACM symposium on Parallelism
in algorithms and architectures, ACM, 2013, pp. 326–335.
[5] A. Czumaj and V. Stemann, Randomized allocation processes, Random Structures & Algorithms, 18 (2001), pp. 297–331.
[6] S. Fu, C.-Z. Xu, and H. Shen, Randomized load balancing strategies with churn resilience in
peer-to-peer networks, Journal of Network and Computer Applications, 34 (2011), pp. 252–
261.
[7] S. Janson, Tail bounds for sums of geometric and exponential variables, Technical report,
(2014).
26
J. AUGUSTINE, W. K. MOSES JR., A. REDLICH, AND E. UPFAL
[8] M. Mitzenmacher and E. Upfal, Probability and computing: Randomized algorithms and
probabilistic analysis, Cambridge University Press, 2005.
[9] A. Nag, Problems in Balls and Bins Model, master’s thesis, Indian Institute of Technology
Madras, India, 2014.
[10] Y. Peres, K. Talwar, and U. Wieder, The (1+ β)-choice process and weighted balls-into-bins,
in Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms,
Society for Industrial and Applied Mathematics, 2010, pp. 1613–1619.
[11] M. Raab and A. Steger, “balls into bins”–a simple and tight analysis, in Randomization and
Approximation Techniques in Computer Science, Springer, 1998, pp. 159–170.
[12] H. Shen and C.-Z. Xu, Locality-aware and churn-resilient load-balancing algorithms in structured peer-to-peer networks, Parallel and Distributed Systems, IEEE Transactions on, 18
(2007), pp. 849–862.
[13] X.-J. Shen, L. Liu, Z.-J. Zha, P.-Y. Gu, Z.-Q. Jiang, J.-M. Chen, and J. Panneerselvam,
Achieving dynamic load balancing through mobile agents in small world p2p networks,
Computer Networks, 75 (2014), pp. 134–148.
[14] K. Talwar and U. Wieder, Balanced allocations: A simple proof for the heavily loaded case,
in Automata, Languages, and Programming, Springer, 2014, pp. 979–990.
[15] B. Vöcking, How asymmetry helps load balancing, Journal of the ACM (JACM), 50 (2003),
pp. 568–589.
| 8cs.DS
|
arXiv:1611.08266v1 [cs.CE] 24 Nov 2016
分类号
密级
UDC
编号
中国科学院研究生院
博士学位论文
三维并行自适应有限元动态负载平衡及 hp 自适应策略研究
刘 辉
指导教师
张林波 研究员、博士
中国科学院数学与系统科学研究院
申请学位级别
论文提交日期
培养单位
学位授予单位
博士
学科专业名称
计算数学
论文答辩日期
中国科学院数学与系统科学研究院
中国科学院研究生院
答辩委员会主席
摘
要
本文工作围绕着 PHG 平台展开. PHG 是“科学与工程计算国家重点实验
室”正在发展的一个并行自适应有限元软件平台. PHG 适合分布式存储并行计
算机, 其设计目标是支持使用 h, p 及 hp 自适应有限元方法和有限体积方法求解
二维和三维偏微分方程组. PHG 支持多种网格及曲面单元, 支持多种网格输入格
式, 提供常用的基函数, 支持多种线性方程组解法器和预条件子, 提供接口以调用
外部线性方程组和特征值问题求解软件包, 同时提供前处理及后处理功能等. 本
文的工作主要涉及两方面的内容. 第一方面内容是研究、实现 PHG 中的网格划
分算法及动态负载平衡模块, 包括研究哈密尔顿路的构造和希尔伯特序的编码解
码算法, 以及基于这些算法的网格划分和动态负载平衡方法. 第二方面内容是基
于 PHG 平台研究 hp 自适应有限元算法中的自适应策略并开展数值实验. 具体
内容如下：
(1) 针对 PHG 所处理的协调四面体网格, 本文在网格的对偶图连通的条件下证
明了过顶点哈密尔顿圈和任意两个单元之间的过顶点哈密尔顿路的存在性,
并在此基础上设计了线性复杂度的四面体网格上过顶点哈密尔顿圈和过顶点
哈密尔顿路的构造算法. 文中还对相应并行算法的构造进行了探讨.
(2) 设计了任意维空间中具有线性复杂度的希尔伯特序编码解码算法并提出了希
尔伯特空间填充曲线的一种变体. 本文同时对编码解码算法进行了改进，设
计了复杂度更低的算法, 降低了计算量. 文中给出的希尔伯特空间填充曲线
的变体保证曲线的编码顺序不随曲线阶数的改变而变化.
(3) 设计实现了加密树及空间填充曲线划分方法并设计了 PHG 的动态负载平衡
模块, 同时实现并改进了 Zoltan 中的一个一维划分算法用于对排序后的单元
进行一维划分. 此外, 本文还实现了 PLUM 中设计的子网格 – 进程映射算法,
用于减少动态负载平衡中的数据迁移量. 数值实验表明, 本文在 PHG 中所实
现的空间填充曲线划分算法不论在划分时间还是划分质量上均优于其它一些
软件包 (其中包括由美国 Sandia 国家实验室开发的著名的网格划分和数据迁
移软件包 Zoltan) 中所提供的相应算法. 本文所实现的基于希尔伯特序的网
格划分算法已经成为 PHG 中默认的网格划分方法, 经受了一些并行规模达
数千子网格、计算规模超过十亿单元的大型应用的考验, 表现出很高的效率
和很强的稳健性.
(4) 设计了一个基于误差下降预测的 hp 自适应策略. 当前, hp 自适应有限元算法
i
ii
学位论文
研究中的一个难点是自适应策略的设计, 目前尚没有成熟的、已被广泛接受
的算法. 本文对 hp 自适应策略进行了研究，在前人提出的几种基于误差下
降预测的 hp 自适应策略的基础上给出了一个新的 hp 自适应加密策略, 该策
略适用于二维三角形、四边形和三维四面体、六面体等不同类型的单元, 适
用于正则加密、二分加密等不同自适应加密方式. 数值实验表明, 该策略可
以达到最优的误差指数下降, 并在数值解的精度和计算效率上优于文献中的
一些策略. 这部分工作的另外一个目的是验证 PHG 的 hp 自适应模块的正确
性和适用性.
关键词: PHG, 自适应有限元, 并行计算, 哈密尔顿路, 空间填充曲线, 动态
负载平衡, hp 自适应策略
Researches on Dynamic Load Balancing Algorithms and
hp Adaptivity in 3-D Parallel Adaptive Finite Element
Computations
Liu Hui (majored in computational mathematics)
Directed by Prof. Zhang Lin-bo
This work is related to PHG (Parallel Hierarchical Grid). PHG is a toolbox for
developing parallel adaptive finite element programs, which is under active development
at the State Key Laboratory of Scientific and Engineering Computing. PHG is designed
for distributed memory parallel computers, and its purpose is to support development
of parallel algorithms and codes for solving real world application problems using h, p,
or hp adaptive finite element methods. PHG supports importing meshes from several
mesh file formats. PHG provides solvers and preconditioners, as well as interfaces to
many external packages, for solving linear systems of equations and eigenvalue problems
resulted from adaptive finite element discretization.
This work is divided into two parts. The first part consists of design and realization
of the dynamic load balancing module of PHG, including studies on mesh partitioning
and data migration algorithms. The second part studies hp adaptive strategies for finite
element computations.
The main results of this work are as follows.
(1) For the tetrahedral meshes used in PHG, under reasonable assumptions, we proved
the existence of through-vertex Hamiltonian paths between arbitrary two vertices,
as well as the existence of through-vertex Hamiltonian cycles, and designed an
efficient algorithm with linear complexity for constructing through-vertex Hamiltonian paths. The resulting algorithm has been implemented in PHG, and is used
for ordering elements in the coarsest mesh for the refinement tree mesh partitioning
algorithm.
(2) We designed encoding and decoding algorithms for high dimensional Hilbert order. Hilbert order has good locality, and it has wide applications in various fields
in computer science, such as memory management, database, and dynamic load
balancing. We analysed existing algorithms for computing 2D and 3D Hilbert order, and designed improved algorithms for computing Hilbert order in arbitrary
iii
学位论文
iv
space dimensions. We also proposed an alternate form of Hilbert space filling curve
which has the advantage of preserving the ordering between different levels. The
algorithms have been implemented in PHG and are used in mesh partitioning.
(3) We implemented refinement tree and space filling curve based mesh partitioning
algorithms in PHG, and designed the dynamic load balancing module of PHG. The
refinement tree based partitioning algorithm was originally proposed by Mitchell,
the one implemented in PHG was improved in several aspects. The space filling
curve based mesh partitioning function in PHG can use either Hilbert or Morton
space filling curve. We also implemented a submesh to process mapping algorithm
from the PLUM package in PHG, and use it to reduce the amount of data migration
during mesh redistribution. Numerical experiments show that the dynamic load
balancing functions work well on thousands of processes with more than one billion
elements, and the space filling curve based mesh partitioning module of PHG is
faster and yields better results than the corresponding function in other well known
mesh partitioning packages.
(4) We studied existing hp adaptive strategies in the literature, and proposed a new
strategy. Numerical experiments show that our new strategy achieves exponential
convergence, and is superior, in both precision of the solutions and computation
time, to the strategy compared. This part of the work also serves to validate the
hp adaptivity module of PHG.
Keywords: PHG, adaptive finite element method, parallel computing, Hamiltonian path, Hilbert order, dynamic load balancing, hp adaptive strategy
目
摘
录
要
第一章
§1.1
§1.2
§1.3
§1.4
§1.5
第二章
§2.1
§2.2
§2.3
§2.4
§2.5
第三章
§3.1
§3.2
§3.3
§3.4
§3.5
§3.6
第四章
§4.1
i
引言
自适应有限元方法 . . . . . . . . .
动态负载平衡 . . . . . . . . . . . .
并行自适应有限元软件平台 PHG
论文主要工作 . . . . . . . . . . . .
论文内容安排 . . . . . . . . . . . .
哈密尔顿路和哈密尔顿圈
协调四面体网格 . . . . . .
哈密尔顿路与哈密尔顿圈 .
存在性定理 . . . . . . . . .
§2.3.1 基本定义 . . . . . .
§2.3.2 存在性定理 . . . . .
算法及数值算例 . . . . . .
本章小结 . . . . . . . . . .
空间填充曲线
基础知识 . . . . . . . . . .
希尔伯特序 . . . . . . . . .
希尔伯特序编码解码算法 .
§3.3.1 预备知识 . . . . . .
§3.3.2 编码算法 . . . . . .
§3.3.3 解码算法 . . . . . .
希尔伯特曲线的新变体 . .
算法的具体化 . . . . . . . .
本章小结 . . . . . . . . . .
1
. . . . . . . . . . . . . . . . . . . .
1
. . . . . . . . . . . . . . . . . . . .
4
. . . . . . . . . . . . . . . . . . . .
5
. . . . . . . . . . . . . . . . . . . .
7
. . . . . . . . . . . . . . . . . . . .
8
9
. . . . . . . . . . . . . . . . . . . . . . . .
9
. . . . . . . . . . . . . . . . . . . . . . . .
10
. . . . . . . . . . . . . . . . . . . . . . . .
11
. . . . . . . . . . . . . . . . . . . . . . . .
11
. . . . . . . . . . . . . . . . . . . . . . . .
13
. . . . . . . . . . . . . . . . . . . . . . . .
15
. . . . . . . . . . . . . . . . . . . . . . . .
16
19
. . . . . . . . . . . . . . . . . . . . . . . .
19
. . . . . . . . . . . . . . . . . . . . . . . .
22
. . . . . . . . . . . . . . . . . . . . . . . .
25
. . . . . . . . . . . . . . . . . . . . . . . .
25
. . . . . . . . . . . . . . . . . . . . . . . .
28
. . . . . . . . . . . . . . . . . . . . . . . .
31
. . . . . . . . . . . . . . . . . . . . . . . .
33
. . . . . . . . . . . . . . . . . . . . . . . .
35
. . . . . . . . . . . . . . . . . . . . . . . .
36
PHG 动态负载平衡模块的设计与实现
39
PHG 的层次网格结构与动态负载平衡 . . . . . . . . . . . . . . . . .
39
§4.1.1 二分自适应网格加密、放粗与 PHG 的层次网格结构
§4.1.2 网格划分
. . . .
39
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
v
学位论文
vi
PHG 的分布式二叉树 . . . . . . . . . . . . . . . . . . . . . . .
43
§4.1.4 动态负载平衡 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
§4.2.1 加密树划分方法 . . . . . . . . . . . . . . . . . . . . . . . . . .
46
§4.2.3 一维划分算法 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
§4.2.5 划分质量度量 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
§4.1.3
§4.2 网格划分方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
§4.2.2 空间填充曲线划分方法 . . . . . . . . . . . . . . . . . . . . . .
49
§4.2.4 子网格―进程映射 . . . . . . . . . . . . . . . . . . . . . . . . .
52
PHG 的动态负载平衡模块 . . . . . . . . . . . . . . . . . . . . . . . .
55
§4.3.1 动态负载平衡函数 . . . . . . . . . . . . . . . . . . . . . . . . .
55
§4.4 数值算例 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
§4.4.2 划分质量对自适应有限元计算性能的影响 . . . . . . . . . . .
57
§4.3
§4.3.2 网格划分和数据迁移 . . . . . . . . . . . . . . . . . . . . . . .
55
§4.4.1 划分质量比较 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
§4.5 本章小结 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
第五章
hp 自适应策略设计
69
§5.1 基础知识 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
§5.1.2 有限元离散 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
§5.2 hp 自适应有限元 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
§5.2.2 hp 自适应加密策略 . . . . . . . . . . . . . . . . . . . . . . . .
76
§5.3.1 hp 自适应策略 . . . . . . . . . . . . . . . . . . . . . . . . . . .
78
§5.5 本章小结 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
§5.1.1 模型问题
§5.1.3 自适应有限元 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
§5.2.1 有限元解空间和变分形式 . . . . . . . . . . . . . . . . . . . . .
75
§5.3 一种新的 hp 自适应策略 . . . . . . . . . . . . . . . . . . . . . . . . .
78
§5.4 数值算例 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
第六章
总结与展望
参考文献
87
89
发表文章目录
105
致
107
谢
第一章
引言
偏微分方程出现在十八世纪, 由物理学家在求解实际问题时提出, 目前已广
泛应用到物理、化学、生物以及经济等领域. 由于存在复杂的初始条件、边界条
件等, 许多问题得不到解析解, 只可以用数值方法求出近似解, 常用方法有有限差
分方法、有限元方法、有限体积方法等. 它们先将问题离散, 然后使用计算机求
解. 随着科技的进步, 需要求解的问题规模越来越大, 对计算精度及运行时间的
要求越来越高. 正是这些因素的存在, 促使着人们发展更有效的数值方法、计算
机硬件和软件. 数值方法的发展和并行计算机软硬件的进步, 为我们求解挑战性
问题带来了机遇.
§1.1
自适应有限元方法
有限元方法是求解偏微分方程的一种数值方法, 此方法在 20 世纪 50 年代
初由工程师们提出, 并用于求解简单的结构问题. 有限元方法的原始思想来自
Courant, 他在 1943 年提出在三角形网格上用分片线性函数去逼近 Dirichlet 问
题. 有限元方法作为一种系统的数值方法, 并奠定数学基础, 则是在 60 年代中期,
由以冯康先生为代表的中国学者与西方学者独立并行地完成 [207].
有限元方法从物理问题出发, 利用变分原理, 整体地考虑物理问题, 得到方程
的变分表达式, 同时对物理问题的计算区域做剖分形成计算网格, 然后在剖分区
域 (单元) 上用简单函数去逼近原问题的解. 有限元方法将一个无穷维问题转化
成有限维问题, 通过求解这个问题, 得到近似解. 有限元方法的优点之一是能方
便地处理复杂区域.
来源于实际问题的偏微分方程往往存在局部奇性, 采用传统有限元方法求解
时, 在解有奇性的区域需要使用非常细的网格才能保证有限元解的精度. 传统有
限元由于使用一致网格, 会导致网格规模非常大. 而解光滑的区域并不需要这么
细的网格, 从而会浪费计算资源. 更有效的做法是仅加密解有奇性的区域. 这是
自适应有限元方法思想的由来.
自适应有限元方法根据解的性质动态地调整网格/基函数. 对自适应有限元
方法 [199], 要解决的问题是如何确定有奇性的区域以及如何对数值解给出一个可
信的误差估计. 标准的先验误差估计只能给出数值解的渐近误差行为, 并且需要
知道解的正则性, 而解的正则性往往是不知道的. 自适应有限元方法需要从数值
解及已知的数据导出误差估计 – 后验误差估计, 并且后验误差估计的计算代价需
要远小于数值解求解的代价. 此外, 误差估计需要是局部的, 并且是真实误差的
1
学位论文
2
上下界, 其中上界表明解的精度是可信的, 小于事先给定的误差, 下界表明误差估
计是精确的, 网格/基函数将在需要调整的地方被调整. 自适应过程是一个计算数
值解、计算误差、调整网格/基函数、重新计算的迭代过程, 具体步骤如图 1.1 所
示 [204]. 常用的自适应方法有: h 自适应方法, p 自适应方法, hp 自适应方法以及
r 自适应方法.
图 1.1 自适应有限元流程图
为了能让自适应流程 1.1 高效运行, 需要有正确的离散方法、求解器、误差
指示子、网格标记策略和合适的加密策略, 同时也要选取适当的参数. 误差指示
子是自适应有限元方法的关键之一, 常用的误差指示子主要有: 残量型误差指示
第一章 引言
3
子, 对偶权重残量误差指示子, 多重基误差指示子, 平均化方法以及平衡残量误差
指示子等.
在给定误差指示子的情况下, 如何选取单元进行加密/放粗同样会影响自适
应有限元的收敛性和效率. 常用的单元标记策略主要有最大误差策略 (max stratedy), 误差等分布策略 (equidistribution stratedy), 保证误差下降策略 (guaranteed
error reduction strategy) [39], MNS (MNS-refinement strategy) 策略 [40] 等. 后面章
节将对标记策略做详细的介绍.
自适应有限元方法另一个关键模块是网格单元类型及加密策略的选取. 二维
区域常选取四边形或者三角形作为基本单元. 四边形网格常采取正则加密, 加密
部分单元会导致悬点的出现, 否则只能进行一致加密. 当执行这种加密时, 在一条
边上可能会出现多个悬点, 这会对有限元空间及程序的数据结构有很多的限制.
为了克服这个问题, 开发人员常采用 Bank[41, 42] 设计的一悬点 (1-irregular) 和三
邻居 (three-neighbors) 原则. 一悬点原则限制每条边上的悬点最多只有一个, 如果
某条边由于邻居单元的加密引入新的悬点, 则该单元需要被加密以保证一悬点原
则. 三邻居原则指出如果四边形的三条边都有悬点, 则该单元需要被加密. 对三角
形单元, 常用的加密算法有正则加密和二分加密 [43]. 正则加密 [44] 是连接三角
形的三个中点, 将一个三角形单元加密成四个子单元, 对于出现的悬点, 则是对
邻居使用一个临时的二分加密以保证协调性. 二分加密是连接顶点和顶点对应的
边, 将一个三角形加密成两个子单元. 二分加密算法有最长边二分加密 [45, 46, 47]
和最新顶点二分加密 [48]. 由 Rivara 提出的最长边二分加密算法是将最长边作为
加密边, 将其中点与相对的顶点相连, 把单元分成两个子单元. 由 Sewell 提出的
最新顶点加密算法是将三角形中最新产生的顶点作为加密顶点, 连接对应边的中
点, 将三角形加密成两个子单元.
三维网格常选取四面体或者六面体作为单元. 六面体的自适应要复杂一些,
有的算法基于八叉树, 即将一个六面体剖分成八个子六面体, 由于会出现悬点, 接
着对邻居做临时的剖分, 将这些邻居单元剖分成金字塔和棱柱, 当临时单元被标
记为加密单元时, 先把它们恢复成六面体, 然后剖分 [49]. 有的做法是定义加密模
板, 当某个单元被标记为加密单元时, 为该单元及邻居选择合适的模板, 生成一个
没有悬点的六面体网格 [50]. 六面体网格的自适应是很复杂的, 程序的运行时间
也较长, 而四面体网格的自适应则要方便得多. 四面体常采用二分加密算法, 和
三角形一样, 可采用最长边加密和最新顶点加密. 四面体二分加密时选取一条边
作为加密边将其一分为二, 产生两个子单元. 最长边二分加密算法由 Rivara 提出
[45, 46, 47]. 最新顶点二分加密算法首先由 Sewell 针对二维区域提出 [48], 然后由
Bänsch 扩展到三维区域 [51]. 最新的进展可以参考 Kossaczky[52], Liu 和 Joe [54],
学位论文
4
及 Arnold [53] 等人的文章.
在网格二分加密的并行算法方面, 许多科研人员做出了有意义的工作. Rivara
等人 [55] 为四面体网格的全局加密设计了并行算法, 但不适合局部加密. Pébay 和
Thompson [57] 对基于边分割的四面体网格设计了并行算法. Jones 和 Plassmann
[43] 为三角形网格设计了并行局部自适应加密算法. Barry、Jones 及 Plassmann
[56] 同时给出了一个实现并行自适应有限元程序的框架. Castãnos 和 Savage [58]
为四面体网格设计了一种并行局部自适应加密算法, 并且应用在软件包 Pared 中.
最近, 张 [1] 为四面体最新顶点二分加密策略设计了并行算法, 并且应用在并行自
适应有限元软件平台 PHG (Parallel Hierarchical Grid) [59] 中.
§1.2
动态负载平衡
偏微分方程的并行数值求解的关键因素之一是处理好进程间的数据分布. 数
据分布指根据进程数对数据进行划分, 将数据分布存储在各进程中. 计算过程中
数据分布保持不变的程序, 如传统有限元程序, 在整个计算过程中仅需在开始做
一次静态划分. 然而对自适应有限元程序, 由于每个进程的负载是在不断地变化
的, 并且这个变化是不可以或很难预测的. 这种情况下, 就需要动态地调整数据分
布以保持负载平衡. 好的数据划分方法既要使得每个进程拥有相同的计算负载,
又要减少进程间的通信.
并行自适应有限元计算中最常用的数据划分方法是划分计算网格, 每个进程
负责一个子网格上的计算, 为保证可扩展性, 每个进程仅存储该子网格及定义在
其上的数据, 不在本地的数据在需要时通过进程间通信获得. 这种计算模型包含
三个模块: 数据划分、维护本地数据结构以方便计算和进程间通信、在需要的时
候进行通信. 在自适应有限元程序中, 工作负载是与网格的单元、面、边及顶点
相关的. 数据划分问题就是将这些对象划分到不同的进程中, 使每个进程拥有尽
可能相同数目的对象. 在有限元计算中, 以单元划分数据是最常用的. 然而每个
单元的负载可能并不相同, 例如在不同的单元上求解不同的物理问题或者像在 p
自适应有限元中每个单元的自由度数可能并不相同, 可以给每个单元一个权重以
表示计算负载. 数据划分的目标就是使得每个进程拥有相同的权重, 同时使得有
限元计算中的通信量最小.
在计算过程中根据各进程中负载的变化调整数据的划分以保持负载平衡称
为动态负载平衡. 并行自适应有限元计算中, 由于各进程的负载不断变化, 动态
负载平衡是必需的. 动态负载平衡中的数据划分算法需满足一些条件: 1) 能在分
布式的数据上并行执行; 2) 执行速度快, 因为它们是纯开销, 并且可能被频繁地
第一章 引言
5
调用; 3) 最好是增量的, 即网格微小的变化仅导致划分的微小变化; 4) 内存使用
是适度的; 5) 容易使用.
对粒子模拟, 粒子间的作用是由距离决定的, 通常距离近的相互影响较大, 按
照距离划分数据会降低通信量. 而对于基于网格的计算, 如有限元、有限体积等
方法, 通信一般发生在邻居之间, 邻接关系决定了通信模式. 动态负载平衡中的
数据划分算法也分为两类: 基于几何性质的算法和基于图的算法. 基于几何性质
的算法利用对象的几何信息, 如坐标信息, 主要算法有递归坐标二分算法 (RCB,
Recursive Coordinate Bisection) [60] 、递归惯量二分算法 (RIB, Recursive Inertial
Bisection) [61]、空间填充曲线算法 (SFC, Space-Filling Curve) [62] 等; 基于图的算
法利用对象的拓扑关系, 如邻居关系或者顶点边的连接信息等, 主要方法有递归
图二分算法 (RGB, Recursive Graph Bisection) [60]、贪婪算法 [63]、递归谱二分方
法 (RSB, Recursive Spectral Bisection) [60]、K-L 方法 (Keinighnan-Lin Algorithm)
[64] 等. 著名的网格划分软件包有: ParMETIS[14], Zoltan[19], JOSTLE[65] 等.
§1.3
并行自适应有限元软件平台 PHG
PHG (Parallel Hierarchical Grid) 是一个并行自适应有限元软件包, 由科学与
工程计算国家重点实验室研制. 该软件包主要用 C 语言开发 (部分使用 C++), 底
层的通信由 MPI 完成, 面向分布式存储并行机, 具有很强的可移植性. PHG 对用
户隐藏了通信接口, 无需用户干预通信, 顶层代码和串行代码一样. PHG 提供了
一个可扩展的并行自适应有限元程序框架, 它提供了网格输入、网格划分、网格
加密/放粗、自由度操作、线性系统组装和求解、特征值求解、可视化、性能统计
等功能. PHG 采取了面向对象的设计, 方便用户添加新的功能. 下面简要介绍一
下 PHG 的功能及特点:
(1) 网格对象. PHG 目前处理的网格对象为协调四面体网格, 输入文件可以为 ALBERTA [66], MEDIT [67], GAMBIT [68] 等格式, 可以处理周期网格及镜像网
格. PHG 提供了丰富的边界类型及边界管理函数, 可以满足各种常规的需求.
网格局部自适应采用最新顶点二分加密算法 [1], 网格的加密/放粗层次结构
被保留, 并分布在不同的进程中. 在周期网格中, PHG 自动维护网格的周期
性.
(2) 动态负载平衡. PHG 提供动态负载平衡功能, 提供了加密树划分方法 [84]、空
间填充曲线划分方法 [62] 等网格划分方法. PHG 同时提供了接口以调用外
部函数库, 如 Zoltan, ParMETIS 等, 方便用户根据需要选取不同的算法.
学位论文
6
(3) 自由度管理. 自由度对象是 PHG 的基本数据结构之一, 用于描述和处理分布
在网格上的各种数据. 它可以用于定义有限元函数、空间, 也可以用于存储、
处理任何其他分布在网格上的数据, 如几何信息等. PHG 提供了自由度的创
建、插值、赋值、复制、基本算术运算等操作, 同时提供了有限元函数的微积
分运算. PHG 还提供了一些特殊类型的自由度, 如常量自由度、解析自由度、
几何量自由度等. PHG 提供了统一的创建自由度的模板用来创建各种自由
度.
(4) 数值积分. 为方便有限元计算 PHG 提供了三角形和四面体上任意阶的数值
积分公式 [69], 这些积分公式的权重都是非负的, 并且所有的积分点均不在单
元外部. PHG 提供了基函数缓存机制, 极大减少重复计算, 加快积分过程.
(5) 向量和稀疏矩阵管理. PHG 提供了一组函数用于管理连续分块存储的分布
式向量, 以及基于行划分、行压缩存储的分布式稀疏矩阵. 向量和矩阵的分
布通过映射进行描述与管理. 这些模块的提供, 使得线性系统可以独立于网
格而存在, 亦使得 PHG 可以独立作为线性方程组解法器使用.
(6) 线性解法器. PHG 提供了一套统一的接口用于求解线性方程组. PHG 本身
实现了一些解法器, 如 PCG, GMRES 等, 和一些预条件子. PHG 同时提供了
丰富的接口, 以供用户调用外部解法器软件包, 如 PETSc, HYPRE, SuperLU,
MUMPS, SPC, LASPack, SPOOLES 等. PHG 定义了线性解法器对象 SOLVER,
用户可以通过该对象方便地定义、操作线性方程组的未知量、系数矩阵和右
端项. PHG 定义的解法器对象独立于特定的方法, 方便用户按照自己的需要
添加解法器及预条件子.
(7) 特征值解法器. PHG 目前没有实现自己的特征值求解器, 但提供了与 PARPACK,
JDBSYM, LOBPCG, SLEPc, Trilinos/Anasazi 和 PRIMME 等软件包的接口, 用
于计算广义特征值和特征向量. 特征值、特征向量的计算采用统一接口, 可
在运行程序时通过命令行选项选择任何一个可用的特征值求解器.
(8) 可视化. PHG 目前实现了 OpenDX[70] 以及 VTK[71] 等格式可视化文件的输
出, 可以输出自由度数据及网格数据.
(9) 参数控制. PHG 提供了灵活的命令行选项管理机制用于控制程序中的参数.
PHG 的选项按照目录分类, 可读性很强.
(10) 性能统计. PHG 提供了获取程序运行时浮点性能及内存使用的函数, 方便用
户获取程序的实际性能.
第一章 引言
7
PHG 同时提供了内存管理、信息输出、输入等功能, 其中信息输出按照用户
指定的输出等级, 满足不同的调试要求. 完整的 PHG 功能以及具体的函数接口
在这里不再一一介绍, 用户可以在 PHG 网站 [59] 中获得.
§1.4
论文主要工作
本论文的工作是围绕 PHG 平台开展的, 主要包括两部分内容. 第一部分内容
涉及网格划分和动态负载平衡算法, 论文中研究了哈密尔顿路、希尔伯特序、网
格划分算法等, 并在此基础上设计和实现了 PHG 的动态负载平衡模块. 第二部
分内容涉及 hp 自适应策略的研究, 论文针对二分加密网格设计了 hp 自适应策
略, 并与已有策略进行了比较. 论文所完成的主要工作具体如下:
(1) 证明了过顶点哈密尔顿路和哈密尔顿圈的存在性并设计实现了线性复杂度的
算法. 针对 PHG 所处理的协调四面体网格, 本文在网格的对偶图连通的条件
下证明了过顶点哈密尔顿圈和任意两个单元之间的过顶点哈密尔顿路的存在
性, 并在此基础上设计了线性复杂度的四面体网格上过顶点哈密尔顿圈和过
顶点哈密尔顿路的构造算法. 文中还对相应并行算法的构造进行了探讨. 本
文在 PHG 平台中实现了该算法, 并将其应用于加密树网格划分方法中初始
网格单元的排序. 将来如果需要, 该算法亦可作为单元排序方法直接应用于
网格划分中.
(2) 设计了任意维空间中具有线性复杂度的希尔伯特序的编码解码算法并提出了
希尔伯特空间填充曲线的一种变体. 希尔伯特序具有良好的空间局部性, 广
泛应用于内存管理、数据库索引、动态负载平衡等领域. 本文总结、分析了文
献中已有的生成二维和三维空间希尔伯特序的算法, 设计了任意维空间中具
有线性复杂度的希尔伯特序编码、解码高效算法, 同时并对算法进行了改进,
设计了复杂度更低的算法, 降低了计算量. 文中还给出了希尔伯特空间填充
曲线的一个变体, 它可保证曲线的编码顺序不随曲线阶数的改变而变化. 本
文在 PHG 平台中实现了二维及三维算法, 并成功将其应用于网格划分中.
(3) 设计实现了加密树及空间填充曲线划分方法并设计了 PHG 的动态负载平衡
模块. 加密树划分方法最初由 Mitchell 提出, 本文在实现该算法的同时亦对
算法进行了改进, 降低了算法的计算复杂度, 提高了算法实现的效率. 同时实
现了空间填充曲线划分框架, 可以使用希尔伯特空间填充曲线或 Morton 空
间填充曲线, 并实现并改进了 Zoltan 中的一个一维划分算法用于对排序后的
单元进行一维划分. 此外, 本文还实现了 PLUM 中设计的子网格 – 进程映射
学位论文
8
算法, 用于减少动态负载平衡中的数据迁移量. 数值实验表明, 本文在 PHG
中所实现的空间填充曲线划分算法不论在划分时间还是划分质量上均优于其
它一些软件包 (其中包括由美国 Sandia 国家实验室开发的著名的网格划分和
数据迁移软件包 Zoltan) 中所提供的相应算法. 本文所实现的基于希尔伯特
序的网格划分算法已经成为 PHG 中默认的网格划分方法, 经受了一些并行
规模达数千子网格、计算规模超过十亿单元的大型应用的考验, 表现出很高
的效率和很强的稳健性.
(4) 设计了一个基于误差下降预测的 hp 自适应策略. 当前, hp 自适应有限元算法
研究中的一个难点是自适应策略的设计, 目前尚没有成熟的、已被广泛接受
的算法. 本文对 hp 自适应策略进行了研究，在前人提出的几种基于误差下
降预测的 hp 自适应策略的基础上给出了一个新的 hp 自适应加密策略, 该策
略适用于二维三角形、四边形和三维四面体、六面体等不同类型的单元, 适
用于正则加密、二分加密等不同自适应加密方式. 数值实验表明, 该策略可
以达到最优的误差指数下降, 并在数值解的精度和计算效率上优于文献中的
一些策略. 这部分工作的另外一个目的是验证 PHG 的 hp 自适应模块的正确
性和适用性.
§1.5
论文内容安排
论文分为六章. 第一章为引言, 本章对自适应有限元及并行自适应有限元软
件平台 PHG 进行简要介绍. 第二、三、四章研究动态负载平衡方法, 其中第二章
研究哈密尔顿路和哈密尔顿圈. 第三章研究希尔伯特序的编码解码算法. 第四章
介绍 PHG 的动态负载平衡及数据迁移算法. 第五章研究 hp 自适应策略. 最后一
章给出总结及未来工作.
第二章
哈密尔顿路和哈密尔顿圈
在三维自适应有限元计算中, 协调四面体网格是最常用的网格类型之一. 四
面体单元适合非结构网格, 适合处理复杂边界, 其数据结构也简单. 在并行自适
应有限元计算中, 一种常用的网格划分方法是先将网格线性化, 即将单元进行排
序, 接着根据进程数将单元按照线性化后的顺序分成几段, 每一段映射给不同的
进程. 哈密尔顿路是常用的单元排序方法之一. 本章研究 PHG 中的非结构协调
四面体网格上哈密尔顿路的基本性质及构造方法. 目前在 PHG 中, 哈密尔顿路
被用于 RTK 网格划分方法 (加密树划分方法) 中初始单元顺序的确定，将来如果
需要亦可将它直接用于网格划分。
§2.1
协调四面体网格
本章中, 单元是一个四面体, 它被看作 R3 中的闭区域. 一个四面体有四个顶
点、六条边和四个三角形面, 如图 2.1 所示. 本章中涉及到的四面体都是非退化
的, 即四个顶点不在一个平面上.
图 2.1 四面体单元
定义 2.1.1. 令 Ω 为 R3 中一个连通的开区域, 一个四面体网格 M 是四面体单元
o
o
o
的集合, 记作 {Ti }, 这些单元满足如下条件: Ω = ∪Ti , Ti ∩ Tj = ∅ (i 6= j), 其中 T
表示单元 T 的内部, ∅ 为空集.
一个顶点如果位于区域 Ω 的边界上, 则被称作边界点, 否则是内部点. |M|
表示网格 M 的单元数目. 两个单元 Ti 和 Tj (i 6= j) 如果有一个公共面, 则它们
9
学位论文
10
被称为邻居.
定义 2.1.2. 网格 M 被称为协调的, 如果对 M 中任意两个不同的单元 Ti 和 Tj ,
Ti ∩ Tj 是一个顶点、一条边、一个面或者空集.
协调四面体网格的生成方法主要有八叉树方法 [76, 77]、Delaunay 方法 [78, 79]
等, 常用的四面体网格生成软件有 Gmsh [80]、Netgen [81]、TetGen [82] 等. 协调四
面体网格局部自适应加密算法主要有最长边二分加密算法 [45, 46, 47]、最新顶点
加密算法 [1, 48, 51, 53, 52, 54] 等, 其质量度量标准主要有最小 – 最大立体角、最
小平面角、半径比、边长比等.
§2.2
哈密尔顿路与哈密尔顿圈
自适应有限元方法根据数值解调整计算网格/基函数. 在分布式存储并行计
算机中, 自适应会导致负载不平衡, 网格划分对并行程序而言是一个重要的模块.
它被用来对网格进行初始划分, 并且当网格变化时, 对网格进行重划分, 以保证并
行程序在计算过程中有均衡的负载. 一类重要的网格划分方法是: 先把网格单元
排序, 形成一个一维的列表, 然后把这个列表划分成段, 每段具有相等的权重. 哈
密尔顿路 (Hamiltonian path) 是这类方法中的排序方法之一 [62, 125, 129].
哈密尔顿路是图论中的概念, 它是一条通过所有结点的路, 且每个结点仅通
过一次. 在四面体网格中, 哈密尔顿路被定义成一个单元序列, 每个单元仅出现
一次. Heber 等人 [85] 对二维三角形单元网格证明了哈密尔顿路的存在性, 其中
哈密尔顿路通过一个顶点或者一条边进入下一个单元, 同样通过一个顶点或者一
条边离开一个单元, 定理的证明需要如下条件: 三角形网格是协调的并且不能存
在局部割点 (local cut vertex) [83]. Mitchell [83] 针对二维及三维空间研究了三角
形、四边形、四面体、六面体网格的哈密尔顿路和哈密尔顿圈的存在性. 对于
四面体网格, 由于仅通过面的哈密尔顿路不一定存在, 因此 Mitchell 的文章中引
入了过顶点哈密尔顿路 (through-vertex Hamiltonian path) 以及过顶点哈密尔顿圈
(through-vertex Hamiltonian cycle) 的概念. Mitchell 证明了: 如果协调四面体网格
没有局部割点和局部割边, 那么过顶点哈密尔顿路和哈密尔顿圈存在. 证明是构
造性的, 可以直接得到构造算法.
本文将在一个假设下, 即网格的对偶图连通, 证明过顶点哈密尔顿路和哈密
尔顿圈的存在性. 证明同样是构造性的, 可以直接导出线性复杂度的算法.
第二章 哈密尔顿路和哈密尔顿圈
11
存在性定理
§2.3
本节将研究四面体网格上过顶点哈密尔顿路和哈密尔顿圈的存在性, 内容分
为两部分, 第一部分先给出基本定义, 介绍一些记号和概念; 第二部分给出存在性
定理.
§2.3.1
基本定义
为了叙述的方便, 在这里将介绍所用到的概念及记号, 其中部分概念来源于
Mitchell 的文章 [83].
定义 2.3.1. 协调四面体网格 M 的对偶图定义为一个无向图, 在这个图中, 网格
M 中每一个单元对应图中的一个顶点, M 中的两个邻居单元则对应图中的一条
边.
定义 2.3.2. 在一个图中, 如果任意两个顶点之间存在一条通路, 则称图是连通的.
定义 2.3.3. 在协调四面体网格 M 中, 一条长度为 n 的路 (path), 是一个单元序
列, 记作 T1 T2 . . . Tn , Ti ∈ M, i = 1, . . . , n, 其中 Ti ∩ Ti+1 6= ∅, 并且 Ti 6= Ti+1 ,
i = 1, n − 1.
定义 2.3.4. 在协调四面体网格 M 中, 一个长度为 n 的圈 (cycle), 是一条长度为
n + 1 的路, 其中 T1 = Tn+1 .
定义 2.3.5. 协调四面体网格 M 上的一条哈密尔顿路 (Hamiltonian path) 是一条
路, 其中网格 M 的每个单元正好出现一次.
定义 2.3.6. 协调四面体网格 M 上的一条部分哈密尔顿路 (partial Hamiltonian
path) 是一条路, 其中每个单元至多出现一次.
定义 2.3.7. 协调四面体网格 M 上的一个哈密尔顿圈 (Hamiltonian cycle) 是一个
圈, 其中网格 M 的每个单元正好出现一次.
定义 2.3.8. 一条过顶点路 (through-vertex path) 是一条路, 其中总是通过顶点进
入或离开一个单元, 并且对同一个单元, 进入或者离开此单元的顶点不是同一个.
在上述定义下, 可以引申出过顶点哈密尔顿路.
定义 2.3.9. 一条(部分) 过顶点哈密尔顿路 (through-vertex Hamiltonian path) 是
一条过顶点路, 并且是一条(部分) 哈密尔顿路.
学位论文
12
过顶点哈密尔顿路可以表示为 H = T1 v1 T2 · · · vn−1 Tn , 其中 Ti 6= Tj (i 6= j) 且
vi 6= vi+1 (1 ≤ i ≤ n − 2), vi 表示连接 Ti 和 Ti+1 的顶点. 过顶点哈密尔顿圈可以
类似表示.
令 M 为区域 Ω 上的一个网格, v 为 M 中的一个顶点, 如果 Ω\v 是非连通的,
则称 v 为 M 中的一个割点. 如果 v 满足如下条件: ∃R > 0 使得 ∀ r (0 < r ≤ R),
(B(v, r) ∩ Ω) \ v 是非连通的, 则称 v 为 M 的局部割点 [83], 其中 B(v, r) 表示以 v
中心的半径为 r 的球. 割点和局部割点如图 2.2 和 2.3 所示. 图 2.2 中顶点 v 是
割点, 图 2.3 中顶点 v 是局部割点. 为了简便, 我们使用了二维图形. 类似的, 可
以定义割边及局部割边.
v
图 2.2 含有割点 v 的网格
图 2.3 含有局部割点 v 的网格
从定义可以很容易看出, 对偶图连通的四面体网格是没有割点的, 但可能会
有局部割点. 没有割点及局部割点的四面体网格由于可能存在割边或者局部割
第二章 哈密尔顿路和哈密尔顿圈
13
边, 对偶图亦不一定连通.
§2.3.2
存在性定理
对协调四面体网格, 本节将证明过顶点哈密尔顿路和哈密尔顿圈的存在性.
本节先给出三个引理, 它们将在后面的证明过程中被用到. 第一个引理给出邻居
间共享顶点的个数. 它是我们构造和扩充哈密尔顿路的基础.
引理 2.1. 假设网格 M 是协调四面体网格, 其对偶图是连通的并且 |M| ≥ 2, 则
有如下结论成立:
(1) 若 M1 和 M2 是 M 的非空子集, 且 M1 ∩ M2 = ∅, M1 ∪ M2 = M, 则存在
单元 T1 ∈ M1 及 T2 ∈ M2 , 使得它们的交集 T1 ∩ T2 是一个公共面, 即它们是
邻居.
(2) 如果 T , T1 和 T2 是 M 中不同的单元, 并且 T1 和 T2 均是 T 的邻居, 则这三
个单元拥有两个公共顶点.
(3) 如果 T , T1 , T2 和 T3 是 M 中不同的单元, 并且 T1 , T2 和 T3 均是 T 的邻居,
则这四个单元拥有一个公共顶点.
证明. 结论 (1) 是对偶图连通的直接推论.
结论 (2): 因为单元 T1 和 T2 都是 T 的邻居, 它们中的每一个都分别和 T 共
享一个面, 并且很容易验证这两个面共享一条边, 所以 T1 , T2 及 T 三个单元共享
一条边, 拥有两个相同的顶点, 结论成立.
结论 (3): 因为 T1 , T2 和 T3 都是 T 的邻居, 它们中的每一个都和 T 共享一
个面, 在四面体中, 任意三个面共享一个顶点, 结论成立.
引理 2.2. 若协调四面体网格 M 的对偶图是连通的, H = T1 v1 T2 v2 · · · Tk−1 vk−1 Tk
(k ≥ 2) 为 M 的一个部分过顶点哈密尔顿路. 那么,
(1) 如果 T1 和 T2 是邻居, 那么 T1 的任何不属于 H 的邻居可以插入到 H 中, 并
且位于 T1 和 T2 之间.
(2) 如果 Tk−1 和 Tk 是邻居, 那么 Tk 的任何不属于 H 的邻居可以插入到 H 中,
并且位于 Tk−1 和 Tk 之间.
证明. 由于两个结论是相似的, 仅给出结论 (1) 的证明过程.
令 T 为 T1 的不属于 H 的邻居. 根据引理 2.1 可知, T1 , T2 和 T 有两个共同
的顶点, 记作 u1 、u2 .
学位论文
14
如果 H 仅包含 T1 和 T2 , 即 k = 2, 那么 T 可以按照如下的方式插入到 H 中:
H = T1 u1 T u 2 T2 .
现在假设 k > 2. 如果 v1 ∈ {u1 , u2 }, 不失一般性, 假设 v1 = u2 , 那么 T 可以
按照如下方式插入到 H 中: H = T1 u1 T v1 T2 v2 · · · Tk . 如果 v1 ∈
/ {u1 , u2 }, 那么 u1
和 u2 中必定有一个顶点和 v2 是不同的, 记作 u1 , 在这种情况下 T 可以按照如下
方式插入到 H 中: H = T1 v1 T u1 T2 v2 · · · Tk . 重复这个过程, T1 的任意一个邻居都
可以插入到 T1 和 T2 之间.
由第二个引理可以知道, 如果创建合适的初始路, 那么我们可以保持该哈密
尔顿路的头部和尾部不变, 把它们所有的邻居单元插入到这两个单元之间. 接下
来介绍第三个引理. 这个引理指出, 如果某个部分过顶点哈密尔顿路没有包含所
有的单元, 则可以通过插入新的单元来扩展这条路, 直至包含所有的单元.
引理 2.3. 若协调四面体网格 M 的对偶图是连通的, H = T1 v1 · · · vn−1 Tn 是 M
中一条非空的部分过顶点哈密尔顿路, 它包含了 T1 及 Tn 的所有邻居. 令 M1 =
{T1 , · · · , Tn }, 如果 M1 6= M, 则 H 可以通过插入新单元来扩展, 并且插入的单元
位于 T1 和 Tn 之间.
证明. 记 M2 = M\M1 . 如果 M2 是非空的, 根据引理 2.1, 存在 T ∈ M2 及
Ti ∈ M1 (1 < i < n) , 使得 T 和 Ti 是邻居.
由于 Ti 有四个顶点, 那么 Ti 和 T 共有的三个顶点之一必定是 Ti 的入口顶点
vi−1 或者出口顶点 vi , 不失一般性, 假定它是 vi (否则的话, 可以反转 H, 在完成插
入之后, 再重新反转), 并且 Ti 和 T 的三个共有顶点之一必定不是 vi−1 或者 vi , 记
作 v. 那么新的部分过顶点哈密尔顿路可以构造为: H = T1 v1 . . . Ti vT vi Ti+1 . . . Tn .
定理得证.
注记: 引理 2.3 的一个条件: 部分过顶点哈密尔顿路包含 T1 及 Tn 的全部邻居是
必需的. 反例, 假定 H = T1 vTn , 其中 T1 和 T2 仅共享一个顶点 v, 如果 T1 有不含
有顶点 v 的邻居 T , 那么不可能将 T 插入到 T1 和 Tn 之间.
定理 2.4. 假设协调四面体网格 M 的对偶图是连通的, 且 |M| ≥ 2. 对网格 M
中任意两个单元 T1 和 Tn , T1 6= Tn , 存在一条 T1 到 Tn 的过顶点哈密尔顿路.
证明. 由于 M 的对偶图是连通的, 可以在网格 M 中找到单元序列 T1 T2 · · · Tn ,
使得序列中任意两个相邻单元是邻居, 并且可以创建初始的过顶点哈密尔顿路
H = T1 v1 T2 v2 · · · vn−1 Tn (采用归纳法, 假设已经创建 H = T1 v1 · · · Ti−1 vi−1 Ti . 对
第二章 哈密尔顿路和哈密尔顿圈
15
单元 Ti+1 , 它和 Ti 有三个公共顶点, 其中之一和 vi−1 是不同的, 记作 vi , 然后 H
可以扩展成 H = T1 v1 · · · Ti−1 vi−1 Ti vi Ti+1 . 重复这个过程, 便可构造出包含 T1 , · · · ,
Tn 的初始部分过顶点哈密尔顿路). 利用引理 2.2, 可以把 T1 和 Tn 的所有邻居插
入到 H 中, 并且全部位于 T1 和 Tn 之间. 接着利用引理 2.3, 可将所有剩余的单
元插入到 H 中. 证明完毕.
下面一个定理是定理 2.4 的直接推论, 它给出协调四面体网格中过顶点哈密
尔顿圈的存在性.
定理 2.5. 如果协调四面体网格 M 的对偶图是连通的, 并且 |M| ≥ 2, 那么在网
格 M 中存在过顶点哈密尔顿圈.
证明. 选取两个邻居单元 T1 和 Tn . 根据定理 2.4, 存在一个哈密尔顿路 H =
T1 v1 T2 v2 . . . vn−1 Tn . 由于 T1 和 Tn 是邻居, 它们有三个公共顶点, 其中一个是
和 v1 及 vn−1 不同的, 记作 vn . 那么可以按照如下方式创建哈密尔顿圈, H =
T1 v1 T2 v2 . . . vn−1 Tn vn T1 .
最后一个定理, 定理 2.6, 对于将来设计哈密尔顿路及哈密尔顿圈的并行算法
是有用的. 由于并行程序中网格是分布在不同进程中的, 这个定理指出, 每个进
程可以独立地创建自己的哈密尔顿路和哈密尔顿圈, 然后将这些部分哈密尔顿路
和哈密尔顿圈组合起来, 形成一个完整的哈密尔顿路和哈密尔顿圈.
定理 2.6. 若四面体网格 M 的对偶图是连通的, M1 ⊂ M 及 M2 ⊂ M 是网格
M 的子网格, 满足条件: |M1 | ≥ 1, M2 ≥ 1, M1 ∩ M2 = ∅ 及 M1 ∪ M2 = M. 如
果 M1 和 M2 的对偶图也是连通的, 那么可以分别对两个子网格构建哈密尔顿
路, 记作 H1 和 H2 , 使得 H = H1 vH2 是 M 的一个过顶点哈密尔顿路, 其中 v 是
M1 和 M2 的一个公共顶点.
证明. 由于 M 的对偶图是连通的, 根据引理 2.1, 存在 T1 ∈ M1 及 T2 ∈ M2
使得 T1 和 T2 是邻居. 由于两个子网格的对偶图均是连通的, 根据定理 2.4, 在
两个子网格中分别存在过顶点哈密尔顿路 H1 = T0 · · · v1 T1 及 H2 = T2 v2 · · · Tn .
因为 T1 和 T2 是邻居, 它们有三个公共顶点, 其中必然有一个顶点和顶点 v1 及
v2 是不同的, 记作 v, 从而 M 上的过顶点哈密尔顿路可以按照如下的方式创建
H = H1 vH2 = T0 · · · v1 T1 vT2 v2 · · · Tn .
§2.4
算法及数值算例
上一节中定理的证明是构造性的, 它们可以直接转化成构建过顶点哈密尔顿
学位论文
16
路和哈密尔顿圈的高效算法. 由证明过程也可以看出, 几个定理的证明是很相似
的, 算法同样也是类似的. 为简便起见, 本节仅给出构造过顶点哈密尔顿路的算
法. 由于证明本身是构造性的, 在此仅给出主要步骤, 具体细节可以在定理的证
明过程中得到. 算法共有三个步骤. 第一步是创建初始的哈密尔顿路; 第二步是
插入头部和尾部单元所有的邻居; 第三步是扩展哈密尔顿路使得这条路包含网格
所有的单元.
算法 1: 哈密尔顿路构造算法
假设 M 是一个协调四面体网格, 其对偶图是连通的, n = |M| ≥ 2. 对任意给定的四
面体 T1 和 Tn , 一个过顶点哈密尔顿路 H = T1 v1 T2 · · · vn−1 Tn 可以按如下三个步骤
构造:
Step 1. 创建一条初始路. 这条路中, 任意两个相邻的单元都有一个公共面. 具体的
创建办法见定理 2.4.
Step 2. 把 T1 和 Tn 的所有邻居插入到初始哈密尔顿路中. 具体做法参看引理 2.2
的证明.
Step 3. 将所有剩余的单元插入到哈密尔顿路中. 具体做法参看引理 2.3 的证明.
下面分析算法 1 的计算复杂度. 第一步是一个图论问题, 这个问题可以由宽
度优先搜索 (breadth fist search) 算法实现, 复杂度为 O(n), 其中 n 为网格的单元
数目. 由于每个单元的邻居是有限的, 被 O(n) 控制, 因此第二步插入邻居过程的
复杂度是 O(n). 在第三步时, 插入每个单元的代价是 O(1), 而剩余单元的数目是
有界的, 不超过 n, 因此第三步的计算复杂度是 O(n). 综合以上分析, 算法的计算
复杂度是 O(n).
我们在自适应有限元工具包 PHG (Parallel Hierarchical Grid) [1, 59] 中实现
了算法 1. 下面给出一些测试结果, 测试平台是一台 DELL PowerEdge 2950 服务
器 (2 个四核 1.60GHz Intel E5310 CPU, 4MB 二级缓存). 我们选取了五组不同的
网格, 单元数目从十多万到几百万不等. 这些网格均由 J. Schöberl 开发的网格生
成软件 NETGEN [81] 生成. 表 2.1 给出了程序的计算时间统计, 可以看出即使单
元数达到三百万的规模, 用于生成过顶点哈密尔顿路的时间仍然少于四秒. 同时,
程序的运行时间是与网格规模成比例增长的, 表明算法的计算复杂度是线性的.
§2.5
本章小结
本章证明了协调四面体网格上过顶点哈密尔顿路和哈密尔顿圈在一定条件
下的存在性, 并实现了线性复杂度的算法, 该算法可用于网格划分中单元的排序.
第二章 哈密尔顿路和哈密尔顿圈
17
表 2.1 程序运行时间
单元数
面数
顶点数
时间 (seconds)
单元/秒
Mesh 1
155,456
323,264
32,238
0.127770
1,216,686
Mesh 2
239,616
480,704
40,781
0.240720
995,413
Mesh 3
453,736
923,336
85,338
0.515140
880,801
Mesh 4
1,243,648
2,536,704
232,288
1.265506
982,727
Mesh 5
3,211,264
6,498,304
573,628
3.736147
859,512
18
学位论文
第三章
空间填充曲线
空间填充曲线提供了一类有效的网格划分方法. 这类曲线出现于 19 世纪, 由
于其特殊的性质一直吸引数学家进行研究, 研究主要集中在空间填充曲线的数学
性质上, 如连续性、可微性等. 空间填充曲线具有良好的空间局部性, 近几十年,
人们将其广泛应用在内存管理、数据库索引、动态负载平衡、图像处理等领域, 在
空间填充曲线的算法研究方面取得了长足的进步. 本章对高维希尔伯特空间填充
曲线的编码解码算法进行研究.
§3.1
基础知识
George Cantor 在 1878 年指出任意两个有限维光滑流形具有相同的势 [86].
该结果暗示从单位区间 [0, 1] 到单位正方形区域 [0, 1]2 存在双射. 问题是这种类
型的映射是不是连续的. 数学家 Netto 在 1879 年给出一个著名结果 [86], 指出这
样的双射是间断的.
定理 3.1. (Netto 定理) 如果 f 是一个从 m 维光滑流形到 n 维光滑流形的双射,
其中 m 6= n, 那么 f 是间断的.
有了这个定理之后, 问题变成: 如果去掉双射的限制, 是否存在一个满射将
区间 [0, 1] 映射到正方形区域 [0, 1]2 上. Giuseppe Peano (1858–1932) 在 1890 年发
现了第一个这样的曲线. 因此, 具有这样性质的曲线不仅被称为空间填充曲线, 又
被称为 Peano 曲线. 曲线的广义定义如下:
定义 3.1.1. 一个 n 维空间填充曲线是一个连续的满射, 将单位区间 [0, 1] 映射到
n 维超立方 [0, 1]n .
在 Peano 之后, 其他数学家也对这个问题进行了深入研究, 进一步的例子分
别被 D. Hilbert (1891), E.H. Moore (1900), W. F. Osgood (1903), H. Lebesgue (1904),
W. Sierpiński (1912), G. Pólya (1913), K. Knopp (1917) 等人给出 [86]. 一般情况下,
空间填充曲线由线段序列迭代, 当迭代次数趋于无穷时生成. 用希尔伯特空间填
充曲线做例子, 如图 3.1, 3.2 及 3.3 所示: 图 3.1 是经过一次迭代产生的曲线; 图
3.2 是经过二次迭代产生的曲线; 图 3.3 是经过六次迭代产生的曲线.
空间填充曲线主要有: Peano 空间填充曲线、希尔伯特空间填充曲线、Moore
空间填充曲线、Sierpiński 空间填充曲线、Morton 空间填充曲线、Lebesgue 空间填
充曲线、Osgood 空间填充曲线等. 其中 Sierpiński 空间填充曲线 [105] 如图 3.4 所
示, Morton 空间填充曲线 [88] 如图 3.5 所示 (高维曲线同样存在, 出于简洁的考
19
学位论文
20
图 3.1 一阶希尔伯特空间填充曲线
图 3.2 二阶希尔伯特空间填充曲线
虑, 仅给出了二维图). 这些曲线大多是在一个世纪前发现的, 但它们的具体应用
却是在近几十年. 空间填充曲线具有自相似性、空间局部性等特点, 被广泛用在
数学、算法、地理信息系统、图像处理、数据库、电路设计、加密、内存管理、数
据压缩、科学计算及动态负载平衡等领域 [89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 100,
101, 102, 103, 104]. Velho 等人利用空间填充曲线设计了图像抖动算法, 应用于激
光打印机等 [87, 104, 118]. Salmon 等人利用空间填充曲线为 N - 体模拟程序进行
数据划分及重排序以改进程序的空间局部性 [103]. Challacombe 等人利用空间填
充曲线在并行机上对稀疏矩阵进行划分 [95]. Mellor-Crummey 等人 [100, 101] 以
第三章 空间填充曲线
21
图 3.3 六阶希尔伯特空间填充曲线
图 3.4 Sierpiński 空间填充曲线 (一阶, 二阶, 三阶)
及 Hu 等人 [96] 利用空间填充曲线对数据及计算重排序以改善内存使用. Matias
等人 [98] 利用空间填充曲线对视频进行处理. Hwansoo Han[99] 等人将空间填充
曲线用在 CFD 计算等科学计算领域中. 空间填充曲线有良好的空间局部性, 也
被用在数据压缩 [106]、索引 [107] 等方面. 在并行计算方面, 空间填充曲线划分
算法也是动态负载平衡常用的算法 [18, 19, 62, 123], 其中希尔伯特空间填充曲线,
Morton 空间填充曲线, Sierpiński 空间填充曲线具有最好的空间局部性, 应用也是
最广泛的. 由于 Morton 空间填充曲线的算法简单, 人们首先使用这种曲线, 应用
领域相当广泛, 但由图 3.5 可以看出, 该曲线具有跳跃性, 其空间局部性相对于另
外两种曲线要差一些. 随着希尔伯特空间填充曲线和 Sierpiński 空间填充曲线的
编码解码算法日趋成熟, 这两种空间填充曲线越来越占有主导地位, 尤其是希尔
学位论文
22
图 3.5 Morton 空间填充曲线 (二阶)
伯特空间填充曲线.
§3.2
希尔伯特序
G. Peano 第一个发现了空间填充曲线, 但第一个给出空间填充曲线一般化生
成方法的人却是希尔伯特 (D. Hilbert, 1891) [86, 118]. 在这里, 用二维希尔伯特空
间填充曲线作为例子来阐述空间填充曲线的生成过程, I 表示区间 [0, 1], Q 表示
正方形 [0, 1]2 . 希尔伯特空间填充曲线的生成方式是: 将区间 I 剖分成四个子区
间, 将 Q 剖分成四个子正方形, 第一次迭代的结果如图 3.1 所示; 接着将 I 的每
个子区间剖分成四个子区间, I 子区间的个数为 16, 将 Q 的每个子正方形剖分成
四个子正方形, 子正方形的个数同样为 16, 第二次迭代的结果如图 3.2; 经过六次
迭代产生的曲线如图 3.3 所示. 该过程一直持续, 当趋于无穷时, 产生一个充满整
个正方形的曲线. 这些子正方形的安排按照如下方式: 相邻子区间对应的子正方
形有一条公共边.
假设迭代次数为 m, 迭代过程会将单位区间 I 及单位正方形 Q 剖分成 22m
个子区间和 22m 个子正方形. 对区间 I 的每个子区间编号, 编号取值为整数, 其
范围为 I2m = {0, 1, · · · , 22m − 1} , 单位正方形 Q 被分成了 22m 个子正方形, 每个
m
子正方形赋予一个整数坐标, 其取值范围为 Qm
2 = {(x, y) : 0 ≤ x, y < 2 − 1}. 迭
代次数 m 又称为阶 (order)、分辨率 (resolution) 或深度 (depth). 在下面的讨论
中, 阶这个术语将会被一直使用. 我们用 Hnm 表示 m- 阶 n- 维希尔伯特空间填充
曲线. 仍用二维希尔伯特空间填充曲线做例子, 取 m 为 2, 集合 Q22 中每一个整数
第三章 空间填充曲线
23
坐标均对应集合 I22 中的一个编号, H22 如图 3.6 所示, 曲线的起始点为 0, 终点为
15, 用黑色实心圆表示. 我们可以看出 H22 是由 H21 生成的: H22 包含了四个经过
空间变换的 H21 , 如图 3.6 所示; H23 也是由四个经过空间变换的 H22 得到. 从这个
观察可以得到一般化的希尔伯特空间填充曲线生成方法: H2m 可以由四个 H2m−1
拼接得到, 只是这四个 H2m−1 分别需要做一次空间变换. 具体的变换方式将在接
下来的内容中仔细分析.
6
5
4
9
10
8
7
11
3
2
13
12
0
1
14
15
图 3.6 希尔伯特序 (H22 , 二阶)
0
1
2
3
8
9
10
11
4
6
12
14
5
7
13
15
图 3.7 Z-order (Morton order, 二阶)
24
学位论文
对任意的正整数 m, 上述方法为集合 Qm
2 定义了一种顺序 (order). 由于该顺
序是由希尔伯特空间填充曲线定义的, 又被称为希尔伯特序 (Hilbert order, Hilbert
code). 对于不同类型的空间填充曲线, 可以定义不同的排序方法 (ordering), 如
Morton 空间填充曲线, 由于其曲线与 Z 很相似, 又称为 Z- 序 (Z-order, Morton
order, Morton code) [88], 图 3.7 是一个二阶 Morton 空间填充曲线定义的 Z- 序.
m
m
m
Qm
2 与 I2 是一个一一对应关系, 将 Q2 中一个坐标映射成 I2 中一个整数的过程
称为编码 (enconding), 编码过程为每一个坐标建立一个索引 (index); 将 I2m 中一
个整数映射成 Qm
2 中一个坐标的过程称为解码 (decoding), 解码过程是一个曲线
遍历的过程. 高维空间填充曲线的编码解码定义和二维是一样的. 希尔伯特空间
填充曲线和希尔伯特序是同一个内容, 只是名称不一样.
接下来的内容集中在希尔伯特序 (希尔伯特空间填充曲线) 上. 很多研究人员
对希尔伯特空间填充曲线的计算方法进行了研究, 提出了许多有效的算法, 它们可
以划分为两类: 递归算法 [109, 110, 111, 112] 和迭代算法 [62, 113, 114, 115, 116, 118].
对递归算法, Butz, Goldschlager[109, 110] 等人设计了简洁的算法, 这些算法是空间
填充曲线递归算法的基础. Breinholt 和 Schierz[108] 设计了效率更高的递归算法.
Witten 等人使用 GROPER 语言描述了一种递归技术 [111]. Cole 在一个图形系统
上实现了递归过程, 并为 Peano 及 Sierpiński 空间填充曲线描述了递归算法 [112].
在通常情况下, 迭代算法的速度要快于递归算法, 尤其是表驱动 (table-driven) 的
迭代算法 [62, 113, 114, 115, 116, 118]. Fisher[113], Liu 和 Schrack[115, 116] 为二维
和三维问题设计了基于二进制操作的迭代算法. Faloutsos 等人 [117] 通过分析希
尔伯特空间填充曲线和 Z- 序之间的关系, 实现了编码解码算法. 在文 [62] 中, 作
者展示了基于表的算法, 其中使用了两个表分别存储了各个象限的变换信息及状
态信息, 由于仅需要查表而不需要重新计算, 极大提高了生成速度. 以上提到的算
法一般都被限制在二维和三维空间, 对高维空间填充曲线, 由于复杂性很高, 这方
面的工作比较少. Butz[109] 的递归算法可以对高维曲线进行计算, 但复杂度较高,
速度慢, 不适合大规模计算. Lawder[119] 重新对 Butz 的算法进行研究, 并给出了
改进. 最近 Kamata 等人 [120] 提出了基于硬件的高维算法. Li 等人 [121] 分析了
高维希尔伯特空间填充曲线不同象限的变换关系.
下面介绍本文作者在希尔伯特序编码和解码算法方面的工作. 本文的工作基
于文章 [118, 121]. Li 等人设计了算法分析高维希尔伯特序的演化规则, 这些规
则包含了坐标变换信息, 由于问题的复杂性, 作者并没有给出编码解码算法. 本
文对这些演化规则进行了细致地分析和理解, 并给出这些变换的数学表达式和解
释, 在此基础上, 设计了编码和解码算法. 对于编码和解码算法, 分别给出了基于
位运算的算法和基于算术运算的算法. 基于位运算的算法是基于算术运算算法
第三章 空间填充曲线
25
的改进, 计算效率更高. 算法的时间计算复杂度为 O(mn). 对于给定的编码解码
问题, 算法可以在 m 步给出结果. Chen 等人 [118] 利用希尔伯特空间填充曲线特
有的性质, 针对二维问题设计了复杂度更低的编码解码算法, 计算复杂度为 O(k),
其中 k 定义为 k = log2 (max(x, y)) + 1, 与 m (k ≤ m) 的大小无关. 利用类似的性
质, 可以将三维希尔伯特序的编码解码算法的复杂度降为 O(k) [123], 其中 k 定义
为 k = log2 (max(x, y, z)) + 1. 本章给出了这种思想的理论依据, 并推广到高维希
尔伯特空间填充曲线, 得到一系列复杂度更低的算法. 同一个坐标, 在不同阶数
的希尔伯特空间填充曲线中的序并不一定相同, 我们对传统的希尔伯特空间填充
曲线做了改造, 得到希尔伯特曲线新的变体, 该曲线满足性质: 同时属于不同阶
曲线的一组坐标在这些曲线中的顺序是一样的. 本章为该曲线变体设计了编码解
码算法.
§3.3
希尔伯特序编码解码算法
本节将介绍编码算法和解码算法. 我们先给出预备知识, 然后分析文章 [121]
中介绍的算法, 分析高维希尔伯特空间填充曲线的性质, 给出理论结果, 在此基础
上介绍编码和解码算法, 其中每类算法将使用四种不同的方式表达, 两种使用算
术运算表达, 两种使用位运算表达, 并且具有不同的计算复杂度. 这里介绍的算法
对任意维数 (n ≥ 2) 的希尔伯特空间填充曲线均适用. 本文给出的是算法框架, 可
以作为模板应用到不同维数的问题中.
§3.3.1
预备知识
本节所有的操作均在非负整数上进行. 需要声明的是操作同样可以在实数
m
域上进行, 只需将输入及输出做一次转换即可. 令 Qm
n 为整数坐标集合, Qn =
{(xn , · · · , x2 , x1 )|0 ≤ xi < 2m , 1 ≤ i ≤ n}, (xn , · · · , x1 ) (∈ Qm
n ) 是希尔伯特空间填充
m
曲线的坐标, xi 为坐标的第 i 个分量. 记 In = {0, 1, · · · , 2nm − 1}. 当 n 为 2 时,
即通常的平面 (此时有 x2 = x, x1 = y); 当 n 为 3 时, 即通常的三维空间 (此时有
x3 = x, x2 = y, x1 = z). 当所有的 xi 和 yi (1 ≤ i ≤ n) 均为 0 或者 1 时, 定义如下
逻辑运算 ∧
(xn , · · · , x2 , x1 ) ∧ (yn , · · · , y2 , y1 ) = (xn ∧ yn , · · · , x2 ∧ y2 , x1 ∧ y1 ),
在这里 ∧ 代表异或运算符 XOR.
(3.1)
学位论文
26
引入符号 (a1 a2 · · · ak )d 表示 d 进制数字, 其中 0 ≤ ai < d, 1 ≤ i ≤ k, 且 k 可
以为任意的正整数. 当 d 为 2 时, 表示二进制; 当 d 为 10 时, 表示通常的十进制.
各种不同的数字表示方式可以互相转换. 当 d 为 2 时, 定义取反运算符 Rek 如下:
Rek (j) = Rek ((a1 a2 · · · ak )2 ) = (b1 b2 · · · bk )2 ,
bi = 1 − ai , 1 ≤ i ≤ k,
(3.2)
其中 j 是一个十进制数, j = (a1 a2 · · · ak )2 为 j 的二进制表示, j 满足条件 0 ≤ j <
2k . 对二进制数, 可以引入通常的按位与 (and), 右移, 左移运算符, 分别记作 &, À
和 ¿. 引入定义 [122]
pin (a1 , a2 , · · ·
, an ) = (
i
X
aj ) mod 2,
(3.3)
j=1
其中 ai 等于 0 或者 1 (1 ≤ i ≤ n). 借助 pin , 定义映射 fn
fn (a1 , a2 , · · · , an ) = (b1 b2 · · · bn )2 = j,
b1 = a1 ,
bi =
(
ai , if pni−1 = 0
1 − ai , if pni−1 = 1
, (3.4)
其中 ai 等于 0 或者 1, 并且 j 为十进制数. 该函数将一个向量 (a1 , a2 , · · · , an ) 映
射成一个整数. 同时定义 fn 的逆映射 bn
bn (j) = bn ((a1 a2 · · · an )2 ) = (b1 , b2 , · · · , bn ),
b1 = a1 ,
bi =
(
ai , if ai−1 = 0
,
1 − ai , if ai−1 = 1
(3.5)
其中 j 为一个十进制数且等于 (a1 a2 · · · an )2 .
注记: 在实际应用中, 对特定的 n, 映射 bn 及 fn 可以事先计算好存储在表中, 需
要的时候直接查询. 对于 bn , 可以采用如下方式快速计算
bn ((a1 · · · an )2 ) = (a1 , · · · , an ) ∧ (0, a1 , · · · , an−1 ).
(3.6)
由于 fn 和 bn 互为逆映射, 计算 bn 之后, 可以直接得到 fn 的结果.
下面部分概念来自文章 [121]. n- 维希尔伯特单元 (n-dimensional cell) 是一个
1 阶的 n- 维希尔伯特空间填充曲线. n- 维希尔伯特因子列表 (Hilbert gene list) 是
一系列坐标变换指令, 它们给出了从 (m−1)- 阶希尔伯特空间填充曲线到 m- 阶曲
线的变换指令. 坐标的变换指令分为两类: 交换指令 (exchange command) 和反转
指令 (reverse command). 它们可以解释为坐标关于某个超平面的对称变换. 希尔
第三章 空间填充曲线
27
伯特空间填充曲线是有方向的, 它有一个入口和一个出口. 当对一个希尔伯特单
元做坐标变换时, 入口和出口的空间坐标可能会改变. 对一个 n- 维空间, 它有 2n
个象限, 例如二维平面有四个象限, 三维空间有八个象限, 编号均从 0 开始, 编号
方式由 fn 直接计算. 每个象限均有不同的变换指令. 引入 Hni,0 和 Hni,1 (0 ≤ i < 2n )
表示单元经过 i- 象限的指令变换后的入口和出口坐标. 在文章 [121] 中, 作者介
绍了如何生成希尔伯特单元和希尔伯特因子列表.
希尔伯特单元在文章 [121] 中是按照递归的方式生成的, 如果要生成 n- 维希
尔伯特单元, 需先生成 (n − 1)- 维希尔伯特单元. 本文分析了其生成方式, 基于理
论分析, 设计了映射 bn 直接生成单元, 同时引入了 bn 的逆映射 fn . bn 将在解码
算法中被使用, fn 在编码算法中被使用.
下面我们对希尔伯特因子列表进行分析, 给出其数学解释并给出具体的变换
表达式. 希尔伯特因子列表包含两种指令: 交换指令和反转指令. 这两组指令均
可以解释为关于某个超平面的对称变换. 对于交换指令, 如果变换发生在分量 xi
与 xj (i 6= j) 之间, 则其对称面为超平面 xi −xj = 0, 坐标 (xn , · · · , xi , · · · , xj , · · · , x1 )
经过变换之后则变为 (xn , · · · , xj , · · · , xi , · · · , x1 ), 该结果可以由简单的数学演算直
接推导出来, 简单说就是交换第 i 个与第 j 个位置的分量, 这是称其为交换指令
的原因. 对反转指令, 假设变换的位置为 i 并且当前希尔伯特空间填充曲线的阶
m
数为 m, 则其对称面为超平面 xi = 2 2−1 , 坐标 (xn , · · · , xi , · · · , x1 ) 经过变换后变
为 (xn , · · · , 2m − 1 − xi , · · · , x1 ). 该操作与映射 Rem (xi ) 是等价的. 反转变换改变
的仅是 xi 位置的坐标值, 结果与指令的执行顺序是无关的. 这两种变换的逆变换
均是自身, 它们均是保长度的, 并且保持曲线上各个点的相对位置不变, 即对给定
的曲线, 变换只会改变曲线在空间中的位置信息, 但不会改变各个坐标对应的序
的相对编号. 总结如下:
性质 3.2. 交换变换的本质是镜面对称变换, 如果变换发生在分量 xi 与 xj (i 6= j)
之间, 其对称面为超平面 xi − xj = 0; 反转变换的本质也是镜面对称变换, 如果变
m
换发生在 xi 分量且当前曲线的阶数为 m, 则对称面为超平面 xi = 2 2−1 , 如果有
多个反转变换, 结果与执行顺序无关. 交换和反转变换的逆变换均为自身.
文章 [121] 给出了生成这些指令的算法, 为了表达的方便, 本文对生成算法进
行了重写, 表达式如下:
n
i,0
i,1
Gi,0
n = (bn (0) ∧ bn (2 − 1)) ∧ (Hn ∧ Hn ),
i,0
Gi,1
n = bn (0) ∧ Hn ,
(0 ≤ i < 2n ),
(0 ≤ i < 2n ).
(3.7)
(3.8)
学位论文
28
Gni,0 和 Gi,1
n 表示第 i 象限的交换指令和反转指令. 根据 bn 的定义, 可以将 bn (0) 及
bn (2n −1) 分别写成如下形式 (0, · · · , 0) 和 (1, 0, 0, · · · , 0), 有关系 bn (0)∧bn (2n −1) =
(1, 0, 0, · · · , 0) 成立. 希尔伯特单元的出口和入口坐标仅有一个分量不同, 并且两
种变换是保持曲线上各点的相对位置不变, 经过变换后的单元的入口坐标和出口
坐标仅有一个分量的值是不同的, 所以入口坐标 Hni,0 与出口坐标 Hni,1 做异或运
算, 得到的结果中除了一个分量为 1 外全部为 0. 根据式 (3.7), Gi,0
n 的分量要么全
部为 0 要么仅有两个不为 0. 这意味着, 每个象限上要么不执行交换指令要么仅
需要执行一次. 如果 Gi,0
n 的第 i 个与第 j 个分量为 1, 则在这两个分量之间执行
交换指令. 反转指令则可能需要执行多次.
在此介绍希尔伯特曲线的一个特殊性质. 分析象限 0 的性质, 根据定义可
以把 bn (0), bn (2n − 1), Hn0,0 及 Hn0,1 分别写成如下形式: (0, · · · , 0), (1, 0, · · · , 0),
0,1
(0, · · · , 0) 和 (0, · · · , 0, 1). 根据式 (3.7) 及 (3.8), 有 G0,0
n = (1, 0, · · · , 0, 1) 和 Gn =
(0, · · · , 0). 这意味着在第 0 象限, 有且仅有交换变换, 变换发生在第一个和最后一
个分量之间. 这是设计计算复杂度更低的算法的理论基础.
性质 3.3. 在每个象限中, 最多执行一次交换变换, 反转变换可以被执行多次. 在
0,1
0 象限中, G0,0
n = (1, 0, · · · , 0, 1), Gn = (0, · · · , 0), 有且仅有交换变换.
§3.3.2
编码算法
本文给出 n 维空间 (n ≥ 2) 希尔伯特空间填充曲线的编码算法. 本小节仅给
出算法模板, 当给定具体的维数 n 后, 可以根据这个模板导出相应的算法.
前面已经提到, 希尔伯特空间填充曲线是自相似的, 曲线的生成是一个迭代
(递归) 过程, m- 阶希尔伯特空间填充曲线 Hnm 的生成需要 (m − 1)- 阶曲线 Hnm−1 .
具体过程可以描述为: n 维空间有 2n 个象限, 其每一个象限包含一个 (m − 1)- 阶
希尔伯特空间填充曲线, 对每个象限的 (m − 1)- 阶希尔伯特空间填充曲线实施该
象限对应的交换指令和反转指令, 然后连接相邻象限中曲线的入口和出口, 即可
得到一个 m- 阶的希尔伯特空间填充曲线. 如图 3.1 及 3.2 所示, 可以看出 H22 的
四个子正方形中的每个均包含一个 H21 . 从曲线的入口出发, 遍历曲线, 可以为曲
线中每个点建立一个序, 这是编码过程.
下面详细分析编码过程。对任意给定的坐标 P ，它的序等于从入口开始遍历
曲线到指定位置所经过的点的个数. 假设当前曲线的阶数为 m, 它由 2n 个 (m−1)阶曲线生成, P 必定位于某个 (m − 1)- 阶曲线中, 假设位于第 i 象限中, 遍历此曲
线, 在它到达所处的 (m − 1)- 阶曲线之前, 要经过 i 个 (m − 1)- 阶曲线. 使用记号
Xm (P ) 表示点 P 在 m- 阶曲线中的希尔伯特序, 那么它的希尔伯特序可以由下式
表示:
第三章 空间填充曲线
29
em−1 (P ),
Xm (P ) = i × 2n(m−1) + X
em−1 (P ) 表示点 P 在经过变换的 (m − 1)- 阶希尔伯特曲线中的序. 前面已
其中 X
经讲过, 交换指令和反转指令是保持长度的, 不会改变曲线的序, 所以我们将 P
所处的某个 (m − 1)- 阶曲线做逆变换, 变换成一个没有经过改变的 (m − 1)- 阶曲
线, 上式相应地变为
Xm (P ) = i × 2n(m−1) + Xm−1 (P ).
(3.9)
该式是设计迭代算法的出发点, 它保证我们可以使用统一的算法框架计算希尔伯
特序.
假设当前希尔伯特空间填充曲线的阶数为 m, 任意一个点 (xn , · · · , x1 ) (∈
m m−1 · · · x1 ) . 希
Qm
n ), 其每个分量 xi (1 ≤ i ≤ n) 可以写成二进制表示 xi = (xi xi
i 2
m
尔伯特序保存在整数 (rm rm−1 · · · r1 )2n (∈ In ) 中. 算法中用到的变换信息 Gi,0
n 及
i,1
n
Gn (0 ≤ i < 2 ) 需事先计算好, 具体的过程可以从文章 [121] 得到.
希尔伯特空间填充曲线的生成是从低阶到高阶的递归过程, 这样的方式可以
生成一个完整的曲线. 由于编码解码问题仅需计算一个点, 这样的方式是浪费时
间和空间的, 根据式 (3.9), 本文采取从高阶到低阶的一个回溯过程, 以迭代的方
式计算. 变换指令有两类, 在编码过程中, 反转指令先执行, 然后是交换指令. 下
面给出第一个算法, 它基于算术运算, 具体过程如下:
编码算法 1.1.
m
m
(1) 如果 m = 0, 程序终止. 否则, rm = fn (xm
n , xn−1 , · · · , x1 ).
m−1 .
(2) 对每个整数 i (1 ≤ i ≤ n), 如果 xm
i = 1, 那么 xi = xi − 2
(3) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Gnrm ,1 的第 i 个分量是 1, 那么 xi =
2m−1 − 1 − xi .
(4) 交换. 如果 Grnm ,0 存在两个非零分量, 其位置为 i 和 j, 则交换 xi 和 xj .
(5) m = m − 1, 转至 (1).
在编码算法 1.1 中, 如果将整数写成二进制数, 会发现在步骤 (2) 中, xi =
是被 (xim−1 · · · x1i )2 所代替, 仅丢掉一个高位有效数字, 该运算可以由
简单的按位与操作 (&) 完成. 步骤 (3) 中的算术运算等价于一个按位取反运算,
1
(xm
i · · · xi )2
30
学位论文
等价于 Rem−1 操作. 在实际应用中, 位运算的效率要高于算术运算, 上述算法可
以按照位运算的方式重写:
编码算法 2.1.
m
m
(1) 如果 m = 0, 程序终止. 否则, rm = fn (xm
n , xn−1 , · · · , x1 ).
1
(2) 对每个整数 i (1 ≤ i ≤ n), xi = (xm−1
· · · x1i )2 = (xm
i · · · xi )2 & (011 · · · 1)2 .
i
(3) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Gnrm ,1 的第 i 个分量是 1, 那么 xi =
Rem−1 (xi ) = (xm−1
· · · x1i )2 ∧ (1 · · · 1).
i
(4) 交换. 如果 Grnm ,0 存在两个非零分量, 其位置为 i 和 j, 则交换 xi 和 xj .
(5) m = m − 1, 转至 (1).
编码算法 1.1 和 编码算法 2.1 只是表达不同, 它们具有相同的复杂度. 下
面分析它们的计算复杂度. 算法在步骤 (2) 和 (3) 需要更新 n 个坐标, 复杂度为
O(n), 步骤 (4) 的计算复杂度为 O(1), 算法会在 m 步内终止, 因此算法的时间复
杂度为 O(nm). 对于空间复杂度, 算法需要存储 2n 个象限的变换信息, 空间复杂
度为 O(2n ). 问题本身的复杂度是按照指数增长的, 这是前人的工作大多集中在
二维及三维的原因. 当 n 较小时, 存储变换信息所占的内存是可以接受的, 例如
维数 n 为 20 时, 所占内存大概为 2M. 当维数较高时, 高效实现仍然是挑战性的.
根据 fn 的定义, 有 r = fn (0, 0, · · · , 0) = 0. 由性质 3.3 知道在象限 0 中仅有
交换指令. 观察 编码算法 1.1, 会发现仅有步骤 (4) 被执行. 这个过程将一直执行
到 r 是非零值, 在这个过程仅有坐标分量 xn 与 x1 交换, 并且交换两次的结果是
保持坐标不变. 这个过程可以跳过, 如果跳过的次数为偶数, 保持坐标不变; 如果
次数为奇数, 则交换 xn 与 x1 . 通过这个观察, 可以设计计算复杂度更低的算法.
定义 k = floor(log2 (max{xn , · · · , x1 })) + 1, 并令 log2 (0) 等于 0. 上述两个算法可以
分别改写成如下形式:
编码算法 1.2.
(1) 设定 (rm · · · r1 )2n 为 0. 如果 m 与 k 具有不同的奇偶性, 则交换 x1 与 xn .
m = k.
m
m
(2) 如果 m = 0, 程序终止. 否则 rm = fn (xm
n , xn−1 , · · · , x1 ).
m−1 .
(3) 对每个整数 i (1 ≤ i ≤ n), 如果 xm
i 为 1, 则 xi = xi − 2
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnm ,1 的第 i 个分量是 1, 那么 xi =
2m−1 − 1 − xi .
(5) 交换. 如果 Grnm ,0 存在两个非零分量, 其位置为 i 和 j, 则交换 xi 和 xj .
第三章 空间填充曲线
31
(6) m = m − 1, 转至 (2).
编码算法 2.2.
(1) 设定 (rm · · · r1 )2n 为 0. 如果 m 与 k 具有不同的奇偶性, 则交换 x1 与 xn .
m = k.
m
m
(2) 如果 m = 0, 程序终止. 否则 rm = fn (xm
n , xn−1 , · · · , x1 ).
1
(3) 对每个整数 i (1 ≤ i ≤ n), xi = (xm−1
· · · x1i )2 = (xm
i · · · xi )2 & (011 · · · 1)2 .
i
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnm ,1 的第 i 个分量是 1, 那么 xi =
Rem−1 (xi ).
(5) 交换. 如果 Grnm ,0 存在两个非零分量, 其位置为 i 和 j, 则交换 xi 和 xj .
(6) m = m − 1, 转至 (2).
编码算法 1.2 和 编码算法 2.2 的时间计算复杂度为 O(kn), 对某个给定的坐
标, 其时间复杂度是与曲线的阶数 m 无关的. 对阶数 m 较高的应用, 这两个算法
更有优势. 它们的空间复杂度与 编码算法 1.1 及 编码算法 2.1 是一样的, 都是
O(2n ).
§3.3.3
解码算法
解码过程是给定整数 z (∈ Inm ) 及希尔伯特空间填充曲线的阶 m, 求出整数
对应的空间坐标 (xn , · · · , x2 , x1 ) (∈ Qm
n ) 的过程. 解码算法所需要的理论基础均已
在上面给出, 此处仅给出每个编码算法对应的解码算法, 不再对解码算法做详细
的讨论.
记 m 为希尔伯特空间填充曲线的阶, (rm rm−1 · · · r1 )2n 为整数 z 的 2n - 进制
数表示. 引入 k, 它表示 (rm rm−1 · · · r1 )2n 中第一个非零元素的位置, 即 rk > 0 且
ri = 0 (i > k). 同时假定当 z 为 0 时, k 为 1. 假定 log2n (0) 为 0, k 的定义等价于
k = floor(log2n (z)) + 1. 计算结果保存在坐标 (xn , · · · , x2 , x1 ) (∈ Qm
n ) 中.
在解码算法中, 交换指令先执行, 然后是反转指令. 下面分别给出上述四个
编码算法对应的解码算法.
解码算法 1.1.
(1) 初始化 (xn , · · · , x1 ), (xn , · · · , x2 , x1 ) = bn (r1 ). v = 2.
(2) 如果 v > m, 程序终止. 否则 (sn , · · · , s2 , s1 ) = bn (rv ).
(3) 交换. 如果 Grnv ,0 在位置 i 及 j 有非零值, 则交换 xi 与 xj .
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnv ,1 的第 i 分量为 1, 则 xi = 2v−1 −1−xi .
32
学位论文
(5) 对每个整数 i (1 ≤ i ≤ n), 如果 si 等于 1, 则 xi = xi + 2v−1 .
(6) v = v + 1, 转至 (2).
解码算法 2.1.
(1) 初始化 (xn , · · · , x1 ), (xn , · · · , x2 , x1 ) = bn (r1 ). v = 2.
(2) 如果 v > m, 程序终止. 否则 (sn , · · · , s2 , s1 ) = bn (rv ).
(3) 交换. 如果 Grnv ,0 在位置 i 及 j 有非零值, 则交换 xi 与 xj .
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnv ,1 的第 i 分量为 1, 则 xi = Rev−1 (xi ).
(5) 对每个整数 i (1 ≤ i ≤ n), 如果 si 等于 1, 则 xi = (xi ) ∧ (1 ¿ (v − 1)).
(6) v = v + 1, 转至 (2).
解码算法 1.2.
(1) 初始化 (xn , · · · , x1 ), (xn , · · · , x2 , x1 ) = bn (r1 ). v = 2.
(2) 如果 v > k, 程序终止. 否则 (sn , · · · , s2 , s1 ) = bn (rv ).
(3) 交换. 如果 Grnv ,0 在位置 i 及 j 有非零值, 则交换 xi 与 xj .
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnv ,1 的第 i 分量为 1, 则 xi = 2v−1 −1−xi .
(5) 对每个整数 i (1 ≤ i ≤ n), 如果 si 等于 1, 则 xi = xi + 2v−1 .
(6) v = v + 1, 转至 (2).
(7) 如果 m 与 k 具有不同的奇偶性, 则交换 xn 与 x1 . 最终的坐标为 (xn , · · · , x1 ).
解码算法 2.2.
(1) 初始化 (xn , · · · , x1 ), (xn , · · · , x2 , x1 ) = bn (r1 ). v = 2.
(2) 如果 v > k, 程序终止. 否则 (sn , · · · , s2 , s1 ) = bn (rv ).
(3) 交换. 如果 Grnv ,0 在位置 i 及 j 有非零值, 则交换 xi 与 xj .
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnv ,1 的第 i 分量为 1, 则 xi = Rev−1 (xi ).
(5) 对每个整数 i (1 ≤ i ≤ n), 如果 si 等于 1, 则 xi = (xi ) ∧ (1 ¿ (v − 1)).
(6) v = v + 1, 转至 (2).
(7) 如果 m 与 k 具有不同的奇偶性, 则交换 xn 与 x1 . 最终的坐标为 (xn , · · · , x1 ).
这四个算法的时间计算复杂度与对应的编码算法的复杂度是一样的. 解码算
法 1.1 和 解码算法 2.1 的时间计算复杂度为 O(nm), 其余两个的时间计算复杂
度为 O(nk). 它们具有相同的空间复杂度, 均为 O(2n ).
第三章 空间填充曲线
§3.4
33
希尔伯特曲线的新变体
由前面的分析可知, Hnm 可以由 Hnm−1 得到, Hnm 的第 0 象限包含的并不是
Hnm−1 本身, 而是经过交换指令变换得到的曲线. 用二维曲线做例子, 如图 3.8 所
示, H22 的第 0 象限中的曲线并不是 H21 , H23 的第 0 象限的曲线并不是 H22 . 这就说
明, 同一个坐标在不同阶的曲线中的序可能是不同的. 当曲线的阶数改变时, 需
要重新计算希尔伯特序.
Hnm 第 0 象限的曲线由 Hnm−1 经过一次交换指令得到, 而交换变换的逆变换
是自身，即同一个交换指令执行两次保持曲线不变. 从图 3.8 可以看出, H11 恰巧
是 H23 的起始曲线. 这说明只要对奇数或者偶数阶的曲线做一次 xn 与 x1 间的交
e m 第 0 象限的曲线是 H
e m−1 自身. 本文选择对偶数
换变换, 就能保证变体曲线 H
n
n
阶曲线进行一次 xn 与 x1 之间交换变换, 得到了希尔伯特空间填充曲线的一个变
e m . 在曲线 H
e m 中, 同一坐标在不同阶曲线中对应的序都是相同的, 同一个
体H
n
n
序对应的坐标值也是一样的. 同样取二维空间做例子, 如图 3.9 示, 可以看到此时
e 3 第 0 象限的曲线是 H
e2 , H
e 2 第 0 象限的曲线是 H
e 1.
H
2
2
2
2
图 3.8 希尔伯特曲线 (H21 , H22 , H23 )
下面给出该变体曲线的编码及解码算法. 由于仅偶数阶希尔伯特曲线被实施
了一次交换变换, 奇数阶曲线没有改变, 可以这么做: 奇数阶的曲线没有被改变,
编码解码算法和原来的算法相同; 对偶数阶的曲线, 可以先实施一次交换变换回
到原来的希尔伯特曲线, 然后调用原来的算法. 这样的好处是不用重写各种算法,
不用重新计算不同象限的变换指令.
由于变体曲线的算法与传统的算法非常相似, 本文仅给出一个编码算法和解
学位论文
34
e 2, H
e 3)
e 1, H
图 3.9 希尔伯特曲线变体 (H
2
2
2
码算法作为代表, 其他基于位运算或者复杂度更低的算法, 都可以类似得到. 算
法描述如下:
编码算法 3.1.
(1) 如果 m 偶数, 交换 xn 与 x1 .
m
m
(2) 如果 m = 0, 程序终止. 否则, rm = fn (xm
n , xn−1 , · · · , x1 ).
m−1 .
(3) 对每个整数 i (1 ≤ i ≤ n), 如果 xm
i = 1, 那么 xi = xi − 2
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Gnrm ,1 的第 i 个分量是 1, 那么 xi =
2m−1 − 1 − xi .
(5) 交换. 如果 Grnm ,0 存在两个非零分量, 其位置为 i 和 j, 则交换 xi 和 xj .
(6) m = m − 1, 转至 (2).
解码算法 3.1.
(1) 初始化 (xn , · · · , x1 ), (xn , · · · , x2 , x1 ) = bn (r1 ). v = 2.
(2) 如果 v > m, 程序终止. 否则 (sn , · · · , s2 , s1 ) = bn (rv ).
(3) 交换. 如果 Grnv ,0 在位置 i 及 j 有非零值, 则交换 xi 与 xj .
(4) 反转. 对每个整数 i (1 ≤ i ≤ n), 如果 Grnv ,1 的第 i 分量为 1, 则 xi = 2v−1 −1−xi .
(5) 对每个整数 i (1 ≤ i ≤ n), 如果 si 等于 1, 则 xi = xi + 2v−1 .
(6) v = v + 1, 转至 (2).
(7) 如果 m 偶数, 交换 xn 与 x1 .
第三章 空间填充曲线
35
这两个算法的时间计算复杂度均为 O(nm), 它们的空间复杂度为 O(2n ). 这
两个算法需要的空间变换信息和前一节的是一样的. 复杂度更低的算法亦可以类
似得到, 这里不再给出.
§3.5
算法的具体化
前几节对希尔伯特空间填充曲线做了理论分析并给出了算法模板, 适用于任
意维数的希尔伯特空间填充曲线. 下面对算法做具体化, 同时提供一些程序优化
建议. 我们仅选取两个算法模板做例子, 出于简单的考虑, 将选择二维希尔伯特空
间填充曲线, 高维空间是完全类似的. 选取编码算法 1.1 和 解码算法 1.1.
现代计算机内存容量较大, 一个常见的编程思想就是通过增加内存使用而加
i,1
n
快程序运行速度. 上述所有算法都需要计算 bn , fn , Gi,0
n 及 Gn (0 ≤ i < 2 ), 对特
定的维数 n, 这些信息是保持不变的, 可以把这些信息事先计算好存储在源文件
中, 在计算时直接查询, 这种技术会大大提高程序的运行速度.
2
1
(0, 1)
(1, 1)
(0, 0)
(1, 0)
0
3
Y
X
图 3.10 希尔伯特单元 (H21 )
在上节的算法描述中使用 (x2 , x1 ) 表示坐标, 本文用更符合习惯的 (x, y) (x =
x2 , y = x1 ) 代替, 先给出二维希尔伯特单元 H21 , 如图 3.10 所示, 并给出了曲线方
向及每个点的整数坐标. 映射 b2 及 f2 根据定义可以直接求得, 它们存储在表 3.1
中, 同时标注在图 3.10 中. 变换指令和反转指令存储在表 3.2 中. 直接观察可以
学位论文
36
i,1
看出, f2 , b2 , Gi,0
2 及 G2 (0 ≤ i < 4) 按照数组的方式存储有较好的查询效率.
表 3.1 左边表格为映射 f2 , 右边表格为映射 b2
f2 (0, 0)
0
b2 (0)
(0, 0)
f2 (0, 1)
1
b2 (1)
(0, 1)
f2 (1, 1)
2
b2 (2)
(1, 1)
f2 (1, 0)
3
b2 (3)
(1, 0)
表 3.2 左边的表格为交换指令, 右边的表格为反转指令
G0,0
2
(1, 1)
G20,1
(0, 0)
G1,0
2
(0, 0)
G21,1
(0, 0)
G2,0
2
(0, 0)
G22,1
(0, 0)
G3,0
2
(1, 1)
G23,1
(1, 1)
观察 编码算法 1.1, 坐标 (xn , · · · , x1 ) 在步骤 (2) ∼ (4) 中被更新, 其中用了
三个步骤描述使算法清晰, 而在实际应用中, 可以把三个步骤组合到一块, 用一个
表达式来表示, 将成倍加快程序的速度. 记当前曲线的阶为 v, xnew , ynew 分别为
更新后的坐标. 取第三象限做例子, 在这个象限中, 先执行反转指令, 然后是交换
指令. 步骤 (2) 是一个消去 x, y 的高位有效位的过程, 有 (xv , y v ) = (1, 0), 所以
(xnew , ynew ) = (x − 2v−1 , y). 步骤 (3) 是反转变换, 有 G23,1 = (1, 1), 需要对两个坐
标做反转, 有 (xnew , ynew ) = (2v − 1 − x, 2v−1 − 1 − y). 步骤 (4) 是一个交换的过程,
v−1 − 1 − y, 2v − 1 − x). 类似地可以推导
由于 G3,0
2 = (1, 1), 最终有 (xnew , ynew ) = (2
出另外三个象限的更新规则. 我们把这四个象限的更新规则总结到表 3.3 中. 对
解码算法可做类似的优化, 我们把解码算法的坐标更新规则总结在表 3.4 中. 在
实际计算中, 将 编码算法 1.1的步骤 (2) ∼ (4) 用一个坐标更新步骤代替, 将 解
码算法 1.1的步骤 (3) ∼ (5) 同样用一个坐标更新步骤代替. 至此便完成了二维空
间填充曲线编码解码算法的具体化。
§3.6
本章小结
本文简要地介绍了空间填充曲线的知识, 并在这些知识的基础上给出高维曲
线的编码解码算法. 对编码问题, 给出了四种算法, 两种基于算术运算, 两种基于
位运算, 它们的时间计算复杂度分别为 O(nm) 和 O(kn), 其中复杂度 O(kn) 与曲
第三章 空间填充曲线
37
表 3.3 编码过程的坐标更新规则
Quadrant
xnew
ynew
0
y
x
1
x
y − 2v−1
2
x − 2v−1
y − 2v−1
3
2v−1 − 1 − y
2v − 1 − x
表 3.4 解码过程的坐标更新规则
Quadrant
xnew
ynew
0
y
x
1
x
y + 2v−1
2
x + 2v−1
y + 2v−1
3
2v − 1 − y
2v−1 − 1 − x
线的阶数 m 无关, 在高阶编码解码应用中有优势. 解码问题是编码问题的逆问
题, 同样给出了四个算法.
此外, 本文还给出了希尔伯特曲线的一个变体, 该变体相对于传统的希尔伯
特曲线有一个优势, 即任意一个坐标或者序, 它对应的序或者坐标, 不随曲线的阶
数改变而变化. 该性质有助于在实际应用中减少计算时间. 相应的, 给出了这个
变体的编码解码算法.
最后通过二维的例子展示了将本章给出的算法具体化的过程, 并给出了程序
优化的建议.
38
学位论文
第四章
PHG 动态负载平衡模块的设计与实现
偏微分方程的并行求解, 关键问题之一是网格划分, 它不仅要求每个进程拥
有相等的计算负载, 同时要求有良好的划分质量, 以减少进程间通信. 在自适应
有限元计算过程中, 网格/基函数不断调整, 会导致负载不平衡, 必须动态地调整
网格分布, 从而实现动态负载平衡. 动态负载平衡算法需要速度快、网格划分质
量高、具有增量性质并且是并行运行的. 本章介绍 PHG 中动态负载平衡的算法
及实现.
PHG 的层次网格结构与动态负载平衡
§4.1
PHG 的计算网格为协调四面体网格. PHG 采用的是最新顶点二分加密算法,
存储了四面体单元所有的前辈, 以二叉树的形式表示. PHG 将网格分布到不同进
程中, 每个进程拥有一个子网格, 子网格的单元 (叶子) 是互不重叠的, 同时进程
中存储了单元所有的前辈, 形成一个分布式层次网格结构. 本节介绍 PHG 的层
次网格结构、网格划分及分布式层次网格.
§4.1.1
二分自适应网格加密、放粗与 PHG 的层次网格结构
PHG 的自适应网格加密方法为最新顶点二分加密算法 [1], 也称为“边加密”
算法. 对一个单元加密时, 将它一条边 (称为加密边) 的中点与与之相对的两个顶
点相连, 将单元一分为二, 如图 4.1 所示. 一个单元被加密后, 加密产生的单元称
为它的子单元, 该单元称为加密所产生的子单元的父单元.
Parent
Child 0
Child 1
图 4.1 四面体二分加密
39
学位论文
40
PHG 的四面体网格经过多次二分加密后, 形成一个以单元为结点的二叉树
结构, 一个单元经过加密后产生两条分枝, 分别对应它的两个子单元. 所有叶子
单元构成当前网格层, PHG 保证当前网格层总是协调的. 下面用二维协调三角形
网格做例子 (三维协调四面体网格的二叉树结构与二维协调三角形网格相同), 如
图 4.2 所示. 在这个例子中, 初始网格包含一个单元 e0 , 将它加密一次得到两个新
单元 e1 和 e2 , 再将 e1 和 e2 各加密一次得到四个新单元 e3 , e4 , e5 , e6 , 最后将 e3 ,
e4 和 e5 各加密一次得到新单元 e7 , e8 , e9 , e10 , e11 和 e12 , 最终网格由 e6 , e7 , e8 , e9 ,
e10 , e11 和 e12 构成.
V
V1
e0
e1
e4
V2
e5
e2
e3
e6
V
e0
e1
e10 e12
e3
e9 e11
e7
e8
e2
e6
e7
e4
e8
e9
e5
e10
e11
e6
e12
图 4.2 层次网格结构的二叉树表示
层次网格结构简单, 数据组织直观, 直接反映了 PHG 的内存管理模式. PHG
设计了单元数据结构 SIMPLEX, 提供了子单元成员分别指向两个子单元, 通过这个
成员实现了二叉树. PHG 实现了深度优先搜索算法遍历整个加密树, 同时提供了
函数接口访问所有网格单元; PHG 提供了邻居成员指向各个面对应的邻居单元,
通过指针实现邻居单元的快速访问; PHG 提供了 parent 成员指向父单元, 对根
单元而言 parent 为空指针, 在该成员的帮助下, 可以实现自下而上的回溯访问.
§4.1.2
网格划分
PHG 采用的划分方法是以单元为单位将网格划分为子网格, 每个进程负责
一个子网格, 每个进程仅存储它所拥有的子网格上的数据, 不在本地的数据通过
进程间通信获得. 对于传统并行有限元, 由于计算过程中网格保持不变, 仅需做一
次网格划分, 划分质量要求很高, 但对划分时间的限制较少. 而对并行自适应有
限元或者涉及时间的问题, 网格/基函数根据数值解调整, 会导致负载不平衡, 程
序需要根据计算负载对网格分布做动态调整. 在这类应用中, 网格划分方法需要
第四章 PHG 动态负载平衡模块的设计与实现
41
满足以下条件: 1) 能在分布式数据上并行执行; 2) 执行速度快, 因为它们可能会
被频繁调用; 3) 能产生高质量的划分; 4) 最好是增量的, 即网格的微小变化仅导
致划分的微小变化; 5) 内存使用是适量的.
并行计算中的数据划分方法可以分为两类: 基于几何性质的划分方法和基于
图的划分方法 [125, 126, 127, 128, 129]. 基于几何性质的划分方法是利用对象 (粒
子、单元等) 的几何信息划分数据, 如坐标, 常用的方法有递归坐标二分方法 RCB
(Recursive Coordinate Bisection) [60, 140, 141] 、递归惯量二分方法 RIB (Recursive
Inertial Bisection) [61, 142]、空间填充曲线划分方法 SFC (Space-Filling Curve) [62,
130, 131, 145, 146] 等. 基于图的划分方法是利用对象的拓扑信息, 如网格中单元的
邻居关系, 常用方法有递归图二分方法 (Recursive Graph Bisection) [60]、贪婪算法
[63]、递归谱二分方法 RSB (Recursive Spectral Bisection) [60, 142, 143, 144]、K-L 算
法 (Keinighnan-Lin Algorithm) [64]、多水平方法 (multilevel methods) [16, 135, 136,
137, 132, 133, 134] 和扩散方法 (diffusive methods) [18, 138, 139] 等. 图方法和几何
方法各有优缺点. 图方法运行时间长, 由于显式地控制通信量, 划分质量高. 几何
方法在空间局部性很重要或者拓扑结构不存在的情况下是非常有效的, 几何方法
没有显式控制通信量, 它只是根据空间位置信息划分, 可能会导致通信量很大, 但
几何划分方法简单, 它们容易实现并且速度快. 同时, 像递归坐标二分方法、空间
填充曲线划分方法等, 它们是隐式增量的 [125], 数据迁移的代价比较小. 在并行
自适应有限元计算中, 最终目标是极小化总体计算时间, 包括划分时间、数据迁
移时间及有限元求解时间. 它们是相关的: 质量高的划分会减少有限元计算时间,
但可能会需要更多的划分时间; 质量差的划分需要时间少, 但可能增加有限元计
算时间. 在实际应用中, 需要根据实际情况综合考虑这些因素. 下面给出几种常
用划分方法的细节:
• 递归坐标二分方法 (RCB) [60, 140, 141]. 交替地在每个空间方向上用一个平
面 (3D) 或者一条直线（2D) 将对象分成两个子集, 使得两个子集中的对象数
目 (或权重) 是近似相等的. 划分的平面或者直线垂直于给定的方向. 对于一
个待划分的集合, RCB 选择最长方向进行分割. RCB 方法没有考虑集合的形
状及长宽比. Jones 和 Plassmann[141] 对这个算法做了改进, 他们将新的方法
叫做 URB (Unbalanced Recursive Bisection). 由于通信量与长宽比有很大的关
系, URB 将待分集合分成两个长宽比接近的集合. 假设有 n 个进程, URB 会
从子集大小比例为 1 : (n - 1), 2 : (n - 2), 3 : (n - 3) 等划分中选择满足上述条
件的划分. Jones 和 Plassmann 在文章中指出该划分会降低通信量. RCB 及
URB 是隐式增量的.
42
学位论文
• 递归惯量二分方法 (RIB) [61, 142]. RIB 也是递归地选择一个平面或者直线将
对象集合分成两个大小相等的子集, 但它选择的平面或者直线不是垂直于坐
标轴的, 而是根据对象的分布, 在空间中选择一个最长方向, 用一个垂直于这
个方向的平面或者直线将对象分成两个子集. 该方法对惯量轴敏感, 因而不
是增量的划分方法.
• 空间填充曲线划分方法 (SFC) [62, 130, 131, 145, 146]. 该方法由 Patra 和 Oden
首先应用在网格自适应加密问题中 [145, 146]. SFC 方法是根据空间填充曲线
将 n- 维问题映射成 1- 维问题, 然后对 1- 维问题进行划分. 对于映射, 可以选
择希尔伯特序, Morton 序等. 该方法是增量的, 方法简单, 容易实现. SFC 方
法划分质量不如图划分质量高, 但运行速度快, 在动态负载平衡中很有竞争
力.
• 多水平方法 (multilevel method) [16, 132, 133, 134, 135, 136, 137]. 多水平方法
是图划分中最流行的方法. 多水平方法类似 V- 循环多重网格方法, 通过将一
些顶点缩成一个顶点, 将图放粗成一个更小的图, 形成一个放粗序列, 然后在
最小的图上进行划分, 最后通过加密过程得到划分结果.
• 递归谱二分方法 (RSB) [60, 142, 143, 144]. 谱划分方法先构造图的 Laplacian
矩阵, 然后求该矩阵的第二小的特征值对应的特征向量 (Fiedler vector), 每个
顶点给定一个对应于这个特征向量的值, 从而将图划分问题变成一维划分问
题. 谱方法划分质量高, 但特征值特征向量的求解代价高, 一般用在静态划分
中.
网格划分是并行计算的核心算法之一. 常用的图划分及几何划分软件如下:
• PaToh[151]. PaToH (Partitioning Tools for Hypergraph) 是一款串行多水平超图
划分软件, 号称是最快的串行超图划分软件. 它可以划分普通的超图, 也可以
划分带有多个限制条件的超图.
• Chaco[152]. Chaco 是一个串行多水平图划分软件. 主要功能有: 扩展了谱方
法, 允许使用两个或者三个 Laplacian 特征向量将图四等分或者八等分, 并为
谱方法提供了高效健壮的特征值特征向量解法器; 提供了通用的 KernighanLin/Fiduccia-Mattheyses 算法, 可以处理带权图; 改进了图和并行体系结构之
间的映射算法; 提供了各种后处理功能以改进划分质量.
• SCOTCH[154] 及 PT-SCOTCH[153]. SCOTCH 是一个串行软件, 提供了图划
第四章 PHG 动态负载平衡模块的设计与实现
43
分、静态映射、稀疏矩阵排序、网格和超图划分; PT-SCOTCH 是 SCOTCH
的并行化版本.
• Parkway[150]. Parkway 是一个并行超图划分软件, 它基于 C++ 和 MPI 实现
了多水平超图划分算法.
• METIS[17], hMETIS[155] 及 ParMETIS[14]. METIS (Serial Graph Partitioning
and Fill-reducing Matrix Ordering) 是一个串行程序集合, 提供了图划分、有
限元网格划分以及稀疏矩阵填充约化 (fill-reducing) 排序功能. METIS 中实
现的算法是基于多水平二分递归算法、k- 路多水平算法及多限制划分算法
等. ParMETIS (Parallel Graph Partitioning and Fill-reducing Matrix Ordering)
是 METIS 的并行版本, 基于 MPI, 适合并行自适应计算及大规模并行处理.
hMETIS (Hypergraph & Circuit Partitioning) 是一个串行超图划分软件, 它基
于多水平超图划分算法实现.
• Zoltan [19, 18]. Zoltan 是美国 Sandia 国家实验室开发的一个软件, 它包含一
系列并行工具以简化并行程序的编程. 它提供了几何划分、图划分、超图划
分功能, 同时提供了数据迁移工具、并行图染色、非结构通信服务及动态内
存管理等. 它为程序在不同划分算法之间切换提供了方便.
• JOSTLE [65]. JOSTLE 是为在分布式并行机上划分非结构网格 (有限元网格、
有限体积网格等) 设计的, 同时可以对网格进行重划分. JOSTLE 将网格看成
非结构图, 使用最新的图划分算法实现, 提供了多水平图划分算法及扩散算
法, 运行速度很快, 拥有串行和并行两个版本.
§4.1.3
PHG 的分布式二叉树
PHG 的网格是分布式存储的, 网格划分基于单元 (二叉树中的叶子) 进行, 当
用 p 个进程并行处理时, PHG 将网格划分成 p 个子网格, 即将网格中单元集合分
解成 p 个互不相交的子集, 每个子集构成一个子网格. 相应地, 描述网格层次结构
的二叉树被划分成 p 棵子树, 每棵子树包含子网格中的叶子单元和它们所有的前
辈 (上层单元). 图 4.3 是将图 4.2 中的网格划分为两个子网格 (p = 2) 时的子网
格和子树示意图, 其中第一个子网格包含叶子单元 e7 , e8 和 e9 , 第二个子网格包
含叶子单元 e6 , e10 , e11 和 e12 . 在不同子树中, 叶子单元互不相同, 但非叶子单元
则可能有重复. 这种子树结构既方便在网格层次间进行遍历, 又具有并行可扩展
性. 缺点是动态负载平衡时需要迁移子树及进行子树的重组, 大大增加了动态负
载平衡实现的复杂度.
学位论文
44
e0
e10 e12
e9
e7
e8
submesh 0
e0
e1
e11
e1
e3
e6
e4
e7
submesh 1
e8
e9
e2
e4
e5
e10
subtree 0
e11
e6
e12
subtree 1
图 4.3 分布式层次网格
§4.1.4
动态负载平衡
在自适应有限元计算中, 网格/基函数调整后, 如果进程的计算负载不平衡,
需要重新调整网格和数据分布, 使得各个进程保持相等量的计算负载, 保证每个
进程都处于繁忙状态, 同时尽可能减少它们之间的通信. PHG 中动态负载平衡与
有限元计算的关系如图 4.4 所示. PHG 中, 负载不平衡因子定义如下,
LIF =
Pp−1
Wi
,
p × max{W0 , · · · , Wp−1 }
i=0
(4.1)
其中 Wi 表示进程 i 的计算负载. PHG 的动态负载平衡主要由如下几个步骤构
成:
(1) 检查网格在各进程中的分布是否平衡, 即判断 LIF 是否大于事先设定的阈值.
如果 LIF 大于事先设定的阈值, 则调用相关的函数重新进行网格划分; 否则,
则继续使用当前的网格划分;
(2) 网格划分: 调用划分程序重新划分网格, 使得每个进程拥有相等的计算负载,
同时使得进程间的通信量最少.
(3) 子网格 – 进程映射: 对新的划分, 计算子网格到进程的映射, 使得从老划分到
新划分的数据迁移量最少.
(4) 网格单元迁移和二叉树重组: 网格分布发生了改变, 将单元按新划分发送到
所属的进程, 重组子网格和二叉树.
(5) 自由度数据迁移: 将单元上的自由度数据迁移到对应的进程上, 重建有限元
空间, 维护数据的正确性和完整性.
本节仅给出动态负载平衡的几个主要步骤, 后面内容将详细介绍网格划分算
法、子网格 – 进程映射、数据迁移等.
第四章 PHG 动态负载平衡模块的设计与实现
45
图 4.4 PHG 中的动态负载平衡
§4.2
网格划分方法
PHG 提供了与软件包 ParMETIS[14, 16] 和 Zoltan[18, 19] 的接口, 前者提供
了多水平图划分方法, 后者提供了超图划分方法以及常用的几何划分方法, 如递
归坐标二分方法、递归惯量二分方法、希尔伯特空间填充曲线划分方法. 除此之
外, 我们在 PHG 中实现两种网格划分方法: 加密树划分方法 [83, 84] 和空间填充
曲线划分方法 [62, 125, 126, 127, 128]. 本节下面部分对这两个方法及实现进行介
绍.
学位论文
46
§4.2.1
加密树划分方法
PHG 存储自适应过程中生成的加密树, 利用加密树划分方法实施网格划分是
一个很自然的想法. 加密树划分方法 (refinement-tree partition method) 由 William
Mitchell[84] 提出, 该方法基于加密/放粗过程中产生的二叉树, 它调用深度优先遍
历算法访问二叉树, 按照访问顺序对网格单元排序. 二叉树的遍历需要满足条件:
先访问左孩子, 然后是右孩子. 由于要求当前叶子结点与下一个被访问的叶子结
点有一个共享面, 加密树划分方法有良好的划分质量. 取二维区域做例子, 如图
4.5 所示, 这棵加密树与图 4.2 中的加密树是相同的, 只是调整了孩子结点的顺序,
对这棵加密树采用常规的深度优先遍历, 按照访问顺序相邻两个叶子单元之间有
一个公共边 (三维为面). 如果加密树不能使上述条件成立, 则需要动态地调整孩
子结点的访问次序使条件成立.
V
V1
e0
e1
e4
V2
e5
e2
e3
e6
V
e0
e1
e10 e12
e3
e9 e11
e7
e8
e2
e6
e7
e4
e8
e9
e5
e10
e11
e6
e12
图 4.5 二分加密产生的加密树
在加密树中按下面方法给每一个结点赋予一个权重 w: 叶子结点的权重由
用户给定, 非叶子结点的权重为以该结点为根的子树的所有叶子结点的权重和.
Mitchell 的算法由两步组成, 第一步先计算每一个结点的权重, 第二步将加密树二
分, 使得每个集合 (set) 中的叶子数目相等, 递归调用二分算法来完成网格划分.
按照 Mitchell 的分析, 算法的计算复杂度为 O(N log p + p log N ), 其中 N 为叶子结
点数目. 细节如算法 2 所示.
上述算法需要计算每一个结点的权重, 由于父单元在多个进程中同时存在,
在计算父单元的权重时, 通信比较复杂. 我们重新设计了算法, 为每一个叶子结
点定义前缀和 (prefix sum). 按照叶子结点的访问顺序, 叶子结点的前缀和定义为:
在到达该叶子结点之前访问的所有叶子结点的权重之和. 下面给出形式化定义,
第四章 PHG 动态负载平衡模块的设计与实现
47
算法 2: 二分加密树划分算法
algorithm bisect
compute subtree weights
bisect subtree(root)
end algorithm bisect
algorithm bisect subtree(node)
if node is a leaf then
assign node to the smaller set
elseif node has one child then
bisect subtree(child)
else node has two children
select a set for each child
for each child examine the sum of the subtree weight with
the accumulated weight of the selected set
for the smaller of the two sums, assign the subtree rooted
at that child to the selected set and add the subtree weight
to the weight of the set
bisect subtree(other child)
endif
end algorithm bisect subtree
按照叶子结点的访问顺序, 为每一个叶子结点赋予一个位置编号, 从 0 开始, 则第
i 个叶子结点的前缀和为
Si =
i−1
X
wj
(4.2)
j=0
其中 wj 为叶子结点 j 的权重, 式 (4.2) 可以改写为
Si =
i−1
X
j=0
wj = Sq +
i−1
X
wj , (q < i).
(4.3)
j=q
假设所有叶子结点的权重之和为 W , 进程数为 p, 那么所有前缀和属于区间 [W
i/p, W (i+1)/p) 的叶子单元分配给子网格 i (0 ≤ i < p). 通过分析可以看出, 只要
计算出叶子结点的前缀和, 就可以通过计算叶子结点所属区间来完成划分.
学位论文
48
当网格位于一个进程中时, 仅需遍历一次加密树即可完成划分, 算法复杂度
为 O(N ). 当网格分布在不同进程中时, 假设进程数为 p, 每个进程含有的叶子结
点数目分别为 ni (0 ≤ i < p), 进程 i 上叶子结点权重之和为 Wi (0 ≤ i < p). 定义
叶子单元在本进程中的局部位置编号, 即叶子结点按照在本进程中访问顺序从 0
开始的编号. 定义 Si,j 为进程 i 上局部编号为 j 的叶子单元的前缀和, 那么由式
(4.3) 可以推导出
Si,j =
i−1
X
k=0
Wk +
j−1
X
wj = Si,j−1 + wj−1 ,
(4.4)
k=0
上式表明, 只要知道了 Wi (0 ≤ i < p) 就可以在每一个进程中计算出所有叶子结
点的前缀和. 在每个进程上, 遍历一次加密树就可以知道本进程所有叶子结点的
权重和 Wi (0 ≤ i < p), 然后通过第二次遍历就可以同时完成计算前缀和及网格
划分. 算法如下所示.
算法 3: PHG 实现的加密树划分算法
Step 1. 每个进程访问本地的子树, 计算所有本地叶子结点的权重和 Wi .
Step 2. 调用 MPI_Scan 操作, 为每个进程收集其需要的 Wi (0 ≤ i < p).
Step 3. 访问本地子树, 根据式 (4.4) 计算每一个叶子结点的前缀和, 并同时完成划
分.
本文设计的算法简洁, 仅需遍历两次加密树以及一次 MPI_Scan 通信, 算法的
计算复杂度为 O(N ). 对加密树划分算法, 由于初始网格会包含许多单元, 因而也
会有多个加密树, 加密树的访问顺序是按照根结点的顺序, 只需对初始网格 (根结
点) 排序, 并在整个自适应过程中维护这个序, 以保证在整个自适应过程中访问子
树的顺序是相同的. PHG 提供了三种单元排序方式: 哈密尔顿路排序, 希尔伯特
序和 Morton 序. 哈密尔顿路排序是将单元按照第二章设计的哈密尔顿路算法排
序; 希尔伯特序和 Morton 序方法是将单元的重心坐标按照这两种序的定义映射
到一维区间, 在一维区间上按照数值大小排序.
函数 phgPartitionRTK 实现了加密树划分算法, 接口如下:
BOOLEAN phgPartitionRTK(GRID *g, int np, DOF *w, FLOAT p),
其中 np 给出子网格数量, 自由度对象 w 给出单元的权重, p 为权重的指数, 单元
e 实际的权重为 wep .
PHG 设计了如下辅助函数用于对初始网格排序:
第四章 PHG 动态负载平衡模块的设计与实现
49
(1) 函数 phgHamiltonPath 实现了哈密尔顿路排序, 接口如下:
int phgHamiltonPath(GRID *g),
其中返回 0 表示成功, 非 0 表示程序失败;
(2) 函数 phgHilbertOrder 实现了希尔伯特序排序, 接口如下:
int phgHilbertOrder(GRID *g),
其中返回 0 表示成功, 非 0 表示程序失败;
(3) 函数 phgMortonOrder 实现了 Morton 序排序, 接口如下:
int phgMortonOrder(GRID *g),
其中返回 0 表示成功, 非 0 表示程序失败;
(4) 函数 phgGridInitOrder 负责调用上面三个函数之一完成对初始网格的排序,
接口如下:
int phgGridInitOrder(GRID *g, BOOLEAN dist),
其中 dist 变量表明网格是否为分布式的, 返回 0 表示成功, 非 0 表示程序失
败.
目前上述几个排序程序都是串行的.
§4.2.2
空间填充曲线划分方法
网格划分中利用空间填充曲线将高维空间映射成一维空间, 完成单元的排
序, 从而将高维划分问题转化为一维划分问题. Morton 空间填充曲线和希尔伯特
空间填充曲线均具有较好的空间局部性, 划分质量比较好, 是最常用的两种曲线
[62, 130, 131, 145, 146].
空间填充曲线划分方法分为三个步骤. 第一个步骤是计算空间填充曲线, 首
先将计算区域映射到立方体 (0, 1)3 , 然后调用空间填充曲线生成算法将立方体映
射到区间 (0, 1); 第二个步骤是划分区间 (0, 1), 使得属于每个区间单元的权重和
相等; 第三个步骤是调用子网格 – 进程映射算法, 对子网格重新编号以减少数据
迁移.
对计算区域 Ω, 存在一个长方体 (三维, 在二维情况下为长方形) 包围盒 (bounding box) 包含这个区域, 假设包围盒在 x, y, z 三个方向的长度分别为 lenx , leny , lenz ,
学位论文
50
将区域映射到 (0, 1)3 的通常的做法是
x1 = (x − x0 ) / lenx , y1 = (y − y0 ) / leny , z1 = (z − z0 ) / lenz ,
其中 (x0 , y0 , z0 ) 是包围盒坐标最小的顶点, (x1 , y1 , z1 ) 为坐标 (x, y, z) 经过变换后
在立方体 (0, 1)3 中的坐标. 上述变换改变了区域的长宽比, 使得变换后的长宽比
为 1 : 1 : 1, 破坏了区域的空间局部性. PHG 中取 len = max(lenx , leny , lenz ), 采用
变换
x1 = (x − x0 ) / len, y1 = (y − y0 ) / len, z1 = (z − z0 ) / len,
该变换保持区域的空间局部性, 改善了划分质量, 这一点将在后面用数值算例加
以验证.
对空间填充曲线划分算法, 我们提供了两种空间填充曲线生成程序: Morton
空间填充曲线和希尔伯特空间填充曲线. 前者的算法简单, 但曲线本身有较大的
跳跃, 使得空间局部性略差. 希尔伯特空间填充曲线的空间局部性要好很多, 但
生成算法复杂.
利用空间填充曲线对单元排序, 然后调用 §4.2.3 中介绍的一维划分算法及
§4.2.4 中介绍的子网格 – 进程映射算法便可完成网格划分. 我们在 PHG 中实现
了下述空间填充曲线生成及网格划分函数:
(1) 函数 phgSFCInvHilbert2D 及 phgSFCInvHilbert3D 实现了二维和三维希尔伯
特空间填充曲线向一维区间 [0, 1] 的映射, 接口如下:
BOOLEAN phgSFCInvHilbert2D(SFC *hsfc, INT n),
BOOLEAN phgSFCInvHilbert3D(SFC *hsfc, INT n),
其中数组 hsfc 存储需要转换的坐标 (∈ [0, 1]3 ) 和映射, n 为数组的长度, 数据
结构 SFC 定义如下:
typedef struct {
SFC_FLOAT co[3];
SFC_FLOAT sfc;
INT index.
} SFC;
其中成员 co[3] 给出空间点的坐标 (∈ [0, 1]3 ), sfc 返回映射结果, 辅助成员
index 存储着单元的本地索引.
第四章 PHG 动态负载平衡模块的设计与实现
51
(2) 函数 phgSFCInvMorton2D 及 phgSFCInvMorton2D 实现了二维和三维 Morton 空
间填充曲线向一维区间 [0, 1] 的映射, 接口如下:
BOOLEAN phgSFCInvMorton2D(SFC *msfc, INT n);
BOOLEAN phgSFCInvMorton3D(SFC *msfc, INT n);
其中参数 msfc 和 n 的意义与希尔伯特空间填充曲线相同. 这两种曲线的阶
数都是 25, 在源文件中直接修改 MAXLEVEL 的数值可以获得阶数 20 到 30 之
间的空间填充曲线.
(3) 函数 phgPartitionSFC 实现了空间填充曲线划分方法, 接口如下:
BOOLEAN phgPartitionSFC(GRID *g, int np, DOF *w, FLOAT p);
其中 g 为网格信息, 划分编号直接存储在单元 e 的成员 mark 中, np 为子网
格数量, 自由度对象 w 为权重, p 为权重的指数, e 的实际权重取为 wep .
§4.2.3
一维划分算法
上节的空间填充曲线划分方法及其他一些网格划分方法最终均将问题转化
为一维划分问题, 因此实现一个高效健壮的一维划分程序是必要的.
假设当前进程数为 p 以及给定区间 [a, b), 这里的一维划分问题为: 如何将 [a,
b) 划分成 p 个子区间 [a, a1 ), [a1 , a2 ), · · · , [ap−2 , ap−1 ), [ap−1 , b), 使得每个区间上包
含的对象的权重和相等.
从问题的描述可以看出, 一维划分需要计算 p − 1 个分割点 a1 , a2 , · · · , ap−2 ,
ap−1 . 我们在 PHG 中实现了一个一维划分算法求解该问题, 其基本算法来源于
Zoltan[18], 它是二分搜索算法的推广. 在二分搜索算法中, 只需求解一个点 a1 , 将
区间分成两个子区间. 如果将区间 k (k ≥ 2) 分而不是二分, 那么 a1 将位于某个
子区间中, 接着将这个子区间 k 分, 重新得到 [a, b) 的一个划分, a1 将位于上述 k
个更小的子区间中的一个. 重复这个过程, 可以在给定的精度下搜索到 a1 . 将这
个算法推广, 将区间 [a, b) 划分成 N (N À p) 个子区间, ai (0 < i < p) 将位于某
个子区间, 然后将子区间细分, 最终在给定的精度下可以搜索到 a1 , a2 , · · · , ap−2 ,
ap−1 .
具体实现时, 我们取子区间数为 N = (p − 1) ∗ k + 1, 程序为每一个划分点
ai (0 < i < p) 维护一个包围盒 (bounding box). 在每一次划分之前先更新这些
包围盒, 缩小 ai (0 < i < p) 的取值范围, 重新划分区间 [0, 1) 的时候, 直接划
分 ai 所在的包围盒, 而不是 ai (0 < i < p) 所在的某个子区间, 这样会加速求解
ai (0 < i < p) 的过程. 对该一维划分算法, 我们实现了两个不同的版本: 第一个版
学位论文
52
本 phgPartition1DV1 的子区间数目始终为 N ; 第二个版本 phgPartition1DV2 是
一个自适应版本, 其子区间数目不是固定的, 当某个 ai (0 < i < p) 达到精度要求
之后, 其所在的包围盒将不再细分, 子区间数目是一个递减的过程, 通过减少不必
要的划分, 加快划分速度. 两个函数的接口如下:
FLOAT phgPartition1DV1(DOTS *x, INT lenx, int np, MPI_Comm comm),
FLOAT phgPartition1DV2(DOTS *x, INT lenx, int np, MPI_Comm comm).
其中 x 是对象数组, 包含当前进程所有对象的权重、坐标 (∈ [0, 1), lenx 是数组的
长度, np 是子网格数, comm 是当前的通信器. DOTS 的定义如下:
typedef struct {
FLOAT key;
FLOAT w;
int pn;
} DOTS;
其中 key (∈ (0, 1)) 是一维坐标, w (≥ 0) 是对象的权重, 通常情况下为正数, pn 是
对象所在划分的编号, 用于返回结果.
§4.2.4
子网格―进程映射
网格划分完成后, 需要将子网格映射到进程上, 该映射尽量使得从老划分到
新划分的数据迁移量最小 [148]. Oliker 和 Biswas 设计了一个启发式算法处理这
个问题, 他们的算法可以得到次优解 [72, 147, 148]. 我们在 PHG 中实现了该算法.
Oliker 和 Biswas 的算法先对数据迁移建模. 用相似矩阵 S (similarity matrix)
表示数据在所有进程中的分布情况. S 在 PHG 中是一个 pold × p 的矩阵, pold 是
当前子网格数, p 为新子网格数 (在通信器不变的情况下, pold 与 p 是相等的), 相
似矩阵 S 的元素 Si,j 表示编号为 i 的子网格中需要迁移到编号为 j 的子网格中
的数据量. 实际计算中, 每个进程并发地计算相似矩阵的一行, 然后通过一个主
进程收集这些信息创建一个相似矩阵, 根据这个相似矩阵计算子网格 – 进程映射
关系, 最后通过一个广播操作将映射关系发送给所有进程.
优化子网格 – 进程映射时, 需要针对不同的体系结构上需要建立不同的代价
方程 (cost function) 以优化数据迁移开销. 常用的两种度量是: TotalV 和 MaxV.
TotalV 极小化所有进程数据迁移量之和; MaxV 极小化单个进程的最大数据迁移
量. Oliker 和 Biswas 的算法中使用的度量是 TotalV. 给定一个相似矩阵 S, 其数
据总和是确定的, 极小化数据迁移与极大化保持数据不动是等价的, 目标是寻找
一个映射 i → pi (i = 0, · · · , p − 1) 极大化如下代价方程
第四章 PHG 动态负载平衡模块的设计与实现
F =
X
53
Si,j ,
pj =i
0≤j<p
其中 (p0 , p1 , · · · , pp−1 ) 是 (0, 1, · · · , p-1) 的一个置换.
Oliker 和 Biswas 设计了一个启发式贪婪算法解决该问题, 所得到的映射是次
优的, 算法的伪代码如下所示, 本文对算法稍微做了修改以适应 PHG 中的应用情
况.
算法 4: 启发式子网格―进程映射算法
for (j = 0; j < p; j++) part_map[j] = unassigned;
for (i = 0; i < p; i++) proc_unmap[i] = 1;
generate list L of entries in S in descending order;
count = 0;
while( count < p) {
find next entry S[i][j] in L such that
proc_unmap[i] > 0 and part_map[j] = 1;
proc_unmap[i] = 0;
part_map[j] = assigned;
count++;
map partition j to processor i;
}
我们在 PHG 中实现了三个子网格 – 进程映射函数, 接口分别如下:
int phgPartitionRemap(GRID *g, int nprocs, MPI_Comm comm),
int phgPartitionRemapV2(GRID *g, int nprocs, MPI_Comm comm),
int phgPartitionRemapV3(double ds[],int perm[],int np, MPI_Comm comm),
前两个函数中, g 为 PHG 的网格, 对其中每个单元 e, e->mark 进入函数时存储了
它在新划分中所属的子网格编号, 而从函数返回时则存储着重映射后的新编号.
学位论文
54
phgPartitionRemapV3 是一个通用的函数接口, ds 存储了当前进程迁往每个进程
的数据迁移量, 映射结果在 perm 中返回, np 为新划分的个数, 数组 ds 与 perm 的
长度均为 np. 该函数除可用于子网格 – 进程映射外, 还可用于计算其他一些应用
中优化通信的映射.
函数 phgPartitionRemap 和 phgPartitionRemapV2 的功能是一样的, 后者仅
存储相似矩阵的非零元, 有着更好的空间复杂度. 如果重映射后子网格编号发生
了变化, 则它们返回一个大于零的值; 如果编号没有变化, 则返回零; 如果计算失
败, 则它们返回小于零的值, 此时子网格编号保持不变.
§4.2.5
划分质量度量
对并行自适应有限元计算而言, 网格划分质量直接影响总体计算时间. 最重
要的划分质量度量指标是进程的计算负载的平衡. 此外, 还需要考虑极小化通信
代价, 该目标对应着极小化子网格共享面的数目, 因为共享面上的数据近似等于
计算中需要通信的数据. 在图划分中, 共享面与割边 (cut edge) 的概念是等价的.
这里, 我们采用两个度量指标来衡量划分质量: 表面比 (surface index) 和进程间
连通系数 (interprocess connectivity). 表面比衡量的是通信量的大小, 进程间连通
系数衡量的是需要通信的进程数 [62, 149].
表面比定义为子网格的边界面的数目与所有面的数目的比值 (在二维中为边
的数目比值). 如果共有 n 个子网格, bi 和 fi 分别代表子网格 i 中边界面的数目
和总面数, 那么最大局部表面比 (maximum local surface index) 定义为
bi
.
0≤i<n fi
rM = max
(4.5)
假如 bt 和 ft 分别表示所有子网格的边界面数目之和与整个网格中面的总数目,
那么可以定义全局表面比 (global surface index)
rG =
bt
.
ft
(4.6)
而平均表面比 (average surface index) 则定义为
n−1
rA =
1 X bi
.
n
fi
(4.7)
i=0
进程间连通系数是指与给定子网格相互连通的子网格的数量. 它衡量与该进
程进行通信的进程数目. 最大进程间连通系数 (maximum interprocess connectivity)
是程序可扩展性的一个重要衡量指标 [62].
第四章 PHG 动态负载平衡模块的设计与实现
55
PHG 的动态负载平衡模块
§4.3
前面介绍了 PHG 中的网格划分算法, 本节将介绍 PHG 的动态负载平衡模
块及相应的用户接口函数, 并给出一些实现细节.
§4.3.1
动态负载平衡函数
PHG 中, 网格的负载不平衡因子保存在网格对象中, 并由 PHG 自动计算及
维护. PHG 的动态负载平衡函数由用户调用, 一般在网格/基函数调整后立即调
用, PHG 的动态负载平衡函数接口如下:
int phgBalanceGrid(GRID *g, FLOAT lif, INT sm, DOF *w, FLOAT p),
其中 g 为网格对象; lif 是动态负载平衡的阈值, 网格的负载不平衡因子大于该
阈值则进行网格重划分; sm 用来控制子网格的数目, 使得子网格的平均单元数目
不小于这个值, 用户可以将其设置负值或者零, 此时表示子网格的最小平均单元
数目为 1; w 和 p 两个参数给出单元的计算负载, 即单元 K 的计算负载为 wpK , w
可以为 NULL, 表示所有单元的计算负载相等.
§4.3.2
网格划分和数据迁移
PHG 的网格对象与一个 MPI 通信器相关联, 通信器中的进程数等于当前网
格划分的子网格数, 当动态负载平衡函数被调用时, PHG 将首先根据 sm 参数、当
前网格中的单元数、最大可用的进程数等判断是否需要调整通信器, 如果需要将
创建一个包含更多进程的通信器. 接着调用划分程序划分网格, 划分结果直接保
存在每个单元中, 单元 e 的成员 mark 存储了该单元所属的子网格编号.
网格划分函数结束时, 返回一个逻辑值说明网格分布是否被改变, FALSE 表明
网格分布保持不变, TRUE 表明网格分布发生改变. 如果网格分布发生改变, 动态
负载平衡函数将调用数据重分布函数重组子网格以及迁移自由度数据等. 其中,
重组子网格包括迁移网格单元、为新的子网格建立相应的分布式二叉树、建立子
网格内部及子网格间的邻居关系等, 包含许多复杂的通信和处理, 在此不做详述.
§4.4
数值算例
本节给出数值算例, 比较不同划分方法的性能、划分质量以及对有限元计
算性能影响. 数值算例分为两个部分, 第一部分考察不同划分方法的划分质量,
衡量指标采用前面介绍的表面比 (surface index) 和进程间连通系数 (interprocess
connectivity). 第二部分使用自适应有限元方法求解偏微分方程, 比较不同划分方
学位论文
56
法对程序性能的影响.
§4.4.1
划分质量比较
本节比较不同划分方法划分质量的优劣, 划分方法包括: ParMETIS, HSFC
(PHG 的希尔伯特空间填充曲线划分方法), MSFC (PHG 的 Morton 空间填充曲线
划分方法), RCB (Zoltan 的递归坐标二分划分方法) 和 Zoltan 的 HSFC. 其中第
一个是图划分方法, 其余几个是几何划分方法.
本节使用了两个网格, 第一个网格 Ω1 是一个圆柱体, 如图 4.6 所示, 直径小
但长度较长, 有较大的长宽比, 网格的单元数目为 2,522,624. 第二个网格 Ω2 是一
个多孔薄板区域, 面积较大, 但很薄, 同时有很多的小孔, 如图 4.7 所示, 单元数目
为 3,713,792. 两个网格均由 Netgen[81] 生成.
图 4.6 测试网格 Ω1 (圆柱体)
表 4.1 给出各种划分方法产生的子网格的最大进程间连通系数. 对第一个
网格 Ω1 , ParMETIS 的划分质量最好, 其次是 RCB, MSFC, PHG/HSFC, 最后是
Zoltan/HSFC, 当子网格数目小于或等于 64 时, PHG/HSFC 的最大进程间连通系
数明显低于 Zoltan/HSFC. 对第二个网格结论是类似的. 需要说明的是这两个区
域比较规则, 很适合 RCB 这类算法.
表 4.2 给出子网格的最大表面比和平均表面比, 这两个指标直接衡量了通信
量的大小, 其中最大表面比衡量了单个进程最大的数据发送/接收量, 平均表面
比衡量了全局的数据发送/接收量, 比值越小表明划分质量越好. 从表中可以看出
ParMETIS 的边界面所占的比例始终最低, 说明其划分质量最好; Zoltan/HSFC 的
数值总是最大, 表明其划分质量在这几个划分方法里最差; MSFC 的划分质量比
PHG/HSFC 略差, 但同样优于 Zoltan/HSFC; 由于区域为规则区域, RCB 的划分
质量优于 PHG/HSFC, MSFC 及 Zoltan/HSFC. 从表 4.1–4.2 可以看出, 图划分方
法的划分质量始终好于几何方法的划分质量, 但前者计算划分的时间亦远大于后
者.
第四章 PHG 动态负载平衡模块的设计与实现
57
图 4.7 测试网格 Ω2 (多孔薄板)
§4.4.2
划分质量对自适应有限元计算性能的影响
本节将 ParMETIS, Zoltan/HSFC, PHG/HSFC, PHG/RTK (PHG 的加密树划
分方法), MSFC (PHG 中的 Morton 空间填充曲线划分方法), RCB (Zoltan 的递归
坐标二分方法) 六种划分方法应用到自适应有限元计算中, 以研究不同划分方法
对并行程序总体性能的影响. 我们将统计划分时间, 比较各个方法的划分速度; 统
计动态负载平衡时间, 比较划分质量对数据迁移的影响; 统计有限元离散线性方
程组求解时间, 比较划分质量对通信的影响; 统计每一个自适应步的总计算时间,
比较不同划分方法对并行程序总体性能影响. 本节将计算三个不同的算例, 分别
求解三种不同类型的偏微分方程.
测试的硬件平台为科学与工程计算国家重点实验室的浪潮 TS10000 高性能集
群 (LSSC-III), 计算结点为浪潮 NX7140N 刀片, 每个刀片包含两颗 Intel X5550 四
核处理器和 24GB 内存, 其单核双精度浮点峰值性能为 10.68Gflops, 282 个计算结
点的总浮点峰值性能为 24Tflops. 所有结点同时通过千兆以太网和 DDR Infiniband
网络互联.
学位论文
58
表 4.1 Ω1 与 Ω2 的最大进程间连通系数
Ω1
# submeshes
16
32
64
128
160
192
ParMETIS
2
2
5
8
8
9
RCB
2
2
5
11
12
13
MSFC
4
5
10
16
19
21
PHG/HSFC
3
6
13
23
24
23
Zoltan/HSFC
12
18
21
23
24
24
Ω2
# submeshes
16
32
64
128
160
192
ParMETIS
7
7
7
10
11
12
RCB
7
7
8
11
12
12
MSFC
9
13
18
21
22
25
PHG/HSFC
8
10
13
18
20
21
Zoltan/HSFC
12
19
23
25
27
31
例 4.4.1. 本例求解如下 Dirichlet 边界条件 Helmholtz 方程,
−∆u + u = f
u(x, y, z) = g
(x, y, z) ∈ Ω
on ∂Ω.
计算区域为圆柱体区域 Ω1 , 取解析解如下:
u = cos(2πx) cos(2πy) cos(2πz),
该算例中解是光滑的, 因此网格加密基本上是均匀的.
测试中使用了 32 个计算结点, 128 个进程. 计算区域如图 4.6 所示, 初始网
格 T1 由 Netgen [81] 生成, 含有 4,927 个四面体单元. 使用三阶协调拉格朗日元对
方程进行离散, 调用数值代数软件包 Hypre 的 BoomerAMG 解法器 [22] 求解有限元
离散产生的线性系统. 网格自适应使用一个后验型的误差指示子, 其定义如下
2
ηK
= h2K kfh + ∆uh + uh k2L2 (K) +
X
f ⊂∂K∩Ω
hf ∂uh 2
k[
]k 2 .
2 ∂nf L (f )
在网格划分时, 单元的权重取为 1, 为了观察网格划分, 负载不平衡因子阈值设置
得较小 (1.05), 划分次数较多, 有利于统计时间. 网格单元标记策略为最大标记策
略 (maximum strategy), 参数 θ 取 0.5 (参看 §5.1.3.2).
第四章 PHG 动态负载平衡模块的设计与实现
59
表 4.2 Ω1 与 Ω2 的表面比
maximum surface index (%, Ω1 )
# submeshes
16
32
64
128
160
192
ParMETIS
0.74 1.48 2.88 5.00 5.78 5.71
RCB
1.0
2.18 4.07 6.06 7.06 7.60
MSFC
4.2
6.51 10.6 16.2 18.9 20.7
PHG/HSFC
3.80 8.31 15.5 19.5 20.4 21.5
Zoltan/HSFC 13.9 19.1 25.4 28.1 30.9 35.8
average surface index (%, Ω1 )
# submeshes
16
32
64
128
160
192
ParMETIS
0.64 1.28 2.39 3.62 4.05 4.42
RCB
0.85 1.68 3.43 5.17 6.03 6.58
MSFC
2.93 5.14 7.44
PHG/HSFC
2.78 5.03 7.18 9.34 10.2 10.8
9.8
10.7 11.5
Zoltan/HSFC 8.72 11.8 16.2 20.6 22.0 23.6
maximum surface index (%, Ω2 )
# submeshes
16
32
64
128
160
192
ParMETIS
2.45 2.56 3.30 5.32 6.13 6.10
RCB
2.57 4.74 6.29 9.16 9.66 11.6
MSFC
3.0
PHG/HSFC
2.86 4.74 7.03 9.78 10.6 11.3
5.11 7.06 9.74 11.1 11.8
Zoltan/HSFC 8.89 12.3 16.5 19.5 21.2 22.8
average surface index (%, Ω2 )
# submeshes
16
32
64
128
160
192
ParMETIS
1.12 1.45 2.18 3.43 3.87 4.21
RCB
1.89 3.04 3.93 5.90 6.58
MSFC
2.24 3.61 5.21 7.34 8.22 8.92
PHG/HSFC
2.19 3.29 4.88 6.92 7.81 8.27
7.2
Zoltan/HSFC 5.10 6.86 9.47 12.6 14.1 15.3
学位论文
60
cylinder, 128 procs
0.35
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
0.3
Partitioning time (s)
0.25
0.2
0.15
0.1
0.05
0
0
2e+06
4e+06
6e+06
8e+06
1e+07
1.2e+07
1.4e+07
Number of elements
图 4.8 网格划分时间 (例 4.4.1)
图 4.8 给出划分时间, 从中可以看出 RTK 方法的划分速度最快, 其次是
MSFC, PHG/HSFC, Zoltan/HSFC. PHG 实现的 HSFC 及 MSFC 均比 Zoltan/HSFC
速度要快. ParMETIS 和 RCB 最慢. 从图上可以看出网格分布对 ParMETIS 的划
分时间有较大影响, 随着网格调整, 其划分时间出现剧烈震荡. 总体来讲, 几何划
分方法对网格分布不敏感, RCB, MSFC, PHG/HSFC, Zoltan/HSFC 的划分时间随
着网格规模的增大以较稳定的速率增加.
图 4.9 给出动态负载平衡时间, 其中包含了网格划分和数据迁移时间, 其中数
据迁移时间占的比重较大. 一个好的划分方法应该是增量的, 即网格的微小变化
仅导致划分的微小变化, 这样会使得数据迁移量比较小. 从图上可以看出 RTK 的
时间最短, 表明其数据迁移量较少, RTK 的时间增长也很稳定. 其次是 ParMETIS,
ParMETIS 同样表现出震荡. 接着是 MSFC, PHG/HSFC, RCB, 用时最长的是
Zoltan/HSFC. 它们的时间增长趋势均比较稳定, 说明这些方法均有较好的可扩
展性和比较低的计算复杂度.
一般情况下线性方程组的求解时间在自适应有限元计算中占统治地位, 图
4.10 和 4.11 分别给出线性方程组求解时间和每一个自适应步的时间. 从图中可
以看出, RCB, ParMETIS, RTK 运行时间最短. 对 RCB 而言, 其划分质量一般要比
ParMETIS 及 RTK 差一些, 这个算例是特例, 因为计算区域是一个长圆柱体, 非常
第四章 PHG 动态负载平衡模块的设计与实现
61
cylinder, 128 procs
2.2
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
2
time of dynamic load balancing time
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
2e+06
4e+06
6e+06
8e+06
1e+07
1.2e+07
1.4e+07
Number of elements
图 4.9 动态负载平衡时间 (例 4.4.1)
cylinder, 128 procs
180
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
160
time of solving linear system (s)
140
120
100
80
60
40
20
0
0
1e+07
2e+07
3e+07
4e+07
5e+07
Number of DOF
图 4.10 有限元离散系统求解时间, 横坐标为自由度数 (例 4.4.1)
6e+07
学位论文
62
cylinder, 128 procs
180
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
160
overall computational time (s)
140
120
100
80
60
40
20
0
0
1e+07
2e+07
3e+07
4e+07
5e+07
6e+07
Number of DOF
图 4.11 每一个自适应步时间 (例 4.4.1)
适合 RCB. 其次是 MSFC, PHG/HSFC, 用时最长的是 Zoltan/HSFC. Zoltan/HSFC
与 PHG/HSFC 是同一类型的划分方法, 由于 PHG 在映射时采取了保持区域长宽
比的方式, 维持了区域的空间局部性, PHG/HSFC 的划分质量比 Zoltan/HSFC 高,
求解时间也较少.
表 4.3 总体运行时间及划分次数 (例 4.4.1)
Method
total running time(s)
# of repartitionings
RCB
3049.60
60
ParMETIS
3360.73
189
RTK
3465.63
59
MSFC
4088.01
58
PHG/HSFC
4493.43
59
Zoltan/HSFC
8954.21
50
表 4.3 给出该算例采用不同划分方法时的运行时间及划分次数, 算例包含 190
个自适应步. RCB 运行时间最短, 说明对计算区域比较规则的问题, 简单的几何
划分方法是非常有效的. Zoltan/HSFC 的运行时间最长, 是别的方法的两倍以上,
第四章 PHG 动态负载平衡模块的设计与实现
63
表现最差. ParMETIS 的划分次数最多, 是别的方法的三倍. 划分次数过多会导致
数据迁移频繁, 当网络带宽较低时, 就需要考虑数据迁移问题, 但由于其划分质量
好, 线性方程组的求解时间较短.
例 4.4.2. 本例求解一个带有 Dirichlet 边界条件的 Poisson 方程
−∆u = f
u(x, y, z) = g
(x, y, z) ∈ Ω
on ∂Ω.
其计算区域为多孔薄板区域 Ω2 , 取解析解为
u = 4(1 − e−100x − (1 − e−100 )x)y(1 − y)e−z ,
它在 x = 0 处呈现边界层现象.
计算区域如图 4.7 示, 初始网格是一致网格, 由 Netgen 生成, 含有 49,271 个
四面体单元. 测试使用了 32 个计算节点, 128 个进程, 与前一个算例一样, 使用三
阶拉格朗日元进行离散, 使用 PCG (Preconditioned Conjugate Gradient method) 求
解离散得到的线性方程组, 采用残量型误差指示子, 形式如下
2
ηK
= h2K kfh + ∆uh k2L2 (K) +
X
f ⊂∂K∩Ω
hf ∂uh 2
k[
]k 2 .
2 ∂nf L (f )
标记策略为最大标记策略 (maximum strategy), 参数取为 0.5 (参看 §5.1.3.2).
图 4.12–4.15 分别给出了划分时间、动态负载平衡时间、线性方程组求解时
间及每个自适应步计算时间. 划分时间如图 4.12 所示, 最快的仍然为 RTK, 其次
为 MSFC, PHG/HSFC, Zoltan/HSFC, 最后是 ParMETIS 和 RCB. RCB 最慢, 同算
例 4.4.1 一样, ParMETIS 又表现出震荡.
动态负载平衡时间如图 4.13 示, RTK 用时最少, 并且 RTK 的运行时间随着
网格单元数量以线性增长; 其次是 ParMETIS, 它的运行略慢, 但其划分质量高,
通过减少数据迁移减少了负载平衡时间; 接着是 MSFC, PHG/HSFC; 运行最慢的
是 Zoltan/HSFC 及 RCB.
线性方程组求解及每个自适应步的时间如图 4.14 和 4.15 所示, MSFC 运行
时间最短, Zoltan/HSFC 运行时间最长, 其他几个划分方法的运行时间相差不大.
RCB 在这个例子中的表现不如在例 4.4.1 中的表现, 因为在前者的计算区域是一
个狭长的圆柱体, 较适合 RCB.
表 4.4 给出了总体运行时间及划分次数, 总的自适应次数为 238 次. 运行时
间最长的为 Zoltan/HSFC, 几乎是别的方法的 2 倍. 算例 4.4.1 及 4.4.2 均表明: 在
学位论文
64
thin plate, 128 procs
2.5
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
Partitioning time (s)
2
1.5
1
0.5
0
0
1e+07
2e+07
3e+07
4e+07
5e+07
6e+07
7e+07
8e+07
6e+07
7e+07
8e+07
Number of elements
图 4.12 网格划分时间 (例 4.4.2)
thin plate, 128 procs
14
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
time of dynamic load balancing time
12
10
8
6
4
2
0
0
1e+07
2e+07
3e+07
4e+07
5e+07
Number of elements
图 4.13 动态负载平衡时间 (例 4.4.2)
第四章 PHG 动态负载平衡模块的设计与实现
65
thin plate, 128 procs
700
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
time of solving linear system (s)
600
500
400
300
200
100
0
0
5e+07
1e+08
1.5e+08
2e+08
2.5e+08
3e+08
3.5e+08
Number of DOF
图 4.14 有限元离散系统求解时间, 横坐标为自由度数 (例 4.4.2)
thin plate, 128 procs
800
ParMETIS
RTK
PHG/HSFC
MSFC
Zoltan/HSFC
RCB
700
overall computational time (s)
600
500
400
300
200
100
0
0
5e+07
1e+08
1.5e+08
2e+08
2.5e+08
Number of DOF
图 4.15 每一个自适应步时间 (例 4.4.2)
3e+08
3.5e+08
学位论文
66
表 4.4 总体运行时间及划分次数 (例 4.4.2)
Method
total running time(s)
# of repartitionings
MSFC
6269.41
83
RTK
6514.84
89
PHG/HSFC
6570.44
83
RCB
6667.12
89
ParMETIS
6758.62
237
Zoltan/HSFC
12079.33
74
映射的过程中, 保持区域的空间局部性对划分质量是很重要的. 其他几个方法的
运行时间比较接近. 值得一提的是 ParMETIS, 和上一个例子一样, 其划分次数最
多, 这是由于问题的奇点是局部的, 网格的自适应局部加密会引起负载不平衡, 需
要调整网格分布. 一般说来, Morton 空间填充曲线的空间局部性要弱于希尔伯特
空间填充曲线, 但在这两个例子中, MSFC 的计算效果却始终优于希尔伯特空间
填充曲线划分方法 (PHG/HSFC & Zoltan/HSFC), 所以在选择划分方法时应根据
问题本身的特点、计算区域和划分质量等多个因素确定合适的划分策略, 以加快
运算速度, 优化计算资源.
例 4.4.3. 本例求解的方程为线性抛物型方程
ut − ∆u = f
取解析解为
u = exp((25((x −
u=g
in Ω × (0, T )
on Γ × [0, T ].
(4.8)
1 2
1 2
− sin(8πt))2 + (y − − cos(8πt))2 + (z − 1)2 ) + 0.9)−1 − 2.5).
2 5
2 5
计算区域为 Ω3 = (0, 1)3 , 时间区间为 [0, 1]. 解的极大值在 z = 1 平面上运动, 网
格在极值附近最密. 由于这是一个依赖于时间的问题, 网格自适应同时包含加密
和放粗.
初始计算网格包含六个四面体, 时间离散采用向后欧拉格式, 空间离散采用
二阶拉格朗日元, 使用 PCG (Preconditioned Conjugate Gradient method) 求解得到
的线性方程组, 在时间和空间方向上同时采用自适应 [66, 156], 误差指示子为残
量型的, 标记策略为最大标记策略 (maximum strategy), 加密和放粗的参数分别选
为 0.6 和 0.1.
第四章 PHG 动态负载平衡模块的设计与实现
67
表 4.5 程序运行时间 (例 4.4.3, 128 进程)
Method
Time TAL(s)
Time DLB(s)
Time SOL (s)
Time STP(s)
PHG/HSFC
6525
0.0734
0.1886
0.9192
Zoltan/HSFC
6744
0.0917
0.1928
0.9501
MSFC
6902
0.0730
0.1966
0.9724
RTK
7015
0.0738
0.2050
0.9884
RCB
7131
0.1126
0.1938
1.0046
ParMETIS
7151
0.1421
0.2114
1.0075
表 4.6 程序运行时间 (例 4.4.3, 192 进程)
Method
Time TAL(s)
Time DLB(s)
Time SOL(s)
Time STP(s)
PHG/HSFC
6597
0.0932
0.1808
0.9294
MSFC
6601
0.0936
0.1863
0.9299
Zoltan/HSFC
6646
0.1046
0.1862
0.9362
RCB
7197
0.1176
0.1862
1.0139
RTK
7185
0.0799
0.2124
1.0123
ParMETIS
7218
0.1982
0.1942
1.0169
这里给出两组测试结果. 第一组测试使用 32 个计算结点、128 个进程, 共
运行了 7098 个时间步, 每个时间步的平均单元数量为 663,151, 平均自由度数为
919,036. 结果见表 4.5, 其中第一项时间为总运行时间 (TAL), 第二项为单次动态
负载平衡的平均时间 (DLB), 第三项为单次线性方程组的平均求解时间 (SOL),
第四项为单个时间步的平均计算时间 (STP). 从表 4.5 中可以看出, 几何划分方
法的效果优于图划分方法, PHG/HSFC, Zoltan/HSFC, 及 MSFC 均优于 RTK 及
ParMETIS, RCB 除外. 其中 PHG/HSFC 运行时间最短, ParMETIS 运行时间最
长. 可以看出, 当网格变化较剧烈时, 几何划分方法也有优势. 一般来说, 在静态
划分中, 图划分方法划分质量最好, 最有优势; 在动态划分中, 情况则不一定.
第二组测试数据如表 4.6 所示, 其中使用了 32 结点、192 进程. 结果与第一
组测试类似, 几何划分方法用时较少, 仍优于图划分方法. 在本算例的两个表格
中, PHG/HSFC 均比 Zoltan/HSFC 用时略少, 说明 PHG 实现的版本的划分质量
略高, 但没有前面两个算例的差别那么大, 因为在这个算例中, 计算区域为 (0, 1)3 ,
Zoltan/HSFC 的映射结果和 PHG/HSFC 是类似的. 两个版本的差别在区域的长
学位论文
68
宽比较大时体现得更为明显.
§4.5
本章小结
本章介绍了 PHG 的层次网格结构, 网格划分方法及动态负载平衡模块, 并通
过数值实验比较了各种网格划分方法在并行自适应有限元计算中的优劣.
通常来讲, 几何划分方法的速度快、实现简单, 但划分质量不高; 图划分方法
的速度慢、算法复杂, 但划分质量非常高. 对静态问题或者网格变化较小的问题,
图划分方法是最好的选择; 对网格频繁调整的问题, 几何划分方法是很有竞争力
的. 在实际应用中, 应综合考虑网格划分、数据迁移、问题求解等多个因素确定
合适的划分方法.
第五章
hp 自适应策略设计
hp 自适应有限元方法同时调整网格和单元上的基函数, 它具有 h 自适应有
限元方法和 p 自适应有限元方法的优点. 在适当的策略下, hp 自适应有限元方法
可以达到指数阶收敛. 关于 hp 自适应有限元方法已有许多研究成果, 但自适应
策略仍旧是其难点之一, 目前尚没有成熟的、广泛被接受的算法, 本章将对 hp 自
适应策略进行研究.
基础知识
§5.1
在实际应用中, 很多科学和工程问题都可以使用偏微分方程或者常微分方程
进行建模, 由于问题的解析解很难获得, 通常通过求解数值解来给出解析解的近
似. 自适应有限元是求解偏微分方程的高效数值方法之一, 它根据数值解及已知
数据对误差作出估计, 并且根据误差指示子调整网格/基函数, 利用最少的计算资
源达到最好的计算结果. 自适应有限元方便处理奇性问题, 也方便处理复杂区域,
是当今科学与工程计算的热点研究领域之一 [157, 158, 159, 165, 199, 200, 201, 202].
§5.1.1
模型问题
考虑椭圆形偏微分方程,
−∇ · (A(x)∇u) =
f in Ω,
u|ΓD = gD on ΓD ,
(5.1)
∂u
= gN on ΓN ,
A(x)
∂ν
T
S
其中 A 为对称正定矩阵, ΓD , ΓN ⊂ ∂Ω, ΓD ΓN = ∅, ΓD ΓN = ∂Ω. 方程 (5.1) 的
变分形式为: 求解 u ∈ H 1 (Ω), 其中 u = ū + G, ū ∈ V = {v ∈ H 1 (Ω), v|ΓD = 0}, G ∈
H 1 (Ω), G|ΓD = gD , 并且
Z
Ω
Z
A∇ū · ∇vdx =
f vdx −
Z
a(u, v) =
Z
Ω
Ω
引入如下符号:
∇G · ∇vdx +
Ω
l(v) =
Z
Ω
f vdx −
Z
Ω
Z
ΓN
gN vdS, ∀v ∈ V.
A∇u · ∇vdx,
∇G · ∇vdx +
69
(5.2)
(5.3)
Z
ΓN
gN vdS.
(5.4)
学位论文
70
则问题变为求解 ū ∈ V , 使得下述变分形式成立,
∀v ∈ V.
a(ū, v) = l(v),
(5.5)
双线性型 a(·, ·) 是 V - 椭圆的并且有界, l 是有界的线性泛函, 由 Lax-Milgram
定理可知, u 存在唯一 [158, 207]. 为了简单起见, 后文将不再区分 ū 与 u.
§5.1.2
有限元离散
在变分问题 (5.5) 中, V 是无限维 Banach 空间, 计算机无法处理无限维空
间, 需要构造 V 的有限维子空间 Vh 进行近似, 并求解如下有限维离散问题: 求
u ∈ Vh ⊂ V , 满足如下变分形式
a(uh , vh ) = l(vh ),
∀vh ∈ Vh .
(5.6)
同样由 Lax-Milgram 定理可知, 该问题的解存在唯一.
为了构造子空间 Vh , 首先对计算区域 Ω 做剖分.
定义 5.1.1. (区域 Ω 的网格剖分) 将 Ω̄ 剖分成有限个闭子区域 K1 , · · · , Km , 每一
个子区域 Ki (1 ≤ i ≤ m) 称为一个单元, 全部单元组成的集合为 Th 称为 Ω 的一
个网格剖分, 简称剖分. 单元满足如下性质:
o
(1) 对每个单元 K ∈ Th , K 是闭集, 其内部 K 是非空且连通的;
(2) 单元 K 的边界 ∂K 是 Lipschitz 连续的;
(3) Ω̄ =
S
K∈Th
K;
o
(4) 任何两个不同的单元 K1 , K2 ∈ Th , K1
T
o
K2 = ∅;
(5) 每个 K ∈ Th , ∂K 或者是 ∂Ω 的一部分, 或者是相邻单元的边界.
剖分 Th 称为区域 Ω 上的一个计算网格. 在二维区域中, 常用单元类型为三
角形和四边形; 在三维区域中, 常用单元类型为四面体和六面体. 并行自适应有限
元软件平台 PHG 采用的单元类型为四面体.
在网格 Th 中, 任意单元 K, 记 PK 为某种多项式集合. 可以构造有限元空间
Vh 使得其中的函数在网格 Th 每一个单元上是一个多项式, 形式如下
Xh = {vh : vh |K ∈ PK , ∀K ∈ Th },
Vh = {vh ∈ Xh : vh |ΓD = 0},
第五章 hp 自适应策略设计
71
当限定多项式次数时, Vh 为有限维空间, 设其维数为 n, 则可以选取 n 个基
函数 {ϕ1 , · · · , ϕn }. 令
uh =
n
X
u i ϕi ,
(5.7)
i=1
则变分问题 (5.6) 等价于如下线性代数方程组
n
X
a(ϕi , ϕj )ui = l(ϕj ),
i=1
j = 1, · · · , n.
(5.8)
其系数矩阵为刚度矩阵 A = [a(ϕi , ϕj )]ni,j=1 , 未知量为 u = (u1 , · · · , un )T , 右端项为
b = (l(ϕ1 ), · · · , l(ϕn ))T , 问题最终可以写成
Au = b.
(5.9)
至此便得到了变分问题 (5.6) 的有限元离散.
§5.1.3
自适应有限元
误差估计和自适应加密由 Babuška 和 Rheinholdt 首先引入到有限元计算中
[160, 161]. 他们的方法基于估计数值解的残留, 使用残留得到一个局部更精确的
新解, 同时定义了网格加密过程. 受他们的启发, 其他计算数学家深入地研究了
这个方法, 设计了更有效的离散格式, 误差指示子等, 关于自适应有限元的理论结
果及最新进展见文献 [162, 163, 164, 165, 166, 195, 199, 200, 202, 203].
§5.1.3.1
误差指示子和自适应策略
误差指示子是自适应有限元方法重要的组件, 它真实反映了实际误差. 一个
好的误差指示子是真实误差的一个上界和下界, 上界表明真实误差被误差指示子
控制, 即计算是可信的; 下界表明误差指示子被真实误差控制, 即误差指示子是有
效的. 常用的误差指示子有如下几类:
• 残量型误差指示子 (residual estimate) [201, 202, 203]: 通过计算残量的范数来
估计数值解的误差.
• 对偶权重残量误差指示子 (dual weighted residual estimate) [199]: 先计算数值
解的残量, 然后乘以一个表示误差依赖关系的权重. 这种类型的误差指示子
需要求解一个对偶变分问题.
• 多重基误差指示子 (hierarchical basis error estimate) [200]: 通过计算高阶元或
者加密网格的有限元解, 估计两个解之间的残量.
学位论文
72
• 平均化方法 (averaging methods) [200]: 使用局部外插或者数值解梯度平均估
计误差子.
• 平衡残量误差指示子 (equilibrated residual error estimate) [200]: 将问题划分为
自平衡的系统, 然后在每个点或者单元的一个 patch 中求解一个局部的 Dirichlet 或者 Neumann 问题, 通过局部问题给出误差指示子.
自适应过程中, 如果全局后验误差估计小于给定值, 则有限元解被认为达到
了精度要求, 自适应循环结束; 否则, 局部自适应加密进程被调用, 通过调整单元
分布、单元中的基函数及加密误差比较大的单元得到一个新的有限元空间, 计算
在新空间上继续进行. 主要的自适应方法有如下几种:
• h 自适应方法: 通过将误差大的网格单元加密成多个子单元实现, 它保持基
函数的阶不变. 该方法的收敛阶是代数的, 是自适应有限元中最成熟最常用
的方法.
• p 自适应方法: 保持网格不变, 通过改变基函数的阶实现自适应. 该方法收敛
速度快, 如果网格选取适当并且解足够光滑, 可以达到指数阶收敛. 但它对网
格的要求比较高, 得到的线性方程组的条件数较高, 方程求解速度慢.
• hp 自适应方法: hp 自适应方法同时调整网格和基函数. 优点是收敛速度快,
在合适的策略下, 可以达到指数阶收敛, 缺点是理论不成熟, 缺乏稳健有效的
自适应策略.
• r 自适应方法: 移动网格方法, 保持网格规模不变, 根据数值解的误差分布修
改网格点的位置. 在误差大的地方放置更多的网格点. 该方法在流体、燃烧、
热传导等问题中应用较广泛.
§5.1.3.2
单元标记策略
本节仅限于讨论如何选取加密单元, 放粗单元的选取是类似的, 只需要改变
一些不等式的方向即可. 许多研究人员对标记策略进行了探讨并提出了一些有效
的策略. Dörfler 在文 [39] 中针对 Poisson 问题提出了保证误差下降策略, Morin 等
人在文 [40] 中提出了 MNS 加密标记策略. 目前常用的单元加密标记策略主要有:
最大策略 (maximum strategy, MAX), 误差等分布策略 (equidistribution strategy,
EQDIST), 保证误差下降策略 (guaranteed error reduction strategy, GERS), 和 MNS
加密策略 (MNS-refinement strategy, MNS). 本节先给出这些策略的数学描述.
第五章 hp 自适应策略设计
73
首先介绍几个符号: Th 表示当前网格的单元集合; S 表示当前网格的边 (二
维) 或面 (三维) 的集合; ηK (K ∈ Th ) 和 ηs (s ∈ S) 表示 (定义在单元和面上的) 误
差指示子; oscA (A ⊂ Th ) 是刻划数据振荡的量 (data oscillation); ε 是停止标准, 即
期望计算达到的精度 (tolerance); p 表示误差指示子的指数, 与误差所使用的范数
有关 (通常等于 2).
我们希望误差渐进相等地分布在各个单元上, 因而在标记加密单元时, 应尽
量选择误差指示子较大的单元. 标记策略设计都遵循这一原理. 常用标记策略描
述如下:
最大策略 (maximum strategy, MAX)
选取 θ ∈ (0, 1), 将所有满足下述条件的单元标记为加密单元:
ηK ≥ θ max
ηK 0
0
(5.10)
K ∈Th
误差等分布策略 (equidistribution strategy, EQDIST)
P p 1/p
ηK ) , 假如误差指示子
假定 Th 中单元的数目为 N , 全局误差为 η = (
K∈Th
ηK 在所有单元中均相等, 则有 η =
N 1/p ηK .
由此, 可以导出下述两个策略.
EQDIST1
选取 θ ∈ (0, 1), 标记满足下述条件的单元为加密单元:
ηK ≥ θ
η
N 1/p
(5.11)
EQDIST2
选取 θ ∈ (0, 1), 标记满足下述条件的单元为加密单元:
ηK ≥ θ
ε
N 1/p
(5.12)
保证误差下降策略 (guaranteed error reduction strategy, GERS)
Dörfler 在 [39] 中对于 Poisson 问题提出了能保证误差下降的策略, 该策略同
样可以应用到其他问题. 给定 θ ∈ (0, 1), 则加密单元的集合为满足下述条件
的 Th 的最小子集 A (⊆ Th ):
X
K∈A
p
ηK
≥ θp
X
p
ηK
(5.13)
K∈Th
MNS 策略加密 (MNS-refinement strategy, MNS)
该策略是 Morin 等人在证明自适应有限元的收敛性时提出的 [40]. 策略使用
学位论文
74
定义在网格的边 (二维) 或面 (三维) 上的误差指示子 ηs 和定义在单元上、刻
划数据振荡的量 osc. 首先, 根据给定的参数 θ1 ∈ (0, 1), 构造满足下述条件的
S 的最小子集 S1 :
X
s∈S1
ηsp ≥ θ1p
X
ηsp
(5.14)
s∈S
将所有至少有一条边 (二维) 或面 (三维) 属于 S1 的单元标记为加密单元, 并
记加密单元的集合为 A1 . 对于 osc, 则根据给定的参数 θ2 ∈ (0, 1), 扩大 A1 , 使
得加密单元集合是满足下述条件的 Th 的最小子集 A (A1 ⊆ A ⊆ Th ):
oscA ≥ θ2 oscTh .
(5.15)
对 MAX 和 EQDIST 策略, 参数 θ 的值越大, 标记的单元的数目越少; 而对于
GERS 和 MNS 加密策略, 参数 θ 的值越大, 标记的单元则越多.
§5.1.3.3
标记策略的并行算法
本节讨论 GERS 和 MNS 加密策略在基于消息传递的分布式存储并行计算
机上的实现问题, 同时介绍并行自适应有限元软件平台 PHG (Parallel Hierarchical
Grid) 中实现这些标记策略的统一函数接口. 由于 MAX 策略和 EQUIDIST 策略
很容易实现, 这里不作讨论.
根据上一节的介绍可知, GERS 策略和 MNS 加密策略的实现均可归结为如
下问题: 给定集合 G, 定义在 G 上的误差指示子 ηK (K ∈ G), 及参数 θ ∈ (0, 1), 求
满足下式的 G 的最小子集 A:
X
K∈A
ηK ≥ θ
X
ηK .
(5.16)
K∈ G
这里, 集合 G 被划分成一些互不相交的子集, 分布在各个进程中.
上述问题最直接的算法是先将 G 中的元素按照误差指示子 ηK 的大小排序,
然后从大到小依次选出 A 中的元素, 算法的计算复杂度为 O(N log N ), N = |G|.
PHG 没有采用这一算法, 因为它需要进行并行排序, 通信量较大.
给定 γ ∈ [0, 1], 定义集合 Aγ 如下:
Aγ = {K | K ∈ G, ηK ≥ γ max
ηK 0 }
0
记 ηγ =
P
K∈Aγ
K ∈G
ηK (它是集合 Aγ 中误差指示子的和), η =
P
ηK , 则问题转化为给
K∈G
定 θ 求 γ 使得 ηγ = θη. 易知 γ 是 θ 的减函数. 我们用二分法计算满足要求的 γ.
第五章 hp 自适应策略设计
75
上述算法很容易并行实现, 每次二分计算中的通信仅为一次规约操作, 算法
的计算复杂度为 O(N ). 为了提高通信粒度, 也可以采用 k 分 (k > 2) 的方法, 通
过适当增加计算量来减少通信次数.
我们为 PHG 设计了统一的函数接口来完成单元标记, 该函数可以同时对加
密和放粗进行标记, 函数接口定义如下:
void phgMarkElements(STRATEGY strategy_r, DOF *eta_r, FLOAT theta,
DOF *osc, FLOAT zeta, int depth_r, STRATEGY strategy_c,
DOF *eta_c, FLOAT gamma, int depth_c, FLOAT a),
其中, strategy_r 和 strategy_c 分别表示加密和放粗策略, 可以取值 MARK_MAX,
MARK_EQDIST, MARK_GERS, MARK_MNS, MARK_ALL, MARK_NONE, 其中最后两个分别表示
标记全部单元和不标记任何单元; eta_r, osc 和 eta_c 分别给出加密误差指示子,
数据振荡和放粗误差指示子; theta, zeta 和 gamma 给出标记参数; 而 depth_r
和 depth_c 则给出被标记的单元加密和放粗次数. 参数 a 仅被 EQDIST 策略
使用, 对于加密而言, 标记的单元集合为 {K | ηK ≥ theta × a}. strategy_r 和
strategy_c 也可以取值为 MARK_DEFAULT, 它允许用户在运行程序时通过命令行选
项 -strategy, -refine-strategy 和 -coarsen-strategy 指定标记策略. 默认策略
为 MARK_GERS.
在这些策略中, GERS 是一个较好的选择, 通常情况下它表现出比其他策略
更好的稳健性.
hp 自适应有限元
§5.2
hp 自适应有限元由 Ivo Babuška 等人最早提出, 它收敛速度快, 受到研究人
员的重视. hp 自适应策略仍然是 hp 自适应有限元方法的难点之一. 本节对自适
应策略进行探讨.
§5.2.1
有限元解空间和变分形式
本节接下来的内容中网格限定为协调四面体网格. 在经典 h 自适应有限元
中, 每个单元上的基函数的阶是相同的, 其有限元空间 Vh 的构造相对简单. hp 自
适应有限元中, 每个单元上的基函数的阶不一定相同. 用符号 Vh,p 和 Th,p 分别表
示有限元空间和网格. 在协调四面体网格中, 单元上的基函数按照其所在的位置
或者定义方式分为顶点基函数、边基函数、面基函数和体基函数, 它们分别定义
在四面体的顶点、边、面和单元内部. 假设给定每个单元中体基函数的阶, 我们采
用最小原则 (minimum rule) [167] 来定义各类基函数的阶, 它可使得有限元空间独
学位论文
76
立于具体的基函数 [158]. 根据最小原则, 顶点、边或面基函数的阶取为包含该顶
点、边或面的所有单元体基函数的阶的最小值.
所有这些基函数张成 Th,p 上的协调有限元空间 Vh,p . hp 自适应有限元的变
分问题为: 求解 uh,p ∈ Vh,p , 使得下式成立,
a(uh,p , vh,p ) = l(vh,p ),
∀vh,p ∈ Vh,p .
(5.17)
有限元空间 Vh,p 的基函数有很多选择, 但一般要求基函数满足两个条件:
• 基函数的支集 (support) 应该是局部的, 从而使得最终的刚度矩阵是稀疏的;
• 基函数应该是层次的 (hierarchic), 低阶基函数集合是高阶基函数集合的子集.
层次性用于保证有限元空间的协调性.
常用的层次基函数基于 Jacobi 多项式, Legendre 多项式和 Lobatto 函数构造, 详
细内容参考 [168, 169, 170].
§5.2.2
hp 自适应加密策略
hp 自适应有限元的难点之一是自适应加密策略, 即在单元上如何选取 h 加
密和 p 加密. 原则是, 应该在方程的解光滑的区域选择 p 加密, 在解有奇性的区
域选择 h 加密. 不幸的是, 解的性质在通常情况下是不知道的. 研究人员设计了
不同的方法确定解的正则性. 下面介绍一些常见的 hp 自适应加密策略:
(1) 利用解的正则性先验信息 (use of a priori knowledge of solution regularity).
p 加密可以在解的光滑区域达到指数阶收敛, 但在有奇性的区域却没有 h 加
密有效. 由于这个原因, 许多 hp 加密策略在解有奇性的区域采用 h 加密, 别
的区域采取 p 加密. 最简单的策略是利用正则性的先验信息. 例如, 如果线性
椭圆偏微分方程的系数是光滑的并且边界条件是分片解析的, 那么在区域的
凹角处有点奇性 [172]. Ainsworth 和 Senior [172], Bernardi 等人 [184] 以及
Valenciano 等人 [185] 利用区域及已知数据的先验信息在有奇性的区域采用 h
加密, 在其他区域采用 p 加密.
(2) 类型参数 (type parameter).
Gui 和 Babuška [182] 及 Affia 等人 [186] 使用类型参数的概念, 设计了 hp 自
适应策略. 给定误差指示子 ηK,pK 和 ηK,pK −1 , 定义
R(K) =
ηK,pK
ηK,pK −1 ,
0,
ηK,pK −1 6= 0
ηK,pK −1 = 0,
第五章 hp 自适应策略设计
77
R 用来评估解的光滑性. 给定参数 γ (0 ≤ γ < ∞), 如果 R(K) ≥ γ, 单元 K 将
采取 h 加密; 如果 R(K) < γ, 单元 K 将采取 p 加密. 另外, 如果 γ = 0, 单元
K 将只采取 h 加密, 而如果 γ = ∞, 则单元 K 将只采取 p 加密.
(3) 用低阶函数空间估计解的正则性 (estimate regularity using smaller p estimates).
Süli, Houston 和 Schwab [173] 利用两个低阶基函数空间计算误差下降阶, 根
据下降阶估计解的正则性. 设单元基函数的阶为 p, 如果在单元上计算基函
数阶分别为 p − 1, p − 2 时的误差 ηp−1 和 ηp−2 , 有下面近似表达式成立
ηp−1
p − 1 1−m
≈(
)
,
ηp−2
p−2
于是正则度 m 可定义为:
m≈1−
log(ηp−1 /ηp−2 )
.
log((p − 1)/(p − 2))
当 p ≤ m − 1 时, 采取 p 加密; 否则采取 h 加密.
(4) 利用高阶函数空间估计解的正则性 (estimate regularity using larger p estimates).
Ainsworth 和 Senior[177] 给出利用高阶基函数空间估计解正则性的策略. 设当
前单元 K 上基函数阶为 p, 分别计算基函数阶为 p + 1, p + 2, p + 3 时的误
差. 该策略将问题 (5.17) 转化为满足平衡条件 (equilibrium condition) 的局部
Neumann 问题. 假设基函数阶为 p + j 时的误差为 ej , 那么有下面近似表达式
成立
ke0 k2K − kej k2K ≈ C 2 (p + j)−2α ,
其中 C, α 为正的常数. 局部正则度 m 定义为 m = 1 + α. 如果 p ≤ m − 1 = α,
则采取 p 加密, 否则采取 h 加密.
(5) 参考解 (reference solution).
Demkowicz 及其合作者 [190, 191, 192] 设计了一个 hp 自适应策略. 假定当前计
算网格及数值解分别为 Th,p 和 uh,p , 对网格和基函数做一致的 h 和 p 加密, 得
到一个更细的网格 Th/2,p+1 , 在这个网格上可以求得一个新的数值解 uh/2,p+1 .
uh,p 与 uh/2,p+1 差值的某个范数作为全局误差, 如
η = kuh,p − uh/2,p+1 kH 1 (Ω) .
下一步是决定每个单元的最优加密方式, 考虑所有可能的 h 加密方式和 p 加
密方式, 从中选择一个最优的加密方法. 该策略是有效的, 但非常复杂, 计算
量特别大. Šolı́n 等人 [193] 给出了一个简化的版本.
学位论文
78
(6) 误差下降预测 (predict error estimate on assumption of smoothness).
Melenk 等人 [174] 以及 Heuveline 等人 [176] 给出了启发式加密策略. 对二维四
边形和三维六面体单元, 他们假设解是光滑的, 根据近似理论, 可以预测 h 加
密及 p 加密的误差下降趋势, 而根据这个预测判断解是否是光滑的. 如果实
际误差达到预测, 表明解是光滑的; 如果达不到预测, 则表明解不够光滑. 对
光滑的区域采取 p 加密, 对非光滑区域采取 h 加密.
hp 自适应有限元的加密策略没有 h 自适应有限元那样成熟, 构造简单有效
的自适应加密方法仍然是一个很大的挑战, hp 自适应有限元基本理论也在发展
之中. 上面给出了一些自适应策略的简单描述, 详细的过程可以参考相关文献, 还
有一些其他的策略, 可以参考文献 [167, 171, 175, 177, 183, 187, 188, 189, 194, 196].
一种新的 hp 自适应策略
§5.3
本节介绍我们设计的 hp 自适应加密策略, 该策略基于误差下降预测. 该策略
中, 我们对不同的加密方式预测误差下降的速度, 如果实际误差与预测一致, 表明
解是光滑的, 对单元采取 p 加密; 如果误差下降没有达到预期, 表明解不够光滑,
对单元采取 h 加密. 这个策略很大程度上受到 Melenk 等人 [174] 以及 Heuveline
等人 [176] 工作的启发. 本文对 h 加密及 p 加密的误差下降估计采用不同的假设,
得到不同的误差下降估计公式, 这个策略不仅适合二维四边形和三维六面体, 也
适合二维三角形单元和三维四面体单元.
本节考虑线性椭圆型偏微分方程
−4u = f in Ω,
u=
(5.18)
0 on ∂Ω,
其中 Ω ⊂ Rd , d = 2, 3, f ∈ L2 (Ω). 我们的目标是设计和定义计算网格 Th,p 和一个
有限元空间 Vh,p ⊂ H01 (Ω), 计算数值解 uh,p ∈ Vh,p 使得变分形式 (5.17) 成立, 并达
到事先设定的停止标准.
p
下面介绍一些记号. 记 k · k 为能量模, kuk = a(u, u); ε 为停止标准; ηK 为
P
1
定义在单元 K 上的误差指示子; η 定义为 η = ( K∈Th,p ηK ) 2 . 用 pK 和 hK 分别
表示单元 K 上的基函数的阶及单元的尺寸. 如果一个单元被加密成多个子单元,
那么 cK 表示其子单元的个数. 最后引入符号 N 表示自由度数目.
§5.3.1
hp 自适应策略
Melenk 等人 [174] 以及 Heuveline 等人 [176] 提出的 hp 自适应策略基于如下近
似理论: 如果对网格做一致剖分, 并且所有单元上基函数的阶均为 p, 记 e = u−uh,p
第五章 hp 自适应策略设计
79
为误差, 如果 u ∈ H m (Ω), 那么当采用分片多项式基函数时有 [177, 178]
kekH 1 (Ω) ≤ C
hµ
p(m−1)
kukH m (Ω) ,
(5.19)
其中 h 是网格的尺寸, µ = min(p, m − 1), C 是独立于 h 及 p 的常数.
Melenk 及其合作者 [174] 针对二维四边形网格提出了一种基于误差下降预测
的 hp 自适应策略: 在四边形网格中, 假设当前单元的误差指示子为 ηK , 如果单元
被加密成四个子单元, 那么每一个子单元上的误差预测为 0.5 · 0.5pK ηK ; 如果单元
上的基函数阶提高一阶, 那么误差预测 γp ηK , γp (∈ (0, 1)) 为常数; 如果单元没有
被调整 (包括网格及基函数), 则误差预测为 ηK , 即保持不变. 在自适应过程中, 一
个单元被标记为加密单元, 则判断其误差下降是否达到预期: 如果误差下降达到
预期, 表明解是光滑的, 下一个自适应步将采取 p 加密; 否则将采取 h 加密.
Heuveline 及其合作者 [176] 针对二维四边形和三维六面体也提出了基于误差
下降预测的策略. 他们的策略是尽可能地多采用 p 加密, 该策略分为几种情况: 1)
如果被标记加密的单元在上一个自适应步中没有被加密过, 则采取 p 加密; 2) 如
old , 如果误差小于这个预测,
果单元 K 是通过 p 加密得到的, 那么误差预测为 hk ηK
则采取 p 加密, 否则采取 h 加密; 3) 如果单元是通过 h 加密得到的, 那么误差预
old , 其中 K
测为 2−pK ηK
M 为其父单元, 如果误差小于这个预测, 则采取 p 加密, 否
M
则采取 h 加密.
这两个策略适合二维四边形单元和三维六面体单元, 不能直接应用于三角形
的局部自适应二分加密及四面体的局部自适应二分加密. 另外 Heuveline 及其合
作者设计的策略存在对 hK 的依赖, 策略对单元尺寸敏感. 当单元尺寸较大的时
候 (hK ≥ 1), 自适应过程中会有过多的 p 加密, 不能适当地调整网格; 而当尺寸较
小的时候 (hK ¿ 1), 则会有太多的 h 加密, 收敛速度不一定能达到指数. 我们希
望将这两个策略推广到网格二分加密的情况, 并且使得策略与网格尺寸无关.
这里给出一个 hp 自适应策略, 该策略允许单元采取 h, p 及 hp 自适应加密.
对 h 及 p 自适应加密做不同的假设: 对 h 自适应加密, 假设可以达到代数阶收敛;
对 p 自适应加密, 假设可以达到指数阶收敛, 然后使用不同的误差估计来预测误
差下降因子. 本文中一个单元允许同时采取 h 及 p 自适应加密, 并给出了这种情
况的误差下降预测.
在 h 自适应有限元方法中, 误差大的单元被加密成几个子单元, 其最优收敛
速度是代数的, 表达式可以写成
kek ≤ C1 N
−p
d
,
(5.20)
学位论文
80
其中 p 是单元上基函数的阶, C1 为常数. 将不等式 (5.20) 写成等式, 具有如下形
式
kek = C1 N
−p
d
= C2 (
X
1
2 2
ηK
) ,
(5.21)
K∈T
其中 C2 是依赖于所采用的误差指示子的常数. 这是本文自适应策略的基本出发
点.
如果对网格 Th,p 采取一致 h 加密, 每个单元被加密成 cK 个子单元, 得到新
1 , 同时假设拥有相同父单元的子单元拥有相同的误差指示子. 现在自
的网格 Th,p
由度的数目近似为 cK N . 如果解是光滑的, 并且达到最优代数收敛, 那么有下式
近似成立
C1 (cK N )
−p
d
= C2 (
X
1
2 2
ηK
) = C2 (cK λ2h
1
K∈Th,p
X
1
2 2
ηK
) ,
(5.22)
K∈Th,p
其中 λh 是误差下降因子. 综合式 (5.21) 及式 (5.22), 对 h 自适应加密, 可以得到
误差下降因子
λ2h =
1 1 2p
( )d.
cK cK
我们采用和文章 [176] 一样的想法: 即尽可能多地采取 p 加密. 对 h 自适应加密,
本文使用一个放大的值做误差下降因子, 表达式如下
λh = (
1 p
)d .
cK
(5.23)
下面推导 p 加密的误差下降因子. 误差估计式 (5.19) 中需要知道解析解 u 的
正则性 m, 由于一般情况下 m 是不知道的. 这里采用了一个启发式的技术. 假设
p 大于等于 m, 并将不等式 (5.19) 写成等式, 则当基函数提高一阶时, 有
kekH 1 (Ω) = C(
p m−1 hµ
)
kukH m (Ω) ,
p+1
p(m−1)
从而误差下降因子为
λp = (
p m−1
)
.
p+1
对给定的问题, 需要使用合适的值近似 m − 1. 在单元 K 上, 我们使用单元 K 上
的基函数的阶近似解在单元上的局部正则性. 取 m − 1 = p/2. 这个选择是合适
的, 因为假设选取的值过低, 表明解的局部正则性比估计的好, 那么下一步自适应
加密将采取 p 加密, 否则下一个自适应步的加密将采取 h 加密. 最终得到 p 加密
误差下降因子为
λp = (
p p
)2 .
p+1
(5.24)
第五章 hp 自适应策略设计
81
在实际应用中, 当某一个单元采取 h 加密时, 为了保证网格的协调性, 有可
能使邻居单元也被 h 加密, 如果这个邻居单元被标记了 p 加密, 将允许这个单元
同时进行 h 自适应加密及 p 自适应加密. 组合上面两个误差下降因子得到 hp 加
密的误差下降因子为,
λhp = (
p p 1 p
)2 ( )d .
p+1
cK
(5.25)
误差下降估计式 (5.23), (5.24) 及 (5.25) 对网格的 h 加密类型并没有限制, 仅需要
知道每个单元 h 加密后产生的子单元的个数 cK . 该策略消除了 Heuveline 及其合
作者提出的策略对 h 的依赖, 使得我们的策略与网格尺寸无关.
接下来给出我们的 hp 自适应加密策略, 此处使用 Heuveline 及其合作者 [176]
设计的一个框架, 并做了一些修改. 自适应过程如下:
(1) 对当前计算网格 Th,p , 如果误差指示子 η 满足停止标准 (η ≤ ε), 则停止自适
应循环.
(2) 使用最大策略标记加密单元.
(3) 对网格 Th,p 中每一个单元 K, 如果 K 被标记为加密单元:
• 如果 K 是通过上一个自适应步的 h 加密得到的, 并且基函数的阶没有改
变. 记其父单元为 KM , 检查是否有
2
2
ηK
≤ λ2h ηK
.
M
如果上式成立, 将单元 K 标记为 p 加密, 将基函数的阶提高一阶; 否则,
将单元 K 标记为 h 加密;
• 如果 K 是通过上一个自适应步的 p 加密得到的, 记其父单元为 KM , 检
查是否有
2
2
ηK
≤ λ2p ηK
.
M
如果上式成立, 将单元 K 标记为 p 加密, 将基函数的阶提高一阶; 否则,
将单元 K 标记为 h 加密;
• 如果 K 是通过上一个自适应步的 hp 加密得到的, 记其父单元为 KM , 检
查是否有
2
2
ηK
≤ λ2hp ηK
.
M
如果上式成立, 将单元 K 标记为 p 加密, 将基函数的阶提高一阶; 否则,
将单元 K 标记为 h 加密;
学位论文
82
• 如果单元 K 在上一个自适应步没有被加密 (单元大小及基函数阶保持不
变), 则将单元 K 标记为 p 加密, 将基函数的阶提高一阶.
(4) 根据单元的加密类型, 对单元实施 p 加密及 h 加密, 转至 (1).
§5.4
数值算例
本节用三个数值算例验证我们设计的 hp 自适应策略. 数值算例中比较了三
个不同的自适应策略: h 自适应方法, 本文设计的 hp 自适应策略及 Melenk 及其
合作者设计的 hp 自适应框架 [174]. 测试程序利用 PHG 编制, 测试环境同前一
章.
由于问题是三维的, d 为 3, 网格局部自适应加密算法为二分加密算法, 一个
单元被加密成两个子单元, 因此 cK 为 2. 因而误差下降因子变为
1 p
λh = ( ) 3 ,
2
p p
)2 ,
p+1
(5.27)
p p 1 p
)2 ( )3 .
p+1
2
(5.28)
λp = (
λhp = (
(5.26)
这里使用的误差指示子由 Melenk 及其合作者设计 [174], 该误差指示子原为二维
问题设计, 同样可以用在三维问题上, 其形式如下
2
ηK
=
h2K
kfpK + ∆uN k2L2 (k) +
p2K
X
f ⊂∂K∩Ω
hf ∂uN 2
k[
]k 2 .
2pf ∂nf L (f )
(5.29)
其中 fpK 为 f 在 p − 1 阶基函数空间上的 L2 (K) 投影, hf 为单元 K 上面 f 的直
径, pf = max(p|K1 , p|K2 ), 其中 f 为单元 K1 及 K2 的共享面.
最大策略中参数 θ 选取为 0.5, 基函数为 H 1 层次基, 基函数初始阶数为 2, 最
大阶数为 15, 解法器为预条件共轭梯度法 (PCG, Preconditioned Conjugate Gradient
method), 预条件子为 block Jacobi.
在三维问题中, 合适的 hp 自适应策略可以使误差达到指数阶收敛, 最优收敛
有如下形式 [179]
kek ≤ Ce−γN
1/5
,
(5.30)
其中 C 及 γ 为常数, N 为自由度数. 本节的图形中, 横轴以及纵轴的变量分别为
自由度数和能量误差, 它们的刻度分别为 N 1/5 和对数. 如果误差达到指数阶收
敛, 那么误差曲线应该是一条直线.
第五章 hp 自适应策略设计
83
图形中, HAFEM 表示仅使用 h 加密的结果, HP/PHG 和 HP/MK 则分别表示我们
的策略和 Melenk 及其合作者的策略, 二者的一个不同点在于: 如果一个被标记的
单元在上一个自适应步中没有被调整 (包括网格和基函数), 那么 HP/PHG 将对这
个单元采取 p 自适应加密, 而 Melenk 等人的做法则是如果单元在这个自适应步
中的误差指示子小于上一个自适应步中的误差指示子则采取 p 自适应加密, 否则
采取 h 自适应加密.
例 5.4.1. 取解析解为 u = e−z (1 + e100(x+y−1) )−1 , 它在平面 x + y − 1 = 0 上呈现内
层现象, 计算区域为 Ω = (0, 1)3 , 初始网格为一致网格, 单元数为 24.
10
HAFEM
HP/PHG
HP/MK
Energy error
1
0.1
0.01
3125
1e+5
7.60e+5
3.2e+6
9.77e+6
Number of DOF
图 5.1 误差曲线 (例 5.4.1)
表 5.1 不同策略的最终单元数、自由度数及能量误差 (例 5.4.1)
# elements
# DOF
Energy error
HP/PHG
23,920
750,009
1.08e-2
HP/MK
86,520
1,271,317
1.46e-2
HAFEM
6,969,762
9,528,910
1.44e-2
例 5.4.1 的计算结果如图 5.1 所示, 策略 HP/PHG 及 HP/MK 均达到了指数收敛,
但 HP/PHG 的收敛速度明显快于 HP/MK. 从表 5.1 也可以看出, HP/MK 的网格规模及
学位论文
84
自由度规模远大于 HP/PHG 的规模, 说明 HP/MK 更多地做 h 自适应加密, 而 HP/PHG
更多地做 p 自适应加密, 效率更高. hp 自适应有限元与 h 自适应有限元相比, 前
者的收敛速度远快于后者.
例 5.4.2. 取解析解为 u = cos(2πx) cos(2πy) cos(2πz), 该问题的难点在于解的高阶
导数, 其奇数阶与偶数阶导数在每个点的行为不一样, p 自适应加密不一定能改
善数值解 [196]. 计算区域为 L- 形区域, Ω = (−1, 1)3 \ [0, 1] × [−1, 0] × [−1, 1], 初始
网格是一致的, 含 144 个单元.
10
HAFEM
HP/PHG
HP/MK
Energy error
1
0.1
0.01
0.001
0.0001
32
1024
7776
3.28e+4
1e+5
2.49e+5
5.38e+5
1.05e+6
1.89e+6
3.2e+6
Number of DOF
图 5.2 误差曲线 (例 5.4.2)
表 5.2 不同策略的最终单元数、自由度数及能量误差 (例 5.4.2)
# elements
# DOF
Energy error
HP/PHG
3,772
246,046
1.01e-4
HP/MK
35,696
1,171,216
1.67e-4
HAFEM
1,663,068
2,263,137
3.57e-2
计算结果见图 5.2 和表 5.2, 两种 hp 自适应策略均达到指数收敛, 但差别更
明显, HP/MK 策略产生的最终网格单元数比 HP/PHG 高出近一个数量级, 自由度数
目大概是 5 倍, 但最终的误差仍然大于 HP/PHG 的误差, 说明 HP/PHG 策略的效率
第五章 hp 自适应策略设计
85
高于 HP/MK. 从表 5.2 可以看出, h 自适应有限元产生的单元数量比 HP/PHG 高出
两个数量级, 自由度数高一个数量级, 误差却大了两个数量级, hp 自适应有限元
方法相对于 h 自适应有限元方法的优势明显.
1
例 5.4.3. 解析解为 u = (x2 + y 2 + z 2 ) 4 , 求解区域是 Fichera 区域, Ω = (−1, 1)3 \
[0, 1]3 . 这个问题的计算区域有奇性, 同时解析解有一个点奇性. 初始网格为一致
网格, 单元数目为 172.
1
HAFEM
HP/PHG
HP/MK
Energy error
0.1
0.01
0.001
0.0001
1e-05
32
1024
7776
3.28e+4
1e+5
2.49e+5
5.38e+5
1.05e+6
1.89e+6
3.2e+6
Number of DOF
图 5.3 误差曲线 (例 5.4.3)
表 5.3 不同策略的最终单元数、自由度数及能量误差 (例 5.4.3)
# elements
# DOF
Energy error
3,429
155,812
1.07e-5
HP/MK
163,204
1,158,279
1.20e-4
HAFEM
1,377,588
1,904,054
4.44e-4
HP/PHG
例 5.4.3 的计算结果见图 5.3 和表 5.3. 两种 hp 自适应策略均达到指数收敛,
并且收敛曲线更光滑, 说明误差下降预测是准确的. 自适应策略 HP/MK 的单元数
是策略 HP/PHG 的 50 倍, 自由度数目大约是 8 倍, 误差却大了一个数量级, 表明我
学位论文
86
们的自适应策略更有效. h 自适应有限元仅达到了代数收敛, 而 hp 自适应有限元
方法达到了指数收敛.
§5.5
本章小结
本章介绍了 hp 自适应有限元及常用的 hp 自适应策略. 在研究现有 hp 自适
应策略的基础上提出了一个新的 hp 自适应策略, 该策略基于误差下降预测. 我
们给出了不同加密策略下误差下降速度的预测公式. 数值实验表明, 我们提出的
hp 自适应策略对所测试的问题可以达到指数收敛, 并且优于 Melenk 及其合作者
提出的策略.
第六章
总结与展望
本文对并行自适应有限元计算中的动态负载平衡及 hp 自适应策略进行了研
究. 对动态负载平衡, 我们研究了过顶点哈密尔顿圈及哈密尔顿路, 给出了存在性
定理, 并设计了线性复杂度的构造算法; 设计了任意维希尔伯特序的编码解码算
法, 并提出该曲线的一个变体; 改进并实现了加密树以及空间填充曲线两种网格
划分方法. 对 hp 自适应有限元方法, 我们提出了一个新的自适应加密策略.
动态负载平衡是并行自适应计算中的一个重要的、必不可少的环节. 在并行
自适应有限元计算中，动态负载平衡通常通过网格重划分实现, 网格重划分的速
度、网格重划分的质量以及网格重划分后数据迁移的开销是影响整个并行自适应
有限元计算并行效率和并行可扩展性的重要因素. 此外, 网格重划分后, 新的子网
格的重组和数据迁移涉及到复杂的数据处理和通信, 其并行程序设计亦是一项非
常具有挑战性的工作. 本文介绍了作者在网格划分算法和 PHG 动态负载平衡模
块的设计与实现方面的工作. 本文所实现的 PHG 动态负载平衡模块已经在实际
应用中经受住了考验, 可以胜任数千 CPU 核、十亿以上单元规模的并行自适应
有限元计算的要求 (该部分内容可以参见崔涛的博士论文), 并且在整个自适应有
限元计算时间中只占很小的比例，展现出很高的效率和很好的稳健性, 基本能够
满足当前应用的需求. 在未来, 我们将根据随着并行规模的增大及应用的复杂化
而对动态负载平衡提出的新要求, 继续对相关算法及模块进行改进和完善.
hp 自适应有限元方法能够达到很高的计算效率, 是一类非常有前景的算法,
是当前科学计算研究的热点问题之一. 当前, hp 自适应有限元方法远未成熟, 离
实际应用尚有很大的距离, 其中的难点之一便是 hp 自适应策略的设计. 本文利用
PHG 平台对 hp 自适应有限元的支持, 对 hp 自适应策略进行了研究和数值实验.
文中提出了一种新的 hp 自适应策略, 它在所适用的网格及网格加密方式方面比
文献中提出的一种策略有所扩展, 并且数值实验的结果表明其在计算效果上亦优
于文献中的策略. 此外, 这些数值实验亦验证了 PHG 平台中与 hp 自适应相关的
模块的正确性和适用性. 这些研究是初步的, 有待于未来从理论上和应用上进一
步深入和展开.
87
88
学位论文
参考文献
[1] L. Zhang, A Parallel Algorithm for Adaptive Local Refinement of Tetrahedral Meshes
Using Bisection, Numer. Math.: Theory, Methods and Applications, 2009(2): 65–89
[2] G. A. Geist, J. A. Kohl and P. M. Papadopoulos, PVM and MPI: a Comparison of
Features, Calculateurs Paralleles, 1996(8): 137–150
[3] Al Geist, Adam Beguelin, Jack Dongarra, Weicheng Jiang, Robert Manchek and Vaidy
Sunderam, PVM: Parallel Virtual Machine A Users’ Guide and Tutorial for Networked
Parallel Computing, The MIT Press, 1994
[4] PVM: Parallel Virtual Machine, http://www.csm.ornl.gov/pvm/pvm home.html
[5] MPI: Message Passing Interface, http://www.mcs.anl.gov/research/projects/mpi/index.htm
[6] The MPI Standard, http://www.mcs.anl.gov/research/projects/mpi/standard.html
[7] MPI-2: Extensions to the Message-Passing Interface,
http://www.mcs.anl.gov/research/projects/mpi/mpi-standard/mpi-report-2.0/mpi2report.htm
[8] MPICH–An Implementation of MPI, http://www-unix.mcs.anl.gov/mpi/mpich/
[9] Bill Gallmeister, POSIX 4 Programmers Guide: Programming for the Real World,
O’Reilly Media, 1995
[10] Dick Buttlar, Jacqueline Farrell and Bradford Nichols, PThreads Programming: A
POSIX Standard for Better Multiprocessing , O’Reilly Media, 1996
[11] Barbara Chapman, Gabriele Jost and Ruud van der Pas, Using OpenMP: Portable Shared
Memory Parallel Programming, The MIT Press, 2007
[12] Rhit Chandra, Leonardo Dagum, Dave Kohr, Dror Maydan, Jeff McDonald and Ramesh
Menon, Parallel Programming in OpenMP, Morgan Kaufmann Publisher, 2001
[13] George Karypis and Vipin Kumar, METIS–Family of Multilevel Partioning Algorithms,
http://glaros.dtc.umn.edu/gkhome/metis/metis/overview
[14] George Karypis and Vipin Kumar, PARMETIS–Parallel Graph Partitioning and Fillreducing Matrix Ordering,
http://glaros.dtc.umn.edu/gkhome/metis/parmetis/overview
[15] Kirk Schloegel, George Karypis and Vipin Kumar， A Unified Algorithm for Loadbalancing Adaptive Scientific Simulations，Supercomputing, 2000
89
学位论文
90
[16] Kirk Schloegel, George Karypis and Vipin Kumar, Parallel Multilevel Algorithms for
Multi-Constraint Graph Partitioning，EuroPar, 2000
[17] George Karypis, Kirk Schloegel and Vipin Kumar, ParMETIS: Parallel Graph Partitioning and Sparse Matrix Ordering Library version 3.1, 2003
[18] E. Boman, K. Devine, R. Heaphy, B. Hendrickson, V. Leung, L.A. Riesen, C. Vaughan,
U. Catalyurek, D. Bozdag, W. Mitchell and J. Teresco, Zoltan v3: Parallel Partitioning,
Load Balancing and Data-Management Services, User’s Guide, Sandia National Laboratories Tech. Rep. SAND2007-4748W, 2007
[19] E. Boman, K. Devine, R. Heaphy, B. Hendrickson, V. Leung, L.A. Riesen, C. Vaughan, U.
Catalyurek, D. Bozdag and W. Mitchell, Zoltan v3: Parallel Partitioning, Load Balancing
and Data-Management Services, Developer’s Guide, Sandia National Laboratories Tech.
Rep. SAND2007-4749W, 2007
[20] Karen Devine, Erik Boman, Robert Heaphy, Bruce Hendrickson and Courtenay Vaughan,
Zoltan Data Management Services for Parallel Dynamic Applications, Computing in
Science and Engineering, 2002(4): 90–97
[21] Erik Boman, Karen Devine, Lee Ann Fisk, Robert Heaphy, Bruce Hendrickson, Vitus
Leung, Courtenay Vaughan, Umit Catalyurek, Doruk Bozdag and William Mitchell,
Zoltan home page, http://www.cs.sandia.gov/Zoltan
[22] R. Falgout, V. E. Henson, U. M. Yang, T. Kole, B. Lee, J. Painter, C. Tong and P.
Vassilevski, HYPRE–High Performance Preconditioners,
http://www.llnl.gov/CASC/linear solvers/
[23] Satish Balay, Kris Buschelman, William D. Gropp, Dinesh Kaushik, Matthew G. Knepley, Lois Curfman McInnes, Barry F. Smith and Hong Zhang, PETSc Web page,
http://www.mcs.anl.gov/petsc
[24] Satish Balay, Kris Buschelman, Victor Eijkhout, William D. Gropp, Dinesh Kaushik,
Matthew G. Knepley, Lois Curfman McInnes, Barry F. Smith and Hong Zhang, PETSc
Users Manual, ANL-95/11 – Revision 3.0.0, Argonne National Laboratory, 2008
[25] Satish Balay, William D. Gropp, Lois Curfman McInnes and Barry F. Smith, Efficient
Management of Parallelism in Object Oriented Numerical Software Libraries, Modern
Software Tools in Scientific Computing, Birkhäuser Press, 1997: 163–202
[26] P. Bastian, K. Birken, K. Johannsen, S. Lang, N. Neuss, H. Rentz-Reichert and C.
Wieners, UG–A Flexible Software Toolbox for Solving Partial Differential Equations,
Computing and Visualization in Science, 1997(1): 27–40
参考文献
91
[27] P. Bastian, K. Birken, K. Johannsen, S. Lang, V. Reichenberger, C. Wieners, G. Wittum,
and C. Wrobel, A parallel software-platform for solving problems of partial differential
equations using unstructured grids and adaptive multigrid methods, High performance
computing in science and engineering, Springer, 1999: 326–339
[28] P.R. Amestoy, I.S. Duff and J.-Y. L’Excellent, Multifrontal parallel distributed symmetric
and unsymmetric solvers, Methods in Appl. Mech. Eng., 2000(184): 501–520.
[29] P.R. Amestoy, I.S. Duff, J.-Y. L’Excellent and J. Koster, A fully asynchronous multifrontal solver using distributed dynamic scheduling, SIAM Journal on Matrix Analysis
and Applications, 2001(23[1]): 15–41
[30] Xiaoye S. Li and James W. Demmel, SuperLU DIST: A Scalable Distributed-Memory
Sparse Direct Solver for Unsymmetric Linear Systems, ACM Trans. Mathematical Software, 2003(29[2]): 110–140
[31] L. S. Blackford, J. Choi, A. Cleary, E. D’Azevedo, J. Demmel, I. Dhillon, J. Dongarra, S.
Hammarling, G. Henry, A. Petitet, K. Stanley, D. Walker and R. C. Whaley, ScaLAPACK
Users’ Guide, Society for Industrial and Applied Mathematics, 1997
[32] Kristi Maschhoff and Danny Sorensen, A portable implementation of ARPACK for distributed memory parallel architectures, Copper Mountain Conference on Iterative Methods, 1996
[33] Vicente Hernandez, Jose E. Roman and Vicente Vidal, SLEPc: A Scalable and Flexible
Toolkit for the Solution of Eigenvalue Problems, ACM Transactions on Mathematical
Software, 2005(31[3]): 351–362
[34] V. Hernandez, J. E. Roman and V. Vidal, SLEPc: Scalable Library for Eigenvalue Problem Computations, Lecture Notes in Computer Science, 2003(2565): 377–391
[35] Steve Benson, Lois Curfman McInnes, Jorge Moré, Todd Munson and Jason Sarich,
TAO User Manual (Revision 1.9), Mathematics and Computer Science Division, Argonne
National Laboratory, 2007, http://www.mcs.anl.gov/tao
[36] Peter MacNeice, Kevin M. Olson, Clark Mobarry, Rosalinda de Fainchtein and Charles
Packer, PARAMESH : A parallel adaptive mesh refinement community toolkit, Computer
Physics Communications, 2000(126): 330–354
[37] M.J. Hagger, Automatic domain decomposition on unstructured grids (DOUG), Advances in Computational Mathematics, 1997: 28–54
学位论文
92
[38] James Ahrens, Berk Geveci and Charles Law, ParaView: An End-User Tool for Large
Data Visualization, In the Visualization Handbook, Elsevier, 2005,
http://www.paraview.org
[39] W. Dörfler, A convergent adaptive algorithm for Poisson’s equation, SIAM J. Numer.
Anal, 1996(33): 1106–1124.
[40] P. Morin, R. H. Nochetto and K. G. Siebert, Convergence of Adaptive Finite Element
Methods, SIAM Review, 2002(44[2]), 631–658.
[41] R.E. Bank, The efficient implementation of local mesh refinement algorithms, Adaptive
Computational Methods for Partial Differential Equations, SIAM, 1983: 74–81
[42] R.E. Bank, PLTMG: A software package for solving elliptic partial differential equations,
Users’ guide 7.0, volume 15 of Frontiers in Applied Mathematics, SIAM, 1994
[43] Mark T. Jones and Paul E. Plassmann, Parallel algorithms for adaptive mesh refinement,
SIAM Journal on Scientific Computing, 1997(18[3]): 686–708
[44] R. E. Bank, A. H. Sherman and A. Weiser, Refinement algorithms and data structures
for regular local mesh refinement, in Scientific Computing, IMACS/North-Holland Publishing Company, Amsterdan, 1983: 3–17
[45] M. C. Rivara, Mesh refinement processes based on the generalized bisection of simplices,
SIAM Journal of Numerical Analysis, 1984(21): 604–613
[46] M. C. Rivara and P. Inostroza, Using Longest-side Bisection Techniques for the Automatic
Refinement fo Delaunay Triangulations, International Journal for Numerical Methods in
Engineering, 1997(40): 581–597
[47] M. C.Rivara and M. Venere, Cost Analysis of the longest-side (triangle bisection) Refinement Algorithms for Triangulations, Engineering with Computers, 1996(12): 224–234
[48] E. G. Sewell, A finite element program with automatic user-controlled mesh grading, in
Advances in Computer Methods for Partial Differential Equations III, New Brunswick,
1979: 8–10
[49] Rupak Biswas and Roger C. Strawn, Tetrahedral and hexahedral mesh adaptation for
CFD problems, Applied Numerical Mathematics, 1998(26[1-2]): 135–151
[50] Yoshitaka Wada and Hiroshi Okuda, Effective adaptation tenique for hexahedral mesh,
Concurrency and Computation: Practice and Experience, 2002(14[6-7]): 451–463
[51] E. Bänsch, An adaptive finite element strategy for the three dimensional time dependent
Navier-Stokes equations, J. Comput. Appl. Math., 1991(36): 3–28
参考文献
93
[52] I. Kossaczky, A Recursive Approach to Local Mesh Refinement in Two and Three dimensions, J. Comput. Appl. Math., 1994(55): 275–288
[53] D. N. Arnold, A. Mukherjee and L. Pouly, Locally Adapted Tetrahedral Meshes Using
Bisection, SIAM J. Sci. Comput., 2000(22[2]), 431–448
[54] A. Liu and B. Joe, Quality Local Refinement of Tetrahedral Meshes Based on Bisection,
SIAM J. Sci. Comput., 1995(16[6]): 1269–1291
[55] M. C. Rivara, D. Pizarro and N. Chrisochoides, Parallel refinement of tetrahedral
meshes using terminaledge bisection algorithm, Proceedings 13th International Meshing
Roundtable, Williamsbourg USA, 2004
[56] W. J. Barry, M. T. Jones and P. E. Plassmann, Parallel adaptive mesh refinement techniques for plasticity problems, Advances in Engineering Software, 1998(29[3]): 217–225
[57] Philippe P. Pébay and David C. Thompson, Parallel Mesh Refinement Without Communication, Proceedings 13th International Meshing Roundtable, Williamsbourg USA
September, 2004
[58] J. G. Castãnos and J. E. Savage, Parallel Refinement of Unstructured Meshes, Proceedings of the IASTED International Conference on Parallel and Distributed Computing
and Systems, MIT, Boston, USA, 1999
[59] PHG: Parallel Hierachical Grid, http://lsec.cc.ac.cn/phg
[60] R. D. Williams, Performance of dynamic load balancing algorithms for unstructured mesh
calculations, Concurrency: Practice and Experience, 1991(3): 457–481
[61] J. de Keyser and D. Roose, Grid partitioning by inertial recursive bisection, Report TW
174, Dept of Computer Science, Belgium, 1992
[62] P. M. Campbell, K. D. Devine, J. E. Flaherty, L. G. Gervasio and J. D. Teresco, Dynamic
octree load balancing using space-filling curves, Technical Report, CS-03-01, 2003
[63] C. Farhat, A simple and efficient automatic FEM domain decomposer, Computer and
Structures, 1988(28): 579–602
[64] C. M. Fiduccia and R. M. Mattheyses, A linear-time heuristic for improving network
partitions, ACM IEEE 19 Design Automation Conference Proceedings, 1982(1982): 175–
181
[65] C. Walshaw and M. Cross, JOSTLE: Parallel Multilevel Graph-Partitioning Software An Overview, In F. Magoules, editor, Mesh Partitioning Techniques and Domain Decomposition Techniques, Civil-Comp Ltd., 2007: 27–58
学位论文
94
[66] A. Schmidt and K. G. Siebert, ALBERTA: An Adaptive Hierarchical Finite Element
Toolbox, http://www.alberta-fem.de/
[67] Pascal J. FREY, Medit: an interactive mesh visualization software,
http://www.ann.jussieu.fr/~frey/logiciels/medit.html
[68] ANSYS Fluent, http://www.ansys.com/products/fluid-dynamics/
[69] Linbo Zhang, Tao Cui and Hui Liu, Symmetric Quadrature Rules on Triangles and
Tetrahedra, J.Comp.Math, 2009(27[1]): 89–96
[70] OpenDX: Open Visualization Data Explorer, http://www.opendx.org
[71] VTK: Visualization Toolkit, http://www.vtk.org
[72] Leonid Oliker and Rupak Biswas, PLUM: parallel load balancing for adaptive unstructured meshes, Journal of Parallel and Distributed Computing, 1998(52[2]): 150–177
[73] Dia, http://projects.gnome.org/dia/
[74] A.O. Cifuentes and A. Kalbag, A performance study of tetrahedral and hexahedral elements in 3-D finite element structure analysis, Finite Elements in Analysis and Design,
1992(12): 313–318
[75] Steven E. Benzley, Ernest Perry, Karl Merkley, Brett Clark and Greg Sjaardama, A
Comparison of all Hexagonal and all Tetrahedral Finite Element Meahes for Elastic and
Elasto-plastic Analysis
[76] Mark A.Yerry and Mark S. Shephard, Three-Dimensional Mesh Generation by Modified
Octree Technique, International Journal for Numerical Methods in Engineering, 1984(20):
1965–1990
[77] Mark S. Shephard and Marcel K. Georges, Three-Dimensional Mesh Generation by Finite
Octree Technique, International Journal for Numerical Methods in Engineering, 1991(32):
709–749
[78] C. L. Lawson, Software for C1 Surface Interpolation, Mathematical Software III, 1977:
161–194
[79] David F. Watson, Computing the Delaunay Tesselation with Application to Voronoi
Polytopes, The Computer Journal, 1981(24[2]): 167–172
[80] C. Geuzaine and J.-F. Remacle, Gmsh: a three-dimensional finite element mesh generator with built-in pre- and post-processing facilities, International Journal for Numerical
Methods in Engineering, 2009(79[11]): 1309–1331
参考文献
95
[81] J. Schöberl, NETGEN: An advancing front 2D/3D-mesh generator based on abstract
rules, Computing and Visualization in Science, 1997(1[1]): 41–52
[82] TetGen: A Quality Tetrahedral Mesh Generator and a 3D Delaunay Triangulator,
http://tetgen.berlios.de/
[83] W. F. Mitchell, Hamiltonian paths through two- and three-dimensional grids, Journal of
Research of the NIST, 2005(110): 127–136
[84] W. F. Mitchell, The Refinement-Tree Partition for Parallel Solution of Partial Differential
Equations, Journal of Research of the NIST, 1998(103): 405–414
[85] G. Heber, R. Biswas and G. R. Gao, Self-avoiding walks over two-dimensional adaptive unstructured meshes, NAS Techinical Report, NAS-98-007, NASA Ames Research
Center, 1998
[86] Hans Sagan, Space-Filling Curves, Springer-Veriag, New York, 1994
[87] L. Velho and J Gomes de Miranda, Digital halftoning with space-filling curves, Computer
Graphics, 1991(25): 81–90
[88] G. M. Morton, A computer Oriented Geodetic DataBase and a New Technique in File
Sequencing, Technical Report, Ottawa, Canada, 1966
[89] Guohua Jin and John Mellor-Crummey, Using Space-filling Curves for Computation
Reordering, In Proceedings of the Los Alamos Computer Science Institute Sixth Annual
Symposium, 2005
[90] C. J. Alpert and A. B. Kahng, Multi-way partitioning via spacefilling curves and dynamic
programming, In Proceedings of the 31st Annual Conference on Design Automation
Conference, 1994: 652–657
[91] D. Abel and D. Mark, A comparative analysis of some two-dimensional orderings, International J. of Geographical Information and Systems, 1990(4[1]): 21–31
[92] J. Bartholdi III and P. Goldsman, Vertex-labeling algorithms for the Hilbert spacefilling
curve, Software: Practice and Experience, 2001(31): 395–408
[93] C. Böhm, S. Berchtold and D. A. Keim, Seaching in hign-dimensional spaces: Index
structures for improving the performance of multimedia databases, ACM Computing
Surveys, 2001(33): 322–373
[94] A. R. Butz, Space filling curves and mathematical programming, Information and Control, 1968(12): 314–330
学位论文
96
[95] M. Challacombe, A general parallel sparse-blocked matrix multiply for linear scaling scf
theory, Computer Physics Communications, 2000(128[1-2]):93–107
[96] Y. C. Hu, A. Cox and W. Zwaenepoel, Improving fine-grained irregular shared-memory
benchmarks by data reordering, In Proceedings of the 2000 ACM/IEEE Conference on
Supercomputing, Dallas, TX, 2000
[97] G. Jin, J. Mellor-Crummey and R. Fowler, Increasing temporal locality with skewing and
recursive blocking, In Proceedings of the 2001 ACM/IEEE Conference on Supercomputing, Denver, CO, Nov 2001
[98] Y. Matias and A. Shamir, A video scrambling technique based on space filling curve, In
Advances in Cryptology–CRYPTO’89, 1987: 398–416
[99] Hwansoo Han and Chau-wen Tseng, Improving Locality For Adaptive Irregular Scientific
Codes,In 13 th Int’l Workshop on Languages and Compilers for Parallel Computing, 1999
[100] J. Mellor-Crummey, D. Whalley and K. Kennedy, Improving memory hierarchy performance for irregular applications, In Proceedings of the 13th ACM International Conference on Supercomputing, Rhodes, Greece, June 1999: 425–433
[101] J. Mellor-Crummey, D. Whalley and K. Kennedy, Improving memory hierarchy performance for irregular applications using data and computation reorderings, International
Journal of Parallel Programming, 2001(29[3]): 217–247
[102] L. K. Platzman and J. J. Bartholdi III, Spacefilling curves and the planar travelling
salesman problem, J. of the ACM, 1989(36): 719–737
[103] J. Salmon and M.Warren, Parallel out-of-core methods for n-body simulation, In Proceedings of the 8th SIAM Conference on Parallel Processing for Scientific Computing,
1997
[104] Y. Zhang and R. E. Webber, Space diffusion: An improved parallel halftoning technique
using spacefilling curves, In Proceedings of the 20th Annual Conference on Computer
Graphics and Interactive Techniques, 1993: 305–312
[105] Sierpiński Curve, http://mathworld.wolfram.com/SierpinskiCurve.html
[106] David Salomon, Data Compression: The Complete Reference, Springer-Verlag, 2000
[107] Jochen Alber and Rolf Niedermeier, On Multi-Dimensional Hilbert Indexings, Theory of
Computing Systems, Springer, 1998: 329–338
[108] G. Breinholt and C. Schierz, Algorithm 781: Generating Hilbert’s space-filling curve by
recursion, ACM Trans. on Mathematical Software, 1998(24[2]): 184–189
参考文献
97
[109] A. R. Butz, Altrnative algorithm for Hilbert’s space-filling curve, IEEE Transactions on
Computers, 1971(20): 424–426
[110] L. M. Goldschlager, Short algorithms for space-filling curves, Software—Practice and
Experience 1981(11): 99–100
[111] I. H. Witten and B. Wyvill, On the generation and use of space-filling curves, Software—
Practice and Experience 1983(13): 519–525
[112] A. J. Cole, A note on space filling curve, Software—Practice and Experience 1983(13):
1181–1189
[113] A. J. Fisher, A new algorithm for generation hilbert curves, Software: Practice and
Experience,1986(16): 5–12
[114] J. G. Griffiths, Table-driven algorithms for generating space-filling curves, ComputerAided Design 1985(17[1]): 37–41
[115] X. Liu and G. F. Schrack, Encoding and decoding the Hilbert order, Software—Practice
and Experience 1996(26[12]): 1335–1346
[116] X. Liu and G. F. Schrack, An algorithm for encoding and decoding the 3-D Hilbert order,
IEEE transactions on image processing 1997(6): 1333–1337
[117] C. Faloutsos and S. Roseman, Fractals for secondary key retrieval, In Pr oceedings of the
8th ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems,
Philadelphia, Pensylvania, USA, 1989: 247–252
[118] N. Chen, N. Wang and B. Shi, A new algorithm for encoding and decoding the Hilbert
order, Software—Practice and Experience 2007(37[8]): 897–908
[119] J. K. Lawder, Calculation of Mappings Between One and n-dimensional Values Using
the Hilbert Space-filling Curve, Technical Report, 2000
[120] S. Kamata, R. O. Eason and Y. Bandou, A new algorithm for N-dimensional Hilbert
scanning, IEEE Trans on Image Processing, 1999(8[7]): 964–973
[121] C. Li and Y. Feng, Algorithm for analyzing n-dimensional Hilbert curve, vol. 3739,
Springer Berlin/Heidelberg, 2005: 657–662
[122] Robert W. Doran, The Gray Code, Journal of Universal Computer Science, 2007(13[11]):
1573–1597
学位论文
98
[123] H. Liu, Dynamic load balancing on adaptive unstructured meshes, High Performance
Computing and Communications, 10th IEEE International Conference on 2008(0): 870–
875
[124] A. Schmidt and K. G. Siebert, ALBERTA: An Adaptive Hierarchical Finite Element
Toolbox, http://www.alberta-fem.de/
[125] Bruce Hendrickson and Karen Devine, Dynamic load balancing in computational mechanics, Computer Methods in Applied Mechanics and Engineering, 2000(184): 485–500
[126] P. K. Jimack, An overview of dynamic load-balancing for parallel adaptive computational mechanics codes, Parallel and Distributed Processing for Computational Mechanics, Saxe-Coburg Publications, 1997
[127] Vol Nr and Per-Olof Fjällström, Algorithms for Graph Partitioning: A Survey, 1998
[128] Y. F. Hu and R. J. Blake, Load balancing for unstructured mesh applications, Progress
in computer research, 2001: 117–148
[129] K. D. Devine, E. G. Boman, R. T. Heaphy, B. A. Hendrickson, J. D. Teresco, J. Faik, J.
E. Flaherty and L. G. Gervasio, New challenges in dynamic load balancing, Appl. Numer.
Math, 2004(52): 133–152
[130] P. M. Campbell, The performance of an octree load balancer for parallel adaptive finite
element computation, Master’s thesis, Computer Science Dept., Rensselaer Polytechnic
Institute, Troy, NY, 2001
[131] Umit Catalyurek and Cevdet Aykanat, Decomposing Irregularly Sparse Matrices for
Parallel Matrix-Vector Multiplication, Proceedings of the Third International Workshop
on Parallel Algorithms for Irregularly Structured Problems, 1996: 75–86
[132] T. Bui and C. Jones, A heuristic for reducing Fill in sparse matrix factorization, in Proc.
6th SIAM Conf. Parallel Processing for Scientific Computing, SIAM, 1993: 445–452
[133] B. Hendrickson and R. Leland, A multilevel algorithm for partitioning graphs, in Proc.
Supercomputing’95, ACM, Dec. 1995
[134] G. Karypis and V. Kumar, A fast and high quality multilevel scheme for partitioning
irregular graphs, Tech. Rep. CORR 95-035, University of Minnesota, Dept. Computer
Science, Minneapolis, MN, June 1995
[135] C. Chang, T. Kurc, A. Sussman, U. Catalyurek and J. Saltz, A hypergraph-based workload partitioning strategy for parallel data aggregation, In: Proc. of 11th SIAM Conf.
Parallel Processing for Scientific Computing, SIAM, Philadelphia, PA.
参考文献
99
[136] G. Cybenko, Dynamic load balancing for distributed memory multiprocessors, Journal
of Parallel and Distributed Computing, 1989(7[2]): 279–301
[137] Karen Devine, Erik Boman, Robert Heapby, Bruce Hendrickson and Courtenay Vaughan,
Zoltan Data Management Service for Parallel Dynamic Applications, Computing in Science and Engineering, 2002(4[2]): 90–97
[138] Karen D. Devine and Joseph E. Flaherty, Parallel adaptive hp-refinement techniques for
conservation laws, Applied Numerical Mathematics, 1996(20[4]): 367–386
[139] Karen Devine and Bruce Hendrickson, Tinkertoy parallel programming: a case study with
Zoltan, International Journal of Computational Science and Engineering, 2005(1[2-4]):
64–72
[140] M. J. Berger and S. H. Bokhari, A partitioning strategy for nonuniform problems on
multiprocessors, IEEE Trans. Computers, 1987(C-36): 570–580
[141] M. T. Jones and P. E. Plassmann, Computational results for parallel unstructured mesh
computations, Computing Systems in Engineering, 1994(5): 297–309
[142] H. D. Simon, Partitioning of unstructured problems for parallel processing, in Proc. Conference on Parallel Methods on Large Scale Structural Analysis and Physics Applications,
Pergammon Press, 1991
[143] R. Van Driessche and D. Roose, Dynamic load balancing with a spectral bisection algorithm for the constrained graph partitioning problem, in High Performance Computing
and Networking, no. 919 in Lecture Notes in Computer Science, Springer, 1995: 392–397
[144] A. Pothen, H. Simon and K. Liou, Partitioning sparse matrices with eigenvectors of
graphs, SIAM J. Matrix Anal., 1990(11): 430–452
[145] J. T. Oden, A. Patra and Y. Feng, Domain decomposition for adaptive hp finite element
methods, in Proc. Seventh Intl. Conf. Domain Decomposition Methods, State College,
Pennsylvania, October 1993
[146] A. Patra and J. T. Oden, Problem decomposition for adaptive hp finite element methods,
J. Computing Systems in Engg., 1995(6[2]): 97–109
[147] Leonid Oliker and Rupak Biswas, Efficient Load Balancing and Data Remapping for
Adaptive Grid Calculations, 9th ACM Symposium on Parallel Algorithms and Architectures, 1997: 33–42
[148] Rupak Biswas and Leonid Oliker, Experiments with Repartitioning and Load Balancing
Adaptive Meshes, 1997
学位论文
100
[149] James D. Teresco and Lida P. Ungar, A Comparison of Zoltan Dynamic Load Balancers
for Adaptive Computation, Department of Computer Science, Williams College, Technical Report CS-03-02, 2003
[150] A. Trifunovic and W. Knottenbelt, Parkway 2.0: A Parallel Multilevel Hypergraph Partitioning Tool, Proc. 19th International Symposium on Computer and Information Sciences, 2004: 789–800
[151] U. V. Catalyürek and C. Aykanat, A hypergraph-partitioning approach for coarse-grain
decomposition, In ACM/IEEE SC2001, Denver, CO, November 2001
[152] Bruce Hendrickson and Robert Leland, The Chaco User’s Guide: Version 2.0, Sandia
Tech Report SAND94–2692, 1994
[153] C. Chevalier and F. Pellegrini, PT-SCOTCH: a tool for efficient parallel graph ordering,
Parallel Computing, 2008(34[6-8]): 318–331
[154] F. Pellegrini and J. Roman, SCOTCH: A Software Package for Static Mapping by Dual
Recursive Bipartitioning of Process and Architecture Graphs, Proceedings of HPCN’96,
Brussels, Belgium. LNCS 1067, Springer, 1996: 493–498
[155] George Karypis and Vipin Kumar, Multilevel k-way Hypergraph Partitioning, 36th Design Automation Conference, 1999: 343–348
[156] Zhiming Chen and Feng Jia, An adaptive finite element method with reliable and efficient
error control for linear parabolic problems, Math. Comp., 2004(73): 1163–1197
[157] E. Zhelezina, Adaptive finite element method for the numerical simulation of electric,
magnetic and acoustic fields, Ph.D. thesis, Univ. Erlangen, 2005
[158] Martin Źitka, hp-FEM for large-scale singular 3D problems, MSc Thesis, Department of
Mathematical Sciences, University of Texas At EL PASO, May 2006
[159] Svatava Vyvialová, Analysis and optimation of a class of hierarchic finite element methods, MSc Thesis, Department of Mathematical Sciences, University of Texas At EL
PASO, May 2006
[160] I. Babuska and W. C. Rheinboldt, A posteriori error estimates for the finite element
method, International Journal for Numerical Methods in Engineering, 1978(11): 1592–
1615
[161] I. Babuska and W. C. Rheinboldt, Adaptive approaches and reliability estimates in
finite element analysis, Computer Methods in Applied Mechanics and Engineering,
1979(17/18): 519–540
参考文献
101
[162] A. K. Noor and I. Babuska, Quality assessment and control of finite element simulations,
Finite Elements and Design, 1987(3): 1–26
[163] O. C. Zhienkiewicz, Achievements and some unsolved problems of the finite element
method, International Journal for Numerical Methods in Engineering, 2000(47): 9–28
[164] J. K. Oden and L. Demkowicz, Advances in adaptive improvements: A survey of adaptive
finite element methods in computational mechanics, In Accuracy Estimates and Adaptive
Refinements in Finite Element Computations (ASME), pages 1987: 1–43
[165] M. Ainsworth and J. T. Oden, A posteriori error estimation in finite element analysis,
Computer Methods in Applied Mechanics and Engineering, 1997(142): 1–88
[166] Z. M. Chen and S. Dai, On the Efficiency of Adaptive Finite Element Methods for Elliptic Problems with Discontinuous Coefficients, SIAM Journal on Scientific Computing,
2002(24): 443–462
[167] P. Šolı́n, K. Segeth and I. Doležel, Higher-Order Finite Element Methods, Chapman &
Hall/CRC Press, 2003
[168] Randolph E. Bank, Hierarchical Bases and the Finite Element Method, Acta Numerica,
1997
[169] S. Adjerid, M. Aiffa and J. E. Flaherty, Hierarchical finite element bases for triangular
and tetrahedral elements, Computer Methods in Applied Mechanics and Engineering,
volume 2001(190[22-23]): 2925–2941
[170] Spencer J. Sherwin and George Em Karniadakis, A new triangular and tetrahedral basis
for high-order hp finite element methods, International Journal for Numerical Methods
in Engineering 1995(38[22]): 3775–3802
[171] W. F. Mitchell and M. A. McClain, A survey of hp-adaptive strategies for elliptic partial
differential equations, Accepted by Annals of the European Academy of Sciences, 2010
[172] M. Ainsworth and B. Senior, hp-finite element procedures on non-uniform geometric
meshes: adaptivity and constrained approximation, Grid Generation and Adaptive Algorithms 1999(113): 1–28
[173] E. Süli, P. Houston and Ch. Schwab, hp-finite element methods for hyperbolic problems,
The Mathematics of Finite Elements and Applications X. MAFELAP, Elsevier, 2000:
143–162.
[174] J. M. Melenk and B. I. Wohlmuth, On residual-based a-posteriori error estimation in
hp-FEM, Advances in Computational Mathematics, 2001(15): 311–331
学位论文
102
[175] T. Eibner and J. M. Melenk, An adaptive strategy for hp-FEM based on testing for
analyticity, Compute Mech, 2007(39): 575–595
[176] V. Heuveline and R. Rannacher, Duality-based adaptivity in the hp-finite element
method, J. Numer. Math, 2003(11): 1–18
[177] M. Ainsworth and B. Senior, An adaptive refinement strategy for hp-finite element computation, Applied Numerical Mathematics, 1998(26): 165–178
[178] I. Babuška and M. Suri, The optimal convergence rate of the p-version of the finite
element method, SIAM Journal on Numerical Analysis, 1987(24): 750–776
[179] I. Babuška, B. Andersson, B. Guo, J. M. Melenk and H. S. Oh, Finite element method for
solving problems with singular solutions, Journal of Computational and Applied Mathematics, 1996(1): 51–70
[180] B. Q. Guo and I. Babuška, The h-p version of the finite element method. Part 1. The
basic approximation results, Comput. Mech, 1986(1): 21–41
[181] B. Q. Guo and I. Babuška, The h-p version of the finite element method. Part 2. General
results and applications, Comput. Mech, 1986(1): 203–226
[182] W. Gui and I. Babuška, The h, p and h-p versions of the finite element method in 1
Dimension. Part III. The adaptive h-p version, Numer. Math, 1986(49): 659–683
[183] P. Houston and E. Süli, A note on the design of hp-adaptive finite element methods
for elliptic partial differential equations, Computer Methods in Applied Mechanics and
Engineering, 2005(194): 229–243
[184] C. Bernardi, N. Fiétier and R. G. Owens, An error indicator for mortar element solutions
to the Stokes problem, IMA J. Num. Anal., 2001(21): 857–886
[185] J. Valenciano and R. G. Owens, An h-p adaptive spectral element method for Stokes
flow, Appl. Numer. Math. 2000(33): 365–371
[186] S. Adjerid, M. Aiffa and J. E. Flaherty, Computational methods for singularly perturbed
systems, in: J.Cronin, R.E. O’Malley (Eds.), Singular Perturbation Concepts of Differential Equations, AMS Providence, 1998
[187] J. T. Oden and A. Patra, A parallel adaptive strategy for hp finite elements, Comput.
Methods. Appl. Mech. Engrg. 1995(121): 449–470
[188] J. T. Oden, A. Patra and Y. Feng, An hp adaptive strategy, Adaptive Multilevel and
Hierarchical Computational Strategies, 1992(157): 23–46
参考文献
103
[189] C. Mavriplis, Adaptive mesh strategies for the spectral element method, Comput. Methods. Appl. Mech. Engrg. 1994(116): 77–86
[190] L. Demkowicz, W. Rachowicz and P. Devloo, A fully automatic hp-adaptivity, J. Sci.
Comp. 2002(17[1-4]): 117–142
[191] W. Rachowicz, L. Demkowicz and J. T. Oden, Toward a universal h-p adaptive finite
element strategy, Part 3. Design of h-p meshes, Comput. Methods. Appl. Mech. Engrg.,
1989(77): 181–212
[192] P. Šolı́n and L. Demkowicz, Goal-oriented hp-adaptivity for elliptic problems, Comput.
Meth. Appl. Mech. Engrg., 2004(193): 449–468
[193] P. Šolı́n, J. Červený and I. Doležel, Arbitrary-level hanging nodes and automatic adaptivity in the hp-FEM, Math. Comput. Simulation, 2008(77): 117–132
[194] P. Houseton and E. Süli, A note on the design of hp-adaptive finite element methods for
elliptic partial differential equations, Comput. Methods Appl. Mech. Engrg., 2005(194):
229–243
[195] M. T. Jones and P. E. Plassmann, Adaptive refinement of unstructured finite-element
meshes, Finite Elements in Analysis and Design, 1997(25[12]): 41–60
[196] G. W. Zumbusch, Simultanous h-p Adaption in Multilevel Finite Elements, Zuse Institute
Berlin, TR-95-14, 1995
[197] Jack Dongarra, Ian Foster, Geoffrey Fox, Wiliam Gropp, Ken Kennedy, Linda Torczon
and Andy White, Sourcebook of Parallel Computing, Morgan Kaufmann Publishers Inc,
San Francisco, CA, USA, 2003
[198] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford Stein, Introduction to Algorithm, Second Edition, The MIT Press, 2001
[199] Wolfgang Bangerth and Rolf Rannacher, Adaptive finite element methods for differential
equations, Birkhäuser Verlag, Basel, Switzerland, 2003
[200] R. Verfürth, Adaptive finite element methods, Lecture notes, 2007
[201] K. Eriksson, D. Estep, P. Hansbo and C. Johnson, Introduction to Adaptive Methods
for Differential Equations, Acta Numerica, 1995: 1–54.
[202] R. H. Nochetto, Adaptive finite element methods for elliptic PDE, CNA Summer School,
2006.
学位论文
104
[203] C. Bernardi and R. Verfürth, Adaptive finite element methods for elliptic equations with
non-smoooth coefficients, Numerische Mathematik 2000(85): 579–608.
[204] 崔涛, PHG 中线性棱单元的实现及三维电磁时谐场问题并行自适应数值模拟实验, 硕
士论文, 中国科学院数学与系统科学研究院, 中国科学院数学与系统科学研究院, 2006
[205] 张林波, 迟学斌, 莫则尧, 李若, 并行计算导论, 清华出版社, 2006
[206] 陈国良, 并行计算 — 结构, 算法和编程, 第一版, 高等教育出版社, 1999
[207] 王烈衡, 许学军, 有限元方法的数学基础, 第一版, 科学出版社, 2004
发表文章目录
[1] Hui Liu, Dynamic Load Balancing on Adaptive Unstructured Meshes, 10th IEEE International Conference on High Performance Computing and Communications, 2008: 870–875.
[2] Linbo Zhang, Tao Cui and Hui Liu, Symmetric Quadrature Rules on Triangles and Tetrahedra, J.Comp.Math, 2009(27[1]): 89–96
[3] 刘辉, 张林波, 自适应有限元常用标记策略及其在 PHG 中的并行实现, 数值计算与计算
机应用, 2009(30[4])
[4] Hui Liu and Linbo Zhang, Existence and construction of Hamiltonian paths and cycles
on conforming tetrahedral meshes, International Journal of Computer Mathematics (accepted)
105
106
学位论文
致
谢
本文是在张林波研究员的指导下完成的, 在论文完成之际, 我衷心地感谢我
的导师张林波研究员, 感谢张老师将我引入科学计算的殿堂, 感谢张老师对我的
指导、关怀和支持. 张老师知识渊博, 治学态度严谨, 思维敏锐, 工作作风精益求
精, 诲人不倦, 平易近人, 对我影响深远. 本论文从选题到完成, 每一步都是在导
师的指导下完成的, 倾注了导师大量的心血. 我的每一点进步, 都离不开张老师的
指导和帮助. 在此, 谨向张老师表示崇高的敬意和衷心的感谢.
感谢林群院士, 曹礼群老师, 郑伟英老师在学习和工作上的支持和帮助, 同
时, 和你们的讨论也使我获益匪浅. 谢谢你们.
感谢计算数学与科学工程计算研究所, 研究所学术氛围浓厚, 学习环境舒适.
本文的工作是在科学与工程计算国家重点实验室完成的, 感谢实验室提供的优越
的科研环境和高水平的计算环境. 感谢白英老师、吴继萍老师、丁如娟老师、张
纪平老师、朱琳爱义老师的支持和帮助. 感谢数学与系统科学研究院的老师及图
书馆等部门老师.
感谢我们课题组的成员. 感谢钱莹师姐, 崔涛师兄, 冷伟, 成杰, 王昆, 谢妍,
感谢你们在学习和工作上给予的帮助和建设性的意见. 感谢钱莹师姐和崔涛师兄
在论文上给予的帮助.
在这五年中, 结识了很多的朋友, 王辛师姐、肖源明师兄、刘歆师兄、阴小波
师兄、翟方曼师姐、初晴、张黎、丁晓东、张静静、杜锐、唐明筠、付云姗、胡娟、
赵旭鹰、陈华杰、刘伟、田霞、陈景润、滕飞、聂宁明、程明厚、张娅、翟淑星、
杨熙、李明、李金, 谢谢你们在学习和生活上给予的支持、鼓励和帮助. 感谢初晴
和她老公王震在工作上给予的宝贵的建议和无私的帮助. 谢谢你们.
感谢家人的养育和支持, 感谢家人无私的奉献. 谢谢你们.
107
| 5cs.CE
|
Effective Differential Lüroth’s Theorem∗
Lisi D’Alfonso♮
Gabriela Jeronimo♯
Pablo Solernó♯
arXiv:1202.6344v2 [math.AC] 1 Jul 2013
♮ Departamento de Ciencias Exactas, Ciclo Básico Común, Universidad de Buenos Aires,
Ciudad Universitaria, 1428, Buenos Aires, Argentina
♯ Departamento de Matemática and IMAS, UBA-CONICET,
Facultad de Ciencias Exactas y Naturales,Universidad de Buenos Aires,
Ciudad Universitaria, 1428, Buenos Aires, Argentina
E-mail addresses: [email protected], [email protected], [email protected]
July 19, 2013
Abstract
This paper focuses on effectivity aspects of the Lüroth’s theorem in differential
fields. Let F be an ordinary differential field of characteristic 0 and F hui be the field
of differential rational functions generated by a single indeterminate u. Let be given
non constant rational functions v1 , . . . , vn ∈ F hui generating a differential subfield
G ⊆ F hui. The differential Lüroth’s theorem proved by Ritt in 1932 states that there
exists v ∈ G such that G = F hvi. Here we prove that the total order and degree of a
generator v are bounded by minj ord(vj ) and (nd(e + 1) + 1)2e+1 , respectively, where
e := maxj ord(vj ) and d := maxj deg(vj ). As a byproduct, our techniques enable us
to compute a Lüroth generator by dealing with a polynomial ideal in a polynomial
ring in finitely many variables.
1
Introduction
In 1876, J. Lüroth in [16] presented his famous result, currently known as the Lüroth’s
Theorem: if k ⊂ L ⊂ k(u) is an extension of fields, where k(u) is the field of rational
functions in one variable u, then L = k(v) for a suitable v ∈ L (see [23, §10.2] for a modern
proof). In 1893 G. Castelnuovo solved the same problem for rational function fields in two
variables over an algebraically closed ground field. For three variables, Lüroth’s problem
has been solved negatively.
In 1932 J.F. Ritt [19] addressed the differential version of this result: Let F be an
ordinary differential field of characteristic 0, u an indeterminate over F and Fhui the
∗
Partially supported by the following Argentinian grants: ANPCyT PICT 2007/816, UBACYT
2002010010041801 (2011-2014) and UBACYT 20020090100069 (2010-2012).
1
smallest field containing F, u and all its derivatives. Then, if G is a differential field such
that F ⊂ G ⊂ Fhui, there is an element v ∈ G such that G = Fhvi. Such an element will
be called a Lüroth generator of the extension F ⊂ G.
In fact, Ritt considered the case of a differential field F of meromorphic functions in an
open set of the complex plane and G a finitely generated extension of F. Later, E. Kolchin
in [14] and [15] gave a new proof of this theorem for any differential field of characteristic
0 and without the hypothesis of finiteness on G. Contrary to the classical setting, the
differential Lüroth problem fails in the case of two variables (see [18]). A possible weak
generalization of Lüroth’s theorem to dimension greater than one is the conjecture in control theory which states that every system linearizable by dynamic feedback is linearizable
by endogenous feedback, or in algebraic terms, that a subextension of a differentially flat
extension is differentially flat ([7], [8, Section 4.2]).
The present paper deals with quantitative aspects of the Differential Lüroth’s Theorem
and the computation of a Lüroth generator v of a finite differential field extension. More
precisely (see Propositions 5 and 19):
Theorem 1 Let F be an ordinary differential field of characteristic 0, u differentially
transcendental over F and G := FhP1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)i, where Pj , Qj ∈ F{u}
are relatively prime differential polynomials of order at most e ≥ 1 (i.e. at least one
derivative of u occurs in Pj or Qj for some j) and total degree bounded by d such that
Pj /Qj ∈
/ F for every 1 ≤ j ≤ n. Then, any Lüroth generator v of F ⊂ G can be written
as the quotient of two relatively prime differential polynomials P (u), Q(u) ∈ F{u} with
order bounded by min{ord(Pj /Qj ); 1 ≤ j ≤ n} and total degree bounded by min{(d +
1)(e+1)n , (nd(e + 1) + 1)2e+1 .
Our approach combines elements of Ritt’s and Kolchin’s proofs (mainly the introduction of the differential polynomial ideal related to the graph of the rational map
u 7→ (Pj /Qj )1≤j≤n ) with estimations concerning the order and the differentiation index of
differential ideals developed in [21], [4] and [5]. These estimations allow us to reduce the
problem of computing a Lüroth generator to a Gröbner basis computation in a polynomial
ring in finitely many variables (see Remark 17).
An algorithmic version of Ritt’s proof of the differential Lüroth’s Theorem is given
in [9]. The authors propose a deterministic algorithm that relies on the computation of
ascending chains by means of the Wu-Ritt’s zero decomposition algorithm; however, no
quantitative questions on the order or the degree of the Lüroth generator are addressed.
For effectiveness considerations of the classical not differential version, we refer the interested reader to [17], [22], [1], [10], [11], [3].
This paper is organized as follows. In Section 2 we introduce the notations, definitions and previous results from differential algebra (mainly concerning the order and the
differentiation index) needed in the rest of the paper. In Section 3 we present a straightforward optimal upper bound on the order of any Lüroth generator and we discuss some
ingredients that appear in the classical proofs of Lüroth’s Theorem by Ritt and Kolchin
2
and that we will use in our arguments. In Section 4, by means of estimates on the differentiation index and the order of an associated DAE system, we reduce the computation
of a Lüroth generator to an elimination problem in effective classical algebraic geometry;
as a byproduct, we obtain upper bounds for the degree of a Lüroth generator. Finally, in
Section 5 we show two simple examples illustrating our constructions.
2
Preliminaries
In this section we introduce the notation used throughout the paper and recall some
definitions and results from differential algebra.
2.1
Basic definitions and notation
A differential field (F, ∆) is a field F with a set of derivations ∆ = {δi }i∈I , δi : F → F.
In this paper, all differential fields are ordinary differential fields; that is to say, they are
equipped with only one derivation δ; for instance, F = Q, R or C with δ = 0, or F = Q(t)
with the usual derivation δ(t) = 1. For this reason, we will simply write differential field
(instead of ordinary differential field).
Let (F, δ) be a differential field of characteristic 0.
The ring of differential polynomials in α indeterminates z := z1 , . . . , zα , which is denoted by F{z1 , . . . , zα } or simply F{z}, is defined as the commutative polynomial ring
(p)
F[zj , 1 ≤ j ≤ α, p ∈ N0 ] (in infinitely many indeterminates), extending the derivation
(i)
(i+1)
of F by letting δ(zj ) = zj
(i)
, that is, zj
stands for the ith derivative of zj (as cus(p)
(p)
tomarily, the first derivatives are also denoted by żj ). We write z (p) := z1 , . . . , zα and
z [p] := z, z (1) , . . . , z (p) for every p ∈ N0 .
The fraction field of F{z} is a differential field, denoted by Fhzi, with the derivation
obtained by extending the derivation δ to the quotients in the usual way. For g ∈ F{z},
(i)
the order of g with respect to zj is ord(g, zj ) := max{i ∈ N0 : zj appears in g}, and the
order of g is ord(g) := max{ord(g, zj ) : 1 ≤ j ≤ α}; this notion of order extends naturally
to Fhzi by taking the maximum of the orders of the numerator and the denominator in a
reduced representation of the rational fraction.
Given differential polynomials H := h1 , . . . , hβ ∈ F{z}, we write [H] to denote the
smallest differential ideal of F{z} containing H (i.e. the smallest ideal containing the
polynomials H and all their derivatives of arbitrary order). The minimum radical differential ideal of F{z} containing H is denoted by {H}. For every i ∈ N, we write
(i)
(i)
H (i) := h1 , . . . , hβ and H [i] := H, H (1) , . . . , H (i) .
A differential field extension G/F consists of two differential fields (F, δF ) and (G, δG )
such that F ⊆ G and δF is the restriction to F of δG . Given a subset Σ ⊂ G, FhΣi denotes
the minimal differential subfield of G containing F and Σ.
An element ξ ∈ G is said to be differentially transcendental over F if the family
of its derivatives {ξ (p) : p ∈ N0 } is algebraically independent over F; otherwise, it is
said to be differentially algebraic over F. A differential transcendence basis of G/F is a
minimal subset Σ ⊂ G such that the differential field extension G/FhΣi is differentially
algebraic. All the differential transcendence bases of a differential field extension have
3
the same cardinality (see [13, Ch. II, Sec. 9, Theorem 4]), which is called its differential
transcendence degree.
2.2
Differential polynomials, ideals and manifolds
Here we recall some definitions and properties concerning differential polynomials and
their solutions.
Let g ∈ F{z} = F{z1 , . . . , zα }. The class of g for the order z1 < z2 < · · · < zα of the
(i)
variables is defined to be the greatest j such that zj appears in g for some i ≥ 0 if g ∈
/ F,
and 0 if g ∈ F. If g is of class j > 0 and of order p in zj , the separant of g, which will
(p)
be denoted by Sg , is ∂g/∂zj and the initial of g, denoted by Ig , is the coefficient of the
(p)
highest power of zj in g.
Given g1 and g2 in F{z}, g2 is said to be of higher rank in zj than g1 if either
(p)
ord(g2 , zj ) > ord(g1 , zj ) or ord(g2 , zj ) = ord(g1 , zj ) = p and the degree of g2 in zj is
(p)
greater than the degree of g1 in zj . Finally, g2 is said to be of higher rank than g1 if g2
is of higher class than g1 or they are of the same class j > 0 and g2 is of higher rank in zj
than g1 .
We will use some elementary facts of the well-known theory of characteristic sets. For
the definitions and basic properties of rankings and characteristic sets, we refer the reader
to [13, Ch. I, §8-10].
Let H be a (not necessarily finite) system of differential polynomials in F{z}. The
manifold of H is the set of all the zeros η ∈ G α of H for all possible differential extensions
G/F.
Every radical differential ideal {H} of F{z} has a unique representation as a finite
irredundant intersection of prime differential ideals, which are called the essential prime
divisors of {H} (see [20, Ch. II, §16-17]).
For a differential polynomial g in F{z} of positive class and algebraically irreducible,
there is only one essential prime divisor of {g} which does not contain Sg ; the manifold of
this prime differential ideal is called the general solution of g (see [20, Ch. II, §12-16]).
2.3
Hilbert-Kolchin function and differentiation index
Let P be a prime differential ideal of F{z}. The differential dimension of P, denoted by
diffdim(P), is the differential transcendence degree of the extension F ֒→ Frac(F{z}/P)
(where Frac denotes the fraction field). The differential Hilbert-Kolchin function of P
with respect to F is the function HP,F : N0 → N0 defined as:
HP,F (i) := the (algebraic) transcendence degree of Frac(F[z [i] ]/(P ∩ F[z [i] ])) over F.
For i ≫ 0, this function equals the linear function diffdim(P)(i + 1) + ord(P), where
ord(P) ∈ N0 is an invariant called the order of P ([13, Ch. II, Sec. 12, Theorem 6]). The
minimum i from which this equality holds is the Hilbert-Kolchin regularity of P.
Let F be a finite set of differential polynomials contained in P of order bounded by a
non-negative integer e. Throughout the paper we assume that e ≥ 1, in other words, all
the systems we consider are actually differential but not purely algebraic.
4
Definition 2 The set F is quasi-regular at P if, for every k ∈ N0 , the Jacobian matrix
of the polynomials F, Ḟ , . . . , F (k) with respect to the variables z [e+k] has full row rank over
the fraction field of F{z}/P.
A fundamental invariant associated to ordinary differential algebraic equation systems
is the differentiation index. There are several definitions of this notion (see [4] and the
references given there), but in every case it represents a measure of the implicitness of the
given system. Here we will use the following definition, introduced in [4, Section 3], in
the context of quasi-regular differential polynomial systems with respect to a fixed prime
differential ideal P:
Definition 3 The P−differentiation index σ of a quasi-regular system F of polynomials
in F{z} of order at most e is
σ := min{k ∈ N0 : (F, Ḟ , . . . , F (k) )Pe+k ∩ F[z [e] ]Pe = [F ]P ∩ F[z [e] ]Pe },
where, for every k ∈ N0 , Pe+k := P ∩ F[z [e+k] ] (i.e. the contraction of the prime ideal P),
F[z [e+k] ]Pe+k denotes the localized ring at the prime ideal Pe+k and (F, Ḟ , . . . , F (k) )Pe+k
is the algebraic ideal generated by F, Ḟ , . . . , F (k) in F[z [e+k] ]Pe+k .
Roughly speaking, the differentiation index of the system F is the minimum number of
derivatives of the polynomials F needed to write all the relations given by the differential
ideal P up to order e.
3
Differential Lüroth’s Theorem
In [19, Chapter VIII] (see also [20] and [13]), the classical Lüroth’s Theorem for transcendental field extensions is generalized to the differential algebra framework:
Theorem 4 (Differential Lüroth’s Theorem) Let F be an ordinary differential field
of characteristic 0 and let u be differentially transcendental over F. Let G be a differential
field such that F ⊂ G ⊂ Fhui. Then, there is an element v ∈ G such that G = Fhvi.
Our goal is the following: for n > 1, let be given differential polynomials P1 , . . . , Pn ,
Q1 , . . . , Qn ∈ F{u}, with Pj /Qj ∈
/ F and Pj , Qj relatively prime polynomials for every
1 ≤ j ≤ n, and denote
G := FhP1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)i,
which is a subfield of Fhui. We want to compute a Lüroth generator of G/F, that is, a
pair of differential polynomials P, Q ∈ F{u} such that Q 6≡ 0 and G = FhP (u)/Q(u)i.
We are also interested in the study of a priori upper bounds for the orders and degrees of
both polynomials P and Q.
An optimal estimate for the order of the polynomials P and Q can be obtained by
elementary computations (see Section 3.1). However, the problem of estimating their
degrees seems to be a more delicate question which requires a more careful analysis that
we will do in the subsequent sections of the paper.
5
3.1
Bound for the order
We start by proving an upper bound for the order of a Lüroth generator.
Proposition 5 Under the previous assumptions and notation, any element v ∈ G such
that G = Fhvi satisfies ord(v) ≤ min{ord(Pj /Qj ) : 1 ≤ j ≤ n}.
Proof. Let v ∈ G be such that G = Fhvi and ε := ord(v).
For j = 1, . . . , n, let vj = Pj (u)/Qj (u). By assumption, vj ∈
/ F. Let T be a new
differential indeterminate over F. Since vj ∈ G = Fhvi, there exists Θj ∈ FhT i such that
vj = Θj (v). Let Nj = ord(Θj ). Then, ord(vj ) ≤ Nj + ε. In addition,
∂vj
∂(Θj (v)) X ∂Θj
∂Θj
∂v (i)
∂v (Nj )
(v)
=
=
=
(v)
.
∂T (i)
∂u(Nj +ε)
∂u(Nj +ε)
∂u(Nj +ε)
∂T (Nj )
∂u(Nj +ε)
i≥0
Since Nj is the order of Θj , it follows that
scendental over F, we have that
∂Θj
∂T (Nj )
∂vj
∂Θj
∂T (Nj )
6= 0, and as v is differentially tran-
(v) 6= 0. Furthermore,
∂v(Nj )
∂u(Nj +ε)
=
∂v
∂u(ε)
6= 0, since
ε = ord(v). We conclude that (Nj +ε) 6= 0 and, therefore, ord(vj ) = Nj + ε.
∂u
The proposition follows.
Note that the above proposition shows that all possible Lüroth generators v have the
same order. In fact, two arbitrary generators are related by an homographic map with
coefficients in F (see for instance [14, §1], [20, Ch. II, §44]).
3.2
Ritt’s approach
Here we discuss some ingredients which appear in the classical proofs of Theorem 4 (see
[20, 14]) and that we also consider in our approach.
Following [20, II. §39 and §40], let y be a new differential indeterminate over the
field Fhui (and, in particular, over G) and consider the differential prime ideal Σ of all
differential polynomials in G{y} vanishing at u:
Σ := {A ∈ G{y} such that A(u) = 0}.
(1)
Lemma 6 The manifold of Σ is the general solution of an irreducible differential polynomial B ∈ G{y}. More precisely, B is a differential polynomial in Σ with the lowest rank
in y.
Proof. Let B ∈ Σ be a differential polynomial with the lowest rank in y. Note that
B ∈ G{y} is algebraically irreducible, since Σ is prime. We denote the order of B by k
and the separant of B by SB = ∂B/∂y (k) . Consider the differential ideal
Σ1 (B) := {A ∈ G{y} | SB A ≡ 0 mod {B}}.
As shown in [20, II. §12], the ideal Σ1 (B) is prime; moreover, we have that A ∈ Σ1 (B) if
a A ≡ 0 mod [B] for some a ∈ N and, in particular, if A ∈ Σ (B) is of order
and only if SB
0
1
at most k = ord(B), then A is a multiple of B (see [20, II. §13]). Furthermore, Σ1 (B) is
6
an essential prime divisor of {B} and, in the representation of {B} as an intersection of
its essential prime divisors, it is the only prime which does not contain SB ([20, II. §15]).
Therefore, the manifold of Σ1 (B) is the general solution of B.
In order to prove the lemma, it suffices to show that Σ = Σ1 (B).
Let A ∈ Σ1 (B). Then, SB A ∈ {B}. Taking into account that Σ is prime, {B} ⊂ Σ,
and SB ∈
/ Σ, it follows that A ∈ Σ.
To see the other inclusion, consider a differential polynomial A ∈ Σ. By the minimality
of B, we have that A is of rank at least the rank of B. Reducing A modulo B, we obtain
a relation of the type
b c
SB
IB A ≡ R mod [B],
where b, c ∈ N0 , IB is the initial of B and R is a differential polynomial whose rank is
lower than the rank of B. Since A and B lie in the differential ideal Σ, it follows that
b I c A ∈ [B] and,
R ∈ Σ, and so, the minimality of B implies that R = 0. In particular, SB
B
c
therefore, IB A ∈ Σ1 (B). Now, IB ∈
/ Σ1 (B) since, otherwise, it would be a differential
polynomial in Σ with a rank lower than the rank of B (recall that Σ1 (B) ⊂ Σ); it follows
that A ∈ Σ1 (B).
Multiplying the polynomial B ∈ G{y} given by Lemma 6 by a suitable denominator,
we obtain a differential polynomial C ∈ F{u, y} with no factor in F{u}. The following
result is proved in [20, II. §42 and §43]:
Proposition 7 If P0 (u) and Q0 (u) are two non-zero coefficients of C (regarded as a
polynomial in F{u}{y}) such that P0 (u)/Q0 (u) ∈
/ F, the polynomial
D(u, y) := Q0 (u)P0 (y) − P0 (u)Q0 (y)
is a multiple of C by a factor in F, and G = FhP0 (u)/Q0 (u)i.
Note that, by the definition of C, the ratio between two coefficients of C coincides with
the ratio of the corresponding coefficients of B.
3.3
An alternative characterization of a Lüroth generator
Under the previous assumptions, consider the map of differential algebras defined by
ψ : F{x1 , . . . , xn , u} → F{P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u), u}
xi
7→
Pi (u)/Qi (u)
u
7→
u
Let P ⊂ F{x, u} be the kernel of the morphism ψ; then, we have an isomorphism
F{P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u), u} ≃ F{x1 , . . . , xn , u}/P.
This implies that P is a prime differential ideal and, moreover, that the fraction field of
F{x1 , . . . , xn , u}/P is isomorphic to Fhui. In addition, the previous isomorphism gives
an inclusion
F{P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)} ֒→ F{x1 , . . . , xn , u}/P,
7
and the inclusion induced from this map in the fraction fields leads to the original extension
G = FhP1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)i ֒→ Fhui.
As above, let y be a new differential indeterminate over Fhui and Σ the ideal of G{y}
introduced in (1). If A ∈ G{y} is a non-zero differential polynomial in Σ, multiplying it by
an adequate element in F{P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)}, we obtain a differential polynomial in F{P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u)}{y}, with the same rank in y as A. Taking
a representative (with respect to ψ) in F{x1 , . . . , xn } for each of its coefficients, we get
b ∈ F{x1 , . . . , xn , y}, with the same rank in y as A, such that
a differential polynomial A
b
A(x1 , . . . , xn , u) ∈ P.
Conversely, given a differential polynomial M ∈ F{x1 , . . . , xn , u} such that M ∈ P and
not every coefficient of M as a polynomial in F{x1 , . . . , xn }{u} lies in P ∩ F{x1 , . . . , xn },
the differential polynomial
f(y) := M (P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u), y) ∈ G{y}
M
(2)
is not the zero polynomial, vanishes at u and has a rank in y no higher than that of M .
We conclude that if M ∈ F{x1 , . . . , xn , u} is a differential polynomial with the lowest rank in u among all the differential polynomials as above, the associated differential
f(y) is a multiple by a factor in G of the minimal polynomial B of u over G
polynomial M
introduced in Lemma 6. Therefore, by Proposition 7, a Lüroth generator of G/F can be
f ∈ G{y} provided that this ratio does
obtained as the ratio of any pair of coefficients of M
not lie in F. Moreover:
Proposition 8 Let M ∈ F{x1 , . . . , xn , u} be a differential polynomial in P\(P∩F{x1 , . . . , xn }){u}
f(y) ∈ G{y} be as in (2). Assume that M
f ∈ F(u[ǫ] ){y}
with the lowest rank in u and let M
for a suitable non-negative integer ǫ. Consider two generic points υ1 , υ2 ∈ Qǫ+1 . Let P (y)
f(y) by substituting u[ǫ] = υ1 and
and Q(y) be the differential polynomials obtained from M
u[ǫ] = υ2 respectively. Then P (u)/Q(u) is a Lüroth generator of G/F.
Proof. By Proposition 7, we have that
f(y) = M (P1 (u)/Q1 (u), . . . , Pn (u)/Qn (u), y) = γ(Q0 (u)P0 (y) − P0 (u)Q0 (y))
M
for some γ ∈ G, and P0 , Q0 are such that G = FhP0 (u)/Q0 (u)i. Then, by means of two
specializations υ1 , υ2 of the variables u[ǫ] so that Q0 , Q1 , . . . , Qn and γ do not vanish and
P0 (υ1 )/Q0 (υ1 ) 6= P0 (υ2 )/Q0 (υ2 ), we obtain polynomials of the form
P (y) = α1 P0 (y) − β1 Q0 (y)
and
Q(y) = α2 P0 (y) − β2 Q0 (y),
where α1 , α2 , β1 , β2 ∈ F and α1 β2 −α2 β1 6= 0. The proposition follows since FhP (u)/Q(u)i =
FhP0 (u)/Q0 (u)i = G.
4
Reduction to a polynomial ring and degree bound
In this section, we will obtain upper bounds for the order and the degree of the differential
polynomial M ∈ F{x1 , . . . , xn , u} (see Proposition 8) involved in our characterization of
a Lüroth generator of G/F. These bounds imply, in particular, an upper bound for the
degrees of the numerator and the denominator of the generator (see Section 4.3).
8
4.1
Bounding the order of a minimal polynomial M
Here we estimate the order in the variables x = x1 , . . . , xn and u of a differential polynomial M (x, u) ∈ P \ (P ∩ F{x1 , . . . , xn }){u} of minimal rank in u, where P is the prime
differential ideal introduced in Section 3.3.
Remark 9 The differential dimension of P equals 1, since the fraction field of F{x1 , . . . , xn , u}/P
is isomorphic to Fhui.
Let e := max{ord(Pj /Qj ) : 1 ≤ j ≤ n}. Without loss of generality, we may assume
that ord(P1 (u)/Q1 (u)) ≥ · · · ≥ ord(Pn (u)/Qn (u)). Consider the elimination order in
F{x1 , . . . , xn , u} with x1 < · · · < xn < u.
Since P1 (u)/Q1 (u) ∈
/ F, it is transcendental over F. Now, as the variable u(e+1)
appears in the derivative (P1 (u)/Q1 (u))′ but it does not appear in P1 (u)/Q1 (u), it follows
that (P1 (u)/Q1 (u))′ is (algebraically) transcendental over F(P1 (u)/Q1 (u)). Continuing in
the same way with the successive derivatives, we conclude that P1 (u)/Q1 (u) is differentially
transcendental over F. This implies that the differential ideal P contains no differential
polynomial involving only the variable x1 .
Thus, a characteristic set of P for the considered elimination order is of the form
R1 (x1 , x2 ), R2 (x1 , x2 , x3 ), . . . , Rn−1 (x1 , . . . , xn ), Rn (x1 , . . . , xn , u).
Furthermore, Rn (x1 , . . . , xn , u) is a differential polynomial in P \ (P ∩ F{x1 , . . . , xn }){u}
with a minimal rank in u, that is, we can take M (x, u) = Rn (x, u). Following [21, Lemma
19], we may assume this characteristic set to be irreducible. Then, by [21, Theorem 24],
we have that ord(Ri ) ≤ ord(P) for every 1 ≤ i ≤ n; in particular,
ord(M ) ≤ ord(P).
(3)
The order of the differential prime ideal P can be computed exactly. In order to do
this, we introduce a system of differential polynomials that provides us with an alternative
characterization of the ideal P which enables us to compute its order.
For every j, 1 ≤ j ≤ n, denote Fj := Qj (u)xj − Pj (u) ∈ F{x, u}, and let F :=
F1 , . . . , Fn . Then we have:
Lemma 10 The ideal P is the (unique) minimal differential prime ideal of [F ] which does
not contain the product Q1 . . . Qn . Moreover, P = [F ] : (Q1 . . . Qn )∞ .
Proof. From the definitions of P and F it is clear that [F ] ⊂ P and Q1 (u) . . . Qn (u) ∈
/ P.
Moreover, since F is a characteristic set for the order in F{x, u} given by u < x1 < · · · < xn
and P ∩ F{u} = (0), we conclude that for any polynomial H ∈ P there exists N ∈ N0
such that (Q1 (u) . . . Qn (u))N H ∈ [F ] (observe that Qj is the initial and the separant of
the polynomial Fj for every j). The proposition follows.
The system F we have introduced has the following property that we will use in the
sequel (recall Definition 2):
Lemma 11 The system F is quasi-regular at P.
9
Proof. Let e be the maximum of the orders of the differential polynomials Fj , 1 ≤ j ≤ n.
For every i ∈ N, let Ji be the Jacobian matrix of the polynomials F [i−1] with respect to
the variables (x, u)[i−1+e] . We have that
(
(k)
0
if h 6= j or h = j, k < l
∂Fj
=
(k−l)
(l)
k
if h = j, k ≥ l
∂xh
l Qj
and so, for every i ∈ N, the minor of Ji corresponding to partial derivatives with respect
to the variables x[i−1] is a scalar multiple of (Q1 (u) . . . Qn (u))i , which is not zero modulo
P.
Now, we apply results from [4] in order to compute the order of P.
Proposition 12 The order of the differential ideal P equals e = max{ord(Pj (u)/Qj (u)) :
1 ≤ j ≤ n}.
Proof. Lemma 10 states that the ideal P is an essential prime divisor of [F ], where
F := F1 , . . . , Fn with Fj (x, u) := Qj (u)xj − Pj (u), 1 ≤ j ≤ n, and, as shown in Lemma 11,
the system F is quasi-regular at P. Therefore, taking into account that e is the maximum
of the orders of the polynomials in F , by [4, Theorem 12], the regularity of the HilbertKolchin function of P is at most e − 1. This implies that the order of P can be obtained
from the value of this function at e − 1; more precisely, since the differential dimension of
P equals 1, we have that
ord(P) = trdegF F[x[e−1] , u[e−1] ]/(P ∩ F[x[e−1] , u[e−1] ]) − e.
In order to compute the transcendence degree involved in the above formula, we observe
first that
F[x[e−1] , u[e−1] ]/(P ∩ F[x[e−1] , u[e−1] ]) ≃ F[(Pj /Qj )[e−1] , u[e−1] ].
It is clear that the variables u[e−1] are algebraically independent in this ring. Then,
if L = F(u[e−1] ), the order of the ideal P coincides with the transcendence degree of
L((Pj /Qj )[e−1] , j = 1, . . . , n) over L. Without loss of generality, we may assume that
ord(P1 /Q1 ) = e. Since the variable u(e) appears in P1 /Q1 , we have that L(u(e) )/L(P1 /Q1 )
is algebraic. Similarly, since u(e+1) appears in (P1 /Q1 )′ , it follows that the extension
L(u(e) , u(e+1) )/L((P1 /Q1 ), (P1 /Q1 )′ ) is algebraic. Proceeding in the same way with the
successive derivatives of P1 /Q1 , we conclude that L(u(e) , u(e+1) , . . . , u(2e−1) ) is algebraic
over L((P1 /Q1 )[e−1] ).
Since ord(Pj /Qj ) ≤ e for j = 1, . . . , n, we have that L((Pj /Qj )[e−1] , j = 1, . . . , n) ⊂
L(u(e) , u(e+1) , . . . , u(2e−1) ) and, by the arguments in the previous paragraph, this extension
is algebraic. Therefore,
trdegL L((Pj /Qj )[e−1] , j = 1, . . . , n) = trdegL L(u(e) , u(e+1) , . . . , u(2e−1) ) = e.
Then, by inequality (3) we conclude:
Proposition 13 There is a differential polynomial M ∈ P \ (P ∩ F{x1 , . . . , xn }){u} with
the lowest rank in u such that ord(M ) ≤ e.
10
4.2
Reduction to algebraic polynomial ideals
As stated in Proposition 8, the Lüroth generator of G/F is closely related with a polynomial
M (x1 , . . . , xn , u) ∈ P \ (P ∩ F{x1 , . . . , xn }){u} with the lowest rank in u.
By Proposition 13, such a polynomial M can be found in the algebraic ideal P ∩
F[x[e] , u[e] ] of the polynomial ring F[x[e] , u[e] ]. The following result will enable us to work
with a finitely generated ideal given by known generators; the key point is the estimation
of the P-differentiation index of the system F := F1 , . . . , Fn (see Definition 3):
Lemma 14 The P-differentiation index of F equals e. In particular, we have that
[F ]P ∩ F[x[e] , u[e] ]Pe = (F, Ḟ , . . . , F (e) )P2e ∩ F[x[e] , u[e] ]Pe ,
where Pe := P ∩ F[x[e] , u[e] ] and P2e := P ∩ F[x[2e] , u[2e] ].
Proof. For every k ∈ N, let Jk be the Jacobian submatrix of the polynomials F, . . . , F (k−1)
with respect to the variables (x, u)(e) , . . . , (x, u)(e+k−1) . The P-differentiation index of F
can be obtained as the minimum k such that rank(Jk+1 ) − rank(Jk ) = n holds, where the
ranks are computed over the fraction field of F{x, u}/P (see [4, Section 3.1]).
Now, since the order of the polynomials F in the variables x is zero, no derivative x(l)
with l ≥ e appears in F, Ḟ , . . . , F (e−1) . This implies that the columns of the Jacobian
submatrices Jk of these systems corresponding to partial derivatives with respect to x(l) ,
l = e, . . . , e + k − 1, are null. On the other hand, as e is the order of the system F , we may
suppose that the variable u(e) appears in the polynomial F1 and so, ∂F1 /∂u(e) 6= 0. Thus,
(i)
(i)
∂F1 /∂u(h) 6= 0 for h − i = e and ∂Fj /∂u(h) = 0 for h − i > e; that is, the matrices Jk ,
k = 1, . . . , e, are block, lower triangular matrices of the form
0 ··· 0 ∗
0 ··· 0 ⋆ 0 ··· 0 ∗
Jk = .
.
.
.
.
.
.
.
.. .. · · ·
..
.. .. ..
..
..
0 ···
0 ⋆ 0 ···
0 ⋆ ···
0 ···
0 ∗
where 0 denotes a zero column vector and ∗ a non-zero column vector. Then, rank(Jk ) = k
for k = 1, . . . , e. Moreover,
!
Je
Je+1 =
∂F (e)
··· ··· 0 ··· 0 ∗
∂x(e)
and, by the diagonal structure of
Q1 (u)
0
···
.
0
∂F (e)
Q2 (u) . .
= .
..
..
∂x(e)
..
.
.
0
···
0
0
..
.
0
Qn (u)
,
we see that rank(Je+1 ) = rank(Je ) + n.
It follows that the P-differentiation index of the system equals e.
11
Notation 15 We denote by V ⊂ A(e+1)n × A2e+1 the affine variety defined as the Zariski
closure of the solution set of the polynomial system
Y
F = 0, Ḟ = 0, . . . , F (e) = 0,
Qj 6= 0,
1≤j≤n
where F = F1 , . . . , Fn and Fj (x, u[e] ) = Qj (u[e] )xj − Pj (u[e] ) for every 1 ≤ j ≤ n.
(e)
∞
The
Q algebraic ideal corresponding to the variety V is (F, Ḟ , . . . , F ) : q , where
q := 1≤j≤n Qj . This ideal is prime, since it is the kernel of the map
[e]
F[x[e] , u[2e] ] → F[(Pj /Qj )1≤j≤n , u[2e] ]
(k)
xj
u(i)
(Pj /Qj )(k)
u(i)
7
→
7→
Then, V is an irreducible variety. Moreover, F, Ḟ , . . . , F (e) is a reduced complete intersection in {q 6= 0} and so, the dimension of V is 2e + 1.
Now, the ideal of the variety V enables us to re-interpret the ideal P ∩ F[x[e] , u[e] ]
where the minimal polynomial M lies:
Proposition 16 The following equality of ideals holds:
(F, Ḟ , . . . , F (e) ) : q ∞ ∩ F[x[e] , u[e] ] = P ∩ F[x[e] , u[e] ].
Proof. We start showing that
(F, Ḟ , . . . , F (e) ) : q ∞ ∩ F[x[e] , u[e] ] = (F, Ḟ , . . . , F (e) )P2e ∩ F[x[e] , u[e] ].
First note that (F, Ḟ , . . . , F (e) ) : q ∞ ⊂ (F, Ḟ , . . . , F (e) )P2e since q ∈
/ P. Conversely,
if h ∈ (F, Ḟ , . . . , F (e) )P2e ∩ F[x[e] , u[e] ], there is a polynomial g ∈ F[x[2e] , u[2e] ] such that
g ∈
/ P and gh ∈ (F, Ḟ , . . . , F (e) ); but g ∈
/ (F, Ḟ , . . . , F (e) ) : q ∞ , since otherwise, q N g ∈
(e)
(F, Ḟ , . . . , F ) ⊂ P for some N ∈ N contradicting the fact that q ∈
/ P and g ∈
/ P. Since
(F, Ḟ , . . . , F (e) ) : q ∞ is a prime ideal, it follows that h ∈ (F, Ḟ , . . . , F (e) ) : q ∞ .
Now, Lemma 14 implies that
(F, Ḟ , . . . , F (e) )P2e ∩ F[x[e] , u[e] ] = [F ]P ∩ F[x[e] , u[e] ].
Finally, since [F ]P F{x, u}P = PP F{x, u}P (see [4, Proposition 3]), we conclude that
[F ]P ∩ F[x[e] , u[e] ] = PP ∩ F[x[e] , u[e] ] = P ∩ F[x[e] , u[e] ];
therefore,
(F, Ḟ , . . . , F (e) ) : q ∞ ∩ F[x[e] , u[e] ] = P ∩ F[x[e] , u[e] ].
The previous proposition can be applied in order to effectively compute the polynomial
M (x1 , . . . , xn , u) ∈ P \ (P ∩ F{x1 , . . . , xn }){u} with minimal rank in u, and consequently
a Lüroth generator of the extension G/F (see Propositions 7 and 8), working over a
polynomial ring in finitely many variables:
12
Remark 17 Consider the polynomial ideal (F, Ḟ , . . . , F (e) ) : q ∞ ⊂ F[x[e] , u[2e] ]. Compute
a Gröbner basis G of this ideal for a pure lexicographic order with the variables x[e] smaller
than the variables u[2e] and u < u̇ < · · · < u(2e) . Then, the polynomial M is the smaller
polynomial in G which contains at least one variable in u[2e] .
4.3
Degree bounds
In order to estimate the degree of a minimal polynomial M (x1 , . . . , xn , u) ∈ P as in the
previous section and, therefore, by Proposition 8, also the degree of a Lüroth generator of
G/F, we will relate M to an eliminating polynomial for the algebraic variety V under a
suitable linear projection.
Let k0 ∈ N0 be the order of u in M . By Proposition 13, we have that k0 ≤ e and
so, M ∈ F[x[e] , u[k0 ] ]. Consider the fields K := Frac(F[x[e] ]/(P ∩ F[x[e] ])) and L :=
Frac(F[x[e] , u[e] ]/(P ∩ F[x[e] , u[e] ])). The minimality of the rank in u of M is equivalent
to the fact that {u[k0 −1] } ⊂ L is algebraically independent over K and M is the minimal
polynomial of u(k0 ) ∈ L over K(u[k0 −1] ).
Let Z ⊂ {x[e] } be a transcendence basis of K over F. Then, if we denote U :=
[k
{u 0 −1] } ⊂ L, we have that {Z, U } ⊂ L is algebraically independent over F and {Z, U, u(k0 ) } ⊂
L is algebraically dependent over F and, since L ⊂ F(V), the same holds in F(V).
Let N be the cardinality of {Z, U }. Consider the projection
π : V → AN +1 ,
π(x[e] , u[2e] ) = (Z, U, u(k0 ) ).
(4)
For the construction of Z, U , the dimension of π(V) equals N and so, the Zariski closure
of π(V) is a hypersurface in AN +1 . Let M0 ∈ F[Z, U, u(k0 ) ] be an irreducible polynomial
defining this hypersurface (recall that V is an irreducible variety). From [12, Lemma 2],
we have the inequality
deg M0 ≤ deg V
(5)
Note that M0 ∈ P \ (P ∩ F{x}){u} and ordu (M0 ) = k0 = ordu (M ). However, M0 may
not have the property of minimal rank in u, as the following simple example shows:
Example 1. Let G = Fhu̇, (u̇)2 i. Here:
• e = 1, n = 2
• F1 = x1 − u̇, F2 = x2 − (u̇)2 , q = 1
Following our previous construction,
V = V (F1 , F2 , Ḟ1 , Ḟ2 ) = V (x1 − u̇, x2 − (u̇)2 , ẋ1 − ü, ẍ2 − 2u̇ü)
and we can take Z = {x2 , ẋ2 } and U = {u}. Then, M0 = (u̇)2 − x2 , which has not minimal
rank in u, since u̇ − x1 vanishes over V. In fact, M = u̇ − x1 .
Even though the polynomial M0 is not the minimal rank polynomial M we are looking
for, the following relation between their degrees will be sufficient to obtain a degree upper
bound for a Lüroth generator.
13
Proposition 18 With the previous assumptions and notation, we have that deg(U,u(k0 ) ) (M ) ≤
deg(U,u(k0 ) ) (M0 ). In particular, deg(U,u(k0 ) ) (M ) ≤ deg(V).
Proof. By construction, M is the minimal polynomial of u(k0 ) over K(U ) and M0 is the
minimal polynomial of u(k0 ) over F(Z, U ). Without loss of generality, we may assume that
M and M0 are polynomials with coefficients in F having content 1 in K[U ] and F[Z, U ]
respectively. Since F(Z, U ) ⊂ K(U ), we infer that M divides M0 in K[U ][u(k0 ) ]. The
proposition follows.
Recalling that a Lüroth generator v of G/F can be obtained as the quotient of two
specializations of the variables x1 , . . . , xn and their derivatives in the polynomial M (see
Proposition 8) and that two arbitrary generators are related by an homographic map
with coefficients in F (see [14, §1], [20, Ch. II, §44]), we conclude that the degrees of
the numerator and the denominator of any Lüroth generator of G/F are bounded by the
degree of the variety V.
We can exhibit purely syntactic degree bounds for V in terms of the number n of given
generators for G/F, their maximum order e, and an upper bound d for the degrees of their
numerators and denominators.
First, since the variety V is an irreducible component of the algebraic set defined by the
(e + 1)n polynomials F, Ḟ , . . . , F (e) of total degrees bounded by d + 1 (here F = F1 , . . . , Fn
and Fj (x, u) = Qj (u)xj − Pj (u)), Bézout’s theorem (see for instance [12, Theorem 1])
implies that
deg V ≤ (d + 1)(e+1)n .
An analysis of the particular structure of the system leads to a different upper bound
for deg(V), which is not exponential in n: taking into account that V is an irreducible
variety of dimension 2e + 1, its degree is the number of points in its intersection with a
generic linear variety of codimension 2e + 1, that is,
deg V = #(V ∩ V (L1 , . . . , L2e+1 )),
where, for every 1 ≤ i ≤ 2e + 1, Li is a generic affine linear form in the variables x[e] , u[2e] ,
X
X
(k)
Li (x[e] , u[2e] ) =
aijk xj +
bik u(k) + ci .
(6)
0≤k≤2e
1≤j≤n
0≤k≤e
For every 1 ≤ j ≤ n, the equation Fj (xj , u) = 0 implies that, generic points of V
satisfy xj = Pj (u)/Qj (u). Proceeding inductively, it follows easily that, generically
(k)
xj
=
Pj (u)
Qj (u)
(k)
=
Rjk (u[e+j] )
Qj (u[e] )k+1
with deg(Rjk ) ≤ d(k + 1)
for every 1 ≤ j ≤ n, 0 ≤ k ≤ e. Substituting these formulae into (6) and clearing
denominators, we deduce that the degree of V equals the number of common solutions of
14
the system defined by the 2e + 1 polynomials
P
Y
Pn [e] [2e]
1
[2e]
e+1
Li
Li (u ) :=
Qj
,...,
,u
Q1
Qn
1≤j≤n
Y
Y
X
Y
X
e+1
(k)
Qe+1
Q
+
c
+
b
u
Qe+1
=
aijk Rjk Qje−k
i
ik
j
j
h
1≤j≤n
0≤k≤e
0≤k≤2e
h6=j
1≤j≤n
1≤j≤n
for generic coefficients aijk , bik , ci and the inequality q 6= 0. From the upper bounds for
the degrees of the polynomials Qj , Rjk , it follows that Li is a polynomial of total degree
bounded by nd(e + 1) + 1 for every 1 ≤ i ≤ 2e + 1. Therefore, Bézout’s bound implies that
deg V ≤ (nd(e + 1) + 1)2e+1 .
We conclude that:
Proposition 19 Under the previous assumptions and notation, the degrees of the numerator and the denominator of any Lüroth generator of G/F are bounded by min{(d+1)(e+1)n ,
(nd(e + 1) + 1)2e+1 }.
5
Examples
As before, let F be a differential field of characteristic 0 and u a differentially transcendental element over F.
Example 1. Let G = Fhu/u̇, u + u̇i. In this case, we have:
• e = 1, n = 2,
• F1 = u̇x1 − u, F2 = x2 − u − u̇, q = u̇,
As the ideal (F1 , F2 , Ḟ1 , Ḟ2 ) = (u̇x1 − u, x2 − u − u̇, üx1 + u̇ẋ1 − u̇, ẋ2 − u̇ − ü) is prime and
does not contain u̇, we have that (F1 , F2 , Ḟ1 , Ḟ2 ) : q ∞ = (F1 , F2 , Ḟ1 , Ḟ2 ) and, therefore,
V = V (F1 , F2 , Ḟ1 , Ḟ2 ) = V (u̇x1 − u, x2 − u − u̇, üx1 + u̇ẋ1 − u̇, ẋ2 − u̇ − ü).
The dimension of V equals 3 and {x1 , x2 , ẋ1 } is a transcendence basis of F(V) over
F. Then, k0 = 0 and so, we look for a polynomial M0 (x1 , x2 , ẋ1 , u) ∈ (F1 , F2 , Ḟ1 , Ḟ2 ). We
compute this polynomial by an elimination procedure:
M0 (x1 , x2 , ẋ1 , u) = (x1 + 1) u − x1 x2 .
Since degu (M0 ) = 1, then M = M0 . Finally, specializing (u, u̇, ü) into u1 = (1, 1, 0)
and u2 = (0, 1, 0), we compute specialization points (1, 2, 1) and (0, 1, 1) respectively for
(x1 , x2 , ẋ1 ); hence, we obtain the following Lüroth generator for G/F:
v=
2u − 2
P (u)
=
.
Q(u)
u
15
In the previous example, the polynomial M lies in the polynomial ideal (F1 , F2 ), that
is, no differentiation of the equations is needed in order to compute it and, consequently, a
Lüroth generator can be obtained as an algebraic rational function of the given generators.
However, this is not always the case, as the following example shows.
Example 2. Let G = Fhu̇, u + üi. We have:
• e = 2, n = 2,
• F1 = x1 − u̇, F2 = x2 − u − u(2) , q = 1,
Following our previous arguments, we consider the ideal
(2)
(2)
(F1 , F2 , Ḟ1 , Ḟ2 , F1 , F2 ) =
(2)
(2)
= (x1 − u̇, x2 − u − u(2) , ẋ1 − u(2) , ẋ2 − u̇ − u(3) , x1 − u(3) , x2 − u(2) − u(4) ).
This is a prime ideal of F[x[2] , u[4] ]. The variety V is the zero-set of this ideal and it has
dimension 5. A transcendence basis of F(V) over F including a maximal subset of x[2] is
(2)
(2)
{x1 , x2 , ẋ1 , ẋ2 , x2 } and the minimal polynomial of u over F(x1 , x2 , ẋ1 , ẋ2 , x2 ) is
M0 = u − x2 + ẋ1 .
Again, since degu (M0 ) = 1, then M = M0 . Two specializations of this polynomial lead us
u+a
. We conclude that G = Fhui.
to a Lüroth generator of G/F of the form v =
u+b
References
[1] C. Alonso, J. Gutiérrez, and T. Recio, A rational function decomposition algorithm
by near-separated polynomials. J. Symb. Comp. 19 (1995), 527544.
[2] P. Bürgisser, M. Clausen, M.A. Shokrollahi, Algebraic Complexity Theory,
Grundlehren der Mathematischen Wissenschaften, vol. 315, Springer, Berlin, 1997.
[3] G. Chèze, Nearly Optimal Algorithms for the Decomposition of Multivariate Rational
Functions and the Extended Lüroth’s Theorem, J. Complexity 26 (2010), no. 4, 344–
363.
[4] L. D’Alfonso, G. Jeronimo, G. Massaccesi, P. Solernó, On the index and the order of quasi-regular implicit systems of diffrential equations. Linear Algebra and its
Applications 430 (2009), pp. 2102–2122.
[5] L. D’Alfonso, G. Jeronimo, P. Solernó, On the complexity of the resolvent representation of some prime differential ideals. J. Complexity 22 (2006), no. 3, 396–430.
[6] L. D’Alfonso, G. Jeronimo, F. Ollivier, A. Sedoglavic, P. Solernó, A geometric index
reduction method for implicit systems of differential algebraic equations. J. Symbolic
Computation 46, Issue 10 (2011), 1114–1138
16
[7] M. Fliess, J. Lévine, Ph. Martin and P. Rouchon. A Lie-Bäcklund approach to equivalence and flatness of nonlinear systems. IEEE Trans. Automat. Control, 44(5): 922–
937, 1999.
[8] M. Fliess, Variations sur la notion de contrôlabilité. Quelques aspects de la théorie
du contrôle, 4786, SMF Journ. Annu., 2000, Soc. Math. France, Paris, 2000.
[9] X. Gao, T. Xu, Lüroth’s theorem in differential fields, Journal of Systems Science and
Complexity 15 (2002), no. 4, 376–383.
[10] J. Gutiérrez, R. Rubio, D. Sevilla, On multivariate rational function decomposition.
Computer algebra (London, ON, 2001). J. Symbolic Comput. 33 (2002), no. 5, 545–
562.
[11] J. Gutiérrez, R. Rubio, D. Sevilla, Unirational Fields of Transcendence Degree One
and Functional Decomposition. Proc. of the 2001 Int. Symposium on Symb. and Alg.
Comput. ISSAC’01. ACM Press, New York (2001), 167–174.
[12] J. Heintz, Definability and fast quantifier elimination in algebraically closed fields.
Theoret. Comput. Sci. 24 (1983), no. 3, 239–277.
[13] E.R. Kolchin, Differential Algebra and Algebraic Groups, Academic Press, NewYork,
1973.
[14] E.R. Kolchin, Extensions of differential fields, II. Ann. of Math. (2) 45 (1944), 358–
361.
[15] E.R. Kolchin, Extensions of differential fields, III. Bull. Amer. Math. Soc. 53 (1947).
397–401.
[16] J. Lüroth, Beweis eines Satzes über rationale curven, Math. Ann. , 9 (1876) pp.
163-165
[17] E. Netto, Über einen Lüroth gordaschen staz. Math. Ann. 46 (1895), 310–318.
[18] F. Ollivier, Une réponse négative au problème de Lüroth différentiel en dimension 2.
C.R. Acad. Sci. Paris, Serie I, t. 327, no.10, Série I, 1998, 881–886.
[19] J.F. Ritt, Differential equations from the algebraic standpoint. Amer. Math. Soc.
Colloq. Publ., Vol XIV, New York, 1932.
[20] J.F. Ritt, Differential Algebra. Amer. Math. Soc. Colloq. Publ., Vol. 33, New York,
1950.
[21] B. Sadik, A bound for the order of characteristic set elements of an ordinary prime
differential ideal and some applications, Appl. Algebra Engrg. Comm. Comput. 10,
no. 3, (2000), 251–268.
[22] T.W. Sederberg, Improperly parametrized rational curves, Computer Aided Geometric Design 3 (1986), 67–75.
[23] B. L. van der Waerden, Modern Algebra, Vol. I, Springer-Verlag, 1991.
17
| 0math.AC
|
arXiv:1703.06092v2 [math.GR] 24 May 2017
Some remarks on finitarily approximable groups
Nikolay Nikolov, Jakob Schneider, and Andreas Thom
Abstract. The concept of a C-approximable group, for a class of finite
groups C, is a common generalization of the concepts of a sofic, weakly
sofic, and linear sofic group.
Glebsky raised the question whether all groups are approximable
by finite solvable groups with arbitrary invariant length function. We
answer this question by showing that any non-trivial finitely generated
perfect group does not have this property, generalizing a counterexample
of Howie. On a related note, we prove that any non-trivial group which
can be approximated by finite groups has a non-trivial quotient that can
be approximated by finite special linear groups.
Moreover, we discuss the question which connected Lie groups can
be embedded into a metric ultraproduct of finite groups with invariant
length function. We prove that these are precisely the abelian ones,
providing a negative answer to a question of Doucha. Referring to a
problem of Zilber, we show that a the identity component of a Lie group,
whose topology is generated by an invariant length function and which
is an abstract quotient of a product of finite groups, has to be abelian.
Both of these last two facts give an alternative proof of a result of
Turing. Finally, we solve a conjecture of Pillay by proving that the
identity component of a compactification of a pseudofinite group must
be abelian as well.
All results of this article are applications of theorems on generators
and commutators in finite groups by the first author and Segal. In Section 4 we also use results of Liebeck and Shalev on bounded generation
in finite simple groups.
1. Introduction
Eversince the work of Gromov on Gottschalk’s Surjunctivity Conjecture
[11], the class of sofic groups has attracted much interest in various areas
of mathematics. Major applications of this notion arose in the work Elek
and Szabó on Kaplansky’s Direct Finiteness Conjecture [5], Lück’s Determinant Conjecture [6], and more recently in joint work of the third author
with Klyachko on generalizations of the Kervaire-Laudenbach Conjecture
and Howie’s Conjecture [15].
Despite considerable effort, no non-sofic group has been found so far. In
view of this situation, attempts have been made to provide variations of the
problem that might be more approachable. In the terminology of Holt and
1
2
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
Rees, sofic groups are precisely those groups which can be approximated by
finite symmetric groups with normalized Hamming length (in the sense of
Definition 1.6 from [27]). It is natural to vary the class of finite groups and
also the metrics that are allowed. Note that our terminology differs from
the one used in [19], where similar concepts were studied.
The strongest form of approximation is satisfied by LEF (resp. LEA)
groups. In this case, it is well known that a finitely presented group is
not approximable by finite (resp. amenable) groups with discrete length
function, i.e. it is not LEF (resp. LEA), if and only if it fails to be residually
finite (resp. residually amenable). Examples of sofic groups which fail to be
LEA (and thus also fail to be LEF) are given in [2] and [14] (see also [26]),
answering a question of Gromov [11].
In [27] the third author proved that the so-called Higman group cannot
be approximated by finite groups with commutator-contractive invariant
length function. In [13] Howie presented a group which, by a result of
Glebsky [8], turned out not to be approximable by finite nilpotent groups
with arbitrary invariant length function.
The present article provides four more results of this type (see Sections
3 and 5). However, in our setting we restrict only the classes of finite groups
and do not impose restrictions on the length functions of the approximating
groups other than being invariant (see Definitions 1 and 3).
Recently, Glebsky [9] asked whether all groups can be approximated
by finite solvable groups (in the sense of Definition 1). In Section 3 we
answer this question by establishing that each non-trivial finitely generated
perfect group is a counterexample (see Theorem 1). The key to this result
is a theorem of Segal [22] on generators and commutators in finite solvable
groups.
In Section 4, using results of the first author from [18] and of Liebeck and
Shalev from [16], we prove that any non-trivial group which is approximable
by finite groups has a non-trivial homomorphism into a metric ultraproduct
of finite simple groups of type PSLn (q) with conjugacy length function (see
Theorem 4).
In Section 5 we discuss the approximability of Lie groups by finite groups.
It is easy to see that R as a topological group is not approximable by symmetric groups, i.e. it is not continuously embeddable into a metric ultraproduct
of symmetric groups with arbitrary invariant length function (see Remark
6). Using a much deeper analysis, we show that a connected Lie group is
appoximable by finite groups (in the sense of Definition 3) precisely when it
is abelian (see Theorem 6). In Question 2.11 of [4] Doucha asked for groups
which can be equipped with an invariant length function such that they do
not embed into a metric ultraproduct of finite groups with invariant length
function. Our result implies that any compact, connected, and non-abelian
Lie group is an example of such a group. Thus the simplest example of a
topological group which is not weakly sofic, i.e. not continuously embeddable
into a metric ultraproduct of finite groups with invariant length function,
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
3
is SO3 (R). However, we remark that every linear Lie group is an abstract
subgroup of the algebraic ultraproduct of finite groups indexed over N (see
Remark 5).
Furthermore, in the same section we answer the question of Zilber [32] if
there exist a compact simple Lie group which is not a quotient of an algebraic
ultraproduct of finite groups. Indeed, we show that a Lie group which
can be equipped with an invariant length function generating its topology
and which is an abstract quotient of a product of finite groups has abelian
identity component (see Theorem 7). Hence any compact simple Lie group
fails to be approximable by finite groups in the sense of Zilber.
A slight variation of Theorem 7 also answers Question 1.2 Pillay [20].
Moreover, we point out that Theorem 6 and Theorem 7 provide an alternative proof of the main result of Turing [30].
Finally, using the same approach as for the previous two results, we
solve the conjecture of Pillay [20] that the Bohr compactification of any
pseudofinite group is abelian (see Theorem 8).
All results of Section 5 follow from a theorem on generators and commutators in finite groups of the first author and Segal [18].
2. Preliminaries
In this section we recall some basic concepts. We introduce the notion
of C-approximable abstract and topological groups, and present examples.
2.1. Metric groups and length functions. By a metric group we
mean a group equipped with a metric. However, we allow a metric to attain the value infinity (this is needed for Definition 2 to make sense). For a
left-invariant metric dG : G×G → [0, ∞] on the group G we define the corresponding length function (or norm) ℓG : G → [0, ∞] by ℓG (g) := dG (1G , g).
It inherits the following three properties from dG :
(i) ℓG (g) = 0 if and only if g = 1G ;
(ii) ℓG (g) = ℓG (g−1 ) for g ∈ G;
(iii) ℓG (gh) ≤ ℓG (g) + ℓG (h) for g, h ∈ G.
Conversely, to any such function ℓG we associate a left-invariant metric
dG : G × G → [0, ∞] defined by dG (g, h) := ℓG (g−1 h). This gives a oneto-one correspondence between length functions and left-invariant metrics.
We indicate that a length function and a left-invariant metric correspond to
each other by equipping them with the same decoration.
The length function ℓG is called invariant if it is constant on conjugacy
classes. This happens precisely when dG is bi-invariant. In this article all
length functions will be invariant (and all metrics are bi-invariant).
Let us introduce the following types of length functions on a finite group
G, which we shall use in this article: The discrete length function ℓdG is the
simplest one. It is defined by ℓdG (g) := 1 for g ∈ G \ {1G } and corresponds
to the discrete metric ddG on G. The conjugacy (pseudo) length function ℓcG
4
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
is defined by
ℓcG (g) :=
(1)
log|g G |
log|G|
for g ∈ G, where g G is the conjugacy class of g in G. It is only a proper
length function if G has trivial center. The projective rank length function
ℓpr
G is defined if G ≤ PGLn (q) for some n ∈ N and q a prime power;
ℓpr
G (g) :=
1
min{rk(1 − ĝ) | ĝ some lift of g}.
n
Finally, the Cayley length function ℓCay,S
with respect to some subset S ⊆ G
G
is defined by
ℓCay,S
(g) := min{n ∈ N | g = s1 · · · sn for si ∈ S ∪ S −1 } ∪ {∞}.
G
We call a family of length functions (ℓi )i∈I on a sequence of finite groups
(Hi )i∈I Lipschitz continuous with respect to a second family (ℓ′i )i∈I on the
same groups if there is L > 0 such that ℓi ≤ Lℓ′i (i ∈ I).
d
For example, since ℓcG , ℓpr
G ≤ ℓG for any finite group G, the conjugacy
length function and the projective rank length function are Lipschitz continuous with respect to the discrete length function (with L = 1).
If (ℓ′i )i∈I is also Lipschitz continuous with respect to (ℓi )i∈I , we call these
families Lipschitz equivalent.
For example, ℓc and ℓpr are Lipschitz equivalent on the class of nonabelian finite simple groups, which follows from [16] (see the argument at
the end of Section 4).
2.2. On C-approximable abstract groups. Now we can define metric approximation of an abstract group by a class of finite groups. Throughout the article let C be such a class.
Definition 1. An abstract group G is called C-approximable if there is
a function δ• : G \ {1G } → (0, ∞] such that for any finite subset S ⊆ G and
ε > 0 there exist a group H ∈ C, an invariant length function ℓH on H, and
a map ϕ : S → H, such that
(i) if 1G ∈ S, then ϕ(1G ) = 1H ;
(ii) if g, h, gh ∈ S, then dH (ϕ(g)ϕ(h), ϕ(gh)) < ε;
(iii) for g ∈ S \ {1G } we have ℓH (ϕ(g)) ≥ δg .
Note that the above definition differs slightly from Definition 1.6 from
the one in [27] as we impose no restrictions on the invariant length functions.
However, it is equivalent to Definition 6 from [9].
Indeed, we may even require that ℓH ≤ 1 and δ• ≡ 1 in the above definition without changing its essence. Namely, choosing ε > 0 small enough, setting c := ming∈S δg , ℓ′H := min{ℓH /c, 1}, and defining δ′ : G \ {1G } → (0, ∞]
by δ•′ := 1, we can replace δ by δ′ and ℓH by ℓ′H . So, in the sense of [12, p. 3],
if we do not impose restrictions on the length functions on the groups from C,
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
5
the terms ‘C-approximation property’, ‘discrete C-approximation property’,
and ‘strong C-approximation property’ coincide.
Moreover, similar to soficity, being C-approximable is a local property.
This is expressed in the following remark.
Remark 1. An abstract group is C-approximable if and only if every
finitely generated subgroup has this property.
Let us now present some examples of C-approximable abstract groups.
Subsequently, denote by Alt (resp. Fin) the class of finite alternating groups
(resp. the class of all finite groups). Indeed, C-approximable abstract groups
(in the above sense) can be seen as a generalization of sofic (resp. weakly
sofic) groups as it is shown in Section 2 of [9]:
Example 1. A group is sofic (resp. weakly sofic) if and only if it is
Alt-approximable (resp. Fin-approximable) as an abstract group.
Groups approximable by certain classes of finite simple groups of Lie
type have been studied in [1] and [28, 29].
Every C-group G is certainly C-approximable, since we can take H := G,
ϕ to be the restriction of the identity on G to S, and ℓH := ℓdH to be the
discrete length function on H in Definition 1. Hence Remark 1 implies:
Example 2. Every locally C-group is C-approximable as an abstract
group.
2.3. Metric ultraproducts of groups. Since there is another common equivalent characterization of C-approximable groups via metric ultraproducts of C-groups with invariant length function, we recall this concept
next (for more details on the algebraic and geometric structure of such ultraproducts see also [25] and [28, 29]).
Definition 2. Let (Hi , ℓi )i∈I be a sequence of finite groups with invariant length function and
Q be U an ultrafilter on the index set I.Q The metric
ultraproduct (H, ℓ) := U (Hi , ℓi ) is defined as the group H := i∈I Hi modulo the normal subgroup NU := {(hi ) ∈ H | limU ℓi (hi ) = 0} of null sequences
equipped with the (invariant) length function ℓ defined by ℓ(h) := limU ℓi (hi ),
where h = (hi ) is a representative.
Here it is important that length functions are [0, ∞]-valued, since otherwise ℓ would not be well-defined. Some authors use
Q a slightly different
definition by restricting H to the sequences (hi ) ∈ i∈I Hi of uniformly
bounded length, i.e. supi∈I ℓi (hi ) < ∞. However, we prefer the above definition since then the ultraproduct is always a quotient of a product of the
finite groups Hi (i ∈ I).
Another thing we mention here is that an ultraproduct as in Definition 2
is always a topological group, i.e. the group operation and taking inverses are
continuous with respect to the topology induced by ℓ. This holds because ℓ
is invariant.
6
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
Lastly, we remark that the algebraic ultraproduct of a family of finite
groups is isomorphic to the metric ultraproduct of these groups equipped
with the discrete length function with respect to the same ultrafilter. In this
sense we can view every algebraic ultraproduct as a metric ultraproduct.
Now we can point out the announced characterization of C-approximable
abstract groups via metric ultraproducts. Let us call the class C trivial if
either C = ∅ or C = {1}. Here is the promised characterization.
Lemma 1. If C is a non-trivial class, every abstract C-approximable
group
G is isomorphic to a discrete subgroup of a metric ultraproduct (H, ℓ) =
Q
U (Hi , ℓi ) of C-groups Hi with invariant length function ℓi (i ∈ I) such that
diam(Hi , ℓi ) = 1 and the distance between the images of any two different
elements of G is one. If G is countable, I can be chosen to be N with the
natural order and U to be some non-principal ultrafilter.
Conversely, any subgroup of a metric ultraproduct of C-groups with invariant length function is C-approximable as an abstract group.
The proof of this result is identical to the corresponding proof in the
sofic case, which is well known. Hence we omit it here.
2.4. On C-approximable topological groups. In view of Lemma 1
it is natural to generalize the notion of a C-approximable group to topological
groups using ultraproducts:
Definition 3. A topological group is called C-approximable if it embeds
continuously into a metric ultraproduct of C-groups with invariant length
functions.
Lemma 1 indicates the following class of examples of C-approximable
topological groups:
Example 3. Every C-approximable abstract group equipped with the
discrete topology is C-approximable as a topological group.
Conversely, a C-approximable topological group is C-approximable as an
abstract group when we ‘forget’ its topology.
To present more classes of examples, we need an auxiliary result. The
following lemma gives a sufficient condition for a metric group to be isomorphic to an ultraproduct of finite metric groups. Its proof is trivial.
Lemma 2. Let (G, ℓG ) be a group with invariant length function, I an
index set, U an ultrafilter on I, (Ki , ℓi )i∈I a sequence of finite groups with
Q
invariant length function, and (K, ℓ) := U (Ki , ℓi ) its metric ultraproduct.
(i) Assume there are mappings ϕi : G → Ki , which are isometric and a
homomorphism in the U -limit, i.e.
lim di (ϕi (g), ϕi (h)) = dG (g, h) and lim di (ϕi (g)ϕi (h), ϕi (gh)) = 0.
U
U
Then there is an isometric embedding ϕ : (G, ℓG ) ֒→ (K, ℓ) in the
ultraproduct defined by ϕ(g) := (ϕi (g)).
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
7
(ii) The
Q embedding ϕ is surjective if and only if for every (ki ) ∈ K :=
i∈I Ki there exists g ∈ G such that limU di (ϕi (g), ki ) = 0.
(iii) It surjects onto the subgroup of elements of finite length of (K, ℓ) if the
previous assertion holds for all (ki ) ∈ K with supi∈I ℓi (ki ) < ∞.
Let C P and C SP be the class of finite products of C-groups and the class
subgroups of finite products of C-groups, respectively. Now, we investigate
which profinite groups are C-approximable as topological groups. The standard example is the given by the following lemma:
Lemma
3. Let Hi (i ∈ N>0 ) be C-groups. Then the profinite group
Q
P
:=
P
i∈N Hi is isomorphic to a metric ultraproduct of C -groups and so
P
C -approximable as a topological group.
Proof. We want to apply (i) and (ii) of Lemma 2 to G := P . Equip
G with the invariant length function ℓG (h) := max{1/i | i ∈ N>0 , hi 6=
1Hi } ∪ {0}, where h = (hi ). Let I := N>0 and U be some non-principal
ultrafilter on N. Set Ki := H1 × · · · × Hi ≤ G and let ℓi be the restriction of
ℓG to Ki . Define ϕi : G → Ki in such a way that for every g ∈ G the distance
dG (ϕi (g), g) is minimal. By definition of ℓG we have that dG (ϕi (g), g) < 1/i.
Hence it is easy to verify that condition (i) of Lemma 2 is fulfilled. For (ii)
define g as the U -limit of (ki ) ∈ K (which exists by compactness). Then
limU di (ϕi (g), ki ) ≤ limU dG (ϕi (g), g) + limU dG (g, ki ) = 0. This ends the
proof.
From the previous example we derive the following result.
Lemma 4. For a pro-C SP group P the following are equivalent:
(i) P is C P -approximable as a topological group.
(ii) P is metrizable.
(iii) P is first-countable.
(iv) P is the inverse limit of a countable inverse system of C SP -groups with
all maps surjective.
(v) P is a closed topological subgroup of a countable product of C-groups.
Proof. The implications (i)⇒(ii)⇒(iii) are trivial. (iii)⇒(iv): Let B
be a countable system of open neighborhoods at 1G . For each B ∈ B we can
find an open normal subgroup N ⊆ B such that P/N is a subgroup of a C SP group, so itself a C SP -group (by Proposition 0.3.3(a) and Proposition
1.2.1
T
of [31]). Let N be the collection of these subgroups. Since B = {1P }, as
P is Hausdorff, the same holds for N . Moreover, for M, N ∈ N we have
P/(M ∩N ) ≤ P/M ×P/N , so P/(M ∩N ) is a C SP -group too. Hence we may
assume that N is closed for finite intersections and apply Proposition 1.2.2
of [31] to obtain that P is the inverse limit of the C SP -groups P/N (N ∈ N )
with respect to the natural maps P/M → P/N for M ≤ N (M, N ∈ N ).
(iv)⇒(v): By the standard construction of the inverse limit, it embeds
into a countable product of C SP -groups, which (by definition) embeds into
a countable product of C-groups. For (v)⇒(i) we only need to show that a
8
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
countable product of C-groups is C P -approximable. This is Lemma 3. The
proof is complete.
Remark 2. The previous lemma implies that if a pro-C SP group embeds
continuously into a metric ultraproduct of C P -groups with invariant length
function, then it already embeds into such an ultraproduct of countably
many groups.
We are now able to present the following important example:
Example 4. If P = hx1 , . . . , xr i is a topologically finitely generated
pro-C SP group, then P is C P -approximable as a topological group.
Proof. Indeed, P embeds
continuously into the product of all its conQ
tinuous finite quotients N P/N and finite generation implies that there
are only countably many of these. By Proposition 1.2.1 of [31] we can restrict this map to a product of subgroups of C SP -groups (which are itself
C SP -groups) such that it still is an embedding. But, the latter embeds into
a countable product of C-groups. Hence P is C P -approximable (by Lemma
4).
However, it is also simple to find examples of profinite groups that are
not approximable by finite groups:
Example 5. Uncountable products of (non-trivial) finite groups are not
metrizable and hence not approximable by finite groups.
Now we turn to Lie groups. The following example demonstrates that
connected abelian Lie groups can always be approximated by finite abelian
groups in the sense of Definition 3. Henceforth, let Abd be the class of finite
abelian groups which are a direct sum of at most d cyclic groups.
Lemma 5. Every d-dimenional connected abelian Lie group L = Rm ×
(R/Z)n (m + n = d) equipped with the ‘euclidean’ length function ℓL is isometrically isomorphic to the subgroup of elements of finite length of an ultraproduct of Abd -groups with length function and hence Abd -approximable.
Proof. We wish to apply (i) and (iii) of Lemma 2 to G := L with
euclidean length function ℓG := ℓL . Let I := N and U be some non-principal
ultrafilter on N. For i ∈ N>0 set
n
m
2
1
i2
−i −i2 + 1
×
,
,...,
Z/Z ,
Si :=
i
i
i
i
and Ki := (Z/(4i2 ))m ×(Z/(i))n . Define αi : {−i2 /i, (−i2 +1)/i, . . . , i2 /i}m →
(Z/(4i2 ))m by x 7→ ix, let βi : ( 1i Z/Z)n → (Z/(i))n be the canonical isomorphism, and set γi : Si → Ki to be the map (x, y) 7→ (αi (x), βi (y)). Moreover,
equip Ki with the unique length function that turns γi into an isometry. Let
δi : G → Si be a map such that dG (δi (g), g) is minimal for all g ∈ G. Now
define ϕi := γi ◦ δi . Clearly, dG (ϕi (g), g) tends to zero for all g ∈ G. Hence
condition (i) of Lemma 2 holds. Condition (iii) follows from compactness of
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
9
closed balls of finite radius in G by the same argument as at the end of the
proof of Lemma 3. The proof is complete.
We will see in Theorem 6 of Section 5 that connected abelian Lie groups
are the only Fin-approximable connected Lie groups.
3. On Sol-approximable groups
Subsequently, let Sol (resp. Nil) be the class of finite solvable (resp. nilpotent) groups. In this section we establish the following theorem.
Theorem 1. Any non-trivial finitely generated and perfect group is not
Sol-approximable.
As a consequence, a finite group is Sol-approximable if and only if it
is solvable: Indeed, any finite solvable group is Sol-approximable. On the
other hand, a non-solvable finite group contains a non-trivial perfect subgroup and hence cannot be Sol-approximable by Remark 1 and Theorem
1.
Initially, Howie proved in [13] that the group hx, y | x−2 y −3 , x−2 (xy)5 i
is not Nil-approximable. We mimic his proof for any non-trivial finitely
generated perfect group and then extend it by establishing that these groups
are not even Sol-approximable using techniques of Segal [22, 23].
In preparation of the proof of Theorem 1, we need an auxiliary result
from [8]. Recall that the pro-C topology on a group K is the initial topology induced by all homomorphisms to C-groups equipped with the discrete
topology. Hence the closure S of a subset S ⊆ K in this topology is characterized as follows: An element k ∈ K lies in this closure if and only if
ϕ(k) ∈ ϕ(S) for all homomorphism ϕ : K → H, where H is a C-group.
Adapting Theorem 4.3 of [8], one can prove the following theorem relating C-approximable groups to the pro-C topology on a free group of finite
rank:
Theorem 2. Let F/N be a presentation of a group G, where rk(F) < ∞.
Then, if G is C-approximable, for each finite sequence n1 , . . . , nk ∈ N it holds
F
that nF
1 · · · nk ⊆ N (in the pro-C topology on F). The converse holds if C is
closed with respect to finite products and subgroups.
Remark 3. If C is closed with respect to finite products and subgroups,
the previous theorem implies that residually C-groups are C-approximable
as abstract groups, since if G = F/N is a finitely generated such group, for
F
each finite sequence n1 , . . . , nk ∈ N we obtain nF
1 · · · nk ⊆ N = N (in the
pro-C topology on F).
In view of Theorem 2, when C is closed for subgroups, to prove the existence of a non-C-approximable group, it suffices to find a normal subgroup
N E F of a free group of finite rank, an element x ∈ F \ N , and a sequence
n1 , . . . , nk ∈ N such that ϕ(x) ∈ ϕ(n1 )H · · · ϕ(nk )H for any surjective homomorphism ϕ : F → H to a C-group.
10
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
As both classes Nil and Sol are closed with respect to subgroups, we
shall construct a situation as described before.
Subsequently, let F be freely generated by x1 , . . . , xr . Fix a presentation
F/N of some non-trivial perfect group P , and an element x ∈ F \ N . The
assumption that P is perfect is equivalent to the fact that F = F′ N . Hence
we can find n1 , . . . , nr , n ∈ N such that xi ≡ ni (i = 1, . . . , r) and x ≡ n
modulo F′ . Consider a surjective homomorphism ϕ : F → H to some finite
group H (later H will be assumed to be nilpotent resp. solvable). Writing
yi := ϕ(xi ), hi := ϕ(ni ) (i = 1, . . . , r), y := ϕ(x), and h := ϕ(n), the above
translates to yi ≡ hi (i = 1 . . . , r) and y ≡ h modulo H ′ . Clearly, h1 , . . . , hr
generate H modulo H ′ (as ϕ is surjective). Now we need a lemma. To state
it, it becomes necessary to introduce some notation. In a group G define the
commutator of two elements g, h ∈ G by [g, h] := g −1 h−1 gh = g −1 gh , for
S ⊆ G and g ∈ G write [S, g] for the set {[s, g] | s ∈ S}, and for subgroups
K, L ≤ G write [K, L] for the subgroup generated by {[k, l] | k ∈ K, l ∈ L}.
Lemma 6 (Proposition 1.2.5 from [23]). Let L E G be groups, and
suppose that G = G′ hg1 , . . . , gm i. Then
[L, G] = [L, g1 ] · · · [L, gm ][L,l G]
for all l ≥ 1. Here, [L,l G] denotes the subgroup [L, G, . . . , G] of G.
| {z }
l
Proof of Theorem 1, Part 1. We apply the previous lemma to G :=
L := H, m := r, and gi := hi (i = 1, . . . , r). Moreover, we choose l ≥ 1 to be
an integer such that γl (H) = γω (H)
T (here γl (H) is the lth term in the lower
central series of H and γω (H) = {γl (H) | l ∈ N>0 }). Hence there exist
lij , lj ∈ H (i, j = 1 . . . , r) such that yi ≡ hi [li1 , h1 ] · · · [lir , hr ] (i = 1, . . . , r)
and y ≡ h[l1 , h1 ] · · · [lr , hr ] modulo γω (H). Assuming H is nilpotent (so
γω (H) = 1), the last congruence shows that
y = ϕ(x) ∈ ϕ(n′1 )H · · · ϕ(n′kNil )H ,
Nil
where kNil = 2r + 1 and (n′j )kj=1
is a fixed sequence with entries in the set
±1
{n, n±1
1 , · · · , nr }. Thus P cannot be Nil-approximable.
To prove that P is not Sol-approximable, we need the following deeper
result of Segal on finite solvable groups:
Theorem 3 (Theorem 2.1 from [22]). Assume G is a finite solvable
group and γω (G)hg1 , . . . , gm i = G for some m ∈ N. Moreover, assume that
′
G is generated by d elements. Then there is a fixed sequence (ij )m
j=1 of
indices in {1, . . . , m}, whose entries and length m′ only depend on d and m
such that
m′
Y
[γω (G), gij ].
γω (G) =
j=1
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
11
Proof of Theorem 1, Part 2. Assume that H is solvable. We want
to apply Theorem 3 to G := H. Since ϕ is surjective, the elements y1 =
ϕ(x1 ), . . . , yr = ϕ(xr ) generate H, so we may set d := r. We still have
to define the elements g1 , . . . , gm ∈ G. From the above congruences we
conclude that the sequence
l11
−1 l1r
−1 lr1
−1 lrr
h1 , . . . , hr , (h−1
1 ) , · · · , (hr ) , · · · , (h1 ) , · · · , (hr )
is a good choice for g1 , . . . , gm . Thus m := r(r + 1) is bounded in terms of
r.
The theorem gives us, similarly as in the nilpotent case, a fixed sequence
′
±1
′′
Sol
with entries in {n, n±1
(nj )kj=1
1 , · · · , nr }, whose length kSol = kNil + 2m
is bounded in terms of r, such that
y = ϕ(x) = ϕ(n′′1 )H · · · ϕ(n′′kSol )H .
Thus P cannot be Sol-approximable.
Note that finite generation is crucial here. Indeed, there exist countably
infinite locally finite-p groups which are perfect and even characteristically
simple [17]. By Example 2, these groups are Nil-approximable (since finite
p-groups are nilpotent), but by definition they are not finitely generated. It
is known that locally finite-solvable groups cannot be non-abelian simple [21,
p. 154], but it seems to be an open problem if there exist Sol-approximable
simple groups.
4. On Fin-approximable groups
Let PSL be the class of simple groups of type PSLn (q), i.e. n ∈ N≥2
and q is a prime power and (n, q) 6= (2, 2), (2, 3), and recall that Fin is the
class of all finite groups. In this section we prove the following result.
Theorem 4. Any non-trivial finitely generated Fin-approximable group
has a non-trivial PSL-approximable quotient. In particular, every simple
Fin-approximable group is PSL-approximable.
To prove Theorem 4 we need some preparation. At first we recall a
classical lemma of Goursat [10]:
Lemma 7 (Goursat’s Lemma). Let G ≤ K × L be a subdirect product,
i.e. the restricted projection maps πK : G → K, πL : G → L are surjective.
Set M := ker(πL ) and N := ker(πK ). Then M E K, N E L, and the image
of G in K/M × L/N is the graph of an isomorphism.
We need the preceding lemma for the following auxiliary result. Recall
that a profinite group is called semisimple if it is the direct product of finite
non-abelian simple groups. Moreover, a finite group G is almost simple if it
has a unique minimal normal subgroup N which is non-abelian simple; in
this case N E G ≤ Aut(N ).
12
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
Q Lemma 8. Let G be a closed subdirect product of a profinite group A =
i∈I Ai , where Ai is almost simple (i ∈ I). Then G contains a closed normal
semisimple subgroup H such that G/H is solvable of derived length at most
three and each simple factor of H is normal in G.
Q
Proof. For J ⊆ I let πJ : A → j∈J Aj be the projection maps. Then
by Proposition 1.2.2 of [31] G is the inverse limit of the groups πJ (G) (J ⊆ I
finite) together with the natural maps πJ (G) → πJ ′ (G) for J ′ ⊆ J.
Using Goursat’s Lemma one can show by induction on |J| that for J ⊆ I
finite there exist r ∈ N and finite non-abelian simple groups S1 , . . . , Sr such
that S1 × · · · × Sr E πJ (G) ≤ Aut(S1 ) × · · · × Aut(Sr ). In this situation, for
j0 ∈ I \ J the projection πJ∪{j0 } (G) → πJ (G) either is an isomorphism or
there exists a finite non-abelian simple group Sr+1 such that S1 ×· · ·×Sr+1 E
πJ∪{j0 } (G) ≤ Aut(S1 ) × · · · × Aut(Sr+1 ) and the restriction of πJ∪{j0 } (G) →
πJ (G) to the socle S1 × · · · Sr+1 of πJ∪{j0 } is the natural projection onto
S1 × · · · × Sr (the socle of πJ (G)).
Now it is clear that G, as the inverse limit of the groups πJ (G) (J ⊆ I
finite) and the maps πJ (G) → πJ ′ (G), contains the inverse limit H of the
socles of these groups together with the restricted maps. It is routine to
check that H has the desired properties. The fact that G/H is solvable of
derived length at most three is implied by Schreier’s conjecture.
Now we can start with the proof of Theorem 4: We will prove that our
group is PSL-approximable where we endow the groups PSLn (q) with the
conjugacy metric – see Equation (1) for a definition.
If the group in the theorem is not perfect, it has a non-trivial cyclic
quotient which clearly has the desired property. So let P = F/N be perb be the
fect, where F is freely generated by x1 , . . . , xr and N E F. Let F
profinite completion of F and M := hhN iiFb be the normal closure of N in
b Identifying F with its image in the profinite completion, it follows from
F.
Theorem 2 that P is Fin-approximable if and only if N = M ∩ F, since for
a sequence n1 , . . . , nk ∈ N we have
b
b
F
F
F
nF
1 · · · nk = F ∩ n1 · · · nk ,
where the closure on the left is taken in F. This is equivalent to saying that
b
b
the map F → F/M
induces an embedding
T of P in F/M .
For a profinite group G set G0 := {O Eo G | G/O is almost simple}.
b 0M .
Claim 1. It holds that F 6≤ F
Proof. Assume the contrary. Then by perfectness of P there are yi ∈ N
such that xi F′ = yi F′ (i = 1, . . . , r). By assumption there are also zi ∈ M
b 0 = zi F
b 0 (i = 1, . . . , r). Set Y := {y ±1 , . . . , yr±1 , z ±1 , . . . , zr±1 }.
such that xi F
1
1
b 0 is closed, by definition, we have that
As F
b 0 hY i = F
b 0 hY i and
b=F
F
b ′ hY i,
b = F′ hY i = F
F
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
13
b Hence by Theorem 1.7 of [18] applied
where all closures are taken in F.
b 0 , F]
b ≥ F
b ′ . Since
b and H := F
b 0 we get that M ≥ hhY ii b ≥ [F
to G := F
0
F
b 0 M/M = F
b 0 /(F
b 0 ∩ M ) is abelian by the preceding argument, we cannot
F
b 0 M , since otherwise P would be abelian. Contradiction proving
have F ≤ F
the claim.
b F
b 0M .
Claim 1 implies that P has a non-trivial homomorphism to F/
b F
b 0 as a subdirect product of all almost simple
Apply Lemma 8 to G := F/
b
b 0 be the semisimple group provided by the
quotients of F. Let H = K/F
b
b 0 M , otherwise the image
lemma. As F/K
is solvable, we cannot have K ≤ F
b
b
of P in F/F0 M would be trivial, contradicting Claim 1.
b 0 M )/F
b 0 is a proper normal subgroup of the semisimple
Hence (K ∩ F
Q
b0 =
group H = K/F
i∈I Si , where Si (i ∈ I) are the simple factors, so by
b 0 of
Theorem 5.12 of [18] it is contained in a maximal normal subgroup L/F
the former. By the same result, K/L is isomorphic (as an abstract group) to
a metric ultraproduct of the Si with the conjugacy length function ℓi := ℓcSi
b since ℓi is left
(i ∈ I). Note that in this situation L is even normal in F,
b F
b 0 by Lemma 8 (i ∈ I).
invariant under Aut(Si ) and Si E F/
Claim 2. In this setting F 6≤ LM .
Proof. Otherwise [F, K] ≤ [LM, K] ≤ L[M, K] = L(M ∩ K) = L.
Here the first inclusion holds by assumption, whereas the second follows
from the commutator identity [lm, k] = [l, k][[l, k], m][m, k] for k ∈ K, l ∈ L
and [l, k] ∈ L, since L E K, and [[l, k], m], [m, k] ∈ [K, M ] = [M, K]. The
b 0 , and the last by the choice of L.
second last equality holds as M, K E F
b
Hence F0 [F, K] ≤ L.
b F
b 0 maps continuously on the finite
Let S be a simple factor of H. F/
discrete group Aut(S) via the conjugation action. The image of this map
clearly contains the inner automorphism, since these are induced by S itself.
b F
b 0 , which must
The elements x1 , . . . , xr generate a dense subgroup of F/
induce all inner automorphisms of S by the previous fact.
= |[S, xi0 ]| ≥ |S|1/r for some
As S has trivial center, we have |S/CS (xi0 )| Q
i0 ∈ {1, . . . , r}. Lemma 3.5 of [18] implies that ri=1 [S, xi ][S, xi −1 ] contains
the normal subsets [S, xi ]S ⊆ S for i = 1, . . . , r. Since |[S, xi0 ]S | ≥ |S|1/r , by
Theorem 1.1 of [16] there is e ∈ N only depending on r such that
!∗e
r
Y
−1
[S, xi ][S, xi ]
= S.
i=1
b 0 [F, K], since e is independent of the simple factor
This implies that K ≤ F
S. But then K ≤ L, a contradiction.
From the previous claim we deduce that P still has a non-trivial hob
b
b
momorphism to F/LM
. Since F/KM
is solvable as a quotient of F/K,
14
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
this homomorphism restricts to KM/LM , which is a non-trivial homomorphic image of the metric ultraproduct K/L. Since the latter is simple by
Proposition 3.1 of [25], we are only left to show that K/L, which is a metric
ultraproduct of the sequence (Si )i∈I of finite simple groups from above with
conjugacy length function with respect to some ultrafilter U , embeds into a
metric ultraproduct of groups PSLni (qi ) equipped with the conjugacy length
function ℓci (i ∈ I), since then P would have the same property.
Let us briefly sketch the argument for this: Firstly, if the limit of the
ranks of the groups Si (i ∈ I) is bounded along the ultrafilter U (where the
rank of the alternating group An is defined to be n and the sporadic groups
are also considered as groups of bounded rank) the resulting ultraproduct
will be a simple group of Lie type over a pseudofinite field k or an alternating
group An , respectively. In the first case it clearly embeds into PSLn (k) for
n ∈ N appropriately chosen. However, the latter is a metric ultraproduct of
groups PSLn (qi ) with conjugacy length function ℓci (i ∈ I) for some sequence
(qi )i∈I of prime powers. The second case is similar.
Hence we may assume that our ultraproduct does not involve finite simple groups from families of bounded rank.
We can further assume that it contains no alternating groups as we
can replace any alternating group An by PSLn (q) for q = pe a power of a
prime with p > n. Namely, the natural embedding An ֒→ PSLn (q), where
PSLn (q) is equipped with the projective rank length function ℓpr
n , induces
Cay,{τ }
the Cayley length function ℓn
on An with respect to the conjugacy
class of a transposition of the ambient symmetric group Sn . The latter is
Lipschitz equivalent to the conjugacy class length function by Theorem 2.15
of [25].
Hence we can assume that all groups Si (i ∈ I) are classical Chevalley
or Steinberg groups. But it follows from Lemmas 5.4, 6.4, and the end of
Section 7 and Theorem 1.1 of [16] that the conjugacy length function and
the projective rank function (coming from a natural embedding in some
PSLn (q) ≤ PGLn (q)) on such groups are also Lipschitz equivalent.
Hence we can embed our ultraproduct K/L into an ultraproduct of
groups PSLni (qi ) equipped with the projective length function ℓpr
i (i ∈ I).
But by the former Lipschitz equivalence, ℓpr
can
be
replaced
by
the conjui
c
gacy length function ℓi (i ∈ I). This ends the proof.
5. On the approximability of Lie groups
In this section we utilize the following theorem of the first author and
Segal to deduce two results concerning the approximability of Lie groups by
finite groups and one result on compactifications of pseudofinite groups.
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
15
Theorem 5 (Theorem 1.2 of [18]). Let g1 , . . . , gm be a symmetric generating set for the finite group G. If K E G, then
∗e
m
Y
[K, gj ] ,
[K, G] =
j=1
where e only depends on m.
Remark 4. It was remarked in [8] that it was an open problem at the
time of writing to decide whether a finite product of conjugacy classes in a
non-abelian free group is always closed in the profinite topology.
It is a rather straightforward consequence of Theorem 5 that this is not
the case. Indeed, the theorem implies that in F = hx1 , . . . , xm i the profinite
−1
closure of the product of the 2me conjugacy classes of x−1
1 , x1 , . . . , xm , xm
contains the entire commutator subgroup, but it is a well known fact (see
Theorem 3.1.2 of [23]) that the commutator width in this group is infinite
if m > 1.
This implication was first observed by Segal and independently discovered by Gismatullin.
Actually, we shall use the following immediate corollary of Theorem 5:
Corollary 1. Let G be a quotient of a product of finite groups, then
for g, h ∈ G and N ∈ N we have
∗e
[gN , hN ] ∈ [G, g][G, g −1 ][G, h][G, h−1 ]
for some fixed constant e ∈ N.
Recall that Fin denotes the class of all finite groups. At first we prove
the following theorem:
Theorem 6. A connected Lie group is Fin-approximable as a topological
group if and only if it is abelian.
By Lemma 5 we already know that connected abelian Lie groups are
Fin-approximable. So we are only left to prove that a Fin-approximable
connected Lie group is actually abelian. This will be a consequence of the
following auxiliary result:
Q
Lemma 9. Let ϕ, ψ : R → (H, ℓ) = U (Hi , ℓi ) be continuous homomorphisms into a metric ultraproduct of finite groups Hi with invariant length
function ℓi (i ∈ I). Then for all s, t ∈ R it holds that [ϕ(s), ψ(t)] = 1H .
Let us first prove Theorem 6 using Lemma 9.
Proof of Theorem 6. Assume L is a connected Fin-approximable
Q
Lie group. Then there is an embedding ι : L ֒→ (H, ℓ) =
U (Hi , ℓi )
into a metric ultraproduct of finite groups with invariant length function.
If a, b ∈ L are in the image of the exponential map, Lemma 9 implies that
ι(a) and ι(b) commute. So as ι is injective, a and b commute. Hence by
connectedness L = L0 is abelian. This ends the proof.
16
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
We are still left to prove Lemma 9:
Proof of Lemma 9. For ε > 0 by continuity we can choose N ∈ N>0
large enough such that ℓ(ϕ(s/N )), ℓ(ψ(t/N )) < ε. Set G := H, g := ϕ(s/N )
and h := ψ(t/N ) and apply Corollary 1. This gives
∗e
[ϕ(s), ψ(t)] = [gN , hN ] ∈ [H, g][H, g−1 ][H, h][H, h−1 ] ,
whence ℓ([ϕ(s), ψ(t)]) < 8eε by invariance of ℓ and the triangle inequality.
Since ε > 0 was arbitrary, the proof is complete.
Note that Theorem 6 provides an answer to Question 2.11 of Doucha [4]
whether there are groups with invariant length function that do not embed in
a metric ultraproduct of finite groups with invariant length function. Since
every compact Lie group can be equipped with an invariant length function that generates its topology, every such group with non-abelian identity
component is an example of such a group by the previous theorem. (Indeed,
Theorem 6 does even provide topological types of groups which cannot occur
as subgroups of such a metric ultraproduct.)
Before we continue with our next result, let us state the following two
remarks.
Remark 5. In Theorem 6 the topology of the Lie group matters. Indeed,
any linear Lie group is Fin-approximable as an abstract group by Remark
1, since all its finitely generated subgroups are residually finite by Malcev’s
Theorem and hence Fin-approximable by Remark 3.
Thus any linear Lie group L is embeddable (as an abstract group) into a
metric ultraproduct of finite groups with invariant length function indexed
over, say, the partially ordered set of pairs consisting of a finite subset of L
and a positive rational number. We will now show that we can even choose
this index set to be N. Namely,
Q if L ⊆ SLn (C), then L can be embedded into
the algebraic ultraproduct U SLn (pp! ), where U is a non-principal ultrafilter on the set of prime
Indeed, this ultraproduct is isomorphic to
Q numbers.
p!
SLn (k), where k = U GF(p ) is a pseudofinite
Q field. Now it is straightforward to see that k contains the field k0 = U GF(p) together with its
algebraic closure k′ := k0alg . However, k′ is an algebraically closed field of
characteristic zero and cardinality 2ℵ0 (a result due to Shelah [24]) and hence
isomorphic to C. Note that, if we view the above algebraic ultraproduct as
a metric ultraproduct, the induced topology on SLn (C) is discrete.
Since some non-linear Lie groups admit finitely presented subgroups
which are not residually finite [3], it is clear that such embeddings cannot
exist without the assumption of linearity.
Remark 6. When one approximates with symmetric groups, one can
not even embed the real line R in a metric ultraproduct of such groups
with invariant length function. E.g. for the symmetric group Sn it can be
shown that all invariant length functions ℓ on it satisfy ℓ(σ k ) ≤ 3ℓ(σ), for
every k ∈ Z and σ ∈ Sn . Using this identity, it is simple to deduce that
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
17
the only continuous homomorphism of R into a metric ultraproduct of finite
symmetric groups with invariant length function is trivial.
Referring to the question of Zilber [32, p. 17] (also Question 1.1 of Pillay
[20]) whether a compact simple Lie group can be a quotient of the algebraic
ultraproduct of finite groups, we present the following second application of
Corollary 1:
Theorem 7. A Lie group equipped with an bi-invariant metric generating its topology that is an abstract quotient of a product of finite groups has
abelian identity component.
The proof of this result is almost identical to the proof of Theorem 6.
Proof. Let (L, ℓL ) be such a Lie group with invariant length function
and a, b ∈ L be in the image of the exponential map. For ε > 0 we find
N ∈ N>0 , g, h ∈ L such that ℓL (g), ℓL (h) < ε and gN = a, hN = b. Then
applying Corollary 1 to G := L yields
∗e
[a, b] = [gN , hN ] ∈ [L, g][L, g −1 ][L, h][L, h−1 ] ,
whence ℓL ([a, b]) < 8eε by the invariance of ℓL and the triangle inequality.
This shows that a and b commute. Hence, as L0 is generated by the image
of the exponential map, it must be abelian.
Theorem 7 implies that any compact simple Lie group, the simplest example being SO3 (R), is not a quotient of a product of finite groups, answering Zilber’s question (and hence also answers Question 1.1 of Pillay [20]).
Moreover, Theorem 7 remains valid if we replace the product of finite
groups by a pseudofinite group, i.e. a group which is a model of the theory of
all finite groups. It then also provides a negative answer to Question 1.2 of
Pillay [20], whether there is a surjective homomorphism from a pseudofinite
group to a compact simple Lie group.
Before we state the last theorem of this section, we digress briefly by
pointing out a further application of Theorems 6 and 7.
Referring to [30], we call a compact group G with compatible invariant
length function ℓG Turing-approximable if for all ε > 0 there is a finite set
Sε , a group Hε , and a bijection γε : Sε → Hε such that for all g ∈ G there
is s ∈ Sε with dG (g, s) < ε and dG (gh, γε−1 (γε (g)γε (h))) < ε for g, h ∈ Sε .
Define for g ∈ Hε
X
dG (γε−1 (f gh), γε−1 (f h)).
ℓε (g) := |Hε |−2
f,h∈Hε
It is routine to check that ℓε is an invariant length function on Hε and that
for all g ∈ Hε we have
ℓε (g) − ℓG (γε−1 (g)) < 3ε.
Set δε : G → Sε such that dG (δε (g), g) is minimal for all g ∈ G.
18
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
In this situation we can apply Lemma 2, setting I := N, U to be a
non-principal ultrafilter on N, Ki := H1/i , and ϕi := γ1/i ◦ δ1/i . Again one
checks easily that we may apply (i) and (ii) of the lemma. Hence a Turingapproximable group is isomorphic to a metric ultraproduct of finite groups
with invariant length function. Thus Theorem 6 as well as Theorem 7 imply
that a Turing-approximable Lie group has abelian identity component. This
is the main result of [30]. By Lemma 3.4 of [7] the latter condition is also
sufficient for a compact Lie group to be Turing-approximable.
Let us now turn to pseudofinite groups. By a compactification of an abstract group G, we mean a compact group C together with a homomorphism
ι : G → C with dense image. Pilay conjectured that the Bohr compactification (i.e. the universal compactification) of a pseudofinite group has abelian
identity component (Conjecture 1.7 in [20]). We answer this conjecture in
the affirmative by the following result:
Theorem 8. Let G be a pseudofinite group. Then the identity component of any compactification C of G is abelian.
The proof is again just an easy application of Corollary 1.
Proof. As G is pseudofinite it satisfies the statement of Corollary 1
(and so does its image in C). An easy compactness argument shows that
C has the same property. Now let ̺i : C → Aut(Vi ) be the irreducible
unitary representations of C and Li the image
Q of ̺i (i ∈ I). By the Peter–
Weyl Theorem, C embeds continuously into i∈I Li , and so C 0 embeds into
Q
0
i∈I Li .
But as Li is a compact quotient of C, Corollary 1 holds in it, and so as
in the proof of Theorem 7 it follows that L0i is abelian (i ∈ I). But then C 0
must be abelian as well, from the above embedding.
Acknowledgements
The second and third author want to thank Alessandro Carderi for interesting discussions. The content of this paper is part of the PhD project
of the second author. This research was supported by ERC Consolidator
Grant No. 681207.
After we finished a first version of this article and circulated it among
some experts, it was pointed out that (independently and slightly earlier)
Lev Glebsky found a solution to Zilber’s problem along the same lines.
References
[1] Goulnara Arzhantseva and Liviu Păunescu, Linear sofic groups and algebras, Transactions of the American Mathematical Society 369 (2017), no. 4, 2285–2310. ↑5
[2] Yves Cornulier, A sofic group away from amenable groups, Mathematische Annalen
350 (2011), no. 2, 269–275. ↑2
[3] Pierre Deligne, Extensions centrales non résiduellement finies de groupes
arithmétiques, CR Acad. Sci. Paris Sér. AB 287 (1978), no. 4, A203–A208. ↑16
SOME REMARKS ON FINITARILY APPROXIMABLE GROUPS
19
[4] Michal Doucha, Metric topological groups: their metric approximation and metric ultraproducts, arXiv:1601.07449 (2016), to appear in Groups, Geometry, and Dynamics.
↑2, 16
[5] Gábor Elek and Endre Szabó, Sofic groups and direct finiteness, Journal of Algebra
280 (2004), no. 2, 426–434. ↑1
[6]
, Hyperlinearity, essentially free actions and L2 -invariants. The sofic property,
Mathematische Annalen 332 (2005), no. 2, 421–441. ↑1
[7] Tsachik Gelander, Limits of finite homogeneous metric spaces, Enseign. Math. (2) 59
(2013), no. 1-2, 195–206. ↑18
[8] Lev Glebsky and Luis Manuel Rivera, Sofic groups and profinite topology on free
groups, Journal of Algebra 320 (2008), no. 9, 3512–3518. ↑2, 9, 15
[9] Lev Glebsky, Approximations of groups, characterizations of sofic groups, and equations over groups, Journal of Algebra (2016). ↑2, 4, 5
[10] Edouard Goursat, Sur les substitutions orthogonales et les divisions régulières de
l’espace, Annales scientifiques de l’École Normale Supérieure, 1889, pp. 9–102. ↑11
[11] Mikhael Gromov, Endomorphisms of symbolic algebraic varieties, Journal of the European Mathematical Society 1 (1999), no. 2, 109–197. ↑1, 2
[12] Derek F. Holt and Sarah Rees, Some closure results for C-approximable groups, arXiv
preprint arXiv:1601.01836 (2016). ↑4
[13] James Howie, The p-adic topology on a free group: A counterexample, Mathematische
Zeitschrift 187 (1984), no. 1, 25–27. ↑2, 9
[14] Aditi Kar and Nikolay Nikolov, A non-LEA sofic group, Proceedings-Mathematical
Sciences (2014), 1–5. ↑2
[15] Anton Klyachko and Andreas Thom, New topological methods to solve equations over
groups, Algebraic & Geometric Topology 17 (2017), no. 1, 331–353. ↑1
[16] Martin W. Liebeck and Aner Shalev, Diameters of finite simple groups: sharp bounds
and applications, Annals of mathematics (2001), 383–406. ↑2, 4, 13, 14
[17] Dermot H. McLain, A characteristically simple group, Mathematical Proceedings of
the Cambridge Philosophical Society, 1954, pp. 641–642. ↑11
[18] Nikolay Nikolov and Dan Segal, Generators and commutators in finite groups; abstract
quotients of compact groups, Inventiones mathematicae 190 (2012), no. 3, 513–602.
↑2, 3, 13, 15
[19] Abderezak Ould Houcine and Françoise Point, Alternatives for pseudofinite groups,
Journal of Group Theory 16 (2013), no. 4, 461–495. ↑2
[20] Anand Pillay, Remarks on compactifications of pseudofinite groups, Fundamenta
Mathematicae 236 (2017), no. 2, 193–200. ↑3, 17, 18
[21] Derek J. S. Robinson, Finiteness conditions and generalized soluble groups. Part 1+2,
Springer-Verlag, New York-Berlin, 1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 62+63. ↑11
[22] Dan Segal, Closed subgroups of profinite groups, Proceedings of the London Mathematical Society 81 (2000), no. 1, 29–54. ↑2, 9, 10
[23]
, Words: notes on verbal width in groups, Vol. 361, Cambridge University
Press, 2009. ↑9, 10, 15
[24] Saharon Shelah, On the cardinality of ultraproduct of finite sets, The Journal of Symbolic Logic 35 (1970), no. 1, 83–84. ↑16
[25] Abel Stolz and Andreas Thom, On the lattice of normal subgroups in ultraproducts of
compact simple groups, Proceedings of the London Mathematical Society 108 (2014),
no. 1, 73–102. ↑5, 14
[26] Andreas Thom, Examples of hyperlinear groups without factorization property,
Groups, Geometry, and Dynamics 4 (2010), no. 1, 195–208. ↑2
[27]
, About the metric approximation of Higman’s group, Journal of Group Theory
15 (2012), no. 2, 301–310. ↑2, 4
20
NIKOLAY NIKOLOV, JAKOB SCHNEIDER, AND ANDREAS THOM
[28] Andreas Thom and John S. Wilson, Metric ultraproducts of finite simple groups,
Comptes Rendus Mathematique 352 (2014), no. 6, 463–466. ↑5
, Some geometric properties of metric ultraproducts of finite simple groups,
[29]
arXiv preprint arXiv:1606.03863 (2016). ↑5
[30] Alan M. Turing, Finite approximations to Lie groups, Annals of Mathematics (1938),
105–111. ↑3, 17, 18
[31] John S. Wilson, Profinite groups, Vol. 19, Clarendon Press, 1998. ↑7, 8, 12
[32] Boris Zilber, Perfect infinities and finite approximation, Infinity and Truth. IMS Lecture Notes Series 25 (2014). ↑3, 17
N. Nikolov, University of Oxford, OX2 6GG Oxford, UK
E-mail address: [email protected]
J. Schneider, TU Dresden, 01062 Dresden, Germany
E-mail address: [email protected]
A. Thom, TU Dresden, 01062 Dresden, Germany
E-mail address: [email protected]
| 4math.GR
|
Temporal Multimodal Fusion
for Video Emotion Classification in the Wild
Valentin Vielzeuf∗
Stéphane Pateux
Frédéric Jurie
Orange Labs
Cesson-Sévigné, France
[email protected]
Orange Labs
Cesson-Sévigné, France
[email protected]
Normandie Univ., UNICAEN,
ENSICAEN, CNRS
Caen, France
[email protected]
arXiv:1709.07200v1 [cs.CV] 21 Sep 2017
ABSTRACT
This paper addresses the question of emotion classification. The
task consists in predicting emotion labels (taken among a set of
possible labels) best describing the emotions contained in short
video clips. Building on a standard framework – lying in describing
videos by audio and visual features used by a supervised classifier
to infer the labels – this paper investigates several novel directions.
First of all, improved face descriptors based on 2D and 3D Convolutional Neural Networks are proposed. Second, the paper explores
several fusion methods, temporal and multimodal, including a novel
hierarchical method combining features and scores. In addition, we
carefully reviewed the different stages of the pipeline and designed
a CNN architecture adapted to the task; this is important as the size
of the training set is small compared to the difficulty of the problem,
making generalization difficult. The so-obtained model ranked 4th
at the 2017 Emotion in the Wild challenge with the accuracy of
58.8 %.
KEYWORDS
Emotion Recognition; Multimodal Fusion; Recurrent Neural Network; Deep Learning;
1
Disg u st
Fear
Hap p y
Sad
Neu t ral
Su rp rise
Figure 1: The 7 emotions of the AFEW dataset: each row represents an emotion with a set of faces sampled across a representative video clip. Please note that if most of the video
clips do contain faces, some of them don’t.
INTRODUCTION AND RELATED WORK
Emotion recognition is a topic of broad and current interest, useful
for many applications such as advertising [27] or psychological
disorders understanding [35]. It is also a topic of importance for
research in other areas e.g., video summarization [37] or face normalization (expression removal). Even if emotion recognition could
appear to be an almost solved problem in laboratory controlled
conditions, there are still many challenging issues in the case of
videos recorded in the wild.
This paper focuses on the task of emotion classification in which
each video clip has to be assigned to one and only one emotion,
based on its audio/video content. The classes are usually the six
basic emotions i.e., anger, disgust, fear, happiness, sadness and surprise, in addition to the neutral class, as in the Audio-video Emotion
Recognition sub-task of the Emotion Recognition in the Wild Challenge [12]. More precisely, this paper presents our methodology,
experiments as well as the results we obtained in the 2017 edition
of the Emotion Recognition in the Wild Challenge [10].
Emotion recognition has received a lot of attention in the scientific literature. One large part of this literature deals with the
possible options for defining and representing emotions. If the use
of discrete classes [31] such as joy, fear, angriness, etc. is the most
∗ Also
An g ry
with Normandie Univ., UNICAEN, ENSICAEN, CNRS.
straightforward way to do it, it can be more interesting to represent
emotions by their degrees of arousal and valence, as proposed in [6].
In the restricted case of facial expressions, action units can also be
used, focusing on the activation of different parts of the face [15].
Links can be made between these two representations: the discrete
classes can be mapped into the arousal valence space [28] and can
be deduced from the action units [24].
Another important part of the literature focuses on the ways to
represent audio/video contents by features that can be subsequently
used by classifiers. Early papers make use of (i) hand-crafted features such as Local Binary Patterns (LBP), Gabor features, discrete
cosine transform for representing images, and Linear Predictive
Coding coefficients (LPC), relative spectral transform - linear perceptual prediction (RASTA-PLP), modulation spectrum (ModSpec)
or Enhanced AutoCorrelation (EAC) for audio, (ii) standard classifiers such as SVN or KNN for classification (see [22] for details).
[38], the winner of EmotiW’15, demonstrated the relevance of Action Units for the recognition of emotions. [3] was among the first
to propose to learn the features instead of using hand-crafted descriptors, relying on Deep Convolutional Networks. More recently,
[18], the winner of EmotiW’16, has introduced the C3D feature
which is an efficient spatio-temporal representation of faces.
The literature on emotion recognition from audio/video contents
also addresses the question of the fusion of the different modalities.
A modality can be seen as one of the signals allowing to perceive the
emotion. Among the recent methods for fusing several modalities,
we can mention the use of two-stream ConvNets [32], ModDrop
[29] or Multiple Kernel Fusion [9]. The most used modalities, in the
context of emotion recognition, are face images and audio/speech,
even if context seems also to be of tremendous importance [5]. For
instance, the general understanding of the scene, even based on simple features describing the whole image, may help to discriminate
between two candidate classes.
As most of the recent methods for emotion recognition are supervised and hence requires some training data, the availability of such
resources is becoming more and more critical. Several challenges
have collected useful data. The AVEC challenge [34] focuses on
the use of several modalities to track arousal and valence in videos
recorded in controlled conditions. The Emotionet challenge [7]
proposes a dataset of one million images annotated with action
units and partially with discrete compound emotion classes. Finally,
the Emotion in the Wild challenge [11] deals with the classification of short video clips into seven discrete classes. The videos are
extracted from movies and TV shows recorded "in the wild". The
ability to work with data recorded in realistic situations, including
occlusions, poor illumination conditions, presence of several people
or even scene breaks is indeed very important.
As aforementioned, this paper deals with our participation to
the Emotion In the Wild 2017 (EmotiW) challenge. We build on
the state-of-the-art pipeline of [18] in which i) audio features are
extracted with the OpenSmile toolkit [16], ii) two video features
are computed, one by the C3D descriptor, the other by the VGG16Face model [30] fine-tuned with FER2013 face emotion database
and introduced into a LSTM network. Each one of these 3 features
is processed by its own classifier, the output of the 3 classifiers
being combined through late fusion. Starting from this pipeline,
we propose to improve it in 3 different ways, which are the main
contributions of our approach.
First, the recent literature suggests that late fusion might not be
the optimal way to combine the different modalities (see e.g., [29]).
This paper investigates different directions, including an original
hierarchical approach allowing to combine scores (late fusion) and
features (early fusion) at different levels. It can be seen as a way
to combine information at its optimal level of description (from
features to scores). This representation addresses an important
issue of fusion, which is to ensure the preservation of unimodal
information while being able to exploit cross-modal information.
Second, we investigate several ways to better use the temporal
information in the visual descriptors. Among several contributions,
we propose a novel descriptor combining C3D and LSTM.
Third, it can be observed that the amount of training data (773
labeled short video clips) is rather small compared to the number of
the parameters of standard deep models, considering the complexity
and diversity of emotions. In this context, supervised methods
are prone to over-fitting. We show in the paper how the effect of
over-fitting can be reduced by carefully choosing the number of
parameters of the model and favoring transfer learning whenever
it is possible.
Figure 2: Comparison of the bounding boxes given by our detection/normalization procedure (color images) with those
provided with the AFEW dataset (gray-scale images), on 4
random faces of AFEW. Our bounding boxes are slightly
larger.
The rest of the paper is organized as follows: a presentation
of the proposed model is done in Section 2, detailing the different modalities and the fusion methods. Then Section 3 presents
the experimental validation as well as the results obtained on the
validation and test sets during the challenge.
2
PRESENTATION OF THE PROPOSED
APPROACH
Figure 3 presents an overview of our approach, which is inspired by
the one of [18]. Modalities we consider are extracted from audio and
video, associated with faces analyzed with two different models (2D
CNN, C3D). The two main contributions are the temporal fusion
and the novel C3D/LSTM descriptor.
On overall, our method works as follows: on the one hand, the
OpenSmile library [16] is used to produce 1582 dimensional features
used by a two-layer perceptron to predict classes as well as compact
descriptors (279-d vectors) from audio. On the other hand, video
classification is based on face analysis. After detecting the faces and
normalizing their appearance, one set of features is computed with
the VGG-16 model [30] while another set of features is obtained by
the C3D model [18]. In both cases, 4096-d features are produced.
Temporal fusion of these so-obtained descriptions is done across
the frames of the sequences, producing per modality scores and
compact descriptors (304 dimensional vectors). The 3 modalities
(audio, VGG-faces and C3D-faces) are then combined using both
the score predictions and the compact representations.
After explaining how the faces are detected and normalized,
the rest of the section gives more details on how each modality is
processed, and how the fusion is performed.
2.1
Face Detection and Alignment
The EmotiW challenge provides face detections for each frame of
each video. However, we preferred not to use these annotations
but to detect the faces ourselves. The motivation is twofold. First,
we want to be able to process any given video and not only those
of EmotiW (e.g., for adding external training data). Second, it is
necessary to master the face alignment process for processing the
2
1582 features
1
Fully
connected
OpenSmile
279 features
4096 features
frame by frame
2
Temporal
fusion
VGG-face
304 features
4096 features
Overlap
of 8
frames
C3D
by window of 16 consecutive
frames
3
Temporal
fusion
No overlap
304 features
4
Temporal
fusion
Softmax score of
size 7
Features vector
297 features
Figure 3: Our model includes audio, VGG-LSTM and C3D-LSTM (with or without overlapping) modalities. Each modality gives
its own feature vectors and scores, which can be fused using different methods. It can be noted that the dimensions of the
features vectors are chosen to have balanced contributions between modalities, making a trade off between the number of
parameters and the performance.
2.2
images of face datasets when pre-training the VGG models (see
Section 2.3 where FER dataset is used to pre-train the VGG model).
For face detection, we use an internal detector provided by Orange Labs. We found out from our observations that this face detector detects more faces (20597 versus 19845 on the validation set)
while having a lower false positive rate (179 false positives versus
908 on the validation set) than the one use to provide EmotiW
annotations.
This detector is applied frame per frame. If one or several faces
are detected, a tracking based on their relative positions allows to
define several face tracks. In a second time, an alignment based on
the landmarks [26] is done. We choose the longest face sequence in
each video and finally apply a temporal smoothing of the positions
of the faces to filter out jittering.
Figure 2 compares one of our normalized faces (after detection /
alignment) with the one provided with the challenge data.
Audio features and classifier
The audio channel of each video is fed into the OpenSmile toolkit [16]1 ,
as most of the EmotiW competitors [11], to get a description vector
of length 1582.
A commonly used approach for audio is then to learn a classifier
on top of the OpenSmile features. Support Vector Machine [18] [38]
seems to be the dominant choice for the classification, even if there
are some other approaches like Random Forests [17].
To be able to control more finely the dimensionality of the description vector (the one used later on during the fusion process),
we learn a two-layer Perceptron with reLu activation on the OpenSmile description, using batch normalisation and dropout. During
inference, we extract for each video a description vector of size 279
– the hidden layer of the perceptron – along with the softmax score.
2.3
Representing Faces with 2D Convolutional
Neural Network
A current popular method (see e.g., [18, 23]) for representing faces
in the context of emotion recognition, especially in the EmotiW
1 We
3
use the emobase2010 configuration file.
challenge, is to fine-tune a pre-trained 2D CNN such as the VGGface model [30] on the images of an emotion images dataset (e.g., the
FER 2013 dataset). Using a pre-trained model is a way to balance
the relatively small size of the EmotiW dataset (AFEW dataset).
The images of the FER 2013 dataset [20] are first processed by
detecting and aligning the faces, following the procedure explained
in Section 2.1. We then fine-tune the VGG-face model on FER 2013
dataset, using both the training and the public test set; during
training we use data augmentation by jittering the scale, flipping
and rotating the faces. The aim is to make the network more robust
to small misalignment of the faces. We also apply a strong dropout
on the last layer of the VGG (keeping only 5% of the nodes) to
prevent over-fitting. We achieve a performance of 71.2% on the
FER private test set, which is slightly higher than the previously
published results [18, 20].
To assess the quality of the description given by this fine-tuned
VGG model to emotion recognition, we first benchmarked it on the
validation set of SFEW [13], which is a dataset containing frames
of the videos of AFEW (those used by the EmotiW challenge). We
achieved a score of 45.2% without retraining the model on SFEW,
while the state-of-the-art result [25] is 52.5%, using a committee of
deep models, and the challenge baseline is of 39.7% [13].
The face sequences detected in AFEW videos are resized to
224x224 and fed into the fine-tuned VGG-face model. We then
extract the 4096 length fc6 layer, following the same pipeline as Fan
et al. [17, 18]. We also compute the softmax score for each frame.
2.4
Overlap of
8 frames
Window of 16 frames
Window of 16 frames
Figure 4: In the case of 8 frames overlap, we can see that the
16-frame windows are sharing half of their faces.
During inference, the face sequences detected in AFEW videos
are resized to 112x112. We then split a video into several windows
of 16 consecutive frames2 , with or without overlapping between
the windows (as shown in Figure 4), and fed it into the weighted
C3D. We then extract the second 4096-d fully connected layer and
the last softmax layer of the C3D for each window of the videos.
The method we propose bears similarities with Multiple Instance
Learning (MIL) [1]. Within the framework of MIL, each video can be
considered as a bag of windows, with one single label. A straightforward way to apply MIL would be to train the C3D on each video as
a bag of windows, and add a final layer for choosing the prediction
as the one with the maximum score among all the scores of the
batch. The loss would be then computed from this prediction. The
weights defined in Eq.(1) play this role by selecting iteratively the
best scoring windows.
Representing Faces with 3D Convolutional
Neural Network
3D convolutional neural networks have been shown to give good
performance in the context of facial expressions recognition in
video [2, 33]. Fan et al. [18] fine-tuned a pre-trained ’sport1m’
model on randomly chosen windows of 16 consecutive faces. During
inference, the model is applied to the central frame of the video.
One limitation of this approach is that, at test time, there is no
guaranty that the best window for capturing the emotion is in the
middle of the video. The same problem occurs during training: a
large part of the windows (randomly) selected for training does
not contain any emotion or does not contain the correct emotion.
Indeed, videos are annotated as a whole but some frames can have
different labels or, sometimes, no expression at all.
We address this limitation in the following way: we first finetune a C3D-sport1m model using all the windows of each video,
optimizing the classification performance until the beginning of
convergence. Then, to be able to learn more from the most meaningful windows than from the others, we weigh each window based
on its scores. More precisely, for the i t h video and the j th window
of this video, at epoch t, the weight w i, j is computed as:
w i, j = e
−s i, j
T (t )
2.5
Per Modality Temporal Fusion
Both VGG and weighted C3D representations are applied to each
frame of the videos, turning the videos into temporal sequences of
visual descriptors. We will name the elements of those sequences
as "descriptors", whether it is the description of a frame or of a
window.
To classify these sequences of descriptors, we investigated several methods. The most straightforward one is to score each descriptor and to take the maximum of the softmax scores as the final
prediction. Similarly, the maximum of the means of the softmax
across time can also be considered.
To better take into account temporal dependencies between
consecutive descriptors, another option is to use Long Short-Term
Memory recurrent neural networks (LSTM) [2, 19]. Unlike [17, 18],
we chose to use a variable length LSTM allowing us to take all the
descriptors as inputs.
To prevent over-fitting, we also applied dropout to the LSTM
cells and weight decay to the final fully connected layer. A random
grid search on hyper-parameters is then applied for each one of
these models.
(1)
where T (t) is a temperature parameter decreasing with epoch t and
si, j , the score of the window j of the video i. We then normalize
Í
the weight to ensure that for each video i, j w i, j = 1. A random
grid search on the hyper-parameters, including the temperature
descent, is made on the validation set.
2 The
sequences with fewer than 16 frames were padded by themselves to reach a
sufficient length.
4
Both VGG-LSTM and C3D-LSTM are used to give one description
vector (output of the final fully connected layer of the LSTM) and
one softmax score for each video.
The final VGG-LSTM architecture has 2230 hidden units for each
LSTM-cell and a final fully connected layer with 297 hidden units.
The maximal length of the input sequence is of 272 frames.
The final C3D-LSTM architecture has 1324 hidden units for each
LSTM-cell and a final fully connected layer of 304 hidden units. The
maximal length of the input sequence is of 34 windows (overlap of
8 frames) or 17 windows (no overlap).
M1
h1
M2
Fully
connected
h2
h3
M3
Hidden
representation
Modalities features
2.6
Fully
connected
Multimodal fusion
Score
Figure 5: Fully connected model with ModDrop. The three
modalities are concatenated and a fully connected with modDrop is applied. The obtained hidden representation is then
fed to a second regular fully connected layer, which will output the scores.
Last but not the least, the different modalities have to be efficiently
combined to maximize the overall performance.
As explained in Section 1, two main fusion strategies can be used
i.e. score fusion (late fusion), which consists in predicting the labels
based on the predictions given by each modality, or features fusion
(early fusion), which consists in taking as input the latent features
vectors given by each modality and learning a classifier on the top
of them.
During the last editions of the challenge, most of the papers
focused on score fusion, using SVM [4], Multiple Kernel Fusion [9]
or weighted means [18]. Differently, several authors tried to train
audio and image modalities together (see e.g., [8]), combining early
features and using soft attention mechanism, but they didn’t achieve
state-of-the-art performance. We propose an approach combining
both.
The rest of the section describes the four different fusion methods
we experimented with, including a novel method (denominated as
score trees).
To be more formal, let n be the number of modalities, W 1 be the
weights matrix of the first layer. It can be divided into n2 weights
block matrices W 1k,l , modeling unimodal and intermodal contributions, with k and l, ranging over the number of modalities n. The
first fully connected equation can be written as :
h1
W 11,1 W 11,2 W 11,3 M 1
© ª ©
ª© ª
h 2 ® = W 12,1 W 12,2 W 12,3 ® M 2 ®
«h 3 ¬ «W 13,1 W 13,2 W 13,3 ¬ «M 3 ¬
Then, the term to add to the loss is simply (with γmd decreasing
through the time):
Õ
γmd
||W 1k,l ||2
k ,l
Setting γ S D to very high values during the first iterations leads
to zeroing non-diagonal block matrices. Lowering it later reintroduces progressively these coefficients. From our observation, this
approach provided better convergence on the considered problem.
Baseline score fusion. We experimented with several standard
score fusion, like majority voting, means of the scores, maximum
of the scores and linear SVM.
Fully connected neural network with modality drop. We also experimented with the ModDrop method of Neverova et al. [29].
It consists in applying dropout by modality, learning the crossmodality correlations while keeping unimodal information. [29]
reports state-of-the-art results on gesture recognition. We apply
this method to our audio, C3D-LSTM and VGG-LSTM features,
as shown in Figure 5. According to [29], this is much better than
simply feeding the concatenation of the modalities features into
a fully connected neural network and letting the network learn a
joint representation. Indeed, the fully connected model would be
unable to preserve unimodal information during cross-modality
learning.
An important step to make convergence possible with ModDrop
is to first learn the fusion without cross-modality. For this reason,
[29] conditioned the weight matrix of the first layer so that the
diagonal blocks are equal to zeros and released this constraint after
a well-chosen number of iterations.
To warranty the preservation of the unimodal information, we
explore an alternative method which turned out to be better: we
apply an adapted weight decay, only on the non-diagonal blocks,
and decreased its contribution to the loss through time.
Score trees. Our motivation is to combine the high-level information coming from the scores with the lower-level information
coming from the features, all together. We did it by building what
we call Score Trees (see Figure 6 for an illustration).
A fully connected classification neural network is applied separately to the features of the different modalities, outputting a vector
of size 7. This vector is then concatenated with the scores of the
two other modalities, to create a vector of size 21. A fully connected
classification neural network is then fed with it and outputs a prediction vector of size 7. The aim is to make predictions with respect
to the predictions coming from other modalities. Finally, these three
new prediction vectors are concatenated and fed into a last fully
connected classifier, which gives the overall scores. This method
can be generalized to any number of modalities.
Weighted Mean. The weighted mean is the approach of the winners of the 2016 edition [18]. It consists in weighting the score of
each modality and sum them up.
The weights are chosen by cross validation on the validation set,
selecting the ones giving the best performance.
We applied it on the VGG-LSTM, C3D-LSTM and audio models.
5
Training
Validation
Test
Angry
133 (17.2 %)
64 (16.7 %)
99 (15.2 %)
Disgust
74 (9.6 %)
40 (10.4 %)
40 (6.1 %)
Fear
81 (10.4 %)
46 (12 %)
70 (10.7 %)
Happy
150 (19.4 %)
63 (16.4 %)
144 (22 %)
Sad
117 (15.1 %)
61 (15.9 %)
80 (12.3 %)
Neutral
144 (18.6 %)
63 (16.4 %)
191 (29.2 %)
Surprise
74 (9.6 %)
46 (12 %)
29 (4.4 %)
Total
773
383
653
Table 1: AFEW 7.0: number of video sequences per class.
Fully connected
Fully
connected
1
2
3
Fully
connected
Fully
connected
1
2
Fully
connected
1
3
Fully
connected
2
the bounding boxes and timestamps are kept. We then extract candidates temporal windows of ten seconds around all of these time
stamps and asked human annotators to select and annotate the most
relevant ones. To ensure the quality of our annotations, we evaluated ourselves (as human beings) on the validation set. We reached
a performance from 60 % to 80 % depending on the annotator, which
is compatible with the figure of 60% observed by [22].
3
Fully
connected
3.2
1
2
3
Figure 6: The Score Tree architecture
3
EXPERIMENTAL VALIDATION AND
RESULTS
After introducing the AFEW dataset (the dataset of the challenge),
this section presents the experimental validation of our method.
We first present experiments done on the modalities taken separately, and then present experiments on their fusion. We finally
introduce the experiments done for the EmotiW’17 challenge and
the performance we obtained.
3.1
Experiments on Single Modalities
Each modality has been evaluated separately on the validation set
of AFEW. The VGG-LSTM and C3D-LSTM modalities performs
better than [11].
Regarding the VGG-LSTM and the C3D-LSTM, both unidirectional and bidirectional LSTM architectures [21] were evaluated,
with one or two layers.
In several recent papers [3, 14], bidirectional LSTM is claimed
to be more efficient than the unidirectional one and could be seen
as a way to augment the data. Nevertheless, in our case, we have
observed that the bidirectional LSTM was prone to over-fitting on
the training set and therefore does not perform well on the validation set. The same observation has been made when increasing the
number of layers. The best performing architecture is in both cases
a one-layer unidirectional LSTM.
3.2.1 VGG-LSTM. The VGG model (without LSTM) has first
been evaluated by taking the maximum of the scores over the
sequences, giving the accuracy of 41.4% on the validation set.
Then the different LSTM architectures were tested and the best
performance for each one is given in Table 2. We note a 3% improvement compared to Fan et al. [18]. It can be explained by the
fact that our model uses the whole sequences, feeding the model
with more information. Data augmentation also helps.
The Acted Facial Emotion in the Wild
dataset
Acted Facial Emotion in the Wild (AFEW) is the dataset used by
the EmotiW challenge. The 2017 version of AFEW is composed
of 773 training videos, 383 validation videos and 653 test videos.
Each video is labeled with one emotion among: ’angry’, ’disgust’,
’fear’, ’happy’, ’sad’, ’neutral’ and ’surprise’. In addition to the video,
cropped and aligned faces extracted by [36, 39] are also provided.
Another important specificity of this dataset is the class distribution of the Train/Val/Test subsets (as shown in Table 1). This
difference can make the performance on the Test set different from
the one of the Validation set, as some classes are more challenging
than others.
3.2.2 C3D-LSTM. We then experimented our method with the
C3D modality alone. The performance is given in Table 3.
We observe that our implementation of C3D trained on random
windows and evaluated on central windows is not as good as Fan
et al. [18]. The C3D trained on central windows performed better
but is not state-of-the-art either.
Our proposed C3D (with LSTM) has been tested with and without
overlapping between windows. To evaluate the weighted C3D, the
prediction of the window with the maximal softmax score among
the video is first taken. It performs better without overlapping,
and we observe a lower difference between training and validation
accuracy. It could be explained by the fact that the number of
3.1.1 External training data. To enlarge the training set, we collected external data by selecting 380 video clips from our personal
DVDs movie collection, after checking that there is no overlap
between the selected movies and the ones of AFEW [12]. These
movies were processed by the following pipeline: faces are first
detected using our face detector (see Section 2.1 for details), and
6
Method
Validation
accuracy
Maximum of the scores
41.4 %
Unidirectional LSTM one layer
48.6 %
Bidirectional LSTM one layer
46.7 %
Unidirectional LSTM two layers
46.2 %
Bidirectional LSTM two layers
45.2 %
Fan et al. [18]
45.42 %
Table 2: VGG-LSTM performance on the validation set of
AFEW.
Fusion method
Validation
Test
accuracy
accuracy
Majority vote
49.3 %
_
Mean
47.8 %
_
ModDrop (sub.3)
52.2 %
56.66 %
Score Tree (sub.4)
50.8 %
54.36 %
Weighted mean (sub.6)
50.6 %
57.58 %
Table 4: Performance of the different fusion methods on the
validation and test sets.
Method
3.4
Validation
accuracy
C3D on central window
38.7 %
C3D on random window
34 %
Weighted C3D (no overlap)
42.1 %
Weighted C3D (8 frames overlap)
40.5 %
Weighted C3D (15 frames overlap)
40.1 %
LSTM C3D (no overlap)
43.2 %
LSTM C3D (8 frames overlap)
41.7 %
Fan et al. [18]
39.7 %
Table 3: C3D and C3D-LSTM performance on the validation
set of AFEW.
We submitted the method presented in this paper to the EmotiW’17
challenge. We submitted 7 runs which performance is given in Table
5).
The difference between the runs are as follows:
• Submission 2: ModDrop fusion of audio, VGG-LSTM and a
LSTM-C3D with an 8-frame overlap.
• Submission 3: addition of another LSTM-C3D, with no overlap, improving the performance on the test set as well as on
the validation set.
• Submission 4: fusion based on Score Trees, did not achieve
a better accuracy on the test set, while observing a slight
improvement on the validation set.
• Submission 5: addition of one VGG-LSTM and two other
LSTM-C3D, one with 8 frames overlap and one without.
These new models were selected among the best results in
our random grid search on hyper-parameters according to
their potential complementarity degree, evaluated by measuring the dissimilarity between their confusion matrices.
The fusion method is ModDrop.
• Submission 6: weighted mean fusion of all the preceding
modalities, giving a gain of 1 % on the test set, while losing
one percent on the validation set, highlighting generalization
issues.
• Submission 7: our best submission, which is the same as
the sixth for the method but with models trained on both
training and validation sets. This improves the accuracy by
1.2 %. This improvement was also observed in the former
editions of the challenge. Surprisingly, adding our own data
didn’t bring significant improvement (gain of less than one
percent on the validation set). This could be explained by
the fact that our annotations are not correlated enough with
the AFEW annotation.
windows is lower if there is no overlap, the choice between the
windows is therefore easier. As a second observation, the use of
LSTM with the weighted C3D leads to the highest scores.
At the end, it can be observed that our C3D descriptor performs
significantly better than the one of [18].
3.2.3 Audio. The audio modality gave a performance of 36.5%,
lower than the state-of-the-art method (39.8%) [17]. The use of a
perceptron classifier (worse than the SVM) nevertheless allowed us
to use high-level audio features during fusion.
3.3
Our participation to the EmotiW’17
challenge
Experiments on Fusion
Table 4 summarizes the different experiments we made on fusion.
The simple baseline fusion strategy (majority vote or means of
the scores) does not perform as well as the VGG-LSTM modality
alone.
The proposed methods (ModDrop and Score Tree) achieved
promising results on the validation set, but are not as good as
the simple weighted mean on the test set. This can be explained by
the largest number of parameters used for the modDrop and the
score tree, and by the fact that some parameters cross validated on
the validation set.
The best performance obtained on the validation set has the
accuracy of 52.2 % , which is significantly higher than the performance of the baseline algorithm provided by the organizers – based
on computing LBPTOP descriptor and using a SVR – giving the
accuracy of 38.81 % on the validation set [10].
The proposed method has been ranked 4th in the competition.
We observed that, this year, the improvement of the top accuracy
compared to the previous editions is small (+1.1%), while from 2015
to 2016 the improvement was of +5.2 %. This might be explained by
the fact that the methods are saturating, converging towards human
performance (which is assumed to be around 60 %). However, the
performance of top human annotators (whose accuracy is higher
than 70 %) means there is still some room for improvement.
7
Submission
Test Accuracy
2
55.28 %
3
56.66 %
4
54.36 %
5
56.51 %
6
57.58 %
7
58.81 %
Table 5: Performance of our submissions on the test set.
Angry
Disgust
Fear
Happy
Neutral
Sad
Surprise
Angry
74,5
0
3,1
3,1
18,4
1
0
Disgust
20
0
5
15
40
20
0
Fear
27,1
0
34,3
1,4
21,4
15,7
0
Happy
6,9
0
0,7
82,6
6,3
3,5
0
Neutral
8,3
0
2,1
5,7
69,9
14
0
Sad
12,5
0
7,5
13,8
25
41,3
0
Surprise
21,4
0
21,4
7,1
35,7
14,3
0
REFERENCES
[1] Saad Ali and Mubarak Shah. 2010. Human action recognition in videos using
kinematic features and multiple instance learning. IEEE transactions on pattern
analysis and machine intelligence 32, 2 (2010), 288–303.
[2] Moez Baccouche, Franck Mamalet, Christian Wolf, Christophe Garcia, and Atilla
Baskurt. 2011. Sequential deep learning for human action recognition. In International Workshop on Human Behavior Understanding. Springer, 29–39.
[3] Moez Baccouche, Franck Mamalet, Christian Wolf, Christophe Garcia, and Atilla
Baskurt. 2012. Spatio-Temporal Convolutional Sparse Auto-Encoder for Sequence
Classification. In BMVC.
[4] Sarah Adel Bargal, Emad Barsoum, Cristian Canton Ferrer, and Cha Zhang. 2016.
Emotion recognition in the wild from videos using images. In Proceedings of the
18th ACM International Conference on Multimodal Interaction. ACM, 433–436.
[5] Lisa Feldman Barrett, Batja Mesquita, and Maria Gendron. 2011. Context in
emotion perception. Current Directions in Psychological Science 20, 5 (2011),
286–290.
[6] Lisa Feldman Barrett and James A Russell. 1999. The structure of current affect:
Controversies and emerging consensus. Current directions in psychological science
8, 1 (1999), 10–14.
[7] C Fabian Benitez-Quiroz, Ramprakash Srinivasan, Qianli Feng, Yan Wang, and
Aleix M Martinez. 2017. EmotioNet Challenge: Recognition of facial expressions
of emotion in the wild. arXiv preprint arXiv:1703.01210 (2017).
[8] Linlin Chao, Jianhua Tao, Minghao Yang, Ya Li, and Zhengqi Wen. 2016. Audio
visual emotion recognition with temporal alignment and perception attention.
arXiv preprint arXiv:1603.08321 (2016).
[9] JunKai Chen, Zenghai Chen, Zheru Chi, and Hong Fu. 2014. Emotion recognition
in the wild with feature fusion and multiple kernel learning. In Proceedings of the
16th International Conference on Multimodal Interaction. ACM, 508–513.
[10] Abhinav Dhall, Roland Goecke, Shreya Ghosh, Jyoti Joshi, Jesse Hoey, and Tom
Gedeon. 2017. From Individual to Group-level Emotion Recognition: EmotiW 5.0.
In Proceedings of the 19th ACM International Conference on Multimodal Interaction.
[11] Abhinav Dhall, Roland Goecke, Jyoti Joshi, Jesse Hoey, and Tom Gedeon. 2016.
Emotiw 2016: Video and group-level emotion recognition challenges. In Proceedings of the 18th ACM International Conference on Multimodal Interaction. ACM,
427–432.
[12] Abhinav Dhall, Roland Goecke, Simon Lucey, and Tom Gedeon. 2012. Collecting Large, Richly Annotated Facial-Expression Databases from Movies. IEEE
MultiMedia (2012).
[13] Abhinav Dhall, OV Ramana Murthy, Roland Goecke, Jyoti Joshi, and Tom Gedeon.
2015. Video and image based emotion recognition challenges in the wild: Emotiw
2015. In Proceedings of the 2015 ACM on International Conference on Multimodal
Interaction. ACM, 423–426.
[14] Yong Du, Wei Wang, and Liang Wang. 2015. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference
on computer vision and pattern recognition. 1110–1118.
[15] Paul Ekman and Wallace V Friesen. 1977. Facial action coding system. (1977).
[16] Florian Eyben, Martin Wöllmer, and Björn Schuller. 2010. Opensmile: the munich
versatile and fast open-source audio feature extractor. In Proceedings of the 18th
ACM international conference on Multimedia. ACM, 1459–1462.
[17] Lijie Fan and Yunjie Ke. 2017. Spatiotemporal Networks for Video Emotion
Recognition. arXiv preprint arXiv:1704.00570 (2017).
[18] Yin Fan, Xiangju Lu, Dian Li, and Yuanliu Liu. 2016. Video-based emotion
recognition using cnn-rnn and c3d hybrid networks. In Proceedings of the 18th
ACM International Conference on Multimodal Interaction. ACM, 445–450.
[19] Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. [n. d.]. Learning to forget:
Continual prediction with LSTM. Technical Report.
[20] Ian J Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi
Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun
Lee, et al. 2013. Challenges in representation learning: A report on three machine
learning contests. In International Conference on Neural Information Processing.
Springer, 117–124.
[21] Alex Graves, Santiago Fernández, and Jürgen Schmidhuber. 2005. Bidirectional
LSTM networks for improved phoneme classification and recognition. Artificial
Neural Networks: Formal Models and Their Applications–ICANN 2005 (2005), 753–
753.
[22] Markus Kächele, Martin Schels, Sascha Meudt, Günther Palm, and Friedhelm
Schwenker. 2016. Revisiting the EmotiW challenge: how wild is it really? Journal
on Multimodal User Interfaces 10, 2 (2016), 151–162.
[23] Heysem Kaya, Furkan Gürpınar, and Albert Ali Salah. 2017. Video-based emotion
recognition in the wild using deep transfer learning and score fusion. Image and
Vision Computing (2017).
[24] Pooya Khorrami, Thomas Paine, and Thomas Huang. 2015. Do deep neural networks learn facial action units when doing expression recognition?. In Proceedings
of the IEEE International Conference on Computer Vision Workshops. 19–27.
[25] Bo-Kyeong Kim, Jihyeon Roh, Suh-Yeon Dong, and Soo-Young Lee. 2016. Hierarchical committee of deep convolutional neural networks for robust facial
Figure 7: Confusion matrix obtained with the seventh submission. We can see that ’disgust’ and ’surprise’ classes are
never predicted by our model, while the three dominant
classes (’happy’, ’neutral’ and ’angry’) are well recognized.
The ’neutral’ class has the largest number of false positives.
It underlines the difficulty, even for humans, to draw the
margin between presence and absence of emotion. Rows denote true classes, columns predicted classes.
4
CONCLUSIONS
This paper proposes a multimodal approach for video emotion
classification, combining VGG and C3D models as image descriptors and explores different temporal fusion architectures. Different
multimodal fusion strategies have also been proposed and experimentally compared, both on the validation and on the test set of
AFEW. At the EmotiW’17 challenge, the proposed method ranked
4th with the accuracy of 58.81 %, 1.5 % under the competition
winners. One important observation from this competition is the
discrepancy between the performance obtained on the test set and
the one on the validation set: good performance on the validation
set is not a warranty to good performance on the test set. Reducing
the number of parameters in our models could help to limit overfitting. Using pre-trained fusion models and, moreover, gathering a
larger set of data would also be a good way to face this problem.
Finally, another interesting path for future work would be to add
contextual information such as scene description, voice recognition
or even movie type as an extra modality.
8
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
expression recognition. Journal on Multimodal User Interfaces 10, 2 (2016), 173–
189.
Davis E King. 2009. Dlib-ml: A machine learning toolkit. Journal of Machine
Learning Research 10, Jul (2009), 1755–1758.
Agata Kołakowska, Agnieszka Landowska, Mariusz Szwoch, Wioleta Szwoch,
and Michał R Wróbel. 2013. Emotion recognition and its application in software
engineering. In 2013 The 6th International Conference on Human System Interaction
(HSI). IEEE, 532–539.
Jean Kossaifi, Georgios Tzimiropoulos, Sinisa Todorovic, and Maja Pantic. 2017.
AFEW-VA database for valence and arousal estimation in-the-wild. Image and
Vision Computing (2017).
Natalia Neverova, Christian Wolf, Graham Taylor, and Florian Nebout. 2016.
Moddrop: adaptive multi-modal gesture recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence 38, 8 (2016), 1692–1706.
Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, et al. 2015. Deep Face
Recognition.. In BMVC, Vol. 1. 6.
Robert Plutchik and Henry Kellerman. 2013. Theories of emotion. Vol. 1. Academic
Press.
Karen Simonyan and Andrew Zisserman. 2014. Two-Stream Convolutional
Networks for Action Recognition in Videos.. In NIPS. 568–576.
Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri.
2015. Learning spatiotemporal features with 3d convolutional networks. In
Proceedings of the IEEE international conference on computer vision. 4489–4497.
Michel Valstar, Jonathan Gratch, Björn Schuller, Fabien Ringeval, Dennis Lalanne,
Mercedes Torres Torres, Stefan Scherer, Giota Stratou, Roddy Cowie, and Maja
Pantic. 2016. Avec 2016: Depression, mood, and emotion recognition workshop
and challenge. In Proceedings of the 6th International Workshop on Audio/Visual
Emotion Challenge. ACM, 3–10.
Peter Washington, Catalin Voss, Nick Haber, Serena Tanaka, Jena Daniels, Carl
Feinstein, Terry Winograd, and Dennis Wall. 2016. A wearable social interaction
aid for children with autism. In Proceedings of the 2016 CHI Conference Extended
Abstracts on Human Factors in Computing Systems. ACM, 2348–2354.
Xuehan Xiong and Fernando De la Torre. 2013. Supervised descent method
and its applications to face alignment. In Proceedings of the IEEE conference on
computer vision and pattern recognition. 532–539.
Baohan Xu, Yanwei Fu, Yu-Gang Jiang, Boyang Li, and Leonid Sigal. 2016. Heterogeneous knowledge transfer in video emotion recognition, attribution and
summarization. IEEE Transactions on Affective Computing (2016).
Anbang Yao, Junchao Shao, Ningning Ma, and Yurong Chen. 2015. Capturing
au-aware facial features and their latent relations for emotion recognition in the
wild. In Proceedings of the 2015 ACM on International Conference on Multimodal
Interaction. ACM, 451–458.
Xiangxin Zhu and Deva Ramanan. 2012. Face detection, pose estimation, and
landmark localization in the wild. In Computer Vision and Pattern Recognition
(CVPR), 2012 IEEE Conference on. IEEE, 2879–2886.
9
| 1cs.CV
|
1
Identifiability of generalised Randles circuit models
arXiv:1505.00153v2 [math.OC] 26 May 2016
S.M.Mahdi Alavi, Adam Mahdi, Stephen J. Payne and David A. Howey
Abstract—The Randles circuit (including a parallel
resistor and capacitor in series with another resistor)
and its generalised topology have widely been employed
in electrochemical energy storage systems such as batteries, fuel cells and supercapacitors, also in biomedical
engineering, for example, to model the electrode-tissue
interface in electroencephalography and baroreceptor
dynamics. This paper studies identifiability of generalised Randles circuit models, that is, whether the
model parameters can be estimated uniquely from the
input-output data. It is shown that generalised Randles circuit models are structurally locally identifiable.
The condition that makes the model structure globally
identifiable is then discussed. Finally, the estimation
accuracy is evaluated through extensive simulations.
Index Terms—Randles circuit, Identifiability, System
identification, Parameter estimation.
I. Introduction
Randles proposed an equivalent circuit for the kinetics of
rapid electrode reactions in [1]. Since then, the model has
been developed and has become the basis for the study
of many electrochemical energy storage systems such as
batteries, fuel cells and supercapacitors, [2], [3]. Figure 1
shows a generalised Randles model consisting of an ohmic
resistor, R∞ , in series with a number of parallel resistors
and capacitors, and a capacitor Cw . In electrochemical
applications, the ohmic resistor R∞ represents usually
the conduction of charge carriers through electrolyte and
metallic conductors. The resistors and capacitors in the
parallel pairs represent the charge transfer resistance and
the double layer capacitance, respectively, or are an approximation of a diffusion process, [4], [5]. The number
of parallel R’s and C’s depends on how many of these
pairs are required such that the frequency response of the
generalised Randles model fits with the device impedance
spectra within the frequency range of interests [6]–[8].
For instance, in [8], the number of the parallel pairs was
determined by minimising the error between the model
and measured voltages. The capacitor Cw , also known as
the Warburg term, accounts for a diffusion process, [4], [5];
or in a battery or supercapacitor, it may represent state of
charge [9]. It should be noted that the open-circuit voltage
S.M.M. Alavi was with the Energy and Power Group, Department of Engineering Science, University of Oxford. He is now with
the Brain Stimulation Engineering Laboratory, Duke University,
Durham, NC 27710, USA. Email: [email protected]
A. Mahdi and S.J. Payne are with the Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford,
Old Road Campus Research Building, Oxford, OX3 7DQ, United
Kingdom. Emails: {adam.mahdi, stephen.payne}@eng.ox.ac.uk
D.A. Howey is with the Energy and Power Group, Department of
Engineering Science, University of Oxford, Parks Road, Oxford, OX1
3PJ, United Kingdom. Email: [email protected]
source is not considered in the generalised Randles model
Figure 1 and it is only focused on the impedance model
in this paper.
The generalised Randles circuit models have also been
employed in biomedical engineering. An electric circuit model of the electrode-tissue interface in electroencephalography includes two parallel R-C pairs in series
with two resistors and two voltage sources [10], which is a
special case of the circuit given in Figure 1. It is noted that
Cw can be considered as a parallel R and C with R = ∞.
The viscoelastic analog of the generalised Randles model is
employed in cardiovascular and cerebral haemodynamics
modeling, in describing the viscoelastic properties of the
aortic wall and the coupling of the nerve endings of the
baroreceptor neurons in the carotid sinus and aortic arch
[11], [12]; and in relating the fluctuation of the arterial
blood pressure with the cerebral blood flow velocity [13].
The identification or parameter estimation of the generalised Randles model (with different numbers of parallel
R’s and C’s) is important in condition monitoring, fault
diagnosis and control [3], [9], [14]–[20]. In [14], the authors
showed that the Randles circuit can be used for monitoring
the battery charge transfer overvoltage. In [15], identification tests for parameter estimation of lead-acid batteries
are suggested. In [20], identification through fitting the
impedance spectra in the frequency domain is presented.
The objective of this paper is to study the identifiability
of the generalised Randles model shown in Figure 1,
that is, whether the model parameters can be estimated
uniquely from input-output data. Typically the identifiability problem is divided into two broad areas: parameter
estimation accuracy and structural identifiability. Parameter estimation accuracy considers the practical aspects
of the problem that come with real data such as noise
and bias [21]. In studying structural identifiability, on the
other hand, one assumes that noise-free informative data
is available and therefore it is, in fact, a data-independent
concept. Unidentifiable parameters can be assigned an
infinite number of values yet still lead to identical inputoutput data. Thus, structural identifiability is a necessary
condition for identifiability and parameter estimation. A
number of analytical approaches to structural identifiability have been proposed, including Laplace transform
(transfer function) [22], [23], Taylor series expansion [24],
[25], similarity transformations [26]–[33], and differential
algebra [34], [35]. In linear systems, it has been shown
that controllability and observability properties are closely
related to the concept of structural identifiability, [30]–
[33]. For example, it was shown that single-input singleoutput linear time-invariant systems are structurally identifiable if and only if their observer canonical form is
2
R1
Rn
R2
R∞
+
Cw i(t)
...
C1
C2
Cn
+v1 −
+v2 −
+vn −
+vw −
v(t)
Fig. 1. The generalised Randles equivalent circuit model.
controllable (see Chapter 4 in [33]). However, controllable
and observable systems can still be unidentifiable in the
general case [31].
Recently, there has been a significant interest in the
identifiability analysis of battery models [36]–[43]. In [36],
the structural identifiability of a five-element equivalent
circuit model (ECM) including two capacitors was discussed by comparing the number of unknown parameters
of the Transfer Function (TF) and the circuit. In [37], the
structural identifiability of a more general nonlinear ECM
is analysed based on the observability conditions. It is
shown that cells with serial connections are not observable,
demonstrating that complete estimation of the state-ofcharge and model parameters from lumped measurements
with series cells is not possible trough independent measurements. However, it is shown that lumped models with
parallel connectivities are observable provided that none
of the parallel cells are identical. The identifiability of
battery electrochemical models was discussed in [38]–[41].
In particular, it was shown that some of the electrochemical model parameters are not identifiable given typical
charge-discharge cycles [38]–[40]. In [41], it was shown that
the shape of the charge-discharge cycles plays a crucial
role in the identifiability of battery parameters. It was
also demonstrated that the system identification method
can be employed in monitoring the battery film growth.
The identifiability problem of Randles ECMs has been
studied in [42], [43], where the authors consider models
that include up to two capacitors, and the analysis is
based on the Fisher information matrix (FIM). The FIM
provides some information about the sensitivity of the
measurement to the model parameters by using likelihood
functions, [44]. In [42], a bound of estimation errors was
developed by using the Cramér-Rao theorem. In [43],
a method was proposed to optimally shape the battery
cycles and improve the identifiability.
This paper shows that the generalised Randles model
in Figure 1 is structurally globally identifiable for n =
1, and structurally locally identifiable for any finite n >
1, and becomes globally identifiable assuming an ordering
through the generalised circuit. Finally, the identifiability
of the model is assessed through extensive simulations.
II. The model parameterisation and problem
statement
The state-space and TF parameterised models of the
generalised Randles circuit given in Figure 1 are derived.
Define the electric current as the system input u(t) =
i(t) ∈ R, the terminal voltage as the system output y(t) =
v(t) ∈ R, the voltages across the internal R-C pairs as the
⊤
states x(t) = v1 (t) · · · vn (t) vw (t)
∈ Rn+1 , and
the model parameters as
θ = R∞ R1 · · · Rn C1 · · · Cn Cw , (1)
where θ belongs to some open subset D ⊂ R2n+2 .
A. State-space parametrisation
By using Kirchhoff’s laws, a state-space model structure
of the Randles circuit, parameterised by θ , can be written
as:
d x(t) = A(θθ ) x(t) + B(θθ ) u(t)
(2)
dt
y(t) = C(θθ ) x(t) + D(θθ )u(t)
where A(θθ ) ∈ R(n+1)×(n+1) , B(θθ ) ∈ Rn+1 , C(θθ )⊤ ∈ Rn+1
and D(θθ ) ∈ R are matrices that depend on the parameter
vector θ , and are given by
A(θθ ) =
B(θθ ) =
−a1 (θθ )
0
0
−a2 (θθ )
..
..
.
.
0
0
0
0
b1 (θθ )
b2 (θθ )
..
, C(θθ )⊤
.
bn (θθ )
bw (θθ )
···
···
..
.
···
···
=
0
0
..
.
0
0
..
.
,
−an (θθ ) 0
0
0
1
1
.. , D(θθ ) = d,
.
1
1
with time constants τi = Ri Ci for i = 1, · · · , n and
1
1
1
ai = , b i =
, bw =
, d = R∞ .
τi
Ci
Cw
(3)
(4)
B. Transfer function parametrisation
Denote the model’s TF by T (s, θ ), where s represents
the Laplace operator. Using the formula T (s, θ ) = C(sI −
A)−1 B+D, a parameterised TF of the generalised Randles
circuit can be written as:
n
X
bw (θθ )
bi (θθ )
+
+ D(θθ ).
(5)
T (s, θ ) =
s + ai (θθ )
s
i=1
3
C. Problem statement
Determine the conditions for the model structure (5)
(equivalently in the state-space form (2) and (3)), where θ
is an unknown parameter vector (1), to be locally and/or
globally structurally identifiable.
III. Main results
Following [33] a definition of structural identifiability is
given as follows.
Definition 1: Let M be a model structure with the TF
T (s, θ ), parametrized by θ , where θ belongs to an open
subset DT ⊂ Rm , and consider the equation
∗
T (s, θ ) = T (s, θ )
for almost all s,
(6)
where θ , θ ∗ ∈ DT . Then, the model structure M is said to
be
- globally identifiable if (6) has a unique solution in DT ,
- locally identifiable if (6) has a finite number of solutions in DT ,
- unidentifiabile if (6) has a infinite number of solutions
in DT .
Remark 1: Instead of Definition 1, one can use the
so-called coefficient map defined as follows, [28], [29].
Consider the monic1 TF:
c0 (θθ ) + c1 (θθ )s + · · · + ck1 (θθ )sk1
T (s, θ ) =
,
d0 (θθ ) + d1 (θθ )s + · · · + dk2 −1 (θθ )sk2 −1 + sk2
(7)
and associate with it the following coefficient map CT :
Rm ⊃ DT → Rk1 +k2 +1 defined as
i
h
(8)
CT (θθ ) = c0 (θθ ), . . . , ck1 (θθ ), d0 (θθ ), . . . , dk2 −1 (θθ ) .
The model structure M is globally identifiable if the coefficient map CT is one-to-one (injective); locally identifiable
if CT is many-to-one; and unidentifiable if it is infinitely
many-to-one.
The following lemma will be used in the proof of the
identifiability of the generalised Randles circuit model.
Lemma 1: Let CT (m) be a coefficient map associated
with the TF
m
X
bi
+ d,
T (s, θ ) =
s + ai
i=1
where θ = (a1 , . . . , am , b1 , . . . , bm , d) and ai ∈ R for
i = 1, . . . , m are pairwise different. Then the following
statements hold:
(a) If m = 1, then CT (m) is one-to-one.
(b) If m > 1, then CT (m) is many-to-one.
Proof. Part(a). For m = 1, the identifiability equation
(6) is given by
b∗1
b1
+d=
+ d∗
s + a1
s + a∗1
and has a unique solution (a1 , b1 , d) = (a∗1 , b∗1 , d∗ ), which
proves part (a).
1 In a monic TF, all coefficients are normalised such that the
coefficient of the greatest order in the denominator is 1.
Part(b). For m > 1, the identifiability equation (6) is
given by
m
X
i=1
m
X b∗
bi
i
∗
+d=
∗ +d .
s + ai
s
+
a
i
i=1
(9)
We claim that equation (9) admits only finite (more
precisely, m! = 1 × 2 × · · · × m) number of solutions. To
prove the claim note that (9) is the equality of two rational
functions, which is satisfied provided that
(s + a1 ) × · · · × (s + am ) = (s + a∗1 ) × · · · × (s + a∗m ). (10)
Since m distinct roots uniquely characterise a monic
polynomial of degree m, and there are m! permutations
of n roots of (s + a1 ) × · · · × (s + am ), equation (10)
has m! solutions. Now, let us fix the permutation s :
{1, . . . , m} → {1, . . . , m} and consider an assignment
(a1 , . . . , am ) = (a∗s1 , . . . , a∗sm ). Since ai for i = 1, . . . , m are
assumed to be pairwise distinct, the expressions 1/(s+ai ),
thought of as functions of the variable s, are linearly
independent. Finally, since each side of equation (9) is a
linear combination of linearly independent functions, we
immediately obtain that bj = b∗sj for i = 1, . . . , m. This
concludes the proof of part (b).
Now, we introduce the concept of reparametrisation (see
e.g. [28]), which we will use the proof of our main theorem.
Definition 2: A reparametrisation of the model structure
M with the coefficient map CT is a map R : Rk ⊃ DT →
Rm such that
Im (CT ◦ R) = Im CT ,
(11)
where Im denotes the image of the map. Moreover, the
reparameterisation is identifiable if the map CT ◦ R : Rk ⊃
D → Rk1 +k2 +1 is identifiable.
The main result of this section is the following theorem which describes the identifiability of the generalised
Randles circuit.
Theorem 1: Let MRC (n) denotes the state-space model
structure (2) with the matrices (3) parametrised by (4),
where n is the number of parallel RC elements connected
in series (see Figure 1). Then the following conditions hold:
(a) If n = 1, then the model structure MRC (n) is globally
identifiable.
(b) If n > 1, then the model structure MRC (n) is locally
identifiable.
(c) If n > 1, and there is an ordering through the
generalised circuit as
an < an−1 < · · · < a1 ,
(12)
then the model structure MRC (n) is globally identifiable.
Proof. Consider the model structure MRC (n) and let
T (s, θ ) denote the corresponding TF given by (5), where
the parameter vector θ is given by (1). We write T (s, θ )
as a rational function (7) of degree k1 = k2 = n + 1 and
associate with it the coefficient map CT′ : R2n+2 ⊃ D →
R2n+3 .
4
Part (a). For n = 1, the coefficient map CT′ : R4 ⊃ D →
R can be written explicitly as
h
i
CT′ (θθ ) = c0 (θθ ), c1 (θθ ), c2 (θθ ), d0 (θθ ), d1 (θθ ) ,
A. Generation of informative data set
5
where
c0 (θθ ) =
1
,
R1 C1 Cw
c2 (θθ ) = R∞ ,
R∞ Cw + R1 Cw + R1 C1
,
R1 C1 Cw
1
d1 (θθ ) =
.
R1 C1
c1 (θθ ) =
d0 (θθ ) = 0,
By direct computation, we can check that equation
CT′ (θθ ) = CT′ (θθ ∗ ) admits a unique solution θ = θ ∗ . Thus
the coefficient map is one-to-one, and the model structure
is globally identifiable.
Part (b). For n > 1, the coefficient map CT′ can be
written as the following composition
CT′ = CT ◦ Rc ,
c
2n+2
(13)
2n+2
where the map R : R
⊃ D → R
is the
reparametrisation Rc (θθ ) = a1 , . . . , an , b1 , . . . , bn , bw , d
defined by (4); and CT is the coefficient map associated
with the TF
n
X
bi
bw
+
+ d,
(14)
T (s, θ a,b ) =
s + ai
s
i=1
where θ a,b = (a1 , . . . , an , b1 , . . . , bn , bw , d) ∈ R2n+2 . By
Lemma 1 (for m = n + 1 and an+1 = 0) the map CT is
many-to-one; and Rc is a one-to-one map with an inverse
R∞ = d, Cw =
bi
1
1
, Ri = , Ci = , i = 1, . . . , n.
bw
ai
bi
The map CT′ is many-to-one, since it is a composition of
a one-to-one with many-to-one map. Thus for n > 1, the
model structure MRC (n) is locally identifiable.
Part (c). Finally, the identifiability equation (9) under
the condition (12) admits a unique solution, which concludes part (c).
Remark 2: The same procedure is applicable to the
discrete time model. The Euler’s first order approximation
is the simplest approximation. Its identifiability analysis
is easier because the coefficients of the discrete time TF
equals the number of parameters. However, it might lead
to numerical instability. If higher order approximations are
applied, the challenge remains to prove whether there is
a finite-to-one maps between the discrete time model and
parameters.
IV. Simulations
Consider a 6-element Randles circuit with R∞ , Cw Ri ,
Ci , i = 1, 2. In this section, accuracy of the estimation
in the presence of noise-free and noisy data, subject to
random initial guess of estimations is studied. First the
informative data set used in the simulations and the way
it is generated are described. More details are provided in
references [33], [45]–[47].
A data set is informative if the input is persistently
exciting. A persistently exciting input adequately excites
all modes of the system. In linear systems, the order of
the system determines the order of persistent excitation.
The order of a persistent excitation equals the number of
coefficients of the monic TF that need to be identified, (see
Theorem 13.1. in [33]). The monic TF of the 6-element
Randles circuit is
f 3 s3 + f 2 s2 + f 1 s + f 0
(15)
H(s) = 3
s + g 2 s2 + g 1 s + g 0
with unknown coefficients to be identified. Therefore, the
necessary order of the persistently exciting input is 7. In
the frequency domain, this means that the spectrum of
the excitation input should have at least 7 nonzero points.
This simulation focuses on multi-sine excitation signals
which are widely employed in electrochemical impedance
spectroscopy (EIS) techniques, [3], [4], [48]. The same
procedure can be applied to other inputs. The multi-sine
signal is given by
u(t) =
l
X
mj cos(ωj t + φj ),
(16)
j=1
where l represents the number of sinusoids, and mj , ωj
and φj ∈ [−π, π) denote the magnitude, frequency in
radians per second and phase in radians, respectively. The
spectrum of the multi-sine signal is given by
Φu (ω) = 2π
l
X
m2j
j=1
4
[δ(ω − ωj ) + δ(ω + ωj )],
(17)
where δ(ω) is the delta function or impulse at frequency
ωj . The spectrum of each sinusoid signal contains two
nonzero points, therefore 4 sinusoid signals are enough
to generate the informative data set for the 6-element
Randles circuit model.
The magnitudes mj , frequencies ωj and phases φj are
arbitrary real values. Specific applications might impose
additional constraints. For instance, in EIS techniques, the
magnitude of the input signal may vary from milliampere
to ampere depending on the size of the energy storage
system. For the sake of simplicity, all magnitudes are
assumed equal to mj = 10−3 for j = 1, · · · , 4. The
frequencies could be equally or logarithmically spread over
the frequency band ωmin ≤ ω ≤ ωmax . The values of ωmin
and ωmax depend on the system dynamics. If mj ’s are
equal, the Schroeder phase choice is suggested to reduce
the Crest factor [33], [49].
Remark 3: A multi-sine signal may not exactly be zero
mean. Sometimes a uni-directional current flow is required.
In this case, the multi-sine excitation current needs to be
superimposed on a known constant DC offset current, (see
examples in [3] and [48]). In system identification, the data
are typically pre-filtered to remove these types of offsets.
5
8
×10-3
4
-3
u(t)
×10
0
3
-2
2.5
1.85
1.9
1.95
2
2.05
2
y(t) noisy with σ =1e-4
×10-4
2
0
-2
-4
-6
-8
1.8
2
2.1
Voltage (V)
Current (A)
4
1.8
y(t) noise-free
3.5
2
6
×10-3
1.5
1.85
1.9
1.95
2
2.05
2.1
1
0.5
0
0
-2
-0.5
-4
-1.5
-1
0
20
40
60
80
100
0
20
Time (s)
40
60
80
100
Time (s)
(a) Excitation input u(t).
(b) Associated output y(t).
Fig. 2. The multi-sine input excitation signal and its associated output for R∞ = 0.05 Ω, R1 = 0.2 Ω, C1 = 0.3 F, R2 = 0.4 Ω, C2 = 0.6 F,
and Cw = 300 F. The voltage response, Figure 2(b), is computed using the model (2) in the noise-free and noisy case with standard deviation
σ = 10−4 . The DC offset has been removed to have zero-mean signals.
B. The results discussion
The following operating point is arbitrarily selected [3],
[4]:
R∞ = 0.05 Ω, R1 = 0.2 Ω, C1 = 0.3 F
R2 = 0.4 Ω, C2 = 0.6 F, Cw = 300 F
The multi-sine excitation signal, Figure 2(a), is generated
by using model (16) with l = 4, mj = 10−3 , fmin = 0.2
Hz, fmax = 500 Hz and Schroeder phase with φ1 randomly
chosen at 1.9775. The Crest factor of the signal is 1.9999.
The voltage response, Figure 2(b), has been computed
using the model (2). The DC offset has been removed to
ensure zero-mean signals.
Given the true values, the smallest time constant is
τmin = R1 C1 = 0.06 s. The sampling time should be
several times larger than 1/τmin. In this simulation, the
sampling frequency is chosen at fs = 500 Hz, which is
16.6 times greater than the inverse of the minimum time
constant. The test duration should typically be several
times larger than the maximum time constant, which is
τmax = R∞ Cw = 15 s. In [47], 6 to 8 times τmax is
suggested, however, this might vary in different applications. In this simulation, a test duration of td = 100 s is
applied. The continuous-time TF (15) is identified using
Matlab’s system identification toolbox [50]. The circuit
parameters are calculated directly from the coefficients of
the TF using the formulas shown in Table II and discussed
later. In order to study the consistency of results, each
test is repeated 100 times, every run with a random initial
guess of parameters. The roots of the denominator must
be positive real numbers. Those estimations, which lead to
complex or negative poles for the TF, or to Cw > 1000 or
to Ci > 10, i = 1, 2, are considered outliers and discarded
from the analysis.
The relative mean error, er is defined as follows:
er = 100 ×
true value − mean of estimations
%. (18)
true value
The estimation accuracies from noise-free and noisy
data with a zero-mean noise and standard deviation σ =
10−4 are compared together. The signal to noise ratio is
10−3
10−4 = 10. Table I shows the mean, standard deviation
and relative errors of the estimations. Figure 3 shows the
histograms of the estimations from the noisy data.
Regardless of the random initial condition, in more than
80% of simulations, the noise-free estimations converge
with er less than 10% for Ri , Ci , i = 1, 2 and for Cw .
The largest er is 10.36% for R∞ (Table I).
Figure 3 shows that estimations from noisy data are
more distributed around the true values. However as it is
seen in Table I, the mean values of estimations in both
noise-free and noisy cases remain the same. The relative
mean errors of estimations are again less than 10% for all
parameters except of R∞ , which is again the largest and
10.38%. Also, er ’s increase with noisy data compared to
that of in the noise-free case.
The largest standard deviation is for Cw . This could be
because of the pure integrator associated with Cw , which
appears in the TF, which might require some modifications
of the data set, [51]. For instance, a methodology has
been proposed in [36] that removes the integral
term by
´
modifying the input signal as imodified(t) = i(t)dt.
C. Calculating the coefficients of R − R||C − R||C − C
circuit
Here we show how to compute the parameters of the
6-element Randles circuit. Using (5), the circuit’s TF is:
H(s) =
f 3 s3 + f 2 s2 + f 1 s + f 0
s3 + g 2 s2 + g 1 s + g 0
No. of estimations
6
100
100
80
80
60
60
the following set of equations for X:
1 1
1
f2 − (a1 + a2 )f3
a2 a1 a1 + a2 X =
f 1 − a1 a2 f 3
0 0
a1 a2
f0
40
40
where,
20
20
0
0.045
X=
0
0.05
0.055
0.06
0.065
0
0.1
No. of estimations
R∞
100
100
80
80
60
60
40
40
20
20
0
0
0.1
0.2
No. of estimations
0.3
0.3
0.4
0
0.3
0.4
0.5
100
80
80
60
60
40
40
20
20
b2
bw
⊤
Other parameters of the circuit are subsequently obtained as:
1
τi
1
, Ci = , Ri =
, i = 1, 2.
Cw =
bw
bi
Ci
0.6
V. Conclusions
R2
100
b1
The parameters of different topologies can simply be
calculated using the same approach. Table II provides
formulas for four widely used Randles models.
C1
0
0.2
0.2
R1
We showed that the generalised Randles circuit model
is locally identifiable and the model structure becomes
globally identifiable if an ordering through the circuit is
assumed. The results were confirmed through extensive
simulations. Finally, explicit formulas for the coefficients
of widely used Randles circuits were presented.
0
0.4
0.6
0.8
0
200
C2
400
600
Cw
Fig. 3. Histogram of accepted estimations in 100 runs from noisy
data.
The relationships between the coefficients and the circuit
parameters are given by:
f3 = d, f2 = b1 + b2 + bw + (a1 + a2 )d
f1 = a2 b1 + a1 b2 + (a1 + a2 )bw + a1 a2 d, f0 = a1 a2 bw (19)
g 2 = a1 + a2 , g 1 = a1 a2 , g 0 = 0
where the parameters ai , bi , bw and d are as defined in
(4).
Because the circuit has an integrator, the identification
method should be set up such that a pole of the denominator is fixed at s = 0. The identification software typically
allows to fix a number of poles and zeros at certain values,
[50]. Using the first equation of (19) and (4), R∞ is given
by:
R∞ = f3
The roots of s3 + g2 s2 + g1 s + g0 = 0 are a1 and a2 ,
and the one that has been fixed at s = 0. By using the
condition (12), select the smallest root as a2 , and the
remaining root as a1 . From (4), the circuit’s time constants
are obtained as follows:
1
τi =
for i = 1, 2.
ai
From (19) and (12), b1 , b2 and bw are obtained by solving
Acknowledgements
This work was funded by the University of Oxford
EPSRC Impact Acceleration Account Technology Fund
Awards EP/K503769/1 and EP/K036157/1. The authors
would like to thank the anonymous reviewers and the editor for their fruitful comments that significantly improved
the paper. We would like also to extend our thanks to Ross
Drummond, Stephen Duncan, Xinfan Lin and Shi Zhao for
their feedback on the first version of this work.
References
[1] J. Randles, “Kinetics of rapid electrode reactions,” Discuss.
Faraday Soc., vol. 1, pp. 11–19, 1947.
[2] C. Rahn and C.-Y. Wang, Battery Systems Engineering. John
Wiley Sons, 2013.
[3] S. M. M. Alavi, C. R. Birkl, and D. A. Howey, “Time-domain
fitting of battery electrochemical impedance models,” Journal
of Power Sources, vol. 288, pp. 345–352, 2015.
[4] E. Barsoukov and J. R. Macdonald, Impedance Spectroscopy:
Theory, Experiment, and Applications, 2nd ed. John Wiley &
Sons, 2005.
[5] S. Buller, M. Thele, R. W. A. A. De Doncker, and E. Karden,
“Impedance-based simulation models of supercapacitors and liion batteries for power electronic applications,” IEEE Transactions on Industry Applications, vol. 41, pp. 742–747, 2005.
[6] D. Andre, M. Meiler, K. Steiner, C. Wimmer, T. Soczka-Guth,
and D. U. Sauer, “Characterization of high-power lithium-ion
batteries by electrochemical impedance spectroscopy. I. Experimental investigation,” Journal of Power Sources, vol. 196, pp.
5334–5341, 2011.
[7] Y. Hu, S. Yurkovich, Y. Guezennec, and B. J. Yurkovich,
“Electro-thermal battery model identification for automotive
applications,” Journal of Power Sources, vol. 196, pp. 449–457,
2011.
[8] C. Birkl and D. Howey, “Model identification and parameter
estimation for LiFePO4 batteries,” Hybrid and Electric Vehicles
Conference 2013 (HEVC 2013), pp. 2.1–2.1, 2013.
7
[9] H. Rahimi-Eichi, F. Baronti, and M. Y. Chow, “Online adaptive
parameter identification and state-of-charge coestimation for
lithium-polymer battery cells,” IEEE Transactions on Industrial Electronics, vol. 61, pp. 2053–2061, 2014.
[10] V. Mihajlovic, B. Grundlehner, R. Vullers, and J. Penders,
“Wearable, wireless eeg solutions in daily life applications: What
are we missing?” IEEE Journal of Biomedical and Health Informatics, vol. 19, pp. 6–21, 2015.
[11] S. Bugenhagen, A. Cowley Jr., and D. Beard, “Baroreceptor
dynamics and their relationship to afferent fiber type and hypertension,” Physiological Genomics, vol. 42, pp. 23–41, 2010.
[12] A. Mahdi, J. Sturdy, T. Ottesen, and M. Olufsen, “Modeling
the afferent dynamics of the baroreflex control system,” PLoS
Computational Biology, vol. 9, p. e10033384, 2013.
[13] G. Mader, M. Olufsen, and A. Mahdi, “Modeling cerebral blood
flow velocity during orthostatic stress,” Annals of biomedical
engineering, pp. 1–11, 2014, dOI:10.1007/s10439-014-1220-4.
[14] J. Lee, J. Lee, O. Nam, J. Kim, B. H. Cho, H.-S. Yun, S.-S.
Choi, K. Kim, J. Kim, and S. Jun, “Modeling and Real Time
Estimation of Lumped Equivalent Circuit Model of a Lithium
Ion Battery,” 12th International Power Electronics and Motion
Control Conference, pp. 1536 – 1540, 2006.
[15] N. Moubayed, J. Kouta, A. El-Ali, H. Dernayka, and R. Outbib,
“Parameter identification of the lead-acid battery model,” 2008
33rd IEEE Photovolatic Specialists Conference, pp. 1–6, 2008.
[16] J. Jang and J. Yoo, “Equivalent circuit evaluation method of
lithium polymer battery using bode plot and numerical analysis,” IEEE Transactions on Energy Conversion, vol. 26, pp.
290–298, 2011.
[17] S. Jiang, “A Parameter Identification Method for a Battery
Equivalent Circuit Model,” SAE Technical Paper, 2011.
[18] B. Pattipati, C. Sankavaram, and K. R. Pattipati, “System
identification and estimation framework for pivotal automotive
battery management system characteristics,” IEEE Transactions on Systems, Man and Cybernetics Part C: Applications
and Reviews, vol. 41, pp. 869–884, 2011.
[19] A. Rahmoun and H. Biechl, “Modelling of Li-ion batteries using
equivalent circuit diagrams,” PRZEGLAD ELEKTROTECHNICZNY, vol. 2, pp. 152–156, 2012.
[20] R. Al-Nazer and V. Cattin, “A new optimization algorithm
for a Li-Ion battery equivalent electrical circuit identification,”
9th International Conference of Modeling, Optimization and
Simulation, 2012.
[21] A. Raue, C. Kreutz, T. Maiwald, J. Bachmann, M. Schilling,
U. Klingmüller, and J. Timmer, “Structural and practical identifiability analysis of partially observed dynamical models by
exploiting the profile likelihood,” Bioinformatics, vol. 25, pp.
1923–1929, 2009.
[22] C. Cobelli and J. J. DiStefano, “Parameter and structural
identifiability concepts and ambiguities: a critical review and
analysis.” The American journal of physiology, vol. 239, pp. R7–
R24, 1980.
[23] R. Bellman and K. Å ström, “On structural identifiability,”
Mathematical Biosciences, vol. 7, pp. 329–339, 1970.
[24] H. Pohjanpalo, “System identifiability based on the power series
expansion of the solution,” Mathematical Biosciences, vol. 41,
pp. 21–33, 1978.
[25] M. J. Chappell, K. R. Godfrey, and S. Vajda, “Global identifiability of the parameters of nonlinear systems with specified
inputs: A comparison of methods,” Mathematical Biosciences,
vol. 102, pp. 41–73, 1990.
[26] S. Vajda, K. R. Godfrey, and H. Rabitz, “Similarity transformation approach to identifiability analysis of nonlinear compartmental models.” Mathematical biosciences, vol. 93, pp. 217–248,
1989.
[27] F. Anstett, G. Bloch, G. Millérioux, and L. Denis-Vidal, “Identifiability of discrete-time nonlinear systems: The local state
isomorphism approach,” Automatica, vol. 44, pp. 2884–2889,
2008.
[28] N. Meshkat and S. Sullivant, “Identifiable reparametrizations of
linear compartment models,” Journal of Symbolic Computation,
vol. 63, pp. 46–67, 2014.
[29] A. Mahdi, N. Meshkat, and S. Sullivant, “Structural identifiability of viscoelastic mechanical systems,” PLoS ONE, vol. 9,
2014.
[30] K. Glover and J. Willems, “Parametrizations of linear dynamical
systems: Canonical forms and identifiability,” IEEE Transactions on Automatic Control, vol. AC-19, no. 6, pp. 640 – 646,
1974.
[31] J. Distefano, “On the relationships between structural identifiability and the controllability, observability properties,” IEEE
Transactions on Automatic Control, vol. 22, p. 652, 1977.
[32] J. M. Van Den Hof, “Structural identifiability of linear compartmental systems,” IEEE Transactions on Automatic Control,
vol. 43, pp. 800–818, 1998.
[33] L. Ljung, System Identification Theory for the User.
PTR
Prentice Hall Upper Saddle River NJ, 1987.
[34] L. Ljung and T. Glad, “On global identifiability for arbitrary
model parametrizations,” Automatica, vol. 30, pp. 265–276,
1994.
[35] S. Audoly, G. Bellu, L. D’Angiò, M. P. Saccomani, and C. Cobelli, “Global identifiability of nonlinear models of biological systems,” IEEE Transactions on Biomedical Engineering,
vol. 48, pp. 55–65, 2001.
[36] M. Sitterly, L. Y. Wang, G. G. Yin, and C. Wang, “Enhanced
identification of battery models for real-time battery management,” IEEE Transactions on Sustainable Energy, vol. 2, pp.
300–308, 2011.
[37] M. Rausch, S. Streif, C. Pankiewitz, and R. Findeisen, “Nonlinear observability and identifiability of single cells in battery
packs,” IEEE International Conference on Control Applications
(CCA), pp. 401–406, 2013.
[38] A. P. Schmidt, M. Bitzer, A. W. Imre, and L. Guzzella,
“Experiment-driven electrochemical modeling and systematic
parameterization for a lithium-ion battery cell,” Journal of
Power Sources, vol. 195, pp. 5071–5080, 2010.
[39] J. C. Forman, S. J. Moura, J. L. Stein, and H. K. Fathy, “Genetic
identification and fisher identifiability analysis of the DoyleFuller-Newman model from experimental cycling of a LiFePO
4 cell,” Journal of Power Sources, vol. 210, pp. 263–275, 2012.
[40] S. J. Moura, N. A. Chaturvedi, and M. Krstić, “Adaptive Partial Differential Equation Observer for Battery Stateof-Charge/State-of-Health Estimation Via an Electrochemical
Model,” Journal of Dynamic Systems, Measurement, and Control, vol. 136, p. 011015, 2013.
[41] A. M. D’Amato, J. C. Forman, T. Ersal, A. Ali, J. L. Stein,
H. Peng, and D. S. Bernstein, “Noninvasive battery-health
diagnostics using retrospective-cost identification of inaccessible
subsystems,” ASME 2012 5th Annual Dynamic Systems and
Control Conference Joint with the JSME 2012 11th Motion and
Vibration Conference, pp. 1–9, 2012.
[42] A. Sharma and H. Fathy, “Fisher identifiability analysis for
a periodically-excited equivalent-circuit lithium-ion battery
model,” American Control Conference (ACC), pp. 274–280,
2014.
[43] M. Rothenberger, J. Anstrom, S. Brennan, and H. Fathy,
“Maximizing Parameter Identifiability of an Equivalent-Circuit
Battery Model Using Optimal Periodic Input Shaping,” ASME
2014 Dynamic Systems and Control Conference, pp. 1–10, 2014.
[44] R. L. M. Peeters and B. Hanzon, “Symbolic computation of
Fisher information matrices for parametrized state-space systems,” Automatica, vol. 35, pp. 1059–1071, 1999.
[45] T. Söderström and P. Stoica, System identification. PrenticeHall, 1989.
[46] J. P. Norton, An introduction to identification. Academic Press,
1986.
[47] Y. Zhu, Multivariable System Identification For Process Control. Elsevier, 2001.
[48] D. A. Howey, P. D. Mitcheson, V. Yufit, G. J. Offer, and N. P.
Brandon, “Online measurement of battery impedance using
motor controller excitation,” IEEE Transactions on Vehicular
Technology, vol. 63, pp. 2557–2566, 2014.
[49] M. Schroeder, “Synthesis of low-peak-factor signals and binary
sequences with low autocorrelation,” IEEE Transactions on
Information Theory, vol. 16, pp. 85 – 89, 1970.
[50] L. Ljung, System Identification Toolbox For Use with MATLAB.
The MathWorks, Inc., 1988.
[51] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Dover
Publications, 2005.
8
TABLE I
Estimation results.
Noise free (9 outlier estimates)
Noise σ = 10−4 (11 outlier estimates)
Parameter
True Value
mean
st.d.
er
mean
st.d.
er
R∞
0.05
0.0552
0.0010
10.36
0.0552
0.0012
10.38
R1
0.2
0.1883
0.0393
5.87
0.1847
0.0451
7.63
C1
0.3
0.2919
0.0617
2.68
0.2886
0.0706
3.79
R2
0.4
0.4061
0.0360
1.52
0.4094
0.0415
2.34
C2
0.6
0.5866
0.0734
2.22
0.5803
0.0867
3.28
Cw
300
289.3709
48.3550
3.54
300.9287
87.8931
0.31
TABLE II
The coefficients of four widely used Randles circuit models.
R1
R1
i(t)
R∞
C1
+
R1
Cw
R∞
C1
+
R2
i(t)
R∞
i(t)
C1
+
C2
R1
R2
C1
C2
Cw
R∞
+
v(t)
v(t)
v(t)
v(t)
-
-
-
-
i(t)
Estimated TF:
s+f0
H(s) = f1s+g
0
Estimated TF:
2
+f1 s+f0
H(s) = f2s2s+g
1 s+g0
Set up to identify a pole at s = 0
Estimated TF:
2
+f1 s+f0
H(s) = f2s2s+g
1 s+g0
Estimated TF:
3
+f2 s2 +f1 s+f0
H(s) = f3s3s+g
2
2 s +g1 s+g0
Set up to identify a pole at s = 0
R∞ = f1
R∞ = f2
R∞ = f2
R∞ = f3
a1 = g 0
τ1 =
roots([1
1
a1
g1 g0 ]) ⇒ a1 , 0
τ1 =
roots([1
g1 g0 ]) ⇒ a1 , a2
choose a2 < a1
1
a1
τi =
AX = B
h
i
A=
B=
b 1 = f 0 − a1 f 1
C1 =
R1 =
1
b1
τ1
C1
X=
1
0
h
f1 − a1 f2
f0
b1
bw
C1 =
1
b1
Cw =
1
bw
R1 =
τ1
C1
1
a2
A=
i
⊤
B=
h
b1
1
a1
Ci =
Ri =
τi
Ci ,
b2
1
bi
g0 ]) ⇒ a2 , a1 , 0
a2 < a1
τi =
f1 − (a1 + a2 )f2
f0 − a1 a2 f2
X=
g2 g1
choose
1
ai
AX = B
h
i
1
a1
roots([1
⊤
i = 1, 2
A=
i
1
ai
AX = B
B=
1
a2
0
X=
1
a1
0
1
a1 + a2
a1 a2
f2 − (a1 + a2 )f3
f1 − a1 a2 f3
f0
b1
b2
Ci =
1
bi
Cw =
1
bw
Ri =
τi
Ci ,
bw
⊤
i = 1, 2
| 3cs.SY
|
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
arXiv:1612.06737v2 [math.AC] 7 Feb 2018
JAN DRAISMA AND FLORIAN M. OOSTERHOF
Abstract. We prove that iterated toric fibre products from a finite collection of
toric varieties are defined by binomials of uniformly bounded degree. This implies
that Markov random fields built up from a finite collection of finite graphs have
uniformly bounded Markov degree.
1. Introduction and main results
The notion of toric fibre product of two toric varieties goes back to [Sul07]. It is of
relevance in algebraic statistics since it captures algebraically the Markov random
field on a graph obtained by glueing two graphs along a common subgraph; see
[RS16] and also below. In [Sul07, RS16, RS14] it is proved that under certain conditions, one can explicitly construct a Markov basis for the large Markov random
field from bases for the components. For related results see [Shi12, EKS14, KR14].
However, these conditions are not always satisfied. Nevertheless, in [RS16, Conjecture 56] the hope was raised that when building larger graphs by glueing copies
from a finite collection of graphs along a common subgraph, there might be a
uniform upper bound on the Markov degree of the models thus constructed, independent of how many copies of each graph are used. A special case of this
conjecture was proved in the same paper [RS16, Theorem 54]. We prove the
conjecture in general, and along the way we link it to recent work [SS17] in representation stability. Indeed, an important point we would like to make, apart from
proving said conjecture, is that algebraic statistics is a natural source of problems
in asymptotic algebra, to which ideas from representation stability apply. Our main
theorems are reminiscent of Sam’s recent stabilisation theorems on equations and
higher syzygies for secant varieties of Veronese embeddings [Sam17a, Sam17b].
Markov random fields. Let G = (N, E) be a finite, undirected, simple graph and
for each node j ∈ N let X j be a random variable taking values in the finite set
[d j ] := {1, . . . , d j }. A joint probability distribution on (X j ) j∈N is said to satisfy the
local Markov properties imposed by the graph if for any two non-neighbours j, k ∈ N
the variables X j and Xk are conditionally independent given {Xl | {j, l} ∈ E}.
On the other hand, a joint probability distribution f on the X j is said Q
to factorise
according to G if for each maximal clique C of G and configuration α ∈ j∈C [d j ] of
C
the random variables labelled by
Q C there exists an interaction parameter θα such
that for each configuration β ∈ j∈N [d j ] of all random variables of G:
Y
f (β) =
θCβ|C
C∈mcl(G)
where mcl(G) is the set of maximal cliques of G, and β|C is the restriction of β to C.
1
2
JAN DRAISMA AND FLORIAN M. OOSTERHOF
These two notions are connected by the Hammersley-Clifford theorem, which says
that a positive joint probability distribution on G factorises according to G if and
only if it satisfies the Markov properties; see [HC71] or [Lau98, Theorem 3.9].
The set of all positive joint probability distributions on G that satisfy the Markov
properties is therefore a subset of the image of the following map:
Y
Q
Q
Q
[d
]
[d
]
C
C
ϕG : C C∈mcl(G) j∈C j → C j∈N j ,
(θα )C,α 7→
θβ|C
C∈mcl(G)
β
It is the ideal IG of polynomials vanishing on im ϕG that is of interest in algebraic
statistics. Since the components of ϕG are monomials, IG is generated by finitely
many
binomials (differences of two monomials) in the standard coordinates on
Q
C j∈N [d j ] , and any finite generating set of binomials can be used to set up a Markov
chain for testing whether given observations of the variables X j are compatible
with the assumption that their joint distribution factorises according to the graph
G [DS98]. The zero locus of IG is often called the graphical model of G.
Now suppose we have graphs G1 , . . . , Gs with node sets N1 , . . . , Ns , that Ni ∩ Nk
equals a fixed set N0 for all i , k in [s], and that the graphSinduced on N0 by each
Gi is equal to a fixed graph H. Moreover, for each j ∈ i Ni fix a number d j of
states. We can then glue copies of the Gi along their common subgraph H, by
which we mean first taking disjoint copies of the Gi and then identifying the nodes
labelled by a fixed j ∈ N0 across all copies. For a1 , . . . , aP
s ∈ Z≥0 , we denote
P the
graph obtained by glueing ai copies of graph Gi , i ∈ [s] by aH1 G1 +H · · · +H aHs Gs :
H
P2
H
H
+H
P1
H
=
Theorem 1. Let G1 , . . . , Gs be graphs with a common subgraph H and a number of states
associated to each node. Then there exists a uniform
bound D ∈PZ≥0 such that for all
P
multiplicities a1 , . . . , as , the ideal IG of G = aH1 G1 +H · · · +H aHs Gs is generated by
binomials of degree at most D.
Our proof shows that one needs only finitely many combinatorial types of binomials, independent of a1 , . . . , as , to generate IG . This result is similar in flavour
to the Independent Set Theorem from [HS12], where the graph is fixed but the d j
vary. Interestingly, the underlying categories responsible for these two stabilisation
phenomena are opposite to each other; see Remark 23.
Example 2. In [RS14] it is proved that the ideal IG for the complete bipartite graph
G = K3,N , with two states for each of the random variables, is generated in degree
at most 12 for all N. The graph G is obtained by glueing N copies of K3,1 along their
common subgraph consisting of 3 nodes without any edges.
We derive Theorem 1 from a general stabilisation result on toric fibre products, which
we introduce next.
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
3
Toric fibre products. Fix a ground field K, let r be a natural number, and let
U1 , . . . , Ur , V1 , . . . , Vr be finite-dimensional vector spaces over K.
Define a bilinear operation
Y
Y
Y
(1)
Uj ×
Vj →
(U j ⊗ V j ),
j
j
(u, v) 7→ u ∗ v := (u1 ⊗ v1 , . . . , ur ⊗ vr ).
j
Definition 3. The toric fibre product X ∗ Y of Zariski-closed subsets X ⊆
Q
Y ⊆ j V j equals the Zariski-closure of the set {u ∗ v | u ∈ X, v ∈ Y}.
Q
j
U j and
♣
Remark 4. In [Sul07], the toric fibre product is defined at the level of ideals: if
j
j
(xi )i are coordinate functions on U j and (yk )k are coordinate functions on V j , then
j
j
j
(zi,k := xi ⊗ yk )i,k are coordinate functions on U j ⊗ V j . The ring homomorphism of
coordinate rings
Y
Y
Y
j
j
j
K[(zi,k ) j,i,k ] = K (U j ⊗ V j ) → K
Uj ×
V j = K[(xi ) j,i , (yk ) j,k ]
j
j
j
j
j
j
dual to (1) sends zi,k to xi ·yk . If we compose this homomorphism with the projection
modulo the ideal of X × Y, then the kernel of the composition is precisely the toric
fibre product of the ideals of X and Y as introduced in [Sul07]. In that paper,
multigradings play a crucial role for computing toric fibre products of ideals, but
do not affect the definition of toric fibre products.
♣
The product ∗ is associative and commutative up to reordering tensor factors. We
can iterate this construction and form products like X∗2 ∗ Y ∗ Z∗3 , where Z also lives
Q
in a product of r vector spaces W j . This variety lives in j (U⊗2
) ⊗ V j ⊗ (W ⊗3
).
j
j
We will not be taking toric fibre products of general varieties, but rather Hadamardstable ones. For this, we have to choose coordinates on each U j , so that U j = Kd j .
Definition 5. On Kd the Hadamard product is defined as (a1 , . . . , ad ) (b1 , . . . , bd ) =
(a1 b1 , . . . , adQ
bd ). On U := U1 ×· · ·×Ur , where U j = Kd j , it is defined component-wise.
A set X ⊆ j U j is called Hadamard-stable if X contains the all-one vector 1U (the
unit element of ) and if moreover x z ∈ X for all x, z ∈ X.
♣
Remark 6. By [ES96, Remark after Proposition 2.3], X is Hadamard-stable if and
only if its ideal is generated by differences of two monomials.
P
In particular, the Zariski-closure in U of a subtorus of the j d j -dimensional torus
Q
dj
j (K \ {0}) is Hadamard-stable. These are the toric varieties from the abstract.
d0
Suppose that we
V j = K j and a corresponding Hadamard
Q also fix
Q identifications
Q
multiplication j V j × j V j → j V j . Equipping the spaces U j ⊗ V j with the natural coordinates and corresponding Hadamard multiplication, the two operations
∗ (v v0 ) = (u ∗Q
v) (u0 ∗ v0 ) as well as 1U j ⊗V j = 1U j ⊗ 1V j .
just defined satisfy (u u0 )Q
Consequently, if both X ⊆ j U j and Y ⊆ j V j are Zariski-closed and Hadamardstable, then so is their toric fibre product X ∗ Y. We can now formulate our second
main result.
4
JAN DRAISMA AND FLORIAN M. OOSTERHOF
Theorem 7. Let s, r ∈ Z≥0 . Q
For each i ∈ [s] and j ∈ [r] let di j ∈ Z≥0 and set Vi j := Kdi j .
For each i ∈ [s], let Xi ⊆
j Vi j be a Hadamard-stable Zariski-closed subset. Then
there exists a uniform bound
D
∈ Z≥0 such that for any exponents a1 , . . . , as the ideal of
Q Ns
⊗ai
X1∗a1 ∗ · · · ∗ Xs∗as ⊆ rj=1
V
is generated by polynomials of degree at most D.
i=1 i j
Remark 8. A straightforward generalisation of this theorem also holds, where
each Xi is a closed sub-scheme given by some ideal Ji in the coordinate ring of
Q
j Vi j . Hadamard-stable then says that that the pull-back of Ji under the Hadamard
product lies in the ideal of Xi × Xi , and the toric fibre product is as in Remark 4.
Since this generality would slightly obscure our arguments, we have decided to
present explicitly the version with Zariski-closed subsets—see also Remark 22.
Also, the theorem remains valid if we remove the condition that the Xi contain the
all-one vector, but require only that they be closed under Hadamard-multiplication;
see Remark 21.
♣
Organisation of this paper. The remainder of this paper is organised as follows.
In Section 2 we introduce the categories of (affine) Fin-varieties and, dually, Finop algebras. The point is that, as we will see in Section 4, the iterated toric fibre
products together form such a Fin-variety (or rather a Fins -variety, where Fins is
the product category of s copies of Fin).
Indeed, they sit naturally in a Cartesian product of copies of the Fin-variety of
rank-one tensors, which, as we prove in Section 3, is Noetherian (Theorem 12).
This Noetherianity result is of a similar flavour as the recent result from [SS17]
(see also [DK14, Proposition 7.5] which follows the same proof strategy) that any
finitely generated Finop -module is Noetherian; this result played a crucial role in a
proof of the Artinian conjecture. However, our Noetherianity result concerns certain
Finop -algebras rather than modules, and is more complicated. Finally, in Section 4
we first prove Theorem 7 and then derive Theorem 1 as a corollary.
Acknowledgements. Both authors are (partially) supported by Draisma’s Vici
grant Stabilisation in Algebra and Geometry from the Netherlands Organisation for
Scientific Research (NWO). The current paper is partly based on the second author’s Master’s thesis at Eindhoven University of Technology; see [Oos16]. We
thank Johannes Rauh and Seth Sullivant for fruitful discussions at Prague Stochastics, August 2014.
2. Affine Fin-varieties and Finop -algebras
The category Fin has as objects all finite sets and as morphisms all maps. Its
opposite category is denoted Finop . When C is another category whose objects are
called somethings, then a Fin-something is a covariant functor from Fin to C and a
Finop -something is a contravariant functor from Fin to C. The C-homomorphism
associated to a Fin-morphism π is denoted π∗ (the push-forward of π) in the covariant
case and π∗ (the pull-back of π) in the contravariant case.
More generally, we can replace Fin by any category D. The D-somethings themselves form a category, in which morphisms are natural transformations. In our
paper, D is always closely related to Fin or to Fins , the s-fold product of Fin.
Here are three instances of Fin or Finop -somethings crucial to our paper.
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
5
Example 9. Fix n ∈ Z≥0 . Then the functor S 7→ [n]S is a Finop -set, which to
π ∈ HomFin (S, T) associates the map π∗ : [n]T → [n]S , α 7→ απ, the composition of
α and π.
Building on this example, the functor An : S 7→ K[xα | α ∈ [n]S ], a polynomial
ring in variables labelled by [n]S , is a Finop -K-algebra, which associates to π the
K-algebra homomorphism π∗ : A(T) → A(S), xα 7→ xαπ .
Third, we define an affine Fin-K-vector space Qn by Qn : S 7→ (Kn )⊗S , the space of
n × · · · × n-tensors with factors labelled by S, which sends π ∈ HomFin (S, T) to the
linear morphism π∗ : (Kn )⊗S → (Kn )⊗T determined by
π∗ : ⊗i∈S vi 7→ ⊗ j∈T
i∈π−1 ( j) vi ,
where
is the Hadamard product in Kn . We follow the natural convention that
an empty Hadamard product equals the all-one vector 1 ∈ Kn ; in particular, this
holds in the previous formula for all j ∈ T that are not in the image of π.
The ring An and the space Qn are related as follows: Qn (S) has a basis consisting
of vectors eα := ⊗i∈S eαi , α ∈ [n]S , where e1 , . . . , en is the standard basis of Kn ;
An (S) is the coordinate ring of Qn (S) generated by the dual basis (xα )α∈[n]S ; and
for π ∈ HomFin (S, T) the pullback π∗ : An (T) → An (S) is the homomorphism of
K-algebras dual to the linear map π∗ : Qn (S) → Qn (T). Indeed, this is verified by
the following computation for α ∈ [n]T :
Y
Y
xα (π∗ ⊗i∈S vi ) = xα (⊗ j∈T
=
(vi )απ(i) = π∗ (xα )(⊗i∈S vi ).
i∈π−1 ( j) vi
i∈π−1 ( j) vi ) =
j∈T
αj
i∈S
This is used in Section 3. ♣
In general, by algebra we shall mean an associative, commutative K-algebra with
1, and homomorphisms are required to preserve 1. So a Finop -algebra B assigns to
each finite set S an algebra and to each map π : S → T an algebra homomorphism
π∗ : B(T) → B(S). An ideal in B is a Finop -subset I of B (i.e., I(S) is a subset of B(S) for
each finite set S and π∗ maps I(T) into I(S)) such that each I(S) is an ideal in B(S);
then S 7→ B(S)/I(S) is again a Finop -algebra, the quotient B/I of B by I.
Given a Finop -algebra B, finite sets S j for j in some index set J, and an element
b j ∈ B(S j ) for each j, there is a unique smallest ideal I in B such that each I(S j )
contains b j . This ideal is constructed as:
I(S) = π∗ (b j ) j ∈ J, π ∈ Hom(S, S j )
This is the ideal generated by the b j . A Finop -algebra is called Noetherian if each
ideal I in it is generated by finitely many elements in various I(S j ), i.e., J can be
taken finite.
Example 10. The Finop -algebra A1 is Noetherian. Indeed, A1 (S) is the polynomial
ring K[t] in a single variable t for all S, and the homomorphism A1 (T) → A1 (S)
is the identity K[t] → K[t]. So Noetherianity follows from Noetherianity of the
algebra K[t].
For n ≥ 2 the Finop -algebra An is not Noetherian. For instance, consider the
monomials
u2 := x21 x12 ∈ An ([2]), u3 := x211 x121 x112 ∈ An ([3]), u4 := x2111 x1211 x1121 x1112 ∈ An ([4]),
6
JAN DRAISMA AND FLORIAN M. OOSTERHOF
and so on. For any map π : [k] → [l] with k > l, by the pigeon hole principle
there are two indices i, j ∈ [k] such that π(i) = π( j) =: m ∈ [l]. Then π∗ x1···121···1 ,
where the 2 is in the m-th position, is a variable with at least two indices equal to 2.
Since uk contains no such variable, π∗ ul does not divide uk . So u2 , u3 , . . . generates
a non-finitely generated monomial Finop -ideal in An . (On the other hand, for each
d the piece of An of homogeneous polynomials of degree at most d is Noetherian
as a Finop -module, see [DK14, Proposition 7.5].)
♣
We shall see in the following section that certain interesting quotients of each An
are Noetherian.
3. Rank-one tensors form a Noetherian Fin-variety
Let Q≤1
n (S) be the variety of rank-one tensors, i.e., those of the form ⊗i∈S vi for vectors
vi ∈ Kn . We claim that this defines a Zariski-closed Fin-subvariety Q≤1
n of Qn .
For this, we must verify that for a map π : S → T the map Qn (S) → Qn (T) dual to
≤1
the algebra homomorphism An (T) → An (S) sends Q≤1
n (S) into Qn (T). And indeed,
in Example 9 we have seen that this map sends
⊗i∈S vi 7→ ⊗ j∈T
i∈π−1 ( j) vi .
It is well known that (if K is infinite) the ideal in An (S) of Q≤1
n (S) equals the ideal
In (S) generated by all binomials constructed as follows. Partition S into two parts
S1 , S2 , let αi , βi ∈ [n]Si and write α1 ||α2 for the element of [n]S which equals αi on Si .
Then we have the binomial
(2)
xα1 ||α2 xβ1 ||β2 − xα1 ||β2 xβ1 ||α2 ∈ In (S),
and In (S) is the ideal generated by these for all partitions and all α1 , α2 , β1 , β2 . The
functor S 7→ In (S) is an ideal in the Finop -algebra An ; for infinite K this follows from
the computation above, and for arbitrary K it follows since the binomials above are
mapped to binomials by pull-backs of maps T → S in Fin. Moreover, In is finitely
generated (see also [DK14, Lemma 7.4]):
Lemma 11. The ideal In in the Finop -algebra An is finitely generated.
Proof. In the determinantal equation (2), if there exist distinct j, l ∈ S1 such that
α1 ( j) = α2 (l) and β1 ( j) = β2 (l), then the equation comes from an equation in In (S\{j})
via the map S → S \ {j} that is the identity on S \ {j} and maps j → l. By the pigeon
hole principle this happens when |S1 | > n2 . Similarly for |S2 | > n2 . Hence In is
certainly generated by In ([2n2 − 1]).
The main result in this section is the following.
Theorem 12. For each n ∈ Z≥0 the coordinate ring An /In of the Fin-variety Q≤1
n of
rank-one tensors is a Noetherian Finop -algebra.
Our proof follows the general technique from [SS17], namely, to pass to a suitable
category close to Fin that allows for a Gröbner basis argument. However, the
relevant well-partial-orderedness proved below is new and quite subtle. We use
the category OS from [SS17] (also implicit in [DK14, Section 7]) defined as follows.
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
7
Definition 13. The objects of the category OS (“ordered-surjective”) are all finite
sets equipped with a linear order and the morphisms π : S → T are all surjective
maps with the additional property that the function T → S, j 7→ min π−1 ( j) is
strictly increasing.
♣
Any Fin-algebra is also an OS-algebra, and OS-Noetherianity implies Fin-Noetherianity. So to prove Theorem 12 we set out to prove the stronger statement that
An /In is, in fact, OS-Noetherian.
We get a more concrete grip on the K-algebra An /In through the following construction. Let Mn denote the (Abelian) Finop -monoid defined by
n
n
X
X
[n]×S
Mn (S) :=
α ∈ Z≥0 | ∀j, l ∈ S :
αi j =
αil
,
i=1
i=1
in which the multiplication is given by addition, and where the pull-back of a map
π : S → T is the map π∗ : Mn (T) → Mn (S) sending (αi j )i∈[n],j∈T to (αiπ(j) )i∈[n],j∈S .
So elements of Mn (S) are matrices with nonnegative integral entries and constant
column sum. Let KMn denote the Fin-algebra sending S to the monoid K-algebra
KMn (S). The following proposition is a reformulation of a well-known fact.
Proposition 14. The Finop -algebra An /In is isomorphic to the Finop -algebra KMn (and
the same is true when both are regarded as OSop -algebras).
Proof. For each finite set S, the K-algebra homomorphism ΦS : An (S) → KMn (S)
that sends xα , α ∈ [n]S to the [n] × S matrix in Mn (S) with a 1 at the positions
(α j , j), j ∈ S and zeroes elsewhere is surjective and has as kernel the ideal In (S).
Moreover, if π : S → T is a morphism in Fin, then we have ΦT π∗ = π∗ ΦS , i.e., the
(ΦS )S define a natural transformation.
Choose any monomial order > on Zn≥0 , i.e., a well-order such that a > b implies
a + c > b + c for every a, b, c ∈ Zn≥0 . Then for each object S in OS we define a
linear order > on Mn (S), as follows: α > β if α , β and the smallest j ∈ S with
α.,j , β., j (i.e., the j-th column of α is not equal to that of β) satisfies α.,j > β., j in
the chosen monomial order on Zn≥0 . A straightforward verification shows that this
is a monomial order on Mn (S) (we call the elements of Mn (S) monomials, even
though KMn (S) is not a polynomial ring). Moreover, for various S, these orders are
interrelated as follows.
Lemma 15. For any π ∈ HomOS (S, T) and α, β ∈ Mn (T), we have α > β ⇒ π∗ α > π∗ β.
Proof. If j ∈ T is the smallest column index where α and β differ, then α0 := π∗ α
and β0 := π∗ β differ in column l := min π−1 ( j), where they equal α.,j and β. j ,
respectively, and the former is larger than the latter. Furthermore, if l0 is the
smallest position where α0 , β0 differ, then α.,π(l0 ) , β.,π(l0 ) and hence π(l0 ) ≥ j and
hence l0 = min π−1 (π(l0 )) ≥ min π−1 ( j) = l. Hence in fact l = l0 and π∗ α > π∗ β.
In addition to the well-order ≤ on each individual Mn (S), we also need the following
partial order | on the union of all of them.
8
JAN DRAISMA AND FLORIAN M. OOSTERHOF
Definition 16. Let S, T be objects in OS. We say that α ∈ Mn (T) divides β ∈ Mn (S) if
there exist a π ∈ HomOS (S, T) and a γ ∈ Mn (S) such that β = γ + π∗ α. In this case,
we write α|β.
♣
The key combinatorial property of the relation just defined is the following.
Proposition 17. The relation | is a well-quasi-order, that is, for any sequence α(1) ∈
Mn (S1 ), α(2) ∈ Mn (S2 ), . . . there exist i < j such that α(i) |α( j) .
Proof. First, to each α ∈ Mn (S) we associate the monomial ideal P(α) in the polynomial ring R := K[z1 , . . . , zn ] (here K is but a place holder) generated by the
monomials zα.,j , j ∈ S. The crucial fact that we will use about monomial ideals
in R is that in any sequence P1 , P2 , . . . of such ideals there exist i < j such that
Pi ⊇ P j —in other words, monomial ideals are well-quasi-ordered with respect to
reverse inclusion [Mac01].
To prove the proposition, suppose, on the contrary, that there exists a sequence as
above with α(i) 6 | α( j) for all i < j. Such a sequence is called bad. Then by basic
properties of well-quasi-orders, some bad sequence exists in which moreover
(3)
P(α(1) ) ⊇ P(α(2) ) ⊇ . . .
Among all bad sequences with this additional property choose one in which, for each
j = 1, 2, . . . , the cardinality |S j | is minimal among all bad sequences starting with
α(1) , . . . , α(j−1) .
Write α(j) = (γ( j) |β(j) ), where β(j) ∈ Zn≥0 is the last column (the one labelled by the
largest element of S j ), and γ( j) is the remainder. By Dickson’s lemma, there exists a
subsequence j1 < j2 < . . . such that β(j1 ) , β( j2 ) , . . . increase weakly in the coordinatewise ordering on Zn≥0 . By restricting to a further subsequence, we may moreover
assume that also
(4)
P(γ(j1 ) ) ⊇ P(γ( j2 ) ) ⊇ . . .
Then consider the new sequence
α(1) , . . . , α( j1 −1) , γ(j1 ) , γ( j2 ) , . . .
By (4) and (3), and since P(α( j) ) ⊇ P(γ(j) ), this sequence also satisfies (3). We claim
that, furthermore, it is bad.
Suppose, for instance, that γ(j1 ) |γ(j2 ) . Set ai := max S ji for i = 1, 2. Then there exists
a π ∈ HomOS (S j2 \ {a2 }, S j1 \ {a1 }) such that γ(j2 ) − π∗ γ( j1 ) ∈ Mn (S j2 \ {a2 }). But then
extend π to an element π of HomOS (S j2 , S j1 ) by setting π(a2 ) := a1 ; since β(j1 ) is
coordinate-wise smaller than β( j2 ) we find that α( j2 ) − π∗ α( j1 ) ∈ Mn (S j2 ), so α( j1 ) |α( j2 ) ,
in contradiction to the badness of the original sequence.
On the other hand, suppose for instance that α(1) |γ( j2 ) and write a2 := max S j2 . Then
there exists a π ∈ HomOS (S j2 \ {a2 }, S1 ) such that γ( j2 ) − π∗ (α(1) ) ∈ Mn (S j2 \ {a2 }).
Now, and this is why we required that (3) holds, since P(α( j2 ) ) ⊆ P(α(1) ), there exists an
element s ∈ S1 such that the column β( j2 ) is coordinatewise at least as large as the
s-th column of α(1) . Extend π to an element of HomOS (S j2 , S1 ) by setting π(a2 ) = s.
Since a2 is the maximal element of S j2 , this does not destroy the property that the
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
9
function min π−1 (.) be increasing in its argument. Moreover, this π has the property
that α(j2 ) − π∗ α(1) ∈ Mn (S j2 ), again a contradiction.
Since we have found a bad sequence satisfying (3) but with strictly smaller underlying set at the j1 -st position, we have arrived at a contradiction.
Next we use a Gröbner basis argument.
Proof of Theorem 12. We prove the stronger statement that KMn is Noetherian as
an OS-algebra. Let P be any ideal in An /In = KMn . For each object S in OS, let
L(S) ⊆ Mn (S) denote the set of leading terms of nonzero elements of P(S) relative to
the ordering >. Proposition 17 implies that there exists a finite collection S1 , . . . , SN
and α( j) ∈ L(S j ) such that each element of each L(S) is divisible by some α( j) .
Correspondingly, there exist elements f j ∈ P(S j ) with leading monomial α( j) and
leading coefficient 1. To see that the f j generate P, suppose that there exists an S
such that P(S) is not contained in the ideal generated by the f j , and let g ∈ P(S)
have minimal leading term β among all elements of P(S) not in the ideal generated
by the f j ; without loss of generality g has leading coefficient 1. By construction,
there exists some j and some π ∈ HomOS (S, S j ) such that β − π∗ α(j) ∈ Mn (S). But
now, by Lemma 15, we find that the leading monomial of π∗ f j equals π∗ α( j) , hence
subtracting a monomial times π∗ f j from g we obtain an element of P(S) with smaller
leading monomial that is not in the ideal generated by the f j —a contradiction.
Below, we need the following generalisation of Theorem 12.
Theorem 18. For any n1 , . . . , nr ∈ (Z)≥0 the Fin-algebra (or OS-algebra) (An1 /In1 ) ⊗
· · · ⊗ (An /In1 ) is Noetherian.
Proof. This algebra is isomorphic to B := K(Mn1 × · · · × Mnr ). There is a natural
embedding ι : Mn1 × · · · × Mnr → Mn1 +...+ns =: Mn by forming a block matrix; its
image consists of block matrices with constant partial column sums. And while a
subalgebra of a Noetherian algebra is not necessarily Noetherian, this is true in the
current setting.
The crucial point is that if αi ∈ (Mn1 × · · · × Mnr )(Si ) for i = 1, 2, then a priori ι(α1 )|ι(α2 )
only means that ι(α2 ) − π∗ ι(α1 ) ∈ Mn (S2 ); but since both summands have constant
partial column sums, so does their difference, so in fact, the difference lies in the
image of ι. With this observation, the proof above for the case where r = 1 goes
through unaltered for arbitrary r.
Remark 19. Similar arguments for passing to sub-algebras are also used in [HS12]
and [Dra10].
♣
4. Proofs of the main results
In this section we prove Theorems 1 and 7.
Toric fibre products. To prove Theorem 7, we work with a product of s copies of
the category Fin; one for each of the varieties Xi whose iterated fibre products are
under consideration. Let s, r ∈ Z≥0 . For each i ∈ [s] and j ∈ [r] let di j ∈ Z≥0 and set
10
JAN DRAISMA AND FLORIAN M. OOSTERHOF
Vi j := Kdi j . Consider the Fins -variety V that assigns to an s-tuple S = (S1 , . . . , Ss )
the product
r O
s
Y
i
Vi⊗S
j
j=1
i=1
and to a morphism π = (π1 , . . . , πs ) : S → T := (T1 , . . . , Ts ) in Fins the linear map
V(S) → V(T) determined by
,
(5)
⊗i ⊗k∈Si vi jk
7→ ⊗i ⊗l∈Ti
k∈π−1 (l) vi jk
j∈[r]
j∈[r]
i
where the Hadamard product is the one on Vi j . Let Q (S) be the Zariski-closed
subset of V(S) consisting of r-tuples of tensors of rank at most one; thus Q≤1 is a
Fins -subvariety of V.
Qr
For each i ∈ [s] let Xi ⊆
j=1 Vi j be a Hadamard-stable Zariski-closed subset.
≤1
Then for any tuple S = (S1 , . . . , Ss ) in Fins the variety X(S) := X1∗S1 ∗ · · · ∗ Xs∗Ss is a
Zariski-closed subset of V(S).
Lemma 20. The association S 7→ X(S) defines a Fins -closed subvariety of Q≤1 .
Proof. From the definition of ∗ in (1) it is clear that the elements of X(S) are r-tuples
of tensors of rank at most 1. Furthermore, for a morphism S → T in Fins the linear
map V(S) → V(T) from (5) sends X(S) into X(T)—here we use that if (vi, j,k ) j∈[r] ∈ Xi
for each k ∈ π−1 (l), then also ( k∈π−1 (l) vi jk ) j∈[r] ∈ Xi since Xi is Hadamard-stable.
Now Theorem 7 follows once we know that the coordinate ring of Q≤1 is a Noetherian (Fins )op -algebra. For s = 1 and r = 1 this is Theorem 12 with n equal to d11 .
For s = 1 and general r, this is Theorem 18 with n j equal to d1j .
For r = 1 and general s, Theorem 7 follows from a Fins -analogue of Theorem 12,
which is proved as follows. The coordinate ring of Q≤1 (S1 , . . . , Ss ) is the subring
of K(Md11 (S1 ) × · · · × Mds1 (Ss )) spanned by the monomials corresponding to s-tuples
of matrices with the same constant column sum. Using Proposition 17 and the
fact that a finite product of well-quasi-ordered sets is well-quasi-ordered one finds
that the natural Fins -analogue on Md11 × · · · × Mds1 of the division relation | is a
well-quasi-order; and this implies, once again, that the coordinate ring of Q≤1 is a
Noetherian (OSs )op -algebra.
Finally, for general s and general r, the result follows as in the proof of Theorem 18.
This proves the Theorem 7 in full generality.
Remark 21. The only place where we used that the Xi contain the all-one vector is
in the proof of Lemma 20 when π−1 (l) happens to be empty. If we do not require
this, then the conclusion of Theorem 7 still holds, since one can work directly with
the category OSs in which morphisms π are surjective.
Remark 22. If we replace the Xi by Hadamard-stable closed subschemes rather
than subvarieties, then S 7→ X(S) is still a Fins -closed subscheme of Q≤1 , and since
the coordinate ring of the latter is Noetherian, the proof goes through unaltered.
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
11
Remark 23. In the Independent Set Theorem from [HS12], the graph G = (N, E)
is fixed but the state space sizes d j grow unboundedly for j in an independent set
T ⊆ N and are fixed for j < T. In this case, given a T-tuple of maps (π j : S j → P j ) j∈T
of finite sets, where S j is thought of as the state space of j ∈ T in the smaller model
and P j as the state space in the larger model, we obtain a natural map from the
larger model into the smaller model. Hence then the graphical model is naturally
a (Finop )T -variety and its coordinate ring is a FinT -algebra. Note the reversal of the
roles of these two categories compared to Lemma 20.
Markov random fields. Given a finite (undirected, simple) graph G = (N, E) with a
number d j of states attached to each node j ∈ N, the graphical model is XG := im ϕG ,
where ϕG is the parameterisation
Y
Q
Q
Q
(θCα )C,α 7→
ϕG : C C∈mcl(G) j∈C [d j ] → C j∈N [d j ] ,
θCβ|C .
C∈mcl(G)
β
Lemma 24. For any finite graph G, the graphical model XG is Hadamard-closed.
Proof. The parameterisation ϕ sends the all-one vector in the domain to the all-one
vector 1 in the target space, so 1 ∈ im ϕ. Moreover, if θ, θ0 are two parameter
vectors, then ϕ(θ θ0 ) = ϕ(θ) ϕ(θ0 ), so im ϕ is Hadamard-closed. Then so is its
closure.
Following [Sul07], we relate graph glueing to toric fibre products. We are given
finite graphs G1 , . . . , Gs with node sets N1 , . . . , Ns such that Ni ∩ Nk = N0 for all i , k
in [s]
S and such that each Gi induces the same graph H on N0 . Moreover, for each
j ∈ i Ni we fix a number d j of states.
Q
Q
For each β0 ∈ j∈N0 [d j ] and each i ∈ [s] set Vi,β0 := C j∈Ni \N0 [d j ] , which we interpret
as the ambient space of the part of the probability table of the variables X j , j ∈ Ni
where we have fixed the states of the variables in N0 to β0 —up to scaling, these are
the conditional joint probabilities
for the X j , j ∈ Ni \ N0 given that the X j , j ∈ N0 are
Q
in joint state β0 . For β ∈ j∈Ni [d j ] write β = β0 ||β0 where β0 , β0 are the restrictions of
β to N0 and Ni \N0 , respectively. For each maximal clique
Q C in Gi define C0 = C∩N0
and C0 = C \ N0 . Correspondingly, decompose α ∈ j∈C [d j ] as α = α0 ||α0 , where
α0 , α0 are the restrictions of α to C0 and C0 , respectively.
Then the graphical model XGi := im ϕGi is the closure of the image of the parameterisation
Y
Y
Q
Q
θC(β0 |C ) || (β0 |C0 ) .
ϕi : C C∈mcl(Gi ) j∈C [d j ] →
Vi,β0 ,
(θCα )C,α 7→
0
Q
C∈mcl(G)
β0 ∈
j∈N0 [d j ]
Q
β0 β0
Setting r := j∈N0 d j , we are exactly in the setting of the previous sections: for each
i, k ∈ [s], we have the bilinear map
Y
Y
Y
∗:
Vi,β0 ×
Vk,β0 →
(Vi,β0 ⊗ Vk,β0 ),
((vi,β0 )β0 , (vk,β0 )β0 ) 7→ (vi,β0 ⊗ vk,β0 )β0 ,
β0
β0
β0
and we can takeQiterated products of this type. The space on the right is naturally
isomorphic to C j∈N1 ∪N2 [d j ] , the space of probability tables for the joint distribution
12
JAN DRAISMA AND FLORIAN M. OOSTERHOF
of the variables labelled by the vertices in the glued graph Gi +H Gk . Under this
identification we have the following.
P
P
Proposition 25. For G := aH1 G1 +H · · · +H aHs Gs we have XG = XG∗a11 ∗ · · · ∗ XG∗ass .
Proof. It suffices to prove this for the gluing of two graphs. Note that a clique in
G := G1 +H G2 is contained entirely in either G1 or G2 , or in both but then already
in H. Let θ, η be a parameter vectors in the domains of ϕG1 , ϕG2 , respectively. Then
ϕG1 (θ) ∗ ϕG2 (η)
Y
=
θC(β0 |C ) || (β0 |C0 )
0
C∈mcl(G1 )
Y
=
C∈mcl(G1 )
Y
=
C∈mcl(G1 )
Y
C
∗
η
0| 0 )
)
||
(β
(β
|
0 C0
C
C∈mcl(G )
β0 β0
Q
β0 ∈ j∈N1 \N0 [d j ]
Y
θC(β0 |C
·
) || (β0 |C0 )
0
C∈mcl(G2 )
Y
C
=
µ(β0 |C ) || (β0 |C0 )
0
C∈mcl(G)
2
θC(β0 |C ) || (β0 |C0 )
0
Y
⊗
β0 β0
ηC(β0 |C ) || (β0 |C0 )
0
ηC(β0 |C ) || (β0 |C0 )
0
Q
β0 ∈ j∈(N1 ∪N2 )\N0 [d j ]
Q
C∈mcl(G2 )
Q
β0 ∈ j∈N2 \N0 [d j ]
Q
β0 ∈ j∈(N1 ∪N2 )\N0 [d j ]
= ϕG (µ),
β0
β0
β0
where, for C ∈ mcl(G) and α ∈ j∈C [d j ], the parameter µCα is defined as
θCα
if C ⊆ N1 and C * N0 ,
C
C
µα :=
ηα
if C ⊆ N2 and C * N0 , and
θC ηC if C ⊆ N0 .
α α
This computation proves that XG1 ∗ XG2 ⊆ XG . Conversely, given any parameter
vector µ for G, we can let θ be the restriction of µ to maximal cliques of the first
and third type above, and set ηCα equal to µCα if C is of the second type above and
equal to 1 if it is of the third type. This yields the opposite inclusion.
Proof of Theorem 1. By Proposition 25, the ideal IG is the ideal of the iterated toric
fibre product XG∗a11 ∗ · · · ∗ XG∗ass . By Lemma 24, each of the varieties XGi is Hadamard
closed. Hence Theorem 7 applies, and IG is generated by polynomials of degree
less than some D, which is independent of a1 , . . . , as . Then it is also generated by
the binomials of at most degree D.
References
[DK14]
[Dra10]
[DS98]
Jan Draisma and Jochen Kuttler. Bounded-rank tensors are defined in bounded degree. Duke
Math. J., 163(1):35–63, 2014.
Jan Draisma. Finiteness for the k-factor model and chirality varieties. Adv. Math., 223:243–256,
2010.
Persi Diaconis and Bernd Sturmfels. Algebraic algorithms for sampling from conditional
distributions. Ann. Stat., 26(1):363–397, 1998.
MARKOV RANDOM FIELDS AND ITERATED TORIC FIBRE PRODUCTS
13
[EKS14]
Alexander Engström, Thomas Kahle, and Seth Sullivant. Multigraded commutative algebra
of graph decompositions. J. Algebr. Comb., 39(2):335–372, 2014.
[ES96]
David Eisenbud and Bernd Sturmfels. Binomial ideals. Duke Math. J., 84(1):1–45, 1996.
[HC71] J.M. Hammersley and P. Clifford. Markov fields on finite graphs and lattices. Unpublished,
http://www.statslab.cam.ac.uk/˜grg/books/hammfest/hamm-cliff.pdf, 1971.
[HS12]
Christopher J. Hillar and Seth Sullivant. Finite Gröbner bases in infinite dimensional polynomial rings and applications. Adv. Math., 221:1–25, 2012.
[KR14]
Thomas Kahle and Johannes Rauh. Toric fiber products versus Segre products. Abh. Math.
Semin. Univ. Hamb., 84(2):187–201, 2014.
[Lau98] Steffen L. Lauritzen. Graphical models, volume 17 of Oxford Statistical Science Series. Oxford
Univ. Press., Oxford, 1998.
[Mac01] Diane MacLagan. Antichains of monomial ideals are finite. Proc. Am. Math. Soc., 129(6):1609–
1615, 2001.
[Oos16] Florian M. Oosterhof. Stabilisation of iterated toric fibre products. 2016. Master’s thesis, Eindhoven University of Technology, http://repository.tue.nl/855107.
[RS14]
Johannes
Rauh
and
Seth
Sullivant.
The
Markov
basis
of
k3,N .
\protect\vrule width0pt\protect\href{http://arxiv.org/abs/1406.5936}{arXiv:1406.5936},
2014.
[RS16]
Johannes Rauh and Seth Sullivant. Lifting Markov bases and higher codimension toric fiber
products. J. Symb. Comp, 74:276–307, 2016.
[Sam17a] Steven V. Sam. Ideals of bounded rank symmetric tensors are generated in bounded degree.
Invent. Math., 207(1):1–21, 2017.
[Sam17b] Steven V. Sam. Syzygies of bounded rank symmetric tensors are generated in bounded
degree. Math. Ann., 368(3–4):1095–1108, 2017.
[Shi12]
Takafumi Shibuta. Gröbner bases of contraction ideals. J. Algebr. Comb., 36(1):1–19, 2012.
[SS17]
Steven V. Sam and Andrew Snowden. Gröbner methods for representations of combinatorial
categories. J. Am. Math. Soc., 30(1):159–203, 2017.
[Sul07]
Seth Sullivant. Toric fiber products. J. Algebra, 316(2):560–577, 2007.
(Jan Draisma) Mathematical Institute, University of Bern, Sidlerstrasse 5, 3012 Bern, Switzerland;
and Eindhoven University of Technology, The Netherlands
E-mail address: [email protected]
(Florian M. Oosterhof) Department of Mathematics and Computer Science, Eindhoven University
of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands
E-mail address: [email protected]
| 0math.AC
|
Logical Methods in Computer Science
Vol. 1 (3:2) 2005, pp. 331–353
www.lmcs-online.org
Submitted
Published
Feb. 5, 2005
Dec. 20, 2005
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL
COMPUTING
DANIELE GORLA a , MATTHEW HENNESSY b , AND VLADIMIRO SASSONE c
a
Dip. di Informatica, Univ. di Roma “La Sapienza”
e-mail address: [email protected]
b
Dept. of Informatics, Univ. of Sussex
e-mail address: [email protected]
c
Dept. of Informatics, Univ. of Sussex
e-mail address: [email protected]
Abstract. We propose a simple global computing framework, whose main concern is
code migration. Systems are structured in sites, and each site is divided into two parts: a
computing body, and a membrane which regulates the interactions between the computing
body and the external environment. More precisely, membranes are filters which control
access to the associated site, and they also rely on the well-established notion of trust
between sites. We develop a basic theory to express and enforce security policies via
membranes. Initially, these only control the actions incoming agents intend to perform
locally. We then adapt the basic theory to encompass more sophisticated policies, where
the number of actions an agent wants to perform, and also their order, are considered.
1. Introduction
Computing is increasingly characterised by the global scale of applications and the
ubiquity of interactions between mobile components. Among the main features of the forthcoming “global ubiquitous computing” paradigm we list distribution and location awarness,
whereby code located at specific sites acts appropriately to local parameters and circumstances, that is, it is “context-aware”; mobility, whereby code is dispatched from site to site
to increase flexibility and expressivity; openness, reflecting the nature of global networks
and embodying the permeating hypothesis of localised, partial knowledge of the execution
environment. Such systems present enormous difficulties, both technical and conceptual,
and are currently more at the stage of exciting future prospectives than that of established
of engineering practice. Two concerns, however, appear to clearly have a ever-reaching
2000 ACM Subject Classification: F.1.1, F.3.1, D.3.1, D.4.6.
Key words and phrases: Process calculi, Mobile code, Language-based security, Type systems.
a
This work has been mostly carried on while the first author was at the Dept. of Informatics, Univ. of
Sussex, with a Marie Curie Fellowship.
a,b,c
The authors would like to acknowledge the support of the EC Global Computing projects Mikado and
Myths.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-1 (3:2) 2005
CC
D. Gorla, M. Hennessy, and V. Sassone
Creative Commons
2
D. GORLA, M. HENNESSY, AND V. SASSONE
import: security and mobility control, arising respectively from openness and from massive
code and resource migrations. They are the focus of the present paper.
We aim at classifying mobile components according to their behaviour, and at empowering sites with control capabilities which allow them to deny access to those agents whose
behaviour does not conform to the site’s policy. We see every site of a system
k[[ M |i P ]]
as an entity named k and structured in two layers: a computing body P , where programs
run their code – possibly accessing local resources offered by the site – and a membrane M ,
which regulates the interactions between the computing body and the external environment.
An agent P wishing to enter a site l must be verified by the membrane before it is given a
chance to execute in l. If the preliminary check succeeds, the agent is allowed to execute,
otherwise it is rejected. In other words, a membrane implements the policy each site wants
to enforce locally, by ruling on the requests of access of the incoming agents. This can be
easily expressed by a migration rule of the form:
k[[ M k |i gol.P | Q ]] k l[[ M l |i R ]] → k[[ M k |i Q ]] k l[[ M l |i P | R ]]
if M l ⊢k P
The relevant parts here are P , the agent wishing to migrate from k to l, and l, the receiving
site, which needs to be satisfied that P ’s behaviour complies with its policy. The latter
is expressed by l’s membrane, M l . The judgement M l ⊢k P represents l inspecting the
incoming code to verify that it upholds M l .
Observe that in the formulation above M l ⊢k P represents a runtime check of all
incoming agents. Because of our fundamental assumption of openendedness, such kind of
checks, undesirable as they might be, cannot be avoided. In order to reduce their impact on
systems performance, and to make the runtime semantics as efficient as possible, we adopt
a strategy which allows for efficient agent verification. Precisely, we adopt an elementary
notion of trust, so that from the point of view of each l the set of sites is consistently
partitioned between “good,” “bad,” and “unknown” sites. Then, in a situation like the
one in the rule above, we assume that l will be willing to accept from a trusted site k a
k-certified digest T of P ’s behaviour. We then modify the primitive go and the judgement
⊢k as in the refined migration rule:
k[[ M k |i go T l.P | Q ]] k l[[ M l |i R ]] → k[[ M k |i Q ]] k l[[ M l |i P | R ]]
if M l ⊢kT P
The notable difference is in M l ⊢kT P . Here, l verifies the entire code P against M l only if
it does not trust k, the signer of P ’s certificate T. Otherwise, it suffices for l to match M l
against the digest T carried by go together with P from k, so effectively shifting work from
l to the originator of P .
Our main concern in this paper is to put the focus on the machinery a membrane should
implement to enforce different kinds of policies. We first distill the simplest calculus which
can conceivably convey our ideas and still support a non-trivial study. It is important to
remark that we are abstracting from agents’ local computations. These can be expressed
in any of several well-known models for concurrency, for example CCS [Mil82] or the π–
calculus [Mil99]. We are concerned, instead, with agents’ migration from site to site: our
main language mechanism is go rather than intra-site (i.e. local) communication. Using this
language, we examine four notions of policy and show how they can be enforced by using
membranes. We start with an amusingly simple policy which only lists allowed actions.
We then move to count action occurrences and then to policies expressed by deterministic
finite automata. Note that such policies are only concerned with the behaviour of single
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
3
agents, and do not take into account “coalitional” behaviours, whereby incoming agents –
apparently innocent – join clusters of resident agents – they too apparently innocent – to
perform cooperatively potentially harmful actions, or at least overrule the host site’s policy.
We call resident those policies intended to be applied to the joint, composite behaviour of
the agents contained at a site. We explore resident policies as our fourth and final notion of
policy. In all the cases, the theory adapts smoothly; we only need to refine the information
stored in the membrane and the inspection mechanisms.
Structure of the paper. In Section 2 we define the calculus used in this paper, and start
with the straightforward policy which only prescribes the actions an agent can perform
when running in a site. In Section 3, we enhance the theory to control also how many (and
not only which kind of) actions an agent wants to perform in a site, and their order of
execution. Finally, in Section 4 we extend the theory to control the overall computation
taking place at a site, and not only the behaviour of single agents. The paper concludes in
Section 5 where a comparison with related work is also given. The theoretical results are
proved in Appendix A. With respect to the extended abstract [GHS04], this paper contains
more examples together with complete proofs.
2. A Simple Calculus
In this section we describe a simple calculus for mobile agents, which may migrate
between sites. Each site is guarded by a membrane, whose task is to ensure that every
agent accepted at the site conforms to an entry policy.
2.1. The Syntax.
The syntax is given in Figure 1 and assumes two pairwise disjoint sets: basic agent
actions Act, ranged over by a, b, c, · · · , and localities Loc, ranged over by l, k, h, · · · . Agents
are constructed using the standard action-prefixing, parallel composition and replication
operators from process calculi, [Mil82]. The one novel operator is that for migration,
go T l.P
This agent seeks to migrate to site l in order to execute the code P ; moreover it promises
to conform to the entry policy T. In practical terms this might consist of a certification
that the incoming code P conforms to the policy T, which the site l has to decide whether
or not to accept. In our framework, this certification is a policy that describes the (local)
behaviour of the agent; thus, in go T l.P , T will be called the digest of P .
A system consists of a finite set of sites running in parallel. A site takes the form
l[[ M |i P ]]
where
• l is the site name
• P is the code currently running at l
• M is the membrane which implements the entry policy.
For convenience we assume that site names are unique in systems. Thus, in a given system
we can identify the membrane associated with the site named l by M l . We start with a
very simple kind of policy, which we will then progressively enhance.
4
D. GORLA, M. HENNESSY, AND V. SASSONE
Basic Actions a, b, c, ... ∈ Act
Localities l, h, k, ... ∈ Loc
Agents P, Q, R ::=
|
|
|
|
Systems
N
nil
a.P
go T l.P
P |Q
!P
nil agent
basic action
migration
composition
replication
::= 0
empty system
| l[[ M |i P ]]
site
| N1 k N2
composition
Figure 1: A Simple Calculus
Definition 2.1 (Policies). A policy is a finite subset of Act ∪ Loc. For two policies T1
and T2 , we write
T1 enforces T2
whenever T1 ⊆ T2 .
Intuitively an agent conforms to a policy T at a given site if
• every action it performs at the site is contained in T
• it will only migrate to sites whose names are in T.
For example, conforming to the policy {info, req, home}, where info, req are actions
and home a location, means that the only actions that will be performed are from the
set {info, req} and migration will only occur, if at all, to the site home. With this
interpretation of policies, our definition of the predicate enforces is also intuitive; if some
code P conforms to the policy T1 and T1 enforces T2 then P also automatically conforms
to T2 .
The purpose of membranes is to enforce such policies on incoming agents. In other
words, at a site l[[M |iQ]] wishing to enforce a policy Tin , the membrane M has to decide when
to allow entry to an agent such as go T l.P from another site. There are two possibilities.
• The first is to syntactically check the code P against the policy Tin ; an implementation
would actually expect the agent to arrive with a proof of this fact, and this proof would
be checked.
• The second would be to trust the agent that its code P conforms to the stated T and
therefore only check that this conforms to the entry policy Tin . Assuming that checking
one policy against another is more efficient than the code analysis, this would make entry
formalities much easier.
Deciding on when to apply the second possibility presupposes a trust management framework for systems, which is the topic of much current research. To simplify matters, here
we simply assume that each site contains, as part of its membrane, a record of the level of
trust it has in other sites. Moreover, we assume only three possible levels: bad, unknown
and good. Intuitively, a site is good/bad if it behaves in a reliable/unreliable way, i.e. it
does/doesn’t properly calculate digests. On the other hand, a site tagged as unknown can
behave in a non specified way; thus, for the sake of security, it will be considered as bad.
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
(r-act)
(r-par)
(r-struct)
(r-mig)
l[[ M |i a.P | Q ]]
→
5
l[[ M |i P | Q ]]
N1 → N1′
N1 k N2 → N1′ k N2
N ≡ N1
N1 → N1′
N1′ ≡ N ′
N → N′
k[[ M k |i go T l.P | Q ]] k l[[ M l |i R ]]
k[[ M k
→
|i Q ]] k l[[ M l |i P | R ]]
if M l ⊢kT P
Figure 2: The reduction relation
l[[ M |i P | nil ]] ≡ l[[ M |i P ]]
N k0 ≡ N
l[[ M |i P | Q ]] ≡ l[[ M |i Q | P ]]
N1 k N2 ≡ N2 k N1
l[[ M |i (P | Q) | R ]] ≡ l[[ M |i P | (Q | R) ]]
(N1 k N2 ) k N3 ≡ N1 k (N2 k N3 )
l[[ M |i !P | Q ]] ≡ l[[ M |i P | !P | Q ]]
Figure 3: The structural equivalence
In a more realistic scenario, it would be possible to refine unknown to either good or bad,
upon collection of enough evidence to consider it reliable or not. For the sake of simplicity,
we do not model this framework here.
Definition 2.2 (Membranes). A membrane M is a pair (Mt , Mp ) where
• Mt is a partial function from Loc to {unknown, good, bad}
• Mp is a policy
2.2. The Operational Semantics.
Having defined both policies and membranes, we now give an operational semantics for
the calculus, which formalises the above discussion on how to manage agent migration. This
is given as a binary relation N → N ′ over systems; it is defined to be the least relation which
satisfies the rules in Figure 2. Rule (r-act) says that the agent a.P running in parallel
with other code in site l, such as Q, can perform the action a; note that the semantics does
not record the occurrence of a. (r-par) and (r-struct) are standard. The first allows
reductions within parallel components, while the second says that reductions are relative
to a structural equivalence; the rules defining this equivalence are given in Figure 3. The
interesting reduction rule is the last one, (r-mig), governing migration; the agent goT l.P
can migrate from site k to site l provided the predicate M l ⊢kT P is true. This ‘enabling’
predicate formalises our discussion above on the role of the membrane M l , and requires in
turn a notion of code P satisfying a policy T,
⊢P :T
6
D. GORLA, M. HENNESSY, AND V. SASSONE
(tc-act)
(tc-mig)
⊢ nil : T
⊢P :T
a∈T
⊢ a.P : T
⊢ P : T′
⊢ go T′ l.P : T
(tc-repl)
(tc-par)
⊢P :T
⊢ !P : T
⊢P :T
⊢Q:T
⊢P |Q:T
(tc-empty)
l∈T
Figure 4: Typechecking incoming agents
With such a notion, we can then define M l ⊢kT P to be:
if Mtl (k) = good then (T enforces Mpl ) else ⊢ P : Mpl
(2.1)
In other words, if the target site l trusts the source site k, it trusts that the professed policy
T is a faithful reflection of the behaviour of the incoming agent P , and then entry is gained
provided that T enforces the entry policy Mpl (i.e., in this case, T ⊆ Mpl ). Otherwise, if k
can not be trusted, then the entire incoming code P has to be checked to ensure that it
conforms to the entry policy, as expressed by the predicate ⊢ P : Mpl .
In Figure 4 we describe a simple inference system for checking that agents conform to
policies, i.e. to infer judgements of the form ⊢ P : T. Rule (tc-empty) simply says that
the empty agent nil satisfies all policies. (tc-act) is also straightforward; a.P satisfies a
policy T and if a is allowed by T, and the residual P satisfies T. The rule (tc-par) says
that to check P | Q it is sufficient to check P and Q separately, and similarly for replicated
agents. The most interesting rule is (tc-mig), which checks go T′ l.P . This not only checks
that migration to l is allowed by the policy, that is l ∈ T, but it also checks that the code to
be spawned there, P , conforms to the associated professed policy T′ . In some sense, if the
agent goT′ l.P is allowed entry into a site k, then k assumes responsibility for any promises
that it makes about conformance to policies.
2.3. Safety.
We have just outlined a reduction semantics in which sites seek to enforce policies either
by directly checking the code of incoming agents against entry policies, or more simply by
checking the professed policy of trusted agents. The extent to which this strategy works
depends, not surprisingly, on the quality of a site’s trust management.
Example 2.3. Let home be a site name with the following trust function
Mth : {alice, bob, secure}
7→
good .
Consider the system
△
N = home[[ M h |i P h ]] k bob[[ M b |i P b ]] k alice[[ M a |i P a ]] k secure[[ M s |i P s ]]
in which the entry policy of home, Mph , is {info, req, secure}, and that of secure, Mps ,
is {give, home}. Since Mth (bob) = good, agents migrating from bob to home are trusted
and only their digests are checked against the entry policy Mph . So, if P b contains the agent
goT1 home.(take.Q)
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
7
where T1 enforces Mph , then the entry policy of home will be transgressed.
As another example, suppose alice, again trusted by home, contains the agent
go T1 home.(info.go T2 secure.(take.Q))
where T2 is some policy which enforces the entry policy of secure, Mps . Again because
T1 enforces Mph , the migration is allowed from alice to home, and moreover the incoming
agent conforms to the policy demanded of home. The second migration of the agent is also
successful if secure trusts home: Mts (home) = good and therefore only the digest T2 is
checked against the entry policy of secure. We then have the reduction
N
−
→∗ home[[. . .]] k bob[[. . .]] k alice[[. . .]] k secure[[ M s |i take.Q | P s ]]
in which now the entry policy of secure has been foiled.
The problem in this example is that the trust knowledge of home is faulty; it trusts in
sites which do not properly ensure that professed policies are enforced. Let us divide the
sites into trustworthy and otherwise. This bipartition could be stored in an external record
stating which nodes are trustworthy (i.e. typechecked) and which ones are not. However,
for economy, we prefer to record this information in the membranes, by demanding that
the trust knowledge at trustworthy sites is a proper reflection of this division. This is more
easily defined if we assume the following ordering over trust levels:
unknown <: bad
and
unknown <: good
This reflects the intuitive idea that sites classified as unknown may, perhaps with further
information, be subsequently classified either as good or bad. On the other hand, good or
bad cannot be further refined; sites classified as either, will not be reclassified.
Definition 2.4 (Trustworthy sites and Coherent systems). In a system N , the site k is
trustworthy if Mtk (k) = good. N is coherent if Mtk (l) <: Mtl (l) for every trustworthy site k.
Thus, if a trustworthy site k believes that a site l can be trusted (i.e., Mtk (l) = good),
then l is indeed trustworthy (as represented by Mtl (l) = good). Similarly, if it believes l to
be bad, then l is indeed bad. The only uncertainty is when k classifies l as unknown: then l
may be either good or bad. Of course, in coherent systems we expect sites which have been
classified as trustworthy to act in a trustworthy manner, which amounts to saying that code
running at such a k must have at one time gained entry there by satisfying the entry policy.
Note that by using policies as in Definition 2.1, if P satisfies an entry policy Mpk , then it
continues to satisfy the policy while running at k (cf. Theorem 2.7 below).
This property of coherent systems, which we call well-formedness, can therefore be
checked syntactically. In Figure 5, we give the set of rules for deriving the judgement
⊢ N : ok
of well-formedness of N . There are only two interesting rules. Firstly, (wf-g.site) says that
l[[ M |i P ]] is well-formed whenever l is trustworthy and ⊢ P : Mp . There is a subtlety here;
this not only means that P conforms to the policy Mp , but also that any digests proffered
by agents in P can also be trusted. The second relevant rule is (wf-u.site), for typing
unknown sites: here there is no need to check the resident code, as agents emigrating from
such sites will not be trusted.
Example 2.5. (Example 2.3 continued.) Let us now re-examine the system N in Example 2.3. Suppose home is trustworthy, that is Mth (home) = good. Then, if N is to be
8
D. GORLA, M. HENNESSY, AND V. SASSONE
(wf-g.site)
⊢ 0 : ok
⊢ P : Mp
⊢ l[[ M |i P ]] : ok
(wf-par)
(wf-u.site)
⊢ N1 : ok,
⊢ N2 : ok
⊢ N1 k N2 : ok
⊢ l[[ M |i P ]] : ok
(wf-empty)
l trustworthy
l not trustworthy
Figure 5: Well-formed systems
(lts-act)
a
a.P → P
(lts-repl)
α
(lts-mig)
l
go T l.P → nil
P | !P → P ′
α
!P → P ′
(lts-par)
α
P1 → P1′
α
P1 | P2 →
α
P2 | P1 →
P1′ | P2
P2 | P1′
Figure 6: A Labelled Transition System
coherent, it is necessary for each of the sites bob, alice and secure also to be trustworthy. Consequently, N cannot be well-formed. For example, to derive ⊢ N : ok it would be
necessary to derive
⊢ go T1 home.(take.Q) : Mpb
where Mpb is the entry policy of bob. But this requires the judgement
⊢ take.Q : T1
where T1 enforces Mph . Since take 6∈ Mph , this is not possible.
One can also check that the code running at alice stops the system from being wellformed. Establishing ⊢ N : ok would also require the judgement
⊢ go T1 home.(info.go T2 secure.(take.Q)) : Mpa
which in turn, eventually, requires
⊢ take.Q : T2
for some T2 such that T2 enforces Mps ; this is impossible, again because take is not in Mps .
In well-formed systems we know that entry policies have been respected. So one way of
demonstrating that our reduction strategy correctly enforces these policies is to prove that
• system well-formedness is preserved by reduction
• only legal computations take place within trustworthy sites
The first requirement is straightforward to formalize:
Theorem 2.6 (Subject Reduction). If ⊢ N : ok and N → N ′ , then ⊢ N ′ : ok.
Proof. See Appendix A.1
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
9
To formalise the second requirement we need some notion of the computations of an
agent. With this in mind, we first define a labelled transition system between agents, which
details the immediate actions an agent can perform, and the residual of those actions. The
rules for the judgements
α
P →Q
where we let α to range over Act ∪ Loc, are given in Figure 6, and are all straightforward.
These judgements are then extended to
σ
P →Q
where σ ranges over (Act ∪ Loc)∗ , in the standard manner: σ = α1 , . . . , αk , when there
α
α
exists P0 , . . . , Pk such that P = P0 →1 . . . →k Pk = P ′ . Finally, let act(σ) denote the set of
all elements of Act ∪ Loc occurring in σ.
Theorem 2.7 (Safety). Let N be a well-formed system. Then, for every trustworthy site
σ
l[[ M |i P ]] in N , P → P ′ implies that act(σ) enforces Mp .
Proof. See Appendix A.1
3. Entry Policies
The calculus of the previous section is based on a simple notion of entry policies, namely
finite sets of actions and location names. An agent conforms to such a policy T at a site
if it only executes actions in T before migrating to some location in T. However both the
syntax and the semantics of the calculus are completely parametric on policies. All that is
required of the collection of policies is
• a binary relation T1 enforces T2 between them
• a binary relation ⊢ P : T indicating that the code P conforms to the policy T.
With any collection of policies, endowed with two such relations, we can define the predicate
M ⊢kT P as in (2.1) above, and thereby get a reduction semantics for the calculus. In this
section we investigate two variations on the notion of entry policies and discuss the extent
to which we can prove that the reduction strategy correctly implements them.
3.1. Multisets as Entry Policies. The policies of the previous section only express the
legal actions agents may perform at a site. However in many situations more restrictive
policies are desirable. To clarify this point, consider the following example.
Example 3.1. Let mail serv be the site name of a mail server with the following entry
policy Mpms :
{list, send, retr, del, reset, quit}
The server accepts client agents performing requests for listing mail messages, sending/retrieving/deleting messages, resetting the mailbox and quitting. Now, consider the
system
△
N = mail serv[[ M ms |i P ms ]] k spam[[ M s |i go T mail serv.(!send) ]]
where T = {send}. According to the typechecking of Figure 4, we have that
⊢ ! send : Mpms
10
D. GORLA, M. HENNESSY, AND V. SASSONE
(tc-act)
(tc-mig)
⊢ nil : T
⊢P :T
⊢ a.P : T ∪ {a}
⊢ P : T′
⊢ go T′ l.P : T ∪ {l}
(tc-par)
(tc-repl)
⊢ P : T1
⊢ Q : T2
⊢ P | Q : T1 ∪ T2
⊢P :T
⊢ !P : T′
(tc-empty)
Tω enforces T′
Figure 7: Typechecking with policies as Multisets
However, the agent is a spamming virus and, in practical implementations, should be rejected by mail serv.
In such scenarios it would be more suitable for policies to be able to fix an upper-bound
over the number of messages sent. This can be achieved in our setting by changing policies
from sets of agent actions to multisets of actions. Consequently, predicate enforces is now
multiset inclusion.
First let us fix some notation. We can view a multiset as a set equipped with an
occurrence function, that associates a natural number to each element of the set. To model
permanent resources, we also allow the occurrence function to associate ω to an element
with an infinite number of occurrences in the multiset. Notationally, eω stands for an
element e occurring infinitely many times in a multiset. This notation is extended to sets
and multisets; for any set/multiset E, we let E ω to denote the multiset {eω : e ∈ E}.
Example 3.2. (Example 3.1 continued.) Coming back to Example 3.1, it would be sufficient to define Mpms to be {. . . , sendK , . . .} where K is a reasonable constant. In this way, an
agent can only send at most K messages in each session; if it wants to send more messages,
it has to disconnect from mail serv (i.e. leave it) and then reconnect again (i.e. immigrate
again later on). In practice, this would prevent major spamming attacks, because the time
spent for login/logout operations would radically slow down the spam propagation.
The theory presented in Sections 2.2 and 2.3 can be adapted to the case where policies
are multisets of actions. The judgment ⊢ P : T is redefined in Figure 7, where operator ∪
stands for multiset union. The key rules are (tc-act), (tc-par) and (tc-repl). The first
two properly decrease the type satisfied when typechecking sub-agents. The third one is
needed because recursive agents can be, in general, freely unfolded; hence, the actions they
intend to locally perform can be iterated arbitrarily many times. For instance, agent
△
P = ! send
△
satisfies policy T = {sendω }. Notice that the new policy satisfaction judgement prevents
the spamming virus of Example 3.1 from typechecking against the policy of mail serv
defined in Example 3.2.
The analysis of the previous section can also be repeated here but an appropriate
notion of well-formed system is more difficult to formulate. The basic problem stems from
the difference between entry policies and resident policies. The fact that all agents who
have ever entered a site l respects an entry policy Mp gives no guarantees as to whether the
joint effect with the code currently occupying the site l also satisfies Mp . For instance, in the
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
11
terms of Example 3.2, mail serv ensures that each incoming agent can only send at most K
messages. Nevertheless, two such agents, having gained entry and now running concurrently
at mail serv, can legally send – jointly – up to 2K messages. It is therefore necessary to
formulate well-formedness in terms of the individual threads of the code currently executing
at a site. Let us say P is a thread if it is not of the form P1 | P2 . Note that every agent P can
be written in the form of P1 | . . . |Pn , n ≥ 1, where each Pi is a thread. So the well-formedness
judgment is modified by replacing rule (wf-g.site) in Figure 5 as below.
(wf-g.sitem )
∀i . (Pi a thread and ⊢ Pi : Mp )
⊢ l[[ M |i P1 | . . . |Pn ]] : ok
l trustworthy
Theorem 3.3 (Subject Reduction for multiset policies). If ⊢ N :
⊢ N ′ : ok.
ok
and N → N ′ , then
Proof. Similar to that of Theorem 2.6. The necessary changes are outlined in Appendix A.2.
The statement of safety must be changed to reflect the focus on individual threads
rather than agents. Moreover, we must keep into account also multiple occurrences of
actions in a trace; thus, we let act(σ) return a multiset formed by all the actions occurring
in σ.
Theorem 3.4 (Safety for multiset policies). Let N be a well-formed system. Then, for
σ
every trustworthy site l[[ M |i P1 | . . . |Pn ]] in N , where each Pi is a thread, Pi → Pi′ implies
that act(σ) enforces Mp .
Proof. See Appendix A.2.
3.2. Finite Automata as Entry Policies. A second limitation of the setting presented
in Section 2 is that policies will sometimes need to prescribe a precise order for executing
legal actions. This is very common in client/server interactions, where a precise protocol
(i.e. a pattern of message exchange) must be respected. To this end we define policies as
deterministic finite automata (DFAs, for short).
Example 3.5. Let us consider Example 3.1 again. Usually, mail servers requires a preliminary authentication phase to give access to mail services. To express this fact, we could
implement the entry policy of mail serv, Mpms , to be the automaton associated to the
regular expression below.
usr.pwd.(list + send + retr + del + reset)∗ .quit
The server accepts client requests only upon authentication, via a username/password mechanism. Moreover, the policy imposes that each session is regularly committed by requiring
that each sequence of actions is terminated by quit. This could be needed to save the
status of the transaction and avoid inconsistencies.
We now give the formal definitions needed to adapt the theory developed in Section 2.
We start by defining a DFA, the language associated to it, the enforces predicate between
DFAs and a way for an agent to satisfy a DFA. As usual [HU79], a DFA is a quintuple
△
A = (S, Σ, s0 , F, δ) where
12
D. GORLA, M. HENNESSY, AND V. SASSONE
• S is a finite set of states;
• Σ is the input alphabet;
• s0 ∈ S is a reserved state, called the starting state;
• ∅ ⊂ F ⊆ S is the set of final states (also called accepting states);
• δ : S × Σ → S is the transition relation.
In our framework, the alphabet of the DFAs considered is a finite subset of Act ∪ Loc.
Moreover, for the sake of simplicity, we shall always assume that the DFAs in this paper
are minimal.
Definition 3.6 (DFA Acceptance and Enforcement). Let A be a DFA. Then
• Acps (A) contains all the σ ∈ Σ∗ such that σ leads A from state s to a final state;
• Acp(A) is defined to be Acps0 (A);
• A1 enforces A2 holds true whenever Acp(A1 ) ⊆ Acp(A2 ).
Notice that, as expected, there is an efficient way to extablish A1 enforces A2 , once
given the automata A1 and A2 (see Proposition A.2 in Appendix A.3). We now formally
describe the language associated to an agent by exploiting the notion of concurrent regular
expressions (CRE, for short) introduced in [GR92] to model concurrent processes. For our
purposes, the following subset of CRE suffices:
e
::=
ǫ
| α | e1 .e2
| e1 ⊙ e2
| e⊗
ǫ denotes the empty sequence of characters, α ranges over Act ∪ Loc, ‘.’ denotes concatenation, ⊙ is the interleaving (or shuffle) operator and ⊗ is its closure. Intuitively, if e
represents the language L, then e⊗ represents {ǫ} ∪ L ∪ L⊙L ∪ L⊙L⊙L . . .. Given a CRE
e, the language associated to it, written lang(e), can be easily defined; a formal definition
is recalled in Appendix A.3. Now, given a process P , we easily define a CRE associated to
it. Formally
CRE(nil)
=
△
ǫ
CRE(a.P )
=
△
a.CRE(P )
CRE(go A l.P )
=
△
l
CRE(P1 | P2 )
=
△
CRE(P1 ) ⊙ CRE(P2 )
CRE(!P )
△
CRE(P )⊗
=
Definition 3.7 (DFA Satisfaction). An agent P satisfies the DFA A, written ⊢ P : A, if
lang(CRE(P )) ⊆ Acp(A), and ⊢ Q : A′ holds for every subagent of P of the form go A′ l.Q.
In Proposition A.2, we prove that DFA satisfaction is decidable, although extremely
hard to establish. This substantiate our hypothesis that verifying digests is preferable to
inspecting the full code from the point of view computational complexity. We are now
ready to state the soundness of this variation. It simply consists in finding a proper notion
of well-formed systems. As in Section 3.1, the entry policy can only express properties
of single threads, instead of coalitions of threads hosted at a site. Thus, we modifiy rule
(wf-g.site) from Figure 5 to:
(wf-g.siteA )
∀i . Pi a thread and ∃s ∈ S . lang(CRE(Pi )) ⊆ Acps (Mp )
⊢ l[[ M |i P1 | . . . |Pn ]] : ok
l trustworthy
This essentially requires that the languages associated to each of the threads in l are suffixes
of words accepted by Mp (cf. Theorem 3.9 below). Since this may appear quite weak, it is
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
13
worth remarking that the well-formedness predicate is just a ‘consistency’ check, a way to
express that the agent is in a state from where it will respect the policy of l. The soundness
theorems are reported below and are proved in Appendix A.3.
Theorem 3.8 (Subject Reduction for automata policies). If ⊢ N :
⊢ N ′ : ok.
ok
and N → N ′ , then
Theorem 3.9 (Safety for automata policies). Let N be a well-formed system. Then, for
every trustworthy site l[[M |iP1 | . . . |Pn ]] in N , where each Pi is a thread, σ ∈ lang(CRE(Pi ))
implies that there exists some σ ′ ∈ Acp(Mp ) such that σ ′ = σ ′′ σ, for some σ ′′ .
We conclude this section with two interesting properties enforceable by using automata.
Example 3.10 (Lock/Unlock). We have two actions, lock and unlock, with the constraint
that each lock must be always followed by an unlock. Let Σl = Σ − {lock} and Σu =
Σ−{lock, unlock}. Thus, the desired policy (written using a regular expression formalism)
is
(Σ∗l .(lock.Σ∗u .unlock)∗ )∗
Example 3.11 (Secrecy). Let secret be a secret action; we require that, whenever an
agent performs secret, it cannot migrate anymore (this policy enforces that agents having
performed secret always remain co-located). Let Σs = Σ − {secret} and Σg = Σ − Loc;
thus, the desired policy is
Σ∗s .(ǫ + secret.Σ∗g )
4. Resident Policies
Here we change the intended interpretation of policies. In the previous section a policy
dictated the proposed behaviour of an agent prior to execution in a site, at the point of
entry. This implied that safety in well-formed systems was a thread-wise property (see rules
(wf-g.siteM ) and (wf-g.siteA )). Here we focus on policies which are intended to describe
the permitted (coalitional) behaviour of agents during execution at a site. Nevertheless
these resident policies are still used to determine whether a new agent is allowed access to
the site in question; entry will only be permitted if the addition of this incoming agent to
the code currently executing at the site does not violate the policy.
Let us consider an example to illustrate the difference between entry and resident
policies.
Example 4.1. Let licence serv be the site name of a server that makes available K
licences to download and install a software product. The distribution policy is based on a
queue: the first K agents landing in the site are granted the licence, the following ones are
△
denied. The policy of the server should be Mps = {get licenceK }. However if this policy
is interpreted as an entry policy, applying the theory of Section 3.1, then the system grants
at most K licences to each incoming agent. Moreover this situation continues indefinitely,
effectively handing out licences to all incoming agents.
We wish to re-interpret the policies of the previous section as resident policies and here
we outline two different schemes for enforcing such policies. For simplicity we confine our
attention to one kind of policy, that of multisets.
14
D. GORLA, M. HENNESSY, AND V. SASSONE
4.1. Static membranes.
Our first scheme is conservative in the sense that many of the concepts developed
in Section 3.1 for entry policies can be redeployed. Let us reconsider rule (r-mig) from
Figure 2. There, the membrane M l only takes into consideration the incoming code P , and
its digest T, when deciding on entry, via the predicate M l ⊢kT P . But if the membrane is to
enforce a resident policy, then it must also take into account the contribution of the code
already running in l, namely R. To do so we need a mechanism for joining policies, such as
those of the incoming P and the resident R in rule (r-mig). So let us assume that the set
of policies, with the relation enforces is a partial order in which every pair of elements T1
and T2 has a least upper bound, denoted T1 ⊔ T2 . For multiset policies this is the case as ⊔
is simply multiset union. In addition we need to be able to calculate the (minimal) policy
which a process R satisfies; let us denote this as pol(R). For multiset policies we can adjust
the rules in Figure 7, essentially by eliminating weakening, to perform this calculation; the
resulting rules are given in Figure 8, with judgements of the form P : T.
Lemma 4.2.
• For every P , there is at most one T such that P : T.
• ⊢ P : T implies that there exists some policy T′ such that T′ enforces T and
P : T′ .
Proof. The first statement is proved by structural induction on P ; the second by induction
on the derivation ⊢ P : T.
Definition 4.3. Define the partial function pol(·) by letting pol(P ) be the unique policy
such that P : T, if it exists.
With these extra concepts we can now change rule (r-mig) in Figure 2 to take the
current resident code into account. It is sufficient to change the side condition, from M l ⊢kT
P to M l , R ⊢kT P , where this latter is defined to be
if Mtl (k) = good then (T⊔ pol(R)) enforces Mpl else ⊢ P | R : Mpl
Here if only the digest needs to be checked then we compare T⊔ pol(R), that is the result
of adding the digest to the policy of the resident code R, against the resident policy Mpl .
On the other hand if the source site is untrusted we then need to analyse the incoming
code in parallel with the resident code R. It should be clear that the theory developed in
Section 3.1 is readily adapted to this revised reduction semantics. In particular the Subject
Reduction and Safety theorems remain true; we spare the reader the details. However it
should also be clear that this approach to enforcing resident policies has serious practical
drawbacks. An implementation would need to:
(1) freeze and retrieve the current content of the site, namely the agent R;
(2) calculate the minimal policy satisfied by R to be merged with P ’s digest in order to
check the predicate enforces , or typecheck the composed agent P | R;
(3) reactivate R and, according to the result of the checking phase, activate P .
Even if the language were equipped with a ‘passivation’ operator, as in [SS03], the overall
operation would still be computationally very intensive. Consequently we suggest below
another approach.
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
(ti-act)
(ti-empty)
nil : ∅
(ti-repl)
P :T
a.P : T ∪ {a}
15
(ti-mig)
P : T′
T′ enforces T
go T l.P : {l}
(ti-par)
P :T
!P : Tω
P : T1
Q : T2
P | Q : T1 ∪ T2
Figure 8: Type inference for agents with policies as multisets
4.2. Dynamic membranes.
In the previous approach we have to repeatedly calculate the policy of the current resident code each time a new agent requests entry. Here we allow the policy in the membrane
to “decrease,” in order to reflect the resources already allocated to the resident code. So at
any particular moment in time the policy currently in the membrane records what resources
remain, for any future agents who may wish to enter; with the entry of each agent there
is a corresponding decrease in the membrane’s policy. Formally we need to change the
migration rule rule (r-mig) to one which not only checks incoming code, or digest, against
the membrane’s policy, but also updates the membrane:
(r-mig′ )
k[[ M k |i go T l.P | Q ]] k l[[ M l |i R ]] →
cl |i P | R ]]
k[[ M k |i Q ]] k l[[ M
cl
if M l ⊢kT P ≻ M
cl is defined as
where the judgement M l ⊢kT P ≻ M
T
if Mtl (k) = good
′
let T =
in T′ enforces
pol(P ) otherwise
b l ⊔ T′ ∧ M l = M
cl
Mpl ∧ Mpl = M
t
t
p
First notice that if this migration occurs then the membrane at the target site changes, from
b l . The latter is obtained from the former by eliminating those resources allocated
Mpl to M
p
to the incoming code P . If the source site, k, is deemed to be good this is calculated via
the incoming digest T; otherwise a direct analysis of the code P is required, to calculate
pol(P ).
This revised schema is more reasonable from an implementation point of view, but its
soundness is more difficult to formalise and prove. As a computation proceeds no permanent
record is kept in the system of the original resident policies at the individual sites. Therefore
well-formedness can only be defined relative to an external record of what the resident
policies were, when the system was initiated. For this purpose we use a function Θ, mapping
trustworthy sites to policies; it is sufficient to record the original polices at these sites as we
are not interested in the behaviour elsewhere.
Then we can define the notion of well-formed systems, relative to such a Θ; this is
written as Θ ⊢ N : ok and the formal definition is given in Table 9. The crucial rule is
(wf-g.site), for trustworthy sites. If l is such a site then l[[ M |i P ]] is well-formed relative
to the original record Θ if Mpl ⊔ pol(P ) guarantees the original resident policy at l, namely
Θ(l).
16
D. GORLA, M. HENNESSY, AND V. SASSONE
(wf-g.site)
l trustworthy
(pol(P ) ⊔ Mp ) enforces Θ(l)
Θ ⊢ l[[ M |i P ]] : ok
(wf-u.site)
Θ ⊢ l[[ M |i P ]] : ok
(wf-empty)
Θ ⊢ 0 : ok
(wf-par)
l not trustworthy
Θ ⊢ N1 : ok,
Θ ⊢ N2 : ok
Θ ⊢ N1 k N2 : ok
Figure 9: Well-formed systems under Θ
Theorem 4.4 (Subject Reduction for resident policies). If Θ ⊢ N : ok and N → N ′ , then
Θ ⊢ N ′ : ok.
Proof. Outlined in Appendix A.4.
The introduction of these external records of the original resident policies also enables
us to give a Safety result.
Theorem 4.5 (Safety for resident policies). Let N be a well-formed system w.r.t. Θ. Then,
σ
for every trustworthy site l[[ M |i P ]] in N , P → P ′ implies that act(σ) enforces Θ(l).
Proof. See Appendix A.4.
5. Conclusion and Related Work
We have presented a framework to describe distributed computations of systems involving migrating agents. The activity of agents entering/running in ‘good’ sites is constrained
by a membrane that implements the layer dedicated to the security of the site. We have
described how membranes can enforce several interesting kind of policies. The basic theory
presented for the simpler case has been refined and tuned throughout the paper to increase
the expressiveness of the framework. Clearly, any other kind of behavioural specification of
an agent can be considered a policy. For example, a promising direction could be considering
logical frameworks (by exploiting model checking or proof checkers).
The calculus we have presented is very basic: it is even simpler than CCS [Mil82], as
no synchronization can occur. Clearly, we did not aim at Turing-completeness, but at a
very basic framework in which to focus on the rôle of membranes. We conjecture that, by
suitably advancing the theory presented here, all the ideas can be lifted to more complex
calculi (including, e.g., synchronization, value passing and/or name restriction).
Related Work. In the last decade, several calculi for distributed systems with code
mobility have appeared in literature. In particular, structuring a system as a (flat or
hierarchical) collection of named sites introduced the possibility of dealing with sophisticated
concrete features. For example, sites can be considered as the unity of failure [FG+96,
Ama00], mobility [FG+96, CG00] or access control [HR02, RH03, GP03]. The present
work can be seen as a contribution to the last research line.
As in [GP03], we have presented a scenario where membranes can evolve. However,
the membranes presented in Section 4 only describe ‘what is left’ in the site. On the other
hand, the (dynamically evolving) type of a site in [GP03] always constrains the overall
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
17
behaviour of agents in the site and it is modified upon acquisition/loss of privileges through
computations.
We borrowed from [RH03] the notion of trust between sites, where agents coming from
trusted sites are accepted without any control. Here, we relaxed this choice by examining
the digest of agents coming from trusted sites. Moreover, we have a fixed net of trust; we
believe that, once communication is added to our basic framework, the richer scenario of
[RH03] (where the partial knowledge of a site can evolve during its computation) can be
recovered.
A related paper is [IK01], where authors develop a generic type system that can be
smoothly instantiated to enforce several properties of the π–calculus (dealing with arity
mismatch in communications, deadlock, race control and linearity). They work with one
kind of type, and modify the subtyping relation in order to yield several relevant notions of
safety. The main difference with our approach is that we have different kind of types (and,
thus, different type checking mechanisms) for all the variations we propose. It would be
nice to lift our work to a more general framework closer to theirs; we leave this for future
work.
Our work is also related to [NR05]. Policies are described there as deterministic finite
automata and constrain the access to critical sections in a concurrent functional language.
A type and effect system is provided that guarantees adherence of systems to the policy. In
particular, the sequential behaviour of each thread is guaranteed to respect the policy, and
the interleavings of the threads’ locks to be safe. But unlike our paper [NR05] has no code
migration, and no explicit distribution; thus, only one centralised policy is used.
Membranes as filters between the computing body of a site and the external environment are also considered in [FMP04, Bou04, SS03]. There, membranes are computationally
capable objects, and can be considered as a kind of process. They can evolve and communicate both with the outer and with the inner part of the associated node, in order to
regulate the life of the node. This differs from our conception of membranes as simple tools
for the verification of incoming agents.
To conclude, we remark that our understanding of membranes is radically different from
the concept of policies in [ES99]. Indeed, in loc. cit., security automata control the execution
of agents running in a site by in-lined monitoring. This technique consists of accepting
incoming code unconditionally, but blocking at runtime those actions not abiding the site
policy. Clearly, in order to implement the strategy, the execution of each action must be
filtered by the policy. This contrasts with our approach, where membranes are ‘containers’
that regulate the interactions between sites and their environments. The computation
taking place within the site is out of the control of the membrane, which therefore cannot
rely on in-lined monitoring.
Acknowledgement
The authors wish to acknowledge the reviewers of this paper for their positive attitude
and for their fruitful comments. Joanna Jedrzejowicz kindly answered some questions on
regular languages with interleaving and iterated interleaving.
References
[Ama00] R. Amadio. On modelling mobility. Theoretical Computer Science, 240(1):147–176, 2000.
18
D. GORLA, M. HENNESSY, AND V. SASSONE
[Bou04] G. Boudol. A generic membrane model. In Proc. of Global Computing, volume 3267 of LNCS, pages
208-222. Springer, 2004.
[Bou98] Z. Bouziane. A primitive recursive algorithm for the general Petri net reachability problem. In Proc.
of FOCS’98, pages 130–136. IEEE, 1998.
[CG00] L. Cardelli and A. D. Gordon. Mobile ambients. Theoretical Computer Science, 240(1):177–213,
2000.
[ES99] U. Erlingsson and F. Schneider. SASI Enforcement of Security Policies: A Retrospective. In Proc. of
New Security Paradigms Workshop, pages 87–95. ACM, 1999.
[FMP04] G. Ferrari, E. Moggi, and R. Pugliese. MetaKlaim: a type safe multi-stage language for global
computing. Mathematical Structures in Computer Science, 14(3):367-395, 2004.
[FG+96] C. Fournet, G. Gonthier, J. Lévy, L. Maranget, and D. Rémy. A calculus of mobile agents. In Proc.
of CONCUR’96, volume 1119 of LNCS, pages 406–421. Springer, 1996.
[GR92] V. Garg and M. Raghunath. Concurrent regular expressions and their replationship to Petri nets.
Theoretical Computer Science, 96:285–304, 1992.
[GHS04] D. Gorla and M. Hennessy and V .Sassone. Security Policies as Membranes in Systems for Global
Computing. In Proc. of FGUC’04, ENTCS. Elsevier, 2004.
[GP03] D. Gorla and R. Pugliese. Resource access and mobility control with dynamic privileges acquisition.
In Proc. of ICALP’03, volume 2719 of LNCS, pages 119–132. Springer-Verlag, 2003.
[HR02] M. Hennessy and J. Riely. Resource Access Control in Systems of Mobile Agents. Information and
Computation, 173:82–120, 2002.
[HU79] J. Hopcroft and J. Ullman. Introduction to automata theory, languages and computation. AddisonWesley, 1979.
[IK01] A. Igarashi and N. Kobayashi. A generic type system for the pi-calculus. In Proceedings of POPL
’01, pages 128–141. ACM, 2001.
[May84] E. Mayr. An algorithm for the general Petri net reachability problem. SIAM Journal of Computing,
13(3):441–460, 1984.
[Mil82] R. Milner. A Calculus for Communicating Systems. Springer-Verlag, 1982.
[Mil99] R. Milner. Communicating and Mobile Systems: the π-Calculus. Cambridge University Press, 1999.
[NR05] N. Nguyen and J. Rathke. Typed Static Analysis for Concurrent, Policy-Based, Resource Access
Control. Draft, 2005.
[Pet81] J. Peterson. Petri Net Theory and Modeling of Systems. Prentice Hall, 1981.
[RH03] J. Riely and M. Hennessy. Trust and partial typing in open systems of mobile agents. Journal of
Automated Reasoning, 31:335-370, 2003.
[SS03] A. Schmitt and J. Stefani. The M-calculus: a higher-order distributed process calculus. In Proc. of
POPL’03, pages 50–61. ACM, 2003.
Appendix A. Technical Proofs
We now outline the proofs of the technical results in the paper, section by section.
A.1. Proofs of Section 2.
Lemma A.1 (Subsumption). If ⊢ P : T and T enforces T′ , then ⊢ P : T′ .
Proof. By induction on the derivation of the judgment ⊢ P : T.
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
19
Proof of Theorem 2.6 [Subject Reduction]: The proof is by induction over the inference of N → N ′ . Notice that trustworthiness is invariant under reduction. Therefore
coherence, which is defined in terms of the trustworthiness of sites, is also preserved by
reduction.
We outline the proof when the inference is deduced using rule (r-mig), a typical example. By hypothesis, ⊢ k[[ M k |i go T l.P | Q ]] : ok; this implies that ⊢ k[[ M k |i Q ]] : ok. Thus,
we only need to prove that ⊢ l[[ M l |i R ]] : ok and M l ⊢kT P imply ⊢ l[[ M l |i P | R ]] : ok. We
have two possible situations:
l trustworthy: Judgment ⊢ R : Mpl holds by hypothesis; judgment ⊢ P : Mpl is implied
by M l ⊢kT P . Indeed, because of the coherence hypothesis, Mtl (k) <: Mtk (k). If Mtk (k) 6=
good, then M l ⊢kT P is exactly the required ⊢ P : Mpl . Otherwise, we know that ⊢ go T l.P :
Mpk ; by rule (tc-mig) this implies that ⊢ P : T. Judgment ⊢ P : Mpl is obtained by using
Lemma A.1, since M l ⊢kT P is defined to be T enforces Mpl (see (2.1) in Section 2.2).
Thus, by using (tc-par), we obtain the desired ⊢ P |R : Mpl .
l not trustworthy: This case is simple, because rule (wf-u.site) always allows to derive
⊢ l[[ M l |i P | R ]] : ok.
The case when (r-act) is used is similar, although simpler, and the case when rule (r-par) is
used requires a simple inductive argument. Finally to prove the case when rule (r-struct)
is used, we need to know that coherency of systems is preserved by structual equivalence;
the proof of this fact, which is straightforward, is left to the reader.
σ
Proof of Theorem 2.7 [Safety]: Let l[[ M |i P ]] be a site in N such that P → P ′ . We
have to prove that act(σ) enforces Mp . The statement is proved by induction over the
length of σ. The base case, when σ = ǫ, is trivial since act(ǫ) = ∅.
α
σ′
α
So we may assume σ = ασ ′ and P → P ′′ → P ′ . Let us consider P → P ′′ ; by induction
α
on → , we can prove that α ∈ Mp and that ⊢ l[[ M |i P ′′ ]] : ok. If the transition has been
inferred by using rule (lts-act), then P = a.P ′′ and, by rule (wf-g.site), we have that ⊢
a.P ′′ : Mp ; by definition of rule (tc-act), we have the desired a ∈ Mp and ⊢ P ′′ : Mp . When
(lts-mig) is used the argument is similar, and all other cases follow in a straightforward
manner by induction.
σ′
Thus, we can now apply induction on the number of actions performed in P ′′ → P ′
and obtain that act(σ ′ ) enforces Mp . This sufficies to conclude that act(σ) = (act(σ ′ ) ∪
{α}) enforces Mp .
A.2. Proofs of Section 3.1.
The proofs given in Appendix A.1 can be easily adapted to the setting in which entry
policies are multisets. We outline only the main changes. First recall that enforces
is multiset inclusion, that judgments ⊢ P : T must be now inferred by using the rules in
Figure 7 and that rule (wf-g.siteM ) is used for well-formedness. Then, Lemma A.1 remains
true in this revised setting.
20
D. GORLA, M. HENNESSY, AND V. SASSONE
Proof of Theorem 3.3 [Subject Reduction]: A straightforward adaptation of the corresponding proof in the previous section. The only significant change is to the case when a
replication is unfolded via the rule (r-struct), i.e.
△
N = l[[ M |i !P | Q ]] ≡ l[[ M |i P | !P | Q ]] → N ′′ ≡ N ′
By hypothesis, ⊢ !P : Mp ; therefore, by definition of rule (tc-repl), we have that ⊢ P : T
for some T such that Tω enforces Mp . Since T enforces Tω and because of Lemma A.1,
we have that ⊢ l[[ M |i P | !P | Q ]] : ok. By induction, ⊢ N ′′ : ok. It is easy to prove that
this sufficies to obtain the desired ⊢ N ′ : ok.
Proof of Theorem 3.4 [Safety]: From the rule (wf-g.siteM ) we know that ⊢ Pi : Mp ,
for all i = 1, . . . , n. We now proceed by induction over |σ|. The base case is trivial. For
α
σ′
α
the inductive case, we consider σ = ασ ′ and Pi → Pi′′ → Pi′ . By induction on → , we can
prove that α ∈ Mp and that ⊢ l[[ Mt ; Mp − {α} |i Pi′′ ]]. If the transition has been inferred by
using rule (lts-act), then Pi = a.Pi′′ and, by rule (wf-g.siteM ), we have that ⊢ a.Pi′′ : Mp ;
by definition of rule (tc-act), we have the desired Mp = T ∪ {a} and ⊢ P ′′ : T. When
(lts-mig) is used the case is simpler, and all other cases follow in a straightforward manner
by induction.
Coming back to the main claim, we use the induction and obtain that act(σ ′ ) enforces
Mp − {α}; thus, act(σ) enforces Mp .
A.3. Proofs of Section 3.2.
We start by recalling from [GR92] the formal definition of the language associated to a
CRE, as follows.
lang(ǫ)
=
△
{ǫ}
lang(α)
△
=
{α}
lang(e1 .e2 )
=
△
{x1 x2 : x1 ∈ lang(e1 ) ∧ x2 ∈ lang(e2 )}
lang(e1 ⊙ e2 )
△
=
lang(e⊗ )
=
{x1 y1 · · · xn yn : x1 · · · xn ∈ lang(e1 ) ∧ y1 · · · yn ∈ lang(e2 )}
S
{ǫ}
if i = 0
△
⊗i where L⊗i =
lang(e)
i≥0
L⊗i−1 ⊙ L otherwise
△
Notice that the definition of the lang(e1 ⊙ e2 ) hides a trick: the xi s and the yi s can also be
ǫ. Thus, as expected, we can also consider for interleaving strings of different length.
We start by accounting on the complexity of predicate enforces and the satisfiability
relation when policies are automata. This is stated by the following Proposition.
Proposition A.2.
(1) A1 enforces A2 can be calculated in polynomial time
(2) ⊢ P : A is decidable, but it is super-exponential
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
21
Proof.
(1) Let Ai = (Si , Σ, si0 , Fi , δi ) and let Li = Acp(Ai ). By definition, we have to check whether
L1 ⊆ L2 or not. This is equivalent to check whether L1 ∩ L2 = ∅. The following steps
have been carried out by following [HU79].
(a) calculate the automaton associated to L2 . This can be done in O(|S2 |) and the
resulting automaton has |S2 | states.
(b) calculate the automaton associated to L1 ∩ L2 . This can be done in O(|S1 | × |S2 | ×
|Σ|) and creates an automaton A with |S1 | × |S2 | states.
(c) Checking the emptyness of L1 ∩ L2 can be done by using a breath-first search that
starts from the starting state of (the graph underlying) A and stops whenever a
final state is reached. If no final state is reached, L1 ∩ L2 is empty. This can be
done in O(|S1 | × |S2 | × |Σ|).
Thus, the overall complexity is O(|S1 | × |S2 | × |Σ|).
(2) It has been proved in [GR92] that each CRE e can be represented by a (labelled) Petri
net, in that the language accepted by the Petri net is lang(e). Now, we can easily
construct a DFA accepting the complement of the language accepted by A (see item (a)
of the previous proof). Now, we can construct the product between this DFA (that can
be seen as a Petri net) and the Petri net associated to CRE(P ); this Petri net accepts
lang(CRE(P )) ∩ Acp(A) (see [Pet81]). Now, the emptyness of this language can be
solved with the algorithm for the reachability problem in the corresponding Petri net.
This problem has been proved decidable [May84] and solvable in double-exponential
time [Bou98].
We now prove the subject reduction theorem in the setting where types are DFAs. To
this aim, we need to adapt Lemma A.1 and we need a very simple result on the languages
associated to DFAs and processes.
Lemma A.3. If ⊢ P : A and A enforces A′ , then ⊢ P : A′ .
Proof. By transitivity of subset inclusion.
Lemma A.4.
(1) ασ ∈ Acps (A) if and only if σ ∈ Acpδ(s,α) (A)
(2) If σ ∈ lang(CRE(a.P )) then σ = aσ ′ for σ ′ ∈ lang(CRE(P )). Viceversa, if σ ∈
lang(CRE(P )), then aσ ∈ lang(CRE(a.P )).
Proof. Trivial.
Proof of Theorem 3.3 [Subject Reduction]: Now ⊢ N : ok relies on rule (wf-g.siteA ).
Again, the proof is by induction on the inference of N → N ′ . We only give the base cases,
because inductive steps can be handled with in a standard way. We only consider the cases
of trustworthy sites; the case for non-trustworthy sites is easier. In what follows, we write
⊢s P : A to mean that ⊢ P : A′ , where A′ is the DFA obtained from A by setting s as
starting state.
(r- act) In this case, N = l[[ M |i a.P | Q ]]. By definition of rule (wf-g.siteA ), it holds
that Q = Q1 | . . . |Qk (for Qi threads), ∃s. ⊢s a.P : Mp and ∀i.∃si . ⊢si Qi : Mp .
By definition, we have that lang(CRE(a.P )) ⊆ Acps (Mp ); by Lemma A.4, we have
that lang(CRE(P )) ⊆ Acpδ(s,a) (Mp ). This sufficies to infer the well-formedness of
N ′ = l[[ M |i P | Q ]].
22
D. GORLA, M. HENNESSY, AND V. SASSONE
(r- mig) In this case, N = k[[ M k |i go A l.P | Q ]] k l[[ M l |i R ]] and M l ⊢kA P . We further
identify two sub-cases:
• M l (k) = good. In this case, because of coherence, we know that ⊢ P : A. Moreover,
by definition of M l ⊢kA P , it holds that A enforces Mpl . By Lemma A.3, we have
that ⊢ P : Mpl . This sufficies to conclude.
• M l (k) 6= good. This case is simpler because M l ⊢kA P is defined to be ⊢ P : Mpl .
Proof of Theorem 3.9 [Safety]: The proof is quite easy. Indeed, by rule (wf-g.siteA ),
it holds that ∃si : ⊢si Pi : Mpl . By definition, this implies that every σ ∈ lang(CRE(Pi )) is
in Acpsi (Mpl ). Since the automaton Mpl is minimal, si is a reachable state from the starting
state s0 , say, with a (finite) string σ ′′ . Then, by Definition 3.6 and by Lemma A.4.1, it
holds that σ ′′ σ ∈ Acp(Mpl ). This proves the thesis.
A.4. Proofs of Section 4.
We show here the main things to modify to carry out the proofs given in Appendix A.2.
Obviously, judgment ⊢ P : T must be now replaced everywhere with P : T and, similarly,
⊢ N : ok becomes Θ ⊢ N : ok.
Proof of Theorem 4.4 [Subject Reduction]: The proof is by induction over the inference of N → N ′ . Inductive steps are simple; we only give the base steps.
(r- act) By hypothesis, Θ ⊢ l[[ M |i a.P | Q ]] : ok. If l is not trustworthy, the case is trivial.
Otherwise, we know by hypothesis that (pol(a.P | Q) ⊔ Mp ) enforces Θ(l). Now, by
definition of judgment
(and hence of function pol(·)) we have that pol(a.P | Q) =
pol(P | Q) ∪ {a}. Hence, (pol(P | Q) ⊔ Mp ) enforces Θ(l), as required.
(r- mig) By hypothesis, Θ ⊢ l[[ M l |i R ]] : ok; we only consider the case in which l is
trustworthy. Thus, we know that (pol(R) ⊔ Mpl ) enforces Θ(l). By the premise of rule
cl . We have two possible situations:
(r-mig), it holds that M l ⊢kT P ≻ M
cl is defined to be T enforces Mpl ∧ Mpl =
Mtl (k) = good: In this case, M l ⊢kT P ≻ M
b l ⊔ T ∧ Ml = M
cl is sufficient to preserve coherence.
cl . The fact that M l = M
M
t
t
t
t
p
Moreover, by rule (ti-mig), we know that P : T′ and T′ enforces T. By rule
(ti-par), pol(P |R) = pol(R) ⊔ T′ and (pol(R) ⊔ T′ ) enforces (pol(R) ⊔ T). Then,
b l = (pol(R) ⊔ T′ ⊔ M
b l ) enforces (pol(R) ⊔ T ⊔ M
b l ) = (pol(R) ⊔
pol(P |R) ⊔ M
p
p
p
Mpl ) enforces Θ(l), as required.
Mtl (k) 6= good: In this case, the previous proof should be rephrased by using pol(P )
instead of the digest T.
Proof of Theorem 4.5 [Safety]: We prove a slightly more general result, that easily
implies the claim desired.
Let N be a well-formed system w.r.t. Θ. If l[[M |iP ]] is a trustworthy site of N
σ
such that (pol(P ) ⊔ Mpl ) = T, then P → P ′ implies that act(σ) enforces T.
The proof is by induction over |σ|. The base case is when σ = ǫ and it is trivial. In the
α
σ′
inductive case, we consider σ = ασ ′ and P → P ′′ → P ′ . To start, it is easy to prove that
pol(P ′′ ) ⊔ {α} = pol(P )
(A.1)
SECURITY POLICIES AS MEMBRANES IN SYSTEMS FOR GLOBAL COMPUTING
23
By transitivity of multiset inclusion and by the claim (A.1) above, (pol(P ′′ ) ⊔ Mpl ) = T′ ,
where T = T′ ⊔ {α}. Thus, node l[[ M l |i P ′′ ]] is well-formed (and trustworthy). By induction
we therefore have that act(σ ′ ) enforces T′ . Hence, act(σ) = act(σ ′ ) ⊔ {α} enforces T′ ⊔
{α} = T, as required.
To conclude, the original claim of Theorem 4.5 is obtained from the result just proved
by noticing that, because of well-formedness, T enforces Θ(l).
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
| 6cs.PL
|
1
Real-time Error Control for Surgical Simulation
arXiv:1610.02570v3 [cs.NA] 14 Feb 2017
Huu Phuoc Bui∗† , Satyendra Tomar† , Hadrien Courtecuisse∗ , Stéphane Cotin‡ and Stéphane Bordas†§
∗ University of Strasbourg, CNRS, ICube, F-67000 Strasbourg Fance
† University of Luxembourg, Research Unit of Engineering Science, L-1359 Luxembourg, Luxembourg
‡ Inria Nancy Grand Est, 54603 Villers-les-Nancy, France
§ Cardiff University, School of Engineering, Queens Buildings, The Parade, Cardiff CF24 3AA, Wales, UK
Abstract—Objective: To present the first real-time a posteriori error-driven adaptive finite element approach for realtime simulation and to demonstrate the method on a needle
insertion problem. Methods: We use corotational elasticity and a
frictional needle/tissue interaction model. The problem is solved
using finite elements within SOFA1 . The refinement strategy
relies upon a hexahedron-based finite element method, combined
with a posteriori error estimation driven local h-refinement, for
simulating soft tissue deformation. Results: We control the local
and global error level in the mechanical fields (e.g. displacement
or stresses) during the simulation. We show the convergence
of the algorithm on academic examples, and demonstrate its
practical usability on a percutaneous procedure involving needle
insertion in a liver. For the latter case, we compare the force
displacement curves obtained from the proposed adaptive algorithm with that obtained from a uniform refinement approach.
Conclusions: Error control guarantees that a tolerable error level
is not exceeded during the simulations. Local mesh refinement
accelerates simulations. Significance: Our work provides a first
step to discriminate between discretization error and modeling
error by providing a robust quantification of discretization error
during simulations.
Index Terms—Finite element method, real-time error estimate,
adaptive refinement, constraint-based interaction.
I. I NTRODUCTION
A. Motivation
R
eal-time simulations are becoming increasingly common
for various applications, from geometric design [1], [2]
to medical simulation [3]. Our focus is on real-time simulation
of the interaction of a surgeon or interventional radiologist
with deformable organs. Such simulations are useful to both,
help surgeons train, rehearse complex operations or/and to
guide them during the intervention. In time, reliable simulations could also be central to robotic surgery.
A number of factors are concurrently involved in defining
the “accuracy of surgical simulators: mainly the modeling
error and the discretization error. Most work in the area has
been looking at the above sources of error as a compounded,
lumped, overall error. Little or no work has been done to
discriminate between modeling error (e.g. needle-tissue interaction, choice of constitutive models) and discretization
error (use of approximation methods like FEM). However,
it is impossible to validate the complete surgical simulation
approach and, more importantly, to understand the sources of
error without evaluating both the discretization error and the
modeling error.
1 https://www.sofa-framework.org/
The first ingredient in any mechanical simulation is the
ability to simulate the deformation of the solid of interest. This
deformable solid mechanics problem is usually solved by finite
element method (FEM) [4] or meshless/meshfree methods
[5], which are used to discretize the equilibrium equations.
It is usually uneconomical or prohibitively expensive to use
a fixed mesh for such simulations. Indeed, coarse meshes
are sufficient to reproduce “smooth” behavior, whereas “nonsmooth” behavior such as discontinuities engendered by cuts
or material interfaces, singularities, boundary layers or stress
concentrations require a finer mesh. Adequate approaches are
thus needed to refine the discretization in these areas.
Yet, existing numerical methods used in surgical simulation
use either a fixed discretization (finite element mesh, meshfree
point cloud, reduced order method), or adapt the mesh using
heuristics [6], [7]. To our knowledge, no approach is currently
able to adapt the finite element mesh based on rational a
posteriori error estimates [8].
Our objective is thus to devise a robust and fast approach
to local remeshing for surgical simulations. To ensure that the
approach can be used in clinical practice, the method should
be robust enough to deal, as realistically as possible, with the
interaction of surgical tools with the organ, and fast enough
for real-time simulations. The approach should also lead to
an improved convergence so that an “economical” mesh is
obtained at each time step. The final goal is to achieve optimal
convergence and the most economical mesh, which will be
studied in our future work.
In this paper, we propose and benchmark a local mesh
refinement and coarsening approach which is based on the
estimation of the discretization error incurred by FEM in the
solution of a corotational model representing soft tissues. The
general ideas presented here can be used directly in geometric
design based on deformable models. Our proposed approach
has similitudes with the octree approaches of [9].
B. Error in numerical simulations
It is useful to first review the various sources of error
in numerical simulations. The first error source arises when
a mathematical model is formulated for a given physical
problem: this is known as the modeling error. The second
error arises upon discretization of this mathematical model, for
example using FEM or meshfree methods. Finally, numerical
error is incurred because of the finite precision of computers
and round off errors. In this paper, we focus on the second
2
source of error, namely discretization error. We therefore
assume that the model we use is descriptive of reality, i.e.
we are solving the right problem, and we ask ourselves the
question whether “we are solving the problem right”, in other
words, correctly.
The main difficulty in answering this question comes from
the fact that an exact solution, to which the numerical solution
could be compared, is generally not available. Different approaches exist to address this problem, which are reviewed in
the literature, see e.g., [4], [8], [10]. Simple methods available
in practice to indicate the error distribution can be categorized
into two classes: recovery-based and residual-based.
The first class of indicators assumes that the exact solution
(of stresses) is smooth enough (at least locally). They rely on
the construction of an “improved” numerical solution from the
raw numerical solution, to which the raw numerical solution
can be compared. Where these two solutions are significantly
different (above certain threshold), the error level is high and
the mesh should be refined, and where these two solutions are
close together, the mesh can be kept unchanged or coarsened.
This idea was proposed by Zienkiewicz and Zhu in [11], and
its asymptotic convergence to the exact error is studied in [12],
[13].
The second class of indicators relies on the computation of
the residual of the governing equations within each computational cell, typically each finite element. These “residualbased” error indicators lead to mesh refinement where the
solution leads to large residuals, and keep the mesh constant
or coarsen it where the element residuals are relatively small,
compared to a given tolerance. These estimates were first
proposed by Babuska and Rheinboldt in [14].
Based on these local error estimators, mesh refinement
methodologies can be devised to derive mesh adaptation, see
e.g. [10] for a recent comprehensive presentation. This requires
two key ingredients: a marking strategy that decides which
elements should be refined, and a refinement rule that defines
how the elements are subdivided. For element marking, we
use the maximum strategy, see Section III for details. Other
strategies, such as bulk/equilibration strategy or percentage
strategy, see e.g. [10], can also be used.
C. Simulation of percutaneous operations
Needle-based percutaneous procedures are an important part
of modern clinical interventions such as biopsy, brachytherapy, cryotherapy or regional anesthesia. The success of these
procedures depends on good training and careful planning
to optimize the path to the target, while avoiding critical
structures [15]. In some instances the procedure can also
be assisted by robotic devices. Unfortunately, natural tissue
motion (due to breathing, for instance), and deformation (due
to needle insertion) generally lead to incorrect or inefficient
planning [15]. To address these issues, one must rely on an
accurate simulation of needle insertion. For most problems,
computational speed is also very important, since the simulation is at the core of an optimization algorithm (for the needle
path) or a robotic control loop.
The main works on needle insertion (see the survey by
[16]) propose to model the interaction between the needle
and soft tissues using FEM. In the various methods proposed
in the literature, three main research directions have been
followed: soft tissue model, flexible needle model and needletissue interactions. The needle model is usually not an issue,
both in terms of modeling choice and computational cost.
For instance, in [17], authors report computation times of a
few milliseconds for a FEM needle model composed of 50
serially-linked Timoshenko beam elements. Soft tissue models
are usually based on FEM, and rely on linear or non-linear
constitutive laws [16].
A large body of work covers the modeling and simulation
of soft tissue deformation, even under real-time computation
constraints. But overall, the interaction model between the
needle and tissue remains a major challenge. It combines different physical phenomena, such as puncturing, cutting, sliding
with friction and Poynting’s effect. To capture the essential
characteristics of these interactions, existing methods usually
rely on experimental force data and remeshing techniques
in order to align nodes of the FEM mesh with the needle
path. In [17], a constraint-based approach, avoiding remeshing,
was used to simulate needle-tissue interactions. However, the
simulations did not account for realistic anatomical details. In
addition, Misra et al. [18] showed that needle steering, which
occurs when using asymmetric needle tips, can be modeled
using microscopic observations of needle-tissue interactions.
Unfortunately, if no assumption can be made about the
region of the domain where the needle will be inserted,
simulations involving very detailed meshes become very slow,
which is a real issue in the context presented above. Errorcontrolled real-time simulation of needle insertion is thus an
unsolved problem whose solution requires tackling a number
of difficulties:
•
•
•
•
•
developing needle-tissue models. A review of cutting
simulation is provided in [19]
using these models within discrete approaches like FEM,
mesh-free methods, or others
accelerating the simulation (advanced hardware, model
order reduction)
validating the needle-tissue interaction model combined
with discrete solution (are we solving the right problem?)
verifying the discrete solution, i.e. controlling the discretization error associated with the discrete model (are
we solving the problem right?)
In this paper, we propose to focus on the last point above,
with the aim to model needle-tissue interactions using an
adaptive meshing strategy driven by simple a posteriori error
estimation techniques. Similarly to [17], we do not require
the mesh to conform to the needle path. Mesh subdivision
is only introduced as a means to improve the accuracy of
the needle-tissue interactions. Our mesh refinement method
is guided by the stress/energy error estimate resulting from
the needle-tissue interaction and imposed boundary conditions:
elements of the mesh are subdivided when a numerical error
threshold is reached. The subdivision process is completely
reversible, i.e. refined elements are set back to their initial
topology when refinement is no longer needed. Our refinement
approach does not rely on a usual octree structure (see also
3
[20]), thus allowing a variety of subdivision schemes that are
well suited for needle insertions, as detailed in Section III.
Using this approach, interactive computation times can be
achieved while detailed tissue motion near the needle shaft
or tip can be computed. This opens new possibilities for fast
simulations of flexible needle insertion in soft tissues. We
illustrate the convergence study of the adaptive refinement
scheme and some of the possible scenarios in Section V.
II. M ODEL AND DISCRETIZATION
In this section, we describe the model and the discretization
approach, which are used for needle and soft tissue interaction.
A. Problem statement
During the needle insertion, three types of constraints are
defined, see Fig. 1. Coulomb’s friction law is used to describe
t (see Fig. 1), the needle tip can cut and go through the tissue
or not
λn < µλt +λn0 (stick);
λn ≥ µλt +λn0 (cut and slip). (4)
Finally, needle shaft constraints are defined along the needle
shaft so that the needle shaft is enforced to follow the insertion
trajectory created by the advancing needle tip. Again, the
Coulomb’s friction law is applied to these constraints to
represent the stick and sliding contact between the tissue and
the needle shaft
λn < µλt (stick);
λn = µλt (sliding).
B. Strong form
We model both, the tissue and the needle, as dynamic
deformable objects. Thus, they can be regarded as dynamic
elastic solids, and the governing equations of the model are
formulated as
div σ + b + λ = ρü
in Ω
1
grad u + (grad u)T
=
2
σ = f (, ν)
σ · n = t̄ on Γt ;
u = ū on Γu ,
Fig. 1: Three types of constraints between the needle and
soft tissue: surface puncture (in red), needle tip constraint
(in green) and needle shaft constraints (in blue). A local
coordinate system n-t is defined at each constraint point.
frictional contacts within these three types of constraints. First,
a puncture constraint is defined between the needle tip and
the tissue surface. This constraint satisfies the Kuhn-Tucker
condition in the direction n (normal to the tissue surface)
δn ≥ 0,
λn ≥ 0,
δn · λn = 0,
(2)
the needle can penetrate into the tissue. In the tangent direction
t, Coulomb’s friction law is considered in order to take into
account the stick/slip between the needle tip and tissue surface
λt < µλn
(stick);
λt = µλn
(slip),
(6a)
(6b)
(6c)
(6d)
(6e)
where σ is the Cauchy stress tensor, b is the body force
vector, ρ is the mass density, is the strain tensor, ν =
(ν1 , ν2 , . . . , νn ) is the internal variables, ( ˙ ) denotes the partial
derivative with respect to time, n denotes the outward unit
normal vector on Γt , and λ denotes the contact force between
the needle and the tissue. The object domain and boundary
conditions are shown in Fig. 2a.
(1)
where δn denotes the distance between the needle tip and the
tissue surface in the direction n, and λn denotes the contact
force in that direction. Let λn0 represent the puncture strength
of the tissue. The Kuhn-Tucker condition expresses that the
contact force only exists when the needle tip is in contact
with the tissue surface. When the contact force is higher than
a threshold (puncture strength of the tissue)
λn > λn0 ,
(5)
(3)
where µ denotes the friction parameter.
Second, a needle tip constraint is defined at the tip of the
needle as soon as it penetrates into the tissue. Depending
on the relationship between the contact forces in the normal
direction n (along the needle shaft) and in the tangent direction
Fig. 2: A body Ω subjected to a traction t̄ on its boundary
part Γt , a body force b, and an imposed displacement ū
on boundary part Γu (a); Simplified illustration of FEM
discretization (b).
C. Spatial and temporal discretization
1) Space discretization: The basic idea of FEM is to discretize the domain Ω into finite elements Ωe , e = 1, 2, . . . , Ne ,
by Nn nodes, as depicted in Figure 2b. Based on the discretization concept, see e.g., [21], [22], we obtain the discrete
problem of the element e
Me üe + Ce u̇e + fe (σ) = feext ,
(7)
4
where Me is the element mass matrix, Ke is the element
stiffness matrix, Ce is the damping matrix, feext is the external
force applied to the element e, while the internal force fe (σ)
reads
Z
Z
fe (σ) =
BTe σ dΩ =
BTe EBe ue dΩ
(8)
Ωe
Ωe
= Ke · ue = Ke · (xe − x0e )
where dv = vt+τ − vt .
After solving (12) for dv, the position and velocity are
updated for needle and tissue as
where Be is the strain-displacement matrix, E is the fourthorder stiffness tensor, xe and x0e denote the current and initial
position of the element e, respectively. However, using Eq. (8)
results in inaccuracy for large rotations problems, which is observed by artificially inflated deformation of the elements. To
overcome this, Felippa et al. [23] decomposed the deformation
gradient of the element into the rigid and deformation parts,
and the element nodal internal force becomes
To take into account the needle-tissue interaction when they
are in contact, a constrained dynamic system is solved for the
needle and the tissue. Eq. (12) then becomes
fe = Re Ke (RTe xe − x0e ),
(9)
where R stands for the element rotation matrix of the element
local frame with respect to its initial orientation, being updated
at each time step. Using this corotational formulation results
in no visual artifacts.
The global mass, stiffness and damping matrices of the
system can then be assembled from the element ones, and
Eq. (7) can be rewritten to a global system equation as
Ma = f (x, v),
(10)
where a = ü, x, v = u̇ are the acceleration, position and
velocity vectors, respectively, and f (x, v) = f ext − Ku − Cv
represents the net force (the difference of the external and
internal forces) applied to the object.
In our simulations, a diagonally lumped mass matrix is
employed, and the stiffness matrix K is computed based on the
co-rotational FE formulation described above, which allows
large rotations for both, needle as well as tissue. For higher
accuracy of the computed strain field, the soft tissue domain is
discretized using hexahedral elements. To avoid the complex
issue of generating an exact hexahedral mesh of the domain,
we use a mesh that does not conform to the boundary of the
domain, as in Immersed Boundary Method [24]. The needle,
on the other hand, is modeled using serially-linked beam
elements, as in [17]. In this case, each node of the needle has
6 degrees of freedom (3 translations and 3 rotations), while
the tissue model only uses 3 translational degrees of freedom
per node.
Since the FEM formulation is based on the discretization of
the physical domain, it naturally introduces the discretization
error in the result. To control this error source, in Section III
we present an adaptive refinement scheme.
2) Time discretization: For temporal discretization, we use
an implicit backward Euler scheme [25], which is described
as follows
u̇t+τ = u̇t + τ üt+τ ;
ut+τ = ut + τ u̇t+τ ,
(11)
where τ denotes the time step. Inserting Eq. (11) into Eq. (10)
gives the final discrete system
(M − τ C − τ 2 K) dv = τ f (xt , vt ) + τ 2 Kvt
{z
}
|
{z
}
|
A
b
(12)
vt+τ = dv + vt ;
xt+τ = xt + τ vt+τ .
(13)
D. Constraint enforcement for needle-tissue interaction
Adv = b + Jλ,
(14)
where λ denotes Lagrange multipliers representing the interaction forces between the needle and the tissue, and J
provides the direction of the constraints. Different types of
constraints between needle and tissue are used, and solving
their interaction is detailed in Section IV.
Remark Combining more advanced and clinically relevant
needle-tissue interaction is straightforward in our approach.
III. E RROR ESTIMATE AND ADAPTIVE REFINEMENT
To achieve faster and more accurate FEM simulations, different adaptive techniques have been proposed in the literature.
Octree-based approaches [26] are the most common, but the
refinement procedure is limited to cubic elements which are
recursively subdivided into eight finer elements. To overcome
this limitation, more generic remeshing techniques [27]–[29]
have been proposed. However, they are complex to implement,
and may lead to ill-shaped elements. Our template-based
refinement algorithm is designed to be independent of the
type of element (tetrahedra, hexahedra, others), and produces
a high quality mesh (thanks to the well-shaped elements of
the predefined template).
Starting with an initial, relatively coarse mesh (as required
to achieve real-time simulation), a criterion based on a posteriori error estimate is evaluated to drive the local refinement.
The elements where the stress increases, i.e. dσ/dt > 0, are
considered for refinement, and the elements where the stress
decreases, i.e. dσ/dt < 0, are taken to a lower refinement
(coarsening) level. We define the approximate error of an
element Ωe as
sZ
(h − s )T (σ h − σ s )dΩ,
ηe =
(15)
Ωe
which is the energy norm of the distance between the FEM
solution (denoted by h) and an improved solution (denoted by
s) obtained by the Zienkiewicz-Zhu smoothing procedure [11].
Among the elements with increasing stress, only those elements, where the error exceeds the predefined threshold,
are subdivided (refined). Similarly, among the elements with
decreasing stress, only those elements, where the error is
smaller than the above threshold, are coarsened. Notice that
we are not limited by the regularity of the mesh and can start
from any (reasonable) heterogeneous mesh as a starting point
prior to refinement.
5
A. Zienkiewicz-Zhu error estimate
Using the superconvergent patch recovery (SPR) procedure [11], the smoothed stress field σ s is recovered from
the stresses computed at the element center. The idea of this
technique is based on the fact that the stress and strain at the
superconvergent points (at element center in the case of linear
hexahedral elements) are accurate with higher order than at
the element nodes, and these values are employed to recover
the nodal stress and strain within the least squares sense.
A 2D representation of a patch of 8 hexahedral elements is
shown in Fig. 3. For each component σjh of the FEM solution
σ h , the nodal recovered stresses are computed by defining a
polynomial interpolation within the element patch as
certain threshold, are refined. Let ηM = maxe ηe , where ηe is
defined in Eq. (15). We mark an element for refinement if
ηe ≥ θηM with 0 < θ < 1.
(20)
Other marking strategies, such as bulk/equilibration strategy,
or percentage strategy, see e.g. [10], can also be used. However, the maximum strategy described above is the cheapest
among all, and hence, it is preferred for our use. In the
maximum strategy, a large value of θ leads to small number
of elements marked for refinement, and small value of θ leads
to large number of elements marked for refinement. In our
studies presented in Section V, we set θ = 0.3.
σ̃js = Paj = [1 x y z xy yz zx xyz][a1 a2 a3 a4 a5 a6 a7 a8 ]Tj . (16)
Let Pk ≡ P(xk , yk , zk ). To determine the unknowns aj we
minimize, for 8 sampling points k of an element patch,
Π=
8
X
h
σj (xk , yk , zk ) − Pk ajk .
(17)
k=1
This minimization results in finding aj by
aj = A−1 b
(18)
where
A=
8
X
PTk Pk and b =
k=1
8
X
PTk σjh (xk , yk , zk )
(19)
k=1
Once aj is available, the nodal recovered stress values are
obtained by simply employing Eq. (16) with P evaluated at
the corresponding node. 2
C. Template-based adaptive h-refinement
Once the refinement criterion is satisfied within an element,
the element is replaced by several elements according to a
predefined template. The template is simply a set of nodes
and an associated topology, defined using an isoparametric
formulation. The template nodes are added by using their
natural coordinates. The position xj , in Cartesian coordinates,
of the new node j is defined as xj = xi ξij , where the Einstein
summation convention is applied on the nodes i of the removed
element (i = 1, 2, . . . 8 for hexahedral elements). The shape
function ξij is computed from the barycentric coordinates of
template node j with respect to the node i. The procedure is
summarized below:
1) Remove the element to be refined
2) Add template nodes and then template
elements using the element shape functions
3) Update the topology of the global mesh
4) Compute stiffness matrix of new elements
5) If needed, update the mass and damping
matrices
It is worth mentioning that if, after refinement, a new
element fulfills the refinement criterion, it can be refined again,
using the same predefined template. This results in a multiresolution mesh (see Fig. 4). Conversely, if the coarsening
criterion is satisfied in already refined elements, the coarsening procedure is applied by simply removing respective fine
elements, and updating the associated matrices.
Superconvergent sampling point
(0,1,1)
(0,1,0)
Patch assembly point
Nodal value determined from the patch
Fig. 3: The smoothed gradient is obtained from an element
patch.
η
(0,0,0)
ξ
ζ
(0,0,1)
(1,1,0)
(1,1,1)
(1,0.5,0.67)
(1,0,0)
(1,0,1)
Natural coordinates
B. Element marking strategy
After obtaining error distribution across all the elements,
we employ the maximum strategy to select those elements
which must be refined for the next level mesh. In this strategy,
only those elements, where error (see Eq. (15)) is higher than
2 Similar to the displacements, the recovered stresses σ s can also be
obtained using element shape functions σ s = Nσ̃ s .
y
x
Physical coordinates
z
Fig. 4: Adaptive subdivision process: each element to be
subdivided is topologically transformed in its reference shape,
using a template expressed in natural coordinates. Cartesian
coordinates in the mesh are computed using the element shape
functions. The process can be applied recursively, and is
completely reversible.
6
5
D. T-junction handling
Since elements are refined by using templates, regardless
of their neighboring elements, some T-junctions (incompatible
nodes or hanging nodes) are generated. To avoid discontinuities at the T-junctions during the simulation, these nodes
need to be handled in a special way, in which T-junction
nodes are considered as slave of other independent (master)
degrees of freedom (DOFs). One of the possible options is
to use Lagrange multipliers, but this approach increases the
total number of DOFs (as to solve the unknown Lagrange
multipliers in addition), and usually leads to ill-conditioned
systems. In our approach, we follow the method proposed
by [30], which considers only the reduced system (without Tjunctions) when solving for the new positions. Let T denote
the transformation matrix from the reduced system to the full
one (with T-junctions). The matrix T contains the barycentric
coordinates of the T-junctions (slaves) with respect to their
masters, and contains 1 for all other normal DOFs. The
reduced system matrix Ar is then computed from the full
system matrix Af as
Ar = TT Af T.
(21)
The nodal forces in the reduced space is computed from the
full space as fr = TT ff . The reduced system Ar dvr = fr is
solved to find dvr , the difference of velocity between current
and previous time step. Once dvr is available, the difference of
velocity in the full space are easily deduced as dvf = Tdvr .
The latter is employed to update the new position and velocity
of the object as in Eq. (13).
A heuristic example is shown in Fig. 5 to explicitly illustrate
the method for T-junction handling, especially how the full
and reduced systems are defined, resulting in the computation
of the transformation matrix between them. After subdivision,
node 7 is a T-junction in the full system (see Fig. 5a). Within
this heuristic illustration, considering that each node has only
one DOF and the static condition is applied, the displacement
of the node 7 is expressed from those of node 2 and 3 as
3
0
7
4
5
2
3
6
1
(a) Full system
4
7
2
0
6
1
(b) Reduced system
Fig. 5: Illustration of T-junction handling method on a
schematic example.
Assuming we are using a non-linear constitutive law (e.g.
hyperelastic) or a co-rotational FE formulation, the system
matrix needs to be updated at each time step. Consequently, the
local updating of the topology has a very limited impact on the
computation. The main overhead comes from the T-junction
handling, but it is somewhat compensated by the reduced
dimensions of the linear system to be solved (although the
reduced matrix is denser than the initial one). Our experience
has shown that if we consider that about 10% of the mesh
elements are subdivided, approximately 20% of the nodes in
the resulting mesh are T-junctions. Obviously, the number of
T-junctions depends strongly on the template mesh used by
the refinement, and also on the fact that the elements are
subdivided locally within one or several regions.
IV. N EEDLE - TISSUE INTERACTION ALGORITHM
(23)
To model the interaction of needle and tissue, we consider two different constraints: penetration (puncture) and
sliding [17]. To avoid remeshing when modeling needle-tissue
interaction, we use the same constraints based approached
as described in [17]. However, unlike [17], we solve the
constrained system differently. Before entering the tissue, the
needle-tissue constraint is only created when the needle tip is
in contact with the tissue surface. This penetration constraint
is represented mathematically as P(xn , xt ) ≥ 0, where xn ,
and xt stand for the position of the needle, and tissue,
respectively. Immediately after entering the tissue, sliding
constraint between the needle and tissue is created along
the needle trajectory as S(xn , xt ) = 0. When friction is
considered, both P and S are nonlinear. The constraints P
and S between the tissue (denoted by subscript 1) and needle
(denoted by subscript 2) are expressed through the global
coordinate system using Lagrange multipliers λ as follows
A1 0 JT1
dv1
b1
T
(24)
0 A2 J2 dv2 = b2 ,
λ
0
J1 J2
0
The transformation matrix T for the general 3D case where
each node has three DOFs is built straightforwardly from this
example.
where A1 and A2 are the system matrices for the soft tissue
and needle, respectively; J1 and J2 account for contraint directions between the needle and tissue. In the local coordinate
system attached to the needle, constraints between needle and
tissue are only expressed in two directions orthogonal to the
needle shaft, resulting in a needle-tissue sliding constraint.
The expressions of these constraints in the global coordinate
system J1 and J2 are then built by transforming the local
u7 = 0.5u2 + 0.5u3 .
(22)
Fig. 5b shows the reduced system where the T-junction node
7 is not considered. The displacement fields between these
full and reduced systems are expressed (by taking into account Eq. (22)) through the transformation matrix T as
1
0
u0
0
u1
0
.. =
0
.
0
u7
| {z }
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
0
uf
|
0
0.5 0.5 0
{z
T
0 0
0 0
u
0
0 0
u
1
0 0
. .
0 0
..
1 0 u6
| {z }
0 1
ur
0
0
}
7
constraint expressions from the local coordinate system to
the global one. However, formulating the problem as (24)
leads to a non positive definite global matrix, which makes
the system challenging to solve. An alternative approach, as
proposed in [15], [17], is to solve the interaction problem
in three steps: predictive motion (no interaction constraints),
constraint solving, and corrective motion. However, this alternative requires the computation of matrix inverse A−1
and
1
A−1
2 . This approach is time consuming, especially for large
systems. Unlike this method, we solve the constrained problem
iteratively by using the augmented Lagrangian method [31]
λ
k+1
k
= λ − WJdv
k+1
,
(25a)
L/2
L/2
(25b)
where W is the penalty weight matrix with finite values.
The advantage of this method is that the exact solution of
the needle-tissue interaction can be obtained, as compared to
the penalty method (see e.g. [32]), and no additional DOFs
are needed, as compared to the classical Lagrange multiplier
method [33]. A critical feature of this approach is that the
system matrix in (25a) is positive definite, therefore iterative
solvers, such as the conjugate gradient, can be used efficiently.
It is worth stressing that using the augmented Lagrangian
method for solving needle-tissue interaction, combined with
the T-junction handling in the tissue, is straightforward. Indeed, as mentioned above, it is sufficient to solve (25a) in the
reduced space, and dvk+1 is easily computed from the reduced
solution.
V. R ESULTS AND DISCUSSIONS
To demonstrate the efficiency of our method, we present
several numerical studies. We first present the convergence of
the stress error on a typical L-shaped domain. The motivation
for this test is because of its localized nature, i.e. stress is
concentrated at the corner of the domain, which mimics the
localized scenario of needle insertion. To demonstrate the
computational advantage of adaptive refinement over uniform
refinement, we also present the computational time for this
problem. Then, to point out the benefits of a local mesh
refinement in needle insertion simulation, we study a needle
insertion scenario with friction to show the impact of local
refinement on the displacement field around the needle shaft
and also on the needle-tissue interaction force profile. Finally,
a more complicated scenario is simulated i.e. insertion of a
needle into a liver which is undergoing breathing motion.
In our simulations, the needle and soft tissue follow a linear elastic constitutive law, associated to a co-rotational FE
formulation.
A. Convergence study
To show the advantage of error-controlled adaptive refinement scheme, as compared to the uniform mesh refinement, a
convergence study is performed on a 3D L-shaped domain. As
shown in Fig. 6, the L-shaped domain is clamped at the right
boundary, and simply supported in the vertical direction at the
top boundary. The dimension is set to L = 4 and the thickness
of the domain is L/2 = 2. Young’s modulus, and Poisson’s
L/2
L/2
q
Fig. 6: Boundary conditions of the L-shaped domain test.
We define the relative error η as
qP
Ne R
h
s T
h
s
e=1 Ωe ( − ) (σ − σ )dΩ
qP
η=
.
Ne R
h )T σ h dΩ
(
e=1 Ωe
(26)
In Fig. 7, we show the plots of the relative error versus the
number of DOFs for uniform and adaptive refinement. We see
that for the uniform refinement, the relative error η converges
with a slope of 0.21, which corresponds to the theoretical slope
of 2/9 for singular problems in 3D. By comparison with the
uniform refinement, the adaptive refinement converges with a
higher slope (0.31). Clearly, to achieve certain expected error
of the simulation, the adaptive refinement needs fewer DOFs
than the uniform refinement.
1
η
(A + JT WJ)dvk+1 = b − JT λk
ratio of the tested material are 1 × 103 , and 0.3, respectively.
The domain is subjected to a uniformly distributed traction
force on the left surface boundary.
Starting with the mesh having 8×8×4 hexahedral elements
(excluding the 4 × 4 × 4 corner elements), two types of
refinements are performed. The first one, called uniform refinement, consists of subsequently subdividing every hexahedral
element in to 8 smaller elements. In the second approach,
called adaptive refinement, only those elements which satisfy
the marking condition (20) are refined (subdivided in to 8
smaller elements).
Uniform
Adaptive
0.1
0.21
0.31
0.01
1000
1
10000
1
100000
1 × 106
DOFs
Fig. 7: Convergence of the relative error, comparison between
uniform and adaptive refinements.
To demonstrate the performance of adaptive (local) refinement in terms of computational time, as compared to
8
uniform (full) refinement of the mesh, with the same expected
relative error η = 8%, we again studied the L-shaped domain
problem. The result is reported in Table I. The local refinement
decreases the number of DOFs by a factor 5.9 (7473 vs.
44064) associated with a computational speed-up of 16×
(736.41 ms vs. 12140.8 ms).
DOFs
Full refinement
Local refinement
44064
7473
Time
TC
MS
x
12140.8
319.24
417.17
Total
time
12140.8
736.41
TABLE I: Computational time (in 10−3 s). TC: Topological
Changes, MS: System Matrix Solve.
In view of above observations, it is a strong argument to
support the employment of adaptive refinement scheme while
limiting the discretization error in real-time simulations.
B. Impact of local mesh refinement on displacement field
We now present the results of a simulation of needle
insertion into a homogeneous 3D tissue model. For this study,
we consider the Young’s modulus of 108 for the needle, and
103 for the tissue, whereas the Poisson’s ratio is taken as 0.4
for both. The friction coefficient between the needle and the
tissue is set to 0.9. The displacement field due to frictional
interactions with the needle, viewed from the xy plane of the
tissue, is shown in Fig. 8.
It is shown (by the nonlinear variation of the displacement
field in the vicinity of the needle) that when the mesh is
adaptively refined near the needle shaft, the needle-tissue
interaction is captured as good as in the case of full refinement
(see Fig. 8d). Indeed, closer the position is to the needle shaft,
higher the obtained displacement field. Conversely, when a
coarse element is used, and is not refined during the simulation, the above behavior is not reproduced within the element
(Fig. 8a). It is important to point out that the refinement using
anisotropic template, as in Fig. 8b, is very relevant since
it generates fewer DOFs than using the isotropic template
(Fig. 8c), while still catching the nonlinear displacement field.
C. Impact of local mesh refinement on needle-tissue interaction
In order to gain insight into the nonlinear behavior of
the needle-tissue interaction around the needle shaft, and to
exhibit the effect of the adaptive refinement on a trade-off
between computational time and precision, a needle insertion
simulation into a phantom tissue test is carried out, see Fig. 9.
For this study, we consider the Young’s modulus of 50 MPa
for the needle, and 10 MPa for the tissue, whereas the
Poisson’s ratio is taken as 0.4 for the tissue, and 0.3 for the
needle, respectively. Again, a linear elastic model based on
corotational formulation is employed for the needle as well
as the tissue. The dimension of the tissue is 4 × 2 × 2 cm.
The needle length and radius are of 3.2 cm and 0.1 cm,
respectively.
Three meshing schemes are employed: a coarse mesh with
resolution 10 × 5 × 5 nodes, a fine mesh with resolution 20 ×
10 × 10 nodes and an adaptive mesh (starting with the coarse
mesh 10×5×5 nodes and adaptively refining the mesh during
the simulation). Within the adaptive meshing scheme, the mesh
refinement is again piloted by the error estimate described
in Section III.
To investigate the sensitivity of the needle-tissue interaction
parameters (frictional coefficient and puncture strength) on the
resulting mesh adaptation, and thus on computational output,
two scenarios are studied. The first concerns varying the
puncture strength σn0 parameter (0 N, 10 N, and 20 N), while
keeping the same frictional coefficient µ = 0.5 between the
tissue and the needle shaft. The second is dedicated to study
the influence of the frictional coefficient by setting it to 0.1,
0.3 and 0.5, while keeping the puncture strength unchanged
(σn0 = 10 N). Within these two scenarios, the frictional
coefficient on the tissue surface is set to 0.8.
Fig. 10 shows the plots of the integrated interaction force
along the needle shaft versus the displacement of the needle
tip for the first scenario. The second scenario is depicted
in Fig. 11.
It shows that when the contact force between the needle
tip and the tissue surface is higher than the tissue puncture
strength, the needle penetrates into the tissue. Right after this
penetration event, a relaxation phase can be observed that
induces a decreasing force at the needle bases. Thereafter, it
is observed that as the needle moves forward, the interaction
force increases due to the increasing frictional force along the
needle shaft (which is directly proportional to the insertion
distance). Only when the contact force at the needle tip is
greater than the cutting strength of the soft tissue, the needle
cuts the tissue and continues going ahead. Immediately after
this cutting action, the relaxation phase is observed anew. This
behavior is periodically observed during the needle insertion.
These observations are clearly shown in Fig. 12a, which is
obtained by zooming in Fig. 10b. A typical behavior with
distinguished phases is presented in Fig. 12b.
It is observed that under mesh refinement the resulting
global behavior of needle-tissue system is less stiff. This is
explained by the fact that beneath mesh refinement, a greater
displacement field is obtained, which results from the needletissue interaction (as also observed in Section V-B). It is
also shown that using the adaptive refinement scheme, the
interaction needle-tissue behavior is close to those when using
a fine mesh (see Fig. 10). As can be seen in Fig. 13, an
interesting observation is that the number of DOFs in the
adaptive refinement simulation is significantly fewer than that
of using the fine mesh. This obviously results in an important
gain in terms of computational time. Indeed, the simulation
using the adaptive refinement mesh runs at nearly 45 FPS
compared to 4 FPS of that using the uniform fine mesh.
It is seen from Fig. 11 that the smaller the frictional
coefficient is, the more the behavior of the adaptive refinement
scheme differs from that of the simulation using the fine mesh.
It aligns nicely with the fact that smaller friction force does
not lead to mesh refinement around the needle shaft. Indeed,
as seen in Fig. 14, the refinement in the case of frictional
coefficient µ = 0.1 is mostly due to the penetration force at
the tissue surface.
9
(a)
(b)
(c)
(d)
0
0.01
0.02
(e)
0.03
Fig. 8: Variation of tissue displacement resulting from friction during needle insertion, measured along a vertical line located at
the needle tip. (a) The un-refined mesh; (b) adaptive refinement using an anisotropic template 2 × 3 × 3; (c) adaptive refinement
using an isotropic template 3 × 3 × 3; and (d) full refinement. The graph (e) shows the benefits of the anisotropic refinement.
z
y
x
Fig. 9: Schematic representation of needle insertion simulation
into a phantom tissue. The phantom tissue is clamped at the
right surface.
D. Application to liver
The method proposed in this paper is now applied to a liver
model undergoing breathing motion, to mimic a typical case
of radio-frequency ablation of a tumor. The same Young’s
modulus and Poisson’s ratio for the needle and the tissue
as in Section V-C are employed. The frictional coefficient
is set to 0.5 when the needle is inserted and to 0.1 when
it is pulled back. The puncture force at the tissue surface
is set to 10 N. Induced by error estimate (Eq. (20)), the
needle insertion and constraints applied to the liver lead to
refinements in different regions. The initial mesh has 1179
DOFs. When the needle advances into the liver, combining
with the motion of the liver due to breathing effect, the mesh
is progressively refined to accurately take in to account the
interaction of needle and liver. The maximum number of DOFs
when the needle is completely inserted into the liver is 2961.
And, when the needle is steadily pulled back, the mesh is
then progressively coarsened until the needle is completely
outside of the liver. Thereafter, the refinement process is now
only due to the movement of the liver by breathing effect and
imposed boundary conditions. The number of DOFs at this
stage is 1509. By applying this adaptive refinement/coarsening
procedure, it is not only guaranteed that the discretization
error is fully controlled, but the computational cost is also
kept as small as possible. Indeed, without adaptive remeshing
procedure applied on the initial mesh, the simulation runs at
35 FPS while the discretization error is 12%, whereas when
the adaptive refinement is performed, it runs at 22 FPS while
decreasing the discretization error to 6%. Note that these frame
rates result not only from computational resolutions of needle,
tissue and their interactions but also from their visualization
cost.
In order to investigate the benefits of the adaptive refinement
scheme when the needle is inserted and retracted into the liver
phantom, tests with uniform and adaptive refinement schemes
are carried out. Within the uniform refinement case, a coarse
mesh with 723 DOFs and a fine mesh with 3894 DOFs are
used for the liver discretization. Whereas upon the adaptive
refinement scenario, the simulations start with the coarse mesh
723 DOFs and is adaptively refined by two schemata: each
marked element is refined into (i) 2 × 2 × 2 elements, and
(ii) into 3 × 3 × 3 elements. The integrated needle-tissue
interaction force along the needle shaft is plotted versus the
needle tip displacement when the needle is inserted and pulled
back, see Fig. 16.
It is observed that when the needle is outside the tissue,
there is no interaction force between them. This is also detected when the needle is completely retracted from the tissue.
It is clearly shown that the needle-tissue interaction depends
strongly on the mesh used (especially if the mesh is coarse).
The mesh influence reveals a stronger effect on the insertion
stage than the pullback one. This is fully understood by the
fact that the frictional coefficient between the needle shaft and
the tissue is more important during the insertion steps than that
during the pullback ones (0.5 versus 0.1, respectively). Using
the coarse mesh, the puncture force at the tissue surface is
not well captured compared to the case where the fine mesh
or adaptive refined mesh is employed, see Fig. 16. Under
mesh refinement around the needle shaft guided by error
estimate, the needle-tissue interaction converges to the solution
of the fine mesh. However, the maximum number of DOFs
when using the adaptive refinement schemes 2 × 2 × 2 and
3×3×3 is 1071 and 2193 respectively. Therefore, as observed
in Section V-C, using the adaptive refinement scheme results
in significantly fewer DOFs compared to the employment
of uniform fine mesh. All of these lead to the conclusion
that even starting with the coarse mesh but employing the
adaptive refinement scheme, the needle-tissue interaction can
be simulated more precisely compared to the coarse mesh, and
with significantly lower computational cost compared to the
uniform fine mesh.
VI. C ONCLUSION AND PERSPECTIVES
This paper contributes a structured approach to answering
the important but rarely tackled question of accuracy in
10
100
80
60
40
20
0
120
mesh 10x5x5
mesh 20x10x10
adaptive refinement
100
Force [N]
Force [N]
120
mesh 10x5x5
mesh 20x10x10
adaptive refinement
80
80
60
40
60
40
20
0
0.5
1 1.5 2 2.5
Displacement [cm]
3
0
3.5
mesh 10x5x5
mesh 20x10x10
adaptive refinement
100
Force [N]
120
20
0
0.5
(a) λn0 = 0 N
1 1.5 2 2.5
Displacement [cm]
3
0
3.5
0
0.5
(b) λn0 = 10 N
1 1.5 2 2.5
Displacement [cm]
3
3.5
(c) λn0 = 20 N
Fig. 10: Comparison of needle-tissue interaction forces along the needle shaft within two cases: without refinement (with
different mesh resolutions) and with adaptive refinement. The penetration strength is varied while keeping the same frictional
coefficient µ = 0.5 between the needle shaft and the soft tissue.
20
50
15
10
5
0
120
mesh 10x5x5
mesh 20x10x10
adaptive refinement
60
Force [N]
Force [N]
70
mesh 10x5x5
mesh 20x10x10
adaptive refinement
40
80
30
20
60
40
20
10
0
0.5
1 1.5 2 2.5
Displacement [cm]
3
3.5
0
mesh 10x5x5
mesh 20x10x10
adaptive refinement
100
Force [N]
25
0
(a) µ = 0.1
0.5
1 1.5 2 2.5
Displacement [cm]
3
0
3.5
0
0.5
(b) µ = 0.3
1 1.5 2 2.5
Displacement [cm]
3
3.5
(c) µ = 0.5
Fig. 11: Comparison of needle-tissue interaction forces along the needle shaft within two cases: without refinement (with
different mesh resolutions) and with adaptive refinement. The frictional coefficient between the needle shaft is varied while
keeping the same penetration strength (10 N) at the tissue surface.
30
25
(3)
(3)
10
20
15
10
5
0
mesh 20x10x10
12
Force [N]
Force [N]
14
mesh 10x5x5
mesh 20x10x10
adaptive refinement
8
(3)
(2)
6
(1)
(2)
(3)
(2)
4
(2)
(3)
(2)
2
0
0.2
0.6
0.8
0.4
Displacement [cm]
(a)
1
0
0
0.2
0.6
0.8
0.4
Displacement [cm]
1
(b)
Fig. 12: (a) Puncture, cutting and relaxation behaviors are shown by a zoom-in plot from Fig. 10b. (b) The typical behavior is
shown in phases. Phase (1): The needle is puncturing the tissue surface. Phase (2): Just after the penetration event, the relaxation
occurs. Phase (3): The interaction force increases due to the fact that frictional force increases with insertion distance. When
the needle tip has cut the tissue to advance forward, the relaxation occurs again.
11
agreement No. 279578) RealTCut towards real time multiscale
simulation of cutting in non-linear materials with applications
to surgical simulation and computer guided surgery”.
Inria thanks for funding of European project RASimAs (FP7
ICT-2013.5.2 No610425).
Stéphane P.A. Bordas is grateful for many helpful discussions with Dr. Pierre Kerfriden, Prof. Karol Miller, Prof. Christian Duriez, Prof. Michel Audette and Mr. Diyako Ghaffari.
7000
6000
DOFS
5000
4000
mesh 10x5x5
mesh 20x10x10
adaptive refinement
3000
2000
1000
0
0
0.5
1 1.5 2 2.5
Displacement [cm]
3
3.5
Fig. 13: Number of DOFs during the needle insertion of the
simulation in Fig. 10b.
surgical simulation.
The novelty of our paper is to drive local adaptive mesh
refinement during needle insertion by a robust a posteriori
estimate of the discretization error.
This can be seen as a first step to control the error associated
with acceleration methods in needle surgical simulations, and
to separate the modeling (are we solving the right problem?)
and discretization error (are we solving the problem right?).
Verification of the discrete scheme is guaranteed in our
approach because the a posteriori estimate asymptotically
converges to the exact error. As we use an implicit approach,
we also control the error on the equilibrium equations. As
such, assuming a proper material model and kinematics for
the problem, we can guarantee the accuracy of the solution.
This is not the case in explicit time stepping approaches.
Validation of the approach is not considered. We focus
here only on one source of error (discretization). Whilst this
is a limitation, we do believe that quantifying discretization
errors separately to modeling errors is necessary to devise
accurate and clinically-usable surgical simulators and to better
understand the resulting simulation results.
A natural direction for research, building on recent work
on data-driven simulations is to devise error-controlled approaches able to learn from data as it is acquired during the
simulation [34]–[36]. In such a paradigm, the model would
adapt to the real situation, as opposed to being driven by a
continuous indirect comparison, as is the case in this work,
with an unknown exact solution. In turn, such a data-driven
approach would facilitate patient-specific simulations, which
were not considered here. We are currently investigating such
directions through Bayesian inference for parameter identification and model selection [37] and uncertainty quantification
approaches [38].
ACKNOWLEDGEMENTS
The first and last author are supported by the Fellowship of
the last author as part of the University of Strasbourg Institute
for Advanced Study (BPC 14/Arc 10138).
Stéphane P.A. Bordas and Satyendra Tomar also thank partial funding for their time provided by the European Research
Council Starting Independent Research Grant (ERC Stg grant
R EFERENCES
[1] A. Nealen et al., “Physically based deformable models in computer
graphics,” in Computer graphics forum, vol. 25, no. 4, 2006, pp. 809–
836.
[2] Y. Wang et al., “Linear subspace design for real-time shape deformation,” ACM Transactions on Graphics (TOG), vol. 34, no. 4, p. 57, 2015.
[3] H. Courtecuisse et al., “Real-time simulation of contact and cutting of
heterogeneous soft-tissues,” Medical Image Analysis, vol. 18, no. 2, pp.
394 – 410, 2014.
[4] O. Zienkiewicz et al., The finite element method: Its basis and fundamentals. Elsevier, 2013, vol. 1.
[5] V. P. Nguyen et al., “Meshless methods: a review and computer implementation aspects,” Mathematics and computers in simulation, vol. 79,
no. 3, pp. 763–813, 2008.
[6] X. Wu et al., “Adaptive nonlinear finite elements for deformable body
simulation using dynamic progressive meshes,” in Computer Graphics
Forum, vol. 20, no. 3, 2001, pp. 349–358.
[7] M. Müller et al., “Stable real-time deformations,” in Proceedings of the
2002 ACM SIGGRAPH/Eurographics symposium on Computer animation, 2002, pp. 49–54.
[8] M. Ainsworth and J. T. Oden, A posteriori error estimation in finite
element analysis. John Wiley & Sons, 2011, vol. 37.
[9] M. Seiler et al., “Robust interactive cutting based on an adaptive octree
simulation mesh,” The Visual Computer, vol. 27, no. 6-8, pp. 519–529,
2011.
[10] R. Verfürth, A posteriori error estimation techniques for finite element
methods, ser. Numerical Mathematics and Scientific Computation. Oxford University Press, Oxford, 2013.
[11] O. Zienkiewicz and J. Zhu, “The superconvergent patch recovery (spr)
and adaptive finite element refinement,” Computer Methods in Applied
Mechanics and Engineering, vol. 101, no. 1, pp. 207–224, 1992.
[12] C. Carstensen and S. Bartels, “Each averaging technique yields reliable
a posteriori error control in FEM on unstructured grids. I. Low order
conforming, nonconforming, and mixed FEM,” Math. Comp., vol. 71,
no. 239, pp. 945–969 (electronic), 2002.
[13] S. Bartels and C. Carstensen, “Each averaging technique yields reliable
a posteriori error control in FEM on unstructured grids. II. Higher order
FEM,” Math. Comp., vol. 71, no. 239, pp. 971–994 (electronic), 2002.
[14] I. Babuška and W. C. Rheinboldt, “Error estimates for adaptive finite
element computations,” SIAM J. Numer. Anal., vol. 15, no. 4, pp. 736–
754, 1978.
[15] N. Hamzé et al., “Preoperative trajectory planning for percutaneous procedures in deformable environments,” Computerized Medical Imaging
and Graphics, vol. 47, pp. 16–28, 2016.
[16] N. Abolhassani et al., “Needle insertion into soft tissue: A survey,”
Medical Engineering and Physics, vol. 29, pp. 413–431, 2007.
[17] C. Duriez et al., “Interactive simulation of flexible needle insertions
based on constraint models,” in Lecture Notes in Computer Science,
vol. 5762, no. PART 2, 2009, pp. 291–299.
[18] S. Misra et al., “Mechanics of flexible needles robotically steered
through soft tissue,” Int. J. Rob. Res., vol. 29, no. 13, pp. 1640–1660,
Nov. 2010.
[19] J. Wu et al., “Physically-based simulation of cuts in deformable bodies:
A survey.” in Eurographics (State of the Art Reports), 2014, pp. 1–19.
[20] C. Dick et al., “A hexahedral multigrid approach for simulating cuts in
deformable objects,” IEEE Transactions on Visualization and Computer
Graphics, vol. 17, no. 11, pp. 1663–1675, 2011.
[21] G. R. Liu and S. S. Quek, “Chapter 3 - Fundamentals for Finite
Element Method,” in The Finite Element Method (Second Edition),
second edition ed., G. R. Liu et al., Eds.
Oxford: ButterworthHeinemann, 2014, pp. 43–79.
[22] O. Zienkiewicz and R. Taylor, The Finite Element Method: Solid
mechanics, ser. Referex collection.Mecánica y materiales. ButterworthHeinemann, 2000.
12
(a) µ = 0.1
(b) µ = 0.3
(c) µ = 0.5
Fig. 14: Refinement patterns (colored by stress level) of the adaptive scenario while using different frictional coefficients.
(a)
(b)
(c)
Fig. 15: (a) Simulation of needle insertion in a liver; (b) using dynamic mesh refinement scheme driven by error estimate; (c)
visual depiction. The simulation runs at 22 Hz (on a 4GHz processor).
80
uniform mesh 723 dofs
uniform mesh 3894 dofs
adaptive refinement 2x2x2
adaptive refinement 3x3x3
70
60
Force [N]
50
[27]
[28]
40
30
[29]
20
10
[30]
0
[31]
-10
-20
0
0.5
1
1.5 2 2.5 3
Displacement [cm]
3.5
4
4.5
Fig. 16: Needle-liver phantom interaction force during needle
insertion and pullback. The interaction force varies due to
advancing friction and tissue cutting strength, globally increases (with positive values) during the insertion stage. At
4 cm of the needle-tip displacement, the needle is retracted
and the interaction force changes the direction, varies due to
retrograding friction, globally decreases and gets zero when
the needle is completely pulled out. These form a loop profile
for force-displacement curve resulting from needle insertionretraction.
[23] C. Felippa and B. Haugen, “A unified formulation of small-strain
corotational finite elements: I. theory,” Computer Methods in Applied
Mechanics and Engineering, vol. 194, no. 2124, pp. 2285 – 2335, 2005.
[24] A. Pinelli et al., “Immersed boundary method for generalised finite
volume and finite difference navier-stokes solvers,” Journal of Computational Physics, vol. 229, no. 24, pp. 9073–9091, 2010.
[25] D. Baraff and A. Witkin, “Large steps in cloth simulation,” in Proceedings of SIGGRAPH, 1998, pp. 43–54.
[26] D.-Y. Kwak and Y.-T. Im, “Hexahedral mesh generation for remeshing
[32]
[33]
[34]
[35]
[36]
[37]
[38]
in three-dimensional metal forming analyses,” Journal of Materials
Processing Technology, vol. 138, no. 13, pp. 531 – 537, 2003.
D. Koschier et al., “Adaptive Tetrahedral Meshes for Brittle Fracture
Simulation,” in Symposium on Computer Animation, V. Koltun and
E. Sifakis, Eds. The Eurographics Association, 2014.
D. Burkhart et al., “Adaptive and Feature-Preserving Subdivision for
High-Quality Tetrahedral Meshes,” Computer Graphics Forum, vol. 29,
no. 1, pp. 117–127, 2010.
C. Paulus et al., “Virtual cutting of deformable objects based on efficient
topological operations,” The Visual Computer, vol. 31, no. 6-8, pp. 831–
841, 2015.
E. Sifakis et al., “Hybrid simulation of deformable solids,” in Proc.
Symposium on Computer Animation, 2007, pp. 81–90.
H. Uzawa and K. J. Arrow, “Iterative methods for concave programming,” in Preference, production, and capital. Cambridge University
Press, 1989, pp. 135–148.
I. Babuška, “The finite element method with penalty,” Mathematics of
Computation, vol. 27, no. 122, pp. 221–228, 1973.
P. Papadopoulos and J. M. Solberg, “Recent Advances in Contact Mechanics A Lagrange multiplier method for the finite element solution of
frictionless contact problems,” Mathematical and Computer Modelling,
vol. 28, no. 4, pp. 373–384, 1998.
M. Seiler et al., “Enriching coarse interactive elastic objects with
high-resolution data-driven deformations,” in Proceedings of the ACM
SIGGRAPH/Eurographics Symposium on Computer Animation, 2012,
pp. 9–17.
——, “Data-driven simulation of detailed surface deformations for
surgery training simulators,” IEEE Transactions on Visualization and
Computer Graphics, vol. 20, no. 10, pp. 1379–1391, Oct 2014.
M. U. Seiler et al., “Efficient Transfer of Contact-Point Local Deformations for Data-Driven Simulations,” in Workshop on Virtual Reality
Interaction and Physical Simulation. The Eurographics Association,
2014.
H. Rappel et al., “Bayesian inference for the stochastic identification
of elastoplastic material parameters: Introduction, misconceptions
and additional insight,” CoRR, vol. abs/1606.02422, 2016. [Online].
Available: http://arxiv.org/abs/1606.02422
P. Hauseux et al., “Accelerating monte carlo estimation with
derivatives of high-level finite element models,” Computer Methods in
Applied Mechanics and Engineering, pp. –, 2017. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0045782516313470
| 5cs.CE
|
arXiv:1407.6513v1 [cs.NE] 24 Jul 2014
Convolutional Neural Associative Memories:
Massive Capacity with Noise Tolerance
Amin Karbasi
E-mail: [email protected]
Computer Science Department
Swiss Federal Institute of Technology Zurich, 8092 Zurich, Switzerland
Amir Hesam Salavati
E-mail: [email protected]
Computer and Communication Sciences Department
Ecole Polytechnique Federale de Lausanne, Lausanne, 1015, Switzerland
Amin Shokrollahi
E-mail: [email protected]
Computer and Communication Sciences Department
Ecole Polytechnique Federale de Lausanne, Lausanne, 1015, Switzerland
July 25, 2014
Abstract
The task of a neural associative memory is to retrieve a set of previously
memorized patterns from their noisy versions using a network of neurons.
An ideal network should have the ability to 1) learn a set of patterns as they
arrive, 2) retrieve the correct patterns from noisy queries, and 3) maximize
the pattern retrieval capacity while maintaining the reliability in responding
to queries. The majority of work on neural associative memories has focused
on designing networks capable of memorizing any set of randomly chosen
patterns at the expense of limiting the retrieval capacity.
In this paper, we show that if we target memorizing only those patterns
that have inherent redundancy (i.e., belong to a subspace), we can obtain all
the aforementioned properties. This is in sharp contrast with the previous
work that could only improve one or two aspects at the expense of the third.
More specifically, we propose framework based on a convolutional neural
network along with an iterative algorithm that learns the redundancy among
the patterns. The resulting network has a retrieval capacity that is exponential in the size of the network. Moreover, the asymptotic error correction
performance of our network is linear in the size of the patterns. We then extend our approach to deal with patterns lie approximately in a subspace. This
extension allows us to memorize datasets containing natural patterns (e.g.,
images). Finally, we report experimental results on both synthetic and real
datasets to support our claims.
1
Introduction
The ability of neuronal networks to memorize a large set of patterns and reliably
retrieve them in the presence of noise, has attracted a large body of research over
the past three decades to design artificial neural associative memories with similar capabilities. Ideally, a perfect neural associative memory should be able to
learn patterns, have a large pattern retrieval capacity and be noise-tolerant. This
problem, called ”associative memory", is in spirit very similar to reliable information transmission faced in communication systems where the goal is to efficiently
decode a set of transmitted patterns over a noisy channel.
Despite this similarity and common methods deployed in both fields (e.g.,
graphical models, iterative algorithms, to name a few), we have witnessed a huge
gap between the efficiency achieved by them. More specifically, by deploying modern coding techniques, it was shown that the number of reliably transmitted patterns
over a noisy channel can be made exponential in n, the length of the patterns. This
was particularly achieved by imposing redundancy among transmitted patterns. In
contrast, the maximum number of patterns that can be reliably memorized by most
current neural networks scales linearly in the size of the patterns. This is due to the
common assumption that a neural network should be able to memorize any subset
of patterns drawn randomly from the set of all possible vectors of length n (see,
for example Hopfield, 1982, Venkatesh and Psaltis, 1989, Jankowski et al., 1996,
Muezzinoglu et al., 2003).
Recently, Kumar et al. (2011) suggested a new formulation of the problem
where only a suitable set of patterns was considered for storing. To enforce the
set of constraints, they formed a bipartite graph (as opposed to a complete graph
considered in the earlier work) where one layer feeds the patterns to the network
and the other takes into account the inherent structure. The role of bipartite graph
is indeed similar to the Tanner graphs used in modern coding techniques (Tanner,
1981). Using this model, Kumar et al. (2011) provided evidence that the resulting
network can memorize an exponential number of patterns at the expense of correcting only a single error during the recall phase. By introducing a multi-layer
2
structure, Salavati and Karbasi (2012) could further improve the error correction
performance to constant number of errors.
In this paper, similar to the model considered by Kumar et al. (2011), we only
consider a set of patterns with weak minor components, i.e., patterns that lie in a
subspace. By making use of this inherent redundancy
• We introduce the first convolutional neural associative network with provably exponential storage capacity.
• We prove that our architecture can correct a linear fraction of errors.
• We develop an online learning algorithm with the ability to learn patterns as
they arrive. This property is specifically useful when the size of the dataset
is massive and patterns can only be learned in a streaming manner.
• We extend our results to the case where patterns lie approximately in a subspace. This extension in particular allows us to efficiently memorize datasets
containing natural patterns.
• We evaluate the performance of our proposed architecture and the learning
algorithm through numerical simulations.
We provide rigorous analysis to support our claims. The storage capacity and error
correction performance of our method is information-theoretically order optimum,
i.e., no other method can significantly improve the results (except for constants).
Our learning algorithm is an extension of the subspace learning method proposed
by Oja and Kohonen (1988), with an additional property of imposing the learned
vectors to be sparse. The sparsity is essential during the noise-elimination phase.
The remainder of this paper is organized as follows. In Section 2 we provide
an overview of the related work in this area. In Section 3 we introduce our notation
and formally state the problems that is the focus of this work, namely, learning
phase, recall phase, and storage capacity. We present our learning algorithm in
Section 4 and our error correction method in Section 5. Section 6 is devoted to the
pattern retrieval capacity. We then report our experimental results on synthetic and
natural datasets in Section 7. Finally, all the proofs are provided in Section 8.
2
Related Work
The famous Hopfield network was among the first auto-associative neural mechanisms capable of learning a set of patterns and recalling them subsequently (Hopfield, 1982). By employing the Hebbian learning rule (Hebb, 1949), Hopfield
3
considered a neural network of size n with binary state neurons. It was shown
by McEliece et al. (1987) that the capacity of a Hopfield network is bounded by
C = (n/2 log(n)). Due to the low capacity of Hopfield networks, extension of
associative memories to non-binary neural models has also been explored, with
the hope of increasing the pattern retrieval capacity. In particular, Jankowski et al.
(1996) investigated a complex-valued neural associative memory where each neuron can be assigned a multivalued state from the set of complex numbers. It was
shown by Muezzinoglu et al. (2003) that the capacity of such networks can be increased to C = n at the cost of a prohibitive weight computation mechanism. To
overcome this drawback, a Modified Gradient Descent learning Rule (MGDR) was
devised by Lee (2006).
Recently, in order to increase the capacity and robustness, a line of work considered exploiting the inherent structure of the patterns. This is done by either
making use of the correlations among the patterns or memorizing only those patterns that have some sort of redundancy. Note that they differ from the previous
work in one important aspect: not any possible set of patterns is considered for
learning, but only those with common structures. By employing neural cliques,
Gripon and Berrou (2011) were among the first to demonstrate that considerable
improvements in the pattern retrieval capacity of Hopfield networks is possible,
albeit still not passing the polynomial boundary on the capacity, i.e., C = O(n2 ).
Similar idea was proposed by Venkatesh (1994) for learning semi-random patterns. This boost to the capacity is achieved by dividing the neural network into
smaller fully interconnected
disjoint blocks. Using this idea, the capacity is in
n/b
creased to Θ b
, where b = ω(ln n) is the size of clusters, Nonetheless, it
was observed that this improvement comes at the price of limited noise tolerance
capabilities.
By deploying higher order neural models, in contrast to the pairwise correlation
considered in Hopfield networks, Peretto and Niez (1986) showed that the storage
capacity can be improved to C = O(np−2 ), where p is the degree of correlation.
In such models, the state of the neurons not only depends on the state of their
neighbors, but also on the correlations among them. However, the main drawback
of this work lies in the prohibitive computational complexity of the learning phase.
Recently, Kumar et al. (2011) introduced a new model based on bipartite graphs
to capture higher order (linear) correlations without the prohibitive computational
complexity in the learning phase. The proposed model was further improved later
(Kumar et al., 2014), Under the assumption that the bipartite graph is fully known,
sparse, and expander, the proposed algorithm by Kumar et al. (2011) increased the
pattern retrieval capacity to C = O(an ), for some a > 1. In addition to those
restrictive assumptions, the performance of the recall phase was still below par.
4
In this paper, we introduce a convolutional neural network, capable of memorizing an exponential number of structured patterns while being able to correct a
linear fraction of noisy neurons. Similar to the model considered by Kumar et al.
(2011), we assume that patterns lie in a low dimensional subspace. Note that a
network of size n, where each neuron can hold a finite number of states, is capable
of memorizing at most an exponential number of patterns in n. Also, correcting a
linear fraction of noisy nodes of the network is the best we can hope for. In addition, and more importantly in practice, we extend our results to the set of patterns
that only approximately belong to a subspace.
It is worth mentioning that learning a set of input patterns with robustness
against noise is not just the focus of neural associative memories. For instance,
Vincent et al. (2008) proposed an interesting approach to extract robust features
in autoencoders. Their approach is based on artificially introducing noise during
the learning phase and let the network learn the mapping between the corrupted
input and the correct version. This way, they shifted the burden from the recall
phase to the learning phase. We, in contrast, consider another form of redundancy
and enforce a suitable pattern structure that helps us design faster algorithms and
derive necessary conditions that help us guarantee to correct a linear fraction of
noise without previously being exposed to.
Although our neural architecture is not technically considered a Deep Belief Network (DBN), it shares some similarities. DBNs are typically used to extract/classify features by the means of several consecutive stages (e.g., pooling,
rectification, etc). Having multiple stages help the network to learn more interesting and complex features. An important class of DBNs are convolutional DBNs.
The input layer (also known as the receptive field) is divided into multiple overlapping patches and the network extracts features from each patch (Jarrett et al.,
2009). Since we divide the input patterns into a few overlapping smaller clusters,
our model is similar to those of convolutional DBNs. Furthermore, we also learn
multiple features (in our case dual vectors) from each patch where the feature extractions differ over different patches. This is indeed very similar to the approach
proposed by Le et al. (2010). In contrast to convolutional DBNs, the focus of this
work is not classification but rather recognition of the exact patterns from their
noisy versions. Moreover, in most DBNs, we not only have to find the proper dictionary for classification, but we also need to calculate the features for each input
pattern. This alone increases the complexity of the whole system, especially if denoising is part of the objective. In our model, however, the dictionary is defined
in terms of dual vectors. Consequently, previously memorized patterns are computationally easy to recognize as they yield the all-zero vector in the output of the
feature extraction stage. In other words, a non-zero output can only happen if the
input pattern is noisy. Another advantage of our model over DBNs is a much faster
5
learning phase. More precisely, by using a single layer with overlapping clusters
in our model the information diffuses gradually in the network. The same criteria
is achieved in DBNs by constructing several stages (Socher et al., 2011).
3
Problem Formulation
In this section, we set our notation and formally define the learning phase, recall
phase, and the storage capacity.
3.1
Learning Phase
Throughout this paper, each pattern is denoted by an integer-valued vector x =
(x1 , x2 , . . . , xn ) of length n where xi ∈ Q = {0, . . . , Q − 1} for i = 1, . . . , n
and Q is a non-negative integer. In words, the set Q could be thought of as the
short term firing rate of neurons. Let {si }ni denote the states of neurons in a neural
network G. Each neuron updates its state based on the states of its neighbors.
P
More precisely, a neuron j first computes a weighted some i∈N (j) wj,i si and
then applies a nonlinear activation function f : R → Q, i.e.,
sj = f
X
wj,i si .
i∈N (j)
Here, wj,i is the weight of the neural connection between neurons j and i, and
N (j) denotes the neighbors of neuron j in G. There are several possible activation functions used in the literature including, but not limited to, linear, threshold,
logistic, and tangent hyperbolic functions.
We denote the dataset of the patterns by the C ×n dimensional matrix X , where
patterns are stored as the rows. Our goal in this work is to memorize patterns with
strong local correlation among the entries. More specifically, we divide the entries
of each pattern x into L overlapping sub-patterns of lengths n1 , . . . , nL , so that
P
ni ≥ n. Note that due to overlaps, an entry in a pattern can be a member of
multiple sub-patterns, as shown in Figure 1. We denote the i-th sub-pattern by
(i)
(i)
x(i) = (x1 , x2 , . . . , x(i)
ni ).
To enforce local correlations, we assume that the sub-patterns x(i) form a subspace
of dimension ki < ni . This is done by imposing linear constraints on each cluster.
These linear constraints are captured during the learning phase in the form of
(i)
(i)
(i)
dual vectors. More specifically, we find a set of non-zero vectors w1 , w2 , . . . , wmi
6
G(1)
y1
Constraint neurons
G(2)
y2
y3
y4
(2)
Pattern neurons
x1
(2)
x2
G(3)
y5
(2)
x3
y6
y7
(2)
x4
Figure 1: Bipartite graph G. In this figure we see three subpatterns x(1) , x(2) , x(3)
along with corresponding clusters G(1) , G(2) , G(2) . The subpattern x(2) has overlaps with both G(1) and G(3) . The weights wi,j are chosen to ensure that W · x = 0
for all patterns x lying in a subspace.
that are orthogonal to the set of sub-patterns x(i) , i.e.,
(i)
(i)
yj = hwj , x(i) i = 0,
∀j ∈ [mi ] ∀i ∈ [L],
(1)
where [q] denotes the set {1, 2, · · · , q} and h·, ·i represents the inner product.
The weight matrix W (i) is constructed by placing all dual vectors next to each
other, i.e.,
(i) (i)
(i) >
W (i) = [w1 |w2 | . . . |wm
] .
i
Equation (1) can be written equivalently as
W (i) · x(i) = 0.
Cluster i represents the bipartite graph G(i) with the connectivity matrix W (i) .
In the next section, we develop an iterative algorithm to learn the weight matrices
W (1) , . . . , W (L) , while encouraging sparsity within each connectivity matrix W (i) .
One can easily map the local constraints imposed by the W (i) ’s into a global
constraint by introducing a global weight matrix W of size m × n. The first m1
rows of the matrix W correspond to the constraints in the first cluster, rows m1 + 1
to m1 +m2 correspond to the constraints in the second cluster, and so forth. Hence,
by inserting zero entries at proper positions, we can construct the global constraint
matrix W . We will use both the local and global connectivity matrices to eliminate
noise in the recall phase.
7
y8
3.2
Recall Phase
In the recall phase a noisy version, say x̂, of an already learned pattern x ∈ X is
given. Here, we assume that the noise is an additive vector of size n, denoted by e,
whose entries assume values independently from {−1, 0, +1}1 with corresponding
probabilities p−1 = p+1 = pe /2 and p0 = 1 − pe . In other words, each entry of the
noise vector is set to ±1 with probability pe . The ±1 values are chosen to simplify
the analysis. Our approach can be easily extended to other integer-valued noise
models.
We denote by e(i) , the realization of noise on the sub-pattern x(i) . In formula,
x̂ = x + e.2 Note that W · x̂ = W · e and W (i) · x̂(i) = W (i) · e(i) . Therefore, the
goal in the recall phase is to remove the noise e and recover the desired pattern x.
This task will be accomplished by exploiting the facts that a) we have chosen the
set of patterns X to satisfy the set of constraints W (i) · x(i) = 0 and b) we opted for
sparse neural graphs G(i) during the learning phase. Based on these two properties,
we develop the first recall algorithm that corrects a linear fraction of noisy entries.
3.3
Capacity
The last issue we look at in this work is the retrieval capacity C of our proposed
method. Retrieval or critical storage capacity is defined as the maximum number
of patterns that a neural network is able to store without having (significant) errors
in returned answers during the recall phase. Hence, the storage capacity C(n) is
usually measured in terms of the network size n. It is well known that the retrieval
capacity is affected by certain considerations about the neural network, including
the range of values or states for the patterns, inherent structure of patterns, and
topology of neural networks. In this work, we show that a careful combination
of patterns’ structure and neural network topology leads to an exponential storage
capacity in the size of the network.
4
The Learning Algorithm
In this section, we develop an algorithm for learning the weight matrix W (`) of
a given cluster `. By our assumptions, the sub-patterns lie in a subspace of dimension k` ≤ n` . Hence we can adopt the iterative algorithm proposed by Oja
1
In our experiments, we have considered larger integer values for noise as well, i.e.,
{−q, . . . , 0, . . . , q}, for some q ∈ N. The ±1 noise model here is considered to simplify the notations and analysis.
2
Note that since entries of x̂ should be between 0 and Q − 1, we cap values below 0 and above
Q − 1 to 0 and Q − 1, respectively.
8
and Karhunen (1985) and Xu et al. (1991) to learn the corresponding null space.
However, in order to ensure the success of the denoising algorithm proposed in
Section 5, we require W (l) to be sparse. To this end, the objective function shown
below has a penalty term to encourage sparsity. Furthermore, we are not seeking
an orthogonal basis as in the approach proposed by Xu et al. (1991). Instead, we
wish to find m` vectors w(`) that are orthogonal to the (sub-) patterns. Hence,
the optimization problem for finding a constraint vector w(`) can be formulated as
follows:
minw(`)
s.t.
X
|hx(`) , w(`) i|2 + ηg(w(`) ),
(2)
x∈X
(`)
kw k2 = 1.
(3)
In the above problem, x(`) is a sub-pattern of x drawn from the training set X , h·, ·i
indicates the inner product, η is a positive constant, and g(·) is the penalty term to
favor sparse results. In this paper, we consider
g(w(`) ) =
n
X
(`)
tanh(σ(wi )2 ).
i=1
(`)
(`)
It is easy to see that for large σ, the function tanh(σ(wi )2 ) approximates |sign(wi )|
(as shown in Figure 2). Therefore, the larger σ gets, the closer g(w(`) ) will be to
k · k0 . Another popular choice, widely used in compressed sensing (see, for example, Donoho, 2006 and Candès and Tao, 2006), is to pick g(w(`) ) = kw(`) k1 . Note
that the optimization problem (2) without the constraint (3) has the trivial solution
w(`) = 0 where 0 is the all-zero vector.
To minimize the objective function shown in (2) subject to the norm constraint
(3) we use stochastic gradient descent and follow a similar approach to that of Xu
et al. (1991). By calculating the derivative of the objective function and considering the updates required for each randomly picked pattern x, we will obtain the
following iterative algorithm:
y (`) (t) = hx(`) (t), w(`) (t)i,
(4)
w̃(`) (t + 1) = w(`) (t) − αt 2y (`) (t)x(`) (t) + ηΓ(w(`) (t)) ,
w(`) (t + 1) =
w̃(`) (t + 1)
.
kw̃(`) (t + 1)k2
(5)
(6)
In the above equations, t is the iteration number, x(`) (t) is the subpattern of a
pattern x(t) drawn at iteration t, αt is a small positive constant, and Γ(w(`) ) =
9
1
tanh(σx)
0.5
0
sign(x)
σ=1
σ = 10
σ = 50
−0.5
−1
−1
−0.8 −0.6 −0.4 −0.2
0
x
0.2
0.4
0.6
0.8
1
Figure 2: Approximation of sign(x) by tanh(σx). As we increase the value of σ
the approximation becomes more accurate.
∇g(w(`) ) is the gradient of the penalty term. The function Γ(w(`) ) encourages
sparsity. To see why, consider the i-th entry of Γ(·), namely,
∂g(z)
∂zi
= 2σzi (1 − tanh2 (σzi2 )).
Γi (zi ) =
Note that Γi ' 2σzi for relatively small values of zi , and Γi ' 0 for larger values
of zi (see Figure 3).
Thus, for proper choices of η and σ, equation (5) suppresses small entries of
w(`) (t) towards zero and favors sparser results. To further simplify the iterative
equations (4), (5), (6) we approximate the function Γ(w(`) (t)) with the following
threshold function (shown in Figure 4):
(
Γi (zi , θt ) =
zi if |zi | ≤ θt ;
0 otherwise.
where θt is a small positive threshold. Following the same approach taken by Oja
and Karhunen (1985) we assume that αt is small enough so that equation (6) can be
(`)
(`)
expanded as powers of αt . Also note that the inner product hw
(t), Γ(w (t), θt )i
is small so in the power expansion we can omit the term αt η hw(`) (t), Γ(w(`) (t), θt )i w(`) (t).
By applying the above approximations we obtain an iterative learning algorithm shown in Algorithm 1. In words, y (`) (t) is the projection of x(`) (t) onto
10
1
Γi (zi )
0.5
0
zi
zi (1 − tanh2 (zi2 ))
zi (1 − tanh2 (10zi2 ))
zi (1 − tanh2 (100zi2 ))
−0.5
−1
−1
−0.8 −0.6 −0.4 −0.2
0
zi
0.2
0.4
0.6
0.8
1
Figure 3: The sparsity penalty Γi (zi ) suppresses small values of zi towards zero.
Note that as σ gets larger, the support of Γi (zi ) gets smaller.
w(`) (t). If for a given data vector x(`) (t) the projection y (`) (t) is non-zero, then
the weight vector will be updated in order to reduce this projection.
4.1
Convergence Analysis
Our main idea for proving the convergence of the learning algorithm is to consider
the learning cost function defined as follows:
E(t) = E(w(`) (t)) =
C
2
1 X
hw(`) (t), xµ i
C µ=1
We show that as we gradually learn patterns from the data set X , the cost function
E(t) goes to zero. In order to establish this result we need to specify the learning
P
rate αt . In what follows, we assume that αt = Ω(1/t) so that t αt → ∞ and
P 2
(`)
t αt < ∞. We first show that the weight vector w (t) never becomes zero, i.e.,
(`)
kw (t)k2 > 0 for all t.
Lemma 1 Assume that kw(`) (0)k2 > 0 and α0 < 1/η. Then for all iterations t,
we have αt < α0 < 1/η and kw(`) (t)k2 > 0.
As mentioned earlier, all the proofs are given in Section 8.
The above lemma ensures that if we reach E(t) = 0 for some iteration t, it is
not the case that w(`) (t) = 0. Next, we prove the convergence of Algorithm 1 to a
minimum ŵ(`) for which E(ŵ(`) ) = 0.
11
0.04
f (x; θ)
0.02
0
−0.02
θ = 0.01
θ = 0.05
−0.04
−0.10 −0.08 −0.06 −0.04 −0.02 0.00
x
0.02
0.04
0.06
0.08
0.10
Figure 4: The soft threshold function f (x, θ) for two different values of θ.
Theorem 2 Under the conditions of Lemma 2, the learning Algorithm 1 converges
to a local minimum ŵ(`) for which E(ŵ(`) ) = 0. Moreover ŵ(`) is orthogonal to
all the patterns in the data set X .
We should note here that a similar convergence result can be proven without introducing the penalty term g(w(`) ). However, our recall algorithm crucially depends
on the sparsity level of learned w(`) ’s. As a consequence we encouraged sparsity by
adding the penalty term g(w(`) ). Our experimental results in Section 7 show that
in fact this strategy works perfectly and the learning algorithm results in sparse
solutions.
In order to find m` constraints required by the learning phase, we need to run
Algorithm 1 at least L times. In practice, we can perform this process in parallel, to speed up the learning phase. It is also more meaningful from a biological
point of view, as each constraint neuron can act independently from the others.
Although running Algorithm 1 in parallel may result in redundant constraints, our
experimental results show that by starting from different random initial points, the
algorithm converges to linearly independent constraints almost surely.
5
Recall Phase
Once the learning phase is finished, the weights of the neural graphs are fixed.
Thus, during the recall phase we assume that the connectivity matrix for each cluster i (denoted by W (i) ) has been learned and satisfies (1).
12
The recall phase of our proposed model consists of two parts: intra-cluster and
inter-cluster. During the intra-cluster part, clusters try to remove noise from their
own sub-patterns. As we will see shortly, each cluster succeeds in correcting a
single error with high probability. Such individual error correction performance is
fairly limited. The inter-cluster part capitalizes on the overlap among clusters to
improve the overall performance of the recall phase. In what follows, we describe
both parts in more details.
5.1
Intra-cluster Recall Algorithm
For the intra-cluster part, shown in Algorithm 2, we exploit the fact that the connectivity matrix of the neural network in each cluster is sparse and orthogonal to
the memorized patterns. As a result we have W (`) (x(`) + e(`) ) = W (`) e(`) where
e(`) is the noise added to the sub-pattern x(`) .
Algorithm 2 performs a series of forward and backward iterations to remove
e(`) . At each iteration, the pattern neurons decide locally whether to update their
current state or not: if the amount of feedback received by a pattern neuron exceeds
a threshold, the neuron updates its state, and remains intact, otherwise.3
In order to state our results, we need to define the degree distribution poly(`)
nomial (from the node perspective). More precisely, let Λi be the fraction of
P (`)
pattern neurons with degree i in cluster G(`) and define Λ(`) (x) = i Λi xi to be
the degree distribution polynomial for the pattern neurons in cluster `. In principle
Λ(`) (x) encapsulates all the information we need to know regarding cluster G(`) ,
namely, the degree distribution. The following theorem provides a lower bound
on the average probability of correcting a single erroneous pattern neuron by each
cluster.
Theorem 3 If we sample the neural graph randomly from Λ(`) (x), and let ϕ → 1,
then Algorithm 2 can correct (at least) a single error in cluster G(`) with probability at least
!!n` −1
d¯`
(`)
(`)
Pc = 1 − Λ
,
m`
where d¯` , n` , and m` are the average degree of pattern neurons, the number of
pattern neurons, and the number of constraint neurons in cluster G(`) , respectively.
3
In order to maintain the current value of a neuron, we can add self-loops to pattern neurons in
Figure 1. the self-loops are not shown in the figure for the sake of clarity).
(`)
(`)
(`)
4
In practice, we usually set yi = sign(hi ) only if |hi | > ψ, where ψ is a small positive
threshold.
13
To gain some intuition we can further simplify the expression in the above theorem
as follows
Pc(`) ≥ 1 −
d¯`
m`
!d(`)
n` −1
min
,
(`)
where dmin is the minimum degree of pattern neurons in cluster ` and we assumed
(`)
that dmin ≥ 1. This shows the significance of having high-degree pattern neurons.
(`)
(`)
In the extreme case of dmin = 0 then we obtain the trivial bound of Pc ≥ (1 −
(`)
(`)
Λ0 )n` −1 , where Λ0 is the fraction of pattern neurons with degree equal to 0. In
(`)
(`)
(`)
(`)
particular, for large n` we obtain Pc ≥ e−D0 , where D0 = Λ0 n` is the total
number of pattern neurons with degree 0. Thus, even if only a single pattern neuron
has a zero degree, probability of correcting a single error drops significantly.
While Theorem 3 provides a lower bound on the probability of correcting a
single error when the connectivity graph is sampled according to the degree distribution polynomial Λ(`) (x), the following lemma shows that under mild conditions
(that depends on the neighborhood relationship among neurons), Algorithm 2 will
correct a single input error with probability 1.
Lemma 4 If no two pattern neurons share the exact same neighborhood in cluster
G(`) , and as ϕ → 1, Algorithm 2 corrects (at least) a single error.
(`)
denote this average
For the remaining of the paper we let Pc = E` Pc
probability of correcting one error averaged over all clusters. Lemma 4 suggests
that nuder mild conditions, and in fact in many practical settings discussed later, Pc
is close to 1. Thus, from now on we pessimistically assume that if there is a single
error in a given cluster, Algorithm 2 corrects it with probability Pc and declares a
failure if there are more than one error.
5.2
Inter-cluster Recall Algorithms
As mentioned earlier, the error correction ability of Algorithm 2 is fairly limited.
As a result, if clusters work independently, they cannot correct more than a few
external errors. However, as clusters overlap their combined performance can potentially be much better. Basically, they can help each other in resolving external
errors: a cluster whose pattern neurons are in their correct states can provide truthful information to neighboring clusters. Figure 5 illustrates this idea.
This property is exploited in the inter-cluster recall approach, formally given
by Algorithm 3. In words, the inter-cluster approach proceeds by applying Algorithm 2 in a round-robin fashion to each cluster. Clusters either eliminate their
14
internal noise in which case they keep their new states and can now help other clusters, or revert back to their original states. Note that by such a scheduling scheme,
neurons can only change their states towards correct values.
The inter-cluster algorithm is in spirit similar to a famous decoding algorithm in
communication systems for erasure channels, called the peeling algorithm (Luby
et al., 2001). To make the connection more concrete, we first need to define a
e we
contracted version of the neural graph G as follows. In the contracted graph G,
(`)
compress all constraint nodes of a cluster G into a single super constraint node
v (`) (see Figure 6). Then, each super constraint node essentially acts as a check
node capable of detecting and correcting any single error among its neighbors (i.e.,
pattern neurons). In contrast, it declares a failure if two or more of its neighbors
are corrupted by noise. Once an error is corrected by a cluster, the number of errors
in overlapping clusters may also reduce which in turn help them to eliminate their
errors.
Through introducing the contracted graph, the similarity to the Peeling Decoder is now evident: in the Peeling Decoder, each constraint (so called checksum)
node is capable of correcting a single erasure among its neighbors. Similarly, if
there are more than one erasure among the neighbors, the checksum node declares
erasure. However, once an erasure is eliminated by a checksum node, it helps other
constraint nodes, namely those connected to the recently-eliminated erased node,
as they will have one less erasure among their neighbors to deal with.
Based on the above similarity, we borrow methods from modern coding theory
to obtain theoretical guarantees on the error rate of our proposed recall algorithm.
More specifically, we use Density Evolution (DE), first developed by Luby et al.
(2001) and generalized by Richardson and Urbanke (2008), to accurately bound
the error correction performance.
e i (resp. ρej ) denote the fraction of edges that are adjacent to pattern (resp.
Let λ
e1, . . . , λ
e L } the pattern degree
constraint) nodes of degree i (resp. j). We call {λ
distribution and {ρe1 , . . . , ρen } the super constraint degree distribution. Similar to
Section 5.1, it is convenient to define the degree distribution polynomials as follows:
e
λ(z)
=
X
ρe(z) =
X
e i z i−1 ,
λ
i
ρi z i−1 .
i
Now consider a given cluster v (`) and a pattern neuron x connected to v (`) . The
decision subgraph of x is defined as the subgraph rooted at x and branched out
from the super constraint nodes, excluding v (`) . If the decision subgraph is a tree
up to a depth of τ (meaning that no node appears more than once), we say that
15
G(1)
G(2)
G(3)
G(1)
G
(2)
G
(3)
(1)
G
G
(c) Step 2: cluster 2 fails.
G(1)
G(2)
G
(2)
G
G(2)
G(3)
(d) Step 3: cluster 3 succeeds.
G(1)
G(3)
(e) Step 4: cluster 1 fails again.
(1)
G(3)
(b) Step 1: cluster 1 fails.
(a) Initial step
(1)
G(2)
G(2)
G(3)
(f) Step 5: cluster 2 succeeds.
G(1)
(3)
G
(g) Step 7: cluster 1 succeeds.
G(2)
G(3)
(h) Step 8: Algorithm finishes successfully.
Figure 5: How overlaps among clusters help the neural network to achieve better
error correction performance. We assume that each cluster can correct one input
error. In other words, if the number of input errors are higher than one the cluster
declares a failure.
16
e corresponding to graph G in Figure 1.
Figure 6: Contraction graph G
the tree assumption holds for τ levels. An example of the decision subgraph is
shown in Figure 7. Finally, we say that the node v (`) is unsatisfied if it is connected
to a noisy pattern node. Recall that Pc denotes the average probability of a super
constraint node correcting a single error among its neighbors.
Theorem 5 Assume that G̃ is chosen randomly according to the degree distribution pair λ̃ and ρ̃. Then, as the number of vertices of G grows large, Algorithm 3 will succeed in correcting all errors with high probability as long as
e (1 − Pc ρe(1 − z)) < z for z ∈ (0, pe ).
pe λ
It is worth to make a few remarks about Theorem 5. First, the condition given in
Theorem 5 can be used to calculate the maximal fraction of errors Algorithm 3
e
can correct. For instance, for the degree distribution pair (λ(z)
= z 2 , ρe(z) = z 5 ),
the threshold is p∗e ≈ 0.429, below which Algorithm 3 corrects all the errors with
high probability. Second, the predicted threshold by Theorem 5 is based on the
pessimistic assumption that a cluster can only correct a single error. Third, for a
e and ρe, as
e constructed randomly according to given degree distributions λ
graph G,
the graph size grows the decision subgraph becomes a tree with probability close
to 1. Hence, it can be shown (see, for example Richardson and Urbanke, 2008)
that the recall performance for any such graphs will be concentrated around the
average case given by Theorem 5.
6
Pattern Retrieval Capacity
Before discussing the the pattern retrieval capacity, we should note that the number of patterns C does not have any effect on the learning or recall algorithm except for its obvious influence on the learning time. More precisely, as long as the
17
Figure 7: The decision subgraph of depth 2 for the third edge (from left) in Figure 6.
(sub)patterns lie on a subspace, the learning Algorithm 1 yields a matrix W that
is orthogonal to all the patterns of the training set. Similarly, in the recall phase,
algorithms 2 and 3 only need to compute W · e for the noise vector e.
Remember that the retrieval capacity is defined as the maximum number of
patterns that a neural of size n is able to store. Hence, in order to show that the
pattern retrieval capacity of our method is exponential in n, we need to demonstrate
that there exists a training set X with C patterns of length n for which C ∝ arn ,
for some a > 1 and 0 < r < 1.
Theorem 6 Let X be a C × n matrix formed by C vectors of length n with entries
from Q. Furthermore, let k = brnc for some 0 < r < 1 and k < min` (n` ).
Then, there exists a set of vectors of size C = Ω(arn ) with some a > 1 such that
rank(X ) = k < n. Moreover, Algorithm 1 can learn this set.
The proof is by construction. This construction can be used to synthetically generate patterns that lie in a subspace.
7
Experimental Results
In this section we evaluate the performance of our proposed algorithms over synthetic and natural datasets 5 .
5
The codes used in this paper are all available online at http://goo.gl/ifR14t.
18
7.1
Synthetic Scenario
A systematic way to generate patterns satisfying a set of linear constraints is outlined in the proof of Theorem 6. This proof is constructive and provides an easy
way to randomly sample patterns with linear constraints. In our simulations, we
consider a neural network in which each pattern neuron is connected to approximately 5 clusters. The number of connections should be neither too small (to ensure
information propagation) nor too big (to adhere to the sparsity requirement).
In the learning phase, Algorithm 1 is performed (in parallel) for each cluster in
order to find the connectivity matrix W . In the recall phase (and at each round),
a pattern x is sampled uniformly at random from the training set. Then, each of
its entries are corrupted with ±1 additive noise independently with probability pe .
Algorithm 3 is subsequently used to denoise the corrupted patterns. We average
out this process over many trials to calculate the error rate and compare it to the
analytic bound derived in Theorem 5.
7.1.1
Learning Results
The left and right panels in Figure 8 illustrate the degree distributions of pattern
and constraint neurons, respectively, over an ensemble of 5 randomly generated
datasets. The network size is n = 400, which is divided into 50 overlapping clusters, each of size around 40, i.e., n` ' 40 for ` = 1, . . . , 50. Each pattern neuron
is connected to 5 clusters, on average. The horizontal axis shows the normalized degree of pattern (resp., constraint) neurons and the vertical axis represents
the fraction of neurons with the given normalized degree. The normalization is
done with respect to the number of pattern (resp., constrain) neurons in the cluster. The parameters for the learning algorithm are αt ∝ 0.95/t, η = 0.75/αt and
θt ∝ 0.05/t.
Figure 9 illustrates the same results for a network of size n = 960, which is
divided into 60 clusters, each with size 80, on average. The learning parameters
are the same as before, i.e., αt ∝ 0.95/t, η = 0.75/αt and θt ∝ 0.05/t, and
each pattern neuron is connected to 5 clusters on average. Note that the overall
normalized degrees are smaller compared to the case of n = 400, which indicates
sparser clusters on average. In almost all cases that we have tried, the learning
phase converges within two learning iterations, i.e., by going over the data set only
twice.
7.1.2
Recall Results
Figure 10 illustrates the performance of the recall algorithm. The horizontal and
vertical axes represent the average fraction of erroneous neurons and the final Pat19
0.3
Degree distribution
Degree distribution
0.3
0.2
0.1
0
0.2
0.1
0
0
0.2 0.4 0.6 0.8
Normalized degree
1
0
(a) Pattern neuron degrees
0.2 0.4 0.6 0.8
Normalized degree
(b) Constraint neuron degrees
Figure 8: Pattern and constraint neuron degree distributions for n = 400, L = 50,
and an average of 20 constraints per cluster. The learning parameters are αt ∝
0.95/t, η = 0.75/αt and θt ∝ 0.05/t.
tern Error Rate (PER), respectively. The performance is compared against the theoretical bound derived in Theorem 5 as well as the the two constructions proposed
by Kumar et al. (2011) and Salavati and Karbasi (2012). The parameters used for
this simulation are n = 400, L = 50 and ϕ = 0.82. For the non-overlapping
clusters approach proposed by Salavati and Karbasi (2012), the network size is
n = 400 with 4 clusters in the first level and one cluster in the second level (identical to their simulations). The convolutional neural network proposed in this paper
clearly outperforms the prior art. Note that for the theoretical estimates used in
Figure 10, we both calculated the probability of correcting a single error by each
cluster Pc (via the lower bound in Theorem 3), and by fixing it to Pc = 1. The
corresponding curves in Figure 10 show that the later estimate is tighter, i.e., when
each cluster can correct a single error with probability close to 1.
Figure 11 shows the final PER for the network with n = 960 and L = 60
clusters. Comparing the PER with that of a network with n = 400 neurons and
L = 50 clusters, we witness a degraded performance. At first glance this might
seem surprising as we increased both the network size and the number of clusters.
However, the key point in determining the performance of Algorithm 3 is not the
number of clusters but rather the size of the clusters and the cluster nodes degree
distribution ρe(x). In the network with n = 960, we have around 80 pattern neurons
per cluster, while in the network with n = 400 we have around n = 40 neurons
20
1
0.3
Degree distribution
Degree distribution
0.3
0.2
0.1
0
0.2
0.1
0
0
0.2 0.4 0.6 0.8
Normalized degree
1
0
(a) Pattern neurons degrees
0.2 0.4 0.6 0.8
Normalized degree
(b) Constraint neurons degrees
Figure 9: Pattern and constraint neuron degree distributions for n = 960, L = 60,
and an average of 40 constraints per cluster. The learning parameters are αt ∝
0.95/t, η = 0.75/αt and θt ∝ 0.05/t.
per cluster. Clearly, by increasing the network size without increasing the number of clusters, the chance of a cluster experiencing more than one error increases
(remember, each cluster can correct a single error). This in turn results in an inferior performance. Hence, increasing the network size helps only if the number of
clusters are increased correspondingly.
7.2
Real Datasets
So far, we have tested our proposed method over synthetic datasets where we generated patterns in such a way that they all belong to a subspace. In many real
datasets (e.g., images and natural sounds), however, patterns rarely form a subspace. Rather, due to their common structures, they come very close to forming
one. The focus of this section is to show how our proposed method can be adapted
to such scenarios.
More specifically, let X denote a dataset of C patterns of length n. Here we
assume that patterns are all vectorized and form the rows of the matrix X . The
eigenvalues of the correlation matrix A = X > X indicate how close the patterns are
to a subspace. Note that A is a positive semidefinite matrix, so all eigenvalues are
non-negative. In particular, if we have an eigenvalue 0 with positive multiplicity,
then the patterns belong to a subspace. Similarly, if we have a set of eigenvalues
21
1
Final pattern error rate
1
0.8
0.6
0.4
No clustering†
Non-overlapping clusters‡
Convolutional-Simulation
Convolutional-Theory
Convolutional-Theory (Pc = 1)
0.2
0
0.00
†:
0.03
0.05
0.07
0.10
pe
0.12
0.15
0.17
0.20
0.22
(Kumar et al., 2011), ‡ : (Salavati and Karbasi, 2012)
Figure 10: Recall error rate along with theoretical bounds for different architectures
of network with n = 400 pattern neurons and L = 50 clusters. We compare the
performance of our method with two other constructions where either no notion of
cluster was considered (Kumar et al., 2011) or no overlaps between clusters was
assumed (Salavati and Karbasi, 2012).
22
Final pattern error rate
1
0.8
0.6
n = 400-Simulation
n = 400-Theory
n = 960-Simulation
n = 960-Theory
0.4
0.2
0
0.00
0.05
0.10
pe
0.15
0.20
0.25
Figure 11: Recall error rate and the theoretical bounds for different architectures
of network with n = 960 and n = 400 pattern neurons and L = 60 and L = 50
clusters, respectively..
all close to zero, then the patterns are close to a subspace of the n-dimensional
space. Figure 12 illustrates the eigenvalue distribution of the correlation matrix for
a dataset of C = 10000 gray-scale images of size 32×32, sampled from 10 classes
of the CIFAR-10 dataset (Krizhevsky and Hinton, 2009). Each image is quantized
to 16 levels. Based on our notation, n = 1024 and Q = 16. As evident from the
figure, almost half of the 1024 eigenvalues are less than 0.001, suggesting that the
patterns are very close to a subspace.
7.2.1
Simulation Scenario
In order to adapt our method to this new scenario where patterns approximately
belong to a subspace, we need to slightly modify the learning and the recall algorithms. We use CIFAR-10 datase as the running example, however, the principles
described below can be easily applied to other datasets.
To start, we first alter the way patterns are represented in such a way that is
makes them easier to learn for our algorithm. More specifically, since the images
are quantized to 16 levels, we can represent a 16-level pixel with 4 bits. As such,
instead of having 1024 integer-valued pattern neurons to represent the patterns in
the dataset, we will have 4096 binary pattern neurons. We adopt this modified
description as it facilitates the learning process.
We then apply Algorithm 1 as before to learn the patterns in the dataset. Obviously, since the patterns do not exactly form a subspace, we cannot expect the
23
101
Eigenvalue
10−1
10−3
10−5
0
100
200
300
400 500 600 700
Eigenvalue’s index
800
900 1,000
Figure 12: The eigenvalues of a dataset with 1000 gray-scale images of size 32 ×
32, uniformly sampled from 10 classes of the CIFAR-10 dataset (Krizhevsky and
Hinton, 2009).
algorithm to finish with a weight vector w that is orthogonal to all the patterns.
Nevertheless, by applying the learning algorithm, we will have a weight vector
whose projection on the patterns is rather small. Following the same procedure
as before, we obtain L neural graphs, W (1) , . . . , W (L) , for each of the L clusters.
Note again that W (i) ’s are approximately (rather than exactly) orthogonal to the
(sub-)patterns.
Our main observation is the following. We can interpret the deviation from
the subspace as noise. Consequently, if we apply the intra-cluster recall method
(Algorithm 2) to the patterns, we can find out what the network has actually learned
in response to original patterns from the dataset X . In other words, Algorithm
2 identifies the projection of original patterns to a subspace X 0 . Hence, all the
learned patterns are orthogonal to the connectivity matrix W . This idea is shown in
figure 13, where we have the original image (left), the quantized version (middle),
and the image learned by the proposed algorithm. It is worth observing that what
the network has learned focuses more on the actual objects rather than unnecessary
details.
For the recall phase, the approach is similar to before: we are given a set of
noisy patterns and the goal is to retrieve the correct versions. In our simulations,
we assume that the noise is added to the learned patterns (see above) but we can
also consider the situation where noise is added to the quantized patterns.
24
(a) Original image
(b) Quantized image
(c) Learned image
Figure 13: Original vs. learned images
7.2.2
Learning Results
Figure 14 illustrates the average cost (defined by E(w(`) (t)) in Section 4.1) for
learning one constraint vector versus the number of iterations. In this example, the
learning parameters are α0 = 0.95/t, η = 1 and θt = 0.01/t. The considered
neural netowrk has n = 4096 pattern neurons and L = 401 clusters of size 100.
The learning process terminates if a) an orthogonal weight vector is found or b)
200 iterations is done.
Figure15 illustrates the required number of iterations of Algorithm 1 so that a
weight vector orthogonal the patterns in the dataset X 0 is obtained. As we see from
the figure, in the majority of the cases, one pass over the dataset is enough. As
before, we have α0 = 0.95/t, η = 1 and θt = 0.01/t.
Furthermore, we have also uploaded a short video clip of the learning algorithm
in action (i.e., iteration by iteration) for a few sample images from the dataset. The
clip is available through the following link: http://goo.gl/evcNOh.
7.2.3
Recall Results
Figures 16 and 17 show the recall error rate as we increase the noise level for the
neural network used in Sec 7.2.2. The update thresholds for the recall algorithm are
set to ϕ = 0.85 and ψ = 0.005. The inter-module recall procedure (Algorithm 3)
is performed at most 80 times and the corresponding error rates are calculated by
evaluating the difference between the final state of pattern neurons (after running
25
Learning cost
0.006
0.004
0.002
0.000
0
3
6
9
12
15
18
21
24
27
t
Number of converged vectors
Figure 14: Average cost versus time for learning a weight vector in a network with
n = 4096 pattern neurons and L = 401 clusters. The size of clusters is set to 100
with (around) 50 constraints in each cluster.
15,000
10,000
5,000
0
0
15
30
45
60
75
Required number of iterations
90
Figure 15: Number of the iterations required for Algortihm 1 to learn a vector
orthogonal to the patterns in the dataset X 0 , in a network with n = 4096 pattern
neurons and L = 401 clusters of size 100 and (around) 50 constraints in each
cluster.
26
105
Final pattern error rate
1
0.8
0.6
0.4
C = 5000
C = 10000
C = 50000
0.2
0
0.000
0.002
0.004
pe
0.006
0.008
Figure 16: Recall error rate for a network with n = 4096 pattern neurons and
L = 401 clusters, with cluster size equal to n` = 100. The proposed recal method
was applied to a dataset of 10000 images sampled from the CIFAR-10 database
Algorithm 3) and the noise-free patterns in the dataset X 0 .
Figure 18 illustrates a few instances of the recalled images. In this figure, we
have the original images (first column), the learned images (second column), the
noisy versions (third column), and the recalled images (forth column). The figure
also shows the input and the output Signal to Noise Ratios (SNR) for each example.
Note that in all examples the SNR increases as we apply our recall algorithm.
For this example we chose ϕ = 0.95 and ψ = 0.025. We have also uploaded a
short video clip of the recall algorithm in action for a few sample images from the
dataset, which can be found on http://goo.gl/EHJfds.
8
Analysis
This section contains the proofs of all theorems and technical lemmas we used in
the paper.
8.1
Proof of Lemma 1
We proceed by induction. To this end, assume kw(`) (t)k2 > 0 and let
!
(`)
(`)
ẁ (t) = w (t) − αt y
(`)
y (`) (t)w(`) (t)
(t) x (t) −
.
kw(`) (t)k22
(`)
27
Symbol error rate
0.100
Input BER
C = 5000-Simul.
C = 10000-Simul.
C = 50000-Simul.
0.050
0.000
0.000
0.020
0.040
0.060
0.080
0.100
0.120
pe
Figure 17: Symbol error rates for a network with n = 4096 pattern neurons and
L = 401 clusters, applied to a dataset of 10000 images sampled from the CIFAR10 database.
Note that
kẁ(`) (t)k22 = kw(`) (t)k22 +αt2 y (`) (t)2 kx(`) (t)−
y (`) (t)w(`) (t) 2
k ≥ kw(`) (t)k22 > 0.
kw(`) (t)k22 2
Now,
kw(`) (t + 1)k22 = kẁ(`) (t)k2 + αt2 η 2 kΓ(w(`) (t), θt )k2 − 2αt ηhΓ(w(`) (t), θt ), ẁ(`) (t)i
≥ kẁ(`) (t)k22 + αt2 η 2 kΓ(w(`) (t), θt )k2 − 2αt ηkΓ(w(`) (t), θt )k2 kẁ(`) (t)k2
=
kẁ(`) (t)k2 − αt ηkΓ(w(`) (t), θt )k2
2
Thus, in order to have kw(`) (t + 1)k2 > 0, we must have that
kẁ(`) (t)k2 − αt ηkΓ(w(`) (t), θt )k2 > 0.
Given that
kΓ(w(`) (t), θt )k2 ≤ kw(t)k2 ≤ kẁ(`) (t)k2 ,
it is sufficient to have αt η < 1 in order to achieve the desired inequality. This
proves the lemma.
8.2
Proof of Theorem 2
Let us define the correlation matrix for the sub-patterns that lie within the domain
of cluster ` as follows
A(`) := E{x(`) (x(`) )T |x ∈ X }.
28
Also, let us define
(`)
At := x(`) (t)(x(`) (t))> .
(`)
Hence, we have A(`) = E(At ). Furthermore, recall the learning cost function
C
2
1 X
hw(`) (t), xµ i .
C µ=1
E(t) = E(w(`) (t)) =
From Alg. 1 we have
(`)
(`)
w (t+1) = w (t)−αt y
(`)
y (`) (t)w(`) (t)
(t) x (t) −
kw(`) (t)k22
!
(`)
!
(`)
+ ηΓ(w (t), θt ) .
Let
Y (`) (t) = Ex (X (`) w(`) (t)),
where Ex (·) is the expectation over the choice of pattern x(t) and X (`) is the matrix
of all the sub-patterns corresponding to cluster ` in the dataset. Thus, we will have
Y (`) (t + 1) = Y (`) (t) 1 + αt
w(`) (t)
Noting that E(t) =
1
(`) (t)k2 ,
2
C kY
E(t + 1) = E(t) 1 + αt
+
>
A(`) w(`) (t)
kw(`) (t)k22
(`) (`) (`)
(`)
(`)
− αt X A w (t) + ηX Γ(w (t), θt ) .
we obtain
>
w(`) (t)
2
A(`) w(`) (t)
kw(`) (t)k22
αt2
kX (`) A(`) w(`) (t) + ηX (`) Γ(w(`) (t), θt )k22
C
w(`) (t)
w(`) (t)
− 2αt 1 + αt
− 2αt 1 + αt
>
>
2
A(`) w(t)
(`)
(`)
(`)
w (t)
A
w (t)
kw(`) (t)k22
>
>
A(`) w(`) (t)
(`)
(`)
(`)
η w (t) A Γ(w (t), θt )
kw(`) (t)k22
29
By omitting all the second order terms O(αt2 ), we obtain
>
w(`) (t)
E(t + 1) ' E(t) 1 + 2αt
− 2αt
w(`) (t)
A(`) w(`) (t)
kw(`) (t)k22
>
A(`)
2
>
w(`) (t) + η w(`) (t)
A(`) Γ(w(`) (t), θt )
>
(`) (t)
(`) w (`) (t)
2
>
w
A
A(`) w(`) (t) −
E(t) − 2αt w(`) (t)
E(t)
2
(`)
kw (t)k2
>
=
A(`) Γ(w(`) (t), θt )
− 2αt η w(`) (t)
(7)
Note that
αt ηk w(`) (t)
>
A(`) Γ(w(`) (t), θt )k2 ≤ αt ηkw(`) (t)k2 kA(`) k2 kΓ(w(`) (t), θt )k2
√
≤ αt ηkw(`) (t)k2 kA(`) k2 ( nθt ).
Since we have θt = Θ(αt ) and that
>
αt ηk w(`) (t)
A(`) Γ(w(`) (t), θt )k2 = O(αt2 )
we can further simplify (7) as follows
E(t + 1) ' E(t) − 2αt w(`) (t)
>
A(`)
2
>
w(`) (t)
w(t) −
A(`) w(`) (t)
kw(`) (t)k22
Thus, in order to show that the algorithm converges, we need to show that
(`)
w (t)
>
A(`)
2
w(`) (t) −
>
w(`) (t)
A(`) w(`) (t)
kw(`) (t)k22
E(t) ≥ 0
which in turn implies E(t + 1) ≤ E(t). By noting that
E(t) = w(`) (t)
>
A(`) w(`) (t),
we must show that
w
(`)
>
A
(`)
2
w
(`)
≥
w
30
(`)
>
(`)
A w
(`)
2
/kw(`) k22 .
E(t) .
The left hand side is kA(`) w(`) k22 . For the right hand side, we have
k w(`)
>
A(`) w(`) k22
≤
kw(`) k22
kw(`) k22 kA(`) w(`) k22
= kA(`) w(`) k22 .
2
(`)
kw k2
The above inequality shows that E(t + 1) ≤ E(t), which readily implies that for
sufficiently large number of iterations, the algorithm converges to a local minimum
ŵ(`) where E(ŵ(`) ) = 0. From Lemma 1 we know that kŵ(`) k2 > 0. Thus, the
only solution for E(ŵ(`) ) = kX (`) ŵ(`) k22 = 0 is for ŵ(`) to be orthogonal to the
patterns in the data set.
8.3
Proof of Theorem 3
In the case of a single error, we can easily show that the noisy pattern neuron will
always be updated towards the correct direction in Algorithm 2. For simplicity,
let’s assume the first pattern neuron of cluster ` is the noisy one. Furthermore, let
z (`) = [1, 0, . . . , 0] be the noise vector. Denoting the ith column of the weight
(`)
matrix by Wi , we will have
(`)
(`)
y (`) = sign(z1 W1 ) = z1 sign(W1 ).
(`)
Hence, in Algorithm 2 we obtain g1 = 1 > ϕ. This means that the noisy node
gets updated towards the correct direction.
Therefore, the only source of error would be a correct pattern neuron getting
(`)
updated mistakenly. Let Pi denote the probability that a correct pattern neuron xi
(`)
gets updated. This happens if |gi | > ϕ. For ϕ → 1, this is equivalent to having
(`)
(`)
(`)
hWi , sign(z1 W1 )i = kWi k0 .
(`)
However, in cases where the neighborhood of xi is different from the neighborhood of x1 among the constraint nodes we have
(`)
(`)
(`)
hWi , sign(W1 )i < kWi k0 .
(`)
(`)
More specifically, let N (xi ) indicate the set of neighbors of xi among constraint
neurons in cluster `. Then in the case where
(`)
(`)
(`)
N (xi ) ∩ N (x1 ) 6= N (xi ),
(`)
(`)
there are non-zero entries in Wi while W1 is zero, and vice-versa. Therefore,
(`)
(`)
(`)
by letting Pi0 to be the probability of N (xi ) ∩ N (x1 ) = N (xi ), we note that
Pi ≤ Pi0 .
31
The above inequality help us obtain an upper bound on Pi , by bound Pi0 . Since
there is only one noisy neuron x1 , we know that, on average, this node is connected
to d¯` constraint neurons which implies that the probability of xi and x1 sharing
exactly the same neighborhood is:
Pi0
=
d¯`
m`
!di
,
where di is the degree of neuron xi . By taking the average over the pattern neurons, we obtain the following bound on the average probability of a correct pattern
neuron being mistakenly updated:
Pe0
X (`)
X (`)
Λdi
Λdi Pi0 =
=
di
di
d¯`
m`
!di
(`)
=Λ
d¯`
m`
!
,
(`) i
i Λi x
is the degree distribution polynomial. Therefore, the
where Λ(`) (x) =
(`)
probability of correcting one noisy input is lower bounded by Pc ≥ (1 − Pe0 )n` −1 ,
i.e.,
!!n` −1
d¯`
(`)
(`)
.
Pc ≥ 1 − Λ
m`
P
This proves the theorem.
8.4
Proof of Lemma 4
Without loss of generality, suppose the first pattern neuron is contaminated by an
external error +1, i.e., z (`) = [1, 0, . . . , 0]. As a result
(`)
y (`) = sign W (`) (x(`) + z (`) ) = sign W x(`) + W z (`) = sign W (`) z (`) = sign W1
(`)
where Wi is the ith column of W (`) . Hence, the feedback transmitted by the
(`)
constraint neurons is sign(W1 ). As a result, decision parameters of pattern neuron
(`)
i, i.e., gi in Algorithm 2, will be
(`)
gi
=
(`)
(`)
(`)
(`)
hsign(W1 ), Wi i
.
hsign(Wi ), Wi i
(`)
(`)
(`)
Note that the denominator is simply kWi k0 = hsign(Wi ), Wi i. By assumption, no two pattern neurons in G(`) share the exact same set of neighbors. Therefore, for all i, j ∈ {1, . . . n` } such that i 6= j, there is at least a non-zero entry, say
(`)
(`)
(`)
(`)
k, in Wj for which Wik = 0. Thus, we have gi = 1 if i = 1 and gi < 1 if
i > 1 As a result, for ϕ → 1, only the first neuron (i.e., the noisy one) will update
its value towards the correct state.
32
,
8.5
Proof of Theorem 5
The proof is, in spirit, similar to Theorem 3.50 of (Richardson and Urbanke, 2008).
Consider a message transmitted over an edge from a given cluster node v (`) to a
given noisy pattern neuron at iteration t of Algorithm 3. This message will be a
failure, indicating that the super constraint node being unable to correct the error,
if
1. the super constraint node v (`) receives at least one error message from its
other neighbors among pattern neurons. This event happens it is connected
to more than one noisy pattern neuron.
2. the super constraint node v (`) does not receive an error message from any of
its other neighbors but is unable to correct the single error in the given noisy
neuron. This event happens with probability 1 − Pc .
Let us denote the probability of the above failure message by π (`) (t) and the average probability that a pattern neuron sends an erroneous message to a neighboring
cluster node by z(t). Then, we have
π (`) (t) = 1 − Pc (1 − z(t))d` −1 ,
e
where de` is the degree of the super constraint neuron v (`) in the contracted graph
e Similarly, let π(t) denote the average probability that a super constraint node
G.
sends a message declaring the violation of at least one of its constraint neurons.
Then we have
π(t) = Ede (π (`) (t)) =
`
X
ρei (1 − Pc (1 − z(t))d` −1 ) = 1 − Pc ρe(1 − z(t)).
e
i
Now consider the message transmitted from a given pattern neuron xi with degree
di to a given super constraint node v (`) in iteration t + 1 of Algorithm 3. This
message will indicate a noisy pattern neuron if the pattern neuron was noisy in the
first place (with probability pe ) and all of its other neighbors among super constraint nodes has sent a violation message in iteration t. Therefore, the probability
of this node being noisy will be z(0)π(t)di −1 where z(0) = pe . Hence, the average
probability that a pattern neuron remains noisy at (t + 1)-th iteration is
z(t + 1) = pe
X
e i π(t)i−1 = pe · λ(π(t))
e
e − Pc ρe(1 − z(t))).
λ
= pe · λ(1
i
Note that the denoising operation will be successful if z(t + 1) < z(t), ∀t. Theree − Pc ρe(1 − z)) < z for
fore, we must look for the maximum pe such that pe λ(1
z ∈ [0, pe ].
33
8.6
Proof of Theorem 6
The proof is based on construction: we build a data set X with the required properties such that it can be memorized by the proposed neural network.
Consider a matrix G ∈ Rk×n with rank k = rn where 0 < r < 1 is chosen
such that k < min` (n` ). Let the entries of G be non-negative integers between 0
and γ − 1. Here we assume that γ ≥ 2.
We start constructing the patterns in the data set as follows. We pick a random
vector u ∈ Rk with integer-valued-entries between 0 and υ − 1 where υ ≥ 2.
We set the pattern x ∈ X to be x = G> u if all the entries of x are between 0
and Q − 1. Since both u and G have only non-negative entries, all entries in x
are non-negative. However, we need to design G such that all entries of G> u be
less than Q. Let gj be the j-th column of G. Then, the j-th entry of x is equal to
xj = hu, gj i. Therefore,
xj = u> gj ≤ dj (γ − 1)(υ − 1)
Let d∗ = minj dj . We can choose γ, υ and d∗ such that
Q − 1 ≥ d∗ (γ − 1)(υ − 1)
which in turn ensures that all entries of x are less than Q. Furthermore, we have
selected k in such a way that k < min` (n` ). As a result we are sure that the
set of sub-patterns of the dataset X form a subspace with dimension k in an n` dimensional space. Since there are υ k vectors u with integer entries between 0 and
υ −1, we have υ k = υ rn patterns forming X . This implies that the storage capacity
C = υ rn is an exponential number in n as long as υ ≥ 2.
9
Conclusions and Final Remarks
In this paper, we proposed the first neural network structure that learns an exponential number of patterns (in the size of the network) and corrects up to a linear
fraction of errors. The main observation we made was that natural patterns seem to
have inherent redundancy and we proposed a framework to captured redundancies
that appear in the form of linear (or close to linear) constraints. Our experimental
results also reveal that our learning algorithm can be seen as a feature extraction
method, tailored for patterns with such constraints. Extending this line of thought
through more sophisticated feature extraction approaches, and in light of recent
developments in deep belief networks (Jarrett et al., 2009; Coates and Ng, 2011;
Le et al., 2010; Vincent et al., 2008; Ngiam et al., 2011), is an interesting future
direction to pursue.
34
References
Emmanuel J. Candès and Terence Tao. Near-optimal signal recovery from random projections: Universal encoding strategies? IEEE Transactions on Information Theory, 52
(12):5406–5425, 2006.
Adam Coates and Andrew Ng. The importance of encoding versus training with sparse
coding and vector quantization. In Lise Getoor and Tobias Scheffer, editors, International Conference on Machine Learning (ICML), pages 921–928, New York, NY, USA,
2011. ACM.
D. L. Donoho. Compressed sensing. IEEE Transactions on Information Theory, 52(4):
1289–1306, 2006.
David L. Donoho, A. Maleki, and A. Montanari. Message-passing algorithms for compressed sensing. Proceedings of the National Academy of Sciences (PNAS), 106(45):
18914–18919, 2009.
Vincent Gripon and Claude Berrou. Sparse neural networks with large learning diversity.
IEEE Transactions on Neural Networks, 22(7):1087–1096, 2011.
Donald O. Hebb. The Organization of Behavior: A Neuropsychological Theory. Wiley &
Sons, New York, 1949.
John J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. U.S.A., 79(8):2554–2558, 1982.
Stanislaw Jankowski, Andrzej Lozowski, and Jacek M. Zurada. Complex-valued multistate neural associative memory. IEEE Transactions on Neural Networks and Learning
Systems, 7(6):1491–1496, 1996.
Kevin Jarrett, Koray Kavukcuoglu, Marc’Aurelio Ranzato, and Yann LeCun. What is the
best multi-stage architecture for object recognition? In IEEE International Conference
on Computer Vision (ICCV), pages 2146–2153, 2009.
Amin Karbasi, Amir Hesam Salavati, and Amin Shokrollahi. Iterative learning and denoising in convolutional neural associative memories. In International Conference on
Machine Learning (ICML), pages 445–453, 2013.
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny
images. Master’s thesis, Department of Computer Science, University of Toronto, 2009.
K. Raj Kumar, Amir Hesam Salavati, and Mohammad A. Shokrollah. Exponential pattern
retrieval capacity with non-binary associative memory. In IEEE Information Theory
Workshop (ITW), pages 80–84, 2011.
K. Raj Kumar, Amir Hesam Salavati, and Mohammad A. Shokrollahi. A non-binary associative memory with exponential pattern retrieval capacity and iterative learning. IEEE
Transaction on Neural Networks and Learning systems, 2014. To appear in.
35
Quoc V. Le, Jiquan Ngiam, Zhenghao Chen, Daniel Jin hao Chia, Pang Wei Koh, and
Andrew Y. Ng. Tiled convolutional neural networks. In Advances in Neural Information
Processing Systems (NIPS), pages 1279–1287, 2010.
Donq-Liang Lee. Improvements of complex-valued Hopfield associative memory by using
generalized projection rules. IEEE Transactions on Neural Networks, 17(5):1341–1347,
2006.
Michael Luby, Michael Mitzenmacher, Mohammad Amin Shokrollahi, and Daniel A.
Spielman. Efficient erasure correcting codes. IEEE Transactions on Information Theory, 47(2):569–584, 2001.
Robert J. McEliece, Edward C. Posner, Eugene R. Rodemich, and Santosh S. Venkatesh.
The capacity of the hopfield associative memory. IEEE Transactions on Information
Theory, 33(4):461–482, 1987.
Mehmet Kerem Muezzinoglu, Cuneyt Guzelis, and Jacek M. Zurada. A new design method
for the complex-valued multistate hopfield associative memory. IEEE Transactions on
Neural Networks, 14(4):891–899, 2003.
Jiquan Ngiam, Pang Wei Koh, Zhenghao Chen, Sonia A. Bhaskar, and Andrew Y. Ng.
Sparse filtering. In John Shawe-Taylor, Richard S. Zemel, Peter L. Bartlett, Fernando
C. N. Pereira, and Kilian Q. Weinberger, editors, Advances in Neural Information Processing Systems (NIPS), pages 1125–1133, 2011.
Erkki Oja and Juha Karhunen. On stochastic approximation of the eigenvectors and eigenvalues of the expectation of a random matrix. Math. Analysis and Applications, 106:
69–84, 1985.
Erkki Oja and Teuvo Kohonen. The subspace learning algorithm as a formalism for pattern recognition and neural networks. In IEEE International Conference on Neural
Networks, volume 1, pages 277–284, 1988.
P. Peretto and J. J. Niez. Long term memory storage capacity of multiconnected neural
networks. Biological Cybernetics, 54(1):53–64, 1986.
Tom Richardson and Ruediger Urbanke. Modern Coding Theory. Cambridge University
Press, New York, NY, USA, 2008.
Amir Hesam Salavati. Coding theory and neural associative memories with exponential pattern retrieval capacity. PhD thesis, Ecole Polytechnique federale de Lausanne (EPFL), 2014. URL http://algo.epfl.ch/~amir/PhD_Thesis_
Salavati.pdf.
Amir Hesam Salavati and Amin Karbasi. Multi-level error-resilient neural networks. In
IEEE International Symposium on Information Theory (ISIT), pages 1064–1068, 2012.
36
Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. Semi-supervised recursive autoencoders for predicting sentiment distributions. In
Proceedings of the Conference on Empirical Methods in Natural Language Processing,
pages 151–161. Association for Computational Linguistics, 2011.
R. Tanner. A recursive approach to low complexity codes. IEEE Transactions on Information Theory, 27(5):533–547, 1981.
Joel A. Tropp and Stephen J. Wright. Computational methods for sparse solution of linear
inverse problems. Proceedings of the IEEE, 98(6):948–958, 2010.
Santosh S Venkatesh. Connectivity versus capacity in the hebb rule. In Theoretical Advances in Neural Computation and Learning, pages 173–240. Springer, 1994.
Santosh S. Venkatesh and Demetri Psaltis. Linear and logarithmic capacities in associative
neural networks. IEEE Transactions on Information Theory, 35(3):558–568, 1989.
Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In International Conference on Machine Learning (ICML), pages 1096–1103, New York, NY, USA, 2008.
ACM.
Lei Xu, Adam Krzyzak, and Erkki Oja. Neural nets for dual subspace pattern recognition
method. International Journal of Neural Systems, 2(3):169–184, 1991.
37
Algorithm 1 Iterative Learning
Input: Dataset X with |X | = C, stopping point ε.
Output: w(`)
1 P
(`)
(`)
2
1: while C
x∈X |hx (t), w (t)i| > ε do
2:
Choose pattern x(t) uniformly at from X .
3:
Compute y (`) (t) = hx(`) (t), w(`) (t)i.
4:
Update w(`) (t) as follows
w(`) (t) = w(`) (t−1)−αt
5:
6:
y (`) (t)w(`) (t − 1)
y (`) (t) x(`) (t) −
kw(`) (t − 1)k22
!
!
+ ηΓ(w(`) (t − 1), θt )
t ← t + 1.
end while
Algorithm 2 Intra-cluster Error Correction
Input: Training set X , threshold ϕ, iteration tmax
(`) (`)
(`)
Output: x1 , x2 , . . . , xn`
1: for t = 1 → tmax do
(`)
2:
Forward iteration:
Calculate the weighted input sum hi
Pn`
(`) (`)
(`)
(`)
(`) 4
j=1 Wij xj , for each neuron yi and set yi = sign(hi ).
3:
(`)
Backward iteration: Each neuron xj computes
(`)
gj
4:
(`) (`)
i=1 Wij yi
.
Pm`
(`)
i=1 |Wij |
Pm`
=
Update the state of each pattern neuron j according to
(`)
(`)
(`)
xj = xj − sign(gj )
(`)
only if |gj | > ϕ.
5:
t←t+1
6: end for
38
=
Algorithm 3 Sequential Peeling Algorithm
e G(1) , G(2) , . . . , G(L) .
Input: G,
Output: x1 , x2 , . . . , xn
1: while There is an unsatisfied v (`) , for ` = 1, . . . , L do
2:
for ` = 1 → L do
3:
If v (`) is unsatisfied, apply Algorithm 2 to cluster G(l) .
4:
If v (`) remained unsatisfied, revert the state of pattern neurons connected
to v (`) to their initial state. Otherwise, keep their current states.
5:
end for
6: end while
7: Declare x1 , x2 , . . . , xn if all v (`) ’s are satisfied. Otherwise, declare failure.
39
Learned (c)
(a)
Original (b)
SNRI : 2.10
image
image
(d) SNRO
2.84
:
(a)
Original (b)
Learned (c)
SNRI : 2.15
image
image
(d) SNRO
2.93
:
(a)
Original (b)
Learned (c)
SNRI : 2.07
image
image
(d) SNRO
3.16
:
(a)
Original (b)
Learned (c)
SNRI : 2.09
image
image
(d) SNRO
2.96
:
(a)
Original (b)
Learned (c)
SNRI : 2.08
image
image
(d) SNRO
2.81
:
Figure 18: Examples of the learning and recall phase for images sampled from
CIFAR-10. Here, SNRI and SNRO denote the input and output SNR’s, respectively.
40
| 9cs.NE
|
1
DNA-GA: A New Approach of Network
Performance Analysis
arXiv:1512.05429v2 [cs.IT] 26 Sep 2017
Ming Ding, Data 61, Australia {[email protected]}
David López Pérez, Bell Labs Alcatel-Lucent, Ireland {[email protected]}
Guoqiang Mao, The University of Technology Sydney and Data 61, Australia {[email protected]}
Zihuai Lin, The University of Sydney, Australia {[email protected]}
Abstract—In this paper, we propose a new approach of
network performance analysis, which is based on our previous
works on the deterministic network analysis using the Gaussian approximation (DNA-GA). First, we extend our previous
works to a signal-to-interference ratio (SIR) analysis, which
makes our DNA-GA analysis a formal microscopic analysis tool.
Second, we show two approaches for upgrading the DNA-GA
analysis to a macroscopic analysis tool. Finally, we perform a
comparison between the proposed DNA-GA analysis and the
existing macroscopic analysis based on stochastic geometry. Our
results show that the DNA-GA analysis possesses a few special
features: (i) shadow fading is naturally considered in the DNAGA analysis; (ii) the DNA-GA analysis can handle non-uniform
user distributions and any type of multi-path fading; (iii) the
shape and/or the size of cell coverage areas in the DNA-GA
analysis can be made arbitrary for the treatment of hotspot
network scenarios. Thus, DNA-GA analysis is very useful for
the network performance analysis of the 5th generation (5G)
systems with general cell deployment and user distribution, both
on a microscopic level and on a macroscopic level. 1
I. I NTRODUCTION
Due to their potential for large performance gains, dense
orthogonal deployments of small cell networks (SCNs) within
the existing macrocell networks2 gained much momentum in
the design of the 4th generation (4G) systems [1], and are
envisaged as the workhorse for capacity enhancement in the
5th generation (5G) systems [2]. In this context, new and
more powerful network performance analysis tools are needed
to better understand the performance implications that these
dense orthogonal SCNs bring about [3-10].
Network performance analysis tools can be broadly classified into two groups, i.e., macroscopic analysis [3,4] and
microscopic analysis [5-10]. The macroscopic analysis usually
assumes that user equipments (UEs) and/or base stations (BSs)
are randomly deployed, often following a homogeneous Poisson distribution to invoke the stochastic geometry theory [3,4].
In essence, the macroscopic analysis investigates network
performance at a high level, such as coverage probability and
signal-to-interference ratio (SIR) distribution, by averaging
over all possible UE and BS deployments [3,4]. Instead, the
microscopic analysis allows for a more detailed analysis and
1 1536-1276 c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Please find the final version
in IEEE from the link: http://ieeexplore.ieee.org/document/7511618/. Digital
Object Identifier: 10.1109/ICC.2016.7511618
2 Orthogonal deployment means that small cells and macrocells operating
on different frequency spectrum, i.e., Small Cell Scenario #2a defined in [1].
is often conducted assuming that UEs are randomly placed but
that BS locations are known [5-10]. Generally speaking, the
macroscopic analysis predicts the network performance in a
statistical sense, while the microscopic analysis is useful for
a network-specific study and optimization.
Within the microscopic analysis and paying special attention
to uplink (UL), in [5], the authors considered a single UL
interfering cell with a disk-shaped coverage area and presented
closed-form expressions for the UL interference considering
both path loss and shadow fading. In [6], the authors conjectured that the UL interference in a hexagonal grid based
cellular network may follow a lognormal distribution, which
was verified via simulation. In [7] and [8], we went a step
further and analytically derived an upper bound of the error
in approximating the dB-scale UL interference from a single
cell by a Gaussian distribution. Such error was measured by
the Kolmogorov–Smirnov (KS) distance [11] between the real
cumulative density function (CDF) and the approximate CDF,
and it was shown to be small for practical SCNs. On the basis
of this single-cell interference analysis, we further investigated
the approximate distribution of the aggregate UL interference
in a multi-cell scenario as a power lognormal distribution. For
more practical networks, in [9] and [10], we also investigated
the network performance of SCNs in current 4G networks
using system-level simulations.
In this paper, our objective is to extend the previous works
in [7] and [8] to analyze the UL SIR performance, and create
a novel and compelling approach for network performance
analysis that can unify the macroscopic and the microscopic
analyses within a single framework, while overcoming a few
drawbacks of the current tools. To that end, our work is
composed of the following three steps:
1) The extension of the UL interference analysis in [7]
and [8] to the UL SIR analysis, which makes our
analysis a formal microscopic analysis tool.
2) The upgrade of the developed microscopic analysis tool
to a macroscopic analysis tool.
3) The comparison between the proposed macroscopic
analysis tool and the existing macroscopic analysis based
on stochastic geometry.
Since the macroscopic and the microscopic analyses are
unified in our framework based on a deterministic network
analysis (DNA) using the Gaussian approximation (GA) presented in [7] and [8], our framework will be referred to as
2
the DNA-GA analysis hereafter. As a result of our three-step
work, the contributions of this paper are three-fold, and are
summarized as follows:
1) Based on the Gaussian approximation theorem presented
in [7] and [8], the approximate distributions of the UL
signal power and the UL SIR for the interested UE are
derived in tractable expressions using the Gauss-Hermite
numerical integration [12], giving rise to the DNA-GA
analysis.
2) Although the DNA-GA analysis stands alone as a solid
contribution to the family of microscopic analysis, two
approaches for upgrading the DNA-GA analysis to a
macroscopic analysis are further investigated. The first
one is the semi-analytical approach, which directly averages the performance given by many DNA-GA analyses
over many random BS deployments to obtain the performance of the macroscopic analysis. The second one
is the analytical approach, which constructs an idealistic
and deterministic BS deployment, and then conducts the
DNA-GA analysis on such BS deployment to obtain an
upper-bound performance of the macroscopic analysis.
3) Interesting results on the comparison between the DNAGA analysis and the stochastic geometry analysis [4] are
presented. Our results show that the DNA-GA analysis
qualifies as a new network performance analysis tool
with a few special merits over stochastic geometry: (i)
Shadow fading is naturally considered in the DNA-GA
analysis, while stochastic geometry usually cannot; (ii)
Non-uniform UE distributions and any type of multipath fading can be treated in the DNA-GA analysis,
while stochastic geometry usually cannot; (iii) Apart
from the common assumption on the cell coverage areas
as Voronoi cells made by stochastic geometry, the shape
and/or the size of cell coverage areas in the DNA-GA
analysis can be made arbitrary, making it suitable for
the network performance analysis of hotspot SCNs.
The remainder of the paper is structured as follows. In
Section II, the network scenario and the system model are
described. In Section III, the DNA-GA analysis is presented,
followed by its upgrade to a macroscopic analysis tool in
Section IV. Our results are validated and compared with
those of the stochastic geometry analysis via simulations in
Section V. Finally, the conclusions are drawn in Section VI.
II. N ETWORK S CENARIO AND S YSTEM M ODEL
In this paper, we consider UL transmissions, and assume
that each small cell BS only schedules one UE in each
frequency/time resource, i.e., resource block (RB). This is a
reasonable assumption in line with 4G networks, i.e., Long
Term Evolution (LTE) [13] and Worldwide Interoperability
for Microwave Access (WiMAX) [14]. Note that small cell
BSs serving no UE do not contribute to the UL interference,
thereby those BSs are ignored in the analysis.
Regarding the network scenario, we consider a SCN with
multiple small cells operating on the same carrier frequency, as
shown in Fig. 1. In more detail, the SCN consists of B small
cells, each of which is managed by a BS. The network includes
the small cell of interest denoted by C1 and B − 1 interfering
small cells denoted by Cb , b ∈ {2, . . . , B}. We focus on a
particular RB, and denote by Kb the active UE associated
with small cell Cb in such RB. Moreover, we denote by Rb
the coverage area of small cell Cb , in which its associated
UEs are randomly distributed. Note that the coverage areas of
adjacent small cells may overlap due to the arbitrary shapes
and sizes of {Rb } , b ∈ {2, . . . , B}.
Fig. 1. A schematic model of the considered SCN.
The distance (in km) from the BS of Cb to the BS of
C1 , b ∈ {1, . . . , B}, and the distance from UE Kb to the
BS of Cm , b, m ∈ {1, . . . , B}, are denoted by Db and dbm ,
respectively. Since DNA-GA is a microscopic analysis tool,
we consider a deterministic deployment of BSs, i.e., the set
{Db } is known, while each UE Kb is randomly distributed in
Rb with a distribution function fZb (z) , z ∈ Rb . Hence, dbm is
a random variable (RV), whose distribution cannot be readily
expressed in an analytical form due to the arbitrary shape and
size of Rb , and the arbitrary form of fZb (z). Regarding Rb
and fZb (z), we have two remarks in the following.
Remark 1: Unlike the existing works, e.g., [3-6, 9-10], where
only the uniform UE distribution was considered, DNA-GA
can handle any probability density function (PDF) of general
UE distribution, here denoted by fZb (z), where 0 < fZb (z) <
+∞,
z ∈ Rb and its integral over Rb equals to one, i.e.,
´
f
(z) dz = 1.
Z
b
Rb
Remark 2: Even if fZb (z) is constant with z, we can only
say that the UE distribution is uniform within the small
cell coverage area Rb , but we cannot guarantee that the UE
distribution is uniform within the entire scenario, because no
UEs are deployed outside the hotspot areas {Rb }, which may
cause the non-uniformity of UE distribution within the entire
scenario. Instead, in stochastic geometry [3,4], UEs are usually
assumed to be uniformly distributed within the entire scenario,
creating Voronoi cells, which is less general and practical than
our assumption of Rb and fZb (z). Note that in the sequel, the
characterization of UE distribution is meant within Rb .
Next, we present the modeling of path loss, shadow fading,
UL transmission power, multi-path fading and noise.
Based on the definition of dbm , the path loss in dB from
UE Kb to the BS of Cm is modeled as
Lbm = A + α × 10log10 dbm ,
(1)
where A is the path loss in dB at the reference distance of
dbm = 1 and α is the path loss exponent. In practice, A and α
3
are constants obtainable from field tests [15]. Note that Lbm
is a RV due to the randomness of dbm .
The shadow fading in dB from UE Kb to the BS of
Cm is denoted by Sbm , and is usually assumed to follow a
lognormal distribution [15]. Based on this assumption, Sbm
is modeled as an independently and identically distributed
2
(i.i.d.) zero-mean Gaussian
RV with a variance of σShad
, i.e.,
2
Sbm ∼ N 0, σShad .
The UL transmission power in dBm of UE Kb is denoted
by Pb , and is subject to a semi-static power control (PC)
mechanism, e.g., the fractional path loss compensation (FPC)
scheme [15]. Based on this FPC scheme, Pb is modeled as
Pb = P0 + η (Lbb + Sbb ) ,
(2)
where P0 is the target received power in dBm on the considered RB at the BS,
η ∈ (0, 1] is the FPC factor, and Lbb and
2
Sbb ∼ N 0, σShad
have been discussed above.
The multi-path fading channel from UE Kb to the BS of
Cm is denoted by hbm ∈ C, where we assume that each UE
and each BS are equipped with one omni-directional antenna.
It is important to note that we consider a general type of multipath fading by assuming that the effective channel gain in dB
2
associated with hbm is defined as Hbm = 10 log10 |hbm | ,
which follows an i.i.d. distribution with a PDF of fH (h).
2
For example, |hbm | can be characterized by an exponential
distribution or a Gamma distribution in case of Rayleigh
fading or Nakagami fading, respectively [16]. And hence, the
distribution of Hbm can be derived analytically.
Finally, we ignore the additive noise because the 4G and the
5G SCNs generally work in the interference-limited region [2].
III. T HE P ROPOSED DNA-GA A NALYSIS
The proposed DNA-GA analysis consists of three steps, i.e.,
the interference analysis, the signal power analysis, and the
SIR analysis, which are presented in the following.
A. The Interference Analysis
Based on the definition of RVs discussed in Section II, the
UL received interference power in dBm from UE Kb to the
BS of C1 can be written as
(a)
Ib = Pb − Lb1 − Sb1 + Hb1
= P0 + (ηLbb − Lb1 ) + (ηSbb − Sb1 ) + Hb1
4
= (P0 + L + S) + Hb1 ,
4
(1)
= Ib
+ Hb1 ,
(3)
where (2) is plugged into the step (a) of (3), and L and S are
4
4
defined as L = (ηLbb − Lb1 ) and S = (ηSbb − Sb1 ), respectively. Apparently, L and S are independent RVs. Besides, the
(1) 4
first part of Ib is further defined as Ib = (P0 + L + S). Since
Sbb and Sb1 (b ∈ {2, . . . , B}) are i.i.d. zero-mean Gaussian
RVs, it is easy to show that S is also a Gaussian RV, whose
mean and variance (
are
µS = 0
2 .
(4)
σS2 = 1 + η 2 σShad
From the definition of Ib in (3), the aggregate interference
power in mW from all interfering UEs to the BS of C1 can
be formulated as
B
X
1
(5)
I mW =
10 10 Ib .
b=2
In our previous work [8], we show that the distribution of
I mW can be well approximated by a power lognormal distribution. This approximation is summarized in the following.
(1)
1) The Distribution of Ib in (3): First, we analyze the
(1)
distribution of Ib shown in (3). Considering a small approximation error, upper-bounded by the KS distance [11] provided
(1)
in [8], we approximate Ib by a Gaussian RV Gb , whose mean
and variance are (
µGb = P0 + µL + µS
,
(6)
2
2
σG
= σL
+ σS2
b
2
where µL and σL
are respectively the mean and the variance
of L, which can be obtained using numerical integration
involving fZb (z) and Rb [7,8]. Details are omitted for brevity.
2) The Distribution of Ib in (3): Second, we analyze the
(1)
distribution of Ib = Ib + Hb1 shown in (3). Considering
a small approximation error, upper-bounded by the KS distance [11] provided in [8], we approximate Ib by another
Gaussian RV Qb , whose
mean and variance are
(
µQb = µGb + µHb1
.
(7)
2
2
2
+ σH
= σG
σQ
b1
b
b
2
where µHb1 and σH
are respectively the mean and the
b1
variance of Hb1 . We omit the details for brevity.
Note that the upper bound of the total approximation error
of the above two steps is obtained from the summation of the
individual approximation errors of the two steps in closed-form
expressions [8]. And it has been shown in [8] that the total
approximation error is small for practical SCNs, without any
requirement on (i) the uniformity of UE distribution and/or
the type of multi-path fading; and (ii) the shape and/or size
of cell coverage areas. Intuitively speaking, the results in [8]
show that the larger the variance of the Gaussian RV, i.e., S
in (6) or Gb in (7), the better the approximation in (6) or in
(7), due to the increasing dominance of the Gaussian RV.
3) The Distribution of I mW in (5): Third, we invoke the
main results in [17-18], which indicate that the sum of multiple
independent lognormal RVs can be well approximated by a
power lognormal RV. Accordingly, in our case, since each
Ib , b ∈ {2, . . . , B} is approximated by a Gaussian RV Qb ,
1
their sum 10 10 Qb shown in (5) should be well approximated
1
by a power lognormal RV expressed as IˆmW = 10 10 Q , where
the PDF and CDF of Q [17] can be written as (8) shown
on the top of the next page. In (8), Φ (x) is the CDF of the
standard normal distribution, and the
2parameters λ, µQ and
σQ are obtained from {µQb } and σQ
that are computed
b
by (7). The procedure to obtain λ, µQ and σQ is omitted here
for brevity, but interested readers are referred to Appendix B
of [8] for further details. As a result of (8), the PDF and CDF
of IˆmW can be written as (9) shown on the top of the next
page, where ζ = ln1010 is a scalar factor originated from the
variable change from 10 log10 v to ln v.
4
Finally, we approximate the distribution of I mW by that of
mW
ˆ
I
shown in (9) presented on the top of the next page.
Note that in this step, the approximation error depends on
the approximate error introduced by the power lognormal
approximation, which has been shown to be reasonably small
and good enough in practical cases [17-18].
(
1
fG1 (v) = q
2
2πσG
1
(v − µG1 )
exp −
2
2σG
1
= Pr [G1 + H11 ≤ x]
Based on the definition of RVs discussed in Section II, the
UL received signal power in dBm from UE K1 to the BS of
C1 can be written as
= Pr [H11 ≤ x − G1 ]
+∞
ˆ
FH11 (x − v) fG1 (v) dv
=
(a)
−∞
+∞
ˆ
(a)
= P0 + (ηL11 − L11 ) + (ηS11 − S11 ) + H11
4
= P0 + L̄11 + S̄11 + H11 ,
4
(1)
= X1 + H11 ,
FH11 (x − v) q
=
−∞
(10)
where (2) is plugged into the step (a) of (10). Besides, L̄11 and
4
4
S̄11 are defined as L̄11 = (η − 1) L11 and S̄11 = (η − 1) S11 ,
(1) 4
respectively. The first part of X1 is further defined as X1 =
P0 + L̄11 + S̄11 , and it is easy to show that S̄11 is a Gaussian
RV, whose mean and
( variance are
µS̄11 = 0
.
(11)
2 2
σS̄2 11 = (1 − η) σShad
Similar to the discussion in subsection III-A1, we consider a
small approximation error, upper-bounded by the KS distance
(1)
shown in [8], and we approximate X1 by a Gaussian RV
G1 , whose mean(and variance are
µG1 = P0 + µL̄11 + µS̄11
,
(12)
2
2
+ σS̄2 11
σG
= σL̄
1
11
2
are respectively the mean and the
where µL̄11 and σL̄
11
variance of L̄11 . As a result, (10) can be re-formulated as
4
X1 ≈ G1 + H11 = X̂1 .
(13)
Note that unlike the discussion in subsection III-A2, it is not
accurate to further approximate X̂1 by a Gaussian RV, because
the randomness of the Gaussian distributed RV S11 is largely
removed by the UL transmission power control mechanism,
rendering a less dominant role of the Gaussian distribution of
2
G1 compared with the distribution of H11 . In other words, σG
1
is comparable with or even smaller than the variance of H11 ,
making the approximation error large according to our results
in [8]. Therefore, we derive the approximate distribution of
X1 using a different method, presented in Theorem 1.
Theorem 1. The approximate CDF of X1 is derived as
M0
√
1 X
wm FH11 x− 2σG1 am+µG1 ,
FX1(x)≈FX̂1(x)= √
π m=1
(14)
where M0 is the number of terms employed in the GaussHermite numerical integration [12], and the weights {wm }
and the abscissas {am } are tabulated in Table 25.10 of [12].
Proof: Since G1 is a Gaussian RV with the mean and the
variance shown in (12), the PDF of G1 can be written as
.
(15)
Besides, according to the definition of RVs in Section II,
we assume the CDF of H11 to be FH11 (h). Hence, the CDF
of X1 can be approximated by
FX1 (x) ≈ FX̂1 (x)
B. The Signal Power Analysis
X1 = P1 − L11 − S11 + H11
)
2
1
= √
π
(b)
1
2
2πσG
1
(
2
(v − µG1 )
exp −
2
2σG
1
)
dv
+∞
ˆ
√
FH11 x −
2σG1 y + µG1 exp −y 2 dy
−∞
M0
√
1 X
= √
2σG1 am + µG1
+ RM0
wm FH11 x −
π m=1
(c)
M0
√
1 X
≈ √
wm FH11 x −
2σG1 am + µG1 ,
π m=1
(d)
(16)
where the step (a) of (16) is obtained from (15), and the
step (b)
√ of (16) is computed using the variable change
v =
2σG1 y + µG1 . Moreover, the step (c) of (16) is
derived using the Gauss-Hermite numerical integration [12],
+∞
´
PM0
i.e.,
f (y) exp −y 2 dy =
m=1 wm f (am ) + RM0 ,
−∞
where M0 is the number of terms in the approximation,
the weights {wm } and the abscissas {am } are tabulated in
Table 25.10 of [12] and RM0 is a residual error in the order
0!
[12], which decays very fast as M0 increases.
of 2M0M
(2M0 )!
Finally, the step (d) of (16) is obtained by dropping RM0 . Our
proof is thus completed by comparing (14) and (16).
In case of Rayleigh fading [16], we propose Corollary 2 to
compute the approximate expression of FX1 (x).
Corollary 2. In case of Rayleigh fading, the approximate CDF
of X1 can be computed by (14), where
h
FH11 (h) = 1 − exp − exp
,
(17)
ζ
where ζ =
10
ln 10 .
Proof: As discussed in Section II, on condition of
2
Rayleigh fading, the channel gain |h11 | follows an exponential distribution with unitary mean [16]. Then, our proof
is completed by deriving (17) based on the variable change
2
H11 = 10 log10 |h11 | . Details are omitted for brevity.
In case of Nakagami fading [16], we propose Corollary 3
to compute the approximate expression of FX1 (x).
Corollary 3. In case of Nakagami fading, the approximate
CDF of X1 can be computed by
(14), where
1
h
1
FH11 (h) =
γ k, exp
,
(18)
Γ (k)
θ
ζ
5
n
o
(q−µQ )2
Q
1
PDF of Q : fQ (q) = λΦλ−1 q−µ
√
exp
−
2
2
σQ
2σQ
2πσQ
.
CDF of Q : FQ (q) = Φλ q−µQ
σQ
n
o
2
Q)
PDF of IˆmW : fIˆmW (v) = λΦλ−1 ζ lnσv−µQ √ ζ 2 exp − (ζ ln v−µ
2
2σ
Q
v 2πσQ
Q
.
CDF of IˆmW : F ˆmW (v) = Φλ ζ ln v−µQ
I
σQ
where Γ (·) and γ (·, ·) are respectively the gamma and the
incomplete gamma functions [12], k and θ are respectively
the shape and the scale parameters of the Gamma distribution
associated with the channel gain of Nakagami fading [16].
Proof: As discussed in Section II, on condition of Nak2
agami fading, the channel gain |h11 | follows a Gamma
distribution with parameters k and θ [16]. Then, our proof
is completed by deriving (18) based on the variable change
2
H11 = 10 log10 |h11 | Details are omitted for brevity.
C. The SIR Analysis
From (10), we can approximate the UL SIR in dB by
4
Z dB ≈ X1 − Q = Ẑ dB .
(19)
We derive the approximate distribution of Z dB in Theorem 4.
Theorem 4. The approximate CDF of Z dB is derived as
FZ dB (z) ≈ FẐ dB (z)
M0
√
√
λ X
=√
wm Φλ−1 2am FX1 z + 2σQ am + µQ , (20)
π m=1
where M0 is the number of terms employed in the GaussHermite numerical integration [12], and the weights {wm }
and the abscissas {am } are tabulated in Table 25.10 in [12].
Proof: From (19), the approximate CDF of Z dB can be
derived as
FZ dB (z) ≈ FẐ dB (z)
= Pr [X1 − Q ≤ z]
= Pr [X1 ≤ z + Q]
+∞
ˆ
= FX1(z + q) fQ (q) dq
−∞
+∞
ˆ
(a)
≈
−∞
(
)
2
1
(q−µQ )
q−µQ
q
exp −
dq
FX̂1(z + q) λΦλ−1
2
σQ
2σQ
2πσ 2
Q
+∞
ˆ
√
√
(b) 1
= √
FX̂1 z + 2σQ y + µQ λΦλ−1 2y exp −y 2 dy
π
−∞
M0
√
√
1 X
= √
wm FX̂1 z + 2σQ am + µQ λΦλ−1 2am +RM0
π m=1
(c)
M0
√
√
λ X
√
≈
wm Φλ−1 2am FX̂1 z + 2σQ am + µQ ,
π m=1
(d)
(8)
(9)
√
change q = 2σQ y + µQ . Moreover, the step (c) of (21) is
derived using the Gauss-Hermite numerical integration [12].
Finally, the step (d) of (21) is obtained by dropping RM0 . Our
proof is thus completed by comparing (20) and (21).
In case of Rayleigh fading [16], we propose Corollary 5 to
compute the approximate expression of FZ dB (z).
Corollary 5. In case of Rayleigh fading, the approximate CDF
of Z dB can be computed by (20), where (17) is plugged into
(14) to obtain FX̂1 (x) in (20).
Proof: The proof is completed by applying Corollary 2
to Theorem 1 and Theorem 4. Details are omitted for brevity.
In case of Nakagami fading [16], we propose Corollary 6
to compute the approximate expression of FZ dB (z).
Corollary 6. In case of Nakagami fading, the approximate
CDF of Z dB can be computed by (20), where (18) is plugged
into (14) to obtain FX̂1 (x) in (20).
Proof: The proof is completed by applying Corollary 3
to Theorem 1 and Theorem 4. Details are omitted for brevity.
IV. M ACROSCOPIC U PGRADE OF THE DNA-GA A NALYSIS
With Theorem 4, we have crafted a powerful microscopic
analysis tool based on the proposed DNA-GA analysis that
can deal with a wide range of network assumptions and
system parameters. In this section, we further investigate
two approaches for upgrading the DNA-GA analysis from a
microscopic analysis tool to a macroscopic one, putting DNAGA in the same league as, e.g., stochastic geometry.
A. The Semi-Analytical Approach
The microscopic and the macroscopic analyses are closely
related to each other. The average performance of many
microscopic analyses conducted over a large number of random BS deployments converges to the performance of the
macroscopic analysis, given that the examined realizations of
the deterministic BS deployments follow the BS deployment
assumption used in the macroscopic analysis. Therefore, we
can directly average the performance results obtained by
applying DNA-GA, i.e., Theorem 4, over a large number of
random BS deployments to obtain the performance results of
the macroscopic analysis.
(21)
where the step (a) of (21) is calculated using Theorem 1 and
(8), and the step (b) of (21) is computed using the variable
B. The Analytical Approach
Instead of conducting DNA-GA, i.e., Theorem 4, over many
BS deployments and averaging all the results together to obtain
6
the results of the macroscopic analysis, we can construct
an idealistic BS deployment on a hexagonal lattice with the
equivalent BS density, and perform a single DNA-GA analysis
on such BS deployment to extract an upper-bound of the
SIR performance of the macroscopic analysis. The hexagonal
lattice leads to an upper-bound performance because BSs
are evenly distributed in the scenario and thus very strong
interference due to close proximity is avoided [3,4].
V. S IMULATION AND D ISCUSSION
Fig. 2. Illustration of a 3GPP SCN deployment [15].
In this section, we conduct simulations to validate the
proposed DNA-GA analysis, using both the semi-analytical
and the analytical approaches. For the semi-analytical approach, to obtain the results of the macroscopic analysis, we
average the results given by Theorem 4 over 1000 random
BS deployments. For each BS deployment, 10,000 random
experiments are conducted to go through the randomness of
UE positions. And for each BS deployment and each UE
placement, another 10,000 random experiments are conducted
to go through the randomness of shadow fading and multi-path
fading. For the analytical approach, only one BS deployment
in a hexagonal lattice is examined. M0 is set to 30 for
the computation in the DNA-GA analysis to ensure a good
accuracy of the results [12].
With regard to the scenario and parameters, 3rd Generation Partnership Project (3GPP) recommendations have been
considered [15]. For the semi-analytical approach, 19 dummy
macrocell sites are deployed with a 0.5 km inter-site distance
to guide the small cell deployment. Each macrocell site has the
shape of a hexagon, and is equally divided into 3 macrocells.
Each macrocell contains 4 randomly deployed small cells,
resulting in 19 × 3 × 4 = 228 small cells with a density
around 55.43 cells/km2 . For the analytical approach, the 228
small cells are located in a hexagonal lattice with the same cell
density. In both cases, each small cell has a coverage radius of
0.04 km, and the minimum inter-BS distance and the minimum
BS-to-UE distance are 0.04 km, and 0.01 km, respectively.
Moreover, according to [15], A = 145.4, α = 3.75, P0 = −76
dBm, η = 0.8, and σS = 10 dB.
Fig. 2 illustrates an example of a random BS deployment
according to [15], where small cell BSs are represented by xmarkers, while the coverage areas of dummy macrocells and
small cells are marked by dashed and solid lines, respectively.
UEs are randomly distributed in the mentioned small cell
coverage areas, and it is important to note that although some
small cell coverage areas are disk-shaped, the coverage areas
of most small cells are of irregular shape due to overlapping.
For brevity, in the following subsections, we omit the
detailed investigation on the interference analysis and the
signal power analysis, and directly present SIR results given
by the DNA-GA analysis and the simulation.
A. Validation of the DNA-GA Analysis
In this subsection, we validate the accuracy of the DNA-GA
analysis in terms of the SIR performance when assuming two
cases for UE distribution and multi-path fading, i.e.,
•
•
Case 1: Uniform UE distribution + Rayleigh fading
Case 2: Non-uniform UE distribution + Nakagami fading
For Case 1, we obtain the SIR results through the DNA-GA
analysis using Theorem 4 and Corollary 5, while for Case 2,
we invoke Theorem 4 and Corollary 6.
When considering a non-uniform UE distribution, we assume that fZb (z) = W
ρ , z ∈ Rb , where ρ is the radial
coordinate of z in the polar coordinate system, the origin of
which is placed at the position of´ the BS of Cb and W is
a normalization constant to make Rb fZb (z) dz = 1. In the
resulting non-uniform UE distribution, UEs are more likely
to locate in the close vicinity of the BS of Cb than at the
cell-edge3 . When considering Nakagami fading, we assume
that k = 10 and θ = 0.1, which corresponds to a multi-path
fading with a strong line-of-sight (LoS) component [16].
For both cases, the UL SIR performance is evaluated using
the simulation and the semi-analytical approach discussed in
Subsection IV-A. Moreover, the upper bound of the UL SIR
is also investigated using the simulation and the analytical
approach discussed in Subsection IV-B based on a BS deployment in a hexagonal lattice. The results are shown in Fig. 3.
3 Note that the considered f
Zb (z) is just an example of the non-uniformly
distributed UEs in Rb , which reflects a reasonable network planning, where
small cell BSs have been deployed at the center of UE clusters. Other forms
of fZb (z) can be considered in our DNA-GA analysis as well.
7
1
0.8
1
UL SIR (Simulation)
UL SIR (DNA-GA)
Upper-Bound (Simulation)
Upper-Bound (DNA-GA)
0.8
CDF
0.6
CDF
0.6
UL SIR (Simulation [4])
UL SIR (Analysis [4])
UL SIR (Simulation)
UL SIR (DNA-GA)
Upper-Bound (DNA-GA)
Case 1
0.4
Stochastic
Geometry
DNA-GA
(Case 1)
0.4
0.2
0.2
Case 2
0
-30
-20
-10
0
UL SIR [dB]
10
20
30
Fig. 3. UL SIR in dB (DNA-GA vs. Simulation).
As can be seen from Fig. 3, the SIR results of the proposed DNA-GA analysis match those of the simulation very
well, particularly in the head portion. In the semi-analytical
approach, the maximum deviation between the CDFs obtained
by the DNA-GA analysis and the simulation for both investigated cases are around 1.0∼1.7 percentile. In the analytical
approach, the fitness becomes even better, i.e., the maximum
deviation for both cases is within 0.6 percentile.
More importantly, for both cases, the upper-bound SIR
performance given by the analytical approach is shown to be
within 2.0 ∼ 2.5 dB from the exact performance, indicating
its usefulness in characterizing the network performance with
low-complexity computation. To take Case 1 of DNA-GA as
an example, the numerical results to be plugged into Theorem 4 for the hexagonal BS deployment are µG1 = −93.07,
2
2
σG
= 5.97, λ = 202.66, µQ = −137.71 and σQ
= 212.04.
1
Finally, note that the SIR of Case 2 outperforms that of
Case 1, mainly because UEs tend to stay closer to their serving
BSs in Case 2 as discussed above, leading to a larger signal
power and a lower interference power.
B. Comparison Between DNA-GA and Stochastic Geometry
In this section, we compare the UL SIR results of the DNAGA analysis (Case 1) and those of the stochastic geometry
analysis in Fig. 4, with the same average cell density of 55.43
cells/km2 and the same assumption of Rayleigh fading4 .
4 Note that the stochastic geometry analysis in [4] poses some assumptions
on the model for the sake of tractability, e.g., no shadow fading, Rayleigh
fading only, homogeneous Poisson distribution of both UEs and BSs in
the entire scenario. In contrast, the DNA-GA analysis does not need such
assumptions and works with a more realistic model, considering the hotspot
SCN scenario [15] shown in Fig. 2 and discussed in Remark 2 of Section II.
0
-30
-20
-10
0
UL SIR [dB]
10
20
30
Fig. 4. UL SIR in dB (DNA-GA vs. Stochastic Geometry [4]).
In Fig. 4, a few interesting aspects are noteworthy. First,
both our analysis and the analysis in [4] are only able to give
approximate results. However, the approximation error of our
DNA-GA analysis is shown to be smaller than that of [4].
Second, there is a significant performance gap between
our DNA-GA analysis and the stochastic geometry analysis
in [4]. This is because (i) Our DNA-GA analysis considers
the shadow fading on top of the multi-path fading, which
leads to a large variance of SIR, while the analysis in [4] does
not, which gives a small SIR variance; and (ii) The DNA-GA
analysis studies the hotspot SCN scenario recommended by
the 3GPP [15], where UEs are deployed closer to the serving
BSs than those in Voronoi cells considered in [4].
Third, the purpose of Fig. 4 is not to reproduce the results
in [4] based on Voronoi cells, but to analytically investigate a
more practical 3GPP network scenario. If the shadow fading
is required to be ignored, albeit impractical, the Gamma
approximation of the aggregate interference [10] could be
invoked to make our approach of analysis still valid. Besides,
the DNA-GA analysis can also handle the case where the cell
coverage areas are constructed as Voronoi cells [4]. However,
to do so, it would be more practical to consider an alternative
UE association strategy (UAS), where each UE is connected
to the BS with the smallest path loss plus shadow fading.
Note that such UAS will blur the boundaries of Voronoi cells,
because each UE is no longer always connected to its closest
BS, making the analysis more intricate and realistic.
Finally, note that a three-fold integral computation is needed
in [4] to compute the results, while no integration is required
in Theorem 4 of the DNA-GA analysis. However, many BS
deployments are needed in the semi-analytical approach of the
DNA-GA analysis, while only one in the analytical approach.
VI. C ONCLUSION
We proposed a new approach of network performance
analysis, which unifies the microscopic and the macroscopic
analyses within a single framework. Compared with stochastic
geometry, our DNA-GA analysis considers shadow fading,
general UE distribution and any type of multi-path fading, as
well as any shape and/or size of cell coverage areas. Thus,
DNA-GA can analyze more realistic networks and is very
useful for the network performance analysis of the 5G systems
with general cell deployment and UE distribution.
8
R EFERENCES
[1] 3GPP, “TR 36.872, Small cell enhancements for E-UTRA and EUTRAN - Physical layer aspects,” Dec. 2013.
[2] D. López Pérez, M. Ding, H. Claussen, and A. H. Jafari, “Towards
1 Gbps/UE in cellular systems: Understanding ultra-dense small cell
deployments,” IEEE Commun. Surveys and Tutorials, Jun. 2015.
[3] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
coverage and rate in cellular networks,” IEEE Trans. on Commun., vol.
59, no. 11, pp. 3122-3134, Nov. 2011.
[4] T. D. Novlan, H. S. Dhillon and J. G. Andrews, “Analytical modeling
of uplink cellular networks,” IEEE Trans. on Wireless Commun., vol.
12, no. 6, pp. 2669-2679, Jun. 2013.
[5] Y. Zhu, J. Xu, Z. Hu, J. Wang and Y. Yang, “Distribution of uplink
inter-cell interference in OFDMA networks with power control,” IEEE
ICC 2014, Sydney, Australia, pp. 5729-5734, Jun. 2014.
[6] J. He, Z. Tang, H. Chen, and W. Cheng, “Statistical model of OFDMA
cellular networks uplink interference using lognormal distribution,”
IEEE Wireless Commun. Letters, vol. 2, no. 5, pp. 575-578, Oct. 2013.
[7] M. Ding, D. López-Pérez, G. Mao, and Z. Lin, “Approximation of uplink
inter-cell interference in FDMA small cell networks,” to appear in IEEE
Globecom 2015, arXiv: 1505.01924 [cs.NI], May 2015.
[8] M. Ding, D. López-Pérez, G. Mao, and Z. Lin, “Microscopic analysis
of the uplink interference in FDMA small cell networks,” submitted to
IEEE Trans. on Wireless Commun., arXiv:1508.02808, Aug. 2015.
[9] M. Ding, D. López Pérez, A. V. Vasilakos, and W. Chen, “Dynamic
TDD transmissions in homogeneous small cell networks,” IEEE ICC
2014, Sydney, Australia, pp. 616-621, Jun. 2014.
[10] M. Ding, D. López Pérez, A. V. Vasilakos, and W. Chen, “Analysis
on the SINR performance of dynamic TDD in homogeneous small cell
networks,” IEEE Globecom 2014, pp. 1552-1558, Dec. 2014.
[11] F. J. Massey Jr. “The Kolmogorov-Smirnov test for goodness of fit,”
Journal of the American Statistical Association, vol. 46, no. 253, pp.
68-78, 1951.
[12] M. Abramowitz and I. Stegun, Handbook of mathematical functions with
formulas, graphs, and mathematical tables (Nineth Ed.), Dover, 1972.
[13] 3GPP, “TS 36.213 (V11.2.0): Physical layer procedures,” Feb. 2013.
[14] WiMax Forum, “WiMAX and the IEEE 802.16m Air Interface Standard,” Apr. 2010.
[15] 3GPP, “TR 36.828 (V11.0.0): Further enhancements to LTE Time
Division Duplex (TDD) for Downlink-Uplink (DL-UL) interference
management and traffic adaptation,” Jun. 2012.
[16] J. Proakis, Digital Communications (Third Ed.), New York: McGrawHill, 1995.
[17] Z. Liu, J. Almhana, and R. McGorman, “Approximating lognormal
sum distributions with power lognormal distributions,” IEEE Trans. on
Vehicular Tech., vol. 57, no. 4, pp. 2611-2617, Jul. 2008.
[18] S. S. Szyszkowicz and H. Yanikomeroglu, “Fitting the modified-powerlognormal to the sum of independent lognormals distribution,” IEEE
Globecom 2009, pp. 1-6, Nov. 2009.
| 7cs.IT
|
Classification and Geometry of General Perceptual Manifolds
SueYeon Chung1,2 , Daniel D. Lee2,3 , and Haim Sompolinsky2,4,5
1
arXiv:1710.06487v2 [cond-mat.dis-nn] 26 Feb 2018
Program in Applied Physics, School of Engineering and Applied Sciences,
Harvard University, Cambridge, MA 02138, USA
2
Center for Brain Science, Harvard University, Cambridge, MA 02138, USA
3
School of Engineering and Applied Science, University of Pennysylvania, Philadelphia, PA 19104, USA
4
Racah Institute of Physics, Hebrew University, Jerusalem 91904, Israel
5
Edmond and Lily Safra Center for Brain Sciences, Hebrew University, Jerusalem 91904, Israel
Perceptual manifolds arise when a neural population responds to an ensemble of sensory signals
associated with different physical features (e.g., orientation, pose, scale, location, and intensity) of
the same perceptual object. Object recognition and discrimination require classifying the manifolds
in a manner that is insensitive to variability within a manifold. How neuronal systems give rise
to invariant object classification and recognition is a fundamental problem in brain theory as well
as in machine learning. Here we study the ability of a readout network to classify objects from
their perceptual manifold representations. We develop a statistical mechanical theory for the linear
classification of manifolds with arbitrary geometry revealing a remarkable relation to the mathematics of conic decomposition. We show how special anchor points on the manifolds can be used
to define novel geometrical measures of radius and dimension which can explain the classification
capacity for manifolds of various geometries. The general theory is demonstrated on a number of
representative manifolds, including `2 ellipsoids prototypical of strictly convex manifolds, `1 balls
representing polytopes with finite samples, and ring manifolds exhibiting non-convex continuous
structures that arise from modulating a continuous degree of freedom. The effects of label sparsity
on the classification capacity of general manifolds are elucidated, displaying a universal scaling relation between label sparsity and the manifold radius. Theoretical predictions are corroborated by
numerical simulations using recently developed algorithms to compute maximum margin solutions
for manifold dichotomies. Our theory and its extensions provide a powerful and rich framework
for applying statistical mechanics of linear classification to data arising from perceptual neuronal
responses as well as to artificial deep networks trained for object recognition tasks.
PACS numbers: 87.18.Sn, 87.19.lv, 42.66.Si, 07.05.Mh
I.
INTRODUCTION
A fundamental cognitive task performed by animals and
humans is the invariant perception of objects, requiring the nervous system to discriminate between different objects despite substantial variability in each objects physical features. For example, in vision, the
mammalian brain is able to recognize objects despite
variations in their orientation, position, pose, lighting
and background. Such impressive robustness to physical changes is not limited to vision; other examples include speech processing which requires the detection of
phonemes despite variability in the acoustic signals associated with individual phonemes; and the discrimination
of odors in the presence of variability in odor concentrations. Sensory systems are organized as hierarchies
consisting of multiple layers transforming sensory signals
into a sequence of distinct neural representations. Studies of high level sensory systems, e.g., the inferotemporal cortex (IT) in vision [1], auditory cortex in audition
[2], and piriform cortex in olfaction [3], reveal that even
the late sensory stages exhibit significant sensitivity of
neuronal responses to physical variables. This suggests
that sensory hierarchies generate representations of objects that although not entirely invariant to changes in
physical features, are still readily decoded in an invari-
ant manner by a downstream system. This hypothesis is
formalized by the notion of the untangling of perceptual
manifolds [4–6]. This viewpoint underlies a number of
studies of object recognition in deep neural networks for
artificial intelligence [7–10].
To conceptualize perceptual manifolds, consider a set of
N neurons responding to a specific sensory signal associated with an object as shown in Fig. 1. The neural
population response to that stimulus is a vector in RN .
Changes in the physical parameters of the input stimulus that do not change the object identity modulate the
neural state vector. The set of all state vectors corresponding to responses to all possible stimuli associated
with the same object can be viewed as a manifold in the
neural state space. In this geometrical perspective, object recognition is equivalent to the task of discriminating
manifolds of different objects from each other. Presumably, as signals propagate from one processing stage to
the next in the sensory hierarchy, the geometry of the
manifolds is reformatted so that they become “untangled,” namely they are more easily separated by a biologically plausible decoder [1]. In this paper, we model
the decoder as a simple single layer network (the perceptron) and ask how the geometrical properties of the
perceptual manifolds influence their ability to be separated by a linear classifier.
2
bility of general, finite dimensional manifolds. The summary of our key results is as follows:
Figure 1. Perceptual manifolds in neural state space. (a)
Firing rates of neurons responding to images of a dog shown
at various orientations θ and scales s. The response to a
particular orientation and scale can be characterized by an
N -dimensional population response. (b) The population responses to the images of the dog form a continuous manifold
representing the complete set of invariances in the RN neural activity space. Other object images, such as those corresponding to a cat in various poses, are represented by other
manifolds in this vector space.
Linear separability has previously been studied in the
context of the classification of points by a perceptron, using combinatorics [11] and statistical mechanics [12, 13].
Gardner’s statistical mechanics theory is extremely important as it provides accurate estimates of the perceptron capacity beyond function counting by incorporating
robustness measures. The robustness of a linear classifier
is quantified by the margin, which measures the distance
between the separating hyperplane and the closest point.
Maximizing the margin of a classifier is a critical objective in machine learning, providing Support Vector Machines (SVM) with their good generalization performance
guarantees [14].
The above theories focus on separating a finite set of
points with no underlying geometrical structure and are
not applicable to the problem of manifold classification
which deals with separating infinite number of points geometrically organized as manifolds. This paper addresses
the important question of how to quantify the capacity of the perceptron for dichotomies of input patterns
described by manifolds. In an earlier paper, we have
presented the analysis for classification of manifolds of
extremely simple geometry, namely balls [15]. However,
the previous results have limited applicability as the neural manifolds arising from realistic physical variations of
objects can exhibit much more complicated geometries.
Can statistical mechanics deal with the classification of
manifolds with complex geometry, and what specific geometric properties determine the separability of manifolds?
In this paper, we develop a theory of the linear separa-
• We begin by introducing a mathematical model
of general manifolds for binary classification (Sec.
II). This formalism allows us to generate generic
bounds on the manifold separability capacity from
the limits of small manifold sizes (classification of
isolated points) as that of large sizes (classification
of entire affine subspaces). These bounds highlight
the fact that for large ambient dimension N , the
maximal number P of separable finite-dimensional
manifolds is proportional to N , even though each
consists of infinite number of points, setting the
stage for a statistical mechanical evaluation of the
P
.
maximal α = N
• Using replica theory, we derive mean field equations for the capacity of linear separation of finite
dimensional manifolds (Sec. III) and for the statistical properties of the optimal separating weight
vector. The mean-field solution is given in the form
of self consistent Karush-Kuhn-Tucker (KKT) conditions involving the manifold anchor point. The
anchor point is a representative support vector for
the manifold. The position of the anchor point on
a manifold changes as the orientations of the other
manifolds are varied, and the ensuing statistics of
the distribution of anchor points play a key role in
our theory. The optimal separating plane intersects
a fraction of the manifolds (the supporting manifolds). Our theory categorizes the dimension of the
span of the intersecting sets (e.g., points, edges,
faces, or full manifolds) in relation to the position
of the anchor points in the manifolds’ convex hulls.
• The mean field theory motivates a new definition
of manifold geometry, which is based on the measure induced by the statistics of the anchor points.
In particular, we define the manifold anchor radius
and dimension, RM and DM , respectively. These
quantities are relevant since the capacity of general
manifolds can be well approximated by the capacity
of `2 balls with radii RM and dimensions DM . Interestingly, we show that in the limit of small manifolds, the anchor point statistics are dominated
by points on the boundary of the manifolds which
have minimal overlap with Gaussian random vectors. The resultant Gaussian radius, Rg and dimension, Dg , are related to the well-known Gaussian
mean-width of convex bodies (Sec. IV). Beyond
understanding fundamental limits for classification
capacity, these geometric measures offer new quantitative tools for assessing how perceptual manifolds are reformatted in brain and artificial systems.
• We apply the general theory to three examples,
representing distinct prototypical manifold classes.
3
One class consists of manifolds with strictly smooth
convex hulls, which do not contain facets and are
exemplified by `2 ellipsoids. Another class is that
of convex polytopes, which arise when the manifolds consists of a finite number of data points, and
are exemplified by `1 ellipsoids. Finally, ring manifolds represent an intermediate class: smooth but
nonconvex manifolds. Ring manifolds are continuous nonlinear functions of a single intrinsic variable,
such as object orientation angle. The differences
between these manifold types show up most clearly
in the distinct patterns of the support dimensions.
However, as we show, they share common trends.
As the size of the manifold increases, the capacity and geometrical measures vary smoothly, exhibiting a smooth cross-over from small radius and
dimension with high capacity to large radius and
dimension with low capacity. This crossover occurs as Rg ∝ √1 . Importantly, for many realistic
Dg
cases, when the size is smaller than the crossover
value, the manifold dimensionality is substantially
smaller than that computed from naive second order statistics, highlighting the saliency and significance of our measures for the anchor geometry.
Figure 2. Model of manifolds in affine subspaces. D = 2
manifold embedded in RN . c is the orthogonal translation
vector for the affine space and x0 is the center of the manifold.
As the scale r is varied, the manifold shrinks to the point x0 ,
or expands to fill the entire affine space.
II.
• Finally, we treat the important case of the classification of manifolds with imbalanced (sparse) labels, which commonly arise in problems of object
recognition. It is well known that in highly sparse
labels, the classification capacity of random points
−1
increases dramatically as (f |log f |) , where f 1
is the fraction of the minority labels. Our analysis
of sparsely labeled manifolds highlights the interplay between manifold size and sparsity. In particular, it shows that sparsity enhances the capacity only when f Rg2 1, where Rg is the (Gaussian) manifold radius. Notably, for a large regime
of parameters, sparsely labeled manifolds can approximately be described by a universal capacity
function equivalent to sparsely labeled `2 balls with
radii Rg and dimensions Dg , as demonstrated by
our numerical evaluations (Sec. VI). Conversely,
when f Rg2 1 , the capacity is low and close to
1
D where D is the dimensionality of the manifold
affine subspace, even for extremely small f .
• Our theory provides for the first time, quantitative and qualitative predictions for the perceptron
classification of realistic data structures. However,
application to real data may require further extensions of the theory and are discussed in Section VII.
Together, the theory makes an important contribution to the development of statistical mechanical
theories of neural information processing in realistic conditions.
MODEL OF MANIFOLDS
Manifolds in affine subspaces: We model a set of
P perceptual manifolds corresponding to P perceptual
object. Each manifold M µ for µ = 1, . . . , P consists of a
compact subset of an affine subspace of RN with affine
dimension D with D < N . A point on the manifold
xµ ∈ M µ can be parameterized as:
~ =
xµ (S)
D+1
X
Si uµi ,
(1)
i=1
where the uµi are a set of orthonormal bases of the (D+1)dimensional linear subspace containing M µ . The D + 1
components Si represents the coordinates of the manifold
point within this subspace and are constrained to be in
~ ∈ S. The bold notation for xµ and uµ indicates
the set S
i
~
they are vectors in RN whereas the arrow notation for S
D+1
. The set S defines the
indicates it is a vector in R
shape of the manifolds and encapsulates the affine constraint. For simplicity, we will first assume the manifolds
have the same geometry so that the coordinate set S is
the same for all the manifolds; extensions that consider
heterogeneous geometries are provided in Sec. VI A.
We study the separability of P manifolds into two classes,
denoted by binary labels y µ = ±1, by a linear hyperplane
passing through the origin. A hyperplane is described by
2
a weight vector w ∈ RN , normalized so kwk = N and
the hyperplane correctly separates the manifolds with a
margin κ ≥ 0 if it satisfies,
y µ w · xµ ≥ κ
(2)
4
for all µ and xµ ∈ M µ . Since linear separability is
a convex problem, separating the manifolds is equivalent
to separating o
the convex hulls, conv (M µ ) =
n
~ |S
~ ∈ conv(S) , where
xµ (S)
conv (S) =
(D+1
X
i=1
~i | S
~i ∈ S, αi ≥ 0,
αi S
D+1
X
)
αi = 1 .
i=1
(3)
The position of an affine subspace relative to the origin
can be defined via the translation vector that is closest
to the origin. This orthogonal translation vector cµ is
perpendicular to all the affine displacement vectors in
M µ , so that all the points in the affine subspace have
equal projections on cµ , i.e., ~xµ · cµ = kcµ k for all xµ ∈
M µ (Fig. 2). We will further assume for simplicity that
the P normskcµ k are all the same and normalized to 1.
To investigate the separability properties of manifolds, it
is helpful to consider scaling a manifold M µ by an overall
scale factor r without changing its shape. We define the
~ 0 ∈ S by a scalar r > 0, by
scaling relative to a center S
(D+1
)
X
µ
µ
0
0
~∈S
rM =
S + r Si − S
u |S
(4)
i
i
i
i=1
When P
r → 0, the manifold rM µ converges to a point:
µ
x0 = i Si0 uµi . On the other hand, when r → ∞, the
manifold rM µ spans the entire affine subspace. If the
manifold is symmetric (such as for an ellipsoid), there
is a natural choice for a center. We will later provide
an appropriate definition for the center point for general,
asymmetric manifolds. In general, the translation vector
~ 0 need not coincide as shown in Fig. 2.
~c and center S
However, we will also discuss later the special case of
centered manifolds in which the translation vector and
center do coincide.
Bounds on linear separability of manifolds: For
dichotomies of P input points in RN at zero margin, κ =
0, the number of dichotomies that can be separated by a
linear hyperplane through the origin is given by [11]:
C0 (P, N ) = 2
N
−1
X
k=0
−1
CP
≤ 2P
k
(5)
n!
is the binomial coefficient for n ≥
where Cnk = k!(n−k)!
k, and zero otherwise. This result holds for P input
vectors that obey the mild condition that the vectors
are in general position, namely that all subsets of input
vectors of size p ≤ N are linearly independent.
For large P and N , the probability 21P C0 (P, N ) of a dichotomy being linearly separable depends only upon the
P
and exhibits a sharp transition at the critical
ratio N
value of α0 = 2. We are not aware of a comprehensive extension of Cover’s counting theorem for general
manifolds; nevertheless, we can provide lower and upper
bounds on the number of linearly realizable dichotomies
by considering the limit of r → 0 and r → ∞ under the
following general conditions.
First, in the limit of r → 0, the linear separability of P
manifolds becomes equivalent to the separability of the P
centers. This leads to the requirement that the centers of
the manifolds, xµ0 , are in general position in RN . Second,
we consider the conditions under which the manifolds are
linearly separable when r → ∞ so that the manifolds
span complete affine subspaces. For a weight vector w
to consistently assign the same label to all points on an
affine subspace, it must be orthogonal to all the displacement vectors in the affine subspace. Hence, to realize a
dichotomy of P manifolds when r → ∞, the weight vector w must lie in a null space of dimension N −Dtot where
Dtot is the rank of the union of affine displacement vectors. When the basis vectors uµi are in general position,
then Dtot = min (DP, N ). Then for the affine subspaces
to be separable, P D < N is required and the projections of the P orthogonal translation vectors need also
be separable in the N − Dtot dimensional null space. Under these general conditions, the number of dichotomies
for D-dimensional affine subspaces that can be linearly
separated, CD (P, N ), can be related to the number of
dichotomies for a finite set of points via:
CD (P, N ) = C0 (P, N − P D).
(6)
From this relationship, we conclude that the ability to
linearly separate D-dimensional affine subspaces exhibits
a transition from always being separable to never being
2
P
= 1+2D
for large P and
separable at the critical ratio N
N (see Supplementary Materials, SM, Sec. S1).
For general D-dimensional manifolds with finite size, the
number of dichotomies that are linearly separable will
be lower bounded by CD (P, N ) and upper bounded by
C0 (P, N ). We introduce the notation, αM (κ), to denote
P
the maximal load N
such that randomly labeled manifolds are linearly separable with a margin κ, with high
probability. Therefore, from the above considerations, it
follows that the critical load at zero margin, αM (κ = 0),
is bounded by,
1
1
−1
≤ αM
(κ = 0) ≤ + D.
2
2
(7)
These bounds highlight the fact that in the large N limit,
the maximal number of separable finite-dimensional manifolds is proportional to N , even though each consists of
an infinite number of points. This sets the stage of a
P
statistical mechanical evaluation of the maximal α = N
, where P is the number of manifolds, and is described
in the following Section.
5
III.
STATISTICAL MECHANICAL THEORY
In order to make theoretical progress beyond the bounds
above, we need to make additional statistical assumptions about the manifold spaces and labels. Specifically,
we will assume that the individual components of uµi are
drawn independently and from identical Gaussian distributions with zero mean and variance N1 , and that the
binary labels y µ = ±1 are randomly assigned to each
manifold with equal probabilities. We will study the thermodynamic limit where N, P → ∞, but with a finite load
P
. In addition, the manifold geometries as speciα = N
fied by the set S in RD+1 , and in particular their affine
dimension, D, is held fixed in the thermodynamic limit.
Under these assumptions, the bounds in Eq. (7) can be
extended to the linear separability of general manifolds
with finite margin κ, and characterized by the reciprocal
−1
of the critical load ratio αM
(κ),
−1
α0−1 (κ) ≤ αM
(κ) ≤ α0−1 (κ) + D
(8)
where α0 (κ) is the maximum load for separation of random i.i.d. points with a margin κ given by the Gardner
theory [12],
α0−1 (κ) =
Z
κ
−∞
Dt (t − κ)2
(9)
2
t
√1 e− 2
2π
. For many interwith Gaussian measure Dt =
esting cases, the affine dimension D is large and the gap
in Eq. (8) is overly loose. Hence, it is important to derive an estimate of the capacity for manifolds with finite
sizes and evaluate the dependence of the capacity and
the nature of the solution on the geometrical properties
of the manifolds as shown below.
A.
As shown in Appendix A, we prove that the general form
of the inverse capacity, exact in the thermodynamic limit,
is:
−1
αM
(κ) = hF (T~ )iT~
(11)
2
~ ·S
~ − κ ≥ 0, ∀S
~∈S
~ − T~ | V
where F (T~ ) = minV~
V
and h. . .iT~ is an average over random D + 1 dimensional
vectors T~ whose components are i.i.d. normally distributed Ti ∼ N (0, 1). The components of the vector
~ represent the signed fields induced by the solution
V
vector w on the D + 1 basis vectors of the manifold. The
Gaussian vector T~ represents the part of the variability
~ due to quenched variability in the manifolds basis
in V
vectors and the labels, as will be explained in detail
below.
The inequality constraints in F can be written equivalently, as a constraint on the point on the mani~ . We therefore
fold with minimal projection on V
~
consider
o support function of S, gS (V ) =
n the concave
~ ·S
~|S
~ ∈ S , which can be used to write the
min ~ V
S
constraint for F (T~ ) as
F (T~ ) = min
~
V
~ − T~
V
2
~
| gS (V ) − κ ≥ 0
(12)
~ ) is easily mapped to
Note that this definition of gS (V
the conventional convex support function defined via the
max operation[16].
Karush-Kuhn-Tucker (KKT) conditions: To gain
insight into the nature of the maximum margin solution,
it is useful to consider the KKT conditions of the convex
optimization in Eq. 12 [16]. For each T~ , the KKT condi~ for F (T~ )
tions that characterize the unique solution of V
is given by:
Mean field theory of manifold separation
capacity
~ = T~ + λS̃(T~ )
V
(13)
where
Following Gardner’s framework [12, 13], we compute the
statistical average of log Z, where Z is the volume of the
space of the solutions, which in our case can be written
as:
Z
2
Z = dN wδ kwk − N Πµ,xµ ∈M µ Θ (y µ w · xµ − κ) ,
(10)
Θ (·) is the Heaviside function to enforce the margin constraints in Eq. (2), along with the delta function to en2
sure kwk = N . In the following, we focus on the properties of the maximum margin solution, namely the solution
for the largest load αM for a fixed margin κ, or equivalently, the solution when the margin κ is maximized for
a given αM .
λ≥0
(14)
~)−κ≥0
gS (V
i
~)−κ =0
λ gS (V
h
The vector S̃(T~ ) is
a subgradient of the support function
~ , S̃(T~ ) ∈ ∂gS V
~ [16], i.e., it is a point on the convex
at V
~ . When the suphull of S with minimal overlap with V
~
port function is differentiable, the subgradient ∂gS V
is unique and is equivalent to the gradient of the support
function,
~ ) = arg min V
~ ·S
~
S̃(T~ ) = ∇gS (V
~
S∈S
(15)
6
Since the support function is positively homogeneous,
~ ) = γgS (V
~ ) for all γ > 0; thus, ∇gS (V
~ ) depends
gS (γ V
~ such that
only on the unit vector V̂ . For values of V
~ ) is not differentiable, the subgradient is not unique,
gS (V
but S̃(T~ ) is defined uniquely as the particular subgradient that obeys the KKT conditions, Eq. 13-14. In the
latter case, S̃(T~ ) ∈ conv (S) may not reside in S itself.
From Eq. (13), we see that the capacity can be written
in terms of S̃(T~ ) as,
F (T~ ) = λS̃(T~ )
2
.
(16)
The scale factor λ is either zero or positive, corresponding
~ ) − κ is positive or zero. If gS (V
~ ) − κ is
to whether gS (V
~ = T~ and T~ · S̃(T~ ) −
positive , then λ = 0 meaning that V
~
~
~ 6= T~ . In
κ > 0. If T · S̃(T ) − κ < 0 , then, λ > 0 and V
2
this case, multiplying 13 with S̃(T~ ) yields λ S̃(T~ ) =
T~ · S̃(T~ ) − κ. Thus, λ obeys the self consistent equation,
h
λ=
i
T~ · S̃(T~ ) − κ
+
S̃(T~ )
(17)
2
where the function [x]+ = max(x, 0).
B.
equation, it follows that α−1 =
N
P
=
λ2 S̃
2
which
yields the KKT expression for the capacity, see Eqs. 11
and 16.
The KKT relations above, are self-consistent equations
for the statistics of λµ and S̃ µ . The mean field theory derives the appropriate statistics from self-consistent equations of the fields on a single manifold. To see this,
consider projecting the solution vector w onto the affine
subspace of one of the manifolds, say µ = 1 . We de~ 1 , as V 1 = y 1 w · u1 ,
fine a D + 1 dimensional vector, V
i
i
i = 1, ..., D + 1, which are the signed fields of the solution
on the affine basis vectors of the µ = 1 manifold. Then,
~ 1 = λ1 S̃ 1 + T~ where T~ represents
Eq. (18) reduces to V
the contribution from all the other manifolds and since
their subspaces are randomly oriented, this contribution
is well described as a random Gaussian vector. Finally,
self consistency requires that for fixed T~ , S̃ 1 is such that
~ 1 and represents a point
it has a minimal overlap with V
residing on the margin hyperplane, otherwise it will not
contribute to the max margin solution. Thus Eq. 13
is just the decomposition of the field induced on a specific manifold decomposed into the contribution induced
by that specific manifold along with the contributions
coming from all the other manifolds. The self consistent equations 14 as well as 17 relating λ to the Gaussian
statistics of T~ then naturally follow from the requirement
that S̃ 1 represents a support vector.
Mean field interpretation of the KKT relations
C.
The KKT relations have a nice interpretation within the
framework of the mean field theory. The maximum margin solution vector can always be written as a linear combination of a set of support vectors. Although there are
infinite numbers of input points in each manifold, the solution vector can be decomposed into P vectors, one per
manifold,
w=
P
X
µ=1
λµ y µ x̃µ , λµ ≥ 0
(18)
where x̃µ ∈ conv (M µ ) is a vector in the convex hull of
the µ-th manifold. In the large N limit, these vectors
are uncorrelated with each other. Hence, squaring this
equation yields,
2
kwk = N =
P
X
λ2µ S̃ µ
2
(19)
µ=1
where S̃ µ are the coordinates of x̃µ in the µ-th affine
subspace of the µ-th manifold, see Eq. (1). From this
Anchor points and manifold supports
The vectors x̃µ contributing to the solution, Eq. (18),
play a key role in the theory. We will denote them or
equivalently their affine subspace components, S̃ µ , as the
manifold anchor points. For a particular configuration of
manifolds, the manifolds could be replaced by an equivalent set of P anchor points to yield the same maximum
margin solution. It is important to stress, however, that
an individual anchor point is determined not only by the
configuration of its associated manifold but also by the
random orientations of all the other manifolds. For a
fixed manifold, the location of its anchor point will vary
with the relative configurations of the other manifolds.
This variation is captured in mean field theory by the
dependence of the anchor point S̃ on the random Gaussian vector T~ .
In particular, the position of the anchor point in the convex hull of its manifold reflects the nature of the relation
between that manifold and the margin planes. In general,
a fraction of the manifolds will intersect with the margin
hyperplanes, i.e., they have non-zero λ; these manifolds
are the support manifolds of the system. The nature of
this support varies and can be characterized by the dimension, k, of the span of the intersecting set of conv (S)
7
(a)
(b)
=0
cone(S)
nP
D+1
o
~i | S
~i ∈ S, αi ≥ 0 , see Fig. 3. The shifted
αi S
polar cone of S, denoted Sκ◦ , is defined as the convex set
of points given by,
n
o
~ ∈ RD+1 | U
~ ·S
~ + κ ≤ 0, ∀S
~∈S
Sκ◦ = U
(20)
>0
i=1
cone(S)
S
S
S̃
S̃
S̃
S̃
T~
T~
~c
~
V
S=0
~
V
S>0
~ + λS̃ at (a) zero
Figure 3. Conic decomposition of −T~ = −V
margin κ = 0 and (b) non-zero margin κ > 0. Given a random
~ − T~ is minimized while
~ is found such that V
Gaussian T~ , V
~ being on the polar cone Sκ◦ . λS̃ is in cone(S) and S̃ is
−V
the anchor point, a projection on the convex hull of S.
with the margin planes. Some support manifolds, which
we call touching manifolds, intersect with the margin hyperplane only with their anchor point. They have a support dimension k = 1, and their anchor point S̃ is on the
boundary of S. The other extreme is fully supporting
manifolds which completely reside in the margin hyperplane. They are characterized by k = D + 1. In this case,
~ is parallel to the translation vector ~c of S. Hence, all
V
the points in S are support vectors, and all have the
~ . The anchor point, in this case,
same overlap, κ, with V
is the unique point in the interior of conv(S) which obeys
the self consistent equation, 13, namely that it balances
the contribution from the other manifolds to zero out
~ that are orthogonal to ~c. In the
the components of V
case of smooth convex hulls (i.e., when S is strongly convex), no other manifold support configurations exist. For
other types of manifolds, there are also partially supporting manifolds, whose convex hull intersection with the
margin hyperplanes consist of k dimensional faces with
1 < k < D + 1. The associated anchor points then reside
inside the intersecting face. For instance, k = 2 implies
that S̃ lies on an edge whereas k = 3 implies that S̃ lies
on a planar 2-face of the convex hull. Determining the
dimension of the support structure that arises for various
T~ is explained below.
D.
Conic decomposition
The KKT conditions can also be interpreted in terms of
the conic decomposition of T~ , which generalizes the notion of the decomposition of vectors onto linear subspaces
and their null spaces via Euclidean projection. The convex cone of the manifold S is defined as cone (S) =
and is illustrated for κ = 0 and κ > 0 in Fig. 3. For
κ = 0, Eq. (20) is simply the conventional polar cone
of S [17]. Equation (13) can then be interpreted as
the decomposition of −T~ into the sum of two compo~ , i.e., its Euclidean
nent vectors: one component is −V
projection onto Sκ◦ ; the other component λS̃(T~ ) is located in cone (S). When κ = 0, the Moreau decomposition theorem states
that
the two components are
~
~
perpendicular: V · λS̃(T ) = 0 [18]. For non-zero κ,
the two components
need not be perpendicular but obey
~ · λS̃(T~ ) = κλ.
V
The position of vector T~ in relation to the cones, cone(S)
and Sκ◦ , gives rise to qualitatively different expressions
for F T~ and contributions to the solution weight vector and inverse capacity. These correspond to the different support dimensions mentioned above. In particular,
~ and λ = 0 so the supwhen -T~ lies inside Sκ◦ , T~ = V
port dimension, k = 0. On the other hand, when −T~
~ = κ~c and the manifold is fully
lies inside cone (S), V
supporting, k = D + 1.
E.
Numerical solution of the mean field equations
The solution of the mean field equations consists of two
stages. First, S̃ is computed for a particular T~ and then
the relevant contributions to the inverse capacity are averaged over the Gaussian distribution of T~ . For simple
geometries such as `2 ellipsoids, the first step may be
solved analytically. However, for more complicated geometries, both steps need to be performed numerically.
~ and S̃ for a given T~
The first step involves determining V
by solving the quadratic semi-infinite programming problem (QSIP), Eq. (12), over the manifold S which may
contain infinitely many points. A novel “cutting plane”
method has been developed to efficiently solve the QSIP
problem, see SM (Sec. S4). Expectations are computed
by sampling Gaussian T~ in D + 1 dimensions and taking
the appropriate averages, similar to procedures for other
mean field methods. The relevant quantities corresponding to the capacity are quite concentrated and converge
quickly with relatively few samples.
In following Sections, we will also show how the mean
field theory compares with computer simulations that
numerically solve for the maximum margin solution of
realizations of P manifolds in RN , given by Eq. (2), for
a variety of manifold geometries. Finding the maximum
8
margin solution is challenging as standard methods to
solving SVM problems are limited to a finite number of
input points. We have recently developed an efficient
algorithm for finding the maximum margin solution in
manifold classification and have used this method in the
present work (see [19] and SM (Sec. S5)).
IV.
A.
MANIFOLD GEOMETRY
Longitudinal and intrinsic coordinates
In this section, we address how the capacity to separate a set of manifolds can be related to their geometry, in particular to their shape within the D-dimensional
affine subspace. Since the projections of all points in a
manifold onto the translation vector, cµ , are the same,
~xµ · cµ = kcµ k = 1, it is convenient to parameterize the
D + 1 affine basis vectors such that uµD+1 = cµ . In these
coordinates, the D + 1 dimensional vector representation
~ = (0, 0, . . . , 0, 1). This parameterization is conof cµ is C
venient since it constrains the manifold variability to be
~ while the D +1 coordinate
in the first D components of S
is a longitudinal variable measuring the distance of the
manifold affine subspace from the origin. We can then
write the D + 1 dimensional vectors T~ = (~t, t0 ) where
~ V
~ = (~v , v0 ), S
~ =(~s, 1), S̃ = (s̃, 1) , where
t0 = T~ · C,
lower case vectors denote vectors in RD . We will also refer to the D-dimensional intrinsic vector, s̃, as the anchor
point. In this notation, the capacity, Eq. (16)-(17) can
be written as,
2
Z
Z
−t0 − ~t · s̃(~t, t0 ) + κ +
−1
(21)
αM (κ) = D~t Dt0
2
1 + s̃(~t, t0 )
It is clear from this form that when D = 0, or when all
the vectors in the affine subspace obey s̃(~t, t0 ) = 0 , the
capacity reduces to the Gardner result, Eq. (9). Since
~ · S̃=~v · s̃ + v0 , for all S̃ , g(V
~ ) = g(~v ) + v0 , and the
V
support function can be expressed as,
g(~v ) = min {~v · ~s | ~s ∈ S} .
~
s
ttouch (~t) = −g(~t)
(23)
Non-zero contributions to the capacity only occur outside
this interior regime when g(~t) + t0 − κ < 0, in which case
λ > 0. Thus, for all support dimensions k > 0, the
solution for ~v is active, satisfying the equality condition,
g(~v ) + v0 − κ = 0, so that from Eq. (13):
g(~v ) + t0 − κ − λ = 0,
(24)
outside the interior regime.
Fully supporting manifolds (k = D + 1): When t0
is sufficiently negative, v0 = κ, ~v = 0 and λ = t0 − κ.
The anchor point s̃(T~ ) which obeys the KKT equations
resides in the interior of conv (S),
s̃(~t, t0 ) =
−~t
κ − t0
(25)
For a fixed ~t, t0 must be negative enough, t0 − κ < tfs ,
where
−~t
∈ conv(S)
(26)
tfs (~t) = arg max t0 |
κ − t0
guaranteeing that s̃(~t, t0 ) ∈ conv (S). The contribution
of this regime to the capacity is
(22)
The KKT relations can be written as ~v = ~t +
λ~s where λ = v0 − t0 ≥ 0, g(~v ) ≥ κ − v0 ,
λ [g(~v ) + v0 − κ] = 0, and s̃ minimizes the overlap with ~v . The resultant equation
for λ (or v0 )
2
is λ = −t0 − ~t · s̃(~t, t0 ) + κ / 1 + s̃(~t, t0 )
which
+
agrees with Eq. 17.
B.
in the previous section. To do this, we fix the random
vector ~t and consider the qualitative change in the anchor
point, s̃(~t, t0 ) as t0 decreases from +∞ to −∞.
Interior manifolds (k = 0): For sufficiently positive t0 ,
the manifold is interior to the margin plane, i.e., λ = 0
with corresponding support dimension k = 0. Although
not contributing to the inverse capacity and solution vector, it is useful to associate anchor points to these manifolds defined as the closest point on the manifold to the
margin plane: s̃(~t) = arg min~s∈S ~t ·~s = ∇g(~t) since ~v = ~t.
This definition ensures continuity of the anchor point for
all λ ≥ 0.
This interior regime holds when g(~t) > κ − t0 , or equivalently for t0 − κ > ttouch (~t) where
Types of supports
Using the above coordinates, we can elucidate the conditions for the different types of support manifolds defined
F (T~ ) = ~v − ~t
2
+ (v0 − t0 )2 = ~t
2
+ (κ − t0 )2 , (27)
see Eq. (12). Finally, for values of t0 such that tfs (~t) ≤
t0 − κ ≤ ttouch (~t), the manifolds are partially supporting with support dimension 1 ≤ k ≤ D. Examples of
different supporting regimes are illustrated in Figure 4.
C.
Effects of size and margin
We now discuss the effect of changing manifold size and
the imposed margin on both capacity and geometry. As
9
described by Eq. (4), change in the manifold size corresponds to scaling every s̃ by a scalar r.
Small size: If r → 0, the manifold shrinks to a point
(the center), whose capacity reduces to that of isolated
points. However, in the case where D 1 , the capacity
may be affected by the manifold structure even if r 1,
see section IV F. Nevertheless, the underlying support
structure is simple. When r is small, the manifolds have
only two support configurations. For t0 − κ > 0, the
manifold is interior with: λ = 0 , v0 = t0 and ~v = ~t.
When t0 − κ < 0, the manifold becomes touching with
support dimension k = 1. In that case, because s̃ has a
small magnitude, ~v ≈ ~t, λ ≈ κ − t0 , and v0 ≈ κ. Thus,
in both cases, ~v is close to the Gaussian vector ~t. The
probability of other configurations vanishes.
Large size: In the large size limit r → ∞, separating the manifolds is equivalent to separating the affine
subspaces. As we show in Appendix C, when r 1,
there are two Rmain support structures. With probabil∞
ity H (−κ) = −κ Dt0 the manifolds are fully supporting,
namely the underlying affine subspaces are parallel to the
margin plane. This regime contributes to the inverse capacity an amount H (−κ) D+α0−1 (κ). The other regimes
(touching or partially supporting) are such that the angle between the affine subspace and the margin plane is
almost zero, and contribute an amount H(κ)D to the
inverse capacity. Combining the two contributions, we
−1
obtain for large sizes, αM
= D + α0−1 (κ), consistent with
Eq. (8).
Large margin: For a fixed T~ , Eq. (26) implies that
larger κ increases the probability of being in the supporting regimes. Increasing κ also shrinks the magnitude of s̃
according to Eq. (25). Hence, when κ 1, the capacity
becomes similar to that of P random points and the corresponding capacity is given by αM ≈ κ−2 , independent
of manifold geometry.
Manifold centers: The theory of manifold classification
described in Sec. III does not require the notion of a
manifold center. However, understanding how scaling
the manifold sizes by a parameter r in (4) affects their
capacity, the center points about which the manifolds
are scaled need to be defined. For many geometries, the
center is a point of symmetry such as for an ellipsoid.
For general manifolds, a natural definition would be the
center of mass of the anchor points S̃(T~ ) averaging over
the Gaussian measure of T~ . We will adopt here a simpler
definition for the center provided by the Steiner point for
~ 0 = (~s0 , 1) with
convex bodies [20], S
~s0 = ∇gS (~t)
~
t
(28)
and the expectation is over the Gaussian measure of
~t ∈ RD . This definition coincides with the center of mass
of the anchor points when the manifold size is small. Furthermore, it is natural to define the geometric properties
~v = ~t
(a1)
(a2)
~v
~t
(a3)
~t
~v
Interior (k=0)
(b1)
~v = ~t
s̃
s̃
s̃
~v = 0
Touching (k=1)
(b2)
(b3)
~v
Fully supporting (k=3)
~v
~t
(b4)
~t
~t
s̃
s̃
Interior (k=0)
Touching (k=1)
s̃
s̃
Partially supporting (k=2)
~v = 0
Fully supporting (k=3)
Figure 4. Determining the anchor points of the Gaussian
distributed vector (~t, t0 ) onto the convex hull of the manifold, denoted as s̃(~t, t0 ). Here we show for the same vector
~t, the change in s̃(~t, t0 ) as t0 decreases from +∞ to −∞.
(a) D = 2 strictly convex manifold. For sufficiently negative −t0 the vector −~t obeys the constraint g(~t) − t0 > κ,
hence −~t = −~v and the configuration corresponds to an interior manifold (support dimension k = 0). For intermediate
values where (ttouch > t0 − κ > tfs ), (−~t, −t0 ) violates the
above constraints and s̃ is a point on the boundary of the
manifold that maximizes the projection on the manifold of
a vector (−~v , −v0 ) that is the closest to (−~t, −t0 ) and obeys
g(~v ) + t0 − λ = κ. Finally, for larger values of −t0 , ~v = 0
and s̃ is a point in the interior of the manifold in the direction of −~t (fully supporting with k = 3). (b) D = 2 square
manifold. Here, in both the interior and touching regimes, s̃
is a vertex of the square. In the fully supporting regime, the
anchor point is in the interior and collinear to −~t. There is
also a partially supporting regime when −t0 is slightly below
tfs . In this regime, −~v is perpendicular to one of the edges
and s̃ resides on an edge, corresponding to manifolds whose
intersection with the margin planes are edges (k = 2).
of manifolds in terms of centered manifolds where the
manifolds are shifted within their affine subspace so that
the center and orthogonal translation vector coincide, i.e.
~s0 = ~c with ~s0 = k~ck = 1. This means that all lengths
are defined relative to the distance of the centers from
the origin and the D-dimensional intrinsic vectors ~s give
the offset relative to the manifold center.
D.
Manifold anchor geometry
The capacity equation (21) motivates defining geometrical measures of the manifolds, which we call manifold
anchor geometry. Manifold anchor geometry is based on
the statistics of the anchor points s̃(~t, t0 ) induced by the
Gaussian random vector (~t, t0 ) which are relevant to the
capacity in Eq. (21). These statistics are sufficient for
determining the classification properties and supporting
structures associated with the maximum margin solution.
We accordingly define the manifold anchor radius and dimension as:
Manifold anchor radius: denoted RM , is defined by
the mean squared length of s̃(~t, t0 ),
D
E
2
2
RM
= s̃(~t, t0 )
(29)
t~,t0
10
Manifold anchor dimension: DM , is given by
D
2 E
DM = ~t · ŝ(~t,t0 )
t~,t0
(a)
(30)
where ŝ is a unit vector in the direction of s̃. The anchor
dimension measures the angular spread between ~t and
the corresponding anchor point s̃ in D dimensions.
D 2 E Note
that the manifold dimension obeys DM ≤ ~t
= D.
Whenever there is no ambiguity, we will call RM and DM
the manifold radius and dimension, respectively.
These geometric descriptors offer a rich description of
the manifold properties relevant for classification. Since
~v and s̃ depend in general on t0 − κ, the above quantities
are averaged not only over ~t but also over t0 . For the
same reason, the manifold anchor geometry also depends
upon the imposed margin.
E.
Gaussian geometry
We have seen that for small manifold sizes, ~v ≈ ~t, and the
anchor points s̃ can be approximated by s̃(~t) = ∇gS (t̂).
Under these conditions, the geometry simplifies as shown
in Fig. 5. For each Gaussian vector ~t ∈ RD , s̃(~t) is
the point on the manifold that first touches a hyperplane normal to -~t as it is translated from infinity towards the manifold. Aside from a set of measure zero,
this touching point is a unique point on the boundary of
conv(S). This procedure is similar to that used to define
the well known Gaussian mean-width, which in our notation equals w(S) = −2 gS (~t) ~t[21]. Note that for small
sizes, the touching point does not depend on t0 or κ so its
statistics are determined only by the shape of conv (S)
relative to its center. This motivates defining a simpler
manifold Gaussian geometry denoted by the subscript g,
which highlights its dependence over a D dimensional
Gaussian measure for ~t.
Gaussian radius: denoted by Rg , measures the mean
square amplitude of the Gaussian anchor point, s̃g (~t) =
∇gS (t̂):
D
2E
Rg2 = s̃g ~t
(31)
~
t
where the expectation is over D-dimensional Gaussian ~t.
Gaussian dimension: Dg , is defined by,
D
2 E
Dg = ~t · ŝg ~t
~
t
(32)
where ŝg is the unit vector in the direction of s̃g . While
Rg2 measures the total variance of the manifold, Dg measures the angular spread of the manifold, through the
~
statistics of the angle between
D 2 E t and its Gaussian anchor
point. Note that Dg ≤ ~t
= D.
~t
s̃g (~t)
(b)
~t
s̃g (~t)
(c)
~t
s̃g (~t)
Figure 5. Gaussian anchor points: Mapping from ~t to points
s̃g (~t), showing the relation between −~t and the point on the
manifold that touches a hyperplane orthogonal to ~t . D =
2 manifolds shown: (a) circle; (b) `2 ellipsoid; (c) polytope
manifold. In (c), only for values of ~t of measure zero (when
~t is exactly perpendicular to an edge), s̃g (~t) will lie along the
edge, otherwise it coincides with a vertex of the polytope. In
all cases, s̃g (~t) will be in the interior of the convex hulls only
for ~t = 0. Otherwise, it is restricted to their boundary.
It is important to note that even in this limit, our geometrical definitions are not equivalent to conventional
geometrical measures such as the longest chord or second order statistics induced from a uniform measure over
the boundary of conv (S). For the special
case of Ddimensional `2 balls with radius R, s̃g ~t is the point
on the boundary of the ball in the direction of ~t so that
Rg = R and Dg = D. However, for general manifolds, Dg
can be much smaller than the manifold affine dimension
D as illustrated in some examples later.
We recapitulate the essential difference between the
Gaussian geometry and the full manifold anchor geometry. In the Gaussian case, the radius is an intrinsic property of the shape of the manifold in its affine subspace
and is invariant to changing its distance from the origin. Thus, scaling the manifold by a global scale factor
r as defined in Eq. (4) results in scaling Rg by the factor r. Likewise, the dimensionality Dg is invariant to a
global scaling of the manifold size. In contrast, the anchor geometry does not obey such invariance for larger
manifolds. The reason is that the anchor point depends
on the longitudinal degrees of freedom, namely on the
size of the manifold relative to the distance from the center. Hence, RM need not scale linearly with r, and DM
will also depend on r. Thus, the anchor geometry can
be viewed as describing the general relationship between
the signal (center distance) to noise (manifold variability)
in classification capacity. We also note that the manifold
anchor geometry automatically accounts for the rich support structure described in Section IV D. In particular,
as t0 decreases, the statistics of the anchor points change
from being concentrated on the boundary of conv(S) to
its interior. Additionally for manifolds which are not
strictly convex and intermediate values of t0 , the anchor
statistics become concentrated on k-dimensional facets
of the convex hull corresponding to partially supported
manifolds.
We illustrate the difference between the two geometries
2D Ellipsoid Distribution of S, theta
11
(a)
$% = ', $) = '
(b)
1
Polar-constrained
0.98
Manifold Anchor
(c)
$% = ', $) = ). '
0.2
Gaussian
Gaussian
0.15
$% = ', $) = ). '
0.4
Gaussian
Polar-constrained
0.3
0.96
0.1
0.2
0.02
0.05
0.1
0
0
2.5
!"
5
0
0
2.5
!"
Gaussian
Polar-constrained
5
0
0
10
20
30
40
#
Figure 6. Distribution of ks̃k (norm of manifold anchor vectors) and θ = ∠(−~t, s̃) for D = 2 ellipsoids. (a) Distribution
of s̃ for `2 ball with R = 5. The Gaussian geometry is peaked
at ks̃k = r (blue), while ks̃k < r has non-zero probability for
the manifold anchor geometry (red). (b-c) D = 2 ellipsoids,
with radii R1 = 5 and R2 = 21 R1 . (b) Distribution of ks̃k
for Gaussian geometry (blue) and anchor geometry (red). (c)
Corresponding distribution of θ = ∠(−~t, s̃).
in Fig. 6 with two simple examples: a D = 2 `2 ball and
a D = 2 `2 ellipse. In both cases we consider the distribution of ks̃k and θ = cos−1 −t̂ · ŝ , the angle between
−~t and s̃. For the ball with radius r, the vectors −~t and s̃
are parallel so the angle is always zero. For the manifold
anchor geometry, s̃ may lie inside the ball in the fully
supporting region. Thus the distribution of ks̃k consists
of a mixture of a delta function at r corresponding to
the interior and touching regions and a smoothly varying
distribution corresponding to the fully supporting region.
Fig. 6 also shows the corresponding distributions for a
two dimensional ellipsoid with major and minor radius,
R1 = r and R2 = 21 r. For the Gaussian geometry, the
distribution of ks̃k has finite support between R1 and R2 ,
whereas the manifold anchor geometry has support also
below R2 . Since ~t and s̃ need not be parallel, the distribution of the angle varies between zero and π4 , with
the manifold anchor geometry being more concentrated
near zero due to contributions from the fully supporting
regime. In Section (VI), we will show how the Gaussian geometry becomes relevant even for larger manifolds
when the labels are highly imbalanced, i.e. sparse classification.
F.
Geometry and classification of high dimensional
manifolds
In general, linear classification as expressed in Eq. (21)
depends on high order statistics of the anchor vectors
s̃(~t, t0 ). Surprisingly, our analysis shows that for high dimensional manifolds, the classification capacity can be
described in terms of the statistics of RM and DM alone.
This is particularly relevant as we expect that in many
applications, the affine dimension of the manifolds is
large. More specifically, we define high-dimensional manifolds as manifolds where the manifold dimension is large,
i.e., DM 1 (but still finite in the thermodynamic limit).
In practice we find that DM & 5 is sufficient. Our analysis below elucidates the interplay between size and dimension, namely, how small RM needs to be for high dimensional manifolds to have a substantial classification
capacity.
In the high-dimensional regime, the mean field equations
simplify due to self averaging of terms involving sums
of components ti and s̃i . The two quantity, ~t · s̃, that
appears in the capacity, (21), can be approximated as
−~t · s̃ ≈ κM where
√ we introduce the effective2 manifold
2
,
margin κM = RM DM . Combined with ks̃k ≈ RM
we obtain
αM (κ) ≈ α0
κ + κM
p
2
1 + RM
!
(33)
where α0 is the capacity for P random points, Eq. (9).
To gain insight into this result we note that the effective
margin on the center is its mean distance from the point
closest to the
√ margin plane, s̃, which is roughly the mean
of ~t·s̃ ≈ RM DM . The denominator in Eq. (33) indicates
that
p the margin needs to be scaled by the input norm,
2 . In Appendix B, we show that Eq. (33) can
1 + RM
also be written as
αM (κ) ≈ αBall (κ, RM , DM )
(34)
namely, the classification capacity of a general high dimensional manifold is well approximated by that of `2
balls with dimension DM and radius RM .
Scaling regime: Eq. (33) implies that to obtain a finite capacity in the
√high-dimensional regime, the effective
margin κM = RM DM needs to be of order unity which
requires the
radius
to be small, scaling for large DM as
− 21
RM = O DM . In this scaling regime, the calculation
of the capacity and geometric properties are particularly
simple. As argued above, when the radius is small, the
components of s̃ are small, hence ~v ≈ ~t, and the Gaussian statistics for the geometry suffice. Thus, we can
replace RM and DM in Eqs. (33)- (34) by Rg and Dg ,
respectively. Note that in the scaling regime, the factor
proportional to 1 + Rg2 in Eq. (33) is the next order correction to the overall capacity sincep
Rg is small. Notably,
the margin in this regime, κg = Rg Dg , is equal to half
the Gaussian mean width of convex bodies, κg ≈ 12 w(S)
[22]. As for the support structure, since the manifold size
is small, the only significant contributions arise from the
interior (k = 0) and touching (k = 1) supports.
Beyond the scaling regime: When Rg is not small,
the anchor geometry, RM and DM , cannot be adequately
described by the Gaussian statistics, R
√g and Dg . In this
case, the manifold margin κM = RM DM is large and
Eq. (33) reduces to :
12
αM ≈
−2
RM
1+
DM
(35)
where we have used α0 (κ) ≈ κ−2 for large margins and
assumed that κ κM . For strictly convex high dimensional manifolds with RM = O(1), only the touching
regime (k = 1) contributes significantly to the geometry and hence to the capacity. For manifolds that are
not strictly convex, partially supporting solutions with
k D can also contribute to the capacity.
Finally, when RM is large, fully supporting regimes with
k ≈ D contribute to the geometry, in which case, the
manifold anchor dimension approaches the affine dimen1
,
sion DM → D and Eqs. (33) and (35)reduce to αM ≈ D
as expected.
V.
A.
EXAMPLES
Strictly convex manifolds: `2 ellipsoids
The family of `2 ellipsoids are examples of manifolds
which are strictly convex. Strictly convex manifold have
smooth boundaries that do not contain corners, edges, or
flats (see Appendix B). Thus, the description of the anchor geometry and support structures is relatively simple.
The reason is that the anchor vectors s̃(~t, t0 ) correspond
to either interior (k = 0), touching (k = 1), or fully supporting (k = D+1), while partial support (1 < k < D+1)
is not possible. The `2 ellipsoid geometry can be solved
analytically; nevertheless, because it has less symmetry
than the sphere, it exhibits some salient properties of
manifold geometry such as nontrivial dimensionality and
non-uniform measure of the anchor points.
We assume the ellipsoids are centered relative to their
symmetry centers as described in Sec. IV, and
PDcan be
parameterized by the set of points: M µ = xµ0 + i=1 si uµi
where
(
)
2
D
X
si
S = ~s |
≤1 .
(36)
Ri
i=1
The components of uµi and of the ellipsoid centers x0 µ
are i.i.d. Gaussian distributed with zero mean and variance √1N so that they are orthonormal in the large N
limit. The radii Ri represent the principal radii of the ellipsoids relative to the center. The anchor points s̃(~t, t0 )
can be computed explicitly (details in the Appendix B),
corresponding to three regimes.
Interior (k = 0), occurs when t0 − κ ≥ ttouch (~t) where,
v
uD
uX
ttouch (~t) = t
Ri2 t2i
(37)
i=1
Here λ = 0, resulting in zero contribution to the inverse
capacity.
The touching regime (k = 1) holds when ttouch (~t) > t0 −
κ > tfs (~t) and
v
u D 2
uX ti
~
(38)
tfs (t) = −t
Ri
i=1
Finally, the fully supporting regime (k = D + 1) occurs
when tfs (~t) > t0 − κ. The full expression for the capacity
for `2 ellipsoids is given in Appendix, Section B. Below,
we focus on the interesting cases of ellipsoids with D 1.
High-dimensional ellipsoids: It is instructive to apply the general analysis of high dimensional manifolds to
ellipsoids with D 1. We will distinguish between different size regimes by assuming that all the radii of the
ellipsoid are scaled by a global factor r. In the high dimensional regime, due to self-averaging, the boundaries
of the touching and fully supporting transitions can be
approximated by,
v
uD
uX
Ri2
(39)
ttouch = t
i=1
v
uD
uX
tfs = −t
Ri−2
(40)
i=1
√
both independent of ~t. Then as long as Ri D are not
large (see below), tfs → −∞, and the probability of fully
supporting vanishes. Discounting the fully supporting
regime, the anchor vectors s̃(~t, t0 ) are given by,
s̃i = −λi ti
λi =
(41)
Ri2
Z(1 + Ri2 )
(42)
PD
R2
where the normalization factor Z 2 = i=1 (1+Ri 2 )2 (see
i
Appendix B) . The capacity for high-dimensional ellipsoids, is determined via Eq. (33) with manifold anchor
radius,
2
RM
=
D
X
λ2i
(43)
i=1
and anchor dimension,
P
DM =
D
i=1
PD
i=1
λi
2
λ2i
(44)
13
Scaling regime: In the scaling regime where the radii
are small, the radius and dimension are equivalent to the
Gaussian geometry:
PD
Rg2
=
Ri4
i=1
PD
i=1
Ri2
R)
0.8
2
Ellipso
1.5
0.6
1
0.4
0.5
0.2
0
50
100
150
(c)
0
200
Dimension
100
R
r
101
102
0
% = 200
0
RR "
/R/$
0.8
M
100
0
10-1
1
150
50
10-2
(d)
200
%*
0.6
0.4
0.2
0
10-2 10-1
100
101
10
102
-2
10
(e2)
r=0.01
R0=1.00e-02
0.6
10
2
r
(e3)
r=1
R =1.00e+00
r=100
R =1.00e+02
0
1
0
R0
r
(e1)
(45)
(b)
1
M
The anchor covariance matrix: We can also compute
the covariance matrix of the anchor points. This matrix
is diagonal in the principal directions of the ellipsoid, and
its eigenvalues are λ2i . It is interesting to compare DM
with a well known measure of an effective dimension of a
covariance matrix, the participation ratio [23, 24]. Given
a spectrum of eigenvalues of the covariance matrix, in our
notation λ2i , we can define a generalized participation
P
2
( D λqi )
The conventional
ratio as P Rq = Pi=1
2q , q > 0.
D
i=1 λi
participation ratio uses q = 2, whereas DM uses q = 1 .
r=1
(a)
D
Note that RM /r and DM are not invariant to scaling the
ellipsoid by a global factor r, reflecting the role of the
fixed centers.
0
1
0.4
0.5
0.5
0.2
P
Dg =
D
i=1
Ri2
PD
2
Ri4
i=1
0
κg =
!1/2
Ri2
(47)
i=1
R
As can be seen, Dg and rg are invariant to scaling all the
radii by r, as expected. It is interesting to compare the
above Gaussian geometric parameters with the statistics
induced by a uniform measure on the surface of the ellipsoid. In this case, the covariance matrix has eigenvalues
P
λ2i = Ri2 and the total variance is equal to i Ri2 . In
contrast, in the Gaussian geometry, the eigenvalues of
the covariance matrix, λ2i , are proportional to Ri4 . This
and the corresponding expression (45) for the squared
radius is the result of a non-uniform induced measure on
the surface of the ellipse in the anchor geometry, even in
its Gaussian limit.
Beyond the scaling regime: When Ri = O(1), the
high dimensional ellipsoids all become touching manifolds
since ttouch → −∞ and tfs → ∞. The capacity is small
because κM 1, and is given by Eq. (35) with Eqs.
(43)-(44). Finally, when all Ri 1, we have,
2
RM
= PD
D
−2
i=1 Ri
=
1
hRi−2 ii
1
D+1
0
0
1
D+1
0
0
1
D+1
(46)
and the effective margin is given as,
D
X
0
(48)
Here RM scales with the global scaling factor, r, DM = D,
and α−1 = D. Although the manifolds are touching,
their angle relative to the margin plane is near zero.
Figure 7. Bimodal `2 ellipsoids. (a) Ellipsoidal radii Ri with
r = 1 (b) Classification capacity as a function of scaling factor
r: (blue lines) full mean field theory capacity, (black dashed)
approximation of the capacity given by equivalent ball with
RM and DM , (circles) simulation capacity, averaged over 5
repetitions, measured with 50 dichotomies per each repetition. (c) Manifold dimension DM as a function of r. (d)
Manifold radius RM relative to r. (e) Fraction of manifolds
with support dimension k for different values of r: k = 0 (interior), k = 1 (touching), k = D + 1 (fully supporting). (e1)
Small r = 0.01, where most manifolds are interior or touching
(e2) r = 1, where most manifolds are in the touching regime
(e3) r = 100, where the fraction of fully
qPsupporting manifolds
−2
is 0.085, predicted by H(−tfs ) = H(
i Ri ) (Eq. 40).
√
When Ri ≈ D or larger, the fully supporting transition tfs becomes order one and the probability of fully
supporting is significant.
In Fig. 7 we illustrate the behavior of high D-dimensional
ellipsoids, using ellipsoids with a bimodal distribution of
their principal radii, Ri : Ri = r, for 1 ≤ i ≤ 10 and
Ri = 0.1r, for 11 ≤ i ≤ 200 (Fig. 7(a)). Their properties are shown as a function of the overall scale r. Fig.
7(b) shows numerical simulations of the capacity, the full
mean field solution as well as the spherical high dimensional approximation (with RM and DM ). These calculations are all in good agreement, showing the accuracy
of the mean field theory and spherical approximation.
As seen in (b)-(d), the system is in the scaling regime
for r < 0.3. In this regime, the manifold dimension is
constant and equals Dg ≈ 14, as predicted by the par-
14
(a)
(b)
r=1
0
1
10
Ri
10
10-1
0
10
10
-2
-1
10
200 400 600 800 1000
(c)
M
750
! = 1023
RM/r
1000
D
(d)
Dimension
500
250
0
!'
10-2
100
r
102
12
10
8
6
4
2
0
10-3 10-2 10-1 100 101 102
r
-3
-2
-1
10 10 10
0
1
2
10 10 10
r
Figure 8. Ellipsoids with radii computed from realistic image
data. (a) SVD spectrum, Ri , taken from the readout layer of
GoogLeNet from a class of ImageNet images with N = 1024
and D = 1023. The radii are scaled by a factor r: Ri0 = rRi
for (b) Classification capacity as a function of r. (blue lines)
full mean field theory capacity, (black dashed) approximation
of the capacity as that of a ball with RM and DM from the
theory for ellipsoids, (circles) simulation capacity, averaged
over 5 repetitions, measured with 50 random dichotomies per
each repetition. (c) Manifold dimension as a function of r.
(d) Manifold radius relative to the scaling factor r, RM /r as
a function of r.
ticipation ratio, Eq. (46), and the manifold radius Rg
is linear with r, as expected from Eq. (45). The ratio
Rg
r ≈ 0.9 is close to unity, indicating that in the scaling
regime, the system is dominated by the largest radii. For
r > 0.3 the effective margin is larger than unity and the
system becomes increasingly affected by the full affine
dimensionality of the ellipsoid, as seen by the marked increase in dimension as well as a corresponding decrease
−1
= D.
in RrM . For larger r, DM approaches D and αM
Fig. 7(e1)-(e3) shows the distributions of the support
dimension 0 ≤ k ≤ D + 1. In the scaling regime, the
interior and touching regimes each have probability close
to 12 and the fully supporting regime is negligible. As
r increases beyond the scaling regime, the interior probability decreases and the solution is almost exclusively
in the touching regime. For very high values of r, the
fully supporting solution gains a substantial probability.
1
Note that the capacity decreases to approximately D
for
a value of r below where a substantial fraction of solutions
are fully supporting. In this case, the touching ellipsoids
all have very small angle with the margin plane.
Until now, we have assumed that the manifold affine subspace dimension D is finite in the limit of large ambient
dimension, N . In realistic data, it is likely that the data
manifolds are technically full rank, i.e. D+1 = N , raising
the question whether our mean field theory is still valid in
such cases. We investigate this scenario by computing the
capacity on ellipsoids containing a realistic distribution
of radii. We have taken as examples, a class of images
from the ImageNet dataset [25], and analyzed the SVD
spectrum of the representations of these images in the
last layer of a deep convolutional network, GoogLeNet
[26]. The computed radii are shown in Fig. 8(a) and
yield a value D = N − 1 = 1023. In order to explore the
properties of such manifolds, we have scaled all the radii
by an overall factor r in our analysis. Because of the decay in the distribution of radii, the Gaussian dimension
for the ellipsoid is only about Dg ≈ 15, much smaller
than D or N , implying that for small r the manifolds
are effectively low dimensional and the geometry is dominated by a small number of radii. As r increases above
r & 0.03, κM becomes larger than 1 and the solution
leaves the scaling regime, resulting in a rapid increase
in DM and a rapid falloff in capacity as shown in Fig.
−1
8(b-c). Finally, for r > 10, we have αM
≈ DM ≈ D
approaching the lower bound for capacity as expected.
The agreement between the numerical simulations and
the mean field estimates of the capacity illustrates the
relevance of the theory for realistic data manifolds that
can be full rank.
B.
Convex polytopes: `1 ellipsoids
The family of `2 ellipsoids represent manifolds that are
smooth and strictly convex. On the other hand, there
are other types of manifolds whose convex hulls are
not strictly convex. In this section, we consider a Ddimensional `1 ellipsoids. They are prototypical of convex
polytopes formed by the convex hulls of finite numbers of
points in RN . The D-dimensional `1 ellipsoid, is parameterized by radii {Ri } and specified by the convex set
PD
M µ = xµ0 + i=1 si uµi , with:
)
D
X
si
≤1 .
~s |
Ri
i=1
(
S=
(49)
Each manifold M µ ∈ RN is centered at xµ0 and consists
of a convex polytope with a finite number (2D) of vertices: {xµ0 ± Rk uµk , k = 1, ..., D}. The vectors uµi specify
the principal axes of the `1 ellipsoids. For simplicity, we
consider the case of `1 balls when all the radii are equal:
Ri = R. We will concentrate on the cases when `1 balls
are high-dimensional; the case for `1 balls with D = 2
was briefly described in [15]. The analytical expression
of the capacity is complex due to the presence of contributions from all types of supports, 1 ≤ k ≤ D + 1.
We address important aspects of the high dimensional
solution below.
15
L1 balls (new) – (d1) can be R=1e-2
High-dimensional `1 balls, R = O(1): When the radius R is small as in the scaling regime, the only contributing solution is the touching solution (k = 1). When
R increases, solutions with all values of k, 1 ≤ k ≤ D + 1
occur, and the support can be any face of the convex
polytope with dimension k. As R increases, the probability distribution p(k) over k of the solution shifts to larger
values. Finally, for large R, only two regimes dominate:
fully supporting (k = D+1) with probability H (−κ) and
partially supporting with k = D with probability H(κ).
We illustrate the behavior of `1 balls with radius r and
affine dimension D = 100. In Fig. 9, (a) shows the
linear classification capacity as a function of r. When
r → 0, the manifold approaches the capacity for isolated
1
points, αM ≈ 2, and when r → ∞, αM ≈ D
= 0.01.
The numerical simulations demonstrate that despite the
different geometry, the capacity of the `1 polytope is similar to that of a `2 ball with radius RM and dimension
DM . In (b), for the scaling regime when r < 0.1 = √1D ,
we see RM ≈ r, but when r 1, RM is much smaller
than r, despite the fact that all Ri of the polytope are
equal. This is because when r is not small, the various
faces and eventually interior of the polytope contribute
to the anchor geometry. In (c), we see DM ≈ 2 log D in
the scaling regime, while DM → D as r → ∞. In terms
of the support structures, when r = 0.001 in the scaling
regime (d1), most manifolds are either interior or touching. For intermediate sizes (d2), the support dimension
is peaked at an intermediate value, and finally for very
large manifolds (d3), most polytope manifolds are nearly
fully supporting.
1
0.4
0.5
0.2
0
0
10-2
100
!
0.6
DM
/R
7R8M/&
M
1.5
102
10
10
L
ℓ(1 balls,
*+,,-,R=1.00e-03
& = (0 − 2
0
10
0
2
10-2
100
(d2)
102
R
&
&
R
&
*+,,-,
& = 3. 5
L 1ℓ(balls,
R=2.50e+00
(d3)
Lℓ(balls,
R=1.00e+02
*+,,-,
& = (66
1
0.08
0.4
0.2
Fraction
(d1)
50
2#$%!
-2
r
(50)
In other words, s̃(~t) is a vertex of the polytope corresponding to the component of ~t with the largest magnitude, see Fig. 5(c). The components of ~t are i.i.d.
Gaussian random variables, and for large D, √
its maximum component, tmax , is concentrated around 2 log D.
Hence, Dg = h(ŝ · ~t)2 i = ht2max i = 2 log D which is
much smaller than D. This result is consistent with the
fact that the Gaussian
√ mean width of a D-dimensional
`1 ball scales with log D and not with D [22]. Since
all the points have norm R, we have Rg = √R, and
the effective margin is then given by κg = R 2 log D
, which is order unity in the scaling regime. In this
regime,
the√capacity
√ given by simple relation αM =
is
α0 κ + R 2 log D / 1 + R2 .
100
0.8
Fraction
R sign (ti ) , |ti | > |tj | ∀j 6= i
0
otherwise
(c)
(b)
2
Fraction
s̃i (~t) =
(a)
,
High-dimensional `1 balls, scaling regime: In the
scaling regime, we have ~v ≈ ~t. In this case, we can write
the solution for the subgradient as:
0.06
0.04
0.2
0.1
0.02
0
-1
0
1
2
3
Embedding
dimension
Support
dimension
(:)
0
0
10
20
30
40
Embedding
dimension
Support dimension
(:)
0
90
95
100
Embedding
dimension
Supporting
dimension
Support
dimension
(:)
Figure 9. Separability of `1 balls. (a) Linear classification
capacity of `1 balls as function of radius r with D = 100
and N = 200: (blue) MFT solution, (black dashed) spherical
approximation, (circle) full numerical simulations. (Inset) Illustration of `1 ball. (b) Manifold radius RM relative to the
actual radius r. (c) Manifold dimension DM as a function of r.
In small r limit, DM is approximately 2 log D , while in large
r, DM is close to D, showing how the solution is orthogonal to
the manifolds when their sizes are large. (d1-d3) Distribution
of support dimensions: (d1) r = 0.001, where most manifolds are either interior or touching, (d2) r = 2.5, where the
support dimension has a peaked distribution, (d3) r = 100,
where most manifolds are close to being fully supporting.
C.
Smooth nonconvex manifolds: Ring manifolds
Many neuroscience experiments measure the responses of
neuronal populations to a continuously varying stimulus,
with one or a small number of degrees of freedom. A prototypical example is the response of neurons in visual cortical areas to the orientation (or direction of movement)
of an object. When a population of neurons respond both
to an object identity as well as to a continuous physical
variation, the result is a set of smooth manifolds each
parameterized by a single variable, denoted θ, describing
continuous curves in RN . Since in general the neural responses are not linear in θ, the curve spans more than one
linear dimension. This smooth curve is not convex, and
is endowed with a complex non-smooth convex hull. It is
thus interesting to consider the consequences of our theory on the separability of smooth but non-convex curves.
The simplest example, considered here is the case where θ
corresponds to a periodic angular variable such as orientation of an image, and we call the resulting non-convex
curve a ring manifold. We model the neuronal responses
as smooth periodic functions of θ, which can be parameterized by decomposing the neuronal responsesP
in Fourier
D
modes. Here, for each object, M µ = xµ0 + i=1 si uµi
µ
where x0 represents the mean (over θ) of the population response to the object. The different D components
correspond to the different Fourier components of the an-
16
(a)
The parameters θµi are the preferred orientation angles
for the corresponding neurons and are assumed to be
evenly distributed between −π ≤ θiµ ≤ π (for simplicity,
we assume that the orientation tuning of the neurons are
all of the same shape and are symmetric around the preferred angle). The statistical assumptions of our analysis
assume that the different manifolds are randomly positioned and oriented with respect to the others. For the
ring manifold model, this implies that the mean responses
xµ0 are independent random Gaussian vectors and also
that the preferred orientation θµi angles are uncorrelated.
With this definition, all the vectors ~s ∈ S obey the norP D2
2
malization k~sk = r where r2 =
n=1 Rn . Thus, for
each object the ring manifold is a closed non-intersecting
smooth curve residing on the surface of a D-dimensional
sphere with radius r .
For the simplest case with D = 2, the ring manifold
is equivalent to a circle in two dimensions. However,
for larger D, the manifold is not convex and its convex
hull is composed of faces with varying dimensions. In
Fig. 10, we investigate the geometrical properties of these
manifolds relevant for classification as a function of the
overall scale factor r, where for simplicity we have chosen
Rn = r for all n. The most striking feature is the small
dimension in the scaling regime, scaling roughly as Dg ≈
2 log D. This logarithmic dependence is similar to that of
the `1 ball polytopes. Then, as r increases, DM increases
dramatically from 2 log D to D.
The similarity of the ring manifold to a convex polytope is also seen in the support dimension k of the manifolds. Support faces of dimension k 6= 0, 1, D + 1 are
seen, implying the presence of partially supporting solutions. Interestingly, D
2 < k ≤ D are excluded, indicating
that the maximal face dimension of the convex hull is D
2.
Each face is the convex hull of a set of k ≤ D
points
2
where each point resides in the 2D subspace spanned by
a pair of (real and imaginary) Fourier harmonics. The
ring manifolds are closely related to the trigonometric
moment curve, whose convex hull geometrical properties
have been extensively studied [27, 28].
In conclusion, the smoothness of the convex hulls becomes apparent in the distinct patterns of the support
1.5
15
DM
20
1
0
5
0
10-2 10-1 100 101 102
r
(c)
(d)
1
12
0.8
10-2 10-1 100 101 102
r
10
0.6
0.4
8
6
0.2
0
%
10
0.5
M
where Rn is the magnitude of the n-th Fourier component
for 1 ≤ n ≤ D
2 . The neural responses in Eq. (1) are
determined by projecting onto the basis:
√
uµi
2 cos(nθµi )
(52)
2n =
√
µi
µi
u2n−1 = 2 sin(nθ )
(b)
2
,
(51)
R /r
Rn
s2n (θ) = √ cos(nθ)
2
R
n
s2n−1 (θ) = √ sin(nθ)
2
D=20, Uniform Rn
DM
gular response, so that
6
10-2 10-1 100 101 102
8
10
2log(D)
2!"# $
r
(e1)
&R==(.
(*
0.01
0.4
0.2
0
0
10
20
k
Support dimension
(k)
(e2)
&=
= 20.00
+(
R
(e3)
0.3
0.2
+((
R& ==200.00
0.6
0.4
0.1
0.2
0
0
0
10
20
k
Support dimension
(k)
12
0
10
20
k
Support dimension
(k)
Figure 10. Linear classification
q of D-dimensional ring mani2
folds, with uniform Rn =
r. (a) Classification capacity
D
for D = 20 as function of r with m = 200 test samples (details of numerical simulations in SM, Sec. S9): (blue) mean
field theory, (black dashed) spherical approximation, (black
circles) numerical simulations. (Inset) Illustration of a ring
manifold. (b) Manifold dimension DM , which shows DM ∼ D
in the large r limit, showing the orthogonality of the solution.
(c) Manifold radius relative to the scaling factor r, RrM , as
a function of r. Fact that RrM becomes small implies that
the manifolds are “fully supporting” to the hyperplane, showing the small radius structure. (d) Manifold dimension grows
with affine dimension, D , as 2 log D for small r = 2√2 1log D
in the scaling regime.(e1-e3) Distribution of support dimensions. (d1) r = 0.01, where most manifolds are either interior
or touching, (d2) r = 20, where the support dimension has
(d3) r = 200, where
a peaked distribution, truncated at D
2
most support dimensions are D
or fully supporting at D + 1.
2
dimensions (compare Figs. 8(d) and 10(e)). However,
we see that all manifolds with Dg ∼ 5 or larger share
common trends. As the size of the manifold increases,
the capacity, and geometry vary smoothly, exhibiting a
smooth cross-over from high capacity with low radius and
dimension, to low capacity with large radius
√ and dimension. This crossover occurs as Rg ∝ 1/ Dg . Also, our
examples demonstrate, in many cases when the size is
smaller than the crossover value, the manifold dimensionality is substantially smaller than that expected from
naive second order statistics, highlighting the saliency
17
and significance of our anchor geometry.
VI.
MANIFOLDS WITH SPARSE LABELS
So far, we have assumed that the number of manifolds
with positive labels is approximately equal to the number of manifolds with negative labels. In this section, we
consider the case where the two classes are unbalanced
such that the number of positively-labeled manifolds is
far less than the negatively-labeled manifolds (the opposite scenario is equivalent). This is a special case of the
problem of classification of manifolds with heterogenous
statistics, where manifolds have different geometries or
label statistics. We begin by addressing the capacity of
mixtures of manifolds and then focus on sparsely labeled
manifolds.
A.
Mixtures of manifold geometries
Our theory of manifold classification is readily extended
to heterogeneous ensemble of manifolds, consisting of L
distinct classes. In the replica theory, the shape of the
manifolds appear only in the free energy term, G1 (see
Appendix, Eq. (A13)). For a mixture statistics, the combined free energy is given by simply averaging the individual free energy terms over each class l. Recall that this
free energy term determines the capacity for each shape,
giving an individual inverse critical load αl−1 . The inverse
capacity of the heterogeneous mixture is then,
α−1 = αl−1
l
(53)
where the average is over the fractional proportions of
the different manifold classes. This remarkably simple
but generic theoretical result enables analyzing diverse
manifold classification problems, consisting of mixtures
of manifolds with varying dimensions, shapes and sizes.
Eq. (53) is adequate for classes that differ in their geometry independent of the assigned labels. In general, classes
may differ in the label statistics (as in the sparse case
studied below) or in geometry that is correlated with labels. For instance, the positively labelled manifolds may
consist of one geometry and the negatively labelled manifolds may have a different geometry. How do structural
differences between the two classes affect the capacity of
the linear classification? A linear classifier can take advantage of these correlations by adding a non-zero bias.
Previously, it was assumed that the optimal separating
hyperplane passes through the origin; this is reasonable
when the two classes are statistically the same. However, when there are statistical differences between the
label assignments of the two classes, Eq. (2) should be
replaced by y µ (w · xµ − b) ≥ κ where the bias b is chosen
to maximize the mixture capacity, 53. The effect of optimizing the bias is discussed in detail in the next section
for the sparse labels and in other scenarios in SM (Sec.
S3).
B.
Manifolds with sparse labels
We define the sparsity parameter f as the fraction of
positively-labeled manifolds so that f = 0.5 corresponds
to having balanced labels. From the theory of the classification of a finite set of random points, it is known
that having sparse labels with f 0.5 can drastically
increase the capacity [12]. In this section, we investigate
how the sparsity of manifold labels improves the manifold
classification capacity.
If the separating hyperplane is constrained to go through
origin and the distribution of inputs is symmetric around
the origin, the labeling y µ is immaterial to the capacity.
Thus, the effect of sparse labels is closely tied to having a
non-zero bias. We thus consider inequality constraints of
the form y µ (w·xµ −b) ≥ κ, and define the bias-dependent
capacity of general manifolds with label sparsity f , margin κ and bias b, as αM (κ, f, b). Next, we observe that
the bias acts as a positive contribution to the margin
for the positively-labeled population and as a negative
contribution to the negatively-labeled population. Thus,
the dependence of αM (κ, f, b) on both f and b can be
expressed as,
−1
−1
−1
(κ − b) (54)
(κ + b) + (1 − f )αM
(κ, f, b) ≡ f αM
αM
where αM (x) is the classification capacity with zero bias
(and hence equivalent to the capacity with f = 0.5) for
the same manifolds. Note that Eq. (54) is similar to Eq.
(53) for mixtures of manifolds. The actual capacity with
sparse labels is given by optimizing the above expression
with respect to b, i.e.,
αM (κ, f ) = maxb αM (κ, f, b)
(55)
In the following, we consider for simplicity the effect of
sparsity for zero margin, κ = 0.
Importantly, if D is not large, the effect of the manifold geometry in sparsely labeled manifolds can be much
larger than that for non-sparse labels. For non-sparse labels, the capacity ranges between 2 and (D+0.5)−1 , while
for sparse manifolds the upper bound can be much larger.
Indeed, small-sized manifolds are expected to have capacity that increases upon decreasing f as αM (0, f ) ∝
1
f |log f | , similar to P uncorrelated points [12]. This potential increase in capacity of sparse labels is however
strongly constrained by the manifold size, since when the
manifolds are large, the solution has to be orthogonal to
1
the manifold directions so that αM (0, f ) ≈ D
. Thus,
18
the geometry of the manifolds plays an important role in
controlling the effect of sparse labels on capacity. These
aspects are already seen in the case of sparsely labeled `2
balls (Appendix E). Here we summarize the main results
for general manifolds.
Sparsity and size: There is a complex interplay between label sparsity and manifold size. Our analysis
yields three qualitatively different regimes:
Low Rg : When the Gaussian radius of the manifolds are
small, i.e., Rg < 1, the extent of the manifolds is noticeable only when the dimension is high. Similar to our
previous analysis of high dimensional manifolds, we find
here that the sparse capacity is equivalent to the capacity of sparsely labeled
p random points with an effective
margin given by Rg Dg ,
p
(56)
αM (f ) ≈ α0 (f, κ = Rg Dg )
where α0 (κ, f ) = maxb α0 (κ, f, b) from the Gardner theory. It should be noted that κg in the above equation has
a noticeable effect only for moderate sparsity. It has a
negligible effect when f → 0, since the bias is large and
dominates over the margin.
Moderate sizes, Rg > 2: In this case, the equivalence to
the capacity of points breaks down. Remarkably we find
that the capacity of general manifolds with substantial
size is well approximated by that of equivalent `2 balls
with the same sparsity f and with dimension and radius
equal to the Gaussian dimension and radius of the manifolds, namely
αM (f ) ≈ αBall (f, Rg , Dg )
(57)
Surprisingly, unlike the nonsparse approximation, where
the equivalence of general manifold to balls, Eq 34, is
valid only for high dimensional manifolds, in the sparse
limit, the spherical approximation is not restricted to
large D. Another interesting result is that the relevant
statistics are given by the Gaussian geometry, Rg and Dg ,
even when Rg is not small. The reason is that for small
f , the bias is large. In that case, the positively labeled
manifolds have large positive margin b and are fully supporting giving a contribution to the inverse capacity of
b2 regardless of their detailed geometry. On the other
hand, the negatively labeled manifolds have large negative margin, implying that most of them are far from
the separating plane (interior) and a small fraction have
touching support. The fully supporting configurations
have negligible probability, hence the overall geometry
is well approximated by the Gaussian quantities Dg and
Rg .
Scaling relationship between sparsity and size: A
further analysis of the capacity of balls with sparse labels
shows it retains a simple form α(f¯, D) (see Appendix E
Eq. E10) which depends on R and f only through the
scaled sparsity f¯ = f R2 . The reason for the scaling of
f with R2 is as follows. When the labels are sparse, the
dominant contribution to the inverse capacity comes from
−1
the minority class, and so the capacity is αBall
≈ f b2
for large b. On the other hand, the optimal value of b
depends on the balance between the contributions from
both classes and scales linearly with R, as it needs to
overcome the local fields from the spheres. Thus, α−1 ∝
f R2 .
Combining Eq. E10 with Eq. 57 yields for general
sparsely labeled manifolds,
−1
αM
= f¯b̄2 +
Z
b̄
∞
dtχDg (t)(t − b̄)2
(58)
where scaled sparsity is f¯ = f (1 + Rg2 ) . 1 .
Note we have defined the scaled sparsity f¯ = f (1 + Rg2 )
rather than f R2 to yield a smoother cross-over to the
small Rg regime.
q Similarly, we define the optimal scaled
bias as b̄ = b 1 + Rg2 1. Qualitatively, the function
α is roughly proportional to f¯−1 with a proportionality
constant that depends on Dg . In the extreme limit of
−1
log f¯ Dg , we obtain the sparse limit α ∝ f¯ log f¯ .
Note that if f is sufficiently small, the gain in capacity
due to sparsity occurs even for large manifolds as long as
f¯ < 1.
Large Rg regime: f¯ > 1: Finally, when Rg is sufficiently
large such that f¯ increases above 1, b̄ is of order 1 or
smaller, and the capacity is small with a value that depends on the detailed geometry of the manifold. In particular, when f¯ 1, the capacity of the manifold ap1
proaches αM (f ) → D
, not D1g .
To demonstrate these remarkable predictions, Eq. (57),
we present in Fig. 11 the capacity of three classes of
sparsely labeled manifolds:`2 balls, `1 ellipsoids and ring
manifolds. In all cases, we show the results of numerical
simulations of the capacity, the full mean field solution,
and the spherical approximation, 57, across several orders
of magnitude of sparsity and size as a function of the
scaled sparsity f¯ = f (1 = r2 ). In each example, there
is a good agreement between the three calculations for
the range of f¯ < 1. Furthermore, the drop in α with
increasing f¯ is similar in all cases, except for an overall
vertical shift which is due to the different Dg , similar to
the effect of dimension in `2 balls (a). In the regime of
moderate radii, results for different f and r all fall on a
universal curve which depends only on f¯, as predicted by
the theory. For small r, the capacity deviates from this
scaling as it is dominated by f alone, similar to sparsely
labeled points. When f¯ > 1, the curves deviate from the
spherical approximation. The true capacity (as revealed
by simulations and full mean field) rapidly decreases with
1
f¯ and saturates at D
, rather than to the D1g limit of the
19
( = 10)+
101
'
( = 10),
0
10
'
( = 10)-
D=10
D=100
10-2
-4
10
10
-2
0
10
!̅
2
10
( = 10),
102
( = 10)-
0
10
-1
/1)-
10-2
/ )-
10
10-2
Ring Manifolds
103 ( = 10)*
( = 10)+
100
!̅(
̅
102
104
1
( = 10)+
( = 10),
10
f=$%&$
(a)
f=1e-04
f=1e-03
f=1e-02
f=1e-01
f=$%&'
(b)
0.8
0.1
( = 10)-
Majority
Minority
0.6
0
10
0.05
-1
10
10-3 10-2 10-1 100 101
!(̅
Figure 11. (a-b) Classification of `2 balls with sparse labels:
(a) Capacity of `2 balls as a function of f¯ = f (1 + r2 ), for
D = 100 (red) and D = 10 (blue): (solid lines) mean field theory, (dotted lines) is an approximation interpolating between
Eqs. 56 and 58 (details in SM, Sec. S9). (b–c) Classification
of general manifolds with sparse labels: (b) Capacity of `1
ellipsoids with D = 100, where the first 10 components are
equal to r, and the remaining 90 components are 12 r, as a
function of f¯ = (1 + Rg2 ). r is varied from 10−3 to 103 : (circles) numerical simulations, (lines) mean field theory, (dotted
lines) spherical approximation. (c) Capacity of ring manifolds
with a Gaussian fall-off spectrum with σ = 0.1 and D = 100
(; details in SM). Fig. 11(c) shows the capacity of ring manifolds whose Fourier components
have a Gaussian fall-off, i.e.
Rn = A exp − 12 (2π(n − 1)σ)2 .
spherical approximation. Finally, we note that our choice
of parameters in (c) (SM Section 7) were such that Rg
(entering in the scaled sparsity) was significantly different
from simply an average radius. Thus, the agreement with
the theory illustrates the important role of the Gaussian
geometry in the sparse case.
As discussed above, in the (high and moderate) sparse
regimes, a large bias alters the geometry of the two classes
in different ways. To illustrate this important aspect, we
show in Fig. 12, the effect of sparsity and bias on the
geometry of the `1 ellipsoids studied in Fig. 11(c). Here
we show the evolution of RM and DM for the majority
and minority classes as f¯ increases. Note that despite the
fact that the shape of the manifolds is the same for all
f, their manifold anchor geometry depends on both class
membership and sparsity levels because these measures
depend on the margin. When f¯ is small, the minority
class has DM = D as seen in Fig. 12(c), i.e., the minority
class manifolds are close to be fully supporting due to
the large positive margin. This can also be seen in the
distributions of support dimension shown in Fig.12(a)(b). On the other hand, the majority class has DM ≈
Dg = 2logD1 D, and these manifolds are mostly in the
interior regime. As f increases, the geometrical statistics
for the two classes become more similar. This is seen in
Fig.12(c)-(d) where DM and RM for both majority and
minority classes converge
to the zero margin value for
large f¯ = f 1 + Rg2 .
0
0.4
Majority
20
Minority
40
0
60
Support dimension (")
(c)
0.2
0
5
95
100
Support dimension (")
(d)
1
100
50
Minor
Major
0
10-5 10-4 10-3 10-2 10-1 100
2
f(1+R
f̅ g)
R
RM
/r0
)/R
2
10
(c)
ℓ1 Ellipsoids
2
10
DM
(b)
ℓ2 balls
( = 10)*
'0
(a)
0.5
Minor
Major
0
-5
-4
-3
-2
-1
0
10 10 10 10 10 10
2
f(1+R
f̅ )
g
Figure 12. Manifold configurations and geometries for classification of `1 ellipsoids with sparse labels, analyzed separately
in terms of the majority and minority classes. The radii for
`1 ellipsoids are Ri = r for 1 ≤ i ≤ 10 and Ri = 12 r for
11 ≤ i ≤ 100 with r = 10. (a) Histogram of support dimensions for moderate sparsity f = 0.1: (blue) minority and
(red) majority manifolds. (b) Histogram of support dimensions for high sparsity f = 10−3 : (blue) minority and (red)
majority manifolds. (c) Manifold dimension as a function of
f¯ = f (1 + Rg2 ) as f is varied: (blue) minority and (red) majority manifolds. (d) Manifold radius relative to the scaling
factor r as a function of f¯: (blue) minority and (red) majority
manifolds.
VII.
SUMMARY AND DISCUSSION
Summary: We have developed a statistical mechanical
theory for linear classification of inputs organized in perceptual manifolds, where all points in a manifold share
the same label. The notion of perceptual manifolds is
critical in a variety of contexts in computational neuroscience modeling and in signal processing. Our theory is
not restricted to manifolds with smooth or regular geometries; it applies to any compact subset of a D-dimensional
affine subspace. Thus, the theory is applicable to manifolds arising from any variation in neuronal responses
with a continuously varying physical variable, or from
any sampled set arising from experimental measurements
on a limited number of stimuli.
The theory describes the capacity of a linear classifier
to separate a dichotomy of general manifolds (with a
given margin) by a universal set of mean field equations.
These equations may be solved analytically for simple geometries, but for more complex geometries, we have developed iterative algorithms to solve the self-consistent
equations. The algorithms are efficient and converge fast
as they involve only solving for O(D) variables of a single manifold, rather than invoking simulations of a full
20
system of P manifolds embedded in RN .
Our theory provides for the first time, quantitative and
qualitative predictions for the binary classification of realistic data structures. However, application to real data
may require further extensions of the theory as presented
here. The most important ones (currently the subject of
on going work) include:
Correlations: In the present work we have assumed
that the directions of the affine subspaces of the different manifolds are uncorrelated. In realistic situations
we expect to see correlations in the manifold geometries,
mainly of two types: One is center-center correlations.
Such correlations can be harmful for linear separability [29, 30]. Another is correlated variability in which
the directions of the affine subspaces are correlated but
not the centers. Positive correlations of the latter form
are beneficial for separability. In the extreme case when
the manifolds share a common affine subspace, the rank
of the union of the subspaces is Dtot = D rather than
Dtot = P D, and the solution weight vector need only lie
in the null space of this smaller subspace. Further work is
needed to extend the present theory to incorporate more
general correlations.
Generalization performance: We have studied the
separability of manifolds with known geometries. In
many realistic problems, this information is not readily
available and only samples reflecting the natural variability of input patterns are provided. These samples can be
used to estimate the underlying manifold model (using
manifold learning techniques [31, 32]) and/or to train a
classifier based upon a finite training set. Generalization
error describes how well a classifier trained on a finite
number of samples would perform on other test points
drawn from the manifolds [33]. It would be important
to extend our theory to calculate the expected generalization error achieved by the maximum margin solution
trained on point cloud manifolds, as a function of the size
of the training set and the geometry of the underlying full
manifolds.
Unrealizable classification: Throughout the present
work, we have assumed that the manifolds are separable by a linear classifier. In realistic problems, the
load may be above the capacity for linear separation,
i.e. α > αM (κ = 0). Alternatively, neural noise may
cause the manifolds to be unbounded in extent, with the
tails of their distribution overlapping so that they are
not separable with zero error. There are several ways to
handle this issue in supervised learning problems. One
possibility is to nonlinearly map the unrealizable inputs
to a higher dimensional feature space, via a multi-layer
network or nonlinear kernel function, where the classification can be performed with zero error. The design of
multilayer networks could be facilitated using manifold
processing principles uncovered by our theory.
Another possibility is to introduce an optimization prob-
lem allowing a small training error, for example, using
an SVM with complementary slack variables [14]. These
procedures raise interesting theoretical challenges, including understanding how the geometry of manifolds
change as they undergo nonlinear transformations, as
well as investigating by statistical mechanics, the performance of a linear classifier of manifolds with slack variables [34].
Concluding remarks: The statistical mechanical theory of perceptron learning has long provided a basis for
understanding the performance and fundamental limitations of single layer neural architectures and their kernel
extensions. However, the previous theory only considered a finite number of random points with no underlying geometric structure, and could not explain the performance of linear classifiers on large, possibly infinite
number of inputs organized as distinct manifolds by the
variability due to changes in physical parameters of objects. The statistical mechanical theory presented in this
work can explain the capacity and limitations of linear
classification of general manifolds, and be used to elucidate changes in neural representations across hierarchical sensory systems. We believe the application of this
theory and its corollary extensions will precipitate novel
insights into how perceptual systems, biological or artificial, can efficiently code and process sensory information.
We would like to thank Uri Cohen, Ryan Adams, Leslie
Valiant, David Cox, Jim DiCarlo, Doris Tsao, and Yoram
Burak for helpful discussions. The work is partially supported by the Gatsby Charitable Foundation, the Swartz
Foundation, the Simons Foundation (SCGB Grant No.
325207), the NIH, and the Human Frontier Science Program (Project RGP0015/2013). D. D. Lee also acknowledges the support of the US National Science Foundation,
Army Research Laboratory, Office of Naval Research, Air
Force Office of Scientific Research, and Department of
Transportation.
Appendix A: Replica Theory of Manifold Capacity
In this section, we outline the derivation of the mean
field replica theory summarized in Eqs. (11)-12. We
define the capacity of linear classification of manifolds,
P
, for which with
αM (κ), as the maximal load, α = N
µ
high probability a solution to y w · xµ ≥ κ exists for
a given κ. Here xµ are points on the P manifolds M µ ,
Eq. (1), and we assume that all N P (D + 1) components
of {uµi } are drawn independently from a Gaussian distribution with zero mean and variance N1 , and that the
binary labels y µ = ±1 are randomly assigned to each
manifold with equal probabilities. We consider the therP
modynamic limit where N, P → ∞ but α = N
, and D
are finite.
Note that the geometric margin, κ0 , defined as the dis-
21
µ
µ
tance from the
√ solution hyperplane is given by y w·x ≥
0
0
κ kwk = κ N . However, this distance depends on the
scale of the input vectors xµ . The correct scaling of the
kxk
margin in the thermodynamic limit is κ0 = √
κ. Since
N
µ
we adopted the normalization of kx k = O(1), the correct scaling of the margin is y µ w · xµ ≥ κ .
Evaluation of solution volume: Following Gardner’s
replica framework, we first consider the volume Z of the
solution space for α < αM (κ) . We define the signed projections of the the
vector uµi on the solution
√ ith µdirection
µ
µ
weight as Hi = N y w · ui , where i = 1, ..., D + 1 and
µ = 1, ..., PP
. Then, the separability constraints can be
D+1
written as i=1 Si Hiµ ≥ κ . Hence the volume can be
written as
Z
Z=
~µ
dN wδ(w2 − N ) ΠP
µ=1 Θµ gS (H ) − κ
variance
1
N)
yields,
+
*
D+1
N
XX
X
µα
µ
iĤi (−y µ
exp
wαj ui,j )
i=1 µα
j=1
(A5)
µ
uµ
i ,y
1X
X µα µβ
qαβ
Ĥi Ĥi
= exp −
2
iµ
αβ
PN
where, qαβ = N1 j=1 wαj wβj . Thus, integrating the variables Ĥiµα yields
hZ n i =
Z Y
n
α=1
2
dwα δ(wα
− N)
T
· δ N qαβ − wα
wβ
(A1)
Z
dqαβ Παβ
(A6)
P
(D + 1)
logdetq X
exp −
2
where
where Θ(x) is a Heaviside step function. gS is the
~
support
n function of oS defined for Eq. 12 as gS (V ) =
~ ·S
~|S
~∈S .
min ~ V
X=
S
The volume defined above depends on the the quenched
random variables uµi and y µ through Hiµ . It is well known
that in order to obtain the typical behavior in the thermodynamic limit, we need to average log Z, which we carry
n
out using the replica trick, hlog Zi = limn→0 hZ ni−1 ,
where hi refers to the average over uµi and y µ . For natural
n, we need to evaluate,
Z Y
n
hZ n i =
α
D+1
Y
h
√
2
dwα δ(wα
− N)
2πδ(Hiµα
i
−y
µ
P Z
Y
~ µα
DH
(A2)
Using Fourier representation of the delta functions, we
obtain
hZ i =
Z Y
n
D+1
YZ
i=1
α
2
dwα δ(wα
− N)
and we have used the fact that all manifolds contribute
the same factor.
We proceed by making the replica symmetric ansatz on
the order parameter qαβ at its saddle point, qαβ = (1 −
q)δαβ + q, from which one obtains in the n → 0 limit:
q−1
αβ =
P Z
Y
~ µα
DH
1
q
δαβ −
1−q
(1 − q)2
(A4)
µ
n
oE
dĤiµα D
T µ
√
exp iĤiµα (Hiµα − y µ wα
ui )
µ
uµ
2π
i ,y
Performing the average over the Gaussian distribution
of uµi (each of the N components has zero mean and
(A8)
and
wαT uµi )iuµi ,yµ
(A3)
(A7)
i,α,β
logdetq = n log(1 − q) +
where we have used the notation,
n
α
µ
dHi
~ = ΠD+1 √
~ −κ
DH
Θ gS (H)
i=1
2π
X
1
~ α exp −
Hiα (q−1 )αβ Hiβ
DH
2
Z Y
nq
1−q
(A9)
Thus the exponential term in X can be written as
!2
√ X
α 2
X
X
q
1
(Hi )
1
exp −
+
H α (A10)
2 αi 1 − q
2 i
1−q α i
Using the Hubbard–Stratonovich transformation, we obtain
"Z
(
)#n
Z
√
~2
q
H
1
~ · T~
~ exp −
+
H
X = DT~
DH
21−q 1−q
(A11)
2
T
dT
i
i
where DT~ = Πi √2π exp − 2 .
Completing the
R
square in the exponential and using
DT~ An =
R
~
exp n
the n → 0 limit, we obtain, X =
DT log A in
R
nq(D+1)
exp 2(1−q) + n DT~ log z(T~ ) with
z(T~ ) =
Z
~
DH exp −
1
~ − √q T~ ||2
||H
2(1 − q)
(A12)
22
Combining these terms, we write the last factor in Eq.
(A6) as exp nP G1 where,
Z
G1 =
(D + 1)
log(1 − q)
DT~ log z(T~ ) −
2
(A13)
q
1
ln(1 − q) +
2
2(1 − q)
(A14)
and represents the constraints on the volume of wα due
to normalization and the order parameter q. Combining
the G0 and G1 contributions, we have
hZ n it0 ,t = eN n[G0 (q)+αG1 (q)]
(A15)
The classification constraints contribute αG1 , with Eq.
(A13), and
!
Z
~2
dYi
Y
D+1
~
(A16)
z(T ) = Πi=1 p
exp −
2(1 − q)
2π(1 − q)
√
~)−κ
Θ gS ( q T~ + Y
where, we have written the fields Hi as
√
Hi = qTi + Yi
(A17)
√
Note that qTi represents the quenched random component due to the randomness in the uµi , and Yi is the
“thermal” component due to the variability within the
solution space. The order parameter q is calculated via
∂G1
0
0 = ∂G
∂q + α ∂q .
Capacity: In the limit where α → αM (κ) , the overlap between the solutions become unity and the volume
q
shrinks to zero. It is convenient to define Q = 1−q
and
study the limit of Q → ∞. In this limit the leading order
is
i
Qh
1 − αhF (T~ )iT~
(A18)
hlog Zi =
2
where the first term is the contribution from G0 → Q
2.
Q
~
The second term comes from G1 → − 2 αhF (T )iT~ , where
the average is over the Gaussian distribution of the D + 1
dimensional vector T~ , and
F (T~ ) → −
2
log z(T~ )
Q
(A19)
is independent of Q and is given by replacing the integrals
in Eq. (A16) by their saddle point, which yields
2
~
~
~
~
F (T ) = min V − T | gS V − κ ≥ 0
(A20)
~
V
−1
αM
(κ) = hF (T~ )iT~
~ − T~
F (T~ ) = min V
~
V
The first factors in hZ n i, Eq. (A6), can be written as
exp nN G0 , where as in the Gardner theory, the entropic
term in the thermodynamic limit is
G0 (q) =
At the capacity, log Z vanishes, the capacity of a general
manifold with margin κ, is given by,
(A21)
2
~)−κ≥0
| gS (V
(A22)
Finally, we note that the mean squared ’annealed’ variability in the fields due to the entropy of solutions vanishes at the capacity limit, as 1/Q , see Eq. (A16) . Thus,
2
~ − T~ in the above equation represents
the quantity V
the annealed variability times Q which remains finite in
the limit of Q → ∞.
Appendix B: Strictly Convex Manifolds
B.1. General
Here we evaluate the capacity of strictly convex manifolds, starting from the expression for general manifolds,
Eq. 21. In a strictly convex manifold, S , any point in
the line segment connecting two points ~x and ~y , ~x, ~y ∈ S
, other than ~x and ~y , belongs to the interior of S . Thus,
the boundary of the manifold does not contain edges or
flats with spanning dimension k > 1 except of the entire manifold, k = D + 1 . Therefore, there are exactly
contributions to the inverse capacity. When t0 obeys
ttouch (~t) > t0 − κ > tfs (~t), the integrand of Eq. 21 con2
~ ~
0 +κ)
tributes (−t·s̃(T )−t
, When t0 < κ + tfs (~t), the man2
~
1+ks̃(T )k
ifold is fully embedded. In this case, ~v = 0, and the
integrand reduces to Eq. 27. In summary, the capacity
for convex manifolds can be written as
α−1 (κ) =
κ+ttouch (~
t)
(−~t · s̃(~t, t0 ) − t0 + κ)2
2
κ+tfs (~
t)
1 + s̃(~t, t0 )
(B1)
Z
Z κ+tfs (~t)
h
i
2
2
+ D~t
Dt0 (t0 − κ) + ~t
Z
D~t
Z
Dt0
−∞
where ttouch (~t) and tfs (~t) are given by Eqs. 23 and 26,
respectively, and s̃(~t, t0 ) = arg min~s (~v · ~s), with ~v = ~t +
(v0 − t0 )s̃.
B.2. `2 Balls
In the case of `2 balls with D and radius R, g(~v ) =
−R k~v k . Hence, ttouch (~t) = R ~t and tfs (~t) = −R−1 ~t
23
. Thus, Eq. B1 reduces to hat the capacity of balls is
−1
αBall
=
∞
Z
Z
Dt0
κ−tR−1
0
∞
Z
+
κ+tR
dtχD (t)
Z
~
ttouch (~t) = ~t ◦ R
κ−tR−1
dtχD (t)
−∞
0
(−t0 + tR + κ)2
(1 + R2 )
obeying the inequality t0 − κ ≥ ttouch (~t) with Eq. 23,
yielding
Dt0 (t0 − κ)2 + t2 (B2)
where,
(B6)
Touching regime: Here the anchor point is given by Eq.
B4 where ~v = ~t + λs̃. Substituting ti + λs˜i for vi in the
numerator of that equation, and g(~v ) = κ−v0 = κ−t0 −λ
, yields,
D
21− 2 D−1 − 1 t2
χD (t) =
t
e 2 , t≥0
Γ( D
2)
(B3)
is the D-dimensional Chi probability density function,
reproducing the results of [15]. Furthermore, Eq. B2 can
be approximated
√ by the capacity of points with a2 margin
D, i.e. αBall (κ, RM , DM ) ≈ (1+R )α0 (κ+
increase
of
R
√
R D) (details in [15]).
s̃touch,i (~t, t0 ) = −
ti Ri2
,
λ(1 + Ri2 ) + (−t0 + κ)
(B7)
where the parameter λ is determined by the ellipsoidal
constraint,
1=
D
X
t2i Ri2
i=1
[λ(1 + Ri2 ) + (−t0 + κ)]
2.
(B8)
B.2. `2 Ellipsoids
a.
Anchor points and support regimes
With ellipsoids, Eq. 36, the support function in Eq. 15
~ =
can be computed explicitly as follows. For a vector V
(~v , v0 ), with non zero ~v , the support function g(~v ) is
minimized by a vector ~s which occurs on the boundary of
the ellipsoid, i.e. it obeys the equality constraint f (~s) = 0
PD 2
with ρ(~s) = i=1 Rsii − 1 , Eq. (36). To evaluate g,
PD
we differentiate − i=1 si~vi + ρf (~s) with respect to ~si ,
where ρ is a Lagrange multiplier enforcing the constraint,
yielding
s̃i = −
vi Ri2
g(~v )
(B4)
and
In this regime, the contribution to the capacity is given
by Eq. 16-17 with s̃ in Eq. (B7).
The touching regime holds for ttouch > t0 − κ > tfs ,
where λ → κ − t0 , ~v vanishes, and the anchor point is
the point on the boundary of the ellipsoid antiparallel to
~t . Substituting this value of λ in Eq. B8 , yields,
v
u
uX ti 2
(B9)
tfs (~t) = −t
Ri
i
Fully supporting regime: When t0 − κ < tfs , we have
~v = 0, v0 = κ, and λ = t0 − κ, implying that the center
as well as the entire ellipsoid is fully supporting the max
margin solution. In this case, the anchor point is antiparallel to ~t at the interior point, 25, and its contribution to
the capacity is as in Eq. 27.
Appendix C: Limit of Large Manifolds
~
g(~v ) = − ~v ◦ R
(B5)
~ is the D dimensional vector of the ellipsoid prinwhere R
cipal radii, and we denote ◦refers to pointwise product of
~ i = vi Ri . For a given (~t, t0 ), the vector
vectors, (~v ◦ R)
(~v , v0 ) is determined by Eq. (13) and the analytic solution above can be used to derive explicit expressions for
s̃(~t, t0 ) in the different regimes as follows.
Interior regime:
In the interior regime λ = 0 , ~v = ~t, resulting in zero
contribution to the inverse capacity. The anchor point
is given by the following boundary point on the ellipse,
given by Eq. B4 with ~v = ~t. This regime holds for t0
In the large size limit, ks̃k → ∞ , linear separation
of manifolds reduces to linear separation of P random
D-dimensional affine subspaces. When separating subspaces, all of them must be fully embedded in the margin plane, otherwise they would intersect it and violate
the classification constraints. However, the way large size
manifolds approach this limit is subtle. To analyze this
limit, we note that when ks̃k is large, g(~v ) < ks̃k k~v k and
from the condition that g(~v ) ≥ (−v0 + κ), we have k~v k ≤
−1
(−v0 +κ)
, i.e., k~v k is O(ks̃k ). A small k~v k implies that
ks̃k
the affine basis vectors, except the center direction, are all
either exactly or almost orthogonal to the solution weight
vector. Since λs̃ = −~t+~v ≈ −~t, it follows that s̃ is almost
antiparallel to the gaussian vector ~t, hence DM → D, see
24
Eq. 30. To elucidate the manifold support structure,
we note first that by Eq. 23 ttouch ∝ −||~t||||s̃|| → −∞
hence the fractional volume of the interior regime is negligible, and the statistics is dominated by the embedded regimes. In fact, the fully embedded transition is
given by tembed ≈ −κ , see Eq.26, so that the fractional
R ∞ volume of the fully embedded regime is H(−κ) =
Dt0 and its contribution to inverse capacity is there−κ
i
h
R∞
2
fore −κ Dt0 h ~t i + (t0 + κ)2 = H(−κ)D + α0−1 (κ) .
The remaining summed probability of the touching and
partially embedded regimes (k ≥ 1) is therefore H(κ).
2
2
2
In these regimes, λ2 ks̃k ≈ λ2 ks̃k ≈ ~t , so that this
R −κ
2
regime contributes a factor of −∞ Dt0 h ~t i = H(κ)D.
Combining these two contributions, we obtain for large
−1
sizes, αM
= D + α0−1 (κ), consistent with Eq. (8).
Appendix D: High Dimensional Manifolds
D.1. High Dimensional `2 ball
Before we discuss general manifolds in high dimension,
we focus on the simple case of high dimensional balls,
the capacity of which is given by Eq. √
B2. When D 1
D. Substituting
, χD (t) , B3, is
centered
around
t
=
√
χD (t) ≈ δ(t − D) yields
As stated above, Eq, D3 implies
√ a finite capacity only in
the scaling
regime,
where
R
D = O(1). If, on the other
√
hand, R D 1, Eq. (D2) implies
−1
αBall
=
R2 D
1 + R2
(D4)
(where we have used the asymptote α0−1 (x) → x2 for
large x), reducing to α−1 = D for large R.
The analysis above also highlights the support
structure
√
of the balls at large D. As long as R D the
fraction of balls that lie fully in the margin plane,
√ is negligible, as implied by the fact that tsf ≈ − D/R
√→
−∞.The overall fraction of interior balls is H(κ + R D)
whereas the √
fraction that touch the margin planes is
1 − H(κ + R D) . Despite the fact that there are no
fully supporting balls, the touching balls are almost parallel to the margin planes if R 1 hence the capacity
reaches its lower bound.
Finally, the large manifold √
limit discussed in Appendix
C is realized for as R D . Here, tsf = −κ and
the system is either touching or fully supporting with
probabilities H(κ) and H(−κ), respectively.
D.2. General manifolds.
To analyze the limit of large dimensional general manifolds we utilize the self averaging of terms in Eq. (21),
(R D + κ −
−1
involving
sums of the D components of ~t and s̃, i.e.,
(κ, R, D) =
αBall
Dt0
+
Dt
([t0 −κ]2 +D)
0
√
R2 + 1
−∞
κ− RD
p
~t · s̃ ≈ h||s̃||ih~t · ŝi ≈ RM DM = κM
(D1)
√
As long as R D , the second term in Eq. (D1)
(D5)
vanishes and yields
Z
−1
αBall
(κ, R, D)
√
√
κ+R D
Z
√
κ+R D
=
−∞
√
t)20
Z
κ−
D
R
√
(R D + κ − t0 )2
Dt0
(D2)
R2 + 1
Here we note that the t0 term in the numerator is significant only if R = O(D−1/2 ) or smaller, in which case the
denominator is just 1 . When R is of order 1 the term
t0 in the integrand is negligible. Hence, in both cases we
can write,
−1
αBall
(κ, R, D) ≈ α0−1
√ !
κ+R D
√
1 + R2
Also, ttouch ≈ ~t · s̃ ≈ κM , hence, we obtain for the capacity,
−1
αM
(κ) ≈
−1
≈ αBall
(κ, RM , DM )
(D6)
where the average is wrt the gaussian t0 . Evaluating RM
and DM involves calculations of the self consistent statistics of the anchor points. This calculation is simplified
in high dimension. In particular, λ, Eq. 17 , reduces to
(D3)
√
In this form, the intuition beyond the factor 1 + R2
is clear, stemming from the fact that the distance of a
point from the margin plane scales with its norm, hence
the margin entering the capacity should factor out this
norm.
h[κM + κ − t0 ]2+ it0
2 +1
RM
λ≈
h[κM + κ − t0 ]+ it0
2
1 + RM
(D7)
hence it is approximately constant, independent of T~ .
In deriving the above approximations we used selfaveraging in summations involving the D intrinsic coordinates. The full dependence on the longitudinal gaussian
25
t0 should remain. Thus, RM and DM should in fact be
substituting by RM (t0 ) and DM (t0 ) denoting 29-30 with
averaging only over ~t. This would yield a more complicated expressions than Eq.D6- D7.
The reason why we can replace them by average quantities is the following: the potential dependence of the
anchor radius and dimension on t0 is via λ(t0 ). However,
inspecting Eq. D7 we note two scenarios: One in which
RM is small and κM of order 1 . In this case, because
the manifold radius is small the contribution λs̃ is small
and can be neglected. This is the same argument why in
this case, the geometry can be replaced by the Gaussian
geometry which does not depend on t0 or κ . The second
scenario is that RM is of order 1 and κM 1, in which
case the order 1 contribution from t0 is negligible.
with sparsity f (Eq. (54) and Eq. B2) is given by,
Z ∞
−1
αBall
=f
dtχD (t)x
0
"Z
κ+tR+b
(−t0 + tR + b + κ)2
Dt0
(1 + R2 )
κ−tR−1 +b
#
Z b+κ−tR−1
2
2
+
Dt0 ((t0 − b − κ) + t )
−∞
∞
Z
+ (1 − f )
"Z
κ+tR−b
Dt0
κ−tR−1 −b
Z
+
−b+κ−tR−1
−∞
Appendix E: Capacity of `2 Balls with Sparse Labels
We begin by writing the capacity of random points with
sparse labels [12], as α0 (f, κ) = maxb α0 (f, κ, b), where
Z
+ (1 − f )
Dt(−t + κ + b)2
Z
−∞
Dt(−t + κ − b)2
b+κ
0=f
−∞
#
2
2
Dt0 ((t0 − κ + b) + t )
(E4)
√
(−t0 + R D + b)2
(1 + R2 )
−∞
√
√
Z R D−b
(−t0 + R D + b)2
+ (1 − f )
Dt0
(1 + R2 )
−∞
Z
√
R D+b
=f
Dt0
κ−b
(E1)
(E5)
Optimizing b yields the following equation for b,
Z
−1
αBall
κ+b
−∞
(−t0 + tR + κ − b)2
(1 + R2 )
and the optimal bias b is given by ∂αB /∂b = 0. Here we
analyze these equations in various size regimes assuming
κ = 0.
Small R: For balls with small radius, the capacity is
that of points, unless
√ the dimensionality is high. Thus,
if D is large but R D . 1,
First, we note that the capacity of sparsely labeled points
is α0 (f, κ) = maxb α0 (f, κ, b), where
α0−1 (f, κ, b) = f
dtχD (t)x
0
Dt(−t + κ + b) + (1 − f )
Z
that is,
√
αBall (f, R, D) ≈ α0 (f, κ = R D)
b−κ
−∞
Dt(−t + κ − b)
(E2)
In the limit of f → 0, b 1 . The first equation reduces
to
α0−1 ≈ f b2 + exp −b2 /2
(E3)
p
yielding for the optimal b ≈ 2| log f | . With this b, the
inverse capacity is dominated by the first term which is
α0−1 ≈ 2f | log f | .
The capacity for `2 balls of radius R with sparse labels
(E6)
As noted above when f → 0 the optimal bias diverges, hence the presence of order 1 the induced√margin is noticeable only for moderate f such that R D =
O(| log f |).
Small f and large R: Here we analyze the above equations in the limit of small f and large R. We assume that
f is sufficiently small so that the optimal bias b is large.
The contribution of the minority class to the inverse ca−1
pacity αB
is dominated by
Z
f
0
∞
Z
b
dtχD (t)
−∞
Dt0 ((−t0 + b)2 + t2 ) ≈ f b2
(E7)
−1
The dominant contribution of the majority class to αB
is,
26
(1 − f )
Z
∞
dtχD (t)
Dt0
−b−Rt−1
0
≈ (1 − f )
−b+tR
Z
Z
b̄
∞
(−t0 + tR − b)2
(1 + R2 )
(E8)
dtχD (t)(t − b̄)2
(E9)
where b̄ = b/R . In deriving the last equation we used
(−t0 +tR−b)2
→ (t−b̄)2 as R, b → ∞ . Second, the integrals
(1+R2 )
are of substantial value only if t ≥ b̄ in which case the
Rb
integral of t0 is −R(b̄−t) Dt0 ≈ 1 and the integral over
t is from b̄ to ∞. Combining the two results yields the
following simple expression for the inverse capacity,
−1
αBall
= f¯b̄2 +
Z
∞
dtχD (t)(t − b̄)2
[7]
[8]
[9]
[10]
(E10)
[11]
where scaled sparsity is f¯ = f R2 . The optimal scaled
bias is given by,
[12]
f¯b̄ =
Z
b̄
∞
b̄
dtχD (t)(t − b̄)
(E11)
Note that R and f affect the capacity only through the
scaled sparsity. When f¯ → 0, capacity is proportional
to (f¯| log f¯|)−1 . In a realistic regime of small f¯ (between
10−4 to 1) capacity decreases with f¯ roughly as 1/f¯ with
a proportionality constant which depends on D, as shown
in the examples of Fig. 12 (a). Finally, when R is sufficiently large that f¯ > 1, b̄ is order 1 or smaller. In this
case, the second term in Eq. E10 dominates and contributes ht2 i = D yielding a capacity that saturates as
D−1 as shown in Fig. 11(a). It should be noted however, that when b is not large, the approximations Eq.
E10-E11 don’t hold anymore.
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[1] James J DiCarlo and David D Cox. Untangling invariant object recognition. Trends in cognitive sciences,
11(8):333–341, 2007.
[2] Jennifer K Bizley and Yale E Cohen. The what, where
and how of auditory-object perception. Nature Reviews
Neuroscience, 14(10):693–707, 2013.
[3] Kevin A Bolding and Kevin M Franks. Complementary
codes for odor identity and intensity in olfactory cortex.
eLife, 6:e22630, 2017.
[4] H Sebastian Seung and Daniel D Lee. The manifold ways
of perception. Science, 290(5500):2268–2269, 2000.
[5] James J DiCarlo, Davide Zoccolan, and Nicole C Rust.
How does the brain solve visual object recognition? Neuron, 73(3):415–434, 2012.
[6] Ben Poole, Subhaneil Lahiri, Maithreyi Raghu, Jascha
Sohl-Dickstein, and Surya Ganguli. Exponential expressivity in deep neural networks through transient chaos.
[22]
[23]
[24]
[25]
In Advances In Neural Information Processing Systems,
pages 3360–3368, 2016.
Marc Aurelio Ranzato, Fu Jie Huang, Y-Lan Boureau,
and Yann LeCun. Unsupervised learning of invariant
feature hierarchies with applications to object recognition. In Computer Vision and Pattern Recognition, 2007.
CVPR’07. IEEE Conference on, pages 1–8. IEEE, 2007.
Yoshua Bengio. Learning deep architectures for ai. Foundations and trends R in Machine Learning, 2(1):1–127,
2009.
Ian Goodfellow, Honglak Lee, Quoc V Le, Andrew Saxe,
and Andrew Y Ng. Measuring invariances in deep networks. In Advances in neural information processing systems, pages 646–654, 2009.
Charles F Cadieu, Ha Hong, Daniel LK Yamins, Nicolas
Pinto, Diego Ardila, Ethan A Solomon, Najib J Majaj,
and James J DiCarlo. Deep neural networks rival the
representation of primate it cortex for core visual object
recognition. PLoS Comput Biol, 10(12):e1003963, 2014.
Thomas M Cover. Geometrical and statistical properties
of systems of linear inequalities with applications in pattern recognition. IEEE transactions on electronic computers, (3):326–334, 1965.
Elizabeth Gardner. The space of interactions in neural
network models. Journal of physics A: Mathematical and
general, 21(1):257, 1988.
E. Gardner. Maximum storage capacity in neural networks. EPL (Europhysics Letters), 4(4):481, 1987.
Vladimir Vapnik. Statistical learning theory. Wiley New
York, 1998.
SueYeon Chung, Daniel D Lee, and Haim Sompolinsky.
Linear readout of object manifolds. Physical Review E,
93(6):060301, 2016.
Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.
Ralph Tyrell Rockafellar. Convex analysis. Princeton
university press, 2015.
Jean-Jacques Moreau. Décomposition orthogonale d’un
espace hilbertien selon deux cônes mutuellement polaires.
CR Acad. Sci. Paris, 225:238–240, 1962.
SueYeon Chung, Uri Cohen, Haim Sompolinsky, and
Daniel D Lee. Learning data manifolds with a cutting
plane method. Neural Computation (Accepted), 2018.
Branko Grünbaum and Geoffrey C Shephard. Convex
polytopes. Bulletin of the London Mathematical Society,
1(3):257–300, 1969.
AA Giannopoulos, VD Milman, and M Rudelson. Convex
bodies with minimal mean width. In Geometric aspects
of functional analysis, pages 81–93. Springer, 2000.
Roman Vershynin. Estimation in high dimensions: a geometric perspective. In Sampling theory, a renaissance,
pages 3–66. Springer, 2015.
Kanaka Rajan, LF Abbott, and Haim Sompolinsky.
Stimulus-dependent suppression of chaos in recurrent
neural networks. Physical Review E, 82(1):011903, 2010.
Ashok Litwin-Kumar, Kameron Decker Harris, Richard
Axel, Haim Sompolinsky, and LF Abbott. Optimal degrees of synaptic connectivity. Neuron, 93(5):1153–1164,
2017.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages
248–255. IEEE, 2009.
27
[26] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going
deeper with convolutions. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 1–9, 2015.
[27] Seung-Hyeok Kye. Faces for two qubit separable states
and the convex hulls of trigonometric moment curves.
arXiv preprint arXiv:1302.2226, 2013.
[28] Zeev Smilansky. Convex hulls of generalized moment
curves. Israel Journal of Mathematics, 52(1):115–128,
1985.
[29] Rémi Monasson. Properties of neural networks storing spatially correlated patterns. Journal of Physics A:
Mathematical and General, 25(13):3701, 1992.
[30] B Lopez, M Schroder, and M Opper. Storage of corre-
[31]
[32]
[33]
[34]
lated patterns in a perceptron. Journal of Physics A:
Mathematical and General, 28(16):L447, 1995.
Sam T Roweis and Lawrence K Saul. Nonlinear dimensionality reduction by locally linear embedding. Science,
290(5500):2323–2326, 2000.
Joshua B Tenenbaum, Vin De Silva, and John C Langford. A global geometric framework for nonlinear dimensionality reduction. science, 290(5500):2319–2323, 2000.
Shun-ichi Amari, Naotake Fujita, and Shigeru Shinomoto. Four types of learning curves. Neural Computation, 4(4):605–618, 1992.
Sebastian Risau-Gusman and Mirta B Gordon. Statistical mechanics of learning with soft margin classifiers.
Physical Review E, 64(3):031907, 2001.
| 9cs.NE
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
An End-to-End Deep Learning Histochemical
Scoring System for Breast Cancer Tissue
Microarray
arXiv:1801.06288v1 [cs.CV] 19 Jan 2018
Jingxin Liu, Bolei Xu, Chi Zheng, Yuanhao Gong, Jon Garibaldi, Daniele Soria, Andew Green, Ian O. Ellis,
Wenbin Zou, Guoping Qiu
Abstract—One of the methods for stratifying different molecular classes of breast cancer is the Nottingham Prognostic Index Plus
(NPI+) which uses breast cancer relevant biomarkers to stain tumour tissues prepared on tissue microarray (TMA). To determine the
molecular class of the tumour, pathologists will have to manually mark the nuclei activity biomarkers through a microscope and use a
semi-quantitative assessment method to assign a histochemical score (H-Score) to each TMA core. However, manually marking
positively stained nuclei is a time consuming, imprecise and subjective process which will lead to inter-observer and intra-observer
discrepancies. In this paper, we present an end-to-end deep learning system which directly predicts the H-Score automatically. The
innovative characteristics of our method is that it is inspired by the H-Scoring process of the pathologists where they count the total
number of cells, the number of tumour cells, and categorise the cells based on the intensity of their positive stains. Our system imitates
the pathologists’ decision process and uses one fully convolutional network (FCN) to extract all nuclei region (tumour and non-tumour),
a second FCN to extract tumour nuclei region, and a multi-column convolutional neural network which takes the outputs of the first two
FCNs and the stain intensity description image as input and acts as the high-level decision making mechanism to directly output the
H-Score of the input TMA image. In additional to developing the deep learning framework, we also present methods for constructing
positive stain intensity description image and for handling discrete scores with numerical gaps. Whilst deep learning has been widely
applied in digital pathology image analysis, to the best of our knowledge, this is the first end-to-end system that takes a TMA image as
input and directly outputs a clinical score. We will present experimental results which demonstrate that the H-Scores predicted by our
model have very high and statistically significant correlation with experienced pathologists’ scores and that the H-Scoring discrepancy
between our algorithm and the pathologits is on par with that between the pathologists. Although it is still a long way from clinical use,
this work demonstrates the possibility of using deep learning techniques to automatically and directly predicting the clinical scores of
digital pathology images.
Index Terms—H-Score, Immunohistochemistry, Diaminobenzidine, Convolutional Neural Network, Breast Cancer
F
1
I NTRODUCTION
Breast cancer (BC) is a heterogeneous group of tumours
with varied genotype and phenotype features [1]. Recent
research of Gene Expression Profiling (GEP) suggests that
BC can be divided into distinct molecular tumour groups
[2], [3]. Personalised BC management often utilizes robust
commonplace technology such as immunohistochemistry
(IHC) for tumour molecular profiling [4], [5].
Diaminobenzidine (DAB) based IHC techniques stain
the target antigens (detected by biomarkers) with brown
colouration (positive) against a blue colouration (negative)
counter-stained by Hematoxylin (see Fig.1 for some example images). To determine the biological class of the tumour, pathologists will mark the nuclei activity biomarkers
through a microscope and give a score based on a semi•
•
•
•
•
•
•
J. Liu, B. Xu, W. Zou, and G. Qiu are with the College of Information
Engineering, Shenzhen University, China; G. Qiu is also with School of
Computer Science, The University of Nottingham, UK
C. Zheng is with Ningbo Yongxin Optics Co., LTD, Zhejiang, China
Y. Gong is with Computer Vision Laboratory , ETH Zurich, Switzerland
J. Garibaldi is with School of Computer Science, The University of
Nottingham, UK
D. Soria is with Department of Computer Science, The University of
Westerminster, UK
A. Green and I. O. Ellis are with Faculty of Medicine & Health Sciences,
The University of Nottingham, United Kingdom.
G. Qiu is the corresponding author, E-mail:[email protected]
quantitative assessment method called the modified histochemical scoring (H-Score) [6], [7]. The H-Scores of tissue
samples stained with different biomarkers are combined
together to determine the biological class of a case. Clinical decision making is to choose an appropriate treatment
from a number of available treatment options according
to the biological class of the tumour. For instance, one of
the methods for stratifying different molecular classes is
the Nottingham Prognosis Index Plus (NPI +) [1] which
uses 10 breast cancer relevant biomarkers to stain tumour
tissues prepared on tissue microarray (TMA). Tissue samples stained by each of these 10 biomarkers are given a
histochemical score (H-Score) and these 10 scores together
will determine the biological class of the case.
Therefore, H-Score is one of the most important pieces of
information for molecular tumour classification. When the
tumour region occupies more than 15% of the TMA section,
a H-Score is calculated based on a linear combination of the
percentage of strongly stained nuclei (SSN ), the percentage
of moderately stained nuclei (M SN ) and the percentage of
weakly stained nuclei (W SN ) according to equation (1):
H − Score = 1 × WSN + 2 × MSN + 3 × SSN
(1)
The final score has a numerical value ranges from 0 to
300. Thus, the histochemical assessment of the TMA’s is
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
2
Fig. 1: Top: Example images extracted from digital TMA slides. Each red circle contains one TMA core stained by
Diaminobenzidine-Hematoxylin (DAB-H). The brown colours indicate positive and the blue colours indicate negative.
Bottom: A schematic illustration of the traditional manual H-Scoring procedure: It needs to first count the total number of
nuclei, then the number of strongly stained, moderately stained and weakly stained tumour nuclei, respectively. The final
H-Score is then calculated according Eq.1.
based on the following semi-quantitative information: the
total number of cells, the number of tumour cells and
the stain intensity distributions within the tumour cells.
In clinical practice, diagnosis requires averaging two experienced pathologists’ assessments. Manually marking the
positively stained nuclei is obviously a time consuming
process. As visual assessment of the TMA’s is subjective,
there is the problem of inter-observer discrepancy and the
issue of repeatability. The semi-quantitative nature of the
method (strongly stained, moderately stained and weakly
stained, the definitions of strong, moderate and weak cannot
be precise and subjective), makes it even more difficult to
ensure inter-subject as well as intra-subject consistency.
With the increasing application of clinicopathologic
prognosis, Computer Aided Diagnosis (CAD) systems have
been proposed to support the pathologists’ decision making.
The key parameters in tissue image assessment include the
number of tumour cells, the positive staining intensities
within these cells and the total number of all cells in the
image. To classify the positively stained pixels and their
stain intensity, methods such as colour deconvolution that
perform mathematical transformation of the RGB image [8]
[9] are widely used to separate positive stains from negative
stains. Numerous computer-assisted approaches have been
proposed for cell or nuclei detection and segmentation [10].
Most literature on histopathology image analysis perform
various low-level quantification steps, there is still little
attempt to perform end-to-end assessment of the image
directly.
In this paper, we ask this question: is it possible to
develop a CAD model that would directly give a highlevel assessment of a digital pathological image, just like
an experienced pathologist would, for example, to give out
a H-Score directly? In an attempt to answer this question,
we propose an end-to-end deep learning system for directly
predicting the H-Scores of breast cancer TMA images, see
Fig. 6. Instead of pushing the raw digital images into the
neural network directly, we follow a similar process that
pathologists use for H-Score estimation. We first construct
a stain intensity nuclei image (SINI) which only contains
nuclei pixels and their corresponding stain intensity information, and a stain intensity tumour image (SITI) which
only contains tumour nuclei pixels and their corresponding
stain intensity information. The SINI and SITI block irrelevant background pixels while only retain useful information
for calculating the H-Score. These two H-Score relevant images are then fed into a dual-channel convolutional neural
network with two input pipelines, which are finally merged
into one pipeline to give an output (H-Score). To the best
of our knowledge, this is a first work that attempts to
develop deep learning based TMA processing model that
directly outputs the histochemical scores. We will present
experimental results which demonstrate that the H-Scores
predicted by our model have high and statistically significant correlation with experienced pathologists’ scores and
that the H-Scoring discrepancy between our algorithm and
the pathologists is on par with that between the pathologists. Although it is still perhaps a long way from clinical
use, this work nevertheless demonstrates the possibility of
automatically scoring cancer TMA’s based on deep learning.
2
R ELATED W ORKS
Researchers have proposed various computer-assisted analysis methods for histopathological images [11]. For pixellevel positive stain segmentation, Pham [12] adapted Yellow
channel in CMYK model, which is believed to have strong
correlation with the DAB stain; Ruifrok [8] presented the
brown image calculated based on mathematical transformation of the RGB image. Yao [13] employed Hough forest for
mitotic cell detection, which is a combination of generalized
Hough transform and random decision trees. Shu et al. [14]
proposed utilizing morphological filtering and seeded watershed for overlapping nuclei segmentation. Object-based
CAD systems have also been developed for tubule detection
in breast cancer [15], glandular structure segmentation [16],
and etc.
With the development of deep learning techniques, various deep neural network based CAD models have been
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
published. Deep convolutional networks with deeper architectures can be used to build more complex models which
will result in more powerful solutions. Li [17] used a 88layer residual network for human epithelial type 2 (HEp-2)
cell segmentation and classification. AggNet with a novel
aggregation layer is proposed for mitosis detection in breast
cancer histology images [18]. Google brain presented a multi
scale CNN model to aid breast cancer metastasis detection
in lymph nodes [19]. A deep learning-based system is
proposed for the detection of metastatic cancer from whole
slide images, which won the Camelyon Grand Challenge
2016 [20]. Shah et al. [21] presented the first completely
data-driven model integrated numerous biologically salient
classifiers for invasive breast cancer prognosis. A symmetric
fully convolutional network is proposed by Ronneberger for
microscopy image segmentation [22].
Digital pathology is relative new compared with other
type of medical imaging such as X-ray, MRI, and CT. Deep
learning as one of the most powerful machine learning
techniques emerged in recent years has seen widespread
applications in many areas. Yap et al. [23] investigated three
deep learning models for breast ultrasound lesion detection.
Moeskops [24] introduced a single CNN model with triplanar input patches for segmenting three different types of
medical images, brain MRI, breast MRI and cardiac CTA. A
combination of multi-channel image representation and unsupervised candidate proposals is proposed for automatic
lesion detection in breast MRI [25].
Most existing high-level CAD frameworks directly follow the assessment criteria by extracting quantitative information from the digital images. Masmoudi et al. [26] proposed an automatic Human Epidermal Growth Factor Receptor 2 (HER2) assessment method, which is an assemble
algorithm of colour pixel classification, nuclei segmentation
and cell membrane modelling. Gaussian-based bar filter was
used for membrane isolation after colour decomposition in
[27]. Trahearn et al. [28] established a two-stage registration
process for IHC stained WSI scoring. Thresholds were defined for DAB stain intensity groups, and tumour region and
nuclei were detected by two different detectors. Recently,
Zhu [29] proposed to train an aggregation model based
on deep convolutional network for patient survival status
prediction.
3
P ROBLEM AND M ETHOD
An immunohistochemical assessment can be formulated as
a model F that maps the input images from the input space
I to the a label space L. Given an input image I ∈ I , its label
l ∈ L is assigned according to the quantitative information
of positive staining intensity Ps , the number of tumour cells
Nt and total number of cells Ne in the image x:
l = F(I|Ps , Nt , Ne ),
(2)
Traditional assessment methods have at least three unsolved issues for both the pathologists and the CAD systems. Firstly, the positive staining intensity needs to be
categorized into four classes: unstained, weak, moderate, and
strong. However, there is no standard quantitative criterion
for classifying the DAB stain intensity. Thus, two pathologists often classify the same staining intensity into two
(a)
(b)
(c)
Fig. 2: Examples of challenging cases of quantitative measurement of biomarkers based on visual assessment. (a) A
variety of stain intensities; (b) unclear staining and overlapping of nucleus; (c) Size differences between different type
of nucleus.
different categories or two different intensities into the same
category. Furthermore, the human visual system may pay
more attention to strongly stained regions but they are often
surrounded by a variety of staining intensities [28], which
may also affect the assessment results. Secondly, cell/nuclei
instance counting is a very important parameter in the
assessment. Nevertheless, both human and computer still
cannot deal with the difficulty of counting overlapping
cells very well. Moreover, variability in the appearance of
different types of nucleus, heterogeneous staining, and the
complex tissue architectures make individually segmenting
cell/nuclei a very challenging problem. Thirdly, the apparent size differences between tumour nuclei and normal
nuclei will affect the quantitative judgement of tumour
nuclei assessment. Examples of these challenging cases are
illustrated in Fig. 2.
To tackle the problem mentioned above, we propose
to develop a convolutional neural network (CNN) based
CAD framework for biomarker assessment of TMA images.
Instead of using CNN as a feature extractor or for low
level processing such as cell segmentation only, we have
developed an end-to-end system which directly predicts the
biomarker score (H-Score). The innovative characteristic of
our method is that it is inspired by the H-Scoring process
of the pathologists where they count the total number of
nuclei and the number of tumour nuclei and categorise
tumour nucleus based on the intensity of their positive
stains. In the complete system, as illustrated in Fig. 6, one
fully convolutional network (FCN) is used to extract all
nuclei region which acts as the step of counting all nucleus
and capture all foreground information, another FCN is
used to extract tumour nuclei region which acts as the step
of counting all tumour nucleus. To mimic the process of
categorising tumour nuclei based on their positive stain
intensities, we derive a stain intensity image which together
with the outputs of the two FCNs are presented to another
deep learning network which acts as the high-level decision
making mechanism to directly output the H-Score of the
input TMA image.
3.1
Stain Intensity Description
Although various DAB stain separation methods have been
proposed [9], [30], few work studied the stain intensity
description and grouping. Since there is no formal definitions for the boundaries between stain intensity groups
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Original image
4
DAB channel image
Stain intensity discription
image
Fig. 3: A comparison of different images generated during the process of stain intensity description. The highlighted
subimage contains strongly stained nuclei.
(e.g, strong, moderate, weak), previous works used manually
defined thresholds for pixel-wise classification to segment
positive stains into each stain group [28].
Most previous works set a single threshold on IDAB
to separate positively stained tissues. However, as shown
in Fig.3, the deeply stained positive nuclei can have dark
and light pixel values on the DAB channel image, since the
strongly stained pixels will have significantly broader hue
spectrum. Furthermore, as illustrated in Fig.4, the same DAB
channel value can correspond to different pixel colours.
Also, from Fig.4, it is clear that in order to separate the
positive stain (brown colour) from the negative stain (blue
colour), the DAB channel thresholds should be set based on
the luminance values. In this paper, we use the Luminance
Adaptive Multi-Thresholding (LAMT) method developed
by the authors [32] to classify positively stained pixels.
Specifically, the transformed pixel IDAB (m, n) is divided
into K equal intervals according to the luminance:
k
k
IDAB
(m, n) = {IDAB
(m, n) ∈ IDAB )|ξ k < Il (m, n) ≤ ζ k }
Fig. 4: Visualization of pixel colours of the DAB-H images
along the luminance axis and the colour deconvolution DAB
axis.
In this work, we propose to directly use the luminance
values of the image to describe the staining intensity instead of setting artificial intensity category boundaries. The
original RGB image I is first transformed into three-channel
stain component image (IDAB−H = [IDAB , IH , IOther ]) using
colour deconvolution [9]:
IDAB−H = M −1 IOD ,
(3)
where M is the stain matrix composed of staining colours
equal to
0.268 0.570 0.776
0.650 0.704 0.286
(4)
0.0
0.0
0.0
for DAB-H stained images, and IOD is Optical Density
converted image calculated according Lambert-Beers law:
IOD = −log(
I
),
I0
(5)
I0 = [255, 255, 255] is the spectral radiation intensity for a
typical 8bit RGB camera [31].
Only the DAB channel image IDAB from the three colour
deconvolution output channels is used, which describes the
DAB stain according the chroma difference.
(6)
where k = 1, 2, ..., K ; ξi and ζi are lower and upper
boundary respectively of k th luminance interval. Il is the
luminance image of the original RGB image calculated
according to Rec. 601 [33]:
Il = 0.299 × IR + 0.587 × IG + 0.114 × IB .
(7)
The transformed pixels are thresholded with different values according to its luminance instead of a single threshold,
the threshold tk is assigned as follows:
k
tk = argmax P (c|IDAB
(m, n))
(8)
c∈C
where C = {cDAB , cH } is the stain label.
Once we have separated the positive stain from the
negative stain, we need to find a way to describe the stain
intensity. As we have already seen in Fig.3 and Fig.4, the
pixel values of IDAB can not describe the biomarker stain
intensity. We propose to use a scheme described in Eq.9 to
assign stain intensity values to pixels:
Ila (m, n) =
(
Il (m, n)
255 + (255 − Il (m, n))
if IDAB (m, n) is positive
(9)
if IDAB (m, n) is negative
where Ila (m, n) is the stain intensity description image.
The idea is that for the positive stain pixels, Ila (m, n) is
the same as the luminance component of the original image
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
in order to preserve the morphology of the positive nuclei;
for the negative stain pixels, Ila (m, n) will have a higher
value for strongly stained pixels (darker blue colour) and a
lower value for weakly stained pixels (lighter blue colour).
In order to separate the positive and negative pixel values
clearly, we add an offset of 255 to the negatively stained
pixels (most negative stain pixels will have a high Il (m, n)
and positive stain pixels will have a low Il (m, n), the value
of positive and negative pixels will be clearly separated in
Ila (m, n)). Therefore, the larger Ila (m, n) is, the weaker is
the stain, the smaller Ila (m, n) is, the stronger is the stain.
When Ila (m, n) is below or equal to 255, it is a positive
stain pixel. In this way, we have obtained an image which
gives a continuous description of the stain intensity of the
image. Instead of setting artificial boundaries to separate
the different degrees of stain intensity, we have now a
continuous description of the stain intensity (see Fig.5). Note
that the pixel values of final image are normalized to the
range from 0 to 1.
3.2
Nuclei and Tumour Maps
As discussed above, the important information pathologists
use to come up with the H-Score is the number of nuclei
and the number of tumour nuclei in the TMA image. We
therefore need to extract these two pieces of information and
we use two separate FCNs, one for segmenting all nucleus
and the other for segmenting tumour nucleus only.
To segment the tumour region, we use our own manually pixel-wise labelled tumour TMA images to train the
FCN. While for segmenting general nuclei which detects
both tumour and non-tumour nuclei, we utilize a transfer
learning strategy to train another FCN. For general nuclei
detection, the training data is obtained from three different datasets: immunofluorescence (IIF) stained HEp-2 cell
dataset [34], Warwick hematoxylin and eosin (H&E) stained
colon cancer dataset [35], and our own DAB-H TMA images.
Since these three image sets are stained with different types
of biomarker, we transform the colour image into grayscale
for training. Training on a mixed image set could help to
reduce overfitting on limited medical dataset and further
boost the performance and robustness [36].
For both the general nuclei detection network and the
tumour nuclei detection network, we use the symmetric U
shape network architecture (U-Net) [22] with skip connection. The high resolution features from the contracting path
are combined with the output from the upsampling path,
which allows the network to learn the high resolution contextual information. The loss function is designed according
the Dice coefficient as:
P
2 m,n ωτ
P
Lmask = −log P
,
(10)
2
2
m,n ω +
m,n τ
where ω is the predicted pixel and τ is the ground truth.
5
3.3
The H-Score Prediction Framework
The overview of the H-Score prediction framework is illustrated in Fig.6. It consists of three stages: 1) Nuclei segmentation, tumour segmentation, and stain intensity description; 2) Constructing the Stain Intensity Nuclei Image (SINI)
and the Stain Intensity Tumour Image (SITI); 3) Predicting
the final histochemical score (H-Score) by the Region Attention Multi-column Convolutional Neural Network (RAMCNN). The rationale of this architecture is as follows: as
only the number of nuclei, the number of tumour nuclei
and the stain intensity of the tumour nuclei are the useful
information for predicting H-Score, we therefore first extract
these information. Rather than setting artificial boundaries
for the categories of stain intensity, we retain a continuous
description of the stain intensity. Only the information useful for predicting the H-Score is presented to a deep CNN to
estimate the H-Score of the input image. This is in contrast
to many work in the literature where the whole image is
thrown to the CNN regardless if a region is useful or not for
the purpose.
The detail of the first stage have been described in
Section 3.1 and 3.2. As illustrated in Fig.6, an input TMA
image I(m, n) is processed by the tumour detection network
which will output a binary image mask, T (m, n), marking
all the tumour nuclei, where T (m, n) = 1 if I(m, n) is a
part of a tumour nuclei and T (m, n) = 0 otherwise; by the
general nuclei detection network which will output another
binary image mask, G(m, n), marking all tumour and nontumour nuclei, where G(m, n) = 1 if I(m, n) is a part of
a nuclei and G(m, n) = 0 otherwise; and by the colour
deconvolution and stain intensity labelling operation of
Equation (8) to produce the stain intensity description image
Ila (m, n). In the second stage, we construct SINI and SITI
by multiplying the nuclei mask image G(m, n) and tumour
mask image T (m, n) with the stain intensity description
image Ila (m, n), i.e. SIN I = Ila (m, n) × G(m, n), and
SIT I = Ila (m, n) × T (m, n). Hence, all background pixels
are zero, while only region of interests (ROI) are retained
in SINI and SITI. All necessary information is preserved
for histochemical assessment. Removing the background
and only retaining ROI will enable the RAM-CNN convolutional layers to focus on foreground objects [37] which
will significantly reduce computational costs and improve
performance.
The proposed RAM-CNN is a deep regression model
with dual input channels. The architecture of RAM-CNN
is shown in Table 1. Two inputs correspond to SINI and SITI
respectively, and the input size is 512 × 512. The parameters
of the two individual branches are updated independently
for extracting cell and tumour features respectively, without
interfering with each other. The two pipelines are merged
into one after two convolutional layers for H-Score prediction. The loss function for H-Score prediction is defined as:
Fig. 5: An illustration of the value of lla and its corresponding stain intensity. The red dot lines are the thresholds of stain
intensity groups [28].
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
Fig. 6: The overview of our proposed H-Score prediction framework. An input TMA image is first processed by two
FCNs to extract tumour cells and all cells (tumour and non-tumour) to produce two mask images. The input image is
also processed by colour deconvolution and positive stain classification to output a stain intensity description image. The
two mask images are used to filter out irrelevant information in the stain intensity description image and only the useful
information is fed to a deep convolutional neural network for the prediction of H-Score of the input TMA.
6
Image Number
5
4
3
2
1
Augmented Image Number
0
0
15
30
45
60
75
90
0
15
30
45
60
75
90
105
120
135
150
165
180
195
210
225
240
255
270
285
300
40
35
30
25
20
10
15
5
0
105
120
135
150
165
180
195
210
225
240
255
270
285
300
H-Score
Fig. 7: The top graph is the original dataset label histogram; The bottom is the augmented label histogram.
Layer
Input
Conv1
MaxPooling
Conv2
MaxPooling
Conv3
MaxPooling
Conv4
MaxPooling
FC1
FC2
Input / Filter Dimensions
512 × 512 × 1 512 × 512 × 1
8×7×7
8×7×7
2×2
2×2
16 × 5 × 5
16 × 5 × 5
2×2
2×2
64 × 3 × 3
2×2
64 × 3 × 3
2×2
2048
1024
4
4.1
TABLE 1: The architecture of Region Attention Multichannel Convolutional Neural Network (RAM-CNN).
Lscore =
N
1 X
kF
(SIN Ii , SIT Ii ) − li k2 ,
N i=1 RAM
(11)
where FRAM (SIN Ii , SIT Ii ) is the estimated score generated by RAM-CNN. li is the ground truth H-Score.
E XPERIMENTS AND R ESULTS
dataset
The H-Score dataset used in our experiment contains 105
TMA images of breast adenocarcinomas from the NPI+
set [1]. Each image contains one whole TMA core. The
tissues are cropped from a sample of one patient which are
stained with three different nuclei activity biomarkers: ER,
p53, and PgR. The original images are captured at a high
resolution of 40× optical magnification, and then resized to
1024×1024 pixels. The dataset is manually marked by two
experienced pathologists with H-Score based on common
practice. For each TMA core, the pathologists give the
percentage of nuclei of different stain intensity levels, and
then calculate the H-Score using Eq.1. The final label (HScore) is determined by averaging two pathologists’ scores,
if the difference between two pathologists is smaller than
20. The dataset is available from the authors on request.
For training the general nuclei detection network, we
transform Warwick H&E colon adenocarcinoma [35] and
NPI+ images to grayscale; the green channel was extracted
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
7
Fig. 8: Examples of intermediate images in the automatics H-Score prediction pipeline. From left to right: the original RGB
image, luminance labelled stain intensity image, nuclei mask image, and tumour mask image respectively.
from HEp-2 cell dataset [34]. As HEp-2 cell images are IIF
stained, the gray value should be inversed.
4.2
Data and Label Augmentation
As in typical medical imaging applications, the dataset sizes
are relatively small. In developing deep learning based
solutions, it is a common practice to augment the training
dataset for training. The training images for general nuclei
detection network and tumour detection network are augmented by randomly cropping sub-images as input samples.
For the H-Score dataset, rotation with random angles and
randomly shifting the image horizontally and vertically
within 5% of image height and width are performed to
augment the training set.
As shown in the top row of Fig.7, the distribution of
the label (H-Score) in the original dataset is unbalanced,
some labels (H-Scores) have far more samples than others.
Furthermore, one of the biggest problems is that because
we have only limited number of samples, the H-Score
values are discrete and discontinuous. There are many gaps
between two H-Scores that has no data. Also, the values
of the TMA image score given by the pathologists have a
quantitative step-size of 5. Therefore, if an image has a score
of 125, it means it has a value of around 125, the values in
the vicinity of 125, i.e., 126 or 124 should also be suitable for
labelling that image. In order to solve the ambiguity issue,
we introduce Distributed Label Augmentation (DLA) which
was inspired by the work of [38], [39].
In the traditional regression method, a given dataset
{(I1 , l1 ), (I2 , l2 ), · · · , (ID , lD )} pairs the instance Id for 1 ≤
d ≤ D with
one single ld from the finite class label
space L = l0 , l1 , · · · , lC , where C is the label size (e.g.,
C = 301 for H-Score). In this paper, the label is augmented so that one instance is associated with a num-
ber of labels. Formally, the dataset can be described as
{(I1 , Y1 ), (I2 , Y2 ), · · · , (ID , YD )}, and Yd ⊆ Y is a set of
(1) (2)
(S)
labels {yd , yd , · · · , yd }, where S is the augmented label
(s)
number for Id . yd is sampled repeatedly from L based
on a probability density function of following Gaussian
distribution:
(lc − µ)2
1
√ exp(−
)
(12)
2σ 2
σ 2π
where µ is the mean which equal to ld and σ is stanPS
(s)
dard deviation. Thus,
s=1 p(yd ) = 1 for each original
TMA image. Consequentially, for an image xi from the
augmented training set, its ground truth labels are assigned
by repeatedly sampling from L according to Eq.12. The
augmented label histogram is shown at the bottom row of
Fig.7.
(s)
p(yd = lc ) =
4.3
Implementation Details
The network architecture for both the tumour nuclei detection and general nuclei detection models is the same as
the U-Net [22] with a input size of 224 × 224. The filter
size of tumour detection net is half narrower than that of
general cell detection net. All networks use rectified linear
unit (ReLU) activation function for the convolutional layer.
The final cell and tumour region maps are predicted using
sliding window.
A leave 5 out cross validation strategy is used for RAMCNN model training, which means that in each round of
testing, we randomly sample 5 TMAs as testing and the
other 100 TMAs as training images. As explained previously,
the training set is augmented via rotation and shift. The images are firstly resized to 512×512 before fed into the RAMCNN. We set σ = 0.9 to generate the H-Score distribution
for ground truth label augmentation. We also add dropout
8
300
300
250
250
200
200
Predicted Score
Predicted Score
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
150
150
100
100
50
50
0
0
50
100
150
200
250
0
300
0
50
100
Pathologist's Score
(a) NAP
200
300
250
250
250
200
200
200
Predicted Score
300
150
150
100
100
50
50
50
0
50
100
150
200
Pathologist's Score
(c) RGB-CNN
250
300
0
0
300
150
100
0
250
(b) NNP
300
Predicted Score
Predicted Score
150
Pathologist's Score
50
100
150
200
250
300
0
0
50
100
Pathologist's Score
150
200
250
300
Pathologist's Score
(d) RA-CNN
(e) RAM-CNN
Fig. 9: Scatter plots of the predicted scores of different models vs the pathologists’ manual scores.
4.4.1 Experimental Results
Fig.8 shows some examples of the intermediate images in
the automatic H-Score prediction pipeline. It is seen that
the luminance labelled stain intensity image marks a sharp
distinction between positive and negative stains. This shows
that our maximum a posteriori (MAP) classifier based Luminance Adaptive Multi-Thresholding (LAMT) method [32]
can reliably separate positive DAB stains for a variety of images. It also shows that our stain intensity labelling strategy
can preserve the morphology of the nuclei, separate positive
and negative stains while retaining a continuous description
of the positive stain intensity.
Fig.10 shows the training curves of the Dice coefficient
for the general nuclei detection network and the tumour
nuclei detection network respectively. Both networks converged after 170 epochs respectively. The nuclei mask images (see Fig.8) show that the deep convolutional network
trained with mixed datasets using transfer learning can
successfully detect the nuclei in our H-Score dataset. The
tumour segmentation network is able to identify tumour region from normal tissues. It is worth noting that the ground
truth masks for the two detection networks are different. All
nucleus in Warwick colon cancer images are labelled with
circular masks with a uniform size, while tumour region
masks are pixel level labelled. Therefore, the final predicted
maps generated by the two networks for the same nucleus
are different. In addition, it is found that the mask dilation
become evident with the increase of DAB stain intensity.
0.9
0.9
Results and Discussions
0.85
0.8
0.8
Dice coefficient
4.4
One possible reason is that the strong homogeneous stain
makes the nuclei texture and edge feature difficult to extract.
Dice coefficient
layers after two fully connected layers with the rates of 0.3
and 0.5 respectively. The regression network is optimized by
Adam [40] with an initial learning rate of 0.001.
0.7
0.75
0.6
0.7
0.65
0.5
0.6
0.55
0.4
training
test
0.3
0
25
50
75
100
Epoch
(a)
125
150
175
training
test
0.5
200
0.45
0
25
50
75
100
Epoch
125
150
175
200
(b)
Fig. 10: Training results for general nuclei detection network
(a) and tumour nuclei detection network (b).
To evaluate the performance of our proposed RAMCNN and the two H-Score relevant images SINI and SITI,
we compare our model with two traditional single input
pipeline CNNs: RGB-CNN and RA-CNN (i.e., region attention CNN). The RGB-CNN takes the original RGB TMA
image with the shape of 512 × 512 × 3 as input, and output
the H-Score prediction. To investigate the effect of multicolumn architecture, we combine SINI and SITI as a two
channel image of 512 × 512 × 2 for the input of RA-CNN.
The architectures of RGB-CNN and RA-CNN are the same
as a single pipeline RAM-CNN (see Table.1).
We also calculate the H-Score using Eq.1 based on the
nuclei area percentage (NAP) and nuclei number percentage
(NNP). Specifically, the luminance labelled stain intensity
description image Ila is first calculated according to the
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
(a) 0-49
(b) 50-99
(c) 100-149
9
(d) 150-199
(e) 200-249
(f) 250-300
Fig. 11: Example TMA images extracted from different H-Score groups.
description in Section 3.1. The pre-defined thresholds [28]
are utilized for categorizing the pixels into different DAB-H
stain intensity groups. For the NAP method, the predicted
H-Score is calculated according to the percentages of area
in different stain intensity groups. NNP employs the NIH
ImageJ tool [41] with a multi-stage cell segmentation technique [14] for cell detection. The detected cells are classified
into unstained, weak, moderate, and strong groups using the
pre-defined thresholds for H-Score calculation .
Model
NAP
NNP
RGB-CNN
RA-CNN
RAM-CNN
Human
MAE
47.09
46.48
32.01
27.22
21.33
20.63
SD
46.03
55.18
44.46
35.72
29.14
30.55
CC
0.87
0.82
0.87
0.92
0.95
0.95
P value
< 0.001
< 0.001
< 0.001
< 0.001
< 0.001
< 0.001
TABLE 2: Performance comparison with different regression
models. The last line (Human) are difference between the HScores given by the two pathologists.
pathologists (MAE = 20.63, SD = 30.55, CC = 0.95). While the
MAE between the RAM-CNN and humans is slightly higher
than that between humans, the SD between humans is
higher than that between RAM-CNN and humans. The CC
between humans and machine and that between humans
are the same.
Fig. 9 illustrates the scatter plots between the model
predicted scores and the pathologists’ scores. Most of the
predicted scores of NAP are lower than the ground truth.
At the lower end, NNP predicted scores are lower than the
ground truth while at the higher end the predicted scores
are higher than the ground truth. These two methods are
affected by several low-level processing components including the pre-defined stain intensity thresholds and the nuclei segmentation accuracy. Our proposed framework gives
more accurate prediction results compared to traditional
single pipeline CNN, further demonstrating that imitating
the pathologists’ H-Scoring process by only keeping useful
information is an effective approach.
4.4.2
In this paper, Mean Absolute Error (MAE), Standard
Deviation (SD) and the correlation coefficient (CC) between
the predicted H-Score and the average H-Score of the two
pathologists are used as the evaluation metrics. As a reference, we also calculate the MAE, SD and CC between the HScores given by the two pathologists of all original diagnosis
data. Results are shown in Table.2.
As can be seen, the NAP based prediction gives the highest MAE with large deviations in the cross validation, which
is followed by NNP. Our RAM-CNN framework achieves
the lowest prediction error (21.33); a traditional CNN setting
with the proposed SINI and SITI as input gives the second
lowest prediction error (27.22). This verifies the effectiveness
of our proposed approach to filtering out irrelevant pixels
and only retain H-Score relevant information in SINI and
SITI. All deep learning based methods outperform NAP
and NNP by a large margin. To investigate the statistical
significance of automatically predicted H-Scores, the correlation of the predicted and those of the pathologists scores
and its P value are also calculated. The correlation between
pathologists scores and those predicted by RAM-CNN is
0.95 with a P value of < 0.001 which means there is strong
evidence against the null hypothesis [42].
It is interesting to observe that the difference between
our RAM-CNN predicted H-Scores and the average of the
two pathologists H-Scores (MAE = 21.33, SD = 29.14, CC
= 0.95) are on par with the difference between the two
Discussions
In this paper, we introduced an end-to-end system to predicted H-Score. To investigate the reason for scoring discrepancy between the proposed algorithm and the pathologists, we firstly compare the H-Score prediction results
for different biomarkers as shown in Table.3. The proposed
framework gives the best accuracy in all three biomarker
images. The performances are slightly different for different
biomarkers. This is to be expected because different markers
will stain the tissues differently. Although the difference
is not large, whether it will be useful to train a separate
network for different biomarkers is something worth investigating in the future.
Biomarker
No. of TMA
NAP
NNP
RGB-CNN
RA-CNN
RAM-CNN
ER
32
42.02
43.53
24.90
25.43
21.01
P53
33
50.68
46.72
31.57
23.39
16.66
PgR
40
48.17
48.92
38.19
31.82
25.44
TABLE 3: Comparing MAE of different methods on three
different biomarkers.
To see how the algorithms perform differently across the
dataset, we divide the TMA images into 6 groups according
to their pathologists’ scores. Example TMA images of each
group are illustrated in Fig.11. For each group, we count
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
30
25
20
15
10
5
0
30
25
20
15
10
5
0
30
25
20
15
10
5
0
30
25
20
15
10
5
0
30
NAP
10
AE >= 30
10 <= AE < 30
AE < 10
NNP
RGB-CNN
(a)
RA-CNN
RAM-CNN
25
20
15
10
5
0
0-49
50-99
100-149
150-199
200-249
250-300
Fig. 12: Comparison of performances of different methods
in different H-Score groups.
the number of TMAs with absolute error (AE) smaller than
10, between 10 and 30, and larger than 30 respectively. The
results of different methods are shown in Fig.12. It is seen
that in the low H-Score group of 0-49, traditional methods
of NAP and NNP give more accurate predicted scores than
CNN based methods. It is found that most low score TMAs
are unstained or weakly stained as shown in Fig.11(a). The
accurate predictions from NAP and NNP indicate that the
predefined threshold for separating unstained and weak (see
Fig. 5) is compatible with pathologists’ criteria. The deep
learning based methods do not set stain intensity thresholds
explicitly and their performances across the six groups are
relatively even.
The accuracies of NAP and NNP decrease rapidly with
the increase of the H-Score. As shown in Fig.11, the stain
intensity and image complexity increase with the H-Score
which directly affect the performance of traditional methods. The result also indicates that the pre-defined stain
intensity thresholds for moderate and strong classes (see
Fig. 5) are less compatible with the pathologists’ criteria.
Furthermore, the large coefficients of moderate and strong
stain (see Eq.1) would magnify the errors of area and nuclei
segmentation in NAP and NNP respectively.
Three deep learning based methods give worse results
on the groups with fewer images (i.e., group 50-99 and 250300), which indicates the importance of a large training data
size. In addition, the uneven distribution of original dataset
may also affect the predicted accuracy.
We further analyse the TMAs individually to investigate
the effect of image quality on the proposed algorithm. We
found that for those TMAs where the tissues are clearly
stained, and the cellular structure is clear without severe
overlap (see Fig.13), our algorithm can give very accurate
prediction. On the other hand, poor image quality causes
errors. In the images that are most easily mis-scored by
(b)
Fig. 13: Examples of accurately scored TMAs by proposed
algorithm. The absolute errors generated by RAM-CNN of
both (a) and (b) are smaller than 2.
our algorithms, we found three significant characteristics as
shown in Fig.14.
The TMA core in Fig.14(a) contains large out-of-focus
regions, which happens more commonly on strongly stained
tissues. The blur regions directly affect the performance of
nuclei segmentation, as well as the nuclei and tumour detection accuracy. They also hinder the final regression network
from extracting topological and morphological information.
Tissue folds (see Fig.14(b)) occurs when a thin tissue slice
folds on itself, and it can happen easily during slide preparation especially in TMA slides. Tissue-fold would cause outof-focus during slide scanning. Furthermore, a tissue fold
in a lightly stained image can be similar in appearance to
a tumour region in a darkly stained image [43]. Hence, the
segmentation accuracy of colour deconvolution would be
greatly affected in tissue-fold regions.
Heterogeneity and overlapping as shown in Fig.14(c)
also affect the automatic scoring performance. The stain
heterogeneity gives rise to a large discrepancy of stain
intensity in a single nucleus, and nuclei overlapping adds
to the difficulty.
These three difficulties directly affect the predicted results of the proposed method, and we found that most large
mis-scored TMAs contain one or more of these characteristics. We found that there were 9 low image quality TMAs
in our dataset and if we exclude these 9 lowest-quality
TMA images, the average MAE of our RAM-CNN is 18.79.
Therefore, future works need to overcome these issues in
order to achieve a high prediction performance. To solve
the problem of out-of-focus, heterogeneity and overlapping,
adding corresponding images in the training set to promote
robustness is one potential quality assurance methods. In
addition, the deep learning based scoring system can be
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
(a)
11
(b)
(c)
Fig. 14: Examples of sources of big scoring discrepancy between algorithm and pathologist. (a) out of focus; (b) tissue folds;
(c) heterogeneity and overlapping.
developed to add nuclei number estimation function for accurate assessment. It is also necessary to add the function of
automated detection and elimination of tissue-fold regions
before H-Score assessment.
[6]
[7]
5
C ONCLUDING R EMARKS
In this paper, we have developed a deep learning framework for automatic end-to-end H-Score assessment for
breast cancer TMAs. Experimental results show that automatic assessment for TMA H-Score is feasible. The H-Scores
predicted by our model have a high correlation with HScores given by experienced pathologists. We show that
the discrepancies between our deep learning model and the
pathologits are on par with those between the pathologists.
We have identified image out of focus, tissue fold and overlapping nuclei as the three major sources of error. We also
found that the major discrepancies between pathologists
and machine predictions occurred in images that will have
a high H-Score value. These findings have suggested future
research directions for improving accuracy.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
E. Rakha, D. Soria, A. R. Green, C. Lemetre, D. G. Powe, C. C.
Nolan, J. M. Garibaldi, G. Ball, and I. O. Ellis, “Nottingham
prognostic index plus (npi+): a modern clinical decision making
tool in breast cancer,” British journal of cancer, vol. 110, no. 7, pp.
1688–1697, 2014.
C. M. Perou, S. S. Jeffrey, M. Van De Rijn, C. A. Rees, M. B.
Eisen, D. T. Ross, A. Pergamenschikov, C. F. Williams, S. X. Zhu,
J. C. Lee et al., “Distinctive gene expression patterns in human
mammary epithelial cells and breast cancers,” Proceedings of the
National Academy of Sciences, vol. 96, no. 16, pp. 9212–9217, 1999.
T. O. Nielsen, J. S. Parker, S. Leung, D. Voduc, M. Ebbert, T. Vickery, S. R. Davies, J. Snider, I. J. Stijleman, J. Reed et al., “A comparison of pam50 intrinsic subtyping with immunohistochemistry and
clinical prognostic factors in tamoxifen-treated estrogen receptor–
positive breast cancer,” Clinical cancer research, pp. 1078–0432, 2010.
D. Soria, J. M. Garibaldi, F. Ambrogi, A. R. Green, D. Powe,
E. Rakha, R. D. Macmillan, R. W. Blamey, G. Ball, P. J. Lisboa et al.,
“A methodology to identify consensus classes from clustering algorithms applied to immunohistochemical data from breast cancer
patients,” Computers in biology and medicine, vol. 40, no. 3, pp. 318–
330, 2010.
A. Green, D. Powe, E. Rakha, D. Soria, C. Lemetre, C. Nolan,
F. Barros, R. Macmillan, J. Garibaldi, G. Ball et al., “Identification
of key clinical phenotypes of breast cancer using a reduced panel
of protein biomarkers,” British journal of cancer, vol. 109, no. 7, p.
1886, 2013.
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
K. McCarty Jr, L. Miller, E. Cox, J. Konrath, and K. McCarty Sr,
“Estrogen receptor analyses. correlation of biochemical and immunohistochemical methods using monoclonal antireceptor antibodies.” Archives of pathology & laboratory medicine, vol. 109, no. 8,
pp. 716–721, 1985.
H. Goulding, S. Pinder, P. Cannon, D. Pearson, R. Nicholson,
D. Snead, J. Bell, C. Elston, J. Robertson, R. Blamey et al., “A new
immunohistochemical antibody for the assessment of estrogen
receptor status on routine formalin-fixed tissue samples,” Human
pathology, vol. 26, no. 3, pp. 291–294, 1995.
A. C. Ruifrok, “Quantification of immunohistochemical staining
by color translation and automated thresholding.” Analytical and
quantitative cytology and histology/the International Academy of Cytology [and] American Society of Cytology, vol. 19, no. 2, pp. 107–113,
1997.
A. C. Ruifrok, D. A. Johnston et al., “Quantification of histochemical staining by color deconvolution,” Analytical and quantitative
cytology and histology, vol. 23, no. 4, pp. 291–299, 2001.
H. Irshad, A. Veillard, L. Roux, and D. Racoceanu, “Methods
for nuclei detection, segmentation, and classification in digital
histopathology: a review current status and future potential,” IEEE
reviews in biomedical engineering, vol. 7, pp. 97–114, 2014.
S. Kothari, J. H. Phan, T. H. Stokes, and M. D. Wang, “Pathology
imaging informatics for quantitative analysis of whole-slide images,” Journal of the American Medical Informatics Association, vol. 20,
no. 6, pp. 1099–1108, 2013.
N.-A. Pham, A. Morrison, J. Schwock, S. Aviel-Ronen, V. Iakovlev,
M.-S. Tsao, J. Ho, and D. W. Hedley, “Quantitative image analysis of immunohistochemical stains using a cmyk color model,”
Diagnostic pathology, vol. 2, no. 1, p. 8, 2007.
A. Yao, J. Gall, C. Leistner, and L. Van Gool, “Interactive object
detection,” in Computer Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on. IEEE, 2012, pp. 3242–3249.
J. Shu, H. Fu, G. Qiu, P. Kaye, and M. Ilyas, “Segmenting overlapping cell nuclei in digital histopathology images,” in Engineering in
Medicine and Biology Society (EMBC), 2013 35th Annual International
Conference of the IEEE. IEEE, 2013, pp. 5445–5448.
A. Basavanhally, E. Yu, J. Xu, S. Ganesan, M. Feldman,
J. Tomaszewski, and A. Madabhushi, “Incorporating domain
knowledge for tubule detection in breast histopathology using
ocallaghan neighborhoods,” in SPIE Medical Imaging, vol. 7963,
no. 1. International Society for Optics and Photonics, 2011, p.
796310.
H. Fu, G. Qiu, J. Shu, and M. Ilyas, “A novel polar space random field model for the detection of glandular structures,” IEEE
transactions on medical imaging, vol. 33, no. 3, pp. 764–776, 2014.
Y. Li, L. Shen, and S. Yu, “Hep-2 specimen image segmentation
and classification using very deep fully convolutional network,”
IEEE Transactions on Medical Imaging, 2017.
S. Albarqouni, C. Baur, F. Achilles, V. Belagiannis, S. Demirci,
and N. Navab, “Aggnet: deep learning from crowds for mitosis
detection in breast cancer histology images,” IEEE transactions on
medical imaging, vol. 35, no. 5, pp. 1313–1321, 2016.
Y. Liu, K. Gadepalli, M. Norouzi, G. E. Dahl, T. Kohlberger,
A. Boyko, S. Venugopalan, A. Timofeev, P. Q. Nelson, G. S. Corrado et al., “Detecting cancer metastases on gigapixel pathology
images,” arXiv preprint arXiv:1703.02442, 2017.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[20] D. Wang, A. Khosla, R. Gargeya, H. Irshad, and A. H. Beck, “Deep
learning for identifying metastatic breast cancer,” arXiv preprint
arXiv:1606.05718, 2016.
[21] M. Shah, C. Rubadue, D. Suster, and D. Wang, “Deep learning
assessment of tumor proliferation in breast cancer histological
images,” arXiv preprint arXiv:1610.03467, 2016.
[22] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional
networks for biomedical image segmentation,” in International
Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241.
[23] M. H. Yap, G. Pons, J. Martı́, S. Ganau, M. Sentı́s, R. Zwiggelaar,
A. K. Davison, and R. Martı́, “Automated breast ultrasound lesions detection using convolutional neural networks,” IEEE journal of biomedical and health informatics, 2017.
[24] P. Moeskops, J. M. Wolterink, B. H. van der Velden, K. G. Gilhuijs,
T. Leiner, M. A. Viergever, and I. Išgum, “Deep learning for
multi-task medical image segmentation in multiple modalities,” in
International Conference on Medical Image Computing and ComputerAssisted Intervention. Springer, 2016, pp. 478–486.
[25] G. Amit, O. Hadad, S. Alpert, T. Tlusty, Y. Gur, R. Ben-Ari, and
S. Hashoul, “Hybrid mass detection in breast mri combining
unsupervised saliency analysis and deep learning,” in International
Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2017, pp. 594–602.
[26] H. Masmoudi, S. M. Hewitt, N. Petrick, K. J. Myers, and M. A.
Gavrielides, “Automated quantitative assessment of her-2/neu
immunohistochemical expression in breast cancer,” IEEE transactions on medical imaging, vol. 28, no. 6, pp. 916–925, 2009.
[27] B. H. Hall, M. Ianosi-Irimie, P. Javidian, W. Chen, S. Ganesan,
and D. J. Foran, “Computer-assisted assessment of the human
epidermal growth factor receptor 2 immunohistochemical assay
in imaged histologic sections using a membrane isolation algorithm and quantitative analysis of positive controls,” BMC Medical
Imaging, vol. 8, no. 1, p. 11, 2008.
[28] N. Trahearn, Y. W. Tsang, I. A. Cree, D. Snead, D. Epstein, and
N. Rajpoot, “Simultaneous automatic scoring and co-registration
of hormone receptors in tumor areas in whole slide images of
breast cancer tissue slides,” Cytometry Part A, 2016.
[29] Z. Xinliang, Y. Jiawen, Z. Feiyun, and H. Junzhou, “Wsisa: Making
survival prediction from whole slide pathology images,” in CVPR,
2017.
[30] E. M. Brey, Z. Lalani, C. Johnston, M. Wong, L. V. McIntire,
P. J. Duke, and C. W. Patrick Jr, “Automated selection of dablabeled tissue for immunohistochemical quantification,” Journal of
Histochemistry & Cytochemistry, vol. 51, no. 5, pp. 575–584, 2003.
[31] P. Haub and T. Meckel, “A model based survey of colour deconvolution in diagnostic brightfield microscopy: Error estimation and
spectral consideration,” Scientific reports, vol. 5, 2015.
[32] J. Liu, G. Qiu, and L. Shen, “Luminance adaptive biomarker
detection in digital pathology images,” Procedia Computer Science,
vol. 90, pp. 113–118, 2016.
[33] I. REC, “Bt. 601-5: Studio encoding parameters of digital television
for standard 4: 3 and wide-screen 16: 9 aspect ratios,” 1995.
[34] P. Hobson, B. C. Lovell, G. Percannella, A. Saggese, M. Vento, and
A. Wiliem, “Hep-2 staining pattern recognition at cell and specimen levels: datasets, algorithms and results,” Pattern Recognition
Letters, vol. 82, pp. 12–22, 2016.
[35] K. Sirinukunwattana, S. E. A. Raza, Y.-W. Tsang, D. R. Snead,
I. A. Cree, and N. M. Rajpoot, “Locality sensitive deep learning
for detection and classification of nuclei in routine colon cancer
histology images,” IEEE transactions on medical imaging, vol. 35,
no. 5, pp. 1196–1206, 2016.
[36] H. Chen, X. Qi, L. Yu, and P.-A. Heng, “Dcan: Deep contour-aware
networks for accurate gland segmentation,” in Proceedings of the
IEEE conference on Computer Vision and Pattern Recognition, 2016,
pp. 2487–2496.
[37] X. Li, Z. Liu, P. Luo, C. C. Loy, and X. Tang, “Not all pixels
are equal: Difficulty-aware semantic segmentation via deep layer
cascade,” 2017.
[38] B.-B. Gao, C. Xing, C.-W. Xie, J. Wu, and X. Geng, “Deep label
distribution learning with label ambiguity,” IEEE Transactions on
Image Processing, 2017.
[39] Z.-H. Zhou, M.-L. Zhang, S.-J. Huang, and Y.-F. Li, “Multi-instance
multi-label learning,” Artificial Intelligence, vol. 176, no. 1, pp. 2291–
2320, 2012.
[40] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
12
[41] C. A. Schneider, W. S. Rasband, and K. W. Eliceiri, “Nih image to
imagej: 25 years of image analysis,” Nature methods, vol. 9, no. 7,
pp. 671–675, 2012.
[42] R. L. Wasserstein and N. A. Lazar, “The asa’s statement on pvalues: context, process, and purpose,” 2016.
[43] S. Kothari, J. H. Phan, and M. D. Wang, “Eliminating tissue-fold
artifacts in histopathological whole-slide images for improved
image-based prediction of cancer grade,” Journal of pathology informatics, vol. 4, 2013.
| 1cs.CV
|
1
Construction of Capacity-Achieving Lattice
Codes: Polar Lattices
arXiv:1411.0187v4 [cs.IT] 25 Dec 2017
Yanfei Yan, Ling Liu, Cong Ling, Member, IEEE and Xiaofu Wu, Member, IEEE
Abstract
In this paper, we propose a new class of lattices constructed from polar codes, namely polar lattices, to achieve the
capacity
1
2
log(1 + SNR) of the additive white Gaussian-noise (AWGN) channel for any signal-to-noise ratio (SNR).
Our construction follows the multilevel approach of Forney et al., where we construct a capacity-achieving polar code
on each level. The component polar codes are shown to be naturally nested, thereby fulfilling the requirement of the
multilevel lattice construction. We prove that polar lattices are AWGN-good, in the sense that the error probability (for
infinite lattice decoding) vanishes for any fixed volume-to-noise ratio (VNR) greater than 2πe. Furthermore, using
the technique of source polarization, we propose discrete Gaussian shaping over the polar lattice to satisfy the power
constraint. The proposed polar lattices permit low-complexity multistage successive cancellation decoding. Both the
construction and shaping are explicit, and the overall complexity of encoding and decoding is O(N log N ) for any
fixed target error probability.
Index Terms
AWGN-good lattices, discrete Gaussian shaping, lattice codes, multilevel construction, polar codes.
I. I NTRODUCTION
A fast-decodable, structured code achieving the capacity of the power-constrained additive white Gaussian-noise
(AWGN) channel is the dream goal of coding theory. Polar codes, proposed by Arıkan in [1], can provably achieve
the capacity of binary memoryless symmetric (BMS) channels. There are considerable efforts to extend polar codes
to general discrete memoryless channels, to nonbinary polar codes, and to asymmetric channels [2]–[7]. A largely
theoretical attempt to construct polar codes for the AWGN channel was given in [8], [9], based on nonbinary polar
codes or on the technique for the multi-access channel. However, it is still an open problem to construct practical
polar codes to achieve the capacity of the AWGN channel. In this paper, we propose polar lattices to fulfil this
goal, based on a combination of binary polar codes and lattice codes.
This work was presented in part at the IEEE Inform. Theory Workshop (ITW) 2012, Laussane, Switzerland, September, 2012, and in part
at the IEEE Int. Symp. Inform. Theory (ISIT), Istanbul, Turkey, July, 2013. The work of Yanfei Yan and Ling Liu is supported by the China
Scholarship Council. The work of Xiaofu Wu is supported by the National Science Foundation of China.
Yanfei Yan, Ling Liu and Cong Ling are with the Department of Electrical and Electronic Engineering, Imperial College London, London,
UK (e-mails: y.yan10, [email protected], [email protected]).
Xiaofu Wu is with the Nanjing University of Posts and Telecommunications, Nanjing 210003, China (e-mail: [email protected]).
December 27, 2017
DRAFT
2
Lattice codes are the counterpart of linear codes in the Euclidean space. The existence of lattice codes achieving
the Gaussian channel capacity has been established using the random coding argument [10], [11]. The rich structures
of lattice codes represent a significant advantage in multiterminal communications, such as information-theoretical
security [12], compute-and-forward [13], and distributed source coding [14] (see [15] for an overview). It is well
known that the design of a lattice code consists of two essentially separate problems: AWGN coding and shaping.
AWGN coding is addressed by the notion of AWGN-good lattices [10], [16]. Informally, AWGN-goodness means
that if the fundamental volume of the lattice is slightly greater than that of the “noise sphere”, the error probability
of infinite lattice decoding could be made arbitrarily small. Recently, several new lattice constructions with good
performance have been introduced [17]–[20]. On the other hand, shaping takes care of the finite power constraint
of the Gaussian channel. Capacity-achieving shaping techniques include Voronoi shaping [10] and lattice Gaussian
shaping [11], [21], [22]. Despite these significant progresses, an explicit construction of lattice codes achieving the
capacity of the Gaussian channel is still open (since this paper was submitted, we have become aware of the work
[23] which shows LDA lattices achieve capacity when the signal-to-noise ratio (SNR) > 1).
A. Contributions
In this paper, we settle this open problem by employing the powerful tool of polarization in lattice construction.
The novel technical contribution of this paper is two-fold:
•
The construction of polar lattices and the proof of their AWGN-goodness. We follow the multilevel construction
of Forney, Trott and Chung [24], where for each level we build a polar code to achieve its capacity. A salient
feature of the proposed method is that it naturally leads to a set of nested polar codes, as required by the
multilevel construction. This compares favorably with existing multilevel constructions [17], where extra efforts
are needed to nest the component codes.
•
The Gaussian shaping technique for polar lattices in the power-constrained AWGN channel. This is based
on source polarization. We are able to achieve the capacity
1
2
log(1 + SNR) with low-complexity multistage
successive cancellation (SC) decoding for any given SNR. It is worth mentioning that our proposed shaping
scheme is not only a practical implementation of lattice Gaussian shaping, but also an improvement in the
sense that we successfully remove the restriction SNR > e in [11, Theorem 3].
Both source and channel polarization are employed in the construction, resulting in an integrated approach in the
sense that error correction and shaping are performed by one single polar code on each level. Further, it is worth
pointing out that each aspect may also be of independent interest. AWGN-good lattices have many applications in
network information theory (e.g., the aforementioned compute-and-forward and Wyner-Ziv coding), while lattice
Gaussian shaping, i.e., generating a Gaussian distribution over a lattice, is useful in lattice-based cryptography as
well [25]. Both theoretical and practical aspects of polar lattices are addressed in this paper. We not only prove the
theoretical goodness of polar lattices, but also give practical rules for designing these lattices.
December 27, 2017
DRAFT
3
B. Relation to Prior Works
This paper is built on the basis of our prior attempt to build lattices from polar codes [26], [27], and significantly
extends it by employing Gaussian shaping. We are aware of the contemporary and independent work on polar-coded
modulation [28], which follows the multilevel coding approach of [29]. It is known that Forney et al.’s multilevel
construction is closely related to multilevel coding [24], [29]. The main conceptual difference between lattice coding
and coded modulation is that lattices are infinite and linear in the Euclidean space. The linear structure of lattices
is much desired in many emerging applications, e.g., in network information theory for the purpose of coordination
[13], [14].
This paper may be viewed as an explicit construction of the lattice Gaussian coding scheme proposed in [11],
where it was shown that Gaussian shaping over an AWGN-good lattice is capacity-achieving. Our approach is
different from the standard Voronoi shaping which involves a quantization-good lattice [10]. The proposed Gaussian
shaping does not require such a quantization-good lattice any more.
The sparse superposition code [30], [31] also achieves the Gaussian channel capacity with polynomial complexity.
However, its decoding complexity is considerably higher than that of the polar lattice; moreover, it requires a
random dictionary shared by the encoder and decoder, which incurs substantial storage complexity. In comparison,
the construction of polar lattices is as explicit as that of polar codes themselves, and the complexity is quasilinear:
O(N log2 N ) for a sub-exponentially vanishing error probability and O(N log N ) for a fixed error probability,
respectively.
Following the multilevel approach, it is also possible to obtain a low-complexity capacity-achieving code by
modifying the work of [8], [9]. However, to the best of our knowledge, this has not been reported in literature; the
resultant code would not possess the many useful structures of a lattice code.
C. Organization and Notation
The rest of this paper is organized as follows. Section II presents the background of lattice codes. In Section
III, we construct polar latices based on Forney et al.’s approach and prove their AWGN-goodness. In Section IV,
we propose Gaussian shaping over the polar lattice to achieve the capacity. Section V gives design examples and
simulation results.
All random variables (RVs) will be denoted by capital letters. Let PX denote the probability distribution of a
RV X taking values x in a set X and let H(X) denote its entropy. For multilevel coding, we denote by Xℓ a RV
X at level ℓ. The i-th realization of Xℓ is denoted by xiℓ . We also use the notation xi:j
ℓ as a shorthand for a vector
(xiℓ , ..., xjℓ ), which is a realization of RVs Xℓi:j = (Xℓi , ..., Xℓj ). Similarly, xiℓ: will denote the realization of the i-th
i
RV from level ℓ to level , i.e., of Xℓ:
= (Xℓi , ..., Xi ). For a set I, I c denotes its complement, and |I| represents
its cardinality. For an integer N , [N ] denotes the set of all integers from 1 to N . Following the notation of [1],
we denote N independent uses of channel W by W N . By channel combining and splitting, we get the combined
(i)
channel WN and the i-th subchannel WN . 1(·) denotes an indicator function. Throughout this paper, we use the
binary logarithm, denoted by log, and information is measured in bits.
December 27, 2017
DRAFT
4
II. BACKGROUND
ON
L ATTICE C ODING
A. Definitions
A lattice is a discrete subgroup of Rn which can be described by
Λ = {λ = Bx : x ∈ Zn },
where we assume the generator matrix B has full rank.
For a vector x ∈ Rn , the nearest-neighbor quantizer associated with Λ is QΛ (x) = arg minλ∈Λ kλ − xk, where
ties are resolved arbitrarily. We define the modulo lattice operation by x mod Λ , x − QΛ (x). The Voronoi region
of Λ, defined by V(Λ) = {x : QΛ (x) = 0}, specifies the nearest-neighbor decoding region. The Voronoi cell is
one example of fundamental region of the lattice. A measurable set R(Λ) ⊂ Rn is a fundamental region of the
lattice Λ if ∪λ∈Λ (R(Λ) + λ) = Rn and if (R(Λ) + λ) ∩ (R(Λ) + λ′ ) = ∅ for any λ 6= λ′ in Λ. The volume
of a fundamental region is equal to that of the Voronoi region V(Λ), which is given by V (Λ) = |det(B)|. More
generally, the mod-R(Λ) operation is defined by x 7→ x̌ where x̌ is the unique element of R(Λ) such that x̌−x ∈ Λ.
Obviously, the usual mod-Λ operation corresponds to the case where R(Λ) = V(Λ).
The theta series of Λ (see, e.g., [32, p.70]) is defined as
ΘΛ (τ ) =
X
2
e−πτ kλk ,
τ > 0.
λ∈Λ
In this paper, we are mostly concerned with the block error probability Pe (Λ, σ 2 ) of lattice decoding. It is the
probability P{x ∈
/ V(Λ)} that an n-dimensional independent and identically distributed (i.i.d.) Gaussian noise vector
x with zero mean and variance σ 2 per dimension falls outside the Voronoi region V(Λ). For an n-dimensional lattice
Λ, define the VNR by
2
V (Λ) n
.
γΛ (σ) ,
σ2
Then we introduce the notion of lattices which are good for the AWGN channel without power constraint.
Definition 1 (AWGN-good lattices): A sequence of lattices Λ(n) of increasing dimension n is AWGN-good if,
for any fixed Pe (Λ(n) , σ 2 ) ∈ (0, 1),
lim γΛ(n) (σ) = 2πe.
n→∞
and if, for a fixed VNR greater than 2πe, Pe (Λ(n) , σ 2 ) goes to 0 as n → ∞.
It is worth mentioning here that we do not insist on exponentially vanishing error probabilities, unlike Poltyrev’s
original treatment of good lattices for coding over the AWGN channel [16]. This is because a sub-exponential or
polynomial decay of the error probability is often good enough.
B. Flatness Factor and Lattice Gaussian Distribution
For σ > 0 and c ∈ Rn , the Gaussian distribution of mean c and variance σ 2 is defined as
kx−ck2
1
e− 2σ2 ,
fσ,c (x) = √
( 2πσ)n
December 27, 2017
DRAFT
5
for all x ∈ Rn . For convenience, let fσ (x) = fσ,0 (x).
Given lattice Λ, we define the Λ-periodic function
fσ,Λ (x) =
for x ∈ Rn .
X − kx−λk2
1
e 2σ2 ,
fσ,λ (x) = √
n
(
2πσ)
λ∈Λ
λ∈Λ
X
Note that fσ,Λ (x) is a probability density if x is restricted to a fundamental region R(Λ). It is actually the
probability density function (PDF) of the Λ-aliased Gaussian noise, i.e., the Gaussian noise after the mod-R(Λ)
operation [24]. When σ is small, the effect of aliasing becomes insignificant and the Λ-aliased Gaussian density
fσ,Λ (x) approaches a Gaussian distribution. When σ is large, fσ,Λ (x) approaches a uniform distribution.
This phenomenon is characterized by the flatness factor, which is defined for a lattice Λ as [12]
ǫΛ (σ) , max |V (Λ)fσ,Λ (x) − 1| .
x∈R(Λ)
It can be interpreted as the maximum variation of fσ,Λ (x) from the uniform distribution over R(Λ). The flatness
factor can be calculated using the theta series [12]:
n
1
γΛ (σ) 2
ǫΛ (σ) =
ΘΛ
− 1.
2π
2πσ 2
We define the discrete Gaussian distribution over Λ centered at c ∈ Rn as the following discrete distribution
taking values in λ ∈ Λ:
DΛ,σ,c (λ) =
where fσ,c (Λ) ,
P
λ∈Λ
fσ,c (λ)
,
fσ,c (Λ)
∀λ ∈ Λ,
fσ,c (λ) = fσ,Λ (c). Again for convenience, we write DΛ,σ = DΛ,σ,0 . Fig. 1 illustrates the
discrete Gaussian distribution over Z2 . As can be seen, it resembles a continuous Gaussian distribution, but is only
defined over a lattice. In fact, discrete and continuous Gaussian distributions share similar properties, if the flatness
factor is small. The discrete Gaussian distribution can also be sampled from a shifted lattice Λ − c. Note the relation
DΛ−c,σ (λ − c) = DΛ,σ,c (λ), namely, they are a shifted version of each other.
The following duality relation holds: the Fourier transform of the Λ-aliased Gaussian distribution fσ,Λ (x) is a
discrete Gaussian distribution on the dual lattice Λ∗ [24]. In fact, this relation can be used to derive the flatness
factor [12].
If the flatness factor is negligible, the discrete Gaussian distribution over a lattice preserves the capacity of the
AWGN channel.
Theorem 1 (Mutual information of discrete Gaussian distribution [11]): Consider an AWGN channel Y = X +E
where the input constellation X has a discrete Gaussian distribution DΛ−c,σs for arbitrary c ∈ Rn , and where the
variance of the noise E is σ 2 . Let the average signal power be P so that SNR = P/σ 2 , and let σ̃ , √ σ2s σ
σs +σ2
Then, if ε = ǫΛ (σ̃) <
December 27, 2017
1
2
and
.
πεt
1−ǫt
≤ ε where
q
π
ǫΛ σs /
t ≥ 1/e
π−t ,
q
εt ,
π
(t−4 + 1)ǫ σ /
s
Λ
π−t , 0 < t < 1/e
DRAFT
6
0.015
DZ2,σ(λ)
0.01
0.005
0
10
5
10
5
0
0
−5
λ2
−5
−10
−10
λ1
Fig. 1. Discrete Gaussian distribution over Z2 .
the discrete Gaussian constellation results in mutual information
ID ≥
5ε
1
log (1 + SNR) −
2
n
(1)
per channel use.
The statement of Theorem 1 is non-asymptotical, i.e., it can hold even if n = 1. A lattice Λ or its coset Λ − c with
a discrete Gaussian distribution is referred to as a good constellation for the AWGN channel if ǫΛ (σ̃) is negligible
[11].
It is further proved in [11] that the channel capacity is achieved with Gaussian shaping over an AWGN-good
lattice and MMSE lattice decoding. To this aim, we use a codebook L − c, where L is an AWGN-good lattice
and c is a proper shift. The encoder maps the information bits to points in L − c, which obey the lattice Gaussian
distribution DL−c,σs . Since the lattice points are not equally probable a priori in the lattice Gaussian coding, we
apply maximum-a-posteriori (MAP) decoding. It is proved in [11] that MAP decoding is equivalent to MMSE
lattice decoding
x̂ = QL−c (αy)
where α =
σs2
σs2 +σ2
is asymptotically equal to the MMSE coefficient
(2)
P
P +σ2
and QL−c denotes the minimum
Euclidean-distance decoder for shifted lattice L − c.
C. Construction D
A sublattice Λ′ ⊂ Λ induces a partition (denoted by Λ/Λ′ ) of Λ into equivalence classes modulo Λ′ . The order
of the partition is denoted by |Λ/Λ′ |, which is equal to the number of cosets. If |Λ/Λ′ | = 2, we call this a binary
partition. Let Λ/Λ1 / · · · /Λr−1 /Λ′ for r ≥ 1 be an n-dimensional lattice partition chain. For each partition Λℓ−1 /Λℓ
(1 ≤ ℓ ≤ r with convention Λ0 = Λ and Λr = Λ′ ) a code Cℓ over Λℓ−1 /Λℓ selects a sequence of representatives
aℓ for the cosets of Λℓ . Consequently, if each partition is a binary partition, the codes Cℓ are binary codes.
December 27, 2017
DRAFT
7
Construction D1 requires a set of nested linear binary codes C1 ⊆ C2 · ·· ⊆ Cr [24]. Suppose Cℓ has block length
N and the number of information bits kℓ for 1 ≤ ℓ ≤ r. Choose a basis g1 , g2 , · · · , gN such that g1 , · · · gkℓ span
Cℓ . When n = 1, the lattice L admits the form [24]
)
( r
kℓ
X
X
i
r N
i
ℓ−1
uℓ gi + 2 Z | uℓ ∈ {0, 1}
L=
2
ℓ=1
(3)
i=1
where the addition is carried out in RN . The fundamental volume of a lattice obtained from this construction is
given by
V (L) = 2−N RC V (Λ′ )N ,
where RC =
Pr
ℓ=1
Rℓ =
1
N
Pr
ℓ=1 kℓ
denotes the sum rate of component codes. For convenience, we will often be
concerned with the one-dimensional lattice partition chain Z/2Z/ · · · /2r Z in this paper.
The following is an example of Construction D: Barnes-Wall lattices constructed from Reed-Muller codes [34]2 .
Reed-Muller codes RM(N, k, d) are a class of linear block codes over GF(2), where N is the length of the codeword,
k is the length of the information block and d is the minimum Hamming distance. Conventionally, Reed-Muller
codes are denoted by RM(r′ , m) (0 ≤ r′ ≤ m) with following relation between N , k and d:
m
m
′
+ ···+
, d = 2m−r .
N = 2m , k = 1 +
′
1
r
The m-th member of the family of Barnes-Wall lattices is a N = 2m dimensional complex lattice or 2N
dimensional real lattice. For example, the code formula of the 1024-dimensional Barnes-Wall lattice is:
BW1024 = RM(1, 10) + 2RM(3, 10) + · · · + 25 Z1024 .
III. C ONSTRUCTION
OF
(4)
P OLAR L ATTICES
As reviewed in the preceding section, achieving the channel capacity involves an AWGN-good lattice. Forney
et al. gave single and multilevel constructions of AWGN-good lattices in [24]. We now follow their multilevel
approach to construct polar lattices. Bear in mind that, in order to achieve the capacity of the AWGN channel with
the noise variance σ 2 , the concerned noise variance for the AWGN-good lattice is in fact σ̃ 2 (recall σ̃ , √ σ2s σ
σs +σ2
),
which is the variance of the equivalent noise after MMSE rescaling [11]. This methodology can also be justified
by the equivalence lemma in the next section (see Lemma 10).
A. Forney et al.’s Construction Revisited
A mod-Λ Gaussian channel is a Gaussian channel with an input in V(Λ) and with a mod-V(Λ) operator at the
receiver front end [24]. The capacity of the mod-Λ channel for noise variance σ 2 is
C(Λ, σ 2 ) = log V (Λ) − h(Λ, σ 2 ),
1 The
(5)
case of multi-dimensional lattice partition is also known as “Construction E”, for which the lattice L is of the dimension nL = nN .
In this paper, we refer to both one and multi-dimensional cases generally as Construction D. See also [32, Chap. 5] [33].
2 We
give the example of Barnes-Wall lattices as a benchmark particularly because of the connection between Reed-Muller codes and polar
codes [1]. The advantage of polar codes over Reed-Muller codes will translate into the advantage of polar lattices over Barnes-Wall lattices.
December 27, 2017
DRAFT
8
where h(Λ, σ 2 ) is the differential entropy of the Λ-aliased noise over V(Λ):
Z
fσ,Λ (x) log fσ,Λ (x)dx.
h(Λ, σ 2 ) = −
V(Λ)
Given lattice partition Λ/Λ′ , the Λ/Λ′ channel is a mod-Λ′ channel whose input is restricted to discrete lattice
points in (Λ + a) ∩ R(Λ′ ) for some translate a. The capacity of the Λ/Λ′ channel is given by [24]
C(Λ/Λ′ , σ 2 ) = C(Λ′ , σ 2 ) − C(Λ, σ 2 )
(6)
= h(Λ, σ 2 ) − h(Λ′ , σ 2 ) + log V (Λ′ )/V (Λ).
Further, if Λ/Λ1 / · · · /Λr−1 /Λ′ is a lattice partition chain, then
C(Λ/Λ′ , σ 2 ) = C(Λ/Λ1 , σ 2 ) + · · · + C(Λr−1 /Λ′ , σ 2 ).
(7)
The key idea of [24] is to use a good component code Cℓ to achieve the capacity C(Λℓ−1 /Λℓ , σ 2 ) for each level
ℓ = 1, 2, . . . , r in Construction D. For such a construction, the total decoding error probability with multistage
decoding is bounded by
Pe (L, σ 2 ) ≤
r
X
Pe (Cℓ , σ 2 ) + Pe ((Λ′ )N , σ 2 ).
(8)
ℓ=1
To achieve a vanishing error probability, i.e., to make Pe (L, σ 2 ) → 0, we need to choose the lattice Λ′ such that
Pe ((Λ′ )N , σ 2 ) → 0 and the codes Cℓ for the Λℓ−1 /Λℓ channels whose error probabilities also tend to zero.
Since V (L) = 2−N RC V (Λ′ )N , the logarithmic VNR of L satisfies
2
V (L) nN
γL (σ)
= log
log
2πe
2πeσ 2
2
2
2− n RC V (Λ′ ) n
2πeσ 2
2
2
= − RC + log V (Λ′ ) − log 2πeσ 2 .
n
n
= log
Define
where h(σ 2 ) =
n
2
2
ǫ1 = C(Λ, σ )
ǫ2 = h(σ 2 ) − h(Λ′ , σ 2 )
ǫ = C(Λ/Λ′ , σ 2 ) − R = Pr C(Λ /Λ , σ 2 ) − R ,
3
C
ℓ−1
ℓ
ℓ
ℓ=1
(9)
(10)
log 2πeσ 2 is the differential entropy of the Gaussian noise. We note that, ǫ1 ≥ 0 represents the
capacity of the mod-Λ1 channel, ǫ2 ≥ 0 (due to the data processing theorem) is the difference between the entropy
of the Gaussian noise and that of the mod-Λr Gaussian noise, and ǫ3 ≥ 0 is the total capacity loss of component
codes.
Then we have
log
γL (σ)
2πe
Since ǫ2 ≥ 0, we obtain the upper bound3
log
3 It
=
2
(ǫ1 − ǫ2 + ǫ3 ).
n
≤
γL (σ)
2πe
2
(ǫ1 + ǫ3 ).
n
(11)
was shown in [24] that ǫ2 ≈ πPe (Λ′ , σ2 ), which is negligible compared to the other two terms.
December 27, 2017
DRAFT
9
Since log
γL (σ)
2πe
= 0 represents the Poltyrev capacity, the right hand side of (11) gives an upper bound on the gap
to the Poltyrev capacity. The bound is equal to
6.02
n (ǫ1
+ ǫ3 ) decibels (dB), by conversion of the binary logarithm
into the base-10 logarithm.
To approach the Poltyrev capacity, we would like to have log
γL (σ)
2πe
→ 0 while Pe (L, σ 2 ) → 0. Thus, from
(11), we need that both ǫ1 and ǫ3 are negligible. In Appendix A, we prove the following lemma.
Lemma 1: The capacity of the mod-Λ channel is bounded by
C(Λ, σ 2 ) ≤ log (1 + ǫΛ (σ)) ≤ log(e) · ǫΛ (σ).
(12)
Thus, we have the following design criteria:
•
•
•
The top lattice Λ has a negligible flatness factor ǫΛ (σ).
The bottom lattice Λ′ has a small error probability Pe (Λ′ , σ 2 ).
Each component code Cℓ is a capacity-approaching code for the Λℓ−1 /Λℓ channel.
These conditions are essentially the same as those of Forney et al. [24], except that we impose a slightly stronger
condition on the top lattice. In [24], the top lattice satisfies C(Λ, σ 2 ) ≈ 0. The reason why we require negligible
ǫΛ (σ) is to achieve the capacity of the power-constrained Gaussian channel. This will become clear in the next
section.
Asymptotically, the error probability of a polar code of length N decreases as e−O(
√
N)
[35] and we may desire
the same for the error probability of a polar lattice. In (8), we can let Pe ((Λ′ )N , σ 2 ) decrease exponentially. The
next lemma shows that the first two criteria can be satisfied by r growing with log N (see Appendix B for a proof).
Lemma 2: Consider a partition chain Λ/Λ1 / · · · /Λr−1 /Λ′ . A number of levels r = Θ(log N ) is sufficient to
achieve ǫΛ (σ) = e−Θ(N ) and Pe (Λ′ , σ 2 ) = e−Θ(N ) .
This lemma is mostly of theoretical interest. In practical designs, if the target error probability is fixed, e.g.,
Pe (L, σ 2 ) = 10−5 , a small number of levels will suffice.
B. Polar Lattices
It is shown in [24] that the Λℓ−1 /Λℓ channel is symmetric, and that the optimum input distribution is uniform.
Since we use a binary partition Λℓ−1 /Λℓ , the input Xℓ is binary for ℓ ∈ 1, 2, . . . , r. Associate Xℓ with representative
aℓ of the coset in the quotient group Λℓ−1 /Λℓ . The fact that the Λℓ−1 /Λℓ channel is a BMS channel allows a polar
code to achieve its capacity.
Let Y denote the output of the AWGN channel. Given x1:ℓ−1 , let Aℓ (x1:ℓ ) denote the coset chosen by xℓ , i.e.,
Aℓ (x1:ℓ ) = a1 + · · · + aℓ + Λℓ . Assuming a uniform input distribution for all Xℓ , the conditional PDF of this
Λℓ−1 /Λℓ channel with input xℓ and output ȳℓ = y mod Λℓ is given by [29, (5)]
PȲℓ |Xℓ ,X1:ℓ−1 (ȳℓ |xℓ , x1:ℓ−1 ) =
=
fσ,Λℓ (ȳℓ − a1 − · · · − aℓ )
X
kȳℓ − ak2
1
√
exp −
.
2σ 2
2πσ
(13)
a∈Aℓ (x1:ℓ )
In [24], this conditional PDF is written in a somewhat different from. Namely, the conditional PDF is fσ,Λℓ (ȳℓ −
a−aℓ) with an offset a. Nevertheless, the two forms are equivalent because we can let the offset a = a1 +· · ·+aℓ−1 .
December 27, 2017
DRAFT
10
The regularity (symmetry) and capacity separability [24, Th. 4 and Th. 5] of the Λ/Λ′ channel hold for any offset
on its input. In fact, the offset due to previous input bits x1:ℓ−1 would be removed by the multistage decoder at
level ℓ, which means that the code for level ℓ can be designed according to (13) with x1:ℓ−1 = 0. For this reason,
we will fix x1:ℓ−1 = 0 to prove channel degradation in the following lemma. The reason why we use the form (13)
is for consistency with the case of non-uniform input x1:ℓ−1 in Sect. IV, where one cannot always let x1:ℓ−1 = 0.
Definition 2: (Degradation [36]): Consider two channels W1 : X → Y1 and W2 : X → Y2 . W1 is said to be
(stochastically) degraded with respect to W2 if there exists a distribution Q : Y2 → Y1 such that
X
W1 (y1 |x) =
W2 (y2 |x)Q(y1 |y2 ).
y2 ∈Y2
The proof of the following lemma4 is given in Appendix C.
Lemma 3: Consider a self-similar binary lattice partition chain Λ/Λ1 / · · · /Λr−1 /Λ′ , in which we have Λℓ = T ℓ Λ
for all ℓ, with T = αV for some scale factor α > 1 and orthogonal matrix V . Then, the Λℓ−1 /Λℓ channel is degraded
with respect to the Λℓ /Λℓ+1 channel for 1 ≤ ℓ ≤ r − 1.
Now, we recall some basics of polar codes. Let W̃ (y|x) be a BMS channel with input alphabet X = {0, 1}
and output alphabet Y ⊆ R. Polar codes are block codes of length N = 2m with input bits u1:N . Let I(W̃ )
be the capacity of W̃ . Given the rate R < I(W̃ ), the information bits are indexed by a set of RN rows of the
generator matrix GN = [ 11 01 ]
⊗m
, where ⊗ denotes the Kronecker product. This gives an N -dimensional channel
W̃N (y 1:N |u1:N ). The channel seen by each bit [1] is given by
X
1
(i)
W̃N (y 1:N |u1:N ).
W̃N (y 1:N , u1:i−1 |ui ) =
N −1
2
i+1:N
N −i
∈X
u
Arıkan proved that as N grows,
(i)
W̃N
approaches either an error-free channel or a completely noisy channel. The
set of completely noisy (resp. error-free) subchannels is called the frozen set F (resp. information set I). One sets
ui = 0 for i ∈ F and only sends information bits within I.
The rule of SC decoding is defined as
0 i ∈ F or
i
û =
1 otherwise.
(i)
W̃N (y 1:N , û1:i−1 |0)
(i)
W̃N (y 1:N , û1:i−1 |1)
≥ 1 when i ∈ I,
Definition 3 (Bhattacharyya Parameter for Symmetric Channel [1]): Given a BMS channel W̃ with transition
probability PY |X , the Bhattacharyya parameter Z̃ ∈ [0, 1] is defined as
Pp
PY |X (y|0)PY |X (y|1).
Z̃(W̃ ) ,
y
(i)
Let PB denote the block error probability of a binary polar code. PB can be upper-bounded as PB ≤ Σi∈I Z̃(W̃N ).
It was shown in [35], [37] that for any β < 12 ,
β
1
(i)
{i : Z(W̃N ) < 2−N }
m→∞ N
β
1
(i)
{i : I(W̃N ) > 1 − 2−N }
lim
m→∞ N
lim
4 This
=
I(W̃ )
=
I(W̃ ).
version of the lemma is suggested by an anonymous reviewer.
December 27, 2017
DRAFT
11
This means that the fraction of good channels is about I(W̃ ) as m → ∞. Therefore, constructing polar codes is
equivalent to choosing the good indices. However, the complexity of the exact computation for a BMS channel with
a continuous output alphabet appears to be exponential in the block length. A quantization method was proposed in
[38] which transforms a BMS channel with a continuous output alphabet to that with a finite output alphabet. Also,
[39] proposed an approximation method to construct polar codes efficiently over any discrete-output BMS channel
(13). We combine these two methods together in order to construct polar codes for the Λℓ−1 /Λℓ channel (see [27]
for more details). It was shown in [38], [39] that with a sufficient number of quantization levels, the approximation
error is negligible while the computational complexity is still O(N logN ).
With component polar codes P(N, kℓ ) for all the Λℓ−1 /Λℓ channels (1 ≤ ℓ ≤ r), we stack them as in Construction
D to build the polar lattice. The following lemma shows that these component codes are nested, which is to guarantee
that the multilevel construction creates a lattice [24]. We consider two rules to determine the component codes, for
theoretical and practical purposes, respectively. One is the capacity rule [24], [29], where we select the channel
indices according to a threshold on the mutual information. The other is the equal-error-probability rule [29],
namely, the same error probability for each level, where we select the channel indices according to a threshold on
the Bhattacharyya parameter. The advantage of the equal-error-probability rule based on the Bhattacharyya parameter
is that it leads to an upper bound on the error probability. For this reason, we use the equal-error-probability rule
in the practical design. It is well known that these two rules will converge as the block length goes to infinity [1].
This nesting relation is a consequence of [37, Lemma 4.7].
Lemma 4: For either the capacity rule or the equal-error-probability rule, the component polar codes built in the
multilevel construction are nested, i.e., P(N, k1 ) ⊆ P(N, k2 ) ⊆ · · · ⊆ P(N, kr ).
Proof. Firstly, consider the equal-error-probability rule. By [37, Lemma 4.7], if a BMS channel Ṽ is a degraded
(i)
(i)
(i)
(i)
version of W̃ , then the subchannel ṼN is also degraded with respect to W̃N and Z̃(ṼN ) ≥ Z̃(W̃N ). Let the
β
threshold be 2−N for some β < 1/2. The codewords are generated by x1:N = uI GI , where GI is the submatrix
of G whose rows are indexed by information set I. The information sets for these two channels are respectively
given by
IW̃
(i)
I
Ṽ
(i)
β
={i : Z̃(W̃N ) < 2−N },
(i)
β
={i : Z̃(ṼN ) < 2−N }.
(i)
Due to the fact that Z̃(ṼN ) ≥ Z̃(W̃N ), we have IṼ ⊆ IW̃ . If we construct polar codes P(N, |IW̃ |) over W̃ and
P(N, |IṼ |) over Ṽ , GIṼ is a submatrix of GIW̃ . Therefore P(N, |IṼ |) ⊆ P(N, |IW̃ |).
From Lemma 3, the channel of the ℓ-th level is always degraded with respect to the channel of the (ℓ + 1)-th
level, and consequently, P(N, kℓ ) ⊆ P(N, kℓ+1 ).
Then, consider the capacity rule. The nesting relation still holds if we select the channel indices according to a
(i)
(i)
threshold on the mutual information. This is because, by [37, Lemma 4.7], I(ṼN ) ≤ I(W̃N ) if a BMS channel
Ṽ is a degraded version of W̃ .
December 27, 2017
DRAFT
12
C. AWGN Goodness
For a threshold 2−N
β
of the Bhattacharyya parameter, the block error probability of the polar code with SC
β
decoding is upper-bounded by N 2−N . It can be made arbitrarily small by increasing the block length N . Also,
the capacity loss ǫ3 diminishes as N → ∞. Therefore, we have the following theorem:
Theorem 2: Suppose ǫΛ (σ) is negligible. Construct polar lattice L from the n-dimensional binary lattice partition
chain Λ/Λ1 / · · · /Λr−1 /Λ′ and r nested polar codes with block length N , where r = O(log N ). Then, the error
probability of L under multistage decoding is bounded by
2
Pe (L, σ ) ≤ rN 2
−N β
+N
1−
Z
!
fσ2 (x)dx ,
V(Λ′ )
(14)
with the logarithmic VNR bounded by (11). As N → ∞, L can achieve the Poltyrev capacity, i.e., L is AWGN-good
L (σ)
→ 0 for any fixed Pe (L, σ 2 ).
in the sense that log γ2πe
Remark 1: It is worth pointing out that Theorem 2 only requires mild conditions. The condition ǫΛ (σ) → 0 is
easily satisfied by properly scaling the top lattice Λ. In practice, if the target error probability is fixed (e.g., 10−5 ),
r can be a small constant, namely, r does not have to scale as log N . Thus, the essential condition is N → ∞.
For finite N , however, the capacity loss ǫ3 is not negligible. We investigate the finite-length performance of polar
lattices in the following.
The finite-length analysis of polar codes was given in [40]–[42]. It was proved that polar codes need a polynomial
1
block length with respect to the gap to capacity ǫloss = I(W̃ ) − R = O(N − µ ) [40], [41], where µ is known as
the scaling exponent. The lower bound of the gap is ǫloss ≥ βN
1
−µ
, where β is a constant that depends only on
1
I(W̃ ) and µ = 3.55 [40]. The upper bound of the gap is ǫloss ≤ β̄N − µ̄ , where β̄ is a constant that depends only
on the block error probability PB and µ̄ = 7 was given in [40]. Later this scaling factor µ̄ has been improved to
5.77 [42].
Thus, the gap to the Poltyrev capacity of finite-dimensional polar lattices is
1
2
γL (σ)
≤
ǫ1 + rβ̄N − µ̄
log
2πe
n
with the corresponding block error probability
Pe (L, σ 2 ) ≤ rPB + Pe (Λ′N , σ 2 ),
where the constant β̄ depends only on PB (assuming equal error probabilities for the component polar codes).
Since n ≪ N is fixed, the gap to the Poltyrev capacity of polar lattices also scales polynomially in the dimension
nL = nN .
In comparison, the optimal bound for finite-dimensional lattices is given by [43]
r
1
1
γL (σ)
2 −1
=
.
Q (Pe (L, σ 2 )) −
log nL + O
log
2πe opt
nL
nL
nL
(15)
At finite dimensions, this is more precise than the exponential error bound for lattices constructed from random
linear codes given in [24]. Thus, given Pe (L, σ 2 ), the scaling exponent of optimum random lattices is 2 which
is smaller than that of polar lattices µ̄. The result is consistent with the fact that polar codes require larger block
length than random codes to achieve the same rate and error probability.
December 27, 2017
DRAFT
13
IV. P OLARIZATION -BASED G AUSSIAN S HAPING
To achieve the capacity of the power-constrained Gaussian channel, we can apply Gaussian shaping over the
polar lattice L. However, it appears difficult to do so directly. In this section, we will apply Gaussian shaping to the
top lattice Λ instead, which is more friendly for implementation. This is motivated by Theorem 1, which implies
that one may construct a capacity-achieving lattice code from a good constellation. More precisely, one may choose
a low-dimensional top lattice such as Z and Z2 whose mutual information has a negligible gap to the channel
capacity as bounded in Theorem 1, and then construct a multilevel code to achieve the capacity. We will show that
this strategy is equivalent to implementing Gaussian shaping over the AWGN-good polar lattice. For this purpose,
we will employ the recently introduced polar codes for asymmetric channels [6], [7].
A. Asymmetric Channels in Multilevel Lattice Coding
By Theorem 1, we choose a good constellation DΛ,σs such that the flatness factor ǫΛ (σ̃) is negligible. Let
the binary partition chain Λ/Λ1 / · · · /Λr−1 /Λ′ / · · · be labelled by bits X1 , · · · , Xr , · · · . Then, DΛ,σs induces a
distribution PX1:r whose limit corresponds to DΛ,σs as r → ∞. An example for DZ,σs for σs = 3 is shown in
Fig. 2. In this case, a shaping constellation with M = 32 points are actually sufficient, since the total probability
of these points is rather close to 1.
0.14
DZ,σ =3
s
0.12
N(0,σs=3)
Probability
0.1
0.08
0.06
0.04
0.02
0
−20
−10
0
Z
10
20
(a) DZ,σs for σs = 3
(b) Bit labelling
Fig. 2. Lattice Gaussian distribution DZ,σs and the associated labelling. A probability P (X1 , X2 , ..., Xi ) in (b) is given by that of the coset
P
indexed by bits X1 , X2 , ..., Xi ; for example, P (X1 = 1, X2 = 0) = λ∈4Z+1 Pr{λ}.
By the chain rule of mutual information
I(Y ; X1:r ) =
r
X
ℓ=1
December 27, 2017
I(Y ; Xℓ |X1:ℓ−1 ),
(16)
DRAFT
14
we obtain r binary-input channels Wℓ for 1 ≤ ℓ ≤ r. Given x1:ℓ−1 , denote again by Aℓ (x1:ℓ ) the coset of Λℓ
indexed by x1:ℓ−1 and xℓ . According to [29, (5)], the channel transition PDF of the ℓ-th channel Wℓ is given by
PY |Xℓ ,X1:ℓ−1 (y|xℓ , x1:ℓ−1 )
=
1
P {Aℓ (x1:ℓ )}
X
a∈Aℓ (x1:ℓ )
1
=
fσs (Aℓ (x1:ℓ ))
= exp −
P (a)PY |A (y|a)
X
a∈Aℓ (x1:ℓ )
kyk2
2(σs2 + σ 2 )
1
ky − ak2
kak2
exp −
−
2πσσs
2σ 2
2σs2
(17)
σs2 + σ 2
2σs2 σ 2
σs2
y−a
σs2 + σ 2
a∈Aℓ (x1:ℓ )
X
kαy − ak2
1
1
kyk2
exp
−
.
= exp −
2(σs2 + σ 2 ) fσs (Aℓ (x1:ℓ )) 2πσσs
2σ̃ 2
1
1
fσs (Aℓ (x1:ℓ )) 2πσσs
X
exp −
!
2
a∈Aℓ (x1:ℓ )
where we recall the MMSE coefficient α =
distribution PXℓ |X1:ℓ−1 , unless
σs2
σs2 +σ2
, and σ̃ = √ σ2s σ
. In general, Wℓ is asymmetric with the input
σs +σ2
fσs (Aℓ (x1:ℓ ))/fσs (Aℓ−1 (x1:ℓ−1 )) ≈ 12
which means that ǫΛℓ (σs ) is negligible.
For a finite power, the number of levels does not need to be large. The following lemma determines how large
r should be in order to achieve the channel capacity. The proof can be found in Appendix D.
Lemma 5: If r = O(log log N ), the mutual information of the bottom level I(Y ; Xr |X1:r−1 ) → 0 as N → ∞.
P
Moreover, using the first r levels only incurs a capacity loss ℓ>r I(Y ; Xℓ |X1:ℓ−1 ) ≤ O( N1 ).
Remark 2: The condition r = O(log log N ) is again theoretical. In practice, r can be a small constant so that
the different between I(Y ; X1:r ) and capacity is negligible, as we will see from the example in the next section.
B. Polar Codes for Asymmetric Channels
Since the component channels are asymmetric, we need polar codes for asymmetric channels to achieve their
capacity. Fortunately, polar codes for the binary memoryless asymmetric (BMA) channels have been introduced in
[6], [7] recently.
Definition 4 (Bhattacharyya Parameter for BMA Channel [6], [44]): Let W be a BMA channel with input
X ∈ X = {0, 1} and output Y ∈ Y, and let PX and PY |X denote the input distribution and channel transition
probability, respectively. The Bhattacharyya parameter Z for channel W is the defined as
q
X
Z(X|Y ) = 2
PY (y) PX|Y (0|y)PX|Y (1|y)
y
=
Xq
PX,Y (0, y)PX,Y (1, y).
2
y
Note that Definition 4 is the same as Definition 3 when PX is uniform.
The following lemma shows that adding an observable at the output of W will not increase Z.
Lemma 6 (Conditioning reduces Bhattacharyya parameter Z): Let (X, Y, Y ′ ) ∼ PX,Y,Y ′ , X ∈ X = {0, 1}, Y ∈
Y, Y ′ ∈ Y ′ , we have
Z(X|Y, Y ′ ) ≤ Z(X|Y ).
December 27, 2017
DRAFT
15
Proof.
Z(X|Y, Y ′ )
Xq
PX,Y,Y ′ (0, y, y ′ )PX,Y,Y ′ (1, y, y ′ )
=
2
=
q
XXq
2
PX,Y,Y ′ (0, y, y ′ ) PX,Y,Y ′ (1, y, y ′ )
y,y ′
y
(a)
y′
X sX
sX
′ (0, y, y ′ )
≤
2
=
Xq
PX,Y (0, y)PX,Y (0, y)
2
y
y′
PX,Y,Y
PX,Y,Y ′ (1, y, y ′ )
y′
y
where (a) follows from Cauchy-Schwartz inequality.
Let X 1:N and Y 1:N be the input and output vector after N independent uses of W . For simplicity, denote the
distribution of (X i , Y i ) by PXY = PX PY |X for i ∈ [N ]. The following property of the polarized random variables
U 1:N = X 1:N GN is well known.
Theorem 3 (Polarization of Random Variables [6]): For any β ∈ (0, 0.5),
o
1 n
i
1:i−1
−N β
lim
= H(X),
i
:
Z(U
|U
)
≥
1
−
2
N →∞ N
o
1 n
i
1:i−1
−N β
lim
i
:
Z(U
|U
)
≤
2
= 1 − H(X),
N →∞ N
o
β
1 n
i : Z(U i |U 1:i−1 , Y 1:N ) ≥ 1 − 2−N
lim
= H(X|Y ),
N →∞ N
o
β
1 n
lim
= 1 − H(X|Y ),
i : Z(U i |U 1:i−1 , Y 1:N ) ≤ 2−N
N →∞ N
(18)
and
o
β
β
1 n
= I(X; Y ),
i : Z(U i |U 1:i−1 , Y 1:N ) ≤ 2−N and Z(U i |U 1:i−1 ) ≥ 1 − 2−N
N →∞ N
o
β
β
1 n
lim
= 1 − I(X; Y ).
i : Z(U i |U 1:i−1 , Y 1:N ) ≥ 2−N or Z(U i |U 1:i−1 ) ≤ 1 − 2−N
N →∞ N
lim
(19)
The Bhattacharyya parameter for asymmetric models was originally defined for distributed source coding in [44].
By the duality between channel coding and source coding, it can be also used to construct capacity-achieving polar
codes for BMA channels [6]. Actually, Z(U i |U 1:i−1 ) is the Bhattacharyya parameter for a single source X (without
side information).
The Bhattacharyya parameter of a BMA channel can be related to that of a symmetrized channel. To this aim,
we use a symmetrization technique which creates a binary-input symmetrized channel W̃ from the BMA channel
W . The following lemma was implicit in [6]; here we make it explicit.
Lemma 7 (Symmetrization): Let W̃ be a binary-input channel with input X̃ ∈ X = {0, 1} and output Ỹ ∈ {Y, X },
built from the asymmetric channel W as shown in Fig. 3. Suppose the input of W̃ is uniformly distributed, i.e.,
PX̃ (x̃ = 0) = PX̃ (x̃ = 1) = 12 . Then it holds for the symmetrized channel W̃ that PỸ |X̃ (y, x⊕ x̃|x̃) = PY,X (y, x). 5
5 Note
that the definition of a symmetrized channel is slightly different from that of a conventional symmetric channel [1], since a condition
on the input distribution is imposed here.
December 27, 2017
DRAFT
16
ܺ
ܺ෨
ܻ
ܹ
ْ
෩
ܹ
ܺ ْ ܺ෨
Fig. 3. The relationship between the asymmetric channel W and the symmetrized channel W̃ .
Proof.
PỸ |X̃ (y, x ⊕ x̃|x̃)
=
(a)
=
(b)
=
(c)
=
P
PỸ ,X̃ (y, x ⊕ x̃, x̃)
=
P
PX̃ (x̃)
P
PX̃ (x̃)
x′ ∈X
′
x′ ∈X
PỸ ,X,X̃ (y, x ⊕ x̃, x′ , x̃)
PX̃ (x̃)
PY |X (y|x )PX⊕X̃,X,X̃ (x ⊕ x̃, x′ , x̃)
′
′
′
x′ ∈X PY |X (y|x )PX⊕X̃|X,X̃ (x ⊕ x̃|x , x̃)PX (x )PX̃ (x̃)
PX̃ (x̃)
PY,X (y, x).
The equalities (a)-(c) follow from (a) Y is only dependent on X, (b) X and X̃ are independent to each other and
(c) PX⊕X̃|X,X̃ (x ⊕ x̃|x′ , x̃) = 1(x′ = x).
The following theorem connects the Bhattacharyya Parameter of a BMA channel W and that of the symmetrized
channel W̃ . Denote by WN and W̃N the combining channels of N uses of W and W̃ , respectively.
Theorem 4 (Connection Between Bhattacharyya Parameters [6]): Let X̃ 1:N and Ỹ 1:N = X 1:N ⊕ X̃ 1:N , Y 1:N
be the uniform input and output vectors of W̃ , respectively, and let U 1:N =X 1:N GN and Ũ 1:N =X̃ 1:N GN . The
Bhattacharyya parameter of each subchannel of WN is equal to that of each subchannel of W̃N , i.e.,
Z(U i |U 1:i−1 , Y 1:N ) = Z̃(Ũ i |Ũ 1:i−1 , X 1:N ⊕ X̃ 1:N , Y 1:N ).
Now, we are in a position to construct polar codes for the BMA channel. Define the frozen set F̃ and information
set Ĩ of the symmetric polar codes as follows:
β
frozen set: F̃ = {i ∈ [N ] : Z(U i |U 1:i−1 , Y 1:N ) > 2−N }
information set: Ĩ = {i ∈ [N ] : Z(U i |U 1:i−1 , Y 1:N ) ≤ 2−N β }.
(20)
By Theorem 4, the Bhattacharyya parameters of the symmetrized channel W̃ and the asymmetric channel W are
the same. However, the channel capacity of W̃ is I(X̃; X ⊕ X̃) + I(X̃; Y |X ⊕ X̃) = 1 − H(X) + I(X; Y ), which
is 1 − H(X) more than the capacity of W . To obtain the real capacity I(X; Y ) of W , the input distribution of
W needs to be adjusted to PX . By polar lossless source coding, the indices with very small Z(U i |U 1:i−1 ) should
be removed from the information set Ĩ of the symmetrized channel, and the proportion of this part is 1 − H(X)
as N → ∞. We name the remaining set as the information set I of the asymmetric channel W . Further, there are
some bits which are uniformly distributed and can be made independent from the information bits; we name this
set as the frozen set F . In order to generate the desired input distribution PX , the remaining bits are determined
December 27, 2017
DRAFT
17
Fig. 4. Polarization for symmetric and asymmetric channels.
from the bits in F ∪ I; we call it the shaping set S. This process is depicted in Fig. 4. We formally define the
three sets as follows:
β
frozen set: F = {i ∈ [N ] : Z(U i |U 1:i−1 , Y 1:N ) ≥ 1 − 2−N }
β
β
information set: I = {i ∈ [N ] : Z(U i |U 1:i−1 , Y 1:N ) ≤ 2−N and Z(U i |U 1:i−1 ) ≥ 1 − 2−N }
shaping set: S = (F ∪ I)c .
(21)
To find these sets, one can use Theorem 4 to calculate Z(U i |U 1:i−1 , Y 1:N ) with the known technique for
symmetric polar codes [38], [45]. We note that Z(U i |U 1:i−1 ) can be computed in a similar way: one constructs a
symmetrized channel between X̃ and X ⊕ X̃, which is actually a binary symmetric channel with cross probability
PX (x = 1). The above construction is equivalent to implementing shaping over the polar code for the symmetrized
channel W̃ .
Besides the construction, the decoding can also be converted to that of the symmetric polar code. If X 1:N ⊕
X̃ 1:N = 0, we have U 1:N = Ũ 1:N , which means the decoding result of U 1:N equals to that of Ũ 1:N . Thus, decoding
of the polar code for W can be treated as decoding of the polar code for W̃ given that X ⊕ X̃ = 0. Clearly, the SC
decoding complexity for asymmetric channel is also O(N log N ). We summarize this observation as the following
lemma.
Lemma 8 (Decoding for Asymmetric Channel [6]): Let y 1:N be a realization of Y 1:N and û1:i−1 be the previous
i − 1 estimates of u1:N . The likelihood ratio of ui is given by
(i)
(i)
PU i |U 1:i−1 ,Y 1:N (0|û1:i−1 , y 1:N )
W̃N ((y 1:N , 01:N ), û1:i−1 |0)
,
=
(i)
PU i |U 1:i−1 ,Y 1:N (1|û1:i−1 , y 1:N )
W̃N ((y 1:N , 01:N ), û1:i−1 |1)
(22)
where W̃N denotes the transition probability of the i-th subchannel of W̃N .
In [6], the bits in F ∪ S are all chosen according to PU i |U 1:i−1 (ui |u1:i−1 ), which can also be calculated using
(22) (treating Y as an independent variable and remove it). However, in order to be compatible with polar lattices,
we modify the scheme such that the bits in F are uniformly distributed over {0, 1} while the bits in S are still
December 27, 2017
DRAFT
18
chosen according to PU i |U 1:i−1 (ui |u1:i−1 ). The expectation of the decoding error probability still vanishes with
N . The following theorem is an extension of the result in [6, Theorem 3]. We give the proof in Appendix E for
completeness.
Theorem 5: Consider a polar code with the following encoding and decoding strategies for a BMA channel.
•
Encoding: Before sending the codeword x1:N = u1:N GN , the index set [N ] are divided into three parts:
the frozen set F , the information set I and the shaping set S which are defined in (21). The encoder places
uniformly distributed information bits in I, and fills F with a uniform random {0, 1} sequence which is shared
between the encoder and the decoder. The bits in S are generated by a mapping φS , {φi }i∈S in the family
•
of randomized mappings ΦS , which yields the following distribution:
0 with probability PU i |U 1:i−1 (0|u1:i−1 ),
i
u =
1 with probability PU i |U 1:i−1 (1|u1:i−1 ).
Decoding: The decoder receives y 1:N and estimates û1:N of u1:N according to the rule
ui ,
if i ∈ F
.
ûi = φi (û1:i−1 ),
if i ∈ S
argmax PU i |U 1:i−1 ,Y 1:N (u|û1:i−1 , y 1:N ), if i ∈ I.
u
With the above encoding and decoding, the message rate can be arbitrarily close to I(Y ; X) and the expectation of
the decoding error probability over the randomized mappings satisfies EΦS [Pe (φS )] ≤ N 2−N
β′
for β ′ < β < 0.5.
β′
Consequently, there exists a deterministic mapping φS such that Pe (φS ) ≤ N 2−N .
In practice, to share the mapping φS between the encoder and the decoder, we can let them have access to the
same source of randomness, e.g., using the same seed for the pseudorandom number generators.
C. Multilevel Polar Codes
Next, our task is to construct polar codes to achieve the mutual information I(Y ; Xℓ |X1:ℓ−1 ) for all levels.
The construction of the preceding subsection is readily applicable to the construction for the first level W1 . To
demonstrate the construction for other levels, we take the channel of the second level W2 as an example. This is
also a BMA channel with input X2 ∈ X = {0, 1}, output Y ∈ Y and side information X1 . Its channel transition
probability is shown in (17). To construct a polar code for the second level, we propose the following two-step
procedure.
Step 1: Construct a polar code for the BMS channel with input vector X̃21:N = [X̃21 , X̃22 , ···, X̃2N ] and output vector
Ỹ 1:N = X21:N ⊕ X̃21:N , Y 1:N , X11:N where X̃2i ∈ X = {0, 1} is uniformly distributed. At this step X1 is
P
regarded as the output. Then the distribution of X2 becomes the marginal distribution x1 ,x3:r PX1:r (x1:r ).
Consider polarized random variables U21:N = X21:N GN and Ũ21:N = X̃21:N GN . According to Theorem 3,
December 27, 2017
DRAFT
19
Fig. 5. The first step of polarization in the construction for the second level.
the polarization gives us the three sets F2 , I2′ and S2′ as shown in Fig. 5. Similarly, we can prove that
|I2′ |
N
→ I(Y, X1 ; X2 ) and
|F2 ∪S2′ |
N
→ 1 − I(Y, X1; X2 ) as N → ∞. These three sets are defined as follows:
β
frozen set: F2 = {i ∈ [N ] : Z(U2i |U21:i−1 , Y 1:N , X11:N ) ≥ 1 − 2−N }
β
β
information set: I2′ = {i ∈ [N ] : Z(U2i |U21:i−1 , Y 1:N , X11:N ) ≤ 2−N and Z(U2i |U21:i−1 ) ≥ 1 − 2−N } (23)
c
shaping set: S2′ = (F2 ∪ I2′ ) .
Step 2: Treat X11:N as the side information for the encoder. Given X11:N , the choices of X21:N are further restricted
since X1 and X2 are generally correlated, i.e., PX1 ,X2 (x1 , x2 ) = fσs (A(x1 , x2 ))/fσs (Λ) (cf. Fig. 2). By
removing from I2′ the bits which are almost deterministic given U21:i−1 and X11:N , we obtain the information
set I2 for W2 . Then the distribution of the input X2 becomes the conditional distribution PX2 |X1 (x2 |x1 ).
The process is shown in Fig. 6. More precisely, the indices are divided into three portions as follows:
1
=
1 − I(X̃2 ; X̃2 ⊕ X2 , X1 , Y ) +I(X̃2 ; X̃2 ⊕ X2 , X1 , Y )
{z
}
|
F2
Step1
=
1 − I(X̃2 ; X̃2 ⊕ X2 , X1 , Y ) + I(X̃2 ; X̃2 ⊕ X2 ) + I(X̃2 ; X1 , Y |X̃2 ⊕ X2 )
|
{z
} |
{z
} |
{z
}
S2′
F2
Step2
=
1 − I(X̃2 ; X̃2 ⊕ X2 , X1 , Y ) + I(X̃2 ; X̃2 ⊕ X2 ) + I(X̃2 ; X1 |X̃2 ⊕ X2 ) + I(X̃2 ; Y |X1 , X̃2 ⊕ X2 )
|
{z
} |
{z
} |
{z
} |
{z
}
S2′
F2
=
SX1
I2
1 − I(X̃2 ; X̃2 ⊕ X2 , X1 , Y ) + 1 − H(X2 ) + I(X2 ; X1 ) + I(X2 ; Y |X1 )
{z
} | {z } | {z } |
{z
}
|
S2′
F2
=
I2′
SX1
I2
1 − I(X̃2 ; X̃2 ⊕ X2 , X1 , Y ) + 1 − H(X2 |X1 ) + I(X2 ; Y |X1 )
|
{z
} |
{z
} |
{z
}
F2
S2
I2
We give the formal statement of this procedure in the following lemma.
Lemma 9: After the first step of polarization, we obtain the three sets F2 , I2′ and S2′ in (23). Let SX1 denote the
β
set of indices whose Bhattacharyya parameters satisfy Z(U2i |U21:i−1 , Y 1:N , X11:N ) ≤ 2−N , Z(U2i |U21:i−1 , X11:N ) ≤
β
β
1 − 2−N and Z(U2i |U21:i−1 ) ≥ 1 − 2−N . The proportion of SX1 is asymptotically given by limN →∞
December 27, 2017
|SX1 |
N
=
DRAFT
20
Fig. 6. The second step of polarization in the construction for the second level.
I(X2 ; X1 ). Then by removing SX1 from I2′ , we obtain the true information set I2 for W2 . Formally, the three sets
are obtained as follows:
β
1:i−1
i
, Y 1:N , X11:N ) ≥ 1 − 2−N }
frozen set: F2 = {i ∈ [N ] : Z(U2 |U2
β
β
information set: I2 = {i ∈ [N ] : Z(U2i |U21:i−1 , Y 1:N , X11:N ) ≤ 2−N and Z(U2i |U21:i−1 , X11:N ) ≥ 1 − 2−N }
c
shaping set: S2 = (F2 ∪ I2 ) .
(24)
Proof. Firstly, we show the proportion of set SX1 goes to I(X1 ; X2 ) as N → ∞. Here we define a slightly
β
β
′
different set SX
= {i ∈ [N ] : Z(U2i |U21:i−1 , X11:N ) ≤ 2−N and Z(U2i |U21:i−1 ) ≥ 1 − 2−N }. Suppose we are
1
constructing an asymmetric polar code for the channel from X1 to X2 ; it is not difficult to find that limN →∞
I(X2 ; X1 ) by Theorem 5. Furthermore, by Lemma 6, if
β
Z(U2i |U21:i−1 , X11:N )
≤2
−N β
′
|
|SX
1
N
=
, we can immediately have
′
Z(U2i |U21:i−1 , X11:N , Y 1:N ) ≤ 2−N . Therefore, the difference between the definitions of SX1 and SX
only lies
1
β
β
on Z(U2i |U21:i−1 , X11:N ). Denoting by P̄X1 the unpolarized set with 2−N ≤ Z(U2i |U21:i−1 , X11:N ) ≤ 1 − 2−N , we
have
′
|
|SX1 | |SX
|P̄X1 |
1
−
≤ lim
= 0.
N →∞ N
N →∞ N
N
lim
As a result, limN →∞
|SX1 |
N
= limN →∞
′
|
|SX
1
N
Secondly, we show that SX1 ∪ I2 =
β
I2′ .
(25)
= I(X2 ; X1 ).
β
Again, by Lemma 6, if Z(U2i |U21:i−1 , X11:N ) ≥ 1 − 2−N , we get
Z(U2i |U21:i−1 ) ≥ 1−2−N and the difference between the definitions of SX1 and I2′ only lies on Z(U2i |U21:i−1 , X11:N ).
Observe that the union SX1 ∪ I2 would remove the condition on Z(U2i |U21:i−1 , X11:N ), and accordingly we have
SX1 ∪ I2 = I2′ . It can be also found that the proportion of I2 goes to I(X2 ; Y |X1 ) as N → ∞.
We summarize our main results in the following theorem:
December 27, 2017
DRAFT
21
Theorem 6 (Coding Theorem for Multilevel Polar Codes): Consider a polar code with the following encoding and
decoding strategies for the channel of the second level W2 with the channel transition probability PY |X2 ,X1 (y|x2 , x1 )
shown in (17).
•
Encoding: Before sending the codeword x1:N
= u1:N
2
2 GN , the index set [N ] are divided into three parts:
the frozen set F2 , information set I2 , and shaping set S2 . The encoder first places uniformly distributed
information bits in I2 . Then the frozen set F2 is filled with a uniform random sequence which are shared
between the encoder and the decoder. The bits in S2 are generated by a mapping φS2 , {φi }i∈S2 form a
family of randomized mappings ΦS2 , which yields the following distribution:
0 with probability P i 1:i−1 1:N (0|u1:i−1 , x1:N ),
1
2
U2 |U2
,X1
ui2 =
1 with probability P i 1:i−1 1:N (1|u1:i−1 , x1:N ).
1
2
U |U
,X
2
•
Decoding: The decoder receives y
to the rule
1:N
and estimates
2
û21:N
(26)
1
based on the previously recovered x1:N
according
1
ui2 ,
if i ∈ F2
ûi2 = φi (û1:i−1
),
if i ∈ S2
2
1:N
argmax PU i |U 1:i−1 ,X 1:N ,Y 1:N (u|û1:i−1
, x1:N
), if i ∈ I2
1 ,y
2
u
2
2
.
1
1:N
Note that probability PU i |U 1:i−1 ,X 1:N ,Y 1:N (u|û1:i−1
) can be calculated by (22) efficiently, treating Y
, x1:N
1 ,y
2
2
2
1
and X1 (already decoded by the SC decoder at level 1) as the outputs of the asymmetric channel. With the
above encoding and decoding, the message rate can be arbitrarily close to I(Y ; X2 |X1 ) and the expectation of the
decoding error probability over the randomized mappings satisfies EΦS2 [Pe (φS2 )] ≤ N 2−N
β′
for any β ′ < β < 0.5.
β′
Consequently, there exists a deterministic mapping φS2 such that Pe (φS2 ) ≤ N 2−N .
The proof of this theorem is given in Appendix F.
Obviously, Theorem 6 can be generalized to the construction of a polar code for the channel of the ℓ-th level
1:N
. As a result, we can construct
Wℓ . The only difference is that the side information changes from X11:N to X1:ℓ−1
a polar code which achieves a rate arbitrarily close to I(Y ; Xℓ |X1:ℓ ) with vanishing error probability. We omit the
proof for the sake of brevity.
D. Achieving Channel Capacity
So far, we have constructed polar codes to achieve the capacity of the induced asymmetric channels for all levels.
Since the sum capacity of the component channels nearly equals the mutual information I(Y ; X), and since we
choose a good constellation such that I(Y ; X) ≈
1
2
log(1 + SNR), we have constructed a lattice code to achieve
the capacity of the Gaussian channel. We summarize the construction in the following theorem:
Theorem 7: Choose a good constellation with negligible flatness factor ǫΛ (σ̃) and negligible ǫt as in Theorem
1, and construct a multilevel polar code with r = O(log log N ) as above. Then, for any SNR, the message rate
approaches
1
2
log(1 + SNR), while the error probability under multistage decoding is bounded by
β′
Pe ≤ rN 2−N ,
December 27, 2017
0 < β ′ < 0.5
(27)
DRAFT
22
as N → ∞.
Remark 3: It is simple to generate a transmitted codeword of the proposed scheme. For n = 1, let
#
"
r
X
X
X
X
ℓ−1
i
i
i
χ=
2
u ℓ gi +
u ℓ gi +
u ℓ gi .
ℓ=1
i∈Iℓ
i∈Sℓ
(28)
i∈Fℓ
The transmitted codeword x is drawn from D2r ZN +χ , σs . From the proof of Lemma 5, we know that the probability
of choosing a point outside of the interval [−2r−1 , 2r−1 ] is negligible if r is sufficiently large, which implies there
exists only one point in this interval with probability close to 1. Therefore, one may simply transmit x = χ mod 2r ,
where the modulo operation is applied component-wise with range (−2r−1 , 2r−1 ].
Next, we show that such a multilevel polar coding scheme is equivalent to Gaussian shaping over a coset
L + c′ of a polar lattice L for some translate c′ . In fact, the polar lattice L is exactly constructed from the
corresponding symmetrized channels W̃ℓ . Recall that the ℓ-th channel Wℓ is a BMA channel with the input
distribution P (Xℓ |X1:ℓ−1 ) (1 ≤ ℓ ≤ r). It is clear that PX1:ℓ (x1:ℓ ) = fσs (Aℓ (x1:ℓ ))/fσs (Λ). By Lemma 7 and
(17), the transition probability of the symmetrized channel W̃ℓ is
PW̃ℓ ((y, x1:ℓ−1 , xℓ ⊕ x̃ℓ )|x̃ℓ ) = PY,X1:ℓ (y, x1:ℓ )
= PX1:ℓ (x1:ℓ )PY |Xℓ ,X1:ℓ−1 (y|xℓ , x1:ℓ−1 )
X
1
1
kyk2
= exp −
2
2
2(σs + σ ) fσs (Λ) 2πσσs
a∈Aℓ (x1:ℓ )
(29)
kαy − ak2
exp −
.
2σ̃ 2
Note that the difference between the asymmetric channel (17) and symmetrized channel (29) is the a priori
probability PX1:ℓ (x1:ℓ ) = fσs (Aℓ (x1:ℓ ))/fσs (Λ). Comparing with the Λℓ−1 /Λℓ channel (13), we see that the
symmetrized channel (29) is equivalent to a Λℓ−1 /Λℓ channel, since the common terms in front of the sum will
be completely cancelled out in the calculation of the likelihood ratio6 . We summarize the foregoing analysis in the
following lemma:
Lemma 10 (Equivalence lemma): Consider a multilevel lattice code constructed from constellation DΛ,σs for a
Gaussian channel with noise variance σ 2 . The ℓ-th symmetrized channel W̃ℓ (1 ≤ ℓ ≤ r) which is derived from
the asymmetric channel Wℓ is equivalent to the MMSE-scaled Λℓ−1 /Λℓ channel with noise variance σ̃ 2 .
Thus, the resultant polar codes for the symmetrized channels are nested, and the polar lattice is AWGN-good
for noise variance σ̃ 2 ; also, the multistage decoding is performed on the MMSE-scaled signal αy (cf. Lemma 8).
Since the frozen sets of the polar codes are filled with random bits (rather than all zeros), we actually obtain a
coset L + c′ of the polar lattice, where the shift c′ accounts for the effects of all random frozen bits. Finally, since
we start from DΛ,σs , we would obtain DΛN ,σs without coding; since L + c′ ⊂ ΛN by construction, we obtain a
discrete Gaussian distribution DL+c′ ,σs over L + c′ .
Remark 4: This analysis shows that our proposed scheme is an explicit construction of lattice Gaussian coding
introduced in [11], which applies Gaussian shaping to an AWGN-good lattice (or its coset). Note that the condition
6 Even
if y ∈ Rn in (29), the sum over Aℓ (x1:ℓ ) is Λℓ -periodic. Hence, the likelihood ratio will be the same if one takes ȳ = y mod Λℓ
and uses (13).
December 27, 2017
DRAFT
23
Fig. 7. Channel capacity for partition chain Z/2Z/ · · · /2r Z. The curve for the ℓ-th partition channel is a translate of the curve for the (ℓ − 1)-th
by 6 dB. The discrete BMS approximation is based on the method of [38], [39] with 64 quantization levels.
of negligible ǫΛ (σ̃) in Theorem 7 is the same as the condition on Λ imposed in the construction of polar lattice
in Section III (cf. Theorem 2). Again, it is always possible to scale down the top lattice Λ such that both ǫΛ (σ̃)
and ǫt become negligible in Theorem 1. Thus, Theorem 7 holds for any SNR, meaning that we have removed the
condition SNR > e required by [11, Theorem 3]7 . Moreover, if a good constellation of the form DΛ−c,σs for some
shift c is used in practice (e.g., a constellation taking values in {±1, ±3, . . .}), the proposed construction holds
verbatim.
Remark 5: By [11, Lemma 1], the power P of a discrete Gaussian distribution DL+c′ ,σs is never greater than
σs2 .
Remark 6: A shaping method was proposed in [11, Section IV], where Gaussian shaping is only performed on
the bottom lattice. However, it requires negligible ǫΛ′ (σs ), which does not hold in general.
V. D ESIGN E XAMPLES
In this section, we give design examples of polar lattices based on the one partition chain, with and without the
power constraint. The design follows the equal-error-probability rule. Multistage SC decoding is applied. Since the
complexity of SC decoding is O(N log N ), the overall decoding complexity is O(rN log N ).
A. Design Examples Without Power Constraint
Consider the one-dimensional lattice partition Z/2Z/ · · · /2r Z. To construct a multilevel lattice, one needs to
determine the number of levels of lattice partitions and the actual rates according to the the target error probability
7 The
reason of the condition SNR > e in [11] is that a more stringent condition is imposed on the flatness factor of L, namely, ǫL
√
2
σs
2 +σ 2
σs
is negligible.
December 27, 2017
DRAFT
24
Fig. 8. A polar lattice with two levels, where σ1 = σ.
for a given noise variance. By the guidelines given in Section II-C, the effective levels are those which can achieve
the target error probability with an actual rate not too close to either 0 or 1. Therefore, one can determine the
number of effective levels with the help of capacity curves in Fig. 7. For example, at the given noise variance
indicated by the straight line in Fig. 7, one may choose partition Z/2Z/4Z, i.e., two levels of component codes,
which was indeed suggested in [24].
The multilevel construction and the multistage decoding are shown in Fig. 8. For the ℓ-th level, g1 , g2 , · · · , gkℓ
are a set of code generators chosen from the matrix GN , and σℓ is the standard deviation of the noise.
Now, we give an example for length N = 1024 and target error probability Pe (L, σ 2 ) = 10−5 . Since the bottom
level is a ZN lattice decoder, σ3 ≈ 0.0845 for target error probability
1
3
· 10−5 . For the middle level, σ2 = 2 · σ3 =
0.1690. From Fig. 7, the channel capacity of the middle level is C(Z/2Z, σ22 ) = C(2Z/4Z, σ12 ) = 0.9874. For
the top level, σ = σ1 = 0.3380 and the capacity is 0.5145. Our goal is to find two polar codes approaching the
respective capacities at block error probabilities ≤
1
3
· 10−5 over these binary-input mod-2 channels.
For N = 1024, we found the first polar code with
code with
k2
N
= 0.9 for Pe (C2 , σ22 ) ≈
1
3
k1
N
= 0.23 for Pe (C1 , σ12 ) ≈
1
3
· 10−5 , and the second polar
· 10−5 . Thus, the sum rate of component polar codes RC = 0.23 + 0.9,
implying a capacity loss ǫ3 = 0.3719. Meanwhile, the factor ǫ1 = C(Z, 0.33802 ) = 0.0160. Therefore, the rate
losses at each level are 0.016, 0.285, and 0.087. From (11), the logarithmic VNR is given by
γL (σ)
log
≤ 2 (ǫ1 + ǫ3 ) = 0.7758,
2πe
(30)
which is 2.34 dB. Fig. 9 shows the simulation results for this example. It is seen that the estimate 2.34 dB is very
close to the actual gap at Pe (L, σ12 ) ≈ 10−5 . This simulation indicates that the performance of the component codes
is very important to the multilevel lattice. The gap to the Poltyrev capacity is largely due to the capacity losses of
component codes. Recall that the channel in the first level is degraded with respect to the one at the second level
according to Lemma 3, and the two polar codes in this construction turn out to be nested.
P
(i)
Thanks to density evolution [45], the upper bound i∈A (Z̃(W̃N )) on the block error probability of a polar
December 27, 2017
DRAFT
25
Fig. 9. Block error probabilities of polar lattices of length N = 1024 with multistage decoding. A comparison between polar lattices and
Barnes-Wall (BW) lattices is also presented. The BW lattices are constructed from Reed-Muller codes at each partition level. By changing the
Barnes-Wall rule (base on the Hamming weight) to the capacity rule after channel polarization, it can be seen that the performance of polar
lattices is significantly improved.
0
10
LDPC Lattice−1000
Polar Lattice−1024
LDLC−1000
LDA−1000
Finite−length Bound
−1
10
−2
SER
10
−3
10
−4
10
−5
10
−6
10
0
0.5
1
1.5
VNR/(2πe) (dB)
2
2.5
Fig. 10. SER of lattices with dimension around 1000.
code with finite length can be calculated numerically. According to (27), we plot the upper bound on the block
error probability Pe (L, σ 2 ) of the polar lattice in Fig. 9, which is quite tight.
The performance comparison of competing lattices approaching the Poltyrev capacity with dimension around
1000 is shown in Fig. 10 in terms of the symbol error rate (SER)8 . The simulation curves of other lattices are
8 SER
is defined as the average error probability of the coordinates of the lattice codeword λ, which is commonly used in literature. The
curve for the LDPC lattice was plotted with the normalized block error probability [17].
December 27, 2017
DRAFT
26
obtained from their corresponding papers. We note that the theoretical minimum gap to the Poltyrev capacity is
about 1 dB for dimension 1000 [43]. Among the four types of lattices compared, the LDPC lattice [17] has the
weakest performance, while all other three have similar performance at this dimension (the difference is within 0.5
dB). In contrast to the polar lattice and LDA lattice [18], [19], analytic results of the LDLC [20] are not available;
therefore, they are less understood in theory. The LDA lattice has slightly better performance than the polar lattice at
the expense of higher decoding complexity (O(p2 N log N )) if p-ary LDPC codes are employed. Assuming p ≈ 2r ,
it would rquire complexity O(22r N log N ).
B. Design Examples With Power Constraint
To satisfy the power constraint, we use discrete lattice distribution DZ,σs for shaping. The mutual information
I(Y ; Xℓ |X1:ℓ−1 ) at each level for different SNRs is shown in Fig. 11. We can see that for partition Z/2Z/..., five
levels are enough to achieve the AWGN channel capacity for SNR ranging from −5 dB to 20 dB. Note that this
is more than the number of levels required in the design of the AWGN-good lattice itself.
3.5
AWGN Capacity
3
I(Y;X1)
Mutual Information
I(Y;X2|X1)
2.5
2
I(Y;X3|X1:2)
I(Y;X |X
4
)
1:3
I(Y;X5|X1:4)
1.5
I(Y;X1:5)
1
0.5
0
−5
0
5
10
SNR (dB)
15
20
Fig. 11. Channel capacity for each level as a function of SNR.
December 27, 2017
DRAFT
27
Shaping set
Information set
Frozen set
Level Index
1
2
3
4
5
0
0.2
0.4
0.6
Set Proportion
0.8
1
Fig. 12. The proportions of the shaping set, information set, and frozen set on each level when N = 216 and SNR = 15 dB.
3.5
Lower bound of rate (Bit/dim)
AWGN Capacity
3
10
Bound for N=2
12
Bound for N=2
2.5
14
Bound for N=2
16
2
Bound for N=2
18
Bound for N=2
1.5
20
Bound for N=2
1
0.5
0
−5
0
5
10
15
20
SNR (dB)
Fig. 13. Lower bounds on the rates achieved by polar lattices with block error probability 5 × 10−5 for block lengths 210 , 212 , ..., 220 .
For each level, we estimate a lower bound on the code rate for block error probability PBℓ = 1 × 10−5 . The
is done by calculating an upper bound on the block error probability of the polar code, using the Bhattacharyya
parameter. With this target error probability, the assignments of bits to the information, shaping and frozen sets
on different levels are shown in Fig. 12 for SNR = 15 dB and N = 216 . In fact, X1 and X2 are nearly uniform
such that there is no need for shaping on the first two levels (these levels actually correspond to the AWGN-good
December 27, 2017
DRAFT
28
lattice). The third level is very good, and most bits are information bits. In contrast, the fifth level is mostly for
shaping; since its message rate is already small, adding another level clearly would not contribute to the overall
rate of the lattice code. Finally, lower bounds on the rates achieved by polar lattices with various block lengths are
shown in Fig. 13. We note that the gap to the channel capacity diminishes as N increases, and it is only about 0.1
bits/dimension when N = 220 .
VI. C ONCLUSIONS
In this paper, we have constructed polar lattices to approach the capacity of the power-constrained Gaussian
channel. The construction is based on a combination of channel polarization and source polarization. Without
shaping, the constructed polar lattices are AWGN-good. The Gaussian shaping on a polar lattice deals with the
power constraint but is technically more involved. The overall scheme is explicit and efficient, featuring quasi-linear
complexity.
ACKNOWLEDGMENTS
The authors would like to thank the Associate Editor and anonymous reviewers for helpful comments.
A PPENDIX A
P ROOF
OF
L EMMA 1
Proof. By the definition of the flatness factor, we have
fσ,Λ (x) ≤
1 + ǫΛ (σ)
.
V (Λ)
Thus, the differential entropy of the mod-Λ Gaussian noise is bounded by
Z
fσ,Λ (x) log fσ,Λ (x)dx
h(Λ, σ 2 ) = −
V(Λ1 )
≥−
Z
fσ,Λ (x) log
V(Λ1 )
= − log
1 + ǫΛ (σ)
dx
V (Λ)
1 + ǫΛ (σ)
V (Λ)
= log V (Λ) − log (1 + ǫΛ (σ)).
Therefore, from (5), C(Λ, σ 2 ) is bounded by log (1 + ǫΛ (σ)). The second inequality in (12) follows from the fact
log(1 + x) = log2 (e) · loge (1 + x) ≤ log(e) · x for x > 0.
A PPENDIX B
P ROOF
OF
L EMMA 2
Proof. For this purpose, we assume Λ = aZn and Λ′ = bZn where a, b are scaling parameters to be estimated.
We note that for all partition chains in [24], this is always possible: if the bottom lattice does not take the form of
bZn , one may simply further extend the partition chain (which will lead to an upper bound on r).
December 27, 2017
DRAFT
29
We firstly note that the flatness factor ǫΛ (σ) can be made arbitrarily small by scaling down the top lattice Λ.
To see this, we recall that ǫΛ (σe ) ≤ [1 + ǫΛ0 (σe )]n − 1 [11, Lemma 3] where Λ0 = aZ for the afore-mentioned
scaling factor a. Let Λ∗0 = a1 Z be the dual lattice of Λ0 . By [12, Corollary 1], we have
ǫΛ0 (σ) = ΘΛ∗0 (2πσ 2 ) − 1
X
exp(−2π 2 σ 2 |λ|2 ) − 1
=
λ∈Λ∗
0
=2
X
1
λ∈ a
N
exp(−2π 2 σ 2 |λ|2 )
(31)
2 exp(−2π 2 σ 2 a12 )
1 − exp(−2π 2 σ 2 a32 )
1
≤ 4 exp(−2π 2 σ 2 2 ) for sufficiently small a.
a
≤
Let
1
a
√
= Θ( N ), we have ǫΛ0 (σ) = e−Θ(N ) and hence ǫΛ (σ) = e−Θ(N ) for fixed n.
Secondly, by the union bound, the error probability of the bottom lattice Λ′ is upper-bounded by
b2
b
Pe (Λ′ , σ 2 ) ≤ nQ
≤ ne− 8σ2
2σ
where we apply the Chernoff bound on the Q-function. We want
Pe (Λ′ , σ 2 ) = e−Θ(N ) ,
√
which leads to b = Θ( N ) for fixed n.
For a binary lattice partition, we have (b/a)n = 2r . Thus, we conclude that
b
r = n log
= n log Θ(N ) = Θ(log N ).
a
A PPENDIX C
P ROOF
OF
L EMMA 3
Proof. By the self-similarity of the lattice partition chain, we can scale a Λℓ−1 /Λℓ channel to a Λℓ /Λℓ+1 channel by
multiplying the output of a Λℓ−1 /Λℓ channel with T . Since T = αV for some scaling factor α > 1 and orthogonal
matrix V , the Gaussian noise for each dimension is still independent of each other and the noise variance per
dimension is increased after the scaling. Therefore, a Λℓ−1 /Λℓ channel is stochastically equivalent to a Λℓ /Λℓ+1
channel with a larger noise variance. For our design examples, a Z/2Z channel with Gaussian noise variance σ 2
is equivalent to a 2Z/4Z channel with Gaussian noise variance 4σ 2 , and a Z2 /RZ2 channel with noise variance
σ 2 per dimension is equivalent to a RZ2 /2Z2 channel with noise variance 2σ 2 per dimension. Then our task is to
prove that a Λℓ /Λℓ+1 channel with noise variance σ22 is degraded with respect to a Λℓ /Λℓ+1 channel with noise
variance σ12 if σ12 ≤ σ22 .
To see this, we construct an intermediate channel with an input in R(Λℓ+1 ) and a mod-Λℓ+1 operation at the
receiver’s front end, as depicted in Fig. 14. The noise variance of this mod-Λℓ+1 channel is given by σ22 − σ12 per
December 27, 2017
DRAFT
30
h
i
dimension. By the property [X + Y ] mod Λℓ+1 = X mod Λℓ+1 + Y mod Λℓ+1 , we find that the concatenated
channel consisting of a Λℓ /Λℓ+1 channel with noise variance σ12 followed by the afore-mentioned intermediate
channel is stochastically equivalent to a Λℓ /Λℓ+1 channel with noise variance σ22 . The proof is completed according
to Definition 2.
%#
(a)
&
Mod
%#
(b)
!"#
%'
&
Mod
%#
(c)
&
$
!"#
$(
%'
Mod
Mod
!"#
!"#
$(
intermediate channel
Fig. 14. Relationship between the channels regarding degradation. X ∈ R(Λℓ+1 ) denotes the channel input, while N1 and N2 denote
independent additive Gaussian noises with variances σ12 and σ22 − σ12 , respectively. Clearly, the two Λℓ /Λℓ+1 channels with noise variances
σ12 and σ22 can be described by channel (a) and (b), respectively. By the property of modulo operation, channel (b) is equivalent to channel (c),
which is a concatenated channel consisting of channel (a) and an intermediate mod-Λℓ+1 channel.
A PPENDIX D
P ROOF
OF
L EMMA 5
Proof. For convenience we consider a one-dimensional partition chain Z/2Z/ · · · . The proof can be extended to
the multi-dimensional case by sandwiching the partition in Zn /2Zn / · · · , which reduces to the one-dimensional
case.
For level r, the selected coset Ar can be written as x1 + · · ·2r−1 xr + 2r Z. Clearly, Ar is a subset of Ar−1 .
Let λ1 and λ2 denote the two lattice points with smallest norm in set Ar−1 . Without loss of generality, we assume
λ1 ≤ 0 ≤ λ2 and |λ1 | ≤ |λ2 |. Observe that λ2 − λ1 = 2r−1 . For a Gaussian distribution with variance σs2 , we can
find a positive integer T , making the probability
Z T σs
x2
1
√
exp(− 2 )dx → 1.
2σs
2πσs
−T σs
Actually, this T does not need to be very large. For instance, when T = 6, the above probability is larger than
1 − 2e−9 . Now we assume 2r−1 = 3T σs , and T = δN for some constant δ, then λ1 and λ2 cannot be in the
December 27, 2017
DRAFT
31
interval [−T σs , T σs ] simultaneously. If the two points are both outside of [−T σs , T σs ], then we have
P
(x+λ1 )2
√ 1
exp(−
2σs2 )
2πσs
fσs (Ar−1 (x1:r−1 ))
x∈2r−1 Z
<
P (Ar−1 ) =
√ 1
fσs (Z)
2πσs
≤2
≤2
X
exp(−
x∈2r−1 Z−
(x + λ1 )2
)
2σs2
λ2
2
exp(− 2σ12 )
1 − exp(−
s
(2r−1 )2
2σs2
)
≤2
exp(− T2 )
2
1 − exp(− 9T2 )
,
where Z− represents all non-positive integers. This means the probability of choosing Ar−1 goes to zero when T
(or N ) is large. Therefore, we have that the point λ1 is in the interval [−T σs , T σs ] and λ2 lies outside. Without
loss of generality, we assume that the two cosets corresponding to xr = 0 and xr = 1 are λ1 + 2r Z and λ2 + 2r Z,
respectively. Then we have
P
2
1)
exp(− (x+λ
2σ2 )
P (xr = 0|x1:r−1 )
= P
2
2)
P (xr = 1|x1:r−1 )
exp(− (x+λ
2σ2 )
x∈2r Z
s
s
x∈2r Z
≥
exp(−
2
P
x∈2r Z+
≥
λ21
2σs2
)
2
2)
exp(− (x+λ
2σs2 )
λ2
exp(− 2σ12 )
s
λ2
2 · exp(− 2σ22 )
s
22r
1 − exp − 2
,
2σs
where Z+ represents all non-negative integers. Since λ2 − λ1 = 2r−1 = 3T σs and λ2 + λ1 ≥ T σs , for any T > 1,
we can obtain
P (xr = 0|x1:r−1 )
1
3 2
≥ exp
T (1 − exp(−18T 2))
P (xr = 1|x1:r−1 )
2
2
1
3 2
3 2 2
1
≥ exp
T
δ N .
= exp
4
2
4
2
Assume that
1
4
exp( 23 δ 2 N 2 ) = M , we can get P (xr = 0|x1:r−1 ) ≥
M
M+1
and P (xr = 1|x1:r−1 ) ≤
1
M+1 .
Then we
have,
I(Y ; Xr |X1:r−1 ) ≤ H(Xr |X1:r−1 ) ≤ h2
1
M +1
,
1
where h2 (p) = plog( p1 ) + (1 − p)log( 1−p
) denotes the binary entropy function. By the relationship ln(x) ≤
when x ≥ 1, we finally have
I(Y ; Xr |X1:r−1 ) ≤ log(e)
1
1
√ +
M
M
= log(e)
2
4
+
2r
exp(δ1 2 ) exp(δ2 22r )
x−1
√
x
,
where δ1 and δ2 are two positive constants. Therefore, when r = O(log log N ), we have I(Y ; Xr |X1:r−1 ) → 0,
P
and ℓ≥r I(Y ; Xℓ |X1:ℓ−1 ) ≤ O( N1 ).
December 27, 2017
DRAFT
32
A PPENDIX E
P ROOF
OF
T HEOREM 5
Proof. Let Ei denote the set of pairs of u1:N and y 1:N such that decoding error occurs at the i-th bit, then the block
S
decoding error event is given by E ≡ i∈I Ei . According to our encoding scheme, each codeword u1:N appears
with probability
2−(|I|+|F |)
Y
PU i |U 1:i−1 (ui |u1:i−1 ).
i∈S
Then the expectation of decoding error probability over all random mapping is expressed as
Y
X
2−(|I|+|F |)(
E[Pe ] =
PU i |U 1:i−1 (ui |u1:i−1 ))
u1:N ,y 1:N
i∈S
· PY 1:N |U 1:N (y 1:N |u1:N )1[(u1:N , y 1:N ) ∈ E].
Now we define the probability distribution QU 1:N ,Y 1:N as
QU 1:N ,Y 1:N (u1:N , y 1:N ) = 2−(|I|+|F |)(
Y
i∈S
PU i |U 1:i−1 (ui |u1:i−1 ))PY 1:N |U 1:N (y 1:N |u1:N ).
Then the variational distance between QU 1:N ,Y 1:N and PU 1:N ,Y 1:N can be bounded as
X
|Q(u1:N , y 1:N ) − P (u1:N , y 1:N )|
2kQU 1:N ,Y 1:N − PU 1:N ,Y 1:N k =
u1:N ,y 1:N
(a)
=
X
u1:N ,y 1:N
≤
=
X
i∈I∪F
X
i∈I∪F
(b)
≤
|
X
i
X
(Q(ui |u1:i−1 ) − P (ui |u1:i−1 ))(
u1:N ,y 1:N
X
u1:i−1
X
X
i∈I∪F u1:i−1
i−1
Y
j=1
|Q(ui |u1:i−1 ) − P (ui |u1:i−1 )|(
N
Y
P (ui |u1:i−1 ))(
i−1
Y
j=1
j=i+1
P (ui |u1:i−1 ))(
N
Y
Q(ui |u1:i−1 ))Q(y 1:N |u1:N )|
j=i+1
Q(ui |u1:i−1 ))Q(y 1:N |u1:N )
2P (u1:i−1 )kQU i |U 1:i−1 =u1:i−1 − PU i |U 1:i−1 =u1:i−1 k
q
P (u1:i−1 ) 2ln2D(PU i |U 1:i−1 =u1:i−1 kQU i |U 1:i−1 =u1:i−1 )
(32)
X s
X
≤
P (u1:i−1 )D(PU i |U 1:i−1 =u1:i−1 kQU i |U 1:i−1 =u1:i−1 )
2ln2
i∈I∪F
u1;i−1
X q
≤
2ln2D(PU i |U 1:i−1 ||QU i |U 1:i−1 )
i∈I∪F
≤
Xp
Xp
2ln2(1 − H(U i |U 1:i−1 )) +
2ln2(1 − H(U i |U 1:i−1 ))
i∈I
i∈F
Xp
Xq
≤
2ln2(1 − Z(U i |U 1:i−1 )2 ) +
2ln2(1 − Z(U i |U 1:i−1 , Y 1:N )2 )
i∈I
i∈F
p
≤ 2N 4ln2 · 2−N β ,
December 27, 2017
DRAFT
33
where equality (a) follows from [6, Equation (56)] and Q(y 1:N |u1:N ) = P (y 1:N |u1:N ). D(·||·) in the inequality
(b) is the relative entropy, and this inequality holds because of the Pinsker’s inequality. Then we have
E[Pe ] = QU 1:N ,Y 1N (E)
≤ kQU 1:N ,Y 1:N − PU 1:N ,Y 1:N k + PU 1:N ,Y 1:N (E)
X
≤ kQU 1:N ,Y 1:N − PU 1:N ,Y 1:N k +
PU 1:N ,Y 1:N (Ei ),
(33)
i∈I
where
PU 1:N ,Y 1:N (Ei ) ≤
≤
X
P (u1;i−1 , y 1:N )P (ui |u1:i−1 , y 1:N ) · 1[P (ui |u1:i−1 , y 1:N ) ≤ P (ui ⊕ 1|u1:i−1 , y 1:N )]
X
P (u1;i−1 , y 1:N )P (ui |u1:i−1 , y 1:N )
u1:N ,y 1:N
u1:N ,y 1:N
s
P (ui ⊕ 1|u1:i−1 , y 1:N )
P (ui |u1:i−1 , y 1:N )
β
= Z(U i |U 1:i−1 , Y 1:N ) ≤ 2−N .
√
β
β′
From (32) and (33), we have E[Pe ] ≤ 2N 4ln2 · 2−N β + N 2−N = N 2−N for any β ′ < β < 0.5.
A PPENDIX F
P ROOF
OF
T HEOREM 6
1:N
Proof. Let Ei denote the set of triples of u1:N
and y 1:N such that decoding error occurs at the i-th bit, then
2 , x1
S
the block decoding error event is given by E ≡ i∈I Ei . According to our encoding scheme, each codeword u1:N
2
appears with probability
2−(|I2 |+|F2 |)
Y
i∈S2
PU i |U 1:i−1 ,X 1:N (ui2 |u1:i−1
, x1:N
1 ).
2
2
2
1
Then the expectation of decoding error probability over all random mapping is expressed as
Y
X
2−(|I2 |+|F2 |) (
E[Pe ] =
PU i |U 1:i−1 ,X 1:N (ui2 |u1:i−1
, x1:N
1 ))
2
2
u1:N
,x1:N
,y 1:N
2
1
2
1
i∈S2
1:N
1:N
1:N
1:N
·PY 1:N ,X11:N |U21:N (y 1:N , x1:N
) ∈ E].
1 |u2 )1[(u2 , x1 , y
Now we define the probability distribution QU21:N ,X11:N ,Y 1:N as
1:N
1:N
QU21:N ,X11:N ,Y 1:N (u1:N
) =2−(|I2 |+|F2 |) · QX11:N (x1:N
2 , x1 , y
1 )
Y
1:N 1:N
(
PU i |U 1:i−1 ,X 1:N (ui2 |u1:i−1
, x1:N
|u2 , x1:N
1 )) · PY 1:N |X11:N ,U21:N (y
1 ).
2
2
2
1
i∈S2
Then the variational distance between QU21:N ,X11:N ,Y 1:N and PU21:N ,X11:N ,Y 1:N can be bounded as
X
1:N
1:N
1:N
|Q(u1:N
) − P (u21:N , x1:N
)|
2kQU21:N ,X11:N ,Y 1:N − PU21:N ,X11:N ,Y 1:N k =
2 , x1 , y
1 ,y
u1:N
,x1:N
,y 1:N
2
1
X
=
u1:N
,x1:N
,y 1:N
2
1
(a)
≤
+
X
1:N
1:N
1:N 1:N
1:N 1:N
1:N
1:N 1:N
|Q(u1:N
|u2 , x1:N
|u2 , x1:N
2 |x1 )Q(x1 )Q(y
1 ) − P (u2 |x1 )P (x1 )P (y
1 )|
u1:N
,x1:N
,y 1:N
2
1
X
u1:N
,x1:N
,y 1:N
2
1
December 27, 2017
1:N
1:N 1:N
1:N
1:N 1:N
|Q(u1:N
|u2 , x1:N
2 |x1 ) − P (u2 |x1 )|P (x1 )P (y
1 )
1:N
1:N 1:N
1:N 1:N
|Q(x1:N
|u2 , x1:N
1 ) − P (x1 )|Q(u2 |x1 )P (y
1 )
DRAFT
34
1:N
1:N 1:N
where inequation (a) follows from [6, Equation (56)], Q(y 1:N |u1:N
|u2 , x1:N
2 , x1 ) = P (y
1 ). For the first
summation, following the same fashion as the proof of Theorem 5, we can prove
X
u1:N
,x1:N
,y 1:N
2
1
1:N 1:N
1:N
1:N 1:N
1:N
|Q(u1:N
|u2 , x11:N ) ≤ 2N
2 |x1 ) − P (u2 |x1 )|P (x1 )P (y
p
4ln2 · 2−N β .
According to the result of the coding scheme for level 1, we already have
p
2kQU11:N ,Y 1:N − PU11:N ,Y 1:N k ≤ 2N 4ln2 · 2−N β .
(34)
Since we have PY 1:N |U11:N = QY 1:N |U11:N , we can write
2kQU11:N − PU11:N k ≤ 2N
p
4ln2 · 2−N β .
(35)
Clearly, there is a one to one mapping between U11:N and X11:N , then we immediately have 2kQX11:N − PX11:N k ≤
√
2N 4ln2 · 2−N β . Therefore, for the second summation,
X
1:N
1:N 1:N
1:N 1:N
|Q(x1:N
|u2 , x1:N
1 ) − P (x1 )|Q(u2 |x1 )P (y
1 )
u1:N
,x1:N
,y 1:N
2
1
=
X
x1:N
1
1:N
|Q(x1:N
1 ) − P (x1 )| ≤ 2N
√
Then we have ||QU 1:N ,X 1:N ,Y 1N − PU 1:N ,X 1:N ,Y 1N || ≤ 4N 4ln2 · 2−N β , and
2
1
2
p
4ln2 · 2−N β .
(36)
1
E[Pe ] = QU 1:N ,X 1:N ,Y 1N (E)
2
1
≤ kQU 1:N ,X 1:N ,Y 1N − PU 1:N ,X 1:N ,Y 1N k + PU 1:N ,X 1:N ,Y 1N (E)
2
1
2
1
2
1
X
≤ kQU 1:N ,X 1:N ,Y 1N − PU 1:N ,X 1:N ,Y 1N k +
PU 1:N ,X 1:N ,Y 1N (Ei ),
2
1
2
1
2
(37)
1
i∈I
The rest part of the proof follows the same fashion of the proof of Theorem 5. Finally we have E[Pe ] ≤ N 2−N
β′
for any β ′ < β < 0.5.
R EFERENCES
[1] E. Arıkan, “Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels,”
IEEE Trans. Inform. Theory, vol. 55, no. 7, pp. 3051–3073, July 2009.
[2] E. Şaşoğlu, E. Telatar, and E. Arıkan, “Polarization for arbitrary discrete memoryless channels,” in Proc. IEEE Inform. Theory Workshop
(ITW), Taormina, Italy, Oct. 2009, pp. 144–148.
[3] A. Sahebi and S. Pradhan, “Multilevel channel polarization for arbitrary discrete memoryless channels,” IEEE Trans. Inform. Theory,,
vol. 59, no. 12, pp. 7839–7857, Dec. 2013.
[4] W. Park and A. Barg, “Polar codes for q-ary channels, q = 2r ,” IEEE Trans. Inform. Theory,, vol. 59, no. 2, pp. 955–969, Feb. 2013.
[5] R. Mori and T. Tanaka, “Non-binary polar codes using Reed-Solomon codes and algebraic geometry codes,” in Proc. IEEE Inform. Theory
Workshop (ITW), Dublin, Ireland, Aug. 2010, pp. 1–5.
[6] J. Honda and H. Yamamoto, “Polar coding without alphabet extension for asymmetric models,” IEEE Trans. Inform. Theory, vol. 59,
no. 12, pp. 7829–7838, Dec. 2013.
[7] M. Mondelli, S. H. Hassani, and R. L. Urbanke, “How to achieve the capacity of asymmetric channels,” CoRR, vol. abs/1406.7373, 2014.
[Online]. Available: http://arxiv.org/abs/1406.7373
[8] E. Abbe and E. Telatar, “Polar codes for the m-user multiple access channel,” IEEE Trans. Inform. Theory, vol. 58, no. 8, pp. 5437–5448,
Aug. 2012.
December 27, 2017
DRAFT
35
[9] E. Abbe and A. Barron, “Polar coding schemes for the AWGN channel,” in Proc. IEEE Int. Symp. Inform. Theory (ISIT), Saint-Petersburg,
Russia, July 2011, pp. 194–198.
[10] U. Erez and R. Zamir, “Achieving 1/2 log (1+SNR) on the AWGN channel with lattice encoding and decoding,” IEEE Trans. Inform.
Theory, vol. 50, no. 10, pp. 2293–2314, Oct. 2004.
[11] C. Ling and J.-C. Belfiore, “Achieiving AWGN channel capacity with lattice Gaussian coding,” IEEE Trans. Inform. Theory, vol. 60,
no. 10, pp. 5918–5929, Oct. 2014.
[12] C. Ling, L. Luzzi, J.-C. Belfiore, and D. Stehlé, “Semantically secure lattice codes for the Gaussian wiretap channel,” IEEE Trans. Inform.
Theory, vol. 60, no. 10, pp. 6399–6416, Oct. 2014.
[13] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing interference through structured codes,” IEEE Trans. Inform. Theory, vol. 57,
no. 10, pp. 6463–6486, Oct. 2011.
[14] R. Zamir, S. Shamai, and U. Erez, “Nested linear/lattice codes for structured multiterminal binning,” IEEE Trans. Inform. Theory, vol. 48,
no. 6, pp. 1250–1276, June 2002.
[15] R. Zamir, Lattice Coding for Signals and Networks.
Cambridge, UK: Cambridge University Press, 2014.
[16] G. Poltyrev, “On coding without restictions for the AWGN channel,” IEEE Trans. Inform. Theory, vol. 40, pp. 409–417, Mar. 1994.
[17] M.-R. Sadeghi, A. Banihashemi, and D. Panario, “Low-density parity-check lattices: Construction and decoding analysis,” IEEE Trans.
Inform. Theory, vol. 52, no. 10, pp. 4481–4495, Oct. 2006.
[18] N. Di Pietro, J. J. Boutros, G. Zémor, and L. Brunel, “Integer low-density lattices based on Construction A,” in Proc. IEEE Inform. Theory
Workshop (ITW), Lausanne, Switzerland, Sept. 2012, pp. 422–426.
[19] N. di Pietro, G. Zémor, and J. J. Boutros, “New results on Construction A lattices based on very sparse parity-check matrices,” in Proc.
IEEE Int. Symp. Inform. Theory (ISIT), Istanbul, Turkey, July 2013, pp. 1675–1679.
[20] N. Sommer, M. Feder, and O. Shalvi, “Low-density lattice codes,” IEEE Trans. Inform. Theory, vol. 54, no. 4, pp. 1561–1585, Apr. 2008.
[21] G. Forney and L.-F. Wei, “Multidimensional constellations–Part I: Introduction, figures of merit, and generalized cross constellations,”
IEEE J. Sel. Areas Commun., vol. 7, no. 6, pp. 877–892, Aug 1989.
[22] F. R. Kschischang and S. Pasupathy, “Optimal nonuniform signaling for Gaussian channels,” IEEE Trans. Inform. Theory, vol. 39, no. 3,
pp. 913–929, May 1993.
[23] N. di Pietro, G. Zémor, and J. J. Boutros, “LDA lattices without dithering achieve capacity on the gaussian channel,” CoRR, vol.
abs/1603.02863, 2016. [Online]. Available: http://arxiv.org/abs/1603.02863
[24] G. D. Forney Jr., M. Trott, and S.-Y. Chung, “Sphere-bound-achieving coset codes and multilevel coset codes,” IEEE Trans. Inform. Theory,
vol. 46, no. 3, pp. 820–850, May 2000.
[25] D. Micciancio and O. Regev, “Worst-case to average-case reductions based on Gaussian measures,” in Proc. Ann. Symp. Found. Computer
Science, Oct. 2004, pp. 372–381.
[26] Y. Yan and C. Ling, “A construction of lattices from polar codes,” in Proc. IEEE Inform. Theory Workshop (ITW), Lausanne, Switzerland,
Sept. 2012, pp. 124–128.
[27] Y. Yan, C. Ling, and X. Wu, “Polar lattices: Where Arıkan meets Forney,” in Proc. IEEE Int. Symp. Inform. Theory (ISIT), Istanbul,
Turkey, 2013, pp. 1292–1296.
[28] M. Seidl, A. Schenk, C. Stierstorfer, and J. B. Huber, “Multilevel polar-coded modulation,” IEEE Trans. Commun., vol. 61, no. 10, pp.
4108–4119, Oct. 2013.
[29] U. Wachsmann, R. Fischer, and J. Huber, “Multilevel codes: Theoretical concepts and practical design rules,” IEEE Trans. Inform. Theory,
vol. 45, no. 5, pp. 1361–1391, July 1999.
[30] A. Joseph and A. Barron, “Least squares superposition codes of moderate dictionary size are reliable at rates up to capacity,” IEEE Trans.
Inform. Theory, vol. 58, no. 5, pp. 2541–2557, May 2012.
[31] ——, “Fast sparse superposition codes have near exponential error probability for R < C,” IEEE Trans. Inform. Theory, vol. 60, no. 2,
pp. 919–942, Feb. 2014.
[32] J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and Groups, Second Edition, 1993, Springer-Verlag, New York.
[33] W. Kositwattanarerk and F. Oggier, “On Construction D and related constructions of lattices from linear codes,” in Int. Workshop on
Coding and Cryptography (WCC), 2013.
[34] G. D. Forney Jr., “Coset codes-Part I: Introduction and geometrical classification,” IEEE Trans. Inform. Theory, vol. 34, no. 5, pp. 1123–
1151, Sept. 1988.
December 27, 2017
DRAFT
36
[35] E. Arıkan and E. Telatar, “On the rate of channel polarization,” in IEEE Int. Symp. Inform. Theory (ISIT), Seoul, Korea, July 2009, pp.
1493–1495.
[36] T. M. Cover and J. A. Thomas, Elements of Information Theory.
New York: Wiley, 1991.
[37] S. B. Korada, “Polar codes for channel and source coding,” Ph.D. dissertation, Ecole Polytechnique Fédérale de Lausanne, 2009.
[38] I. Tal and A. Vardy, “How to construct polar codes,” IEEE Trans. Inform. Theory, vol. 59, no. 10, pp. 6562–6582, Oct. 2013.
[39] R. Pedarsani, S. Hassani, I. Tal, and E. Telatar, “On the construction of polar codes,” in Proc. IEEE Int. Symp. Inform. Theory (ISIT),
Saint-Petersburg, Russia, July 2011, pp. 11–15.
[40] S. H. Hassani, K. Alishahi, and R. Urbanke, “Finite-length scaling for polar codes,” IEEE Trans. Inform. Theory, vol. 60, no. 10, pp.
5875–5898, Oct. 2014.
[41] V. Guruswami and P. Xia, “Polar codes: Speed of polarization and polynomial gap to capacity,” in IEEE 54th Annual Symp. Foundations
of Computer Science (FOCS), Oct. 2013, pp. 310–319.
[42] D. Goldin and D. Burshtein, “Improved bounds on the finite length scaling of polar codes,” IEEE Trans. Inform. Theory, vol. 60, no. 11,
pp. 6966–6978, Nov. 2014.
[43] A. Ingber, R. Zamir, and M. Feder, “Finite dimensional infinite constellations,” IEEE Trans. Inform. Theory, vol. 59, no. 3, pp. 1630–1656,
Mar. 2013.
[44] E. Arıkan, “Source polarization,” in Proc. IEEE Int. Symp. Inform. Theory (ISIT), Austin, USA, July 2010, pp. 899–903.
[45] R. Mori and T. Tanaka, “Performance of polar codes with the construction using density evolution,” IEEE Comm. Lett., vol. 13, no. 7, pp.
519–521, July 2009.
December 27, 2017
DRAFT
| 7cs.IT
|
Approximate Correlation Clustering Using
Same-Cluster Queries
Nir Ailon⋆1 , Anup Bhattacharya⋆⋆2 , and Ragesh Jaiswal⋆ ⋆ ⋆2
arXiv:1712.06865v1 [cs.DS] 19 Dec 2017
1
2
Technion, Haifa, Israel.†
Department of Computer Science and Engineering,
Indian Institute of Technology Delhi.‡
Abstract. Ashtiani et al. (NIPS 2016) introduced a semi-supervised
framework for clustering (SSAC) where a learner is allowed to make samecluster queries. More specifically, in their model, there is a query oracle
that answers queries of the form “given any two vertices, do they belong
to the same optimal cluster?”. In many clustering contexts, this kind of
oracle queries are feasible. Ashtiani et al. showed the usefulness of such a
query framework by giving a polynomial time algorithm for the k-means
clustering problem where the input dataset satisfies some separation
condition. Ailon et al. extended the above work to the approximation
setting by giving an efficient (1+ε)-approximation algorithm for k-means
for any small ε > 0 and any dataset within the SSAC framework. In
this work, we extend this line of study to the correlation clustering
problem. Correlation clustering is a graph clustering problem where
pairwise similarity (or dissimilarity) information is given for every pair
of vertices and the objective is to partition the vertices into clusters that
minimise the disagreement (or maximises agreement) with the pairwise
information given as input. These problems are popularly known as
MinDisAgree and MaxAgree problems, and MinDisAgree[k] and MaxAgree[k]
are versions of these problems where the number of optimal clusters is at
most k. There exist Polynomial Time Approximation Schemes (PTAS)
for MinDisAgree[k] and MaxAgree[k] where the approximation guarantee
is (1 + ε) for any small ε and the running time is polynomial in the input
parameters but exponential in k and 1/ε. We get a significant running
time improvement within the SSAC framework at the cost of making a
small number of same-cluster queries. We obtain an (1+ε)-approximation
algorithm for any small ε with running time that is polynomial in the
input parameters and also in k and 1/ε. We also give non-trivial upper
and lower bounds on the number of same-cluster queries, the lower bound
being based on the Exponential Time Hypothesis (ETH). Note that the
existence of an efficient algorithm for MinDisAgree[k] in the SSAC setting
exhibits the power of same-cluster queries since such polynomial time
algorithm (polynomial even in k and 1/ε) is not possible in the classical
⋆
⋆⋆
⋆⋆⋆
†
‡
Nir Ailon acknowledges the generous support of ISF grant number 2021408
Anup Bhattacharya acknowledges the support of TCS fellowship at IIT Delhi.
Ragesh Jaiswal acknowledges the support of ISF-UGC India-Israel Grant 2014.
Email address: [email protected]
Email addresses: {anupb, rjaiswal}@cse.iitd.ac.in
(non-query) setting due to our conditional lower bounds. Our conditional
lower bound is particularly interesting as it not only establishes a lower
bound on the number of same cluster queries in the SSAC framework
but also establishes a conditional lower bound on the running time of
any (1 + ε)-approximation algorithm for MinDisAgree[k].
1
Introduction
Correlation clustering is a graph clustering problem where we are given similarity
or dissimilarity information for pairs of vertices. The input is a graph G on n
vertices. Edges of G are labeled as similar (positive) or dissimilar (negative).
The clustering objective is to partition the vertices into clusters such that edges
labeled ‘positive’ remain within clusters and ‘negative’ edges go across clusters.
However, this similarity/dissimilarity information may be inconsistent with this
objective. For example, there may exist vertices u, v, w such that edges (u, v)
and (u, w) are labeled ‘positive’ whereas edge (v, w) is labeled ‘negative’. In this
case, it is not possible to come up with a clustering of these 3 vertices that would
agree with all the edge labels. The objective of correlation clustering is to come up
with a clustering that minimises disagreement or maximises agreement with the
edge labels given as input. The minimisation version of the problem, known as
MinDisAgree, minimises the sum of the number of negative edges present inside
clusters and the number of positive edges going across clusters. Similarly, the
maximisation version is known as MaxAgree where the objective is to maximise
the sum of the number of positive edges present inside clusters and the number
of negative edges going across clusters. Unlike k-means or k-median clustering,
in correlation clustering, there is no restriction on the number of clusters formed
by the optimal clustering. When the number of optimal clusters is given to
be at most k, these problems are known as MinDisAgree[k] and MaxAgree[k]
respectively.
Bansal et al. [8] gave a constant approximation algorithm for MinDisAgree and
a PTAS for MaxAgree. Subsequently, Charikar et al. [9] improved approximation
guarantee for MinDisAgree to 4, and showed that MinDisAgree is APX-hard.
These results are for correlation clustering on complete graphs as it is known for
general graphs, it is at least as hard as minimum multi-cut problem [9]. Since
MinDisAgree is APX-hard [9], additional assumptions were introduced for better
results. For example [17,15] studied MinDisAgree where the input is noisy and
comes from a semi-random model. When k is given as part of the input, Giotis
and Guruswami [12] gave a PTAS for MinDisAgree[k].
Recently there have been some works [2,7] with a beyond-worst case flavour
where polynomial time algorithms for NP-hard problems have been designed
under some stability assumptions. Ashtiani et al. [3] considered one such stability
assumption called γ-margin. They introduced a semi-supervised active learning
(SSAC) framework and within this framework, gave a probabilistic polynomial
time algorithm for k-means on datasets that satisfy the γ-margin property.
More specifically, their SSAC framework involves a query oracle that answers
queries of the form “given any two vertices, do they belong to the same optimal
cluster?”. The query oracle responds with a Yes/No answer where these answers
are assumed to be consistent with some fixed optimal solution. In this framework,
they studied the query complexity for polynomial time algorithms for k-means
on datasets satisfying the γ-margin property. Ailon et al. [1] extended this work
to study query complexity bounds for (1+ε)-approximation for k-means in SSAC
framework for any small ε > 0 without any stability assumption on the dataset.
They gave almost matching upper and lower bounds on the number of queries
for (1 + ε)-approximation of k-means problem in SSAC framework.
In this work, we study MinDisAgree[k] in the SSAC framework, where the
optimal clustering has at most k clusters and give upper and lower bounds on
the number of same-cluster queries for (1 + ε)-approximation for correlation
clustering for any ε > 0. We also give upper bounds for MaxAgree[k]. Our
algorithm is based on the PTAS by Giotis and Guruswami [12] for MinDisAgree[k].
The algorithm by Giotis and Guruswami involves random sampling a subset
S of vertices and considers all possible ways of partitioning S into k clusters
S = {S1 , . . . , Sk }, and for every such k-partitioning, clusters the rest of the
vertices greedily. Every vertex v ∈ V \ S is assigned a cluster Sj that maximizes
its agreement with the edge labels. Their main result was the following.
Theorem 1 (Giotis and Guruswami [12]). For every k ≥ 2, there is a PTAS
k
2
for MinDisAgree[k] with running time nO(9 /ε ) log n.
Since Giotis and Guruswami considered all possible ways of partitioning
subset S into k clusters, their running time has exponential dependence on k.
Here, we make the simple observation that within the SSAC framework we can
overcome this exponential dependence on k by making same-cluster queries to
the oracle. The basic idea is to randomly sample a subset S of vertices as before
and partition it optimally into k clusters by making same-cluster queries to the
oracle. Note that by making at most k|S| same-cluster queries, one can partition
S optimally into k clusters. Once we have subset S partitioned as in the optimal
clustering (a key step needed in the analysis of Giotis and Guruswami) we follow
their algorithm and analysis for (1 + ε)-approximation for MinDisAgree[k]. Here
is our main result for MinDisAgree[k] in the SSAC framework. We obtain similar
results for MaxAgree[k].
Theorem 2 (Main result: Upper bound). Let ε > 0 and k ≥ 2. There is
a (randomized)
that uses
algorithm in the SSAC framework for MinDisAgree[k]
14
nk14 log k log n
k log k log n
same-cluster queries, runs in time O(
) and outputs
O
ε6
ε6
a (1 + ε)-approximate solution with high probability.
We complement our upper bound result by providing a lower bound on the
number of queries in the SSAC framework for any efficient (1 +ε)-approximation
algorithm for MinDisAgree for any ε > 0. Our lower bound result is conditioned
on the Exponential Time Hypothesis (ETH hypothesis) [13,14]. Our lower bound
result implies that the number of queries is depended on the number of optimal
clusters k. Our main result with respect to query lower bound is given as follows.
Theorem 3 (Main result: Lower bound). Given that Exponential Time
Hypothesis (ETH) holds, there exists a constant δ > 0 such that any (1 + δ)approximation algorithm for MinDisAgree[k] in the SSAC framework that runs
in polynomial time makes Ω( polyklog k ) same-cluster queries.
Exponential Time Hypothesis is the following statement regarding the hardness
of the 3-SAT problem.
Exponential Time Hypothesis (ETH)[13,14]: There does not exist an
algorithm that can decide whether any 3-SAT formula with m clauses is
satisfiable with running time 2o(m) .
Note that our query lower bound result is a simple corollary of the following
theorem that we prove.
Theorem 4. If the Exponential Time Hypothesis (ETH) holds, then there exists
a constant δ > 0 such that any (1+δ)-approximation algorithm for MinDisAgree[k]
k
requires 2Ω( poly log k ) time.
The above lower bound statement may be of independent interest. It was
already known that MinDisAgree is APX-hard. Our result is a non-trivial addition
to the understanding of the hardness of the correlation clustering problem. Given
that our query upper bound result is through making simple observations in the
algorithms of Giotis and Guruswami, our lower bound results may be regarded
as the primary contribution of this work. So, we first give our lower bound results
in the next section and the upper bound results in Section 3. However, before
we start discussing our results, here is a brief discussion on the related works.
Related Works There have been numerous works on clustering problems in semisupervised settings. Balcan and Blum [5] proposed an interactive framework for
clustering which use ‘split/merge’ queries. In this framework, given any abritrary
clustering C = {C1 , C2 , . . . , } as query, oracle specifies some cluster Cl should be
split or clusters Ci and Cj should be merged. Awasthi et al. [4] developed a local
clustering algorithm which uses these split/merge queries. One versus all queries
for clustering were studied by Voevodski et al. [20]. The oracle, on a query s ∈ X,
returns distances from s to all points in X. The authors provided a clustering,
close to optimal k-median clustering, with only O(k) such queries on instances
satisfying (c, ε)-approximation stability property [6]. Fomin et al. [11] gave a
conditional lower bound for the cluster editing problem which can also be stated
as a decision version of the correlation clustering problem. In the p-cluster editing
problem, given a graph G and a budget B, and an integer p, the objective is to
decide whether G can be transformed into a union of p clusters (disjoint cliques)
using at most B edge additions and deletions. Assuming ETH, they showed that
there exists p = Θ(k ω )√for some 0 ≤ ω ≤ 1 such that there is no algorithm
that decides in time 2o( pB) · nO(1) whether G can be transformed into a union
of p cliques using at most B adjustments (edge additions and deletions). It is
not clear whether their exact reduction can be modified into an approximation
preserving reduction to obtain results similar to what we have here. Mazumdar
and Saha [18] studied correlation clustering problem in a similar setting where
edge similarity and dissimilarity information are assumed to be coming from two
distributions. Given such an input, they studied the cluster recovery problem in
SSAC framework, and gave upper and lower bounds on the query complexity.
Their lower bound results are information theoretic in nature. We are, however,
interested in the approximate solutions for the correlation clustering problem.
2
Query Lower Bounds
In this section, we obtain a lower bound on the number of same-cluster queries
that any FPTAS within the SSAC framework needs to make for the problem
MinDisAgree[k]. We derive a conditional lower bound for the minimum number
of queries under the Exponential Time Hypothesis (ETH) assumption. Some
such conditional lower bound results based on ETH can be found in [16]. We
prove the following main theorem in this section.
Theorem 5. If the Exponential Time Hypothesis (ETH) holds, then there exists
a constant δ > 0 such that any (1+δ)-approximation algorithm for MinDisAgree[k]
k
requires 2Ω( poly log k ) time.
The above theorem gives a proof of Theorem 3.
Proof (Proof of Theorem 3). Let us assume that there exists a query-FPTAS that
makes only o( polyklog k ) same-cluster queries. Then, by considering all possible
answers for these queries and picking the best solution, one can solve the problem
k
in 2o( poly log k ) time which contradicts Theorem 5.
In the remaining section, we give the proof of Theorem 5. First, we state the
ETH hypothesis. Our lower bound results are derived assuming this hypothesis.
Hypothesis 1 (Exponential Time Hypothesis (ETH)[13,14]): There does
not exist an algorithm that decides whether any 3-SAT formula with m
clauses is satisfiable with running time 2o(m) .
Since we would like to obtain lower bounds in the approximation domain, we
will need a gap version of the above ETH hypothesis. The following version of
the PCP theorem would be very useful in obtaining a gap version of ETH.
Theorem 6 (Dinur’s PCP Theorem [10]). For some constants ε, d > 0,
there exists a polynomial-time reduction that takes a 3-SAT formula ψ with m
clauses as input and produces one E3-SAT3 formula φ with m′ = O(mpoly log m)
clauses such that
– if ψ is satisfiable, then φ is satisfiable, and
3
Every clause in an E3-SAT formula has exactly 3 literals.
– if ψ is unsatisfiable, then val(φ) ≤ 1 − ε, and
– each variable in φ appears in at most d clauses.
where val(φ) is the maximum fraction of clauses of φ which are satisfiable by any
assignment.
The hypothesis below follows from ETH and the above Theorem 6, and will
be useful for our analysis.
Hypothesis 2: There exists constants ε, d > 0 such that the following
holds: There does not exist an algorithm that, given a E3-SAT formula
ψ with m clauses and each variable appearing in at most d clauses,
distinguishes whether ψ is satisfiable or val(ψ) ≤ (1 − ε), and runs in
m
time better than 2Ω ( poly log m ) .
The lemma given below trivially follows from Dinur’s PCP Theorem 6.
Lemma 1. If Hypothesis 1 holds, then so does Hypothesis 2.
We now give a reduction from the gap version of the E3-SAT problem to the
gap version of the NAE3-SAT problem. A problem instance of NAE3-SAT consists
of a set of clauses (each containing exactly 3 literals) and a clause is said to be
satisfied by an assignment iff at least one and at most two literals in the clause
is true (NAE stands for “Not All Equal"). For any instance φ, we define val′ (φ)
to be the maximum fraction of clauses that can be satisfied in the “not all equal"
sense by an assignment. Note that this is different from val(φ) which is equal to
the maximum fraction of clauses that can be satisfied (in the usual sense). First,
we reduce E3-SAT to NAE6-SAT and then NAE6-SAT to NAE3-SAT.
Lemma 2. Let 0 < ε < 1 and d > 1. There is a polynomial time reduction that
given an instance ψ of E3-SAT with m clauses with each variable appearing in at
most d clauses, produces an instance φ of NAE6-SAT with 4m clauses such that
1. If val(ψ) = 1, then val′ (φ) = 1, and
2. If val(ψ) ≤ (1 − ε), then val′ (φ) ≤ (1 − ε/4), and
3. Each variable in φ appears in at most 4d clauses.
Proof. We construct φ in the following manner: for every variable xi in ψ, we
introduce two variables yi and zi . We will use xi = 1 iff yi 6= zi for every i in our
reduction. For every clause (li , lj , lk ) (with li , lj , lk being literals), we introduce
the following four NAE clauses in φ:
(pi , qi , pj , qj , pk , qk ), (pi , qi , pj , qj , p̄k , q̄k ), (pi , qi , p̄j , q̄j , pk , qk ), (pi , qi , p̄j , q̄j , p̄k , q̄k )
For any index (say i), if li = xi (that is, the variable is in the positive form), then
pi = yi and qi = zi . On the other hand, if li = x̄i , then pi = yi and qi = z̄i . So
for example, for the clause (x2 , x̄7 , x9 ) in ψ, we have the following four clauses:
(y2 , z2 , y7 , z̄7 , y9 , z9 ), (y2 , z2 , y7 , z̄7 , ȳ9 , z̄9 ), (y2 , z2 , ȳ7 , z7 , y9 , z9 ), (y2 , z2 , ȳ7 , z7 , ȳ9 , z̄9 )
Note that property (3) of the lemma holds due to our construction. For property
(1), we argue that for any satisfying assignment for ψ, the assignment of variables
in φ as per the rule xi = 1 iff yi 6= zi is a satisfying assignment of ψ (in the NAE
sense). This is because for every literal l that makes a clause in ψ true, the two
corresponding copies p and q satisfies all the four clauses (in the NAE sense). For
property (2), we prove the contrapositive. Suppose there is an assignment to the
y, z variables in φ that satisfies at least (1 − ε/4) fraction of the clauses. We will
argue that the assignment to the variables of ψ as per the rule xi = 1 iff yi 6= zi
satisfies at least (1 − ε) fraction of clauses of ψ. First, note that for every set of
4 clauses in φ created from a single clause of ψ, either 3 of them are satisfied or
all four are satisfied (whatever the assignment of the variable be). Let m1 be the
number of these 4-sets where all 4 clauses are satisfied and let m2 be the number
of these 4-sets where 3 clauses are satisfied, where m = m1 + m2 . Then we have
4m1 + 3m2 ≥ (1 − ε/4) · (4m) which implies that m1 ≥ (1 − ε)m. Note that for
any of the 4-sets where all 4 clauses are satisfied, the corresponding clause in ψ
is satisfied with respect to the assignment as per rule xi = 1 iff yi 6= zi (since at
least one the p, q pairs will have opposite values). So, the fraction of the clauses
1
satisfied in ψ is at least m
m ≥ (1 − ε).
Lemma 3. Let 0 < ε < 1 and d > 1. There is a polynomial time reduction
that given an instance ψ of NAE6-SAT with m clauses and with each variable
appearing in at most d clauses, produces an instance φ of NAE3-SAT with 4m
clauses such that:
1. If val′ (ψ) = 1, then val′ (φ) = 1, and
2. If val′ (ψ) ≤ (1 − ε), then val′ (φ) ≤ (1 − ε/4).
3. Each variable in φ appears in at most max (d, 2) clauses.
Proof. For every clause Ci = (ai , bi , ci , di , ei , fi ) in ψ, we construct the following
four clauses in φ (let us call it a 4-set): (ai , bi , xi ), (x̄i , ci , yi ), (ȳi , di , zi ), (z̄i , ei , fi ),
introducing new variables xi , yi , zi . Property (3) trivially holds for this construction.
For every satisfying assignment for ψ, there is a way to set the clause variables
xi , yi , zi for every i such that all four clauses in the 4-set corresponding to
clause Ci are satisfied. So, property (1) holds. We show property (2) using
contraposition. Consider any assignment of φ that satisfies at least (1 − ε/4)
fraction of the clauses. Let mj denote the number of 4-sets such that as per
P4
this assignment j out of 4 clauses are satisfied. Then, we have j=0 j · mj ≥
P3
(1 − ε/4) · (4m). This implies that: 3 · j=0 mj + 4m4 ≥ (1 − ε/4) · (4m) which
implies that m4 ≥ (1−ε)m. Now, note that for any 4-set such that all four clauses
are satisfied, the corresponding clause in ψ is satisfied by the same assignment
to the variables. This implies that there is an assignment that makes at least
(1 − ε) fraction of clauses true in ψ.
We come up with the following hypothesis which holds given that Hypothesis
2 holds, and is crucial for our analysis.
Hypothesis 3: There exists constants ε, d > 0 such that the following
holds: There does not exist an algorithm that, given a NAE3-SAT formula
ψ with m clauses with each variable appearing in at most d clauses,
distinguishes whether val′ (ψ) = 1 or val′ (ψ) ≤ (1 − ε), and runs in time
m
better than 2Ω ( poly log m ) .
The lemma given below follows easily from the Lemmas 2 and 3 above.
Lemma 4. If Hypothesis 2 holds, then so does Hypothesis 3.
We now give a reduction from the gap version of NAE3-SAT to the gap version
of monotone NAE3-SAT that has no negative variables. Note that because of the
NAE (not all equal) property, setting all variables to 1 does not necessarily
satisfy the formula.
Lemma 5. Let 0 < ε < 1 and d > 1. There is a polynomial time reduction
that given an instance ψ of NAE3-SAT with m clauses and with each variable
appearing in at most d clauses, produces an instance φ of monotone NAE3-SAT
with O(m) clauses such that:
1. If val′ (ψ) = 1, then val′ (φ) = 1, and
ε
).
2. If val′ (ψ) ≤ (1 − ε), then val′ (φ) ≤ (1 − 1+12d
3. Each variable in φ appears in at most 4d clauses.
Proof. We construct φ in the following manner: Substitute all positive literals
of the variable xi with yi and all negative literals with zi for new variables yi , zi .
Also, for every variable xi , add the following 4d clauses:
{(yi , zi , tji ), (yi , zi , uji ), (yi , zi , vij ), (tji , uji , vij )}dj=1
where tji , uji , vij for 1 ≤ j ≤ d are new variables. Note that the only way to
satisfy all the above clauses is to have yi 6= zi . Let m′ denote the total number
of clauses in φ. So, m′ = m + 4dn. Also, from the construction, each variable in
φ appears in at most 4d clauses. This proves property (3). Property (1) follows
from the fact that for any satisfying assignment for ψ, there is a way to extend
this assignment to variables in φ such that all clauses are satisfied. For all i,
yi = xi and zi = x̄i . All the new variables t, u, v can be set so as to make all the
new clauses satisfied.
We argue property (2) using contraposition. Suppose there is an assignment
to variables in φ that makes at least (1 − ε/(1 + 12d)) fraction of clauses satisfied.
First, note that there is also an assignment that makes at least (1 − ε/(1 + 12d))
fraction of the clauses satisfied and in which for all i, yi 6= zi . This is because 3d
out of 4d of the following clauses can be satisfied when yi = zi :
{(yi , zi , tji ), (yi , zi , uji ), (yi , zi , vij ), (tji , uji , vij )}dj=1
However, if we flip one of yi , zi , then the number of above clauses satisfied can
be 4d and we might lose out on at most d clauses since a variable appears in
at most d clauses in ψ. Let m′ be the number of clauses corresponding to the
original clauses that are satisfied with this assignment. So, we have m′ + 4nd >
ε
(1 − (1+12d)
)(m + 4nd) which gives:
m′ > (1 − ε)m +
4ndε
12mdε
−
≥ (1 − ε)m
1 + 12d 1 + 12d
(since 3m ≥ n)
This completes the proof of the lemma.
We come up with the following hypothesis which holds given that Hypothesis
3 holds.
Hypothesis 4: There exists constants ε, d > 0 such that the following
holds: There does not exist an algorithm that, given a monotone NAE3SAT formula ψ with m clauses with each variable appearing in at most
d clauses, distinguishes whether val′ (ψ) = 1 or val′ (ψ) ≤ (1 − ε), and
m
runs in time better than 2Ω ( poly log m ) .
The lemma below follows easily from Lemma 5 mentioned in above.
Lemma 6. If Hypothesis 3 holds, then so does Hypothesis 4.
We provide a reduction from the gap version of monotone NAE3-SAT to a
gap version of 2-colorability of 3-uniform bounded degree hypergraph.
Lemma 7. Let 0 < ε < 1 and d > 1. There exists a polynomial time reduction
that given a monotone NAE3-SAT instance ψ with m clauses and with every
variable appearing in at most d clauses, outputs an instance H of 3-uniform
hypergraph with O(m) vertices and hyperedges and with bounded degree d such
that if ψ is satisfiable, then H is 2-colorable, and if at most (1 − ε)-fraction
of clauses of ψ are satisfiable, then any 2-coloring of H would have at most
(1 − ε)-fraction of edges that are bichromatic.
Proof. The reduction constructs a hypergraph H(V, E) as follows. The set of
vertices V correspond to the set of variables (all of them positive literals) of the
monotone NAE3-SAT instance ψ. The set of edges E correspond to the set of
clauses all of which have 3 literals, and therefore every hyperedge is of size 3.
The resulting hypergraph is 3-uniform, and since every variable appears in at
most d clauses, the hypergraph H is of bounded degree d, and |V | = O(m) and
|E| = O(m). If there exists a satisfying assignment for ψ, then every edge in H
is bichromatic and the hypergraph would be 2-colorable, and if at most (1 − ε)fraction of clauses are satisfiable by any assignment, then at most (1−ε)-fraction
of edges of H are bichromatic.
We come up with the following hypothesis which holds given that Hypothesis
4 holds.
Hypothesis 5: There exists constants ε, d > 0 such that the following
holds: There does not exist an algorithm that, given a 3-uniform hypergraph
H with m vertices and where every vertex has degree at most d, distinguishes
whether H is bichromatic or at most (1−ε)-fraction of edges are bichromatic,
m
and runs in time better than 2Ω ( poly log m ) .
The lemma below follows easily from Lemma 7 above.
Lemma 8. If Hypothesis 4 holds, then so does Hypothesis 5.
We now give a reduction from 2-colorability in 3-uniform hypergraph H with
constant bounded degree to a correlation clustering instance on a complete graph
G. We use the reduction as given in [9] for our purposes.
Lemma 9 ([9]). Let ε, d > 0. There is a polynomial-time reduction that given a
3-uniform hypergraph H(V, E) with m vertices and where each vertex appears in
at most d hyperedges, outputs an instance of the correlation clustering problem
where the graph G(V ′ , E ′ ) has N = O(m) vertices and M = 2N edges with edges
in E ′ are labeled as ‘positive’ and all the other edges in the complete graph on
V ′ vertices are labeled as ‘negative’ such that the following holds:
1. If H is 2-colorable, then the cost of the optimal correlation clustering is
M − N , and
2. If at most (1 − ε)-fraction of hyperedges of H are bi-chromatic, then the
optimal cost of correlation clustering is at least M − (1 − δ)N , where δ is
some constant.
We come up with the following hypothesis which holds given that Hypothesis
5 holds.
Hypothesis 6: There exists constants ε > 0 such that the following
holds: There does not exist a (1 + ε)-factor approximation algorithm
for the MinDisAgree[k] problem that runs in time better than poly(n) ·
k
2Ω( poly log k ) .
The lemma below follows easily from Lemma 9 given above.
Lemma 10. If Hypothesis 5 holds, then so does Hypothesis 6.
Finally, the proof of Theorem 5 follows from chaining together lemmas 1, 4,
6, 8, and 10.
3
Algorithms for MaxAgree[k] and MinDisAgree[k] in
SSAC Framework
In this section, we give (1 + ε)-approximation algorithms for the MaxAgree[k]
and MinDisAgree[k] problems within the SSAC framework for any ε > 0.
3.1
MaxAgree[k]
In this section, we will discuss a query algorithm that gives (1+ε)-approximation
to the MaxAgree[k] problem. The algorithm that we will discuss is closely related
to the non-query algorithm for MaxAgree[k] by Giotis and Guruswami. See
Algorithm MaxAg(k, ε) in [12]. In fact, except for a few changes, this section
will look extremely similar to Section 3 in [12]. Given this, it will help if we
mention the high-level idea of the Giotis-Guruswami algorithm and point out
the changes that can be made within the SSAC framework to obtain the desired
result. The algorithm of Giotis and Guruswami proceeds in m iterations, where
m = O(1/ε). The given dataset V is partitioned into m equal parts V 1 , . . . , V m ,
and in the ith iteration, points in V i are assigned to one of the k clusters. In order
to cluster V i in the ith iteration, the algorithm samples a set of data points S i ,
and for all possible k-partitions of S i , it checks the agreement of a point v ∈ V i
with the k clusters of S i . Suppose for a particular clustering S1i , . . . , Ski of S i , the
agreement of vertices in V i is maximised. Then the vertices in V i are clustered
by placing them into the cluster that maximises their agreement with respect to
S1i , . . . , Ski . Trying out all possible k-partitions of S i is an expensive operation
i
in the Giotis-Guruswami algorithm (since the running time becomes Ω(k |S | )).
This is where the same-cluster queries help. Instead of trying out all possible
k-partitions of S i , we can make use of the same-cluster queries to find a single
appropriate k-partition of S i in the ith iteration. This is the clustering that
matches the “hybrid” clustering of Giotis and Guruswami. So, the running time
i
of the ith iteration improves from O(k |S | ) to O(k · |S i |). Moreover, the number
th
of same-cluster queries made in the i iteration is k · |S i |, thus making the total
number of same-cluster queries to be O( kε · |S i |). The theorem is given below.
The details of the proof of this theorem is not given since it trivially follows from
Giotis and Guruswami (see Theorem 3.2 in [12]).
Theorem 7. There is a query algorithm QueryMaxAg that behaves as follows:
On input ε, δ and a labelling L of the edges of a complete graph G with n vertices,
with probability at least (1 − δ), algorithm QueryMaxAg outputs a k clustering
of the graph such that the number of agreements induced by this k-clustering is
at least OP T − εn2 /2, where OP T is the optimal number of agreements induced
k
by any k-clustering of G. The running time of the algorithm is O nk
ε3 log ε2 δ .
Moreover, the number of same-cluster queries made by QueryMaxAg is O εk3 log εk2 δ .
Using the simple observation that OP T ≥ n2 /16 (see proof of Theorem 3.1
in [12]), we get that the above query algorithm gives (1 + ε)-approximation
guarantee in the SSAC framework.
3.2
MinDisAgree[k]
In this section, we provide a (1+ε)-approximation algorithm for the MinDisAgree[k]
for any small ε > 0. Giotis and Guruswami [12] provided a (1+ε)-approximation
algorithm for MinDisAgree[k]. In this work, we extend their algorithm to make it
work in the SSAC framework with the aid of same-cluster queries, and thereby
improve the running time of the algorithm considerably. Our query algorithm
will be closely based on the non-query algorithm of Giotis and Guruswami. In
fact, except for a small (but crucial) change, the algorithms are the same. So,
we begin by discussing the main ideas and the result by Giotis and Guruswami.
Lemma 11 (Theorem 4.7 in [12]). For every k ≥ 2 and ε > 0, there is a (1 +
k
2
ε)-approximation algorithm for MinDisAgree[k] with running time nO(9 /ε ) log n.
The algorithm by Giotis and Guruswami builds on the following ideas. First,
from the discussion in the previous section, we know that there is a FPTAS
within SSAC framework for MaxAgree[k]. Therefore, unless OP T , the optimal
value for MinDisAgree[k] is small (OPT= γn2 , for some small γ > 0), the
complement solution for MaxAgree[k] would give a valid (1 + ε)-approximate
solution for MinDisAgree[k]. Since OP T is small, this implies that the optimal
value for MaxAgree[k] is large which means that for any random vertex v in graph
G = (V, E), a lot of edges incident on v agree to the optimal clustering. Suppose
we are given a random subset S ⊆ V of vertices that are optimally clustered
S = {S1 , . . . , Sk }, and let us assume that S is sufficiently large. Since most of
the edges in E are in agreement with the optimal clustering, we would be able
to assign vertices in V \ S to their respective clusters greedily. For any arbitrary
v ∈ V \ S, assign v to Si for which the number of edges that agree is maximized.
Giotis and Guruswami observed that clustering vertices in V \ S in this manner
would work with high probability when these vertices belong to large clusters.
For vertices in small clusters, we may not be able to decide assignments with
high probability. They carry out this greedy assignment of vertices in V \ S
into clusters S1 , . . . , Sk , and filter out clusters that are sufficiently large and
recursively run the same procedure on the union of small clusters.
For any randomly sampled subset S ⊆ V of vertices, Giotis and Guruswami
try out all possible ways of partitioning S into k clusters in order to partition
S optimally into k clusters S1 , . . . , Sk . This ensures that at least one of the
partitions matches the optimal partition. However, this exhaustive way of partitioning
imposes huge burden on the running time of the algorithm. In fact, their algorithm
k
2
runs in nO(9 /ε ) log n time. Using access to the same-cluster query oracle, we
can obtain a significant reduction in the running time of the algorithm. We query
the oracle with pairs of vertices in S and since query answers are assumed to be
consistent with some unique optimal solution, optimal k clustering of vertices
in S is accomplished using at most k|S| same-cluster queries. Once we have a kpartitioning of sample S that is consistent with the optimal k-clusters, we follow
the remaining steps of [12]. The same approximation analysis as in [12] follows
for the query algorithm. For completeness we give the modified algorithm in
Figure 3.1. Let oracle A take any two vertices u and v as input and return ‘Yes’
if they belong to the same cluster in optimal clustering, and ‘No’ otherwise.
Here is the main theorem giving approximation guarantee of the above algorithm.
As stated earlier, the proof follows from the proof of a similar theorem (Theorem
4.7 in [12]) by Giotis and Guruswami.
Theorem 8. Let 0 < ε ≤ 1/2. For any input labelling, QueryMinDisAgree(k, ε/4)
returns a k-clustering with the number of disagreements within a factor of (1 + ε)
of the optimal. 4
4
Readers familiar with [12] will realise that the statement of the theorem is slightly
different from statement of the similar theorem (Theorem 13) in [12]. More
QueryMinDisAgree(k, α)
n
Input: Labeling L :
2
A : n2 → {Yes,No}.
Output: A k-clustering of G
1
Constants: c1 = 20
→
{+, −} of edges of graph G(V, E), Oracle
(1) If k = 1, return 1-clustering.
α2 c2
(2) Run QueryMaxAg on input L with accuracy 32k41 to obtain k-clustering ClusM ax.
5 log n
c1 α
independently and uniformly at
(3) Set β = 16k
2 . Pick sample S ⊆ V of size
β2
random with replacement from V .
(4) Optimally cluster S = {S1 , . . . , Sk } by making same-cluster queries to oracle A.
(5) Let Cj = Sj for 1 ≤ j ≤ k.
(6) For each u ∈ V \ S
(6.1) ∀i = 1, . . . , k: Let liu be the number of edges which agree between u and
nodes in Si .
(6.2) Let ju = arg maxi liu be the index of the cluster which maximizes the above
quantity.
(6.3) Cju = Cju ∪ {u}.
n
(7) Let the set of large and small clusters be Large = {j : 1 ≤ j ≤ k, |Sj | ≥ 2k
} and
Small = [k] \ Large.
(8) Let l = |Large|, and s = k − l.
(9) Cluster W = ∪j∈Small Sj into s clusters {W1 , . . . , Ws } using recursive calls to
QueryMinDisAgree(s, α).
(10) Let ClusM in be clustering obtained by k clusters {Sj }j∈Large and {Wt }{1≤t≤s} .
(11) Return the better of ClusM in and ClusM ax
Algorithm 3.1: Query version of the algorithm by Giotis and Guruswami.
Even though the approximation analysis of the query algorithm remains
the same as the non-query algorithm of Giotis and Guruswami, the running
time analysis changes significantly. Let us write a recurrence relation for the
running time of our recursive algorithm. Let T (k) denote the running time of
the algorithm when n node graph is supposed to be clustered into k clusters with
a given precision parameter α. Using the results of the previous subsection, the
13
running time of step (2) is O( nk α6log k ). The running time for partitioning the set
S is given by k|S| which is O( k
is
5
n
O( nk αlog
).
2
5
log n
α2 ).
Steps (6-8) would cost O(nk|S|) time which
So, the recurrence relation for the running time may be written
13
14
k log n
k log n
as T (k) = T (k − 1) + O( nk log
). This simplifies to O( nk log
). As
α6
α6
far as the same-cluster queries are concerned, we can write a similar recurrence
14
13
relation. Q(k) = Q(k − 1) + O( k logα6k log n ) which simplifies to O( k logα6k log n ).
This completes the proof of Theorem 2.
4
Conclusion and Open Problems
In this work, we give upper and lower bounds on the number of same-cluster
queries to obtain an efficient (1 + ε)-approximation for correlation clustering
on complete graphs in the SSAC framework of Ashtiani et al. Our lower bound
results are based on the Exponential Time Hypothesis (ETH). It is an interesting
open problem to give unconditional lower bounds for these problems. Another
interesting open problem is to design query based algorithms with faulty oracles.
This setting is more practical since in many contexts it may not be known
whether any two vertices belong to the same optimal cluster with high confidence.
Mitzenmacher and Tsourakakis [19] designed a query based algorithm for clustering
where query oracle, similar to our model, answers whether any two vertices
belong to the same optimal cluster or otherwise but these answers may be wrong
with some probability less than 12 . For k = 2, we can use their algorithm to
obtain (1 + ε)-approximation for MinDisAgree[k] with faulty oracle. However, for
k > 2 they needed stronger query model to obtain good clusterings. Designing
an efficient (1+ε)-approximation algorithm for MinDisAgree[k] with faulty oracle
is an interesting open problem.
References
1. Nir Ailon, Anup Bhattacharya, Ragesh Jaiswal, and Amit Kumar. Approximate
clustering with same-cluster queries. CoRR, abs/1704.01862, 2017.
2. Haris Angelidakis, Konstantin Makarychev, and Yury Makarychev. Algorithms
for stable and perturbation-resilient problems. In ACM SIGACT Symposium on
Theory of Computing, pages 438–451, 2017.
specifically, the claim is about the function call with ε/4 as a parameter rather than
ε. This is done to allow the recursive call in step (9) to be made with same value
of precision parameter as the initial call. This does not change the approximation
analysis but is crucial for our running time analysis.
3. Hassan Ashtiani, Shrinu Kushagra, and Shai Ben-David. Clustering with samecluster queries. In NIPS, pages 3216–3224, 2016.
4. Pranjal Awasthi, Maria-Florina Balcan, and Konstantin Voevodski.
Local
algorithms for interactive clustering. In ICML, pages 550–558, 2014.
5. Maria-Florina Balcan and Avrim Blum. Clustering with interactive feedback. In
International Conference on Algorithmic Learning Theory, pages 316–328, 2008.
6. Maria-Florina Balcan, Avrim Blum, and Anupam Gupta. Clustering under
approximation stability. Journal of the ACM (JACM), 60(2):8, 2013.
7. Maria Florina Balcan and Yingyu Liang. Clustering under perturbation resilience.
SIAM Journal on Computing, 45(1):102–155, 2016.
8. Nikhil Bansal, Avrim Blum, and Shuchi Chawla. Correlation clustering. Machine
Learning, 56(1-3):89–113, 2004.
9. Moses Charikar, Venkatesan Guruswami, and Anthony Wirth. Clustering with
qualitative information. Journal of Computer and System Sciences, 71(3):360–383,
2005.
10. Irit Dinur. The PCP theorem by gap amplification. J. ACM, 54(3):12, 2007.
11. Fedor V Fomin, Stefan Kratsch, Marcin Pilipczuk, Michał Pilipczuk, and Yngve
Villanger. Tight bounds for parameterized complexity of cluster editing with a
small number of clusters. Journal of Computer and System Sciences, 80(7):1430–
1447, 2014.
12. Ioannis Giotis and Venkatesan Guruswami. Correlation clustering with a fixed
number of clusters. In ACM-SIAM symposium on Discrete algorithm, pages 1167–
1176, 2006.
13. Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-sat. Journal
of Computer and System Sciences, 62(2):367 – 375, 2001.
14. Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems
have strongly exponential complexity? Journal of Computer and System Sciences,
63(4):512 – 530, 2001.
15. Konstantin Makarychev, Yury Makarychev, and Aravindan Vijayaraghavan.
Correlation clustering with noisy partial information. In COLT, pages 1321–1342,
2015.
16. Pasin Manurangsi. Almost-polynomial ratio eth-hardness of approximating densest
$k$-subgraph. CoRR, abs/1611.05991, 2016.
17. Claire Mathieu and Warren Schudy. Correlation clustering with noisy input. In
ACM-SIAM Symposium on Discrete Algorithms, pages 712–728, 2010.
18. Arya Mazumdar and Barna Saha. Query complexity of clustering with side
information. arXiv preprint arXiv:1706.07719, 2017.
19. Michael Mitzenmacher and Charalampos E Tsourakakis. Predicting signed edges
with o(n1+o(1) log n) queries. arXiv preprint arXiv:1609.00750, 2016.
20. Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and
Yu Xia. Efficient clustering with limited distance information. In Conference on
Uncertainty in Artificial Intelligence, pages 632–640, 2010.
| 8cs.DS
|
Architecture of Environmental Risk Modelling: for a faster
and more robust response to natural disasters
Dario Rodriguez-Aseretto
a,1 ,
Christian Schaerer b , Daniele de Rigo
a,c
a
b
European Commission, Joint Research Centre, Institute for Environment and
Sustainability, Ispra (VA), Italy.
Polytechnic School, National University of Asuncion, San Lorenzo, Central, Paraguay.
c Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria,
Milano, Italy.
arXiv:1409.7966v1 [cs.CE] 28 Sep 2014
Abstract
Demands on the disaster response capacity of the European Union are likely
to increase, as the impacts of disasters continue to grow both in size and
frequency. This has resulted in intensive research on issues concerning
spatially-explicit information and modelling and their multiple sources of
uncertainty. Geospatial support is one of the forms of assistance frequently
required by emergency response centres along with hazard forecast and event
management assessment. Robust modelling of natural hazards requires dynamic simulations under an array of multiple inputs from different sources.
Uncertainty is associated with meteorological forecast and calibration of the
model parameters. Software uncertainty also derives from the data transformation models (D-TM) needed for predicting hazard behaviour and its
consequences. On the other hand, social contributions have recently been
recognized as valuable in raw-data collection and mapping efforts traditionally dominated by professional organizations. Here an architecture overview
is proposed for adaptive and robust modelling of natural hazards, following
the Semantic Array Programming paradigm to also include the distributed
array of social contributors called Citizen Sensor in a semantically-enhanced
strategy for D-TM modelling. The modelling architecture proposes a multicriteria approach for assessing the array of potential impacts with qualitative rapid assessment methods based on a Partial Open Loop Feedback
Control (POLFC) schema and complementing more traditional and accurate a-posteriori assessment. We discuss the computational aspect of environmental risk modelling using array-based parallel paradigms on High
Performance Computing (HPC) platforms, in order for the implications of
urgency to be introduced into the systems (Urgent-HPC).
Keywords: Geospatial, Integrated Natural Resources Modelling and Management,
Semantic Array Programming, Warning System, Remote Sensing, Parallel Application, High Performance Computing, Partial Open Loop Feedback Control
1
E-mail Corresponding Author: [email protected]
1
Cite as:
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental
Risk Modelling: for a faster and more robust response to natural disasters.
3rd Conference of Computational Interdisciplinary Sciences, Asunción, Paraguay
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
1. INTRODUCTION: CONTEXT, PITFALLS AND THE SCIENCEPOLICY INTERFACE
Europe experienced a series of particularly severe disasters in the recent years
[1, 2], with worrying potential impacts of similar disasters under future projected
scenarios of economy, society and climate change [3, 4]. They range from flash
floods [5, 6, 7] and severe storms in Western Europe with an expected increasing
intensity trend [8], large-scale floods in Central Europe [9], volcanic ash clouds
[10, 11, 12] (e.g. after the Eyjafjallajkull eruption), large forest fires in Portugal and
Mediterranean countries [13, 14]. Biological invasions such as emerging plant pests
and diseases have the potential to further interact e.g. with wildfires [15] and to
impact on ecosystem services [16] and economy with substantial uncertainties [17].
It should be underlined that these recent highlights are set in the context of
systemic changes in key sectors [18, 19, 20] which overall may be expected to at
least persist in the next decades. As a general trend, demands on the EU’s resilience in preparedness and disaster response capacity are likely to increase, as
the impacts of disasters continue to grow both in size and frequency, even considering only the growing exposure (societal factors) [21, 22]. The aforementioned
examples of disturbances are often characterised by non-local system feedbacks
and off-site impacts which may connect multiple natural resources (system of systems) [23, 24, 25]. In this particular multifaceted context [26, 27, 28], landscape
[29] and ecosystem dynamics show intense interactions with disturbances [30].
As a consequence, classical disciplinary and domain-specific approaches which
might be perfectly suitable at local-scale may easily result in unacceptable simplifications within a broader context. A broad perspective is also vital for investigating future natural-hazard patterns at regional/continental scale and adapting preparedness planning [31, 32, 33]. The complexity and uncertainty associated with
these interactions – along with the severity and variety of the involved impacts [34]
– urge robust, holistic coordinated [35] and transparent approaches [36, 37]. At
the same time, the very complexity itself of the control-system problems involved
[38, 39, 40] may force the analysis to enter into the region of deep-uncertainty [41].
The mathematization of systems in this context as a formal control problem
should be able to establish an effective science-policy interface, which is not a trivial aspect. This is easily recognised even just considering the peculiarities – which
have been well known for a long time – of geospatially-aware environmental data
[42] and decision support systems [43, 44, 45], their entanglement with growingly
complex ICT aspects [46, 47] and their not infrequent cross-sectoral characterisation. Several pitfalls may degrade the real-world usefulness of the mathematization/implementation process. While it is relatively intuitive how a poor mathematization with a too simplistic approach might result in a failure, subtle pitfalls
may lie even where an “appropriately advanced” theoretical approach is proposed.
2
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
Mathematization should resist silo thinking [48, 49] temptations such as academic
solution-driven pressures [37, 50] to force the problem into fashionable “hot topics”
of control theory: robust approximations of the real-world broad complexity may
serve egregiously instead of state-of-art solutions of oversimplified problems.
Other long-lasting academic claims are “towards” fully automated scientific
workflows in computational science, maybe including self-healing and self-adapting
capabilities of the computational models implementing the mathematization. These
kinds of claims might easily prompt some irony [51] among experienced practitioners in wide-scale transdisciplinary modelling for environment (WSTMe, [52]) as a
never-ending research Pandora’s box with doubtful net advantages [53]. Complex,
highly uncertain and sensitive problems for policy and society, as WSTMe problems typically are, will possibly never be suitable for full automation: even in this
family of problems, “humans will always be part of the computational process”
[54] also for vital accountability aspects [55].
While a certain level of autonomic computing [56] capabilities might be essential for the evolvability and robustness of WSTMe (in particular, perhaps, a higher
level of semantic awareness in computational models and a self-adapting ability
to scale up to the multiple dimensions of the arrays of data/parameters; see next
section), here the potential pitfall is the illusion of fully automating WSTMe. The
domain of applicability of this puristic academic silo – although promising for relatively simple, well-defined (and not too policy-sensitive) case studies – might be
intrinsically too narrow for climbing up to deal with the wicked problems typical
of complex environmental systems [57, 58, 59].
The discussed pitfalls might deserve a brief summary. First, perhaps, is the risk
of “solving the wrong problem precisely” [60] by neglecting key sources of uncertainty – e.g. unsuitable to be modelled within the “warmly supported” solution
of a given research group. During emergency operations, the risks of providing
a “myopic decision support” should be emphasised; i.e. suggesting inappropriate
actions [61] – e.g. inaction or missing precaution – due to the potential overwhelming lack of information [62] or the oversimplification/underestimation of potential
chains of impacts due to the lack of computational resources for a decent (perhaps
even qualitative and approximate) rapid assessment of them.
Overcoming these pitfalls is still an open issue. Here, we would like to contribute to the debate by proposing the integrated use of some mitigation approaches. We focus on some general aspects of the modelling architecture for the
computational science support, in order for emergency-operators, decision-makers,
stakeholders and citizens to be involved in a participatory [63] information and decision support system which assimilates uncertainty and precaution [64, 57]. Since
no silver bullet seems to be available for mitigating the intrinsic wide-extent of complexity and uncertainty in environmental risk modelling, an array of approaches is
3
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
integrated and the computational aspects are explicitly connected with the supervision and distributed interaction of human expertise. This follows the idea that
the boundary between classical control-theory management strategies for natural
resources and hazards (driven by automatic control problem formulations – “minimize the risk score function”) and scenario modelling under deep-uncertainty (by
e.g. merely supporting emergency-operators, decision-makers and risk-assessors
with understandable information – “sorry, no such thing as a risk score function
can be precisely defined”) is fuzzy. Both modelling and management aspects may
be computationally intensive and their integration is a transdisciplinary problem
(integrated natural resources modelling and management, INRMM [65]).
2. ENVIRONMENTAL RISK MODELLING - ARCHITECTURE
Figure 1 illustrates a general modelling conceptualization where the interactions
among natural hazard behaviour, related transdisciplinary impacts, risk management and control strategies are taken into account. The special focus on the many
sources of uncertainty [36] leads to a robust semantically-enhanced modelling architecture based on the paradigm of Semantic Array Programming (SemAP) [66, 67],
with an emphasis on the array of input, intermediate and output data/parameters
and the array of data-transformation modules (D-TM) dealing with them.
ζ
Arrays of hazard models hjf (·), dynamic information forecasts X ζX (i.e. meteorology) and static parametrisation θ ζθ (i.e. spatial distribution of land cover)
are considered. Their multiplicity derives from the many sources on uncertainty
ζ = {ζf , ζX , ζθ } which affect their estimation (or implementation, for the D-TM
ζ
ζ
software modules fi f (·) which are the building blocks of the hazard models hjf (·)).
Furthermore, during emergency modelling support the lack of timely and accurate monitoring systems over large spatial extents (e.g. at the continental scale)
may imply a noticeable level of uncertainty to affect possibly even the location
of natural hazards (geoparsing [68] uncertainty). This peculiar information gap
may be mitigated by integrating remote sensing (e.g. satellite imagery) with a
distributed array of social contributors (Citizen Sensor [69, 70, 71]), exploiting
mobile applications (Apps) and online social networks [72]. Remote sensing and
the Citizen Sensor are here designed to cooperate by complementing accurate (but
often less timely) geospatial information with distributed alert notifications from
citizens, which might be timely but not necessarily accurate. Their safe integration
implies the supervision of human expertise, even if the task may be supported by
automatic tools [73]. Assessing the evolution in the timespan U t = [tbegin , tend ] of a
certain hazard event for the associated array of impacts C k,t may be also complex
(e.g. [41, 74, 75]). In particular, the array of impacts is often irreducible to a
unidimensional quantity (e.g. monetary cost) [76, 77].
4
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
Figure 1 - Modular architecture for environmental risk modelling. Based on
Urgent HPC, it follows the Semantic Array Programming paradigm (image adapted
from [36, 52]) integrating as inputs remote sensing, meteo data and the Citizen Sensor.
The analysis of non-trivial systems subject to environmental risk and natural resources management may naturally lead to multi-objective (multi criteria) control
problems, which might benefit from advanced machine learning techniques for
5
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
mitigating the involved huge computational costs [78]. Indeed, the multiplicity
of modelling dimensions (states; controls; uncertainty-driven arrays of parameters and scenarios; arrays of D-TM modules to account for software uncertainty)
may easily lead to an exponential increase of the required computational processes
(the so called “curse of dimensionality”). A viable mitigation strategy might be
offered by HPC tools (such as Urgent HPC [79, 80, 81]) in order to sample highdimensional modelling space with a proper method.
Box 1 – In a nutshell.
Context Demands on the EU’s resilience in preparedness and disaster response capacity
are likely to increase, as the impacts of disasters continue to grow.
• Classical disciplinary and domain-specific approaches which might be perfectly suitable at local-scale may result in unacceptable simplifications in a broader context.
Pitfalls Mathematization of systems in this contex as a formal control problem should be
able to establish an effective science-policy interface. Academic silo thinking should
stop advertising solution-driven oversimplification to fit control theory “hot topics”.
• Although in this family of problems “humans will always be part of the computational process” (despite any academic potential illusion of fashionable full automation),
• evolvability (for adapting models to new emerging needs and knowledge) and robustness (for supporting uncertainty-aware decision processes) would still need
. a higher level of semantic awareness in computational models and
. a self-adapting ability to scale up to the multiple dimensions of the arrays of
data/parameters.
Multiplicity: uncertainty and complexity In this context, the boundary between classical control-theory management strategies for natural resources and hazards and scenario modelling under deep-uncertainty is fuzzy (INRMM).
• A key aspect of soundness relies on explicitly considering the multiple dimensions
of the problem and the array of uncertainties involved.
• As no silver bullet seems to be available for reliably attacking this amount of
uncertainty and complexity, an integration of methods is proposed.
Mitigating with an integrated approach Array programming is well-suited for easily
managing a multiplicity of arrays of hazard models, dynamic input information, static
parametrisation and the distribute array of social contributions (Citizen Sensor).
• Array-based abstract – thus better scalable – modularisation of the data-transformations (D-TM), and a semantically-enhanced design of the D-TM structure and
interactions (Semantic Array Programming) is proposed to consider also the array
of uncertainties (data, modelling, geoparsing, software uncertainty) and the array of
criteria to assess the potential impacts associated with the hazard scenarios.
• The unevenly available information during an emergency event may be efficiently
exploited by means of a POLFC schema.
• Its demanding computations may become affordable during an emergency event
with an appropriate array-based parallelisation strategy within Urgent-HPC.
6
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
SemAP can simplify WSTMe modelling of nontrivial static [82, 83] and dynamic
[41, 75, 84] geospatial quantities. Under the SemAP paradigm, the generic i-th
D-TM module Yi = fi (θi , Xi ) is subject to the semantic checks sem i as pre-, postconditions and invariants on the inputs θi , Xi , outputs Yi and the D-TM itself fi (·).
The control problem is associated with the unevenly available dynamic updates of
field measurements and other data related to an on-going hazard emergency. An
Emergency Manager may thus be interested in assessing the best control strategy
ut (·) given a set of impacts and their associated costs as they can be approximately
estimated (rapid assessment) with the currently available data. This data-driven
approach can be implemented as Partial Open Loop Feedback Control (POLFC)
approach [85] for minimizing the overall costs associated with the natural hazard
event, from the time t ∈ U t onwards:
1, t 2, t
C C · · · C k, t · · · C n, t
(1)
ut (·) = arg minu ∈ Utu, t
end
where the k-th cost C k, t is linked to the corresponding impact assessment criterion. This POLFC schema within the SemAP paradigm may be considered a
semantically-enhanced dynamic data-driven application system (DDDAS) [41, 75,
84]. Finally, the Emergency Manager may communicate the updated scenarios
of the emergency evolution (by means of geospatial maps and other executive
summary information) in order for decision-makers and stakeholders to be able to
assess the updated multi-criteria pattern of costs and the preferred control options.
This critical communication constitutes the science-policy interface and must be as
supportive as possible. It is designed to exploit web map services (WMS) [86, 87]
(on top of the underpinning free software for WSTMe, e.g. [88]) which may be
accessed in a normal browser or with specific Apps for smart-phones [73].
3. CONCLUDING REMARKS
NSF Cyberinfrastructure Council report reads: While hardware performance has
been growing exponentially - with gate density doubling every 18 months, storage capacity every 12 months, and network capability every 9 months - it has
become clear that increasingly capable hardware is not the only requirement for
computation-enabled discovery. Sophisticated software, visualization tools, middleware and scientific applications created and used by interdisciplinary teams are
critical to turning flops, bytes and bits into scientific breakthroughs [89]. Transdisciplinary environmental problems such as the ones dealing with complexity
and deep-uncertainty in supporting natural-hazard emergency might appear as
seemingly intractable [90]. Nevertheless, approximate rapid-assessment based on
computationally intensive modelling may offer a new perspective at least able
to support emergency operations and decision-making with qualitative or semi7
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
quantitative scenarios. Even a partial approximate but timely investigation on
the potential interactions of the many sources of uncertainty might help emergency
managers and decision-makers to base control strategies on the best available –
although typically incomplete – sound scientific information. In this context, a
key aspect of soundness relies on explicitly considering the multiple dimensions
of the problem and the array of uncertainties involved. As no silver bullet seems
to be available for reliably attacking this amount of uncertainty and complexity,
an integration of methods is proposed, inspired by their promising synergy. Array
programming is perfectly suited for easily managing a multiplicity of arrays of hazard models, dynamic input information, static parametrisation and the distribute
array of social contributions (Citizen Sensor). The transdisciplinary nature of
complex natural hazards – their need for an unpredictably broad and multifaceted
readiness to robust scalability – may benefit (1) from a disciplined abstract modularisation of the data-transformations which compose the models (D-TM), and
(2) from a semantically-enhanced design of the D-TM structure and interactions.
These two aspects define the Semantic Array Programming (SemAP, [66, 67])
paradigm whose application – extended to geospatial aspects [52] – is proposed to
consider also the array of uncertainties (data, modelling, geoparsing, software uncertainty) and the array of criteria to assess the potential impacts associated with
the hazard scenarios. The unevenly available information during an emergency
event may be efficiently exploited by means of a partial open loop feedback control (POLFC, [85]) schema, already successfully tested in this integrated approach
[41, 75, 84] as a promising evolution of adaptive data-driven strategies [91]. Its
demanding computations may become affordable during an emergency event with
an appropriate array-based parallelisation strategy within Urgent-HPC.
References
[1] SIPPEL, S., OTTO, F.E.L., 2014. Beyond climatological extremes - assessing how the odds
of hydrometeorological extreme events in South-East Europe change in a warming climate.
Climatic Change: 1-18.
[2] CIRELLA, G.T., et al, 2014. Natural hazard risk assessment and management methodologies review: Europe. In: Linkov, I. (Ed.), Sustainable Cities and Military Installations.
NATO Science for Peace and Security Series C: Environmental Security. Springer Netherlands, pp. 329-358.
[3] CISCAR, J. C., et al, 2013. Climate impacts in Europe: an integrated economic assessment.
In: Impacts World 2013 - International Conference on Climate Change Effects. Potsdam
Institute for Climate Impact Research (PIK) e. V., pp. 87-96.
[4] CISCAR, J. C., et al, 2014. Climate Impacts in Europe - The JRC PESETA II project.
Vol. 26586 of EUR - Scientific and Technical Research. Publ. Off. Eur. Union, 155 pp.
[5] DANKERS, R., FEYEN, L., 2008. Climate change impact on flood hazard in Europe: An assessment based on high-resolution climate simulations. J.Geophys.Res.113(D19):D19105+.
[6] GAUME, E., et al, 2009. A compilation of data on European flash floods. Journal of
Hydrology 367(1-2): 70-78.
8
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
[7] MARCHI, L., et al, 2010. Characterisation of selected extreme flash floods in Europe and
implications for flood risk management. Journal of Hydrology 394(1-2): 118-133.
[8] FESER, F., et al, 2014. Storminess over the North Atlantic and Northwestern Europe - a
review. Quarterly Journal of the Royal Meteorological Society.
[9] JONGMAN, B., et al, 2014. Increasing stress on disaster-risk finance due to large floods.
Nature Climate Change 4 (4): 264-268.
[10] SELF, S., 2006. The effects and consequences of very large explosive volcanic eruptions.
Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering
Sciences 364(1845): 2073-2097.
[11] SWINDLES, G. T., et al, 2011. A 7000 yr perspective on volcanic ash clouds affecting
northern Europe. Geology 39(9): 887-890.
[12] GRAMLING, C., 2014. As volcano rumbles, scientists plan for aviation alerts. Science
345(6200): 990.
[13] ALLARD, G., et al, 2013. State of Mediterranean forests 2013. FAO, 177 pp.
[14] SCHMUCK, G., et al, 2014. Forest Fires in Europe, Middle East and North Africa 2013.
Publications Office of the European Union. 107 pp.
[15] NIJHUIS, M., 2012. Forest fires: Burn out. Nature 489 (7416), 352-354.
[16] BOYD, I.L., et al, 2013. The consequence of tree pests and diseases for ecosystem services.
Science 342 (6160): 1235773+.
[17] VENETTE, R.C., et al, 2012. Summary of the international pest risk mapping workgroup
meeting sponsored by the cooperative research program on biological resource management
for sustainable agricultural systems. In: 6th International Pest Risk Mapping Workgroup
Meeting: ”Advancing risk assessment models for invasive alien species in the food chain:
contending with climate change, economics and uncertainty”. Organisation for Economic
Co-operation and Development (OECD), pp. 1-2.
[18] MAES, J., et al, 2013. Mapping and Assessment of Ecosystems and their Services - An
analytical framework for ecosystem assessments under action 5 of the EU biodiversity
strategy to 2020. Publications office of the European Union, 57 pp.
[19] EUROPEAN COMMISSION, 2013. Communication from the commission to the European
Parliament, the Council, the European Economic and Social Committee and the Committee of the Regions - A new EU forest strategy: for forests and the forest-based sector.
No. COM(2013) 659 final. Communication from the Commission to the Council and the
European Parliament.
[20] EUROPEAN COMMISSION, 2013. Commission staff working document accompanying the
document: Communication from the commission to the European Parliament, the Council,
the European Economic and Social Committee and the Committee of the Regions - a
new EU forest strategy: for forests and the forest-based sector. Commission Staff Working
Document 2013 (SWD/2013/0342 final), 98pp.
[21] BARREDO, J. I., 2007. Major flood disasters in Europe: 1950-2005. Natural Hazards 42(1):
125-148.
[22] BARREDO, J. I., 2010. No upward trend in normalised windstorm losses in Europe: 19702008. Natural Hazards and Earth System Science 10(1): 97-104.
[23] EVANS, M. R., et al, 2012. Predictive ecology: systems approaches. Philosophical transactions of the Royal Society of London. Series B, Biological sciences 367(1586): 163-169.
[24] PHILLIS, Y. A., KOUIKOGLOU, V. S., 2012. System-of-Systems hierarchy of biodiversity
conservation problems. Ecological Modelling 235-236: 36-48.
[25] LANGMANN, B., 2014. On the role of climate forcing by volcanic sulphate and volcanic
ash. Advances in Meteorology 2014: 1-17.
[26] GOTTRET, M. V., WHITE, D., 2001. Assessing the impact of integrated natural resource
management: Challenges and experiences. Ecology and Society 5 (2): 17+
9
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
[27] HAGMANN, J., et al, 2001. Success factors in integrated natural resource management
R&D: Lessons from practice. Ecology and Society 5 (2), 29+.
[28] ZHANG, X., et al, 2004. Scaling issues in environmental modelling. In: Wainwright, J.,
Mulligan, M. (Eds.), Environmental modelling : finding simplicity in complexity. Wiley.
[29] ESTREGUIL, C., et al, 2013. Forest landscape in Europe: Pattern, fragmentation and
connectivity. EUR - Scientific and Technical Research 25717 (JRC 77295), 18 pp.
[30] TURNER, R., 2010, Disturbance and landscape dynamics in a changing world. Ecology,
91 10: 2833-2849.
[31] Van WESTEN, C.J., 2013. Remote sensing and GIS for natural hazards assessment and
disaster risk management. In: Bishop, M. P. (Ed.), Remote Sensing and GIScience in
Geomorphology. Vol. 3 of Treatise on Geomorphology. Elsevier, pp. 259-298.
[32] URBAN, M.C., et al, 2012. A crucial step toward realism: responses to climate change
from an evolving metacommunity perspective. Evolutionary Applications 5 (2): 154-167.
[33] BAKLANOV, A., 2007. Environmental risk and assessment modelling - scientific needs
and expected advancements. In: Ebel, A., Davitashvili, T. (Eds.), Air, Water and Soil
Quality Modelling for Risk and Impact Assessment. NATO Security Through Science Series.
Springer Netherlands, pp. 29-44.
[34] STEFFEN, W., et al, 2011. The anthropocene: From global change to planetary stewardship. AMBIO 40 (7): 739-761.
[35] WHITE, C., et al, 2012. The value of coordinated management of interacting ecosystem
services. Ecology Letters 15 (6): 509-519.
[36] de RIGO, D., 2013, Software Uncertainty in Integrated Environmental Modelling: the role
of Semantics and Open Science. Geophys. Res. Abstr. 15: 13292+.
[37] de RIGO, D., (exp.) 2014. Behind the horizon of reproducible integrated environmental
modelling at European scale: ethics and practice of scientific knowledge freedom. F1000
Research, submitted.
[38] LEMPERT, R. J., May 2002. A new decision sciences for complex systems. Proceedings of
the National Academy of Sciences 99 (suppl 3): 7309-7313.
[39] RAMMEL, C., et al, 2007. Managing complex adaptive systems - a co-evolutionary perspective on natural resource management. Ecological Economics 63 (1): 9-21.
[40] van der SLUIJS 2012, J. P., 2012. Uncertainty and dissent in climate risk assessment: A
Post-Normal perspective. Nature and Culture 7 (2): 174-195.
[41] de RIGO, D., et al, 2013. An architecture for adaptive robust modelling of wildfire behaviour
under deep uncertainty. IFIP Adv. Inf. Commun. Technol. 413: 367-380.
[42] GUARISO, G., et al, 2000. A map-based web server for the collection and distribution of
environmental data. In: Kosmatin Fras, M., Mussio, L., Crosilla, F., Podobnikar, T. (Eds.),
Bridging the gap: ISPRS WG VI/3 and IV/3 Workshop, Ljubljana, February 2-5, 2000 :
Collection of Abstracts. Ljubljana.
[43] GUARISO, G., WERTHNER, H., 1989. Environmental decision support systems. E. Horwood; Halsted Press.
[44] GUARISO, G., et al, 1985. Decision support systems for water management: The lake
Como case study. European Journal of Operational Research 21(3): 295-306.
[45] SONCINI-SESSA, R., et al, 2007. Integrated and participatory water resources
management theory. Elsevier.
[46] GUARISO, G., PAGE, B. (Eds), 1994. Computer support for environmental impact
assessment: proceedings of the IFIP TC5/WG5.11 Working Conference on Computer
Support for Environmental Impact Assessment, CSEIA 93, Como, Italy, 6-8 October, 1993.
North-Holland.
[47] CASAGRANDI, R., GUARISO, G., 2009. Impact of ICT in environmental sciences: A
citation analysis 1990-2007. Environmental Modelling & Software 24 (7): 865-871.
10
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
[48] COLE, J., 2010. Interoperability in a crisis 2: Human factors and organisational processes.
Tech. rep., Royal United Services Institute.
[49] STERMAN, J.D., 2002. All models are wrong: reflections on becoming a systems scientist.
System Dynamics Review 18 (4): 501-531.
[50] WEICHSELGARTNER, J., KASPERSON, R., 2010. Barriers in the science-policy-practice
interface: Toward a knowledge-action-system in global environmental change research.
Global Environmental Change 20 (2): 266-277.
[51] BAINBRIDGE, L., 1983. Ironies of automation. Automatica 19(6): 775-779.
[52] de RIGO, D., et al, 2013. Toward open science at the European scale: Geospatial
Semantic Array Programming for integrated environmental modelling. Geophysical
Research Abstracts 15: 13245+.
[53] STENSSON, P., JANSSON, A., 2013. Autonomous technology - sources of confusion: a
model for explanation and prediction of conceptual shifts. Ergonomics 57 (3): 455-470.
[54] RUSSELL, D. M., et al, 2003. Dealing with ghosts: Managing the user experience of
autonomic computing. IBM Systems Journal 42 (1): 177-188.
[55] ANDERSON, S., et al, 2003. Making autonomic computing systems accountable: the problem of human computer interaction. In: Database and Expert Systems Applications, 2003.
Proceedings. 14th International Workshop on. IEEE, pp. 718-724.
[56] KEPHART, J. O., CHESS, D. M. 2003. The vision of autonomic computing. Computer 36
(1): 41-50.
[57] van der SLUIJS, J.P., 2005. Uncertainty as a monster in the science-policy interface: four
coping strategies. Water Science & Technology 52 (6): 87-92.
[58] FRAME, B., 2008. ’Wicked’, ’messy’, and ’clumsy’: long-term frameworks for sustainability. Environment and Planning C: Government and Policy 26 (6): 1113-1128.
[59] MCGUIRE, M., SILVIA, C., 2010. The effect of problem severity, managerial and organizational capacity, and agency structure on intergovernmental collaboration: Evidence from
local emergency management. Public Administration Review 70(2): 279-288.
[60] BEA, R., et al, 2009. A new approach to risk: The implications of e3. Risk Management
11 (1): 30-43.
[61] ADAMS, K.M., HESTER, P.T., 2012. Errors in systems approaches. International Journal
of System of Systems Engineering 3 (3/4): 233+.
[62] LARSSON, A., et al, 2010. Decision evaluation of response strategies in emergency management using imprecise assessments. Journal of Homeland Security and Emergency Management 7 (1).
[63] INNOCENTI, D., ALBRITO, P., 2011. Reducing the risks posed by natural hazards and
climate change: the need for a participatory dialogue between the scientific community and
policy makers. Environmental Science & Policy 14 (7): 730-733.
[64] RAVETZ, J., 2004. The post-normal science of precaution. Futures 36 (3): 347-357.
[65] de RIGO, D., 2012. Integrated Natural Resources Modelling and Management: minimal
redefinition of a known challenge for environmental modelling. Excerpt from the Call for a
shared research agenda toward scientific knowledge freedom, Maieutike Research Initiative
[66] de RIGO, D., 2012, Semantic Array Programming for Environmental Modelling:
Application of the Mastrave Library. Int. Congress on Environmental Modelling and
Software. Managing Resources of a Limited Plant, 1167-1176.
[67] de RIGO, D., 2012. Semantic Array Programming with Mastrave - Introduction to Semantic
Computational Modelling.
[68] CORTI, P., et al, 2012. Fire news management in the context of the european forest fire
information system (EFFIS). In: proceedings of ”Quinta conferenza italiana sul software
geografico e sui dati geografici liberi” (GFOSS DAY 2012).
[69] SHETH, A., 2009. Citizen sensing, social signals, and enriching human experience. Internet
Computing, IEEE 13 (4): 87-92.
11
Rodriguez-Aseretto, D., Schaerer, C., de Rigo, D., 2014. Architecture of Environmental Risk
Modelling: for a faster and more robust response to natural disasters. CCIS 2014
[70] ZHANG, D., et al, 2011. The emergence of social and community intelligence. Computer
44 (7): 21-28.
[71] ADAM, N.R., et al, 2012. Spatial computing and social media in the context of disaster
management. Intelligent Systems, IEEE 27 (6): 90-96.
[72] FRATERNALI, P., et al, 2012. Putting humans in the loop: Social computing for water
resources management. Environmental Modelling & Software 37: 68-77.
[73] RODRIGUEZ-ASERETTO, D., et al, (exp.) 2014. Image geometry correction of daily
forest fire progression map using MODIS active fire observation and Citizens sensor. IEEE
Earthzine 7(2). Submitted.
[74] BOSCO, C., et al, 2013. Multi-Scale Robust Modelling of Landslide Susceptibility: Regional
Rapid Assessment and Catchment Robust Fuzzy Ensemble. IFIP Adv. Inf. Commun.
Technol. 413: 321-335.
[75] DI LEO, M., et al, 2013. Dynamic data driven ensemble for wildfire behaviour assessment:
a case study. IFIP Adv. Inf. Commun. Technol. 413: 11-22.
[76] ACKERMAN, F., HEINZERLING, L., 2002. Pricing the priceless: Cost-Benefit analysis
of environmental protection. University of Pennsylvania Law Review 150 (5): 1553-1584.
[77] GASPARATOS, A., 2010. Embedded value systems in sustainability assessment tools and
their implications. Journal of Environmental Management 91 (8): 1613-1622.
[78] de RIGO, D., et al, 2001. Neuro-dynamic programming for the efficient management of
reservoir networks. In: Proceedings of MODSIM 2001, International Congress on Modelling
and Simulation. Vol. 4. Model. Simul. Soc. Australia and New Zealand, pp. 1949-1954.
[79] RODRIGUEZ-ASERETTO, D., et al, 2009. Injecting dynamic Real-Time data into a
DDDAS for forest fire behavior prediction. Lecture Notes in Computer Science 5545: 489499.
[80] CENCERRADO, A., et al, 2009. Support for urgent computing based on resource virtualization. Lecture Notes in Computer Science 5544: 227-236.
[81] JOSHIMOTO, K, K, et al, 2012. Implementations of Urgent Computing on Production
HPC Systems. Procedia Computer Science 9: 1687-1693.
[82] de RIGO, D., BOSCO, C., 2011. Architecture of a Pan-European framework for integrated
soil water erosion assessment. IFIP Adv. Inf. Commun. Technol. 359: 310-318.
[83] de RIGO, D., et al, 2013. Continental-Scale Living Forest Biomass and Carbon Stock: A
Robust Fuzzy Ensemble of IPCC Tier 1 Maps for Europe. IFIP Adv. Inf. Commun. Technol.
359: 271-284.
[84] RODRIGUEZ-ASERETTO, D., et al, 2013. A data-driven model for large wildfire
behaviour prediction in Europe. Procedia Computer Science 18: 1861-1870.
[85] CASTELLETTI, A., et al, On-Line design of water reservoir policies based on inflow
prediction. IFAC-PapersOnLine 17: 14540-14545.
[86] MCINERNEY, D., et al, 2012. Developing a forest data portal to support Multi-Scale
decision making. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 5(6): 1692-1699.
[87] BASTIN, et al, 2012. Web services for forest data, analysis and monitoring: Developments
from EuroGEOSS. IEEE Earthzine 5(2): 389531+.
[88] RODRIGUEZ-ASERETTO, D., et al, 2013. Free and open source software underpinning
the European Forest Data Centre. Geophysical Research Abstracts 15: 12101+.
[89] NATIONAL SCIENCE FOUNDATION, CYBERINFRASTRUCTURE COUNCIL, 2007.
Cyberinfrastructure vision for 21st century discovery. Tech. Rep. NSF 07-28, National
Science Foundation.
[90] ALTAY, N., GREEN, W.G., 2006. OR/MS research in disaster operations management.
European Journal of Operational Research 175 (1): 475-493.
[91] RODRIGUEZ-ASERETTO, D., et al, 2008. An adaptive system for forest fire behavior
prediction. In: Computational Science and Engineering, 2008. CSE ’08. 11th IEEE International Conference on. IEEE, pp. 275-282.
12
| 5cs.CE
|
Emergent failures and cascades in power grids: a statistical physics perspective
Tommaso Nesti and Bert Zwart
CWI, Amsterdam
Alessandro Zocca
arXiv:1709.10166v1 [physics.soc-ph] 28 Sep 2017
California Institute of Technology
We consider complex networks where line failures occur indirectly as line flows are influenced
by fluctuating input at nodes, a prime example being a power grid where power is generated by
renewable sources. We examine the propagation of such emergent failures in the small noise regime,
combining concepts from statistical physics and the physics of power flow. In particular we characterize rigorously and explicitly the configuration of inputs responsible for failures and cascades, and
analyze the propagation of failures, which often is not of nearest-neighbor type.
Cascading failures in complex networks received a lot
of attention in recent years [1–3, 10, 11, 13, 16–19, 26, 29–
32]. Despite proposing different mechanisms for the cascade evolution, a common feature of all these works is
that the cascade is assumed to be triggered by some external initiating event or contingency. This initial contingency, or attack, is chosen either (i) deliberately, e.g. targeting the most vulnerable or crucial network component,
thus aiming for a worst-case analysis, or (ii) uniformly at
random, in order to understand the average reliability
of a network. This distinction led to the insight that
complex networks are resilient to random attacks, but
vulnerable to targeted attacks [8, 9, 19]. All attacks, random or deterministic, lead to direct failure of the target
network line or node.
regime where ε → 0. The parallel with statistical physics
allows us to determine the most likely configuration of
power injections that leads to failures and possibly cascades. Our results are explicit, and yield fundamental
insights into the way cascades occur. Though we focus
on a power grid, our approach is applicable to any network where fluctuations of the node inputs can trigger
line failures.
We now give a detailed model description. Our network is represented by a connected graph G with n nodes
representing the buses and m directed edges modeling the
transmission lines. The stochastic fluctuations of the net
power injections around their nominal values µ ∈ Rn are
modeled as a multivariate n-dimensional Gaussian random vector with mean µ
In the present work, we look at networks where line failures occur indirectly, by small fluctuations at the nodes.
Our inspiration is drawn from the power grid [5, 12], and
its potential vulnerability to fluctuations of renewable energy sources. Specifically, we look at the power grid as a
static noise-perturbed system, where nodes are attacked
globally but indirectly in the form of small fluctuations in
the power injections. Due to the interplay between network structure, correlations in the power injections, and
power flow physics, these fluctuations may cumulate so
that line failures emerge, possibly triggering cascades.
p ∼ Nn (µ, εΣp ),
This is highly relevant due to the increasing penetration of renewable energy sources in modern power grids
and their susceptibility to weather conditions. This development poses several challenges to the design and control of such networks: intermittent and highly correlated
power generation causes random fluctuations in the line
power flows, possibly yielding outages and cascading failures. It is then of crucial importance to develop appropriate physical models that give fundamental insight.
The emerging nature of failures challenges the analysis.
To tackle this, we use ideas from statistical physics and
large deviations theory. We consider a stochastic model
for the network power injections similar to that proposed
in [28], introduce a positive real parameter ε describing
the magnitude of the noise and focus on the small-noise
(1)
with the convention that pi is the net power injected at
node i. Differently for the usual assumption of the fluctuations being independent and identically distributed, we
can allow for heterogeneous standard deviations of the
power injections at the various nodes as well as for dependencies between of the fluctuations in different nodes
by choosing a non-trivial covariance matrix Σp . This is
instrumental, for instance, to model positive correlations
due to geographical proximity of wind turbines and solar panels. Gaussianity is consistent with atmospheric
physics [6, Section 1.7] and wind turbine statistics [4].
Assuming the mean power injection vector µ has zero
sum and using the DC approximation [5, Chapter 4], the
line power flows f are
f = V p,
(2)
where V is a m×n matrix that encodes the power network
topology and weights (modeling susceptances).
The assumption that the mean power injection vector
µ has zero sum does not guarantee
that the total net
Pn
power injected in the network i=1 pi is equal to zero as
p is a random vector. This minor technical issue can be
easily resolved by assuming that the total power injection
mismatch is distributed uniformly among all the nodes
2
(we account for this in the matrix V , for details see the
Supplemental Material).
In view of assumptions (1) and (2), the line power
flows f also follow a multivariate Gaussian distribution,
namely f ∼ Nm (ν, εΣf ). The vector ν = V µ ∈ Rm describes the average line flows, while the covariance matrix
Σf = V Σp V T describes the correlations between line
flows taking in account both by the correlations of the
power injections (encoded by the matrix Σp ) and those
created by the network topology due to the physics of
the power flows (Kirchhoff’s laws) via the matrix V .
A line overloads if the absolute amount of power flowing in it exceeds a given line threshold. We assume that
such an overload immediately leads to the outage of the
corresponding line, to which we will henceforth refer simply as line failure. The failure of a line cause a global
redistribution of the line power flows according to Kirchhoff’s laws, which could trigger further failures and cascades.
Without loss of generality, we assume that the matrix
V also contains information about the line thresholds, so
that f is the vector of normalized line power flows and
the failure of line ℓ is described by the event {|fℓ | ≥ 1}.
We consider a scenario where the power grid operates
on average safely within the limits by assuming that
maxℓ=1,...,m |νℓ | < 1 so that only large fluctuations of
line flows lead to failures, which correspondingly become
rare events as ε → 0. To assess how much a network is
“robust” against initiating failure and identify its most
vulnerable lines, we derive the exponential decay of probabilities of single line failure events, namely {|fℓ | ≥ 1} for
ℓ = 1, . . . , m. The theory of large deviations [27] is concerned precisely with calculating the exponential decay
of rare events probabilities, which are usually referred to
as rate functions or entropy functions. Since line flows
have been normalized, we are interested in evaluating the
rate function describing the failure event of any line ℓ in
a single point, i.e. 1, and we refer to the corresponding
value Iℓ as failure decay rate. Thanks to the fact that the
line power flow are Gaussian, we can explicitly calculate
Iℓ , see [27, Example 3.1], as
(1 − |νℓ |)2
,
2σℓ2
Iℓ = − lim ε log Pε (|fℓ | ≥ 1) =
ε→0
(3)
where σℓ2 = (Σf )ℓ,ℓ is the variance of the line flow fℓ .
Thus, for small ε, we can approximate the probability of
the emergent failure of line ℓ as
P(|fℓ | ≥ 1) ≈ exp(−Iℓ /ε),
(4)
and that of the first emergent failure as
how large the variance σℓ2 is. Note that σℓ2 accounts for
the power injections variability and correlations as well
as for how much the network possibly amplifies or mitigates them. Decay rates can be used to identify lines that
are most susceptible to the system’s noise, as illustrated
by Figure 1.
FIG. 1: Failure decay rates {Iℓ }ℓ=1,...,m heat-map for the
IEEE 118-bus test system. The likelihood of line failures is
visualized using a color gradient in which the red lines are the
most vulnerable ones.
Large deviations theory provides further analytic tools
that give valuable insight in understanding the way a specific rare event occurs. Conditionally on the failure of line
ℓ, the power injections configuration exhibits a sensible
(“large”) deviation from its mean that is characterized by
p(ℓ) =
1
(p − µ)T Σ−1
p (p − µ).
2
p|≥1
arg inf
p∈Rn
: |eT
V
ℓ
(6)
Provided that νℓ 6= 0, the solution is unique and reads
p(ℓ) = µ +
(sign(νℓ ) − νℓ ) p
Σp V T e ℓ ∈ Rn ,
σℓ2
(7)
where sign(a) = 1 if a ≥ 0 and −1 otherwise, while eℓ ∈
Rm is the vector with the ℓ-th entry equal to 1 and zeros
elsewhere (details in the Supplemental Material).
A key finding is that an emergent line failure does not
occur due to large deviations in the power injections of
the neighboring nodes, but as a cumulative effect of small
unusual fluctuations of the power injections in the entire
network “summed up” by power flow physics, see Figure 2.
Our approach allows to differentiate between different
types of line failures: by calculating the line power flow
profile f (ℓ) := V p(ℓ) corresponding to the most likely
power injections configuration leading to the failure of
line ℓ, we can assess whether the most likely way for failure of line ℓ to occur is as
(ℓ)
(5)
• an isolated failure, if |fk | < 1 for any line k 6= ℓ,
or
The decay rate Iℓ depends on how close the (normalized) average power flow νℓ is to the threshold 1 and
• a joint failure of multiple lines together, if there
(ℓ)
exists some other line k 6= ℓ such that |fk | ≥ 1.
P(max |fℓ | ≥ 1) ≈ exp(− min Iℓ /ε).
ℓ
ℓ
3
FIG. 2: Representation of the most likely power injections
configuration of the IEEE 118-bus test system leading to the
failure of the green line. Nodes size is adjusted proportionally
to their deviations from their mean, while their color describes
whether these deviations are positive (blue) or negative (red)
flow on the unique remaining path in the opposite direction. This observation can be made rigorous by showing
(ℓ)
that the redistribution coefficients are φk = −1 for every k 6= ℓ. It is intuitive that neighboring lines have
positive correlated power flows, while distant lines have
negative correlations as power flows must sum to zero by
Kirchhoff’s law. Hence, the most likely power injections
configuration that makes the power flow in line ℓ exceeding the line threshold (say by becoming larger than 1)
also makes the power flows in the antipodal half of the
network negative. These will then go beyond the line
threshold (by becoming smaller than −1) after the power
flow redistributes, see Figure 3.
-13/35-1
-13/35
-3/7
The resulting non-standard power injections configurae (ℓ) (a
tion p(ℓ) redistributes across an altered network G
subgraph of the original graph G) in which line ℓ (and
possible other lines, in case of a joint failure) has been
removed, possibly increasing the stress on the remaining
e(ℓ) is
lines. The way this redistribution happens on G
governed by the power flow physics and we assume that
it occurs instantaneously, without any transient effects.
The power flow redistribution amounts to compute a new
matrix Ve linking the power injections and the new power
e(ℓ)
flows. The most likely power flow configuration on G
(ℓ)
(ℓ)
after redistribution is then fe = Ve p .
In the special case of an isolated failure (say of line ℓ) it
is enough to calculate only a vector φ(ℓ) ∈ Rm−1 of redistribution coefficients, known as line overload distribution
factors (LODF) in power system engineering. Indeed,
e (ℓ) after rethe most likely power flow configuration on G
(ℓ)
(ℓ)
(ℓ)
e
distribution is equal to f = (fk )k6=ℓ ± φ , where the
sign of the second term depends to the direction in which
the power flowed on line ℓ when the overload occurred.
The power flow configuration fe(ℓ) can be efficiently
used to determine which lines will fail with high probability as a consequence of the original (possibly joint) failure,
(ℓ)
by checking for which line indices k we have |fek | ≥ 1, see
the Supplemental Material for more details.
A rigorous probabilistic theory of emergent cascading
failures can be developed by combining these two ingredients, the statistical physics results describing the most
likely power injections configuration leading to the first
failure and the power flow redistribution in the network
afterwards. In particular, this approach explains several
qualitative features of cascades in power grids that have
been observed empirically, the most prominent one being
the non-local propagation of failures.
We now illustrate how our framework can transparently explain this phenomenon using a ring network, in
which there are exactly two paths along which the power
can flow between any two nodes. Upon the failure of
line ℓ, the power originally flowing on line ℓ must now
-3/7
-1/7
-1/7
1/7-1
1/7
-19/35-1
-19/35
-5/7
1/7
-13/35
1
5/7
1/7
3/7
-5/7
1/7
-13/35-1
failed
5/7
1/7-1
3/7
FIG. 3: On the left, the most likely power injections p(ℓ) leading to the failure in black and power flows f (ℓ) in blue. On the
right, the situation after the power flow redistribution with
(ℓ)
three subsequent failures and the values of fe(ℓ) = (fk −1)k6=ℓ
in blue
Non-local failure propagation thus emerges from the interplay between the power flow configuration just before
the line failure and the network structure, which determines the alternative paths along which power could flow
after the line failure. Figure 4 shows an example of nonlocal failure propagation on a IEEE test system.
FIG. 4: The power injection configuration for the IEEE 118bus test system that most likely causes the failure of green
line after the power redistribution causes the red line to fail.
The most likely power injection configuration leading
to the emergent failure of a given line can be used in
combination with power flow redistribution routines to
generate the failures triggered by that initial scenario.
By repeating this procedure for all lines, one can obtain
4
Graph % joint failures E(F1ec )
IEEE14
65.0%
4.40
IEEE30
97.6%
3.73
IEEE39
80.4%
4.78
88.5%
8.00
IEEE57
IEEE96
72.2%
6.70
IEEE118
91.6%
10.40
87.0%
18.13
IEEE300
E(F2ec )
8.40
9.88
11.39
19.00
21.47
24.53
39.19
E(F2cc )
4.95
4.95
4.85
10.44
7.31
7.56
7.42
built using our large deviations approach. The cliques
of the influence graph (i.e. its maximal fully connected
subgraphs) can then be used to identify clusters of cosusceptable lines [34], which are the lines that statistically
fail often in the same cascade event.
Table I: Percentage of joint failures in emergent cascades and
average number of failed lines F1 up to stage 1 and F2 up to
stage 2 for emergent cascades (ec) and classical cascades (cc)
for some IEEE test systems.
insightful statistics of the first two stages of emergent
cascading failures (ec) and compare them with those of
classical cascading failures (cc), obtained using nominal
power injection values rather than the most likely ones
and deterministic removal of the initial failing line. For
these numerical experiments, line thresholds are taken to
be proportional to the average absolute power flow on the
corresponding lines, i.e. Cℓ = (1 + α)|νℓ | with α = 0.25,
and Σp to be the identity matrix.
As shown in Table I, emergent cascades have a very
high percentage of joint failures and an average number
of failures in the first cascade stage much larger than one
(in classical cascades only one line is removed in the first
cascade stage). Furthermore, the expected total number
of failed lines up to the second cascade stage is significantly larger for emergent cascades than for classical cascades. Lastly, failures propagate in emergent cascades on
average a bit less far than in classical cascades, as illustrated by the statistics of the failure jumping distance in
Table II.
Graph E(Dec )
IEEE14 0.388
IEEE30 0.754
IEEE39 0.898
IEEE57 1.210
IEEE96 1.450
IEEE118 0.679
IEEE300 1.408
E(Dcc )
0.987
1.198
1.633
2.507
1.781
1.638
2.580
cv (Dec )
0.600
0.879
0.891
0.863
0.879
0.745
0.806
cv (Dcc )
1.050
1.115
1.149
1.415
0.946
1.169
1.081
Table II: Average and coefficient of variation of the failure
jumping distance D in stage 2 both for emergent cascades
(ec) and classical cascades (cc). The distance between two
lines is measured as the shortest path between any of their
endpoints.
Our approach also gives a constructive way to build
the so-called “influence graph” [14, 15, 22], in which a directed edge connects lines ℓ and ℓ′ if the failure of the line
ℓ triggers (simultaneously or after redistribution) that of
line ℓ′ . Figure 5 shows an example of influence graph
FIG. 5: The influence graph of the IEEE 118-bus test system (in black) built using the first two stages of all cascade
realizations has a deeply different structure than the original
network (in blue)
The proposed viewpoint on endogenous cascade failures can have important practical implications in terms
of power system reliability: the most likely power injection configurations leading to (possibly joint) failures can
be leveraged to improve the current N −1 safety criterion
which uses nominal values of power injection configurations.
[1] R. Albert, I. Albert, and G. Nakarado. Structural vulnerability of the North American power grid. Physical
Review E, 69(2):025103, feb 2004.
[2] R. Albert and A.-L. Barabási. Statistical mechanics of
complex networks. Reviews of Modern Physics, 74(1):47–
97, 2002.
[3] R. Albert, H. Jeong, and A.-L. Barabási.
Error
and attack tolerance of complex networks. Nature,
406(6794):378–382, jul 2000.
[4] J. Berg, A. Natarajan, J. Mann, and E. Patton. Gaussian vs non-Gaussian turbulence: impact on wind turbine
loads. Wind Energy, 19(11):1975–1989, 2016.
[5] D. Bienstock. Electrical Transmission System Cascades
and Vulnerability. SIAM, Philadelphia, dec 2015.
[6] D. Bienstock, M. Chertkov, and S. Harnett. ChanceConstrained Optimal Power Flow: Risk-Aware Network
Control under Uncertainty. SIAM Review, 56(3):461–495,
2014.
[7] H. Cetinay, F. Kuipers, and P. Van Mieghem. A Topological Investigation of Power Flow. IEEE Systems Journal,
pages 1–9, 2016.
[8] R. Cohen, K. Erez, D. Ben-Avraham, and S. Havlin. Resilience of the Internet to random breakdowns. Physical
Review Letters, 85(21):4626–4628, 2000.
[9] R. Cohen, K. Erez, D. Ben-Avraham, and S. Havlin.
Breakdown of the Internet under Intentional Attack.
5
Physical Review Letters, 86(16):3682–3685, apr 2001.
[10] P. Crucitti, V. Latora, and M. Marchiori. Model for cascading failures in complex networks. Physical Review E,
69(4):045104, apr 2004.
[11] P. Crucitti, V. Latora, M. Marchiori, and A. Rapisarda.
Efficiency of scale-free networks: error and attack tolerance. Physica A: Statistical Mechanics and its Applications, 320:622–642, mar 2003.
[12] I. Dobson, B. Carreras, V. Lynch, and D. Newman.
Complex systems analysis of series of blackouts: Cascading failure, critical points, and self-organization.
Chaos: An Interdisciplinary Journal of Nonlinear Science, 17(2):026103, jun 2007.
[13] D. Heide, M. Schäfer, and M. Greiner. Robustness of
networks against fluctuation-induced cascading failures.
Physical Review E, 77(5):056103, may 2008.
[14] P. Hines, I. Dobson, E. Cotilla-Sanchez, and M. Eppstein. ”Dual Graph” and ”Random Chemistry” Methods
for Cascading Failure Analysis. In 2013 46th Hawaii International Conference on System Sciences, pages 2141–
2150. IEEE, jan 2013.
[15] P. Hines, I. Dobson, and P. Rezaei. Cascading Power
Outages Propagate Locally in an Influence Graph that
is not the Actual Grid Topology. IEEE Transactions on
Power Systems, 32(2):1–1, 2016.
[16] R. Kinney, P. Crucitti, R. Albert, and V. Latora. Modeling cascading failures in the North American power grid.
European Physical Journal B, 46(1):101–107, 2005.
[17] B. Mirzasoleiman, M. Babaei, M. Jalili, and M. Safari.
Cascaded failures in weighted networks. Physical Review
E, 84(4):046114, oct 2011.
[18] A. Motter. Cascade control and defense in complex networks. Physical Review Letters, 93(9):1–4, 2004.
[19] A. Motter and Y.-C. Lai. Cascade-based attacks on complex networks. Physical Review E, 66(6):065102, 2002.
[20] L. Powell. Power system load flow analysis. McGraw Hill,
2004.
[21] K. Purchala, L. Meeus, D. Van Dommelen, and R. Belmans. Usefulness of DC power flow for active power flow
analysis. In IEEE Power Engineering Society General
Meeting, pages 2457–2462. IEEE, 2005.
[22] J. Qi, K. Sun, and S. Mei. An Interaction Model for Simulation and Mitigation of Cascading Failures. IEEE Transactions on Power Systems, 30(2):804–819, mar 2015.
[23] M. Schaub, J. Lehmann, S. Yaliraki, and M. Barahona.
Structure of complex networks: Quantifying edge-to-edge
relations by failure-induced flow redistribution. Network
Science, 2(01):66–89, apr 2014.
[24] S. Soltan, D. Mazauric, and G. Zussman. Analysis of
Failures in Power Grids. IEEE Transactions on Control
of Network Systems, 4(2):288–300, jun 2017.
[25] B. Stott, J. Jardim, and O. Alsac. DC Power Flow Revisited. IEEE Transactions on Power Systems, 24(3):1290–
1300, 2009.
[26] S. Sun, Z. Liu, Z. Chen, and Z. Yuan. Error and attack
tolerance of evolving networks with local preferential attachment. Physica A: Statistical Mechanics and its Applications, 373:851–860, jan 2007.
[27] H. Touchette. The large deviation approach to statistical
mechanics. Physics Reports, 478(1-3):1–69, jul 2009.
[28] Z. Wang, A. Scaglione, and R. Thomas. A MarkovTransition Model for Cascading Failures in Power Grids.
In 2012 45th Hawaii International Conference on System
Sciences, pages 2115–2124. IEEE, jan 2012.
[29] D. Watts. A simple model of global cascades on random
networks. Proceedings of the National Academy of Sciences, 99(9):5766–5771, apr 2002.
[30] D. Witthaut, M. Rohden, X. Zhang, S. Hallerberg, and
M. Timme. Critical Links and Nonlocal Rerouting in
Complex Supply Networks. Physical Review Letters,
116(13):138701, mar 2016.
[31] D. Witthaut and M. Timme. Nonlocal failures in complex
supply networks by single link additions. The European
Physical Journal B, 86(9):377, sep 2013.
[32] D. Witthaut and M. Timme. Nonlocal effects and countermeasures in cascading failures. Physical Review E,
92(3):032809, sep 2015.
[33] A. Wood, B. Wollenberg, and G. Sheble. Power generation, operation, and control. John Wiley & Sons, 3rd
edition, 2014.
[34] Y. Yang, T. Nishikawa, and A. Motter. Vulnerability and
Cosusceptibility Determine the Size of Network Cascades.
Physical Review Letters, 118(4):048301, 2017.
SUPPLEMENTAL MATERIAL
Power grid model and DC approximation
We model the power grid network as a connected
weighted graph G with n nodes, modeling buses, and m
edges, representing the transmission lines. Choosing an
arbitrary but fixed orientation of the transmission lines,
the network structure is described by the edge-vertex incidence matrix B ∈ Rm×n definite entry-wise as
if ℓ = (i, j),
1
(8)
Bℓ,i = −1 if ℓ = (j, i),
0
otherwise.
Denote by βℓ = βi,j = βj,i > 0 the weight of edge
ℓ = (i, j), corresponding to the susceptance of that transmission line. By convention, we set βi,j = βj,i = 0 if there
is no transmission line between i and j. Denote by W the
m×m diagonal matrix defined as W := diag(β1 , . . . , βm ).
The network topology and weights are simultaneously encoded in the weighted Laplacian matrix of the graph G,
defined as L := B T W B or entry-wise as
(
−βi,j
if i 6= j,
Li,j = P
(9)
if i = j.
k6=j βi,k
Denote by J ∈ Rn×n the matrix with all entries equal
to one. The relation between any vector of power injections p ∈ Rn and the phase angles θ ∈ Rn they induce in
the network nodes can be written in matrix form as
θ = L+ Sp,
(10)
where L+ ∈ Rn×n is the Moore-Penrose pseudo-inverse of
L and the matrix S := I − n1 J ∈ Rn×n acts as a “global
slack”, ensuring that the net power injection is always
6
identically zero. Exploiting the eigenspace structure of
L, L+ can be calculated as
1 −1 1
(11)
− J,
L+ = L + J
n
n
In the literature, instead of L+ it is commonly used
another matrix L̂, calculated using the inverse of the
(n − 1) × (n − 1) sub-matrix obtained from L deleting the
first row and first column the matrix L. In our method
we are implicitly choosing an average value of zero as a
reference for the nodes voltage phase angles, while in the
classical one the first node is used as reference by setting
is phase angle equal to zero. We remark that these two
procedure are equivalent if one is interested in the line
power flows, as these latter depend only on the phase
angle differences.
We make use of the DC approximation, which is commonly used in high-voltage transmission system analysis [20, 21, 25, 33], according to which the real power
flows fˆ are related with the phase angles θ via the linear
relation fˆ = Bθ. In view of (10), the line power flow fˆ
can be written as a linear transformation of the power
injections p, i.e.
fˆ = W BL+ Sp.
(12)
It is convenient to look at the normalized line power flow
vector f ∈ Rm , defined component-wise as fℓ := fˆℓ /Cℓ
for every ℓ = 1, . . . , m, where Cℓ is the line threshold,
which we define using the nominal average power injection vector µ and choosing a tolerance parameter α > 0
as
Cℓ = (1 + α)|(W BL+ Sµ)ℓ |,
ℓ = 1, . . . , m.
(13)
The relation between line power flows and normalized
power flows can be rewritten as f = C fˆ, where C is the
−1
m×m diagonal matrix C := diag(C1−1 , . . . , Cm
). In view
of (12), the normalized power flows f can be expressed
in terms of the power injections p as
f = V p,
(14)
where V := CW BL+ S ∈ Rm×n .
The most likely power injection configuration p(ℓ) ∈ Rn
given the event {|(fε )ℓ | ≥ 1} is the solution of the variational problem
p(ℓ) =
1
(p − µ)T Σ−1
p (p − µ).
p∈Rn : |eT V p|≥1 2
arg inf
which, when νℓ 6= 0, can be explicitly computed as
p(ℓ) = µ +
(sign(νℓ ) − νℓ ) p
Σp V T e ℓ .
σℓ2
f (ℓ) = V p(ℓ) = ν +
(sign(νℓ ) − νℓ )
V Σp V T eℓ ∈ Rm . (18)
σℓ2
The vector f (ℓ) can be also seen as the conditional expectation of random line power flow vector fε conditional on
the failure event {fℓ = sign(νℓ )}, namely
f (ℓ) = E[fε | (fε )ℓ = sign(νℓ )],
(19)
and thus, in particular, for every k = 1, . . . , m
(ℓ)
fk = νk + (sign(νℓ ) − νℓ )
Cov(fℓ , fk )
.
Var(fℓ )
(20)
Note that the case νℓ = 0 has been excluded only for
compactness. Indeed, in that case the variational problem (16) has two solutions, p(ℓ,+) and p(ℓ,−) . This is easily
explained as if the power flow on line ℓ has mean νℓ = 0,
then it is equally likely for the overload event {|fℓ | ≥ 1}
to occur as {fℓ ≥ 1} or as {fℓ ≤ −1} and the most likely
power injection configurations that trigger them can be
different.
The previous proposition immediately yields the large
deviation principle also for the first line failure event
{kfε k∞ ≥ 1}, which reads
ε→0
In the following statement we write fε to stress the
dependence of the line power flows on the noise parameter
ε.
(17)
The line power flows corresponding to the power injection configuration p(ℓ) can be calculated as
(1 − |νℓ |)2
.
ℓ=1,...,m
2σℓ2
lim ε log P(||fε ||∞ ≥ 1) = − min
Large deviation principles for failure events
(16)
ℓ
(21)
Indeed, the decay rate for the event that at least one
line fails is equal to the minimum of the decay rates for
the failure of each line. The most likely power injections
configuration that leads to the event {kfε k∞ ≥ 1} is
2
∗
ℓ |)
.
p(ℓ ) with ℓ∗ = arg minℓ=1,...,m (1−|ν
2σ2
ℓ
Proposition 1 Assume that maxj=1,...,m |νk | < 1.
Then, for every ℓ = 1, . . . , m, the sequence of line power
flows (fε )ε>0 satisfies the large deviations principle
lim ε log P(|(fε )ℓ | ≥ 1) = −
ε→0
(1 − |νℓ |)2
.
2σℓ2
(15)
Proof of Proposition 1.
Let (Z (i) )i∈N be a sequence of i.i.d. multivariate
Pn normal
vectors Z (i) ∼ Nm (ν, Σ), and let Sn := n1 i=1 Z (i) be
the sequence of the partial sums. By setting ε = n1 , it
d
immediately follows that that fε = Sn . Denote g(p) :=
7
1
T −1
2 (p − µ) Σp (p − µ).
Following [27, Section 3.D], we get
1
log P((Sn )ℓ ≥ 1) =
n
(1 − νℓ )2
=−
inf
g(p) = −
,
(22)
2σℓ2
p∈Rn : eT
V p≥1
ℓ
1
log P((Sn )ℓ ≤ −1) =
lim ε log P((fε )ℓ ≤ −1) = lim
n→∞ n
ε→0
(−1 − νℓ )2
=−
inf
g(p) = −
.
(23)
2σℓ2
p∈Rn : eT
V p≤−1
ℓ
lim ε log P((fε )ℓ ≥ 1) = lim
n→∞
ε→0
The optimizers of problems (22) and (23) are easily computed respectively as as
p(ℓ,+)
p(ℓ,−)
1 − νℓ p
Σp V T e ℓ ,
=µ+
σℓ2
−1 − νℓ p
=µ+
Σp V T e ℓ .
σℓ2
inf
= min
n
g(p) =
inf
p∈Rn : eT
V p≥1
ℓ
g(p),
inf
p∈Rn : eT
V p≤−1
ℓ
Cℓ Ra,j − Ra,i + Rb,i − Rb,j
·
,
Ck
2(1 − βℓ Ri,j )
(26)
It is easy to prove that in a ring network with homogeneous line thresholds and susceptances φℓ,k = −1 for
every ℓ 6= k.
More in general (hence also in the case of joint failures), the most likely power flow configuration fe(ℓ) after
redistribution in general can be written as
φℓ,k = φ(i,j),(a,b) := βk ·
fe(ℓ) = Ve p(ℓ) ,
(27)
where the (m−j(ℓ))×n matrix Ve can be constructed analogously to V , but considering the altered graph G(ℓ) instead of G. The next proposition shows that it is enough
to look at the vector fe(ℓ) to determine whether a line
that survived at the first cascade stage (i.e. that did not
fail jointly with ℓ) will fail with high probability or not
after the power redistribution (e.g. at the second cascade
stage).
Note that trivially
p∈Rn : |eT
V p|≥1
ℓ
where if ℓ = (i, j) and k = (a, b) the coefficient φℓ,k ∈ R
can be computed as
o
g(p) ,
(ℓ)
and thus identities (15) and (16) immediately follow.
Proposition 2 For any δ ∈ (0, 1), define Qδ := {y ∈
Rm−j(ℓ) : |yi | ≤ 1 − δ}. The following statement hold:
(ℓ)
i) If |fek | < 1 for any k 6∈ J(ℓ), then
Power flow redistribution
(ℓ)
lim ε log P(feε ∈
/ Q̊0
|(fε )ℓ | ≥ 1) < 0;
ε→0
For every line ℓ define J (ℓ) to be the collection of lines
that fail jointly with ℓ as
(ℓ)
J (ℓ) := {k : |fk | ≥ 1}.
(24)
(ℓ)
fek = fk + φℓ,k sign(νℓ ),
for every k 6= ℓ,
(ℓ)
ii) If there exists k 6∈ J(ℓ) such that |fek | ≥ 1, then
(ℓ)
lim ε log P(feε ∈ Qδ |(fε )ℓ | ≥ 1) < 0.
ε→0
Let j(ℓ) = |J (ℓ)| its cardinality and note that j(ℓ) ≥ 1
e (ℓ) the
as trivially ℓ always belongs to J (ℓ). Denote by G
graph obtained from G by removing all the lines in J (ℓ).
Let us focus first on the case of the isolated failure of line ℓ, that is when J (ℓ) = {ℓ}. In this case
e (ℓ) = G(V, E \ {ℓ}) is the graph obtained from G after
G
removing the line ℓ = (i, j). Provided that the power injections remain unchanged, the power flows redistribute
among the remaining lines. Using the concept of resistance matrix R ∈ Rm×m and under the DC approximation, in [7, 23, 24] it is proven that alternative paths for
e(ℓ) is still
the power to flow from node i to j exists (i.e. G
connected) if and only if βi,j Ri,j 6= 1. In other words,
βi,j Ri,j = 1 can only occur in the scenario where line
ℓ = (i, j) is a bridge, i.e. its removal results in the disconnection of the original graph G in two components.
e(ℓ) is still a connected graph, the power flows after
If G
redistribution fe(ℓ) ∈ Rm−1 are related with the original
line flows f ∈ Rm in the network G by the relation
(25)
(28)
(29)
Proof of Proposition 2. i) We have
P(fε ∈
/ Q̊(k) (fε )ℓ ≥ 1) =
P(fε ∈
/ Q̊(k) , (fε )ℓ ≥ 1)
,
P((fε )ℓ ≥ 1)
Denote g(p) := 21 (p − µ)T Σ−1
p (p − µ). From large deviations theory it readily follows that
lim ε log P((fε )ℓ ≥ 1) = −
ε→0
inf
p∈Rn : eT
V p≥1
ℓ
g(p)
(ℓ)
lim ε log P(feε ∈
/ Q̊0 , (fε )ℓ ≥ 1) =
(30)
ε→0
=−
inf
g(p).
p∈Rn : eT
ℓ V p≥1,
(31)
e
∃ k6∈J (ℓ) : |eT
k V p|≥1
Define the corresponding decay rates as
Iℓ :=
inf
p∈Rn : eT
V p≥1
ℓ
g(p),
Jℓ :=
inf
p∈Rn : eT
ℓ V p≥1,
e
∃ k6∈J (ℓ) : |eT
k V p|≥1
g(p).
(32)
8
Then we can rewrite
lim ε log P(feε ∈
/ Q̊(k) (fε )ℓ ≥ 1) = −Jℓ + Iℓ .
ε→0
(33)
and, therefore,
lim ε log P(feε ∈
/ Q̊(k) (fε )ℓ ≥ 1) = 0 ⇐⇒ Jℓ = Iℓ .
ε→0
(34)
Notice that the feasible set of the minimization problem
(31) is strictly contained in that of the problem (30), implying that Jℓ ≥ Iℓ .
Recall that we denoted by p(ℓ) the unique optimal solution of (30) and the corresponding line power flow vector
fe(ℓ) = Ve p(ℓ) . Let p̂(ℓ) be an optimal solution of (31) and
define fˆ(ℓ) := Ve p̂(ℓ) . Clearly p̂(ℓ) is feasible also for problem (30). If it was the case that Jℓ = Iℓ , then p̂(ℓ) would
be an optimal solution also for (30), and thus by uniqueness (g(p) is strictly convex) p̂(ℓ) = p(ℓ) and fˆ(ℓ) = fe(ℓ) .
But this leads to a contradiction, since by assumption
(ℓ)
|fek | < 1 for all k 6∈ J (ℓ), while fˆ(ℓ) is by construc(ℓ)
tion such that there exists k 6∈ J (ℓ) such that |fˆk | ≥ 1.
Hence Jℓ > Iℓ and we conclude that
lim ε log P(feε ∈
/ Q̊(k) (fε )ℓ ≥ 1) < 0.
ε→0
(35)
The proof in the case (fε )ℓ ≤ −1 is analogous and so is
that of statement (ii).
| 3cs.SY
|
arXiv:1705.10735v1 [math.ST] 30 May 2017
Submitted to the Annals of Statistics
arXiv: arXiv:0000.0000
THE TWO-TO-INFINITY NORM AND SINGULAR
SUBSPACE GEOMETRY WITH APPLICATIONS TO
HIGH-DIMENSIONAL STATISTICS∗
By Joshua Cape, Minh Tang, and Carey E. Priebe
Johns Hopkins University
The singular value matrix decomposition plays a ubiquitous role
throughout statistics and related fields. Myriad applications including
clustering, classification, and dimensionality reduction involve studying and exploiting the geometric structure of singular values and singular vectors.
This paper contributes to the literature by providing a novel collection of technical and theoretical tools for studying the geometry
of singular subspaces using the 2 → ∞ norm. Motivated by preliminary deterministic Procrustes analysis, we consider a general matrix
perturbation setting in which we derive a new Procrustean matrix
decomposition. Together with flexible machinery developed for the
2 → ∞ norm, this allows us to conduct a refined analysis of the induced perturbation geometry with respect to the underlying singular
vectors even in the presence of singular value multiplicity. Our analysis yields perturbation bounds for a range of popular matrix noise
models, each of which has a meaningful associated statistical inference task. We discuss how the 2 → ∞ norm is arguably the preferred
norm in certain statistical settings. Specific applications discussed
in this paper include the problem of covariance matrix estimation,
singular subspace recovery, and multiple graph inference.
Both our novel Procrustean matrix decomposition and the technical machinery developed for the 2 → ∞ norm may be of independent
interest.
1. Introduction.
1.1. Background. The geometry of singular subspaces is of fundamental
importance throughout a wide range of fields including statistics, machine
∗
This work is partially supported by the XDATA program of the Defense Advanced
Research Projects Agency (DARPA) administered through Air Force Research Laboratory (AFRL) contract FA8750-12-2-0303 and by the DARPA D3M program administered
through AFRL contract FA8750-17-2-0112. This work is also supported by the Acheson J.
Duncan Fund for the Advancement of Research in Statistics at Johns Hopkins University.
MSC 2010 subject classifications: Primary 62H12, 62H25; secondary 62H30
Keywords and phrases: singular value decomposition, perturbation theory, spectral
methods, Procrustes analysis, high-dimensional statistics
1
2
J. CAPE, M. TANG, AND C. E. PRIEBE
learning, computer science, applied mathematics, and network science. Singular vectors (or eigenvectors) together with their corresponding subspaces
and singular values (or eigenvalues) appear throughout various statistical
applications including principal component analysis [2, 5, 22] covariance
matrix estimation [15, 16, 17], spectral clustering [24, 34, 42], and graph
inference [39, 40, 41] to name a few.
Singular subspaces and their geometry are also studied in the random
matrix theory literature which has come to have a profound influence on
the development of high-dimensional statistical theory [1, 31, 45]. Of interest there is the behavior of random matrices themselves, such as the
phenomenon of eigenvector delocalization [35], as well as the spectral behavior of non-random (in particular, low-rank) matrices undergoing random
perturbation [29]. For an overview of recent work on the spectral properties
of random matrices, in particular the behavior of eigenvectors of random
matrices, see the recent survey [30]. For further discussion of how random
matrix theory has come to impact statistics, see the recent survey [31].
From a computational perspective, optimization algorithms are often concerned with the behavior of singular vectors and subspaces in applications
to signal processing and compressed sensing [14]. The study of algorithmic
performance on manifolds and manifold learning, especially the Grassmann
and Stiefel manifolds, motivates related interest in a collection of Procrustestype problems [4, 13]. Indeed, Procrustes analysis occupies an established
area within the theoretical study of statistics on manifolds [9] and arises in
applications including diffusion tensor imaging [11] and shape analysis [12].
See [19] for an extended treatment of both theoretical and numerical aspects
of Procrustes-type problems.
Foundational results from the matrix theory literature concerning the
perturbation of singular values, singular vectors, and singular subspaces date
back to the original work of Weyl [44], Davis and Kahan [10], and Wedin
[43], among others. Indeed, these results form the backbone for much of the
linear algebraic machinery that has since been developed for the purposes of
statistical application and inference. See the classical references [3, 21, 36]
for further treatment of these foundational results and related historical
developments.
1.2. Overview. This paper contributes to the literature by providing a
novel collection of technical and theoretical tools for studying the geometry of singular subspaces with respect to the 2 → ∞ subordinate vector
norm on matrices (described below). We focus on the alignment of singular
subspaces in terms of geometric distance measures between collections of
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY
3
singular vectors (or eigenvectors), especially the classical sin Θ distance. We
prove singular vector perturbation theorems for both low rank and arbitrary
rank matrix settings. We present our main theoretical results quite generally followed by concrete consequences thereof to facilitate direct statistical
applications, specifically to covariance matrix estimation, singular subspace
recovery, and multiple graph inference. Among the advantages of our methods is that we allow singular value multiplicity and require only a population
gap in the spirit of Theorem 2 in [47].
As a special case of our general framework, we recover a strengthened
version of recent results in [17] wherein the authors obtain an ℓ∞ norm
perturbation bound on singular vectors for low rank matrices exhibiting
specific coherence structure. In this way, beyond the stated theorems in this
paper, our results immediately yield analogous applications to, for example,
robust covariance estimation involving heavy-tailed random variables as in
[17].
Our Procrustes analysis complements the recent study of rate-optimal
perturbation bounds for singular subspaces in [6]. When considered in tandem, we demonstrate a setting in which one recovers nearly rate-matching
bounds for a particular Procrustes-type problem.
Yet another consequence of this work is that we extend and complement
current spectral methodology for graph inference and embedding [28, 39].
To the best of our knowledge, we obtain among the first-ever estimation
bounds for multiple graph inference in the presence of edge correlation.
1.3. Setting. More precisely, this paper formulates and analyzes a general matrix decomposition for the aligned difference between real matrices
U and Û consisting of r orthonormal columns (i.e. partial isometries; Stiefel
matrices; orthogonal r-frames) given by
(1.1)
Û − U W,
where W denotes an r × r orthogonal matrix. We focus on (but are not limited to) a particular “nice” choice of W which corresponds to an “optimal”
Procrustes rotation in a sense that will be made precise later. As such, our
results have implications for a class of related Procrustes-type problems.
Along with our matrix decomposition, we develop technical machinery for
the 2 → ∞ subordinate vector norm on matrices, defined for A ∈ Rp1 ×p2 by
(1.2)
kAk2→∞ := max kAxk∞ .
kxk2 =1
Together, these results allow us to obtain a suite of singular vector perturbation bounds for rectangular matrices corresponding to U, Û , and W via
an additive perturbation framework of the singular value decomposition.
4
J. CAPE, M. TANG, AND C. E. PRIEBE
The 2 → ∞ norm provides finer uniform control on the entries of a matrix
than the more commonly encountered spectral or Frobenius norm. As such,
in the presence of additional underlying matrix and/or perturbation structure, the 2 → ∞ norm may well be of greater operational significance and
the preferred norm to consider. In the compressed sensing and optimization
literature, for example, matrices exhibiting the so-called bounded coherence
property in the sense of [7] form a popular and widely-encountered class of
matrices for which the 2 → ∞ norm can be shown to be the “right” choice.
The 2 → ∞ norm is encountered from time to time but is by no means
as pervasive as either the spectral or Frobenius matrix norm. Recently, it
has appeared in the study of random matrices when a fraction of the matrix
entries are modified [33]. Another recent use of the 2 → ∞ norm was in [28]
wherein clustering certain stochastic block model graphs according to the
adjacency spectral embedding is shown to be strongly universally consistent
under mean-squared error. Among the aims of this paper is to advocate for
the more widespread consideration of the 2 → ∞ norm.
1.4. Sample application: covariance matrix estimation. Before proceeding further, we briefly pause to present an application of our work and
methods to estimating the top singular vectors of a structured covariance
matrix. Another result with applications to covariance matrix estimation
will be presented in Section 4.1 (Theorem 4.4).
Denote a random vector Y by its coordinates Y := (Y (1) , Y (2) , . . . , Y (d) )⊤ ∈
d
R and let Y, Y1 , Y2 , . . . , Yn be independent, identically distributed, mean
zero multivariate normal random (column) vectors in Rd with positive semidefinite covariance matrix Γ ∈ Rd×d . Denote the spectral decomposition
of Γ by Γ = U ΣU ⊤ + U⊥ Σ⊥ U⊥⊤ where [U |U⊥ ] ≡ [u1 |u2 | . . . |ud ] ∈ Rd×d is a
unitary matrix and the singular values of Γ are indexed in non-increasing order, σ1 (Γ) ≥ σ2 (Γ) ≥ · · · ≥ σd (Γ), with Σ := diag(σ1 (Γ), σ2 (Γ), . . . , σr (Γ)) ∈
Rr×r and Σ⊥ := diag(σr+1 (Γ), σr+2 (Γ), . . . , σd (Γ)) ∈ Rd−r×d−r where δr (Γ) :=
σr (Γ) − σr+1 (Γ) > 0. Here Σ may be thought of as representing the “signal”
(or “spike”) singular values of Γ while Σ⊥ contains the “noise” (or “bulk”)
singular values. Note that the largest singular values of Γ are not assumed
to be distinct; rather, the assumption δr (Γ) > 0 simply requires a singular
value “population gap” between Σ and Σ⊥ .
n×d let Γ̂
For the matrix of row observations Υ := [Y1 |Y2 | . . . |Yn ]⊤ ∈ R
n
1 Pn
1 ⊤
denote the classical sample covariance matrix Γ̂n := n Υ Υ ≡ n k=1 Yk Yk⊤
with spectral decomposition given by Γ̂n ≡ Û Σ̂Û ⊤ + Û⊥ Σ̂⊥ Û⊥⊤ . Define En :=
Γ̂n −Γ to be the difference between the true and sample covariance matrices.
Further suppose that Γ exhibits bounded coherence in the sense that
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY
5
pr
kU k2→∞ = O
d where O(·) denotes conventional big-O notation. Similarly let Θ(·) and Ω(·) denote conventional big-Theta and big-Omega notation, respectively.
Let WU denote the (random) orthogonal matrix corresponding to the
optimal Frobenius norm Procrustes alignment of U and Û (for further discussion see Section 2.3). Then we have the following performance guarantee
when estimating U , the matrix of top singular vectors of Γ.
Theorem 1.1.
Consider the covariance
of Section 1.4
matrixqsetting
along with
where d ≫ r. Suppose that σr (Γ) = Ω max σ1 (Γ) log(d)
n ,1
p
σ1 (Γ) = Θ(σr (Γ)) and σr+1 (Γ) = O(1). Let ν(Y ) := max1≤i≤d Var(Y (i) ).
Then there exists a constant C > 0 such that with probability at least 1−d−2 ,
!
r
log(d)
ν(Y )r
kÛ − U WU k2→∞ ≤ C p
(1.3)
.
n
σr (Γ)
Similar results hold more generally when the random vector Y is instead
assumed to have a sub-Gaussian distribution.
Remark
p
p1.2. In the setting of Theorem 1.1 one often has ν(Y ) =
O( σ1 (Γ) dr ), in which case the above bound can be written in the simplified form
!
r
r 3 log(d)
kÛ − U WU k2→∞ ≤ C
(1.4)
.
nd
Remark 1.3. Although Theorem 1.1 is stated with respect to the r
largest singular values of the covariance matrix Γ, analogous results may be
formulated for collections of sequential singular values σs (Γ), σs+1 (Γ), . . . , σt (Γ)
that are well-separated from the remainder of the singular values in Σ⊥ , i.e.
when δgap := min(σs−1 (Γ) − σs (Γ), σt (Γ) − σt+1 (Γ)) ≫ 0. To this end, see
Theorem 3.1 and Theorem 7.9.
1.5. Organization. The rest of this paper is organized as follows. Section
2 establishes notation, motivates the use of the 2 → ∞ norm in the context
of Procrustes problems, and presents the perturbation model considered in
this paper. Section 3 collects our general main results which fall under two
categories: matrix decompositions and matrix perturbation theorems. Section 4 demonstrates how this paper improves upon and complements existing
work in the literature by way of considering three statistical applications,
6
J. CAPE, M. TANG, AND C. E. PRIEBE
specifically covariance matrix estimation, singular subspace recovery, and
multiple graph inference. In Section 5 we offer some concluding remarks.
Sections 6 and 7 contain the technical machinery developed for this paper
as well as additional proofs of our main theorems.
2. Preliminaries.
2.1. Notation. In this paper all vectors and matrices are assumed to be
real-valued for simplicity. The symbols := and ≡ are used to assign definitions and denote formal equivalence. The quantity Cα denotes a general
constant depending only on α (either a parameter or an index) which may
change from line to line unless otherwise specified. For any positive integer n,
let [n] := {1, 2, . . . , n}. Additionally, let O(·) denote standard big-O notation
and o(·) denote little-O notation, possibly with an underlying probabilistic
qualifying statement. Similarly let Θ(·) and Ω(·) denote conventional bigTheta and big-Omega notation, respectively.
For (column) vectors x, y ∈ Rp1 where x ≡ (x1 , . . . , xp1 )⊤ , the standard
Euclidean inner product between x and y is denoted by hx, yi. The classical
P
1/p
ℓp vector norms are denoted by kxkp := ( pi=1 |xi |p )
for 1 ≤ p < ∞ and
kxk∞ := maxi |xi |.
Let Op,r denote the set of all p×r real matrices with orthonormal columns
where Op ≡ Op,p denotes the set of orthogonal matrices in Rp×p . For the
rectangular matrix A ∈ Rp1 ×p2 , denote its singular value decomposition
(SVD) by A = U ΣV ⊤ , where the singular values of A are arranged in nonincreasing order and given by Σ = diag(σ1 (A), σ2 (A), . . . ).
This paper makes use of several standard consistent (i.e. sub-multiplicative)
matrix norms,
qP namely kAk2 := σ1 (A) denotes the spectral normPof A,
2
|ai,j |
kAkF :=
i σi (A) denotes the Frobenius norm of A, kAk1 := maxj
P i
denotes the maximum absolute column sum of A, and kAk∞ := maxi j |ai,j |
denotes the maximum absolute row sum of A. We also consider the matrix
norm (more precisely, non-consistent vector norm on matrices) given by
kAkmax := maxi,j |ai,j |.
2.2. Norm relations. A central focus of this paper is on the vector norm
on matrices defined by kAk2→∞ := maxkxk2 =1 kAxk∞ . Proposition 7.1 establishes the elementary fact that this norm corresponds to the maximum
Euclidean row norm of the matrix A. Propositions 7.3 and 7.5 further catalog the relationship between k·k2→∞ and several of the aforementioned more
commonly encountered matrix norms. These propositions, though straightforward, contribute to the machinery for obtaining the main results in this
paper.
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY
7
The 2 → ∞ norm is an attractive quantity due in part to being easily
interpretable and straightforward to compute. Qualitatively speaking, small
values of k · k2→∞ capture “global” (over all rows) and “uniform” (within
each row) matrix behavior in much the same way as do small values of
k · kmax and k · k∞ . This stands in contrast to the matrix norms k · k2 and
k·kF which capture “global” but not necessarily “uniform” matrix behavior.
√
For example, given A := {1/ p2 }p1 ×p2 , observe that kAk2→∞ = 1 while
√
kAk2 = kAkF = p1 .
For A ∈ Rp1 ×p2 , the standard relations between the ℓp norms for p ∈
{1, 2, ∞} permit quantitative comparison of k · k2→∞ to the relative magnitudes of k · kmax and k · k∞ . In particular, the relations between these
quantities depend upon the underlying matrix column dimension, namely.
1
√
kAk2→∞ ≤ kAkmax ≤ kAk2→∞ ≤ kAk∞ ≤ p2 kAk2→∞ .
√
p2
In contrast, the relationship between k · k2→∞ and k · k2 depends on the
matrix row dimension (Proposition 7.3), namely
kAk2→∞ ≤ kAk2 ≤
√
p1 kAk2→∞ .
The consideration of such dimensionality relations plays an important
role in motivating our approach to prove new matrix perturbation results. In
particular, it may be the case that kAk2→∞ ≪ kAk2 when the row dimension
of A is large, as the above example demonstrates, and so bounding kAk2→∞
may be preferred to bounding kAk2 , or, for that matter, to bounding the
larger quantity kAkF .
Given our discussion of matrix norm relations, we also recall the wellknown rank-based relation between the matrix norms k · k2 and k · kF which
allows us to interface the Frobenius and 2 → ∞ norms. In particular, for
any matrix A,
p
kAk2 ≤ kAkF ≤ rank(A)kAk2 .
We pause to note that the 2 → ∞ norm is not in general sub-multiplicative
for matrices. In particular, the “constrained” sub-multiplicative behavior of
k · k2→∞ (Proposition 7.5) together with the non-commutativity of matrix
multiplication and standard properties of common matrix norms—especially
the spectral and Frobenius matrix norms—imply a substantial amount of
flexibility when bounding matrix products and passing between norms. For
this reason, a host of matrix norm bounds follow naturally from our matrix decomposition results in Section 3.1, and the relative strength of these
bounds will depend upon underlying matrix model assumptions.
8
J. CAPE, M. TANG, AND C. E. PRIEBE
2.3. Singular subspaces and Procrustes. Let U and Û denote the corresponding subspaces for which the columns of U, Û ∈ Op,r form orthonormal
bases, respectively. From the classical C-S matrix decomposition, a natural
measure of distance between these subspaces (corresp. matrices) is given via
the canonical (principal) angles between U and Û ([3], Section 7.1). More
specifically, for the singular values of U ⊤ Û , denoted {σi (U ⊤ Û )}ri=1 and indexed in non-increasing order, the canonical angles are given by the main
diagonal elements of the r × r diagonal matrix
Θ(Û , U ) := diag(cos−1 (σ1 (U ⊤ Û )), cos−1 (σ2 (U ⊤ Û )), . . . , cos−1 (σr (U ⊤ Û ))).
For an in-depth review of the C-S decomposition and canonical angles, see
for example [3, 36]. An extensive summary of relationships between sin Θ
distances, specifically k sin Θ(Û , U )k2 and k sin Θ(Û , U )kF , as well as various
other distance measures is provided in the appendix of [6]. This paper focuses
on the sin Θ distance and related Procrustes-type distance measures.
Geometrically, the notion of distance between U and Û corresponds to
discerning the extent of rotational (angular) alignment between these matrices and their corresponding subspaces. As such, Procrustes-type analysis
lends itself to establishing distance measures. More generally, given two matrices A and B together with a set of matrices S and a norm k · k, a general
version of the Procrustes problem is given by the optimization problem
inf kA − BSk.
(2.1)
S∈S
For U, Û ∈ Op1 ,r , this paper considers the two specific instances
(2.2)
inf kÛ − U W k2→∞ and inf kÛ − U W k2 ,
W ∈Or
W ∈Or
with emphasis on the former motivated by insight with respect to the latter.
In each case, the infimum is achieved over Or by the compactness of
Or together with the continuity of the the specified norms. Therefore, let
Wν⋆ ∈ Or denote a Procrustes solution under k · kν for ν ∈ {2 → ∞, 2}
where dependence upon the underlying matrices U and Û is implicit from
context. Unfortunately, neither of the above Procrustes problems admits an
analytically tractable minimizer in general. In contrast, by instead switching
to the Frobenius norm, one arrives at the classical orthogonal Procrustes
problem which does admit an analytically tractable minimizer and which
we denote by WU . Namely, WU achieves
(2.3)
inf kÛ − U W kF .
W ∈Or
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY
9
For the singular value decomposition of U ⊤ Û ∈ Rr×r denoted U ⊤ Û ≡
UU ΣU VU⊤ , the solution WU is given explicitly by WU ≡ UU VU⊤ . Given these
observations, it is therefore natural to study the surrogate quantities
(2.4)
kÛ − U WU k2→∞ and kÛ − U WU k2 .
Towards this end, the sin Θ distance and Procrustes problems are related
in the sense that (e.g. [6], Lemma 1)
√
k sin Θ(Û , U )kF ≤ kÛ − U WU kF ≤ 2k sin Θ(Û , U )kF
and
k sin Θ(Û , U )k2 ≤ kÛ − U W2⋆ k2 ≤ kÛ − U WU k2 ≤
√
2k sin Θ(Û , U )k2 .
Alternatively, as detailed in Lemma 7.8, one can bound kÛ − U WU k2 via
k sin Θ(Û , U )k2 in a manner providing a clearer demonstration that the performance of WU is “close” to the performance of W2⋆ under k · k2 , namely
k sin Θ(Û , U )k2 ≤ kÛ − U W2⋆ k2
≤ kÛ − U WU k2 ≤ k sin Θ(Û , U )k2 + k sin Θ(Û , U )k22 .
Loosely speaking, this says that the relative fluctuation between WU and
W2⋆ in the spectral Procrustes problem are at most O(k sin Θ(Û , U )k22 ).
By simply considering the naı̈ve relationship between k · k2→∞ and k · k2 ,
we similarly observe that
1
⋆
k2→∞
√ k sin Θ(Û , U )k2 ≤ kÛ − U W2→∞
p1
≤ kÛ − U WU k2→∞ ≤ k sin Θ(Û , U )k2 + k sin Θ(Û , U )k22 ,
whereby the lower bound suggests that careful analysis may yield a tighter
upper bound on kÛ − U WU k2→∞ in meaningful settings wherein kÛ −
U WU k2→∞ ≪ kÛ − U WU k2 .
We proceed to link U and Û via the perturbation framework to be established in Section 2.4 so that subsequently Û has the added interpretation of
being viewed as a perturbation of U . In that structured setting, we formulate
a Procrustean matrix decomposition (Section 3.1) by further decomposing
the underlying matrices corresponding to the quantities k sin Θ(Û , U )k2 and
k sin Θ(Û , U )k22 above. Together with machinery for the 2 → ∞ norm and
careful model-based analysis, we subsequently derive a collection of operationally significant perturbation bounds (Sections 3.2, 4.1, 4.2, and 4.3)
which improve upon existing results throughout the statistics literature.
10
J. CAPE, M. TANG, AND C. E. PRIEBE
2.4. Perturbation framework for the singular value decomposition. For
rectangular matrices X̂, X, E ∈ Rp1×p2 , the matrix X shall denote a true,
unobserved underlying matrix, whereas X̂ := X + E represents an observed
perturbation of X under the unobserved additive error E. For X and X̂,
consider their respective partitioned singular value decompositions given in
block matrix form by
⊤
Σ 0
V
(2.5)
X = U U⊥ ·
·
V⊥⊤
0 Σ⊥
and
(2.6)
X̂ := X + E = Û
Û⊥
⊤
Σ̂ 0
V̂
·
.
·
0 Σ̂⊥
V̂⊥⊤
Here U ∈ Op1,r , V ∈ Op2 ,r , [U |U⊥ ] ∈ Op1 , and [V |V⊥ ] ∈ Op2 . The matrices
Σ and Σ⊥ contain the singular values of X where Σ = diag(σ1 (X), . . . , σr (X)) ∈
Rr×r and Σ⊥ ∈ Rp1−r×p2 −r has the remaining singular values σr+1 (X), . . .
on its main diagonal, possibly padded with additional zeros, where σ1 (X) ≥
· · · ≥ σr (X) ≥ σr+1 (X) ≥ · · · ≥ 0. The use of the character ⊥ in Σ⊥ is
a simplifying abuse of notation employed for notational consistency. The
quantities Û , Û⊥ , V̂ , V̂⊥ , Σ̂, and Σ̂⊥ are defined analogously.
We note that this framework can be employed more generally when, for
example, Σ contains a collection of (sequential) singular values of interest
which are separated from the remaining singular values in Σ⊥ .
3. Main results.
3.1. A Procrustean matrix decomposition and its variants. In this section
we present our matrix decomposition and its variants. The procedure for
deriving the matrix decomposition is based on a geometric viewpoint and is
explained in Section 6.1.
Theorem 3.1. In the general rectangular matrix setting of Sections 2.3
and 2.4, the matrix (Û − U WU ) ∈ Rp1 ×r admits the decomposition
(3.1)
Û − U WU = (I − U U ⊤ )EV WV Σ̂−1
+ (I − U U ⊤ )E(V̂ − V WV )Σ̂−1
+ (I − U U ⊤ )X(V̂ − V V ⊤ V̂ )Σ̂−1
+ U (U ⊤ Û − WU ).
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 11
Moreover, the decomposition still holds when replacing the r × r orthogonal matrices WU and WV with any real r × r matrices T1 and T2 , respectively. The analogous decomposition for V̂ − V WV is given by replacing
U, Û , V, V̂ , E, X, WU , and WV above with V, V̂ , U, Û , E ⊤ , X ⊤ , WV , and WU ,
respectively.
For ease of reference we state the symmetric case of Theorem 3.1 as a
corollary. In the absence of a positive semi-definiteness assumption, the diagonal entries of Σ, Σ⊥ , Σ̂, and Σ̂⊥ then correspond to the eigenvalues of X
and X̂.
Corollary 3.2. In the special case when p1 = p2 = p and X, E ∈ Rp×p
are symmetric matrices, Theorem 3.1 becomes
(3.2)
Û − U WU = (I − U U ⊤ )EU WU Σ̂−1
+ (I − U U ⊤ )E(Û − U WU )Σ̂−1
+ (I − U U ⊤ )X(Û − U U ⊤ Û )Σ̂−1
+ U (U ⊤ Û − WU ).
Remark 3.3. To reiterate, note that by construction the orthogonal
matrix WU depends upon the perturbed quantity Û which depends upon
the error E. Consequently, WU is unknown (resp., random) when E is assumed unknown (resp., random). Since we make no distinct singular value
(or distinct eigenvalue) assumption in this paper, in general the quantity
Û cannot hope to recover U in the presence of singular value multiplicity.
Indeed, Û can only be viewed as an estimate of U up to an orthogonal transformation, and our specific choice of WU is natural given the aforementioned
Procrustes-based motivation.
Statistical inference and applications are often either invariant under or
equivalent modulo orthogonal transformations given the presence of nonidentifiability. For example, clustering the rows of U is equivalent to clustering the rows of the matrix U WU . As such, the consideration of WU does
not weaken the strength or applicability of our results in practice.
It will also prove convenient to work with the following modified versions
of Theorem 3.1 stated below as corollaries.
12
J. CAPE, M. TANG, AND C. E. PRIEBE
Corollary 3.4.
(3.3)
The decomposition in Theorem 3.1 can be rewritten as
Û − U WU = (I − U U ⊤ )(E + X)(V̂ − V WV )Σ̂−1
+ (I − U U ⊤ )E(V V ⊤ )V WV Σ̂−1
+ U (U ⊤ Û − WU ).
Corollary 3.5.
(3.4)
Corollary 3.4 can be equivalently expressed as
Û − U WU = (U⊥ U⊥⊤ )E(V⊥ V⊥⊤ )(V̂ − V V ⊤ V̂ )Σ̂−1
+ (U⊥ U⊥⊤ )E(V V ⊤ )V (V ⊤ V̂ − WV )Σ̂−1
+ (U⊥ U⊥⊤ )X(V⊥ V⊥⊤ )(V̂ − V V ⊤ V̂ )Σ̂−1
+ (U⊥ U⊥⊤ )E(V V ⊤ )V WV Σ̂−1
+ U (U ⊤ Û − WU ).
3.2. General perturbation theorems. We are now in a position to obtain a
wide class of perturbation theorems via a unified methodology by employing
Theorem 3.1, its variants, the 2 → ∞ norm machinery in Section 7.1, and
the geometric observations in Section 7.2. The remainder of this section is
devoted to presenting several such general perturbation theorems. Section 4
subsequently discusses several specialized perturbation theorems tailored to
applications in high–dimensional statistics.
Let X, X̂, E ∈ Rp1 ×p2 and WU ∈ Or be defined as in Section 2.4. Let
CX,U and CX,V denote upper bounds on k(U⊥ U⊥⊤ )Xk∞ and k(V⊥ V⊥⊤ )X ⊤ k∞ ,
respectively, and define CE,U , CE,V analogously.
Theorem 3.6 (Baseline 2 → ∞ norm Procrustes perturbation bound).
Suppose σr (X) > σr+1 (X) ≥ 0 and that σr (X) > 2kEk2 . Then
k(U⊥ U⊥⊤ )E(V V ⊤ )k2→∞
(3.5) kÛ − U WU k2→∞ ≤ 2
σr (X)
⊤
k(U⊥ U⊥ )E(V⊥ V⊥⊤ )k2→∞
+2
k sin Θ(V̂ , V )k2
σr (X)
k(U⊥ U⊥⊤ )X(V⊥ V⊥⊤ )k2→∞
+2
k sin Θ(V̂ , V )k2
σr (X)
+ k sin Θ(Û , U )k22 kU k2→∞ .
The following theorem provides a uniform perturbation bound for the
quantities kÛ −U WU k2→∞ and kV̂ −V WV k2→∞ . Corollary 3.8 subsequently
yields a bound in response to Theorem 1 in [17].
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 13
Theorem 3.7 (General perturbation theorem for rectangular matrices).
Suppose σr (X) > σr+1 (X) > 0 and that
σr (X) > max{2kEk2 , (2/α)CE,U , (2/α′ )CE,V , (2/β)CX,U , (2/β ′ )CX,V }
for some constants 0 < α, α′ , β, β ′ < 1 such that δ := (α + β)(α′ + β ′ ) < 1.
Then,
k(U⊥ U⊥⊤ )E(V V ⊤ )k2→∞
(3.6)
(1 − δ)kÛ − U WU k2→∞ ≤ 2
σr (X)
⊤
k(V⊥ V⊥ )E ⊤ U U ⊤ k2→∞
+2
σr (X)
+ k sin Θ(Û , U )k22 kU k2→∞
+ k sin Θ(V̂ , V )k22 kV k2→∞ .
If instead rank(X) = r so σr+1 (X) = 0 and provided
σr (X) > max{2kEk2 , (2/α)CE,U , (2/α′ )CE,V }
for some constants 0 < α, α′ < 1 such that δ := α × α′ < 1, then the above
bound still holds.
Corollary 3.8 (Uniform perturbation bound for rectangular matrices).
Suppose σr (X) > σr+1 (X) = 0 and that
σr (X) > max{2kEk2 , (2/α)CE,U , (2/α′ )CE,V }
for some constants 0 < α, α′ < 1 such that δ := α × α′ < 1. Then
(3.7)
kEk∞ kEk1
,
×max {kU k2→∞ , kV k2→∞ } .
(1−δ)kÛ −U WU k2→∞ ≤ 12×max
σr (X) σr (X)
4. Applications. This section presents several applications of our matrix decomposition perturbation theorems and 2 → ∞ norm machinery to
three statistical settings corresponding to, among others, the recent work in
[17], [6], and [28], respectively. We emphasize that for each statistical application, our Theorems 4.4, 4.5, and 4.9 (as well as Theorem 1.1) are obtained
via individualized, problem-specific analysis within the broader context of
a unified methodology for deriving perturbation bounds. This is made clear
in the proofs of the theorems.
In each statistical application considered in this paper, we demonstrate
how our results strengthen, complement, and extend existing work. In preparation for doing so, first consider the following structural matrix property
introduced in [7] within the context of low-rank matrix recovery.
14
J. CAPE, M. TANG, AND C. E. PRIEBE
Definition 4.1 ([7], Definition 1.2). Let U be a subspace of Rp of dimension r, and let PU be the orthogonal projection onto U . Then the coherence
of U (vis-à-vis the standard basis {ei }) is defined to be
p
maxkPU ei k22 .
(4.1)
µ(U ) :=
r i∈[p]
For U ∈ Op,r , the columns of U span a subspace of dimension r in Rp ,
so it is natural to abuse notation and interchange U with its underlying
subspace U . In this case PU = U U ⊤ , and so Propositions 7.1 and 7.6 allow
us to equivalently write
p
µ(U ) :=
kU k22→∞ .
r
Observe that 1 ≤ µ(U ) ≤ p/r, where the upper and lower bounds are
achieved for U consisting of all standard basis vectors or of vectors all with
√
magnitude 1/ p, respectively. Since the (orthonormal) columns of U each
have unit Euclidean norm (“mass”), the magnitude of µ(U ) can be viewed
as describing the coordinate-wise accumulation of mass for a collection of
orthonormal singular (or eigen) vectors.
For our purposes, the assumption of bounded coherence (equiv. incoherence) as discussed in [7] corresponds to the existence of a positive constant
Cµ such that
r
r
(4.2)
kU k2→∞ ≤ Cµ
.
p
This property arises naturally in, for example, the random orthogonal (matrix) model in [7] and corresponds to the recoverability of a low rank matrix
via nuclear norm minimization when sampling only a subset of the matrix entries. In the study of random matrices, bounded coherence is closely related
to the delocalization phenomenon of eigenvectors [35]. Further examples of
matrices whose row and column spaces exhibit bounded coherence can be
found in the study of networks. Specifically, it is not difficult to check that
this property holds for the top eigenvectors of the (non-random) low-rank
edge probability matrices corresponding to the Erdős-Rényi model and the
balanced k-block stochastic block model, among others.
Remark 4.2. We emphasize that throughout the formulation of our general results in Section 3 we never assumed the matrix X to have bounded
coherence in either of its factors U or V . Rather, by working with the
2 → ∞ norm in a Procrustes setting, our results are consequently particularly strong and interpretable when combined with this additional structural
matrix property.
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 15
4.1. Singular vector perturbation bounds: ℓ∞ and k·k2→∞ norms. In [17],
the authors specifically consider low rank matrices with distinct singular
values (or eigenvalues) whose unitary factors exhibit bounded coherence.
For such matrices, Theorems 1.1 and 2.1 in [17] provide singular vector
(eigenvector) perturbation bounds in the ℓ∞ vector norm which explicitly
depend upon the underlying matrix dimension within the singular value
perturbation setting of Section 2.4.
In this paper Corollary 3.8 formulates a straightforward perturbation
bound that is, upon further inspection, operationally in the same spirit as
Theorem 1.1 of [17]. Moreover, note that our bound on the quantity kÛ −
U WU k2→∞ immediately yields a bound on the quantities kÛ − U WU kmax
and infW ∈Or kÛ − U W kmax , thereby providing ℓ∞ -type bounds for the perturbed singular vectors up to orthogonal transformation, the analogue of sign
flips in [17] for well-separated, distinct singular values (similarly for V , V̂ ,
and WV ). Also observe that controlling the dependence of kÛ − U WU k2→∞
and kV̂ − V WV k2→∞ on one another follows from the “union bound-type”
assumptions implicitly depending upon the underlying matrix dimensions.
Again, note that our perturbation bounds hold for a wider range of model
settings which includes those exhibiting singular value (eigenvalue) multiplicity.
For symmetric matrices, we likewise improve upon [17] (Theorem 2.1).
We now make this explicit in accordance with our notation.
Theorem 4.3 ([17], Theorem 2.1). Let X, E ∈ Rp×p be symmetric matrices with rank(X) = r such that X has the spectral decomposition X =
U ΛU ⊤ where Λ = diag(λ1 , λ2 , . . . , λr ) and the eigenvalues satisfy |λ1 | ≥
|λ2 | ≥ · · · ≥ |λr | > 0. Define γ := min{|λi |−|λi+1 | : 1 ≤ i ≤ r−1}∧min{|λi | :
1 ≤ i ≤ r}.
√
• Suppose that γ > 5p(kEk∞ + 2r pkEU kmax )kU k22→∞ . Then there exists an orthogonal matrix W ∈ Or such that
(4.3)
kÛ − U W kmax ≤ 45r
2
kEk∞
γ
+
√
pkEU kmax
γ
(kU k2→∞ + pkU k32→∞ ).
• Suppose
that there exists a positive constant Cµ such that kU k2→∞ ≤
q
√
r
Cµ p and that γ > 5Cµ2 r(1 + Cµ r)kEk∞ . Then there exists an
orthogonal matrix W ∈ Or such that
(4.4)
kÛ − U W kmax
√
√
kEk∞
2
5
≤ 45Cµ r (1 + Cµ r)(1 + Cµ r)
.
√
γ p
16
J. CAPE, M. TANG, AND C. E. PRIEBE
Theorem 4.4 (Improvement of [17], Theorem 2.1). Consider the setting
of Theorem 4.3 but now where γ = 0 is permitted i.e. we allow repeated
eigenvalues.
• Suppose that |λr | > 4kEk∞ . Then there exists an orthogonal matrix
W ∈ Or such that
(4.5)
kÛ − U W kmax ≤ 14
kEk∞
|λr |
kU k2→∞ .
• Suppose
that there exists a positive constant Cµ such that kU k2→∞ ≤
q
r
Cµ p and that |λr | > 4kEk∞ . Then there exists an orthogonal matrix
W ∈ Or such that
(4.6)
√
kÛ − U W kmax ≤ 14Cµ r
kEk∞
√
|λr | p
.
Theorems 4.3 and 4.4 demonstrate that our refined analysis yields superior
bounds with respect to absolute constant factors, rank-dependent factors,
and eigengap magnitude/multiplicity assumptions.
4.2. Singular subspace perturbation and random matrices. In this section
we provide an example which interfaces our results with the recent rateoptimal singular subspace perturbation bounds obtained in [6].
Consider the setting wherein X ∈ Rp1 ×p2 is a fixed rank-r matrix with
√
r ≤ p1 ≪ p2 and σr (X) = Ω(p2 / p1 ) where E ∈ Rp1 ×p2 is a random matrix
with independent standard normal entries. Theorems 1, 2, and 3 in [6] imply
that in this setting, with high probability, the following bounds hold for the
left and right singular vectors, respectively.
√
√
p1
p2
(4.7) k sin Θ(Û , U )k2 = Θ
and k sin Θ(V̂ , V )k2 = Θ
σr (X)
σr (X)
Observe that the bound is stronger for k sin Θ(Û , U )k2 than for k sin Θ(V̂ , V )k2
with the latter quantity being more difficult to control in general. With an
eye towards the latter quantity, the following theorem demonstrates how our
analysis of kV̂ − V WV k2→∞ allows us to recover upper and lower bounds
for kV̂ − V WV kp
2→∞ in terms of k sin Θ(V̂ , V )k2 that differ
p by a factor of
√
at most Cmax{ r log(p2 ), p1 } in general and at most C r log(p2 ) under
the additional assumption of bounded coherence.
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 17
Theorem 4.5. Let X, E ∈ Rp1 ×p2 be as in Section 2.4 such that rank(X) =
√
3/2
r and r ≤ p1 ≪ p2 where σr (X) = Ω(p2 / p1 ) and p2 = Ω(p1 ). Suppose
that the entries of E are independent standard normal random variables.
Then there exists a constant C > 0 such that with probability at least 1−p−2
2 ,
!
p
√
max{ r log(p2 ), p1 }
k sin Θ(V̂ , V )k2 .
(4.8) kV̂ − V WV k2→∞ ≤ C
√
p2
If in addition kV k2→∞ = O
(4.9)
q
−2
r
p2 , then with probability at least 1 − p2 ,
kV̂ − V WV k2→∞ ≤ C
p
r log(p2 )
√
p2
!
k sin Θ(V̂ , V )k2 .
Note that the lower bound √1p2 k sin Θ(V̂ , V )k2 ≤ kV̂ − V WV k2→∞ always
holds by Proposition 7.3 and Lemma 7.7.
4.3. Statistical inference for random graphs. In the study of networks,
community detection and clustering are tasks of central interest. A network
(or, alternatively a graph G := (V, E) consisting of a vertex set V and edge
set E) may be represented, for example, by its adjacency matrix A ≡ AG
which captures the edge connectivity of the nodes in the network. For inhomogeneous independent edge random graph models, the adjacency matrix
can be viewed as a random perturbation of an underlying (often low rank)
edge probability matrix P where P = E[A] holds on the off-diagonal. In the
notation of Section 2.4, the matrix P corresponds to X, the matrix A − P
corresponds to E, and the matrix A corresponds to X̂. By viewing Û (the
matrix containing the top eigenvectors of A) as an estimate of U (the matrix
of top eigenvectors of P ), our Section 3 theorems immediately apply.
Spectral-based methods and related optimization problems for random
graphs employ the spectral decomposition of the adjacency matrix (or matrixvalued functions thereof, e.g. the Laplacian matrix and its variants). For
example, the recent paper [23] presents a general spectral-based, dimensionreduction community detection framework which incorporates the (spectral
norm) distance between the leading eigenvectors of A and P . Taken in the
context of this recent work and indeed the wider network analysis literature,
our paper complements existing efforts and paves the way for expanding the
toolkit of network analysts to include more Procrustean and 2 → ∞ norm
machinery.
Much of the existing literature for networks and graph models concerns
the popular stochastic block model (SBM) [20] and its variants. The related
18
J. CAPE, M. TANG, AND C. E. PRIEBE
random dot product graph (RDPG) model first introduced in [46] has subsequently been developed in a series of papers as both a tractable and flexible
random graph model amenable to spectral methods [18, 28, 37, 38, 39, 40,
41]. In the RDPG model, the graph eigenvalues and eigenvectors are closely
related to the model’s generating latent positions; in particular, the top
eigenvectors of the adjacency matrix scaled by its largest eigenvalues form
an estimator of the latent positions (up to orthogonal transformation).
Given the existing RDPG literature, the results in this paper extend both
the treatment of the 2 → ∞ norm in [28] and Procrustes matching for
graphs in [39]. Specifically, our 2 → ∞ bounds in Section 3 imply a version
of Lemma 5 in [28] for the (unscaled) eigenvectors that does not require the
matrix-valued model parameter P to have distinct eigenvalues. Our Procrustes analysis also suggests a refinement of the test statistic formulation
in the two-sample graph inference hypothesis testing framework of [39].
It is also worth noting that our level of generality allows for the consideration of random graph (matrix) models which allow edge dependence
structure, such as the (C, c, γ) property in [29] (see below). Indeed, moving
beyond independent edge models represents an important direction for future work in network science and in the development of statistical inference
for graph data.
Definition 4.6 ([29]). A p1 ×p2 random matrix M is said to be (C, c, γ)concentrated if, given a trio of positive constants (C, c, γ), for all unit vectors
u ∈ Rp1 , v ∈ Rp2 , and for every t > 0,
(4.10)
P [|hM v, ui| > t] ≤ C exp(−ctγ ).
Remark 4.7. The proofs of our main theorems demonstrate the importance of bounding the quantities kEV k2→∞ and kU ⊤ EV k2 in the perturbation framework of Section 2.4. Note that when E satisfies the (C, c, γ)concentrated property in Definition 4.6, then the above quantities can be
easily controlled by, for example, naı̈ve union bounds. For further discussion
of the (C, c, γ)-concentrated property and how it holds for a large class of
random matrix models, see [29].
In the network literature, current active research directions include the
development of random graph models exhibiting edge correlation and the
development of inference methodology for multiple graphs. For the purposes
of this paper, we shall consider the ρ-correlated stochastic block model introduced in [26] and the omnibus embedding matrix for multiple graphs introduced in [32] and subsequently employed in [8, 27]. The ρ-correlated stochastic block model provides a simple yet easily interpretable and tractable
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 19
model for dependent random graphs [26] while the omnibus embedding
matrix provides a framework for performing spectral analysis on multiple
graphs by leveraging graph dissimilarities [27, 32] or similarities [8].
Definition 4.8 ([26], Definition 1). Let G n denote the set of labeled, nvertex, simple, undirected graphs. Two n-vertex random graphs (G1 , G2 ) ∈
→
G 1 × G 2 are said to be ρ-correlated SBM(κ, n, b, Λ) graphs (abbreviated ρSBM) if
→
1. G1 := (V, E(G1 )) and G2 := (V, E(G2 )) are marginally SBM(κ, n, b, Λ)
random graphs; i.e. for each i = 1, 2,
(a) The vertex set V is the union of κ blocks V1 , V2 , . . . , Vκ , which
are disjoint sets with respective cardinalities n1 , n2 , . . . , nκ ;
(b) The block membership function b : V 7→ [κ] is such that for each
v ∈ V, b(v) denotes the block of v; i.e., v ∈ Vb(V ) ;
(c) The block adjacency probabilities are given by the symmetric
matrix Λ ∈ [0, 1]κ×κ ; i.e., for each pair of vertices {j, l} ∈ V2 ,
the adjacency of j and l is an independent Bernoulli trial with
probability of success Λb(j),b(l) .
2. The random variables
{I[{j, k} ∈ E(Gi )]}i=1,2;{j,k}∈(V )
2
are collectively independent except that for each {j, k} ∈ V2 , the
correlation between I[{j, k} ∈ E(G1 )] and I[{j, k} ∈ E(G2 )] is ρ ≥ 0.
The following theorem provides a guarantee for estimating the eigenvectors corresponding to the largest eigenvalues of a multiple graph omnibus
matrix when the graphs are not independent. To the best of our knowledge,
Theorem 4.9 is the first of its kind.
→
Theorem 4.9. Let (G1 , G2 ) be a pair of ρ-correlated SBM(κ, n, b, Λ)
graphs as in Definition 4.8 with the corresponding pair of n × n (symmetric,
binary) adjacency matrices (A1 , A2 ). Let the model omnibus matrix O and
adjacency omnibus matrix Ô be given by
"
#
A1 +A2
1
1 1
A
2
O :=
⊗ ZΛZ ⊤ ; Ô := A1 +A2
2
1 1
A
2
where ⊗ denotes the matrix Kronecker product and Z is the n × κ matrix
of vertex-to-block assignments such that P := ZΛZ ⊤ ∈ [0, 1]n×n denotes the
edge probability matrix.
20
J. CAPE, M. TANG, AND C. E. PRIEBE
Let r := rank(Λ) and therefore rank(O) = r. Suppose that the maximum
expected degree of Gi , i = 1, 2, denoted ∆, satisfies ∆ ≫ log4 (n) and that
σr (O) = Ω(∆). As in Section 2.4, let U, Û ∈ O2n,r denote the matrices
whose columns are the normalized eigenvectors corresponding to the largest
eigenvalues of O and Ô, respectively, given by the diagonal matrices Σ and
Σ̂, respectively. Then with probability 1 − o(1) i.e. asymptotically almost
surely in n, one has
!
p
r log(n)
kÛ − U WU k2→∞ = O
.
∆
Remark 4.10. The implicit dependence upon the correlation factor ρ in
Theorem 4.9 can be made explicit by a more careful analysis of the constant
factor and probability statement. This is not our present concern.
5. Discussion. In summary, this paper develops a flexible Procrustean
matrix decomposition and its variants together with machinery for the 2 →
∞ norm in order to study the perturbation of singular subspaces and their
geometry. We have demonstrated the widespread applicability of our framework and results to a host of popular matrix models, namely matrices with
•
•
•
•
independent, identically distributed entries (Section 4.2),
independent, identically distributed rows (Section 1.4 and 4.1),
independent, not-necessarily-identically-distributed entries (Section 4.3),
neither independent nor identically distributed entries (Section 4.3).
We emphasize that in each application discussed in this paper, the underlying problem setting demands model-specific analysis both in terms of
which formulation of the Procrustean matrix decomposition to use and how
to transition between norms. For example, using the rectangular matrix notation in this paper, recall how the assumption of bounded coherence led to
the importance of the product term kEk∞ kV k2→∞ in Section 4.1 whereas in
the case of i.i.d. normal matrices in Section 4.2 the central term of interest is
kEV k2→∞ . Similarly, in the context of covariance matrix estimation (Theorem 1.1 as well as Theorem 4.4), note how discrepancies in model specificity
and assumptions inspired different approaches in deriving the stated bounds.
Moreover, the study of k·k2→∞ directly translates to k·kmax via the relation
(5.1)
inf kÛ − U W kmax ≤ inf kÛ − U W k2→∞ ≤ kÛ − U WU k2→∞ .
W ∈Or
W ∈Or
Ample open problems and applications exist for which it is and will be
productive to consider the 2 → ∞ norm in the future. This paper details
three specific applications, namely
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 21
• singular vector estimation under perturbation (Section 4.1),
• singular subspace recovery under perturbation (Section 4.2),
• statistical estimation and inference for graphs (Section 4.3).
It is our hope that the level of generality and flexibility presented in this
paper will facilitate the more widespread use of the 2 → ∞ norm in the
statistics literature. To this end, we further invite the reader to apply and
adapt our Procrustean matrix decomposition for their own purposes.
6. Proofs.
6.1. Proof of the Procrustean matrix decomposition. Here we explain the
derivation of the matrix decomposition for Û −U WU as presented in Theorem
3.1.
Proof of Theorem 3.1. First observe that the matrices Û and U WU
are equivalently written as X̂ V̂ Σ̂−1 and XV Σ−1 WU , respectively, given the
block matrix formulation in Section 2.4. Next, the explicit correspondence
between WU and U ⊤ Û resulting from Eqn. (2.3) along with subsequent leftmultiplication by the matrix U motivates the introduction of the projected
quantity ±U U ⊤ Û and to write
Û − U WU = (Û − U U ⊤ Û ) + (U U ⊤ Û − U WU )
= (I − U U ⊤ )X̂ V̂ Σ̂−1 + U (U ⊤ Û − WU ).
The matrix U (U ⊤ Û − WU ) is shown to be small in both spectral and
2 → ∞ norm by Lemma 7.8 and via Proposition 7.5. Ignoring U for the
moment, the matrix U ⊤ Û − WU represents a geometric residual measure of
closeness between the matrix U ⊤ Û and the Frobenius-optimal orthogonal
matrix WU .
It is not immediately clear how to control the quantity (I − U U ⊤ )X̂ V̂ Σ̂−1
given the dependence on the perturbed quantity X̂. If instead we replace
X̂ with X and consider the matrix (I − U U ⊤ )X V̂ Σ̂−1 , then by the block
matrix form in Section 2.4 one can check that (I − U U ⊤ )X = X(I − V V ⊤ ).
Together with the fact that (I −U U ⊤ ) is an orthogonal projection and hence
is idempotent, it follows that
(I − U U ⊤ )X V̂ Σ̂−1 = (I − U U ⊤ )X(V̂ − V V ⊤ V̂ )Σ̂−1
So, introducing the quantity ±(I − U U ⊤ )X V̂ Σ̂−1 yields
(I − U U ⊤ )X̂ V̂ Σ̂−1 = (I − U U ⊤ )E V̂ Σ̂−1 + (I − U U ⊤ )X(V̂ − V V ⊤ V̂ )Σ̂−1 .
22
J. CAPE, M. TANG, AND C. E. PRIEBE
Note that by Lemma 7.7 and Proposition 7.5, all of the terms comprising
the matrix product (I − U U ⊤ )X(V̂ − V V ⊤ V̂ )Σ̂−1 can be controlled (submultiplicatively). In certain settings it shall be useful to further decompose
(I − U U ⊤ )X(V̂ − V V ⊤ V̂ )Σ̂−1 into two matrices as
(I − U U ⊤ )X(V̂ − V WV )Σ̂−1 + (I − U U ⊤ )XV (WV − V ⊤ V̂ )Σ̂−1 .
Note that the second matrix above vanishes given that X ≡ U ΣV ⊤ +
U⊥ Σ⊥ V⊥⊤ .
As for the earlier matrix (I − U U ⊤ )E V̂ Σ̂−1 , we do not assume additional
control over the quantity V̂ , so we rewrite the above matrix product in terms
of V and a corresponding residual quantity. A natural choice is therefore
to incorporate the orthogonal factor WV . Specifically, introducing ±(I −
U U ⊤ )EV WV Σ̂−1 produces
(I − U U ⊤ )E V̂ Σ̂−1 = (I − U U ⊤ )E(V̂ − V WV )Σ̂−1 + (I − U U ⊤ )EV WV Σ̂−1 .
Moving forward, the matrix (I −U U ⊤ )EV WV Σ̂−1 becomes the leading term
of interest. Gathering all the terms on the right-hand sides of the above
equations yields Theorem 3.1. Corollaries 3.2 and 3.4 are evident given that
U ⊤ U and V ⊤ V are both simply the identity matrix.
6.2. Proofs of general perturbation theorems.
6.2.1. Theorem 3.6.
Proof of Theorem 3.6. The assumption σr (X) > 2kEk2 implies that
σr (X̂) ≥ 12 σr (X) since by Weyl’s inequality for singular values, σr (X̂) ≥
σr (X) − kEk2 ≥ 12 σr (X). The theorem then follows from Corollary 3.5 together with Proposition 7.5 and Lemma 7.7.
6.2.2. Theorem 3.7.
Proof of Theorem 3.7. By Corollary 3.4, consider the decomposition
Û − U WU = (I − U U ⊤ )(E + X)(V̂ − V WV )Σ̂−1
+ (I − U U ⊤ )E(V V ⊤ )V WV Σ̂−1
+ U (U ⊤ Û − WU ).
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 23
Subsequently applying Proposition 7.5 and Lemma 7.7 yields
!
CE,U + CX,U
kV̂ − V WV k2→∞
kÛ − U WU k2→∞ ≤
σr (X̂)
!
1
+
k(I − U U ⊤ )EV V ⊤ k2→∞
σr (X̂)
+ k sin Θ(Û , U )k22 kU k2→∞
and similarly
kV̂ − V WV k2→∞ ≤
+
CE,V + CX,V
1
σr (X̂)
!
σr (X̂)
!
kÛ − U WU k2→∞
k(I − V V ⊤ )E ⊤ U U ⊤ k2→∞
+ k sin Θ(V̂ , V )k22 kV k2→∞
By assumption
σr (X) > max{2kEk2 , (2/α)CE,U , (2/α′ )CE,V , (2/β)CX,U , (2/β ′ )CX,V }
for constants 0 < α, α′ , β, β ′ < 1 such that δ := (α + β)(α′ + β ′ ) < 1. Note
that the assumption σr (X) > 2kEk2 implies that σr (X̂) ≥ σr (X) − kEk2 ≥
1
2 σr (X) by Weyl’s inequality for singular values. Thus, combining the above
observations, bounds, and rearranging terms yields
2
(1 − δ)kÛ − U WU k2→∞ ≤
k(I − U U ⊤ )EV V ⊤ k2→∞
σr (X)
2(α + β)
+
k(I − V V ⊤ )E ⊤ U U ⊤ k2→∞
σr (X)
+ k sin Θ(Û , U )k22 kU k2→∞
+ (α + β) k sin Θ(V̂ , V )k22 kV k2→∞ ,
whereby the first claim follows since (α + β) < 1.
When rank(X) = r, the matrix (I − U U ⊤ )X vanishes since Σ⊥ is identically zero. Corollary 3.2 therefore becomes
Û − U WU = (I − U U ⊤ )E(V̂ − V WV )Σ̂−1
+ (I − U U ⊤ )E(V V ⊤ )V WV Σ̂−1
+ U (U ⊤ Û − WU )
and similarly for V̂ − V WV , which removes the need for assumptions on
σr (X) with respect to the terms CX,U and CX,V . Hence the bound holds.
24
J. CAPE, M. TANG, AND C. E. PRIEBE
6.2.3. Corollary 3.8.
Proof of Corollary 3.8. By Theorem 3.7, we have the bound
2
(1 − δ)kÛ − U WU k2→∞ ≤
k(I − U U ⊤ )E(V V ⊤ )k2→∞
σr (X)
2
+
k(I − V V ⊤ )E ⊤ (U U ⊤ )k2→∞
σr (X)
+ k sin Θ(Û , U )k22 kU k2→∞
+ k sin Θ(V̂ , V )k22 kV k2→∞ .
Next, by Wedin’s sin Θ theorem together with the general matrix fact that
kEk2 ≤ max{kEk∞ , kEk1 } and the assumption σr (X) > 2kEk2 , we have
that
o
n
2 × max{kEk∞ , kEk1 }
max k sin Θ(Û , U )k2 , k sin Θ(V̂ , V )k2 ≤ min
,1 .
σr (X)
Using properties of the 2 → ∞ norm, we therefore have
k(I − U U ⊤ )E(V V ⊤ )k2→∞ ≤ kEV V ⊤ k2→∞ + k(U U ⊤ )E(V V ⊤ )k2→∞
≤ kEV k2→∞ + kU k2→∞ kU ⊤ EV k2
≤ kEk∞ kV k2→∞ + kU k2→∞ max{kEk∞ , kEk1 }
≤ 2 × max {kEk∞ , kEk1 } × max {kU k2→∞ , kV k2→∞ } .
Similarly,
k(I − V V ⊤ )E ⊤ (U U ⊤ )k2→∞ ≤ kEk1 kU k2→∞ + kV k2→∞ max{kEk∞ , kEk1 }
≤ 2 × max {kEk∞ , kEk1 } × max {kU k2→∞ , kV k2→∞ } .
Combining these observations yields the stated bound.
6.3. Proof of Theorem 1.1.
Proof of Theorem 1.1. In what follows the constant C > 0 may change
from line to line. First, adapting the proof of Theorem 3.6 for symmetric
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 25
positive semi-definite matrices yields the bound
kÛ − U WU k2→∞ ≤
k(U⊥ U⊥⊤ )En (U U ⊤ )k2→∞
σr (Γ̂n )
!
+
k(U⊥ U⊥⊤ )En (U⊥ U⊥⊤ )k2→∞
+
k(U⊥ U⊥⊤ )Γ(U⊥ U⊥⊤ )k2→∞
σr (Γ̂n )
σr (Γ̂n )
!
!
k sin Θ(Û , U )k2
k sin Θ(Û , U )k2
+ k sin Θ(Û , U )k22 kU k2→∞ .
Next we collect several observations.
• k(U⊥ U⊥⊤ )En (U U ⊤ )k2→∞ ≤ kU⊥ U⊥⊤ k∞ kEn U k2→∞ by Proposition 7.5,
• k(U⊥ U⊥⊤ )Γ(U⊥ U⊥⊤ )k2→∞ ≤ k(U⊥ U⊥⊤ )Γ(U⊥ U⊥⊤ )k2 = kU⊥ Σ⊥ U⊥⊤ k2 =
σr+1 (Γ),
• k sin Θ(Û , U )k2 ≤ 2kEn k2 /δr (Γ) by Theorem 7.9,
• The assumption σr (Γ) > 2kEn k implies that σr (Γ̂n ) ≥ 21 σr (Γ),
p
• The bounded coherence assumption on U yields kU k2→∞ ≤ C dr
√
together with kU⊥ U⊥⊤ k∞ ≤ (1 + C) r for some positive constant C.
By Theorems 1 and 2 in [22] applied to the random vectors Yk with covariq
ance matrix Γ, there exists a constant C > 0 such that kEn k ≤ Cσ1 (Γ) log(d)
n
with probability at least 1 − 13 d−2 . Similarly, by applying these theorems
matrix U⊥ Σ⊥ U⊥⊤ , we have
to the random vectors U⊥⊤ Yk with covariance
q
that k(U⊥ U⊥⊤ )En (U⊥ U⊥⊤ )k2 ≤ Cσr+1 (Γ) log(d)
with probability at least
n
1 −2
1 − 3 d . Combining these observations yields that with probability at least
1 − 23 d−2 ,
kÛ − U WU k2→∞
√
C rkEn U k2→∞
Cσ1 (Γ)σr+1 (Γ) log(d)
≤
+
σr (Γ)
δr (Γ)σr (Γ)
n
!
r
Cσ1 (Γ)σr+1 (Γ) log(d)
+
δr (Γ)σr (Γ)
n
r
2
Cσ1 (Γ) r log(d)
+
.
δr2 (Γ)
d n
As for the matrix (En U ) ∈ Rd×r , consider the bound
√
√
kEn U k2→∞ ≤ rkEn U kmax = r maxi∈[d],j∈[r]|hEn⊤ ei , uj i|
26
J. CAPE, M. TANG, AND C. E. PRIEBE
where for each (i, j) ∈ [d] × [r],
n
n h
i 1X
i
1 Xh ⊤
(i)
⊤
⊤
⊤
(uj Yk )(Yk ei ) − uj Γei =
hYk , uj iYk − hΓei , uj i .
hEn ei , uj i =
n
n
k=1
k=1
Denote the sub-gaussian random variable and vector Orlicz ψ2 norms by
√
kY (i) kψ2 := sup p(E[|Y (i) |p ])1/p and kY kψ2 := sup khY, xikψ2 .
p≥1
kxk2 =1
The product of (sub-)Gaussian random variables has a sub-exponential dis(i)
tribution, and in particular the term {hYk , uj iYk − hΓei , uj i} is a centered
sub-exponential random variable which is independent and identically distributed for each 1 ≤ k ≤ n when i and j are fixed. An upper bound for the
sub-exponential Orlicz ψ1 norm of this random variable is given in terms of
the sub-gaussian Orlicz ψ2 norm, ([14], Remark 5.18) namely
(i)
(i)
khYk , uj iYk − hΓei , uj ikψ1 ≤ 2khYk , uj iYk kψ1 ≤ 2khY, uj ikψ2 kY (i) kψ2 .
The random vectors Yk are mean zero multivariate normal, therefore
q
kY (i) kψ2 ≤ Cmax1≤i≤d Var(Y (i) ) := Cν(Y ).
Together with the observation that Var(hY, uj i) = u⊤
j Γuj = σj (Γ) for all
j ∈ [r], then
p
khY, uj ikψ2 ≤ C σ1 (Γ).
By Bernstein’s inequality ([14], Proposition 5.16), it follows that
"
#
r
p
r log(d)
1
P kEn U k2→∞ ≥ C σ1 (Γ)ν(Y )
≤ d−2 .
n
3
Combining this observation with the hypotheses σr+1 (Γ) = O(1) and σ1 (Γ) =
Θ(σr (Γ)) yields that with probability at least 1 − d−2 ,
!
r
C log(d)
Cν(Y )r log(d)
+
kÛ − U WU k2→∞ ≤ p
n
σr (Γ) n
σr (Γ)
!
r
log(d)
C
+
σr (Γ)
n
r
r log(d)
+ C
.
d n
q
log(d)
ν(Y )r
with probability at least
Hence, kÛ − U WU k2→∞ ≤ C √
n
σr (Γ)
1−
d−2 .
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 27
6.4. Proof of Theorem 4.4.
Proof of Theorem 4.4. Specializing Corollary 3.4 for the symmetric
case when rank(X) = r yields the decomposition
Û − U WU = (I − U U ⊤ )E(Û − U WU )Λ̂−1 + (I − U U ⊤ )E(U U ⊤ )U WU Λ̂−1
+ U (U ⊤ Û − WU ).
Rewriting the above decomposition yields
Û − U WU = E(Û − U WU )Λ̂−1 + (U U ⊤ )E(Û − U WU )Λ̂−1
+ EU WU Λ̂−1
+ (U U ⊤ )EU WU Λ̂−1
+ U (U ⊤ Û − WU ).
Applying the technical results in Sections 7.1 and 7.2 yields the term-wise
bounds
kE(Û − U WU )Λ̂−1 k2→∞ ≤ kEk∞ kÛ − U WU k2→∞ |λ̂r |−1 ,
k(U U ⊤ )E(Û − U WU )Λ̂−1 k2→∞ ≤ kU k2→∞ kEk2 kÛ − U WU k2 |λ̂r |−1 ,
kEU WU Λ̂−1 k2→∞ ≤ kEk∞ kU k2→∞ |λ̂r |−1 ,
k(U U ⊤ )EU WU Λ̂−1 k2→∞ ≤ kU k2→∞ kEk2 ||λ̂r |−1 ,
kU (U ⊤ Û − WU )k2→∞ ≤ kU k2→∞ kU ⊤ Û − WU k2 .
By assumption E√ is symmetric, therefore kEk2 ≤ kEk∞ . Furthermore,
kÛ − U WU k2 ≤ 2k sin Θ(Û , U )k2 by Lemma 7.8, and k sin Θ(Û , U )k2 ≤
2kEk2 |λr |−1 by Theorem 7.9. Therefore,
kE(Û − U WU )Λ̂−1 k2→∞ ≤ kEk∞ kÛ − U WU k2→∞ |λ̂r |−1 ,
k(U U ⊤ )E(Û − U WU )Λ̂−1 k2→∞ ≤ 4kEk2∞ kU k2→∞ |λ̂r |−1 |λr |−1 ,
kEU WU Λ̂−1 k2→∞ ≤ kEk∞ kU k2→∞ |λ̂r |−1 ,
k(U U ⊤ )EU WU Λ̂−1 k2→∞ ≤ kEk∞ kU k2→∞ |λ̂r |−1 ,
kU (U ⊤ Û − WU )k2→∞ ≤ 4kEk2∞ kU k2→∞ |λr |−2 .
By assumption |λr | > 4kEk∞ , so |λˆr | ≥ 12 |λr | and kEk∞ |λ̂r |−1 ≤ 2kEk∞ |λr |−1 ≤
28
1
2.
J. CAPE, M. TANG, AND C. E. PRIEBE
Therefore,
1
kÛ − U WU k2→∞ ,
2
≤ 2kEk∞ kU k2→∞ |λr |−1 ,
kE(Û − U WU )Λ̂−1 k2→∞ ≤
k(U U ⊤ )E(Û − U WU )Λ̂−1 k2→∞
kEU WU Λ̂−1 k2→∞ ≤ 2kEk∞ kU k2→∞ |λr |−1 ,
k(U U ⊤ )EU WU Λ̂−1 k2→∞ ≤ 2kEk∞ kU k2→∞ |λr |−1 ,
kU (U ⊤ Û − WU )k2→∞ ≤ kEk∞ kU k2→∞ |λr |−1 .
∞
Hence, kÛ − U WU k2→∞ ≤ 14 kEk
kU k2→∞ .
|λr |
6.5. Proof of Theorem 4.5.
Proof of Theorem 4.5. Note that rank(X) = r implies that the matrix (I − V V ⊤ )X ⊤ vanishes. Therefore, rewriting Corollary 3.4 yields the
decomposition
V̂ − V WV = (I − V V ⊤ )E ⊤ U WU Σ̂−1 + (I − V V ⊤ )E ⊤ (Û − U WU )Σ̂−1
+ V (V ⊤ V̂ − WV ).
Observe that (I − V V ⊤ ) = V⊥ V⊥⊤ and
k(V⊥ V⊥⊤ )E ⊤ U WU Σ̂−1 k2→∞ ≤ k(V⊥ V⊥⊤ )E ⊤ U k2→∞ σr−1 (X̂).
By Proposition 7.5 and Lemma 7.8,
k(V⊥ V⊥⊤ )E ⊤ (Û − U WU )Σ̂−1 k2→∞ ≤ k(V⊥ V⊥⊤ )E ⊤ k2→∞ kÛ − U WU k2 kΣ̂−1 k2
√
≤ 2k(V⊥ V⊥⊤ )E ⊤ k2→∞ k sin Θ(Û , U )k2 σr−1 (X̂).
Furthermore, Proposition 7.5 and Lemma 7.7 yield
kV (V ⊤ V̂ − WV )k2→∞ ≤ k sin Θ(V̂ , V )k22 kV k2→∞ .
Now consider the matrix (V⊥ V⊥⊤ )E ⊤ ∈ Rp2×p1 , and observe that its
columns are centered, multivariate normal random vectors with covariance
matrix (V⊥ V⊥⊤ ). It follows that row i of the matrix (V⊥ V⊥⊤ )E ⊤ is a centered, multivariate normal random vector with covariance matrix σi2 I where
σi2 := (V⊥ V⊥⊤ )i,i ≤ 1 and I ∈ Rp1 ×p1 denotes the identity matrix. By Gaussian concentration and applying a union
p2 ≫ p1 ,
p
p bound with the hypothesis
⊤
⊤
we have that k(V⊥ V⊥ )E k2→∞ = O( p1 log(p1 p2 )) = O( p1 log(p2 )) with
probability at least 1 − 31 p−2
2 .
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 29
As for the matrix (V⊥ V⊥⊤ )E ⊤ U ∈ Rp2×r , the above argument implies
that entry (i, j) is N (0, σi2 ). Hence
as above, we
p
p by the same arguments
have k(V⊥ V⊥⊤ )E ⊤ U k2→∞ = O( r log(rp2 )) = O( r log(p2 )) with probability at least 1 − 13 p−2
2 .
√
√
By hypothesis r ≤ p1 ≪ p2 and σr (X) ≥ Cp2 / p1 where kEk2 = O( p2 )
holds with probability at least 1 − 13 p−2
2 , hence σr (X) ≥ CkEk2 . For this
setting the rate optimal bounds in [6] are given by
√
√
p1
p2
k sin Θ(Û , U )k2 = Θ
and k sin Θ(V̂ , V )k2 = Θ
.
σr (X)
σr (X)
Combining these observations yields
!
k(V⊥ V⊥⊤ )E ⊤ U k2→∞
!
r log(p2 )
≤C
;
σr (X)
σr (X̂)
!
!
p
p1 log(p2 )
k(V⊥ V⊥⊤ )E ⊤ k2→∞
k sin Θ(Û , U )k2 ≤ C
≤C
σr2 (X)
σr (X̂)
√
p1
k sin Θ(V̂ , V )k22 kV k2→∞ ≤ C
kV k2→∞ .
σr (X)
p
3/2 p
p1
log(p2 )
p2 σr (X)
3/2
By assumption p2 = Ω(p1 ), so in the absence of a bounded coherence
assumption √1p2 k sin Θ(V̂ , V )k2 ≤ kV̂ − V WV k2→∞ and
kV̂ − V WV k2→∞ ≤ C
≤C
!
p
√
max{ r log(p2 ), p1 }
σr (X)
!
p
√
max{ r log(p2 ), p1 }
k sin Θ(V̂ , V )k2 .
√
p2
On the other hand, provided kV k2→∞ = O
tion of bounded coherence, then
√
r log(p2 )
√
k sin Θ(V̂ , V )k2 .
C
p2
√1 k sin Θ(V̂
p2
q
r
p2
under the assump-
, V )k2 ≤ kV̂ − V WV k2→∞ ≤
6.6. Proof of Theorem 4.9.
Proof of Theorem 4.9. Again we wish to bound kÛ − U WU k2→∞ .
Observe that the matrix (I − U U ⊤ )O vanishes since rank(O) = r. This fact
!
;
30
J. CAPE, M. TANG, AND C. E. PRIEBE
together with Corollary 3.2 implies the bound
kÛ − U WU k2→∞ ≤ k(I − U U ⊤ )(Ô − O)U WU Σ̂−1 k2→∞
+ k(I − U U ⊤ )(Ô − O)(Û − U WU )Σ̂−1 k2→∞
+ kU k2→∞ kU ⊤ Û − WU k2 .
The above bound can be further weakened to yield
kÛ − U WU k2→∞ ≤ k(Ô − O)U k2→∞ kΣ̂−1 k2
+ kU k2→∞ kU ⊤ (Ô − O)U k2 kΣ̂−1 k2
+ kÔ − Ok2 kÛ − U WU k2 kΣ̂−1 k2
+ kU k2→∞ kU ⊤ Û − WU k2 .
We proceed to bound all of the terms on the right hand side of the above
inequality. To this end, a straightforward calculation reveals that
kÔ − Ok2 ≤ 3 × max{kA1 − P k2 , kA2 − P k2 }.
√
For i = 1, 2, then kAi −P k2 = O( ∆) asymptotically almost surely when the
maximum expected degree of Gi , denoted ∆, satisfies ∆ ≫ log4 (n) [25] as in
the hypothesis. Furthermore, the assumption σr (O) = Ω(∆) implies σr (Ô) =
Ω(∆) asymptotically almost surely in n. Combining these observations with
the proof of Lemma 7.8 and the result of Theorem 7.9 yields the relations
CkÔ − Ok2
1
kÛ − U WU k2 ≤ Ck sin Θ(Û , U )k2 ≤
=O √
.
σr (O)
∆
It is worth noting that the above relations provide a naı̈ve bound for the
underlying quantity of interest, kÛ − U WU k2→∞ .
Next, for the matrix, (Ô − O)U ∈ R2n×r , consider the bound
√
k(Ô − O)U k2→∞ ≤ r maxi∈[2n],j∈[r]|h(Ô − O)uj , ei i|.
Note that Uk+n,j = Uk,j for all 1 ≤ k ≤ n. Now for each 1 ≤ i ≤ n and
1 ≤ j ≤ r,
h(Ô − O)uj , ei i = e⊤
i (Ô − O)uj
n
2n
X
X
1 1
Ai,k−n + A2i,k−n − 2Pi,k−n Uk,j
(A1i,k − Pi,k )Uk,j +
2
k=1
k=n+1
n
X
3 1
1
=
Ai,k + A2i,k − 2Pi,k Uk,j .
2
2
=
k=1
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 31
Observe that for n + 1 ≤ i ≤ 2n, the roles of A1 and A2 are interchanged.
For any 1 ≤ i ≤ n, the above expansion is a sum of independent (in
k), bounded, mean zero random variables taking values in [−2Uk,j , 2Uk,j ].
Hence by Hoeffding’s inequality, with probability tending to one in n,
p
k(Ô − O)U k2→∞ = O( r log(n)).
Similarly, for the matrix U ⊤ (Ô − O)U ∈ Rr×r ,
√
kU ⊤ (Ô − O)U k2 ≤ rkU ⊤ (Ô − O)U k2→∞
≤ r maxi∈[r],j∈[r]|h(Ô − O)uj , ui i|.
In particular for 1 ≤ i, j ≤ r, then
h(Ô − O)uj , ui i =
u⊤
i (Ô −
n X
n
X
O)uj =
(2A1l,k + 2A2l,k − 4Pl,k )Uk,j Ul,i
l=1 k=1
=
X
1≤l<k≤n
4(A1l,k + A2l,k − 2Pl,k )Uk,j Ul,i
This is a sum of independent, mean zero, bounded random variables taking values in [−8Uk,j Ul,i , 8Uk,j Ul,i ]. By another application of Hoeffding’s
inequality, with probability almost one,
p
kU ⊤ (Ô − O)U k2 = O(r log(r)).
Note that kU k2→∞ ≤ 1 always holds (here we do not assume bounded
coherence) and that our hypotheses imply kΣ̂−1 k2 = O(1/∆). Lemma 7.7
behaves as O(1/∆). Hence
bounds kU ⊤ Û − WU k2 by k sin Θ(Û , U )k22 which
√
r log(n)
with probability
our analysis yields that kÛ − U WU k2→∞ = O
∆
1 − o(1) as n → ∞.
32
J. CAPE, M. TANG, AND C. E. PRIEBE
References.
[1] Zhidong Bai and Jack W. Silverstein, Spectral analysis of large dimensional random
matrices, vol. 20, Springer, 2010.
[2] Konstantinos Benidis, Ying Sun, Prabhu Babu, and Daniel P. Palomar, Orthogonal
sparse eigenvectors: A Procrustes problem, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2016), 4683–4686.
[3] Rajendra Bhatia, Matrix analysis, GTM, Springer-Verlag, New York, 1997.
[4] Adam W. Bojanczyk and Adam Lutoborski, The Procrustes problem for orthogonal
stiefel matrices, SIAM Journal on Scientific Computing 21 (1999), no. 4, 1291–1304.
[5] T. Tony Cai, Zongming Ma, and Yihong Wu, Sparse PCA: Optimal rates and adaptive
estimation, The Annals of Statistics 41 (2013), no. 6, 3074–3110.
[6] T. Tony Cai and Anru Zhang, Rate-optimal perturbation bounds for singular subspaces with applications to high-dimensional statistics, preprint arXiv:1605.00353v1,
to appear in The Annals of Statistics (2016).
[7] Emmanuel J. Candès and Benjamin Recht, Exact matrix completion via convex optimization, Foundations of Computational Mathematics 9 (2009), no. 6, 717–772.
[8] Li Chen, Joshua T. Vogelstein, Vince Lyzinski, and Carey E. Priebe, A joint graph
inference case study: The C. elegans chemical and electrical connectomes, Worm 5
(2016), no. 2.
[9] Yasuko Chikuse, Statistics on special manifolds, vol. 174, Springer Science & Business
Media, 2012.
[10] Chandler Davis and William Morton Kahan, The rotation of eigenvectors by a perturbation. iii, SIAM Journal on Numerical Analysis 7 (1970), no. 1, 1–46.
[11] Ian L. Dryden, Alexey Koloydenko, and Diwei Zhou, Non-euclidean statistics for
covariance matrices with applications to diffusion tensor imaging, The Annals of
Applied Statistics 3 (2009), no. 3, 1102–1123.
[12] Ian L. Dryden and Kanti V. Mardia, Statistical shape analysis with applications in
R, John Wiley & Sons, 2016.
[13] Alan Edelman, Tomás A. Arias, and Steven T. Smith, The geometry of algorithms
with orthogonality constraints, SIAM Journal on Matrix Analysis and Applications
20 (1998), no. 2, 303–353.
[14] Yonina C Eldar and Gitta Kutyniok, Compressed sensing: theory and applications,
Cambridge University Press, 2012.
[15] Jianqing Fan, Yuan Liao, and Martina Mincheva, Large covariance estimation by
thresholding principal orthogonal complements, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 75 (2013), no. 4, 603–680.
[16] Jianqing Fan, Philippe Rigollet, and Weichen Wang, Estimation of functionals of
sparse covariance matrices, The Annals of Statistics 43 (2015), no. 6, 2706–2737.
[17] Jianqing Fan, Weichen Wang, and Yiqiao Zhong, An eigenvector perturbation bound
and its application to robust covariance estimation, arXiv:1603.03516v1 (2016).
[18] Donniell E. Fishkind, Daniel L. Sussman, Minh Tang, Joshua T. Vogelstein, and
Carey E. Priebe, Consistent adjacency-spectral partitioning for the stochastic block
model when the model parameters are unknown, SIAM Journal on Matrix Analysis
and Applications 34 (2013), no. 1, 23–39.
[19] John C. Gower and Garmt B. Dijksterhuis, Procrustes problems, no. 30, Oxford University Press, 2004.
[20] Paul W. Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt, Stochastic
blockmodels: First steps, Social networks 5 (1983), no. 2, 109–137.
[21] Roger A. Horn and Charles R. Johnson, Matrix analysis, Cambridge University Press,
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 33
2012.
[22] Vladimir Koltchinskii and Karim Lounici, New asymptotic results in principal component analysis, preprint arXiv:1601.01457 (2016).
[23] Can M. Le, Elizaveta Levina, and Roman Vershynin, Optimization via low-rank approximation for community detection in networks, The Annals of Statistics 44 (2016),
no. 1, 373–400.
[24] Jing Lei and Alessandro Rinaldo, Consistency of spectral clustering in stochastic block
models, The Annals of Statistics 43 (2015), no. 1, 215–237.
[25] Linyuan Lu and Xing Peng, Spectra of edge-independent random graphs, The Electronic Journal of Combinatorics 20 (2013), no. 4, P27.
[26] Vince Lyzinski, Information recovery in shuffled graphs via graph matching, preprint
arXiv:1605.02315 (2016).
[27] Vince Lyzinski, Youngser Park, Carey E. Priebe, and Michael W. Trosset, Fast embedding for jofc using the raw stress criterion, to appear in The Journal of Computational
and Graphical Statistics (2016).
[28] Vince Lyzinski, Daniel L. Sussman, Minh Tang, Avanti Athreya, and Carey E. Priebe,
Perfect clustering for stochastic blockmodel graphs via adjacency spectral embedding,
Electronic Journal of Statistics 8 (2014), no. 2, 2905–2922.
[29] Sean O’Rourke, Van Vu, and Ke Wang, Random perturbation of low rank matrices:
Improving classical bounds, preprint arXiv:1311.2657.
[30]
, Eigenvectors of random matrices: A survey, Journal of Combinatorial Theory, Series A 144 (2016), 361–442.
[31] Debashis Paul and Alexander Aue, Random matrix theory in statistics: A review,
Journal of Statistical Planning and Inference 150 (2014), 1 – 29.
[32] Carey E. Priebe, David J. Marchette, Zhiliang Ma, and Sancar Adali, Manifold matching: Joint optimization of fidelity and commensurability, Brazilian Journal of Probability and Statistics (2013), 377–400.
[33] Elizaveta Rebrova and Roman Vershynin, Norms of random matrices: local and global
problems, preprint arXiv:1608.06953v1.
[34] Karl Rohe, Sourav Chatterjee, and Bin Yu, Spectral clustering and the highdimensional stochastic blockmodel, The Annals of Statistics (2011), 1878–1915.
[35] Mark Rudelson and Roman Vershynin, Delocalization of eigenvectors of random matrices with independent entries, Duke Mathematical Journal 164 (2015), no. 13, 2507–
2538.
[36] G. W. Stewart and Ji-guang Sun, Matrix perturbation theory, Academic Press, 1990.
[37] Daniel L. Sussman, Minh Tang, Donniell E. Fishkind, and Carey E. Priebe, A consistent adjacency spectral embedding for stochastic blockmodel graphs, Journal of the
American Statistical Association 107 (2012), no. 499, 1119–1128.
[38] Daniel L. Sussman, Minh Tang, and Carey E. Priebe, Consistent latent position estimation and vertex classification for random dot product graphs, IEEE Transactions
on Pattern Analysis and Machine Intelligence 36 (2014), no. 1, 48–57.
[39] Minh Tang, Avanti Athreya, Daniel L. Sussman, Vince Lyzinski, Youngser Park, and
Carey E. Priebe, A semiparametric two-sample hypothesis testing problem for random
graphs, Journal of Computational and Graphical Statistics (2016).
[40] Minh Tang, Avanti Athreya, Daniel L. Sussman, Vince Lyzinski, and Carey E.
Priebe, A nonparametric two-sample hypothesis testing problem for random dot product graphs, to appear in Bernoulli (2014).
[41] Minh Tang and Carey E. Priebe, Limit theorems for eigenvectors of the normalized
laplacian for random graphs, preprint arXiv:1607.08601v1.
[42] Ulrike Von Luxburg, A tutorial on spectral clustering, Statistics and Computing 17
34
J. CAPE, M. TANG, AND C. E. PRIEBE
(2007), no. 4, 395–416.
[43] Per-Åke Wedin, Perturbation bounds in connection with singular value decomposition,
BIT Numerical Mathematics 12 (1972), no. 1, 99–111.
[44] Hermann Weyl, Das asymptotische verteilungsgesetz der eigenwerte linearer partieller
differentialgleichungen (mit einer anwendung auf die theorie der hohlraumstrahlung),
Mathematische Annalen 71 (1912), no. 4, 441–479.
[45] Jianfeng Yao, Zhidong Bai, and Shurong Zheng, Large sample covariance matrices
and high-dimensional data analysis, no. 39, Cambridge University Press, 2015.
[46] Stephen J. Young and Edward R. Scheinerman, Random dot product graph models
for social networks, International Workshop on Algorithms and Models for the WebGraph (2007), 138–149.
[47] Yi Yu, Tengyao Wang, and Richard J. Samworth, A useful variant of the Davis–
Kahan theorem for statisticians, Biometrika 102 (2015), no. 2, 315–323.
7. Supplement A. In this supplementary material, we provide technical proofs pertaining to the 2 → ∞ norm, singular subspace geometry, and
a modification of Theorem 2 in [47]. The material here plays an essential
role in the proofs of our main theorems.
7.1. Technical tools for the 2 → ∞ norm. For A ∈ Rp1×p2 , consider the
vector norm on matrices k · k2→∞ defined by
kAk2→∞ := max kAxk∞
(7.1)
kxk2 =1
Let Ai ∈ Rp2 denote the i-th row of A. The following proposition shows that
kAk2→∞ corresponds to the maximum Euclidean norm on the rows of A.
Proposition 7.1.
For A ∈ Rp1×p2 , then kAk2→∞ = max kAi k2 .
i∈[p1 ]
Proof. The definition of k · k2→∞ and the Cauchy-Schwarz inequality
together yield that kAk2→∞ ≤ max kAi k2 , since
i∈[p1 ]
(7.2)
kAk2→∞ := max kAxk∞ = max max |hAx, ei i| ≤ max kAi k2 .
kxk2 =1
kxk2 =1i∈[p1 ]
i∈[p1 ]
Barring the trivial case A ≡ 0, let e⋆ denote the standard basis vector in Rp1
with index given by arg maxi∈[p1 ] kAi k2 > 0, noting that for each i ∈ [p1 ],
−1 ⊤
⊤
Ai = e⊤
i A. Now define the unit-Euclidean norm vector x⋆ := ke⋆ Ak2 (e⋆ A).
Then
(7.3)
kAk2→∞ = max max |hAx, ei i| ≥ |hAx⋆ , e⋆ i| = ke⊤
⋆ Ak2 = max kAi k2 .
kxk2 =1i∈[p1 ]
This establishes the desired equivalence.
i∈[p1 ]
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 35
Remark 7.2. The norm k · k2→∞ is said to be subordinate with respect
to the vector norms k · k2 and k · k∞ , since for any x ∈ Rp2 , kAxk∞ ≤
kAk2→∞ kxk2 . Note, however, that k · k2→∞ is not submultiplicative for matrices in general. For example,
1 1
1 2
A=B=
and AB =
, but
0 1
0 1
√
√
kABk2→∞ = 5 > 4 = kAk2→∞ kBk2→∞ .
Proposition 7.3.
(7.4)
For A ∈ Rp1×p2 , then
√
√
kAk2→∞ ≤ kAk2 ≤ min{ p1 kAk2→∞ , p2 kA⊤ k2→∞ }
Proof. The first inequality is obvious since
kAk2→∞ = max max |hAx, ei i| ≤ max max |hAx, yi| = kAk2 .
kxk2 =1i∈[p1 ]
kxk2 =1kyk2 =1
The second inequality holds by an application of the Cauchy-Schwarz in√
equality together with the vector norm relationship kxk2 ≤ p1 kxk∞ for
x ∈ Rp1 . In particular,
kAk2 = max max |hAx, yi| ≤ max kAxk2 ≤
kxk2 =1kyk2 =1
kxk2 =1
√
p1 max kAxk∞ =
kxk2 =1
√
p1 kAk2→∞ .
By the transpose-invariance of the spectral norm we further have by symmetry that
√
kAk2 = kA⊤ k2 ≤ p2 kA⊤ k2→∞ .
Remark 7.4. The relationship in Proposition 7.3 is sharp. Indeed, for
√
the second inequality, take A := {1/ p2 }p1 ×p2 . Then kAk2→∞ = 1 and
p
√
kA⊤ k2→∞ = p1 /p2 while kAk2 = p1 . In particular, for “tall” rectangular
matrices, the spectral norm can be much larger than the 2 → ∞ norm.
Proposition 7.5.
then
For all A ∈ Rp1×p2 , B ∈ Rp2 ×p3 , and C ∈ Rp4 ×p1 ,
(7.5)
kABk2→∞ ≤ kAk2→∞ kBk2
and
(7.6)
kCAk2→∞ ≤ kCk∞ kAk2→∞ .
36
J. CAPE, M. TANG, AND C. E. PRIEBE
Proof. The subordinate property of k · k2→∞ yields that for all x ∈ Rp3 ,
kABxk∞ ≤ kAk2→∞ kBxk2 , hence maximizing over all unit vectors x yields
Equation (7.5).
In contrast, Eqn. (7.6) follows from Hölder’s inequality coupled with the
fact that the vector norms k·k1 and k·k∞ are dual to one another. Explicitly,
kCAk2→∞ = max max |hCAx, ei i| ≤ max max kC ⊤ ei k1 kAxk∞
kxk2 =1i∈[p1 ]
kxk2 =1i∈[p1 ]
⊤
max kAxk∞ = kC ⊤ k1 kAk2→∞
≤ max kC yk1
kxk2 =1
kyk1 =1
= kCk∞ kAk2→∞ .
Proposition 7.6.
(7.7)
(7.8)
For A ∈ Rr×s , U ∈ Op1 ,r , and V ∈ Op2 ,s , then
kAk2 = kU Ak2 = kAV ⊤ k2 = kU AV ⊤ k2 ,
kAk2→∞ = kAV ⊤ k2→∞ .
Moreover, kU Ak2→∞ need not equal kAk2→∞ .
Proof. The statement follows from Proposition 7.5 and the submultiplicativity of k · k2 together with the observation that U ⊤ U = Ir and
V ⊤ V = Is . In contrast, the matrices
√
√
√ √
2
1/√2 1/ √2
1 1
1/√2
(7.9)
U :=
, A :=
, UA =
0 1
1/ 2 −1/ 2
1/ 2 0
exhibit kU Ak2→∞ =
p
5/2 >
√
2 = kAk2→∞ .
7.2. Singular subspace geometric bounds.
7.3. Technical (deterministic) lemmas. Let U, Û ∈ Op×r and WU ∈ Or
denote the corresponding Frobenius-optimal Procrustes solution from Section 2.3. In what follows, we use the fact that k sin Θ(Û , U )k2 = kU⊥⊤ Û k2 =
k(I − U U ⊤ )Û Û ⊤ k2 ([3], Chapter 7).
Lemma 7.7. Let T ∈ Rr×r be arbitrary. The following relations hold with
respect to U, Û , WU , and T in terms of k · k2 and sin Θ distance.
(7.10)
(7.11)
k sin Θ(Û , U )k2 = kÛ − U U ⊤ Û k2 ≤ kÛ − U T k2 ,
1
k sin Θ(Û , U )k22 ≤ kU ⊤ Û − WU k2 ≤ k sin Θ(Û , U )k22 .
2
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 37
Proof. The matrix (Û − U U ⊤ Û ) ∈ Rp×r represents the residual of Û
after orthogonally projecting onto the subspace spanned by the columns of
U . Note that kAk22 = maxkxk2 =1 hA⊤ Ax, xi, and so several intermediate steps
of computation yield that for any T ∈ Rr×r ,
kÛ − U U ⊤ Û k22 = maxkxk2 =1 h(Û − U U ⊤ Û )⊤ (Û − U U ⊤ Û )x, xi
= maxkxk2 =1 h(I − Û ⊤ U U ⊤ Û )x, xi
≤ maxkxk2 =1 h(I − Û ⊤ U U ⊤ Û )x, xi + k(T − U ⊤ Û )xk22
= maxkxk2 =1 h(Û − U T )⊤ (Û − U T )x, xi
= kÛ − U T k22 .
On the other hand, by Proposition 7.6 it follows that
kÛ −U U ⊤ Û k2 = kÛ Û ⊤ −U U ⊤ Û Û ⊤ k2 = k(I−U U ⊤ )Û Û ⊤ k2 = k sin Θ(Û , U )k2 .
The second matrix (U ⊤ Û − WU ) ∈ Rr×r may be viewed as a residual measure of the extent to which U ⊤ Û is “almost” the optimal rotation matrix
WU , where “optimal” is with respect to the Frobenius norm Procrustes
problem as before. The unitary invariance of k · k2 together with the interpretation of canonical angles between Û and U , denoted {θi }i where
cos(θi ) = σi (U ⊤ Û ) ∈ [0, 1] yields
kU ⊤ Û − WU k2 = kUU ΣU VU⊤ − UU VU⊤ k2 = kΣU − Ir k2 = 1 − mini cos(θi ).
Thus, both
kU ⊤ Û − WU k2 ≤ 1 − mini cos2 (θi ) = maxi sin2 (θi ) = k sin Θ(Û , U )k22
and
1
1
1
kU ⊤ Û − WU k2 ≥ (1 − mini cos2 (θi )) = maxi sin2 (θi ) = k sin Θ(Û , U )k22 .
2
2
2
Lemma 7.8.
(7.12)
The quantity kÛ − U WU k2 can be bounded as follows.
k sin Θ(Û , U )k2 ≤ kÛ − U W2⋆ k2 ≤ kÛ − U WU k2
and
(7.13)
kÛ − U WU k2 ≤ k sin Θ(Û , U )k2 + k sin Θ(Û , U )k22 .
Moreover, together with Lemma 1 in [6],
(7.14)
√
kÛ − U WU k2 ≤ min{1 + k sin Θ(Û , U )k2 , 2}k sin Θ(Û , U )k2 .
38
J. CAPE, M. TANG, AND C. E. PRIEBE
Proof. The lower bound follows from setting T = W2⋆ in Lemma 7.7
together with the definition of W2⋆ . Again by Lemma 7.7 and together with
the triangle inequality,
kÛ − U WU k2 ≤ kÛ − U U ⊤ Û k2 + kU (U ⊤ Û − WU )k2
≤ k sin Θ(Û , U )k2 + k sin Θ(Û , U )k22
The proof of Lemma 1 in [6] establishes that
inf kÛ − U W k2 ≤ kÛ − U WU k2 ≤
W ∈Or
√
2k sin Θ(Û , U )k2 .
This completes the proof.
7.4. Modification of Theorem 2 in [47]. Below we prove a modified version of Theorem 2 in [47] stated in terms of k sin Θ(V̂ , V )k2 rather than
k sin Θ(V̂ , V )kF . Although the original theorem implies a bound on the
quantity k sin Θ(V̂ , V )k2 , here we are able to remove a multiplicative factor depending on the rank of V . Our proof approach combines the original
argument together with classical results in [36]. The statement of the theorem and its proof below interface the notation in Section 2 with the notation
in Section 2 of [47].
Theorem 7.9 (Modification of [47], Theorem 2). Let X, X̂ ∈ Rp×p be
symmetric matrices with eigenvalues λ1 ≥ · · · ≥ λp and λ̂1 ≥ · · · ≥ λ̂p ,
respectively. Write E := X̂ − X and fix 1 ≤ r ≤ s ≤ p. Assume that
δgap := min(λr−1 − λr , λs − λs+1 ) > 0 where λ0 := ∞ and λp+1 := −∞. Let
d = s−r+1 and let V := [vr |vr+1 | . . . |vs ] ∈ Rp×d and V̂ := [v̂r |v̂r+1 | . . . |v̂s ] ∈
Rp×d have orthonormal columns satisfying Xvj = λj vj and X̂ v̂j = λ̂j v̂j for
j = r, r + 1, . . . , s. Then
2kEk2
.
(7.15)
k sin Θ(V̂ , V )k2 ≤
δgap
Proof. Let Λ, Λ̂ ∈ Rd×d be the diagonal matrices defined as Λ :=
diag(λr , λr+1 , . . . , λs ) and Λ̂ := diag(λ̂r , λ̂r+1 , . . . , λ̂s ). Also define Λ⊥ :=
diag(λ1 , . . . , λr−1 , λs+1 , . . . , λp ) and let V⊥ ∈ Op,p−d be such that P :=
[V |V⊥ ] ∈ Op and P ⊤ XP = diag(Λ, Λ⊥ ). Observe that X̂ V̂ − V̂ Λ̂ since
Λ̂ = V̂ ⊤ X̂ V̂ . Then
0 = X̂ V̂ − V̂ Λ̂ = (X V̂ − V̂ Λ) + (X̂ − X)V̂ − V̂ (Λ̂ − Λ).
THE TWO-TO-INFINITY NORM AND SINGULAR SUBSPACE GEOMETRY 39
By an inequality due to Weyl ([36], Corollary IV.4.9) and properties of the
spectral norm, then
kX V̂ − V̂ Λk2 ≤ k(X̂ − X)V̂ k2 + kV̂ (Λ̂ − Λ)k2
≤ kX̂ − Xk2 + kΛ̂ − Λk2
≤ 2kX̂ − Xk2 = 2kEk2 .
In summary,
kX V̂ − V̂ Λk2 ≤ 2kEk2 .
Finally, by an application of Theorem 3.6 in [36], it follows that
!
kX V̂ − V̂ Λk2
.
k sin Θ(V̂ , V )k2 ≤
δgap
Combining the above two inequalities yields the result.
Department of Applied Mathematics and Statistics
Johns Hopkins University
3400 N. Charles Street
Baltimore, Maryland 21218
USA
E-mail: [email protected]
[email protected]
[email protected]
| 10math.ST
|
A Study of Language Usage Evolution in Open Source
Software
Siim Karus
Harald Gall
University of Tartu, Estonia
University of Zurich, Switzerland
University of Zurich
Switzerland
[email protected]
[email protected]
ABSTRACT
The use of programming languages such as Java and C in Open
Source Software (OSS) has been well studied. However, many
other popular languages such as XSL or XML have received minor
attention. In this paper, we discuss some trends in OSS
development that we observed when considering multiple
programming language evolution of OSS. Based on the revision
data of 22 OSS projects, we tracked the evolution of language usage
and other artefacts such as documentation files, binaries and
graphics files. In these systems several different languages and
artefact types including C/C++, Java, XML, XSL, Makefile,
Groovy, HTML, Shell scripts, CSS, Graphics files, JavaScript, JSP,
Ruby, Phyton, XQuery, OpenDocument files, PHP, etc. have been
used. We found that the amount of code written in different
languages differs substantially. Some of our findings can be
summarized as follows: (1) JavaScript and CSS files most often coevolve with XSL; (2) Most Java developers but only every second
C/C++ developer work with XML; (3) and more generally, we
observed a significant increase of usage of XML and XSL during
recent years and found that Java or C are hardly ever the only
language used by a developer. In fact, a developer works with more
than 5 different artefact types (or 4 different languages) in a project
on average.
Categories and Subject Descriptors
D.2.7 [Software Engineering]: Distribution, Maintenance, and
Enhancement – Restructuring, reverse engineering, and
reengineering, version control; D.3.2 [Programming Languages]:
Language Classifications – object-oriented languages, extensible
language; K.2 [Computing Milieux] History of Computing –
Software, People
1. INTRODUCTION
There has been a lot of effort put into studying the use of procedural
languages such as C and object-oriented languages such as Java.
Even less common languages such as Perl, Python, or Ruby have
received their fair share of attention. However, when looking at the
statistics of most used languages, a language far more common than
any of the ones mentioned earlier, strikes out. According to
ohloh.net 1 which tracks more than 400,000 open source software
(OSS) repositories, about 15% of actively developed OSS projects
contain XML while less than 10% contain HTML, and other
languages are present in less than 8 % of projects. Even more, XML
is also the language with the most lines of code changed per month.
The use of XML in OSS projects, however, has not received
considerable attention so far.
As XML is a mark-up language, having only little meaning on its
own, it would be interesting to understand, what other language it is
being used with. Looking at co-evolving file types, we could
investigate that issue. Even more general, the question of which
languages and file types are used together and, therefore, are coevolving in OSS projects can be formulated.
To address this research question, we studied 22 OSS software
repositories over 12 years. Our study focused on two levels of file
type couplings: developer and commit level. On the developer level,
developers in the projects were studied regarding their language
experience in the projects. For that, we addressed the following
questions:
General Terms
Management,
Measurement,
Documentation,
Experimentation, Human Factors, Languages.
Which languages and artefacts are commonly used in OSS
development and in what proportions?
How many file types does a developer typically work with
and are there some usage patterns for file types?
How has the language usage and, as a consequence, the
language expertise requirements for developers changed
during the observation period?
Design,
Keywords
Programming language, Open source software, evolution, software
archives.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
Conference’10, Month 1–2, 2010, City, State, Country.
Copyright 2010 ACM 1-58113-000-0/00/0010…$10.00.
At the commit level, co-changing files appearing together in
commits were studied. For that, we addressed the following
questions:
Which co-evolution patterns can be observed in OSS
projects (e.g., are there distinct dependencies between
languages or artefact types commonly edited together)?
How have the dependencies between file types used in the
projects changed during the observation period?
Additionally, on a more general level of OSS projects studied, we
were interested in what are the most common languages or artefact
1
http://www.ohloh.net
TABLE 1.
OVERVIEW OF OSS PROJECTS USED IN THE STUDY.
Project Name
Type
Period Studied
cocoon
business
2003 - 2003
18
22
99
3575
commons
business
2007 - 2009
45
31
2981
6029
esb
business
2007 - 2009
28
23
1419
922
httpd
business
1996 - 1997
10
6
99
79
Zope
business
1996 - 1997
3
3
100
26
wsas
business
2007 - 2009
37
16
1517
1356
wsf
business
2007 - 2009
37
27
3642
4836
bibliographic
desktop
2003 - 2008
4
7
477
155
bizdev
desktop
2003 - 2009
6
6
129
9
dia
desktop
1997 - 2009
152
12
4196
3042
docbook
desktop
2000 - 2009
30
26
7540
6612
docbook2X
desktop
1999 - 2007
2
16
1082
304
exist
desktop
2002 - 2009
39
30
7116
4651
fbug-read-only
desktop
2007 - 2007
2
12
23
291
feedparser-read-only
desktop
2004 - 2009
5
10
263
4651
gnome-doc-utils
desktop
1999 - 2009
127
11
1032
328
gnucash
desktop
1997 - 2009
21
21
11757
3855
groovy
desktop
2003 - 2009
61
28
8339
5583
nltk-read-only
desktop
2001 - 2001
3
8
98
83
subversion
desktop
2000 - 2000
3
6
99
52
tei
desktop
2001 - 2009
14
26
5772
3832
valgrind
desktop
2002 - 2009
21
15
6857
3339
types in the 22 OSS projects. Our observations clearly show some
trends: (1) JavaScript and CSS files most often co-evolve with
XSL; (2) almost every Java developer but only every second C
developer works with XML; (3) over the years a significant
increase of XSL and XML usage can be observed showing
technological shifts due to framework development.
The paper is organized as follows. In section 2, the OSS projects
used in the study are introduced and described. Section 3 details the
findings about developers and Section 4 discusses our findings
about co-evolution of different types of language usage in OSS
projects. Threats to validity are outlined in Section 5 and related
work is discussed in Section 6. We conclude with our results and
give a brief outlook onto future work.
2. Dataset
To study development patterns, a dataset of 22 OSS projects was
used. The projects were split into desktop type and business (server)
type projects by their nature. That is, projects offering business
functionality such as web services were considered to be business
type projects and projects mainly used in desktop environments
were considered to be of type desktop. Table 1 shows the periods
studied, number of developers (Devs), number of different artefact
types (Art.Types) used, number of commits/revisions (Revs) and
files for each of the projects used in this study. The number of files
stated in the table includes all files including those that were deleted
Dev.-s
Art. Types
Rev.-s
Files
during the course of the projects and are not present in the latest
revision of the corresponding project.
The projects were chosen so that they would represent a wide
spectrum of development projects in terms of type, duration,
development team size, and usage scenario. Whilst business type
projects commons, esb, wsas, and wsf belong to a larger complex
super-project called WSO2 and bizdev and bibliographic are
utilities for OpenOffice, the rest of the projects were mostly
independent from each other. Docbook, docbook2X, and gnomedoc-utils represent documentation development tools. Exist [1],
feedparser-read-only, groovy, tei, subversion, nltk-read-only
(natural language toolkit), fbug-read-only (firebug) and valgrind are
projects for software project development aids or libraries. httpd,
Zope, and cocoon are application development platforms. Gnucash
is an accounting application, and dia is a diagramming solution.
To better understand, how well the dataset represents the
population, the dataset was compared with graphs publicly available
from ohloh.net. In both cases the usage of C/C++ displayed steep
decrease in its usage share and Java presented sudden emergence
and strong yet no longer growing presence. The share of commits to
XML files was increasing and had reached the highest share of file
types used. The main difference between the dataset used and
ohloh.net data was the lower usage of HTML in our dataset. The
dataset used in the study accordingly exhibited higher share of
XML and Java compared to ohloh.net data. The distribution of
Figure 1. Distribution of major file types worked on in projects per year.
project configuration or by the build tools (e.g. ant or
Maven) to store project build configuration. The
distribution of these subtypes is shown in Figure 3.
major artefacts worked on in the projects in out dataset during
different years is shown on Figure 1.
We identified and classified 45 major file types of the most
common file extensions in our repository: Archive, Audio, awk,
Binary, C, C#, C++, Command Script, CSS, Data, DTD, Graphics,
Groovy, HTML, Java, JavaScript, jsp, Makefile, Manifest, MS
Office, No extension, OpenDocument, OpenXML, Patch and Diff,
PDF, Perl, PHP, Plaintext, PostScript, Project, Properties, Python,
Resources, Rich Text, Ruby, sed, Shell Script, SQL, SQML, TeX,
WSDL, XML, XML Schema, XQuery, and XSL. Other languages
were present with only very few files.
Most of the files were classified by their extensions; however, there
were some exceptions:
The category plaintext includes files with the extensions
.txt, .readme, .changes, .install and files named
“README”, “INSTALL”, “TODO”, “COPYING”,
“COPYRIGHT”,
“AUTHORS”,
“LICENSE”,
“ACKNOWLEDGEMENTS”,
“NEWS”,
“NOTES”,
“ChangeLog”, and “CHANGES”. These files contain
project documentation in plain text format.
The category Project contains XML files with root element
“project”. These files are mostly used by IDEs to store
The category Manifest contains files with the extension
“.manifest” and files named “manifest.xml”.
The category Properties contains files with extension
“properties” and XML files with root element “properties”.
These files are used in Java projects to store application
configuration.
The category “Perl” additionally contains extensionless
text files which begin with “#!/usr/bin/perl”.
The category “Shell Scripts” additionally contains
extensionless text files which begin with “#!/bin/sh”, “#!
/bin/sh”, or “#!/bin/bash”.
The category “SGML” additionally includes “catalog”
files.
Every file can belong only to one category at once. For example,
XSL, XAML, XHTML, etc. files were not counted as XML files,
neither are files that are included in other categories due to
exceptions (e.g. files named “manifest.xml”, which belong to
category Manifest). Another special general group is “files without
Figure 2. Proportion of developers generating different types of artefacts during different years.
TABLE 2.
MOST ABUNDANT FILE TYPES USED TOGETHER BY DEVELOPERS.
extensions”, which includes folders due to the differences in how
the repositories present their data.
The data was gathered in May 2009 and contains revision
information from February 1996 to April 2009 (12 years).
3. Developers
To study the habits of developers and find language usage sets
commonly present in the projects, we extracted developer
information from revision data in the revision control systems (CVS
and SVN). We then listed the file types used by each developer and
analyzed the data.
3.1 Languages Used
identified as plaintext files. Files without extensions (mostly
changes to directory structure) were edited by 37% of developer.
Makefiles and XML were used by 34% of the developers, making
these artefact types share the third and the fourth position. Java files
were edited by 26% of developers, followed by the popularity of
project files (21%) and HTML files (19%). Surprisingly, C/C++
files were used by fewer developers (14%) than XSL files (15%).
Considering that XSL has gained popularity while C/C++ has lost
its, it can be said that in the more recent years there are more active
XSL developers than there have been C/C++ developers. The ratio
of developers using different file types throughout the study period
is shown in Figure 2. Note that year 2009 figures only account for
the first quarter of the year (data collection point).
The most popular artefact type used by 64% of developers was
Figure 3. Distribution of project file subtypes in the dataset and the last years (ns marks revisions with explicit namespace).
Figure 4. Additional languages used in at least 20% of commits by C developers.
Table 2 displays the most common artefacts commonly used by the
same developer. The most common combination of file types used
by a developer was Java and XML. This can be explained by both
of these languages being in the top four artefact types encountered.
Among language pairs used by more than 10% of developers,
Plaintext and XML were the most popular second languages.
Interestingly, only 85% of XSL developers modified XML files.
This could be caused by XSL applied on either XML files with
extensions other than “xml” or by XSL being used to transform
documents created by the runtime or received from third party.
XML Schema editors were also active in XML development in 98%
of cases (followed by WSDL, Archive file, Properties file,
JavaScript and CSS developers). XSL was commonly seen together
with languages used in web development (i.e. HTML, XML
Schema, CSS, JavaScript and graphics files).
One should keep in mind that making commits to certain type of
files does not necessarily mean that the developer has expertise in
the responding field. The commits could be deferred from other
developers or be just necessities solved with the help of other
developers. The identification of expertise is a complex task studied
in other works like [2] and [3].
These sets of commonly co-appearing languages along with the
popularity of languages allow us to identify major classes of
developers by the languages they use. The three major classes
defined by the most popular languages are C/C++ developers, Java
developers and XML developers.
3.1.1 C/C++ Developers.
C/C++ developers were frequent users of plaintext files (used by
85% of C/C++ developers) and Makefiles (82%). This is expected
as plaintext files were commonly used to document C/C++ projects
while Makefiles were the chosen technology to control the C/C++
build process. Files without extensions were modified by 79% of
C/C++ developers, which can be explained by a decent folder
structure. The fourth most common language used by C developers
was Shell scripts (58% of C developers) followed closely by XML
(52% of C developers).. Further details about most abundant file
types used by developers can be seen in Table 2. The matrix shows
for developers using file type specified in rows the percentage of
developers also using file type specified by column (e.g. 42% of
developers of archive files also worked with XSL files).
During the late 1990s, most developers had worked with C, written
Makefiles and created some other types of artefacts. Since then, less
than half of the developers have written C or C++ code (after
dropping to 9% in 2005, the percentage of developers using C or
C++ has climbed steadily to 16% in 2009) and Makefiles have been
continuously become less popular dropping from 63% in 2002 to
19% in 2009 (see Figure 2).
The most commonly used language by C developers (apart from C
itself) has almost always been Makefiles and most commonly used
file type “plaintext” (see Figure 4). Nevertheless, the popularity of
plaintext files is slowly decreasing among C developers. Use of
XML has made a strong impression since its adoption in 1998 and
has reached more than 60% of C developers in 2005. This could be
related to more widespread adoption of XML standards and XML
replacing Makefile based building environments.
3.1.2 Java Developers.
A total of 79% of Java developers also worked on XML files,
making XML the most popular language used together with Java.
The second most popular language used together with Java was
Project files, which was used by 72% of Java developers. The top
three also includes files without extensions (directory structure
modifications), which were used by 71% of Java developers. The
next popular file types were used significantly less (see Table 2).
Usage of Java has been on the rise with more than 30% of
developers having used it in the last study period. As shown in
Figure 5, Java developers use more different types of artefacts than
C/C++ developers. The graph also displays that Java developers are
writing XSL by themselves less frequently than they used to. As
XSL has become more popular in general, it can be explained by
XSL being written by developers more focused on XSL and less on
Java
It is wrong to assume that the popularity of XML in Java projects is
mainly due to project build files. In fact, more than half of the .xml
files found in Java projects were of project-specific types. Also, the
use of binary files (including .class and .jar files) by Java
developers has dropped below 20%. This could be a result of using
separate library repositories instead of having all files in revision
control repository.
Figure 5. Additional languages used in at least 20% of commits by Java developers.
3.1.3 XML Developers.
Knowledge of developing XML files has also been on a steady rise
with more than 40% of the developers having used it in the period
2008-2009.
XML developers come from different areas and work with variety
of different artefacts. This is shown by the fact that only there are
lots of different artefact types used in more than 20% of commits by
XML developers (Figure 6). The most popular file type modified by
XML developers is “files without extensions” (used by more than
60% of XML developers), which has been slowly losing its
popularity among XML developers since 2002.
It is a good practice for .xml files to have explicitly defined
namespace(s), which can be used to verify the files. However, 31%
of .xml files did not specify their namespace. The most often
encountered namespaces were http://www.tei-c.org/ns/1.0 (29%)
and http://docbook.org/ns/docbook (17%), showing that XML is
often used for project or domain specific languages. This is
confirmed by the most popular root elements: <refentry> (either in
http://docbook.org/ns/docbook namespace or no namespace
specified, 27% of all root elements) and <elementSpec> (in
http://www.tei-c.org/ns/1.0 namespace, 21% of all root elements).
The most common root element in XML files used with Java was
<project> (8% of XML files). These files were classified into
“Project” category which mostly contained files without explicit
namespace (see Figure 3).
XSL, which is commonly accompanied by XML, has been used by
a steady 10% of developers since its introduction in 2000 with
another 10% gain since 2007.
3.2 The First Commit
While commits in general tell us about the file types used in the
projects, the first commit made by a developer tells a lot about the
initial experience and start-out of developers. We can expect
developers to prefer file types and languages they are more familiar
with when joining the development team.
The first commit also shows the patterns of how developers get
involved or build up their contribution. It is expected that a
developer using more different languages in its first commit needs
to understand the project’s architecture and build practices better
than a developer who starts by just changing a few lines in a single
file.
The number of different types of files in the developers’ first
Figure 6. Additional languages used in at least 20% of commits by XML developers.
commit was usually less than four with almost half being commits
of a single file type. There is also a trend of using fewer files and
file types in the first commit towards the end of the study period
(see Figure 7).
In the 1990s C and Makefile were the most popular choices as the
first commit language, but in the later years XML and Java have
taken the lead. The number of different file types present in the first
commit and the number of files in developers’ first commit are
lower in 2008-2009 than in the 1990s. In 2008-2009 75% of first
commits were of files of single type whilst only 15% of commits
made in 1990-s were of single file type.
The analysis of the most common file type combinations shows that
12% of first commits included both Makefiles and files without
extensions (3% contained no other file types). Most common single
file type commits were for files without extension (8%), Java (8%),
and XML (7%).
When considering all commits, 21% contained both Makefiles and
files without extensions (only 2% contained only these two and 2%
contained also C files). Java and XML files were encountered in 4%
of all commits (1% contained only these). The most common single
type commit file types were Java, XML, HTML and XSL (all
accounted for only 1% of commits). This shows that developers
expand their competences by learning and deploying new languages
during the project; however, these languages will be tightly coupled
causing files of different file types to be changed at a time.
Most developers started with XML (16%), Makefiles (17%) or Java
(15%), but rarely used only one language in their first commit.
4. Co-Changes
To find out, which file types or languages have been used together
and which files are co-changed, the file types for each commit were
analyzed. The common file types committed together were
identified for both project types business and desktop separately.
4.1 Business Type Projects
The most commonly encountered combination present in business
projects was a combination of Java and XML. About 15% of all
commits made to Java files were accompanied by changes to XML
files and 32% of changes to .xml files co-occurred with changes to
Java files.
The strongest bidirectional relation was found to be between
JavaScript and XSL files. More than 40% files of these types were
co-changed. Also, in 42% of cases, a change to a CSS file was
accompanied by a change to an XSL file. Other co-occurrences
were much less frequent even in case of web file types (e.g. changes
to CSS files were accompanied with changes to Javascript files in
only 36% of the cases). One reason for that could be that in the
projects studied XSL was mainly used for generating reports and
data presentations in web applications. This usually results in XSL
being used in place of writing HTML directly and thus gets changed
often during user interface development and testing. On the other
hand, co-change rate of less than 45% with other presentation type
artefacts (Javascript, CSS or graphics files) indicates that business
type projects in the dataset used XML for business document
transformations about as often as they used these for generating
presentations.
Commits, which contained multiple files of the same type were
most frequently commits of files of type graphics (53% of all
graphics commits), C (52% of all C commits), XML Schema
(50%), PHP (48%), Java (47%) and binary and XSL (both 43%).
That is, 53% of commits of graphics files contained more than one
graphics file. XML Schema files were more often changed along
with Java or XML files than any other XML Schema file. Similarly
WSDL files were changed with XML or Java files more often than
with any other WSDL file. Web file types such as CSS and
JavaScript were more often committed with XSL files than with any
other of the same kind. This means that graphics developers,
C/C++, and XML Schema developers are more likely to work in
patches than other developers.
We also found that changes to binary files (e.g. .class, .o and .dll
files) were on average accompanied by changes to files in almost
four other file types while commits to Java files were accompanied
by the average of 0.57 artefacts of other types. This could be caused
by developers committing compiled files along with source code.
Frequently co-changing file types also include XML Schema (3.7
other file types), and WSDL (2.4 file types). Files usually not cochanging were of types PHP (0.1 other file types), C (0.4 other file
types), and Java (0.6 other file types).
The most commonly encountered file types in multiple file type
commits are in order of frequency: Java (on average, present in
27% of commits with files of other type), Project (15% of commits
Figure 7. Number of file types in developer first commit by year.
TABLE 3.
ARTEFACT TYPES COMMIT TOGETHER IN BUSINESS TYPE PROJECTS.
with files of other type), XML (13%), and XSL (11%), and files
without extensions (11%). The details of co-changes in business
type projects can be seen in Table 3. The table shows, how many
commits containing artefacts of the type listed in row header
contained artefacts of the type listed in columns.
Co-change trends. In summary, we have observed the following
language usage trends in our dataset: (1) Java and XML files coevolve most often compared to the other file and language types,
whereas C files rarely co-evolve with any other file type; (2) Binary
files co-evolve with XML and Java files in most of the cases; (3)
WSDL files often co-change with Java and XML files; (4)
JavaScript files co-evolve with XSL files; (5) XML Schema files
co-changes with WSDL, XML and Java files; and (6) XSL files
basically only co-change with JavaScript files.
4.2 Desktop Type Projects
In desktop projects, C was historically the most common language
in our dataset. As such, files representing languages and file types
related to C development were commonly changed together. For
example, changes to C files were accompanied by changes to files
without extensions (e.g. folders, Linux executables) in 32% of the
cases and with changes to Makefiles in 14% of the cases. A similar
observation was made with graphics files, which were committed
together with Makefiles in 29% of the cases and with changes to
files having no extensions in 40% of the cases. As opposed to
business type projects, changes to Java files were accompanied by
changes to XML files only in 3% of the cases. Details of these cooccurrences can be seen in Table 4. The most common co-change
pattern was observed for Groovy files, which were co-changed with
Java files in about half of the cases.
Graphics commits had the most diversity of co-changed artefacts
(2.6 other file types were committed with graphic files on average),
followed by command scripts (2.2 file types), JavaScript (1.9 file
types), and Binary files (1.9 file types). Binary files were committed
together with Java or XML files in more than third of the cases. The
most independent file types were Java (co-changed with 0.4 file
types on average) and XSL (co-changed with 0.4 file types on
average).
Multiple file type commits most often contained files without
extensions, XML, Java, XSL and C files. Multiple graphics files
were co-committed in 62% of the cases (i.e. 62% of commits with
graphics files contained more than one graphics file). Other file
types often changed in a bulk (i.e. with multiple files in a commit)
were C (48% of all C commits), Binary (47%), PHP (46%), and
Java (42%) files.
Co-change trends. In summary, we have observed the following
language co-change trends: (1) Binary files co-change with Java
and XML files; (2) C files with Makefiles; (3) Command Scripts
with Makefiles, Shell Script and XML files; (4) CSS files with
XML; (5) Groovy with Java files; (6) JavaScript with CSS and XSL
files; and (7) Ruby files co-change with XSL.
One of the major differences to business (server) type projects is
that in desktop OSS projects we observed much lower co-evolution
of Java and XML files (in either direction the co-change was half as
likely as in business type projects). On the other hand, CSS files cochanged with XML files twice as often in desktop projects. Other
trends are similar for both OSS project types investigated.
5. Threats to validity
Threats to the validity of our work are confounding and selection
(bias and generalisability).
TABLE 4.
ARTEFACT TYPES COMMIT TOGETHER IN DESKTOP TYPE PROJECTS.
Confounding is an internal threat to the explanations given to some
observations. That is, there might be some event in the society that
have changed the characteristics of developers or languages used
(e.g. companies campaigns to push their technologies), that we can
not directly relate to the dataset, which makes these relationships
difficult or impossible to identify. The impact of these events might
end up attributed to some other change we could find correlation
with. This threat cannot be avoided.
Weißgerber et al. have built a plug-in for Eclipse to show how
likely different files are to be changed together [6]. Their tool does
not exclude any files. However, they aim to visualise patterns
emerging in specific projects regarding the co-evolution of files.
They do not try to describe co-evolution on file type or artefact type
level. As such their tool is useful for monitoring software
development processes. In contrast, our paper explains more
general patterns spanning through OSS software projects.
Selection threat is both internal (bias) and external
(generalisability). It is internal as the selection might be biased
towards certain projects (e.g. by motivation). We do accept that the
dataset studied has somewhat elite collection of projects as there
were no single developer projects, which account to about a half of
the population of all OSS projects [4]. We have validated the
representativeness of our dataset against the data provided by
ohloh.net and found the general characteristics of these datasets to
be similar despite the threat of bias. We found no differences in the
artefact popularity rankings and the biggest difference observed was
the popularity of HTML code. This similarity gives high confidence
to the generalisability and representativeness of the results of this
study.
Dattero et al. conducted a survey during 2000-2001 and looked into
differences by the developer gender [7]. They discovered that
female developers are more likely to work with deprecated
technologies. They also found that female developers tend to be less
experienced and are familiar with only 2.53 languages as opposed
to 3.25 languages male developers were familiar with. These
numbers are similar to our findings, however, we also saw that the
average number of different file types (usually representing
different technologies) used by developers has decreased during the
period studied.
6. Related Work
The idea of studying cross-file co-changes has been addressed by
some research so far. However, these studies have been often
language specific and they rarely look at different file types. Even
studies encompassing multiple file types have been limited to
specific file types. For example, Zimmerman et al. studied how
lines of different files evolve in a project [5]. Their study is limited
to textual files and focused more on visualisation and clustering of
files based on their change history.
The different patterns of evolution of OSS have been outlined by
Nakakoji et al. [8]. They determined that there are three main types
of OSS: exploration-oriented, utility-oriented, and service-oriented.
These types determine how the software evolves and how the
developers behave. The projects studied here spanned over all these
types – business projects being largely exploration-oriented,
gnucash, bibliographic being utility-oriented and eXist, feedparserread-only service-oriented (as in providing stable services not to be
confused with SOA). The study also shows that projects have a
development speed cycle, along which the projects transform from
one type to another. This can be used to explain the fluctuations in
the language and file type shares over time as seen in this study.
Open-source software repositories have been used for studying
various aspects of software development like developer role
identification (core or associate) [9], framework hotspot detection
[10] and other. These works are complementary and help
developing a better understanding of sotware development process
and open-source software. It has also been shown that the number
and size of open-source projects are growing exponentially and
open-source projects are becoming more diverse by expanding into
new domains [11].
7. Conclusions and Future Work
We investigated the revision data of 22 OSS projects and tracked
the evolution of multiple programming language usage. Our
findings can be summarized from a language and a developer
perspective.
First, as for multiple programming language usage, our study
confirmed the ohloh.net data that the most popular (i.e. widely
used) language in OSS software projects is XML followed by Java
and C. XML has increased its popularity steadily over the last
decade while C has lost its high share to various other languages of
which Java has been among the more popular ones. Despite
becoming popular in just a few years, Java has not been able to
grow its share significantly during the last years. XSL has
maintained its share for the last years.
The most commonly co-evolving files are usually of the same type.
These are, ranked in order of co-evolution intensity: Java and XML;
C and plaintext files; and C and Makefiles. The most co-dependent
pair of file types in the business type projects studied was
JavaScript and XSL with a co-change rate (measured in common
commits) of more than 40% of the cases. Java and XML files
(especially those of project specific types) are more likely to be
edited by the same person than Java files and project definition
files.
Based on the projects analyzed, we found that XSL is important for
both generating user interfaces and for document transformations.
similar characteristics to the data available from ohloh.net (e.g.
more HTML code), which has currently the biggest analysed listing
of open source projects and as such is closest to representing the
population. It is not feasible to incorporate all projects listed by
ohloh.net as the data that would need to be analysed would exceed
our capabilities of processing it in timely fashion.
8. Acknowledgments
This research was conducted during a visit of the first author to the
software evolution and architecture lab at the University of Zurich.
We thank the members of the lab for their valuable advice. The
work is also partially funded by ERDF via the Estonian Centre of
Excellence in Computer Science.
9. References
[1] W. Meier, "eXist: An Open Source Native XML Database," in WEB,
WEB-SERVICES, AND DATABASE SYSTEMS, vol. 2593, ERFURT,
GERMANY, 2003, pp. 169-183, 2nd Annual International Workshop
on Web Databases, ERFURT, GERMANY, OCT 07-10, 2002.
[2] O. Alonso, P. T. Devanbu, and M. Gertz, "Expertise identification and
visualization from CVS," in MSR '08: Proceedings of the 2008
international working conference on Mining software repositories,
Leipzig, Germany, 2008, pp. 125-128.
[3] J. Brunnert, O. Alonso, and D. Riehle, "Enterprise People and Skill
Discovery Using Tolerant Retrieval and Visualization," in Advances in
Information Retrieval, 29th European Conference on IR Research,
ECIR 2007, vol. 4425, Rome, Italy, 2007, pp. 674-677.
[4] A. Capiluppi, P. Lago, and M. Morisio, "Characteristics of open source
projects," in Software Maintenance and Reengineering, 2003.
Proceedings. Seventh European Conference on, Benevento, Italy,
2003, pp. 317-327.
[5] T. Zimmermann, S. Kim, A. Zeller, and J. E. J. Whitehead, "Mining
version archives for co-changed lines," in MSR '06: Proceedings of the
2006 international workshop on Mining software repositories,
Shanghai, China, 2006, pp. 72-75.
Second, as for developers, we found that fewer file types are used
by new developers in their first commits, even though most
developers began with experience with multiple file types. Most
developers worked with at least five different file types during the
period studied. 80% of Java developers worked with XML files
while only 40% of C developers did so (60% in the later years).
[6] P. Weißgerber, L. von Klenze, M. Burch, and S. Diehl, "Exploring
evolutionary coupling in Eclipse," in eclipse '05: Proceedings of the
2005 OOPSLA workshop on Eclipse technology eXchange, San Diego,
California, 2005, pp. 31-34.
The study of languages used by developers from 1997 to 2009
showed the decreasing importance of Makefiles and plaintext files
for C developers while the importance of XML increased with
almost any other language. Whilst document type definition
language was being deprecated, XML Schema did not seem to
replace it (neither did any other language), implying that
standardised schemas are being preferred over project specific ones.
[8] K. Nakakoji, Y. Yamamoto, Y. Nishinaka, K. Kishida, and Y. Ye,
"Evolution patterns of open-source software systems and
communities," in IWPSE '02: Proceedings of the International
Workshop on Principles of Software Evolution, Orlando, Florida, 2002,
pp. 76-85.
From the characteristics of developer language usage, we saw that
not just knowing multiple languages is required from the
developers, but developers must also understand different coding
paradigms (e.g. procedural and object-oriented languages are often
used side-by-side with rule and template based extensible
languages). While in the 1990s they needed to know how to code in
C and write Makefiles, the increased variety of languages used in
newer projects and lack of distinct leaders in languages introduced
the need to be familiar with multi-language development.
Future work will address to better describe the population by
including newer/future projects. The ideal dataset would have more
[7] R. Dattero and S. D. Galup, "Programming languages and gender,"
Communications of the ACM, vol. 47, no. 1, pp. 99-102, January 2004.
[9] L. Yu and S. Ramaswamy, "Mining CVS Repositories to Understand
Open-Source Project Developer Roles," in Mining Software
Repositories, 2007. ICSE Workshops MSR '07. Fourth International
Workshop on, Minneapolis, MN, USA, 2007, p. 8.
[10] S. Thummalapenta and T. Xie, "SpotWeb: detecting framework
hotspots via mining open source repositories on the web," in
Proceedings of the 2008 international working conference on Mining
software repositories, Leipzig, Germany, 2008, pp. 109-112.
[11] A. Deshpande and D. Riehle, "The Total Growth of Open Source,"
Open Source Development, Communities And Quality In International
Federation For Information Processing, vol. 275, pp. 197-209, 2008.
| 6cs.PL
|
Real-time deep hair matting on mobile devices
arXiv:1712.07168v2 [cs.CV] 10 Jan 2018
Alex Levinshtein∗ , Cheng Chang∗ , Edmund Phung∗ , Irina Kezele∗ , Wenzhangzhi Guo∗ , Parham Aarabi∗†
∗ ModiFace Inc.
† University of Toronto
Abstract—Augmented reality is an emerging technology in
many application domains. Among them is the beauty industry,
where live virtual try-on of beauty products is of great
importance. In this paper, we address the problem of live
hair color augmentation. To achieve this goal, hair needs to be
segmented quickly and accurately. We show how a modified
MobileNet CNN architecture can be used to segment the hair in
real-time. Instead of training this network using large amounts
of accurate segmentation data, which is difficult to obtain, we
use crowd sourced hair segmentation data. While such data
is much simpler to obtain, the segmentations there are noisy
and coarse. Despite this, we show how our system can produce
accurate and fine-detailed hair mattes, while running at over
30 fps on an iPad Pro tablet.
Keywords-hair segmentation; matting;augmented reality;
deep learning; neural networks
I. I NTRODUCTION
Real-time image segmentation is an important problem in
computer vision with a multitude of applications. Among
them is the segmentation of hair for live color augmentation
in beauty applications (Fig. 1). This use case, however,
presents additional challenges. First, unlike many objects
with simple shape, hair has a very complex structure. For realistic color augmentation, a coarse hair segmentation mask
is insufficient. One needs a hair matte instead. Secondly,
many beauty applications run on mobile devices or in web
browsers, where powerful computing resources are not available. This makes it more challenging to achieve real-time
performance. This paper addresses both these challenges and
introduces a system that can accurately segment hair at over
30 fps on a mobile device.
In line with recent success of convolutional neural networks (CNNs) for semantic segmentation, our hair segmentation methods is based on CNNs. We make two main
contributions. First, most modern CNNs cannot run in realtime even on powerful GPUs and may occupy a large amount
of memory. Our target is real-time performance on a mobile
device. In our first contribution we show how to adapt
the recently proposed MobileNets [1] architecture for hair
segmentation, which is both fast and compact enough to be
used on a mobile device.
In the absence of detailed hair segmentation ground truth,
we train our network on noisy and coarse crowd-sourced
data. A coarse segmentation result, however, is insufficient
for hair color augmentation purposes. For realistic color
augmentation, an accurate hair matte is needed. In our
(a)
(b)
(c)
Figure 1. Automatic hair matting and coloring. (a) Input image. (b) Output
hair matte produced by our method. (c) Recolored hair.
second contribution, we propose a method for obtaining
accurate hair mattes in real-time without the need for accurate hair matte training data. First, we show how to modify
our baseline network architecture to have the capacity for
capturing fine-level details. Next, by adding a secondary
loss function that promotes perceptually appealing matting
results, we show that the network can be trained to yield
detailed hair mattes using only coarse hair segmentation
training data. We compare this approach to simple Guided
Filter [2] post-processing and show that it yields more
accurate and sharper results.
We evaluate our method, showing that it achieves stateof-the-art accuracy while running in real-time on a mobile
device. In the remainder of the paper, we discuss related
work in Sec. II, describe our approach in Sec. III, evaluate
our method Sec. IV, and conclude in Sec. V.
II. R ELATED WORK
Similar to work on general image segmentation, hair
segmentation work can be divided into two categories. The
first category of approaches uses hand-crafted features for
segmentation. Yacoob et al. [3] employ simple pixel-wise
color models to classify hair. Analogous method is employed
by Aarabi [4], while also making use of facial feature
locations and skin information. Khan et al. [5] use more
advanced features with random forests for classification.
Such approaches, however, prove to be insufficiently robust
for real-world applications.
For more spatially consistent segmentation results, a popular method is to formulate segmentation as random field
inference. Lee’s et al. [6] build a Markov Random Field
Figure 2.
Fully Convolutional MobileNet Architecture for Hair Segmentation
over image pixels, while Huang et al. [7] build their model
over superpixels instead. Wang has an alternative method
[8], [9] where overlapping image patches are first segmented
independently and then combined.
Recently, given the success of deep neural networks
(DNNs) in many areas, including semantic segmentation,
DNN-based hair segmentation methods have emerged. Guo
and Aarabi [10] use a heuristic method to mine highconfidence positive and negative hair patches from each
image, and train a separate DNN per image, which is then
used to classify the remaining pixels. Inspired by recent
success of fully convolutional networks (FCN) for semantic
segmentation [11], Chai et al. [12] and Qin et al. [13] employ
FCNs for hair segmentation. Due to the coarseness of raw
FCN segmentation results, similar to [14], both methods post
process the results using dense CRFs [15]. Additionally,
[13] have an extra matting step to obtain high-detail hair
mattes. Finally, Xu et al. [16] propose an end-to-end CNN
architecture for generic image matting yielding state-of-theart results.
Our approach follows this trend, while addressing several
issues. All the aforementioned methods [12], [13], [16] build
upon the VGG16 architecture [17]. A single forward pass
in VGG16 takes around 100ms even on a powerful GPU,
and much longer on a mobile device. Adding dense CRF
inference and matting futher increases the run-time. Moreover, VGG16 occupies approximately 500MB of memory,
which is too much for mobile applications. Instead, we show
how to adapt the recently proposed compact MobileNets
architecture [1] for segmentation and yield real-time matting
results without expensive post processing methods. Finally,
while it may be possible to obtain detailed hair matting data
using semi-automatic labeling techniques [16], we show how
to train our network without the need for such data.
III. A PPROACH
This section describes our contributions in detail. Firstly,
we describe our modifications to the original MobileNet [1]
architecture and the challenges of obtaining training data for
hair segmentation. Secondly, we illustrate our method for
real-time hair matting without the use of matting training
data.
A. Fully Convolutional MobileNet for Hair Segmentation
Inspired by [12], [13], we first tried to use a modified
VGG16 network [17] for hair segmentation, however a forward pass through the network took more than 2 seconds per
frame and the network occupied about 500MB of memory.
This was incompatible with our real-time mobile use case
and therefore we use MobileNets [1] instead, which are
faster and more compact.
We modified the original MobileNet architecture into a
fully convolutional network for segmentation, which we
name HairSegNet. First, we remove the last three layers:
Avg Pool, FC, and Softmax (refer to Table 1 in [1]). Next,
similar to [12], [14], to preserve fine details we increase the
output feature resolution by changing the step size of the
last two layers with step size of 2 to 1. Due to our use of
pre-trained weights on ImageNet, we dilate all the kernels
for the layers with updated resolution by their scale factor
w.r.t. their original resolution. Namely, kernels for layers that
increased by a factor of 2 are dilated by 2 and kernels for
images are still needed. We crowd-source such data using
a hair coloring app where users have to manually mark
their hair. While getting this data is cheap, the resulting
hair segmentation labels are very noisy and coarse. Fig. 3
illustrates this issue. Note that in the 2nd image hair was
labeled very sparsely, while in the 3rd image a photograph
of a pet was submitted. We manually clean this data by only
keeping the images of human faces with sufficiently good
hair masks. This is considerably faster than marking the hair
from scratch or fixing incorrect segmentations.
Figure 3. Crowd-sourced training data for hair segmentation. Top: images.
Bottom: masks. The data is noisy and coarse, with some images having poor
masks (2nd) and some non-face images (3rd).
layers that increased by a factor of 4 are dilated by 4. This
yields a final resolution of 28 × 28.
Next, we build a decoder that takes the above CNN
features as input and upsamples them to a hair mask at the
original 224 × 224 resolution. We tried upsampling using
transposed convolution layers, but saw gridding artifacts in
the resulting masks. Therefore, upsampling is performed by
a simplified version of an inverted MobileNet architecture.
At each stage, we upsample the previous layer by a factor
of 2 by replicating each pixel in a 2 × 2 neighborhood.
Then, we apply separable depthwise convolution, followed
by pointwise 1 × 1 convolutions with 64 filters, followed
by ReLu. The number of filters does not have a large effect
on accuracy, with 64 filters yielding a slightly better performance based on our experiments (see Sec. IV). The previous
block is repeated three times, yielding a 224 × 224 × 64
output. We conclude by adding a 1 × 1 convolution with
softmax activation and 2 output channels for hair / nonhair. The network is trained by minimizing the binary cross
entropy loss LM between predicted and ground truth masks.
The full architecture is illustrated in Fig. 2.
The resulting architecture is considerably more compact
than VGG16, occupying only 15MB. More importantly, a
forward pass takes 300ms when implemented in Tensorflow
on iPad Pro. Using the recently released optimized CoreML
library [18] from Apple, this time can be further reduced to
60ms per frame.
Training deep neural networks requires a large amount
of data. While there are large datasets for general semantic segmentation, these datasets are much less popular
for hair segmentation. Moreover, unlike some objects like
cars, which have a relatively simple shape, hair shape is
very complex. Therefore, obtaining precise ground truth
segmentation for hair is even more challenging.
To cope with this challenge we use a pre-trained network
on ImageNet and fine-tune the entire network on hair segmentation data. Nevertheless, several thousands of training
B. Hair matting
In our second contribution we show how to obtain accurate hair matting results. We solve the matting problem
using a CNN, which we name HairMatteNet, in an end-toend manner. Such an approach faces two challenges. First,
we need an architecture with the capacity to learn high
resolution matting details. The network in Sec. III-A may not
be suitable since the results are still generated by incremental
upsampling of relatively low-res layer (28 × 28). Secondly,
our CNN needs to learn hair matting using only coarse
segmentation training data.
We address the first issue by adding skip connections
between layers in the encoder and corresponding layers in
the decoder, similar to many modern network architectures
[11]. This way, shallower layers in the encoder, which
contain high-res but weak features are combined with lowres but powerful features from deeper layers. The layers
are combined by first applying a 1 × 1 convolution to
the incoming encoder layers to make the output depth
compatible with the incoming decoder layers (64 for the
three outer skip connections and 1024 for the inner skip
connection) and then merging the layers using addition. For
each resolution, the deepest encoder layer at that resolution
is taken for skip connection.
The second issue is addressed by adding a loss function
that promotes perceptually accurate matting output. Motivated by the alpha matting evaluation work of Rhemann
et al. [19], our secondary loss measures the consistency
between image and mask edges. It is minimized when the
two agree. Specifically, we define our mask-image gradient
consistency loss to be:
P
2
Mmag 1 − (Ix Mx + Iy My )
P
,
(1)
LC =
Mmag
where (Ix , Iy ) and (Mx , My ) are the normalized image
and mask gradients respectively, and Mmag is the mask
gradient magnitude. This loss is added to the original binary
cross entropy loss with a weight w, making the overall
loss L = LM + wLC . The combination of the two losses
maintains the balance between being true to training masks
while generating masks that adhere to image edges. Fig. 4
illustrates our new architecture and the combination of the
two loss functions.
Figure 4. Fully Convolutional MobileNet Architecture for Hair Matting. Skip connections are added to increase the network capacity for capturing high
resolution detail. Mask-image gradient consistency loss is added alongside with standard binary cross entropy loss to promote detailed matting results.
We compare HairMatteNet to simple post-processing of
our coarse segmentation mask (HairSegNet) with a Guided
Filter [2]. Qin et al. [13] used a similar approach but
employed a more advanced matting method that is not fast
enough for real-time applications on mobile devices. Guided
Filter is an edge-preserving filter and has a linear runtime complexity w.r.t. the image size. It takes only 5ms to
process a 224 × 224 image on iPad Pro. Fig. 5 compares
the masks with (c) and without (b) the filter. The former is
clearly capturing more details, with individual hair strands
becoming apparent. However, the filter adds detail only
locally near the edges of the mask from CNN. Moreover, the
edges of the refined masks have a visible halo around them,
which becomes even more apparent when the hair color has
lower contrast with its surroundings. This halo causes color
bleeding during hair recoloring. HairMatteNet yields sharper
edges (Fig. 5d) and captures longer hair strands, without the
unwanted halo effect seen in Guided Filter post-processing.
As an additional bonus, HairMatteNet runs twice as fast
compared to HairSegNet, taking only 30ms per frame on
a mobile device and without the need for an extra postprocessing matting step. Due to our use of skip connections,
that help with capturing high resolution detail, HairMatteNet
maintains the original MobileNet encoder structure with the
deepest layers having 7 × 7 resolution. These layers have
many depth channels (1024) and become very expensive to
(a)
(b)
(c)
(d)
Figure 5. Hair segmentation and matting. (a) Input image. (b) HairSegNet.
(c) HairSegNet + Guided Filter. (d) HairMatteNet.
process with increased resolution. Having a 7 × 7 resolution
makes processing much faster compared to the 28 × 28
resolution in HairSegNet.
IV. E XPERIMENTS
We evaluate our method on three datasets. First is our
crowd-sourced dataset, consisting of 9000 training, 380
validation, and 282 testing images. All three subsets include
the original images and their flipped versions. Since our
target is hair matting on mobile devices, we pre-process our
data by detecting the face and cropping a region around it
based on the scale expected for typical selfies.
To compare our method to existing approaches, we evaluate two public datasets: LFW Parts dataset [20] and the
hair dataset of Guo and Aarabi [10]. The former consists of
2927 250 × 250 images, with 1500 training, 500 validation,
and 927 test images. Pixels are labeled into three categories:
hair, skin, and background, generated at the superpixel level.
The latter consists of 115 high-resolution images. Since it
contains too few images to train on, we use our crowdsourced training data when evaluating on this set. To make
this dataset consistent with our training data, we pre-process
it in a similar manner (using face detection and cropping),
adding flipped images as well. Since in a few cases faces
were not detected, the resulting dataset consists of 212
images.
Training is done using a batch size of 4 using the Adadelta
[21] method in Keras [22], with learning rate 1.0, ρ = 0.95,
and = 1e − 7. We use L2 regularization with the weight
2 · 10−5 for convolution layers only. Depthwise convolution
layers and the last convolution layer are not regularized.
We set the loss balancing weight to w = 0.5. In the threeclass LFW data, only the hair class is contributing to the
mask-image gradient consistency loss. We train our model
for 50 epochs and select the best performing epoch using
validation data. Training on crowd-sourced dataset takes 5
hours on Nvidia GeForce GTX 1080 Ti GPU and less than
an hour on LFW Parts due to much smaller training set size.
A. Quantitative evaluation
For quantitative performance analysis, we measure the F1score, Performance [10], IoU, and Accuracy, averaged across
all test images. To measure the consistency of image and
hair mask edges, we also report the mask-image gradient
consistency loss (Eqn. 1). Recall that during the manual
clean-up in Sec. III-A we only filtered images rather than
correcting the masks. As a result, the quality of our hair
annotation is still poor. Therefore, prior to evaluation on our
crowd-sourced data, we manually corrected the test masks,
spending no more than 2 minutes per annotation. This
yielded slightly better ground truth. Three variants of our
method are evaluated on this relabeled data. Table I shows
our results. All three methods perform similarly w.r.t. the
ground truth comparison measures, however, HairMatteNet
is the clear winner in the gradient consistency loss category,
indicating that its masks adhere much better to image edges.
On the LFW Parts dataset, we report an on-par performance with the best performing method in Qin et al. [13],
but achieve it in real-time on a mobile device. We use only
the accuracy measure for evaluation since it is the only
measure used in [13]. Arguably, especially since LFW Parts
was annotated at the superpixel level, the ground truth there
may not good enough for high-accuracy analysis. On the
dataset of Guo and Aarabi [10] we report an F1-score of
0.9376 and Performance of 0.8253. We re-ran HNN, the best
performing method in [10], on this post-processed dataset
and obtained similar performance to that reported by the
authors, with F1-score of 0.7673 and Performance of 0.4674.
Model
F1
Perf.
IoU
Acc. Grad-cons.
Crowd-sourced dataset
HairSegNet
0.9212 0.7833 0.8564 0.9624 0.2696
HairSegNet + GF 0.9251 0.7964 0.8637 0.9646 0.1803
HairMatteNet
0.9219 0.7945 0.8589 0.9619 0.0533
LFW Parts dataset [20]
HairMatteNet
NA
NA
NA
0.965
NA
[13] FCN+CRF
NA
NA
N A 0.9676
NA
HairMatteNet SP
NA
NA
N A 0.9769
NA
[13] FCN+CRF SP N A
NA
N A 0.9732
NA
Guo and Aarabi dataset [10]
HairMatteNet
0.9376 0.8253 0.8848 0.964
0.0532
HNN [10]
0.7673 0.4674 0.6454 0.8793 0.2732
Table I
Q UANTITATIVE EVALUATION
Depth
16
32
64
128
F1
0.9202
0.9217
0.9229
0.9225
Perf.
0.7878
0.8001
0.7939
0.7937
IoU
0.8559
0.8584
0.8608
0.8605
Acc.
0.9585
0.9581
0.9604
0.9596
Table II
D ECODER LAYER DEPTH EXPERIMENTS ON VALIDATION DATA
B. Qualitative evaluation
We evaluate our method on publicly available selfie images for qualitative analysis. Results can be seen in Fig. 6.
HairSegNet (Fig. 6b) yields good but coarse masks. HairSegNet with Guided Filter (Fig. 6c) produces better masks but
with an undesirable blur around hair boundaries. The most
accurate and sharpest results are achieved by HairMatteNet
(Fig. 6d). A failure mode of both Guided Filter postprocessing and HairMatteNet is their under-segmentation of
hair-like objects in the vicinity of hair, such as eyebrows
in case of dark hair or bright background for light hair.
In addition, highlights inside the hair can cause the hair
mask from HairMatteNet to be non-homogeneous, which is
especially apparent in the last three examples in column (h).
C. Network architecture experiments
Decoder layer depth Using our validation data, we have
experimented with number of decoder layer channels, but
observed that it does not have a large effect on accuracy.
Table II illustrates our experiments with the number of
channels in the decoder, with 64 channels yielding the
best results according to most measures. These experiments
were done using the skip connections architecture in Fig. 4
without using the gradient consistency loss.
Input image size Howard et al. [1] observed that MobileNets perform better given higher image resolution. Given
our goal of accurate hair matting, we experimented with
increasing the resolution beyond 224 × 224, which is the
highest resolution MobileNet were trained on ImageNet. For
the 2nd and 3rd image in Fig. 6, Fig. 7 shows qualitative
comparison of masks inferred using HairMatteNet from
(a)
(b)
Figure 6.
(c)
(d)
(e)
(f)
(g)
(h)
Qualitative evaluation. (a,e) Input image. (b,f) HairSegNet. (c,g) HairSegNet + Guided Filter. (d,h) HairMatteNet.
224 × 224 images vs. 480 × 480 images. The 480 × 480
results look more accurate around the hair edges, with
longer hair strands being captured (e.g., the long hair strand
falling on the nose in the first image). However, the issues
mentioned in the previous section are emphasized as well,
with more of the hair mask bleeding into non-hair regions
and the inside of the mask becoming non-homogeneous due
to hair highlights. In addition, processing a larger image is
significantly more expensive.
V. S UMMARY
This paper presented a hair matting method with real-time
performance on mobile devices. We have shown how, given
noisy and coarse data, a modified MobileNet architecture is
trained to yield accurate matting results. While we apply the
proposed architecture for hair matting, it is general and can
be applied to other segmentation tasks. In future work we
will explore fully automatic methods for training from noisy
data without the need for manual filtering. In addition, we
will explore further improvements to matting quality, such as
capturing longer hair strands and segmenting light hair; all
this while keeping the hair mask homogeneous, preventing
it from bleeding into non-hair regions, and maintaining our
real-time performance on mobile devices.
R EFERENCES
[1] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang,
T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017.
[2] K. He, J. Sun, and X. Tang, “Guided image filtering,” TPAMI,
vol. 35, no. 6, pp. 1397–1409, 2013.
[3] Y. Yacoob and L. S. Davis, “Detection and analysis of hair,”
TPAMI, vol. 28, no. 7, pp. 1164–1169, 2006.
[4] P. Aarabi, “Automatic segmentation of hair in images,” in
ISM, 2015, pp. 69–72.
[5] K. Khan, M. Mauro, and R. Leonardi, “Multi-class semantic
segmentation of faces,” in ICIP, 2015, pp. 827–831.
[6] K.-c. Lee, D. Anguelov, B. Sumengen, and S. B. Gokturk,
“Markov random field models for hair and face segmentation,” in FG, 2008, pp. 1–6.
[14] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille, “Semantic image segmentation with deep convolutional nets and fully connected crfs,” in ICLR, 2015.
[15] P. Krähenbühl and V. Koltun, “Efficient inference in fully
connected crfs with gaussian edge potentials,” in NIPS, 2011,
pp. 109–117.
[16] N. Xu, B. Price, S. Cohen, and T. Huang, “Deep image
matting,” in CVPR, 2017.
[17] K. Simonyan and A. Zisserman, “Very deep convolutional
networks for large-scale image recognition,” arXiv preprint
arXiv:1409.1556, 2014.
[18] “Coreml,” https://developer.apple.com/documentation/coreml.
[19] C. Rhemann, C. Rother, J. Wang, M. Gelautz, P. Kohli,
and P. Rott, “A perceptually motivated online benchmark for
image matting,” in CVPR, 2009, pp. 1826–1833.
[20] A. Kae, K. Sohn, H. Lee, and E. Learned-Miller, “Augmenting CRFs with Boltzmann machine shape priors for image
labeling,” in CVPR, 2013.
[21] M. D. Zeiler, “Adadelta: an adaptive learning rate method,”
arXiv preprint arXiv:1212.5701, 2012.
[22] F. Chollet et al., “Keras,” https://github.com/fchollet/keras,
2015.
(a)
Figure 7.
(b)
Network resolution comparison (a) 224 × 224 (b) 480 × 480
[7] G. B. Huang, M. Narayana, and E. Learned-Miller, “Towards
unconstrained face recognition,” in CVPRW. IEEE, 2008,
pp. 1–8.
[8] N. Wang, H. Ai, and S. Lao, “A compositional exemplarbased model for hair segmentation,” in ACCV, 2010, pp. 171–
184.
[9] N. Wang, H. Ai, and F. Tang, “What are good parts for hair
shape modeling?” in CVPR, 2012, pp. 662–669.
[10] W. Guo and P. Aarabi, “Hair segmentation using heuristicallytrained neural networks,” TNNLS, 2016.
[11] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional
networks for semantic segmentation,” in CVPR, 2015, pp.
3431–3440.
[12] M. Chai, T. Shao, H. Wu, Y. Weng, and K. Zhou, “Autohair:
Fully automatic hair modeling from a single image,” TOG,
vol. 35, no. 4, p. 116, 2016.
[13] S. Qin, S. Kim, and R. Manduchi, “Automatic skin and hair
masking using fully convolutional networks,” in ICME, 2017.
| 1cs.CV
|
Efficient Diverse Ensemble for Discriminative Co-Tracking
Kourosh Meshgi, Shigeyuki Oba, Shin Ishii
Graduate School of Informatics, Kyoto University
606–8501 Yoshida-honmachi, Kyoto, Japan
arXiv:1711.06564v1 [cs.CV] 16 Nov 2017
{meshgi-k,oba,ishii}@sys.i.kyoto-u.ac.jp
Abstract
Ensemble discriminative tracking utilizes a committee of
classifiers, to label data samples, which are in turn, used for
retraining the tracker to localize the target using the collective knowledge of the committee. Committee members could
vary in their features, memory update schemes, or training
data, however, it is inevitable to have committee members
that excessively agree because of large overlaps in their
version space. To remove this redundancy and have an effective ensemble learning, it is critical for the committee to
include consistent hypotheses that differ from one-another,
covering the version space with minimum overlaps. In this
study, we propose an online ensemble tracker that directly
generates a diverse committee by generating an efficient set
of artificial training. The artificial data is sampled from the
empirical distribution of the samples taken from both target and background, whereas the process is governed by
query-by-committee to shrink the overlap between classifiers. The experimental results demonstrate that the proposed scheme outperforms conventional ensemble trackers
on public benchmarks.
(a) Typical ensemble state
(b) Conventional update
(c) Partial update
(d) Diversified update
Figure 1. Version space examples for ensemble classifiers. (a) All
hypotheses are consistent with the previous labeled data, but each
represents a different classifier in the version space Vt . In the next
time step, the models are updated with the new data (boxed). (b)
Updating with all of the data tend to make the hypothesis more
overlapping. (c) Random subsets of training data are given to
the hypotheses and they update without considering the rest of the
data, the hypotheses cover random areas of the version space. (d)
Random subsets of training data plus artificial generated data (proposed), trains the hypothese to be mutually uncorrelated as much
as possible, while encouraging them to cover more (unexplored)
area of the version space.
1. Introduction
Tracking-by-detection [2, 20, 23, 6, 21, 4] as one the
most popular approaches of discriminative tracking utilizes
classifier(s) to perform the classification task using object
detectors. In a tracking-by-detection pipeline, several samples are obtained from each frame of the video sequence,
to be classified and labeled by the target detector, and this
information is used to re-train the classifier in a closed feedback loop. This approach advantages from the overwhelming maturity of the object detection literature, both in the
terms of accuracy and speed [13, 15], yet struggles to keep
up with the target evolution as it rises issues such as proper
strategy, rate, and extent of the model update [51, 61, 36].
To adapt to object appearance changes, the tracking-bydetection methods update the decision boundary as opposed
to object appearance model in generative trackers. Imper-
fections of target detection and model update throughout
the tracking, manifest themselves as accumulating errors,
which essentially drifts the model from the real target distribution, hence leads to target loss and tracking failure.
Such imperfections can be caused by labeling noise, selflearning loop, sensitive online-learning schemes, improper
update frequency, non-realistic assumption about the target
distribution, and equal weights for all training samples.
Misclassification of a sample due to drastic target transformations, visual artifacts (such occlusion) or model errors not only degrades target localization accuracy, but also
confuses the classifier [23] when trained by this erroneous
label. Typically in tracking-by-detection, the classifier is
1
retrained using its own output from the earlier tracking
episodes (the self-learning loop), which amplitudes a training noise in the classifier and accumulate the error over
time. The problem amplifies when the tracker lacks a forgetting mechanism or is unable to obtain external scaffolds. Some researchers believe in the necessity of having a
“teacher” to train the classifier [21]. This inspired the use of
co-tracking [55], ensemble tracking [49, 62], disabling updates during occlusions, or label verification schemes [27]
to break the self-learning loop using auxiliary classifiers.
Ensemble tracking framework provides effective frameworks to tackle one or more of these challenges. In such
frameworks, the self-learning loop is broken, and the labeling process is performed by leveraging a group of classifiers
with different views [20, 22, 49], subsets of training data
[43] or memories [62, 42].
The main challenge in ensemble methods is how to
decorrelate ensemble members and diversify learned models [22]. Combining the outputs of multiple classifiers is
only useful if they disagree on some inputs [31], however,
individual learners with the same training data are usually
highly correlated [66] (See Figure 1).
Contributions: We propose a diversified ensemble discriminative tracker (DEDT) for real-time object tracking.
We construct an ensemble using various subsamples of the
tracking data and maintain the ensemble throughout the
tracking. This is possible by devising methods to update
the ensemble to reflect target changes while keeping its diversity to achieve good accuracy and generalization. In addition, breaking the self-learning loop to avoid the potential
drift of the ensemble is applied in a con-tracking framework
with an auxiliary classifier. However, to avoid unnecessary
computation and boost the accuracy of the tracker, an effective data exchange scheme is required. We demonstrate
that learning ensembles with randomized subsets of training data along with artificial data with diverse labels in a
co-tracking framework achieve superior accuracy. This paper offers the following contributions:
• We propose a novel ensemble update scheme that generates necessary samples to diversify the ensemble.
Unlike the other model update schemes that ignore
the correlation between classifiers of an ensemble, this
method is designed to promote diversity.
• We propose a co-tracking framework that accommodates the short and long-term memory mixture, effective collaboration between classification modules, and
optimized data exchange between modules by borrowing the concept of query-by-committee [54] from active learning literature.
Note that this is different from the co-tracking elaborated in
[55], as in that method two classifiers cast a weighted vote
to label the target, and pass the samples they struggle with
Figure 2. Schematic of the system. The proposed tracker, DEDT,
labels the obtained sample using an homogeneous ensemble of
the classifiers, the committee. The samples that the committee
has highest disagreement upon (the uncertain samples) are queried
from the auxiliary classifier, a different type of classifier. The location of the target is then estimated using the labeled target. Each
member of the ensemble is then updated with a random subset of
uncertain samples. By generating the diversity set (r.t Sec 4.2), the
ensemble is then diversified, yielding a more effective ensemble.
to the other one to learn. However, in our tracker, the ensemble passes the disputed samples to an auxiliary classifier
which is trained on all of the data periodically, to provide the
effect of long-term memory while being resistant to abrupt
changes, outliers and label noise. The evaluation results of
DEDT on OTB50 dataset [61] demonstrates competitive accuracy for our method compared to the state-of-the-art of
tracking.
2. Prior Work
Ensemble tracking: Using a linear combination of several weak classifiers with different associated weights has
been proposed in a seminal work by Avidan [3]. Following
this study, constructing an ensemble by boosting [20], online boosting [46, 35], multi-class boosting [48] and multiinstance boosting [4, 63] led to the enhancement of the performance of the ensemble trackers. Despite its popularity,
boosting demonstrates low endurance against label noise
[52] and alternative techniques such as Bayesian ensemble
weight adjustment [6] has been proposed to alleviate this
shortcoming. Recently, ensemble learning based on CNNs
gained popularity. Researchers make ensembles of CNNs
that shares convolutional layers [44], different loss functions for each output of the feature map [59], and repeatedly
subsampling different nodes and layers in fully connected
layers on CNN to build an ensemble [22, 38]. Furthermore,
it is proposed to exploit the power of ensembles such as feature adjustment in ensembles [18] and the addition of the
ensemble’s members [49, 62] over-time.
Ensemble diversity: Empirically, ensembles tend to yield
better results when there is a significant diversity among the
models [32]. Zhou [66] categorizes the diversity generation heuristics into (i) manipulation of data samples based
on sampling approaches such as bagging and boosting (e.g.
in [43]), (ii) manipulation of input features such as online
boosting [20], random subspaces[50], random ferns [47]
and random forests [49] or combining using different layers, neurons or interconnection layout of CNNs [22, 38],
(iii) manipulation of learning parameter, and (iv) manipulation of the error representation. The literature also suggests
a fifth category of manipulation of error function which encourages the diversity such as ensemble classifier selection
based on Fisher linear discriminant [58].
Training data selection: A principled ordering of training examples can reduce the cost of labeling and lead to
faster increases in the performance of the classifier [57],
therefore we strive to use training examples based on their
usefulness, and avoid using on all of them (including noisy
ones and outliers) that may result in higher accuracy [16].
Starting from easiest examples (Curriculum learning) [7],
pruning adversarial examples1 [39], excluding misclassified
samples from next rounds of training [56], sorting samples
by their training value [34] are some of the proposed approaches in the literature. However, the most common setting is active learning, in which the algorithm selects which
training examples to label at each step for the highest gains
in the performance. I this view, it may require to focus on
learning the hardest examples first. For example, following
the criteria of “highest uncertainty”, an active learner select
samples closest to the decision boundary to be labeled next.
This concept can be useful in visual tracking, e.g. to measure the uncertainty caused by bags of samples [65].
Active learning for ensembles: Query-by-committee
(QBC) [54] is one of the most popular ensemble-based active learning approaches, which constructs a committee of
models representing competing hypotheses to label the samples. By defining a utility function on the ensemble (such
as disagreement, entropy, or Q-statistics [66]), this method
selects the most informative samples to be queried from the
oracle (or any other collaborating classifier) in a form of
the query optimization process [53]. Built upon randomized component learning algorithm, QBC involves Gibbs
sampling, which is usually interactable in real-world situations. By extending QBC to use deterministic classifiers with different subsets of data to construct an ensemble, Abe and Mamitsuka[1] proposed practical query-bybagging and query-by-boosting. The set of hypotheses consistent with the data is called version space and by selecting
the most informative samples to be labeled, QBC attempts
to shrink the version space. However, only a committee
of hypotheses that effectively samples the version space of
all consistent hypotheses is productive for the sample se1 Images with tiny, imperceptible perturbations that fool a classifier into
predicting the wrong labels with high confidence.
lection [10]. To this end, it is crucial to promote the diversity of the ensemble [41]. In QBag and QBoost algorithms, all of the classifiers are trained on random subsets
of the similar dataset, which degrade the diversity of the ensemble. Reducing the number of necessary labeled samples
[33], unified sample learning and feature selection procedure [37] and reducing the sampling bias by controlling the
variance [9] are some of the improvements that active learning provides for the discriminative trackers. Moreover, using diversity data to diversify the committee members [41]
and promoting the classifiers that have unique misclassifications [58] are from few samples that active learning was
employed to promote the diversity of the ensemble.
3. Tracking by Detection
By definition, a tracker tries to determine the state of
the target pt in frame Ft (t ∈ {1, . . . , T }) by finding the
transformation yt from its previous state pt−1 . In trackingby-detection formulation, the tracker employs a classifier θt
to separate the target from the background. It is realized
by evaluating possible candidates from the expected target
state-space Yt . The candidate whose appearance resembles
the target the most, is usually considered as the new target
state. Finally, the classifier is updated to reflect the recent
information.
p
◦yj
To this end, first several samples xt t−1 t ∈ Xt are obtained by a transformation ytj ∈ Yt from the previous target
state, pt−1 ◦ ytj . Sample j ∈ {1, . . . , n} indicates the location pt−1 ◦ ytj in the frame Ft , where the image patch
p
◦yj
xt t−1 t is contained. Then, each sample is evaluated by
the classifier scoring function h : Xt → R to calculate the
p
◦yj
score sjt = h(xt t−1 t |θt ). This score is utilized to obtain
a label `jt for the sample, typically by thresholding its score,
j
+1 , st > τu
(1)
`jt = −1 , sjt < τl
0
, otherwise
where τl and τu serves as lower and upper thresholds respectively. Finally, the target location yt is obtained by
comparing the samples’ classification scores. To obtain the
exact target state, the sample with highest score is selected
∗
as the new target, yt = ytj s.t. j ∗ = argmax (sjt ). A
i
subset of the samples and their labels are used to re-train
the classifier’s model θt+1 = u(θt , Dξ(t) ). Here, Dt =
{hXt , Lt i} is the set of samples Xt and their labels Lt , u(.)
is the model update function, and the ξ(t) defines the subset
of the samples that the tracker considers for model update.
An ensemble discriminative tracker employs a set of
classifiers instead of one. These classifiers, hereafter called
(1)
(C)
committee, are represented by Ct = {θt , . . . , θt }, and
are typically homogeneous and independent (e.g., [49, 35]).
Popular ensemble trackers utilize the majority voting of the
committee as their utility function,
sjt =
C
X
p
sign h(xt t−1
◦ytj
(c)
|θt ) .
(2)
c=1
Then eq(1) is used to label the samples.
The model of each classifier is updated independently,
(c)
(c)
θt+1 = u(θt , Dξ(t) ) meaning that all of the committee
members are trained with a similar set of samples and a
common label for them.
4. Diverse Ensemble Discriminative Tracker
We propose a diverse ensemble tracker composed of
a highly-adaptive and diverse ensemble of classifiers C
(the committee), a long-term memory object detector (that
serves as the auxiliary classifier), and an information exchange channel governed by active learning. This allows for
effective diversification of the ensemble, improving the generalization of the tracker and accelerating its convergence
to the ever-changing distribution of target appearance. We
leveraged the complementary nature and long-term memory
of the auxiliary tracker to facilitate effective model update.
One way to diversify the ensemble is to increase the
number of examples they disagree upon [31]. Using bagging and boosting to construct an ensemble out of a fix
sample set, ignores this critical need for diversity as all of
the data are randomly sampled from a shared data distribution. However, for each committee member, there exists a
set of samples that distinguish them from other committee
members. One way to obtain such samples is to generate
some training samples artificially to differ maximally from
the current ensemble [40].
The diversified ensemble covers larger areas of the version space (i.e. the space of consistent hypotheses with the
samples from current frame), however, this radical update of
the ensemble may render the classifier susceptible to drastic
target appearance changes, abrupt motion, and occlusions.
In this case, given the non-stationary nature of the target
distribution2 , the classifier should adapt itself rapidly with
the target changes, yet it should keep a memory of the target
to re-identify if the target goes out-of-view or got occluded
(as known as stability-plasticity dilemma [21]). In addition,
there are samples for which the ensemble is not unanimous
and an external teacher maybe deemed required.
To amend these shortcomings, an auxiliary classifier is
utilized to label the samples which the ensemble dispute
upon (co-tracking). This classifier is batch-updated with all
of the samples less frequently than the ensemble, realizing
2 the non-stationarity means that the appearance of an object may
change so significantly that a negative sample in the current frame looks
more similar to a positive example in the previous frames.[5]
the longer memory for the tracker. Active query optimization is employed to query the label of the most informative
samples from the auxiliary classifier, which is observed to
effectively balance the stability-plasticity equilibrium of the
tracker as well. Figure 2 presents the schematic of the proposed tracker.
4.1. Formalization
In this approach, If the committee comes to a solid vote
about a sample, then the sample is labeled accordingly.
However, when the committee disagrees about a sample, its
(o)
label is queried from the auxiliary classifier θt :
+1
j
`t = −1
p
◦yj (o)
sign h(xt t−1 t |θt )
, sjt > τu
, sjt < τl
(3)
, otherwise
in which sjt is derived from eq(2). The uncertain samples
p
◦yj
list is defined as Ut = {xt t−1 t |τl < sjt ) < τu }.
The committee members are then updates using our proposed mechanism f (.) using the uncertain samples Ut ,
(c)
(1..c)
θt+1 = f (θt
, Ut , Dt )
(4)
Finally, to maintain a long-term memory and slower update
rate for the auxiliary classifier, it is updated every ∆ frames
with all of the samples from t − ∆ to t.
(
(o)
u(θt , Dt−∆..t ) , if t 6= k∆ + 1
(o)
θt+1 =
(5)
(o)
θt
, if t = k∆ + 1
Algorithm (1) summarizes the proposed tracker.
4.2. Diversifying Ensemble Update
The model updates to construct a diverse ensemble either replace the weakest or oldest classifier of the ensemble
[20, 3] or creates a new ensemble in each iteration [41].
While the former lacks flexibility to adjust to the rate of target change, the latter involves a high level of computation
redundancy. To alleviate these shortcomings, we create an
ensemble for the first frame, update them in each frame to
keep a memory of the target, and diversify them to improve
the effectiveness of ensemble. The diversifying update procedure is as follows:
1. The members ensemble Ct is updated with a random
subsets (of size m) of the uncertain data Ut , that make
them more adept in handling such samples, and generate a temporary ensemble Ct0 . Note that for certain
samples (those not in Ut ), the committee is unanimous
about the label and adding them to the training set of
the committee classifiers is redundant [43].
(c)
input : Committee models θt , Auxiliary model θ(o)
input : Target position in previous frame pt−1
output: Target position in current frame pt
for j ← 1 to n do
Sample a transformation ytj ∼ N (pt , Σsearch )
Calculate committee score sjt (eq(2))
if τl < sjt < τu then sample label is uncertain
p
◦yj
`jt = sign h(xt t−1 t |θ(o) )
j
Ut ← Ut ∪ {hxpt−1 ◦yt , `jt i}
else
`jt = sign(sjt )
j
D ← D ∪ {hxpt−1 ◦yt , `jt i}
for c ← 1 to C do
(c)
Uniformly resample m data St from Ut
0(c)
(c)
(c)
θt ← u(θt |St )
Calculate the prediction error of Ct , (Ct ) = |Ut |/|Dt |
Calculate empirical distribution of samples, Π(Xt )
for c ← 1 to C do
do
(c)
Draw m0 samples At from Π(Xt )
ˆ 0)
Calculate class membership probability `(C
t
−1
0
Set the labels of samples ∝ `ˆ (Ct )
0(c)
00(c)
(c)
θt ← u(θt |At )
Calculate new prediction error (Ct00 ) (eq(6))
while (Ct00 ) ≥ (Ct )
0(c)
00(c)
θt ← θt
All diversity sets are applied, Ct+1 ← Ct0
if mod(t, ∆) = 0 then
(o)
(o)
θt+1 ← u(θt , Dt−∆..t )
∗
Target transformation yt = ytj s.t.j ∗ = argmax(sjt )
i
Calculate target position pt = pt−1 ◦ ŷt
Algorithm 1: Diverse Ensemble Discriminative Tracker
2. The label prediction of the original ensemble Ct is then
calculated on Dt w.r.t. the labels given by the whole
tracker (composed of the ensemble and the auxiliary
classifier), and prediction error (Ct ) is obtained.
3. The empirical distribution of training data, Π(Xt ), is
calculated to govern the creation of the artificial data.
4. In an iterative process for each of the committee members, m0 samples are drawn from a Π(Xt ), assuming attribute independence. Given a sample, the class
membership probabilities of the temporary ensemble
ˆ 0 ) is then calculated. Labels are then sampled from
`(C
t
this distribution, such that the probability of selecting a
label is inversely proportional to the temporary ensemble prediction. This set of artificial samples and their
diverse labels are called the diversity set of committee
(c)
member c, At .
5. The classifier c of temporary ensemble is updated with
(c)
00(c)
At , to obtain the diverse ensemble Ct00 = {θt }
and calculate its prediction error (Ct00 ). If this update increases the total prediction error of the ensemble
((Ct00 ) > (Ct )), then the artificial data is rejected and
(c)
new data At should be generated,
(Ct00 ) =
C X
n
X
p
1 `jt 6= h(xt t−1
◦ytj
00(c)
|θt
) .
(6)
c=1 j=1
where 1(.) denotes the step function that returns 1 iff its
argument is true/positive and 0 otherwise.
This procedure creates samples for each member of the
committee that distinguish them from other members of the
ensemble using a contradictory label (therefore improving
the ensemble diversity [41]), but only accepts them when
using such artificial data improves the ensemble accuracy.
4.3. Implementation Details
There are several parameters in the system such as the
number of committee members (C), parameters of sampling
step (number of samples n, effective search radius Σsearch ),
and the holding time of auxiliary classifier (∆). Larger values of m results in temporary committee with a higher degree of overlap, thus less diverse, whereas smaller values
of m tend to miss the latest changes of the quick-changing
target. A Larger number of artificial samples m0 result in
more diversity in the ensemble, but reduce the chance of
successful update (i.e. lowering the prediction error of the
ensemble). These parameters were tuned using a simulated
annealing optimization on a cross-validation set.
In our implementation, we used lazy classifiers with
HOG [12] feature for the ensemble and reused the calculations with a caching mechanism to accelerate the classification method. For the empirical distribution of the data, a
Gaussian distribution is determinedly estimating the mean
and standard variation of the given training set (i.e. HOG
of Xt ). In addition, to localize the target, the samples with
positive ensemble scores are considered, and the one with
the highest sum of confidence scores is selected as the next
target position. The auxiliary classifier is a a part-based detector [17]. The features, part-base detector dictionary, and
the parameters of lazy classifier, thresholds τl , τu , and the
rest of above-mentioned parameters (Except for C that have
been adjusted to control the speed of the tracker, here) have
been adjusted using cross-validation. With C = 15, n =
1000, m = 80, m0 = 250, τu = 0.54, τl = −0.41 DEDT
achieved the speed of 21.97 fps on a Pentium IV PC @ 3.5
GHz and a Matlab/C++ implementation on a CPU.
Table 1. Quantitative evaluation of trackers under different visual
tracking challenges using AUC of success plot. The first, second
and third best methods are shown in color. More data available on
the author’s website.
Attribute
IV
DEF
OCC
SV
IPR
OPR
OV
LR
BC
FM
MB
TLD
0.48
0.38
0.46
0.49
0.50
0.48
0.54
0.36
0.39
0.45
0.41
0.53
0.51
0.50
0.51
0.54
0.53
0.52
0.33
0.52
0.52
0.47
0.62
0.62
0.61
0.58
0.58
0.62
0.68
0.43
0.67
0.65
0.63
0.73
0.69
0.69
0.71
0.69
0.70
0.73
0.50
0.72
0.65
0.65
0.68
0.70
0.69
0.68
0.69
0.67
0.62
0.47
0.67
0.56
0.61
0.70
0.67
0.70
0.71
0.70
0.69
0.66
0.58
0.70
0.63
0.69
0.75
0.69
0.76
0.76
0.72
0.74
0.79
0.70
0.70
0.72
0.72
0.75
0.69
0.72
0.74
0.73
0.74
0.76
0.58
0.73
0.74
0.72
ALL
0.49
0.55
0.62
0.72
0.69
0.70
0.75
0.74
STRK MEEM MUSTer STAPLE SRDCF CCOT
Ours
Table 2. Comparison of trackers based on their success rates
(IoU > 0.50). The first, second and third best methods are shown
in color.
TLD
STRK MEEM MUSTer STAPLE SRDCF CCOT
Ours
Success
45.1
58.2
68.2
66.7
73.9
76.0
81.6
81.8
Avg FPS
21.2
11.3
14.2
8.3
48.1
4.3
0.2
21.9
the regions respectively. We also compare all the trackers
by the success rate at the conventional thresholds of 0.50
(IoU > 0.50) [61].
5.1. Effect of Diversification
To demonstrate the effectiveness of the proposed diversification method we compare the DEDT tracker with two
different versions of the tracker. In the firs version, DEDTbag, the ensemble classifiers are only updated with uniformpicked subsets of the uncertain data (step 1 in section 4.2).
In the other version, DEDT-art, the committee members are
only updated with artificially generated data (steps 2-5 in
the same section). All three algorithms use m + m0 samples to update their classifiers. In addition to the overall
performance of the tracker, we measure the diversity of the
ensemble using the Q-statistics as elaborated in [32]. For
statistically independent classifiers i and j, the expectation
of Qi,k = 0. Classifiers that tend to classify the same
sample correctly will have positive values of Q, and those
which commit errors on different samples have negative Q
(−1 ≤ Qi,k ≤ +1). For the ensemble of C classifiers, the
averaged Q statistics over all pairs of classifiers is
5. Experiments
We perform a benchmark on the 50 videos of the OTB50
[61], along with partial subsets of the dataset with a distinguishing attribute to evaluate the tracker performance
under different situations. These attributes are illumination variation (IV), scale variation (SV), occlusions (OCC),
deformation (DEF), motion blur (MB), fast motion (FM),
in-plane-rotation (IPR), out-of-plane rotation (OPR), outof-view (OV), low resolution (LR), and background clutter
(BC), defined based on the biggest challenges that a tracker
may face throughout tracking. For this comparison, we have
used success and precision plots, where their area under
curve provides a robust metric for comparing tracker performances [61]. The result of the algorithms are reported as
the average of five independent runs.
The precision plot compares the number of frames that
a tracker has certain pixels of displacement, whereas the
overall performance of the tracker is measured by the area
under the surface of its success plot, where the success of
tracker in time t is determined when the normalized overlap
of the tracker target estimation pt with the ground truth p∗t
(also known as IoU) exceeds a threshold τov . Success plot,
graphs the success of the tracker against different values of
the threshold τov and its AU C is calculated as
1
AU C =
T
Z
0
T
1X
t=1
1
|pt ∩ p∗t |
> τov dτov ,
|p∗t ∪ p∗t |
(7)
where T is the length of sequence, |.| denotes the area of
the region and ∩ and ∪ stands for intersection and union of
Qav =
C−1
C
X X
2
Qi,j , s.t.
C(C − 1) i=1 j=i+1
(8)
Qi,j =
N f f N bb − N f b N bf
N f f N bb + N f b N bf
(9)
where N f b is the number of cases that classifier i classified
the sample as foreground, while classifier j detected it as
background, etc.
Figure 3. The effect of diversification procedure employed in the
proposed tracker.
Figure 3 illustrates the effectiveness of the diversification mechanism in contrast with merely generating data or update the classifiers with uninformed subsamples of the data.
From the experiment results,
AU C(DEDT-art) < AU C(DEDT-bag) < AU C(DEDT)
and 0 < Qav (DEDT) < Qav (DET-art) < Qav (DET-bag)
it can be concluded that all of steps of proposed diversification are crucial to maintain an accurate and diverse ensemble. Qav (DEDT-art) < Qav (DEDT-bag) shows that
the diversity of DEDT-art is better than random diversity obtain by DEDT-bag, however, AU C(DEDT-art) <
AU C(DEDT-bag) reveals that merely using artificial data
without the samples gathered by the tracker, does not provide enough data for an accurate model update.
5.2. Effect of using Artificial Data
In the first look, using synthesized data to train the ensemble that will keep track of a real object may not seem
proper. In this experiment, we look for the closest patch
of the real image (frame t of the video) to the synthesized
sample, and use it as the diversity data. To this end, in each
frame, a dense sampling over the frame is performed, the
HOG of these image patches are calculated, and the closest
match to the generated sample (using Euclidean distance)
is selected. The obtained tracker is referred as DEDT-real,
and its performance is compared to the original DEDT.
As Figure 4 shows, the use of this computationallyexpensive version of the algorithm does not improve the
performance significantly. However, it should be noted that
generating adversarial samples of the ensemble [19] for as
the diversity data of individual committee members is expected to increase the accuracy of the ensemble, yet it is out
of the scope of the current research and may be considered
as a future direction for this research.
Figure 5. The “activeness”, i.e. the effect of labeling thresholds on
the performance of the proposed algorithm.
semble or auxiliary classifier with the same chance. To interpret this figure it is prudent to note that δ → 0 forces the
ensemble to label all of the samples without any assistance
from the auxiliary classifier. By increasing δ the ensemble
starts to query highly disputed samples from the auxiliary
classifier, which is desired by design. If this value increases
excessively, the ensemble queries even slightly uncertain
samples from the auxiliary classifier, rendering the tracker
prone to the labeling noise of this classifier. In addition, the
tracker loses its ability to update rapidly in the case of an
abrupt change in the target’s appearance or location, leading to a degraded performance of the tracker. In the extreme
case of δ → 1 the tracker reduces to a single object detector
modeled by the auxiliary classifier.
The information exchange in one way is in the form
of querying the most informative labels from the auxiliary
classifier, and on the other way is re-training it with the labeled samples by the committee (for certain samples). We
observed that this exchange is essential to construct a robust and accurate tracker. Moreover, such data exchange
not only breaks the self-learning loop but also manages the
plasticity-stability equilibrium of the tracker. In this view,
lower values of δ correspond to a more-flexible tracker,
while higher values make it more conservative.
5.4. Comparison with State-of-the-Art
Figure 4. The effect of using artificial data versus real data employed in the proposed tracker.
5.3. Effect of “Activeness”
Labeling thresholds (τl and τu ) control the “activeness”
of the data exchange between the committee and the auxiliary classifier, therefore allowing the ensemble to get
more/less assistance for its collaborator. In our implementation, these two values are treated independently, but for
the sake of argument assume that τl = −δ and τu = +δ
(δ ∈ [0, 1]). Figure 5 compares the effects of different values of the δ, and also a “random” data exchange scheme in
which the labeler gets the label of the sample from the en-
To establish a fair comparison with the state-of-the-art,
some of the most popular discriminative trackers (according
to a recent large benchmark [61] and the recent literature)
are selected: TLD [27], STRK [23], MEEM [62], MUSTer
[26], STAPLE [8], SRDCF [14], and CCOT [15].
Figure 6 presents the success and precision plots of
DEDT along with other state-of-the-art trackers for all sequences. It is shown in this plot that DEDT usually keeps
the localization error under 10 pixels. Table 1 presents the
area under the curve of the success plot (eq(7)) for all the
sequences and their subcategories, each focusing on a certain challenge of the visual tracking. As shown, DEDT has
Figure 6. Quantitative performance comparison of the proposed
tracker, DEDT, with the state-of-the-art trackers using success plot
(top) and precision plot (bottom).
the competitive precision compared to CCOT which employs state-of-the-art multi-resolution deep feature maps,
and performs better than the rest of the other investigated
trackers on this dataset. The performance of DEDT is comparable with CCOT in the case of illumination variation,
deformation, out-of-view, out-of-plane rotation and motion
blur, while it has superior performance in handling background clutter. This indicates the effectiveness of the target vs. background detection and flexibility for accommodating rapid target changes. While the former can be attributed to effective ensemble tracking, the latter is known
to be the effect of combining long and short-term memory.
It is observed in the run-time that for handling extreme rotations, the ensemble heavily relies on the auxiliary tracker,
which although brings the superior performance in the category, a better representation of the ensemble model may reduce the reliance of the tracker to the auxiliary tracker.The
proposed algorithm shows a sub-optimal performance in
low-resolution scenario compared to DCF-based trackers
(SRDCF, and CCOT), and although it does not provide a
high-quality localization for smaller/low-resolution targets,
it is able to keep tracking them. This finding highlights the
importance of further research on the ensemble-based DCF
trackers. Finally, a qualitative comparison of DEDT versus
other trackers is presented in Figure 7.
6. Conclusion
Figure 7. Sample tracking results of evaluated algorithms on several challenging video sequences, in these sequences the red box
depicts the DEDT against other trackers (blue). The ground truth
is illustrated with yellow dashed box. From top to bottom the sequences are Skating1, FaceOcc2, Shaking, Basketball, and Soccer
with drastic illumination changes, scaling and out-of-plane rotations, background clutter, noise and severe occlusions.
In this study, we proposed diverse ensemble discriminative tracker (DEDT) that maintains a diverse committee
of classifiers to the label of the samples and queries the
most disputed labels –which are the most informative ones–
from a long-term memory auxiliary classifier. By generating artificial data with diverse labels, we intended to diversify the ensemble of classifiers, efficiently covering the
version space, increasing the generalization of the ensemble, and as a result, improve the accuracy. In addition, by
using the query-by-committee concept in labeling and updating stages of the tracker, the label noise problem is decreased because of updating classifiers only with the most
informative samples. By using the diverse committee, in
turn, the problem of equal weights for the samples are addressed, and a good approximation of the target location is
acquired even without dense sampling. The active learning
scheme also manages the balance between short-term and
long-term memory by recalling the label from long-term
memory when the short-term memory is not clear about
the label (due to forgetting the label or insufficient data).
This also reduces the dependence of the tracker on a single classifier (i.e., auxiliary classifier), yet breaking the selflearning loop to avoid accumulative model drift.
The result of the experiment on OTB50[61] benchmark
demonstrates the competitive tracking performance of the
proposed tracker compared with the state-of-the-art. In the
next step, we will investigate the strategies to detect or generate more challenging samples for the ensemble (e.g. adversarial samples of the ensemble) to accelerate the model
construction especially in rapidly changing scenarios.
References
[1] N. Abe and H. Mamitsuka. Query learning strategies using
boosting and bagging. In ICML’98, 1998. 3
[2] S. Avidan. Support vector tracking. PAMI, 26(8):1064–1072,
2004. 1
[3] S. Avidan. Ensemble tracking. PAMI, 29, 2007. 2, 4
[4] B. Babenko, M.-H. Yang, and S. Belongie. Visual tracking
with online multiple instance learning. In CVPR’09, 2009.
1, 2
[5] Q. Bai, Z. Wu, S. Sclaroff, M. Betke, and C. Monnier. Randomized ensemble tracking. In ICCV’13, 2013. 4
[6] Y. Bai and M. Tang. Robust tracking via weakly supervised
ranking svm. In CVPR’12, 2012. 1, 2
[7] Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In ICML’09, 2009. 3
[8] L. Bertinetto, J. Valmadre, S. Golodetz, O. Miksik, and P. H.
Torr. Staple: Complementary learners for real-time tracking.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 1401–1409, 2016. 7
[9] A. Beygelzimer, S. Dasgupta, and J. Langford. Importance
weighted active learning. In Proceedings of the 26th Annual
International Conference on Machine Learning, pages 49–
56. ACM, 2009. 3
[10] D. A. Cohn, Z. Ghahramani, and M. I. Jordan. Active learning with statistical models. Journal of artificial intelligence
research, 4(1):129–145, 1996. 3
[11] R. T. Collins, Y. Liu, and M. Leordeanu. Online selection of
discriminative tracking features. PAMI, 27(10):1631–1643,
2005.
[12] N. Dalal and B. Triggs. Histograms of oriented gradients for
human detection. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference
on, volume 1, pages 886–893. IEEE, 2005. 5
[13] M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg. Eco:
Efficient convolution operators for tracking. arXiv preprint
arXiv:1611.09224, 2016. 1
[14] M. Danelljan, G. Hager, F. Shahbaz Khan, and M. Felsberg.
Learning spatially regularized correlation filters for visual
tracking. In ICCV’15, pages 4310–4318, 2015. 7
[15] M. Danelljan, A. Robinson, F. S. Khan, and M. Felsberg.
Beyond correlation filters: Learning continuous convolution
operators for visual tracking. In ECCV’16. 1, 7
[16] F. De la Torre and M. J. Black. Robust principal component
analysis for computer vision. In ICCV’01, 2001. 3
[17] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained partbased models. PAMI, 32, 2010. 5
[18] J. Gall, A. Yao, N. Razavi, L. Van Gool, and V. Lempitsky. Hough forests for object detection, tracking, and action
recognition. PAMI, 2011. 2
[19] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint
arXiv:1412.6572, 2014. 7
[20] H. Grabner, M. Grabner, and H. Bischof. Real-time tracking
via on-line boosting. In BMVC’06, volume 1, page 6, 2006.
1, 2, 3, 4
[21] H. Grabner, C. Leistner, and H. Bischof. Semi-supervised
on-line boosting for robust tracking. In ECCV’08. 2008. 1,
2, 4
[22] B. Han, J. Sim, and H. Adam. Branchout: Regularization
for online ensemble tracking with convolutional neural networks. In Proceedings of IEEE International Conference on
Computer Vision, pages 2217–2224, 2017. 2, 3
[23] S. Hare, A. Saffari, and P. H. Torr. Struck: Structured output
tracking with kernels. In ICCV’11, 2011. 1, 7
[24] S. He, Q. Yang, R. Lau, J. Wang, and M.-H. Yang. Visual tracking via locality sensitive histograms. In CVPR’13,
pages 2427–2434, 2013.
[25] J. F. Henriques, R. Caseiro, P. Martins, and J. Batista. Exploiting the circulant structure of tracking-by-detection with
kernels. In ECCV’12, pages 702–715. Springer, 2012.
[26] Z. Hong, Z. Chen, C. Wang, X. Mei, D. Prokhorov, and
D. Tao. Multi-store tracker (muster): a cognitive psychology inspired approach to object tracking. In CVPR’15. 7
[27] Z. Kalal, K. Mikolajczyk, and J. Matas. Tracking-learningdetection. PAMI, 34(7):1409–1422, 2012. 2, 7
[28] H. Kiani Galoogahi, A. Fagg, and S. Lucey. Learning background-aware correlation filters for visual tracking.
arXiv, 2017.
[29] H. Kiani Galoogahi, T. Sim, and S. Lucey. Correlation filters
with limited boundaries. In CVPR’15, 2015.
[30] M. Kristan, J. Matas, A. Leonardis, and M. Felsberg. The
visual object tracking vot2015 challenge results. In ICCVw’15.
[31] A. Krogh, J. Vedelsby, et al. Neural network ensembles,
cross validation, and active learning. Advances in neural information processing systems, 7:231–238, 1995. 2, 4
[32] L. I. Kuncheva and C. J. Whitaker. Measures of diversity in
classifier ensembles and their relationship with the ensemble
accuracy. Machine learning, 51(2):181–207, 2003. 3, 6
[33] C. H. Lampert and J. Peters. Active structured learning for
high-speed object detection. In PR, pages 221–231. Springer,
2009. 3
[34] A. Lapedriza, H. Pirsiavash, Z. Bylinskii, and A. Torralba.
Are all training examples equally valuable? arXiv, 2013. 3
[35] C. Leistner, A. Saffari, and H. Bischof. Miforests: Multipleinstance learning with randomized trees. In ECCV’10, 2010.
2, 3
[36] A. Li, M. Lin, Y. Wu, M.-H. Yang, and S. Yan. Nus-pro: A
new visual tracking challenge. PAMI, 2016. 1
[37] C. Li, X. Wang, W. Dong, J. Yan, Q. Liu, and H. Zha. Active
sample learning and feature selection: A unified approach.
arXiv preprint arXiv:1503.01239, 2015. 3
[38] H. Li, Y. Li, and F. Porikli. Convolutional neural net bagging for online visual tracking. Computer Vision and Image
Understanding, 153:120–129, 2016. 2, 3
[39] J. Lu, T. Issaranon, and D. Forsyth. Safetynet: Detecting and
rejecting adversarial examples robustly. arXiv, 2017. 3
[40] P. Melville and R. J. Mooney. Constructing diverse classifier ensembles using artificial training examples. In IJCAI,
volume 3, pages 505–510, 2003. 4
[41] P. Melville and R. J. Mooney. Diverse ensembles for active learning. In Proceedings of the twenty-first international
conference on Machine learning, page 74. ACM, 2004. 3, 4,
5
[42] K. Meshgi, S. Oba, and S. Ishii. Active discriminative tracking using collective memory. In MVA’17. 2
[43] K. Meshgi, S. Oba, and S. Ishii. Robust discriminative tracking via query-by-committee. In AVSS’16, 2016. 2, 3, 4
[44] H. Nam, M. Baek, and B. Han. Modeling and propagating
cnns in a tree structure for visual tracking. arXiv preprint
arXiv:1608.07242, 2016. 2
[45] S. Oron, A. Bar-Hillel, D. Levi, and S. Avidan. Locally orderless tracking. IJCV, 2015.
[46] N. C. Oza. Online bagging and boosting. In SMC’05, 2005.
2
[47] C. Rao, C. Yao, X. Bai, W. Qiu, and W. Liu. Online random ferns for robust visual tracking. In Pattern Recognition
(ICPR), 2012 21st International Conference on, pages 1447–
1450. IEEE, 2012. 3
[48] A. Saffari, C. Leistner, M. Godec, and H. Bischof. Robust
multi-view boosting with priors. In ECCV’10. 2010. 2
[49] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof.
On-line random forests. In ICCVw’09. 2, 3
[50] A. Salaheldin, S. Maher, and M. Helw. Robust real-time
tracking with diverse ensembles and random projections. In
Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 112–120, 2013. 3
[51] S. Salti, A. Cavallaro, and L. Di Stefano. Adaptive appearance modeling for video tracking: Survey and evaluation.
IEEE TIP, 2012. 1
[52] J. Santner, C. Leistner, A. Saffari, T. Pock, and H. Bischof.
Prost: Parallel robust online simple tracking. In CVPR’10. 2
[53] B. Settles. Active learning. Morgan & Claypool Publishers,
2012. 3
[54] H. S. Seung, M. Opper, and H. Sompolinsky. Query by committee. In COLT’92, pages 287–294. ACM, 1992. 2, 3
[55] F. Tang, S. Brennan, Q. Zhao, and H. Tao. Co-tracking using
semi-supervised support vector machines. In ICCV’07. 2
[56] A. Vezhnevets and O. Barinova. Avoiding boosting overfitting by removing confusing samples. In ECML’07. 3
[57] S. Vijayanarasimhan and K. Grauman. Cost-sensitive active
visual category learning. IJCV, 2011. 3
[58] I. Visentini, J. Kittler, and G. L. Foresti. Diversity-based
classifier selection for adaptive object tracking. In MCS,
pages 438–447. Springer, 2009. 3
[59] L. Wang, W. Ouyang, X. Wang, and H. Lu. Stct: Sequentially training convolutional networks for visual tracking. In
Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 1373–1381, 2016. 2
[60] X. Wang, G. Hua, and T. Han. Discriminative tracking by
metric learning. ECCV’10, pages 200–214, 2010.
[61] Y. Wu, J. Lim, and M.-H. Yang. Online object tracking: A
benchmark. In CVPR’13, pages 2411–2418. IEEE, 2013. 1,
2, 6, 7, 8
[62] J. Zhang, S. Ma, and S. Sclaroff. Meem: Robust tracking via
multiple experts using entropy minimization. In ECCV’14.
2, 7
[63] K. Zhang and H. Song. Real-time visual tracking via online
weighted multiple instance learning. PR, 2013. 2
[64] K. Zhang, L. Zhang, and M.-H. Yang. Real-time compressive tracking. In ECCV’12, pages 864–877. Springer, 2012.
[65] K. Zhang, L. Zhang, M.-H. Yang, and Q. Hu. Robust
object tracking via active feature selection. IEEE CSVT,
23(11):1957–1967, 2013. 3
[66] Z.-H. Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012. 2, 3
| 1cs.CV
|
arXiv:1710.04369v3 [math.ST] 28 Feb 2018
The Inverse Gamma-Gamma Prior for Optimal
Posterior Contraction and Multiple Hypothesis
Testing ∗
Ray Bai
Malay Ghosh
†
University of Florida
March 2, 2018
Abstract
We study the well-known problem of estimating a sparse n-dimensional
unknown mean vector θ = (θ1 , ..., θn ) with entries corrupted by Gaussian white noise. In the Bayesian framework, continuous shrinkage
priors which can be expressed as scale-mixture normal densities are
popular for obtaining sparse estimates of θ. In this article, we introduce a new fully Bayesian scale-mixture prior known as the inverse
gamma-gamma (IGG) prior. We prove that the posterior distribution
contracts around the true θ at (near) minimax rate under very mild
conditions. In the process, we prove that the sufficient conditions for
minimax posterior contraction given by van der Pas et al. [25] are not
necessary for optimal posterior contraction. We further show that the
IGG posterior density concentrates at a rate faster than those of the
horseshoe or the horseshoe+ in the Kullback-Leibler (K-L) sense. To
classify true signals (θi 6= 0), we also propose a hypothesis test based
on thresholding the posterior mean. Taking the loss function to be the
expected number of misclassified tests, we show that our test procedure asymptotically attains the optimal Bayes risk exactly. We illustrate through simulations and data analysis that the IGG has excellent
finite sample performance for both estimation and classification.
∗
Keywords and phrases: normal means problem, sparsity, nearly black vectors, posterior contraction, multiple hypothesis testing, heavy tail, shrinkage estimation
†
Malay Ghosh (email: [email protected]) is Distinguished Professor, Department of
Statistics, University of Florida. Ray Bai (email: [email protected]) is Graduate Student,
Department of Statistics, University of Florida.
1
1
1.1
Introduction
The Normal Means Problem Revisited
Suppose we observe an n-component random observation (X1 , ..., Xn ) ∈ Rn ,
such that
Xi = θi + i , i = 1, ..., n,
(1)
where i ∼ N (0, 1), i = 1, ..., n. In the high-dimensional setting where n is
very large, sparsity is a very common phenomenon. That is, in the unknown
mean vector θ = (θ1 , ..., θn ), only a few of the θi ’s are nonzero. Under model
(1), we are primarily interested in separating the signals (θi 6= 0) from the
noise (θi = 0) and giving robust estimates of the signals.
This simple framework (1) is the basis for a number of high-dimensional
problems, such as image reconstruction, genetics, and wavelet analysis (Johnstone and Silverman [20]). For example, if we wish to reconstruct an image
from millions of pixels of data, only a few pixels are typically needed to
recover the objects of interest. In genetics, we may have tens of thousands
of gene expression data points, but only a few are significantly associated
with the phenotype of interest. For instance, Wellcome Trust [28] has confirmed that only seven genes have a non-negligible association with Type I
diabetes. These applications demonstrate that sparsity is a fairly reasonable
assumption for θ in (1).
1.2
Scale-Mixture Shrinkage Priors
Scale-mixture shrinkage priors are widely used for obtaining sparse estimates
of θ in (1). These priors typically take the form
θi |σi2 ∼ N (0, σi2 ), σi2 ∼ π(σi2 ), i = 1, ..., n,
(2)
where π : [0, ∞) → [0, ∞) is a density on the positive reals. These scalemixture densities typically contain heavy mass around zero, so that the
posterior density is heavily concentrated around 0 ∈ Rn . However, they
also retain heavy enough tails in order to correctly identify and prevent
overshrinkage of the true signals.
Global-local (GL) shrinkage priors comprise a wide class of scale-mixture
shrinkage priors (2). GL priors take the form
θi |τ, λi ∼ N (0, λi τ ), λi ∼ f, τ ∼ g,
(3)
where τ is a global shrinkage parameter that shrinks all θi ’s to the origin,
while the local scale parameters λi ’s control the degree of individual shrinkage. Examples of GL priors include the Bayesian lasso (Park and Casella
2
[21]), the horseshoe prior (Carvalho et al. [9]), the Strawderman-Berger prior
(Strawderman [24], Berger [4]), the normal-exponential-gamma (NEG) prior
(Griffin and Brown [19]), the Dirichlet-Laplace prior (Bhattacharya et al.
[6]), the generalized double Pareto (GDP) family (Armagan et al. [2]), and
the horseshoe+ prior (Bhadra et al. [5]).
The three parameter beta normal (TPBN) mixture family introduced by
Armagan et al. [1] generalizes several well-known scale-mixture shrinkage
priors. The TPBN family places a beta prime density (also known as the
inverted beta) as the prior on λi in (3), i.e.
π(λi ) =
Γ(a + b) a−1
λ (1 + λi )−(a+b) , i = 1, ..., n,
Γ(a)Γ(b) i
(4)
where a and b are positive constants. Examples of priors that fall under the
TPBN family include the horseshoe prior (a = b = 0.5), the StrawdermanBerger prior (a = 1, b = 0.5), and the normal-exponential gamma (NEG)
prior (a = 1, b > 0).
GL priors have been studied extensively in the context of sparse normal
means estimation. Many authors have shown that the posterior distribution
under GL priors contracts at (near) minimax rate. Most of the past posterior
contraction results have relied on tuning or estimating the global parameter
τ to achieve this rate. The τ in (3) can either be a priori specified with a
specific rate of decay (as in van der Pas et al. [27] or Ghosh and Chakrabarti
[16]) or it can estimated from the data through empirical Bayes or by placing
a prior on τ (as in van der Pas et al. [27], van der Pas et al. [26], and
Bhattacharya et al. [6]).
Moving beyond the global-local framework, van der Pas et al. [25] provided conditions for which the posterior distribution under any scale-mixture
shrinkage prior of the form (2) achieves the minimax posterior contraction
rate, provided that the θi ’s are a posteriori independent. Their result is
quite general and covers a wide variety of priors, including the normalgamma prior (Griffin and Brown [18]), the spike-and-slab LASSO (Ročková
[22]) and the horseshoe+ prior (Bhadra et al. [5]). A thorough discussion of
optimal posterior contraction is given in Section 3.
In addition to robust estimation of θ, we are often interested in identifying the true signals (or non-zero entries) within θ. Here, we are essentially conducting n simultaneous hypothesis tests, H0i : θi = 0 vs. H1i :
θi 6= 0, i = 1, ..., n. Assuming that the true data-generating model is a
two-components mixture density, Bogdan et al. [7] studied the risk properties of a large number of multiple testing rules. Specifically, Bogdan et al.
3
[7] considered a symmetric 0-1 loss function taken to be the expected total number of misclassified tests. By imposing a few regularity conditions
to induce sparsity and to bound the Type I and Type II error probabilities away from zero and one, Bogdan et al. [7] arrived at a simple closed
form for the asymptotic Bayes risk under 0-1 loss. They termed this as the
asymptotically Bayes optimal risk under sparsity (or ABOS risk). They then
provided necessary and sufficient conditions for which a number of classical
multiple test procedures (e.g. the Benjamini and Hochberg [3] procedure)
could asymptotically match the ABOS risk, provided that the true θi ’s are
generated from a point mass-mixture density. A thorough discussion of this
decision theoretic framework is presented in Section 4.1.
Testing rules induced by scale-mixture shrinkage priors – specifically GL
priors (3) – have also been studied in this decision theoretic framework. Assuming that the θi ’s come from a two-components model, Datta and Ghosh
[12] showed that a thresholding rule based on the posterior mean under the
horseshoe prior could asymptotically attain the ABOS risk up to a multiplicative constant. Ghosh et al. [17] generalized this result to a general class
of shrinkage priors of the form (3), including the student-t distribution, the
TPBN family, and the GDP family of priors. Ghosh and Chakrabarti [16]
later showed that their thresholding rule for this same class of priors could
even asymptotically attain the ABOS risk exactly. Bhadra et al. [5] also
extended this same rule for the horseshoe+ prior, showing that a testing
rule based on the horseshoe+ prior could asymptotically attain the ABOS
risk up to a multiplicative constant. In all of the aforementioned papers,
the global parameter τ in (3) was treated as either a tuning parameter that
decays to zero as n → ∞ or set as a plug-in empirical Bayes estimate τb from
van der Pas et al. [27].
In this article, we introduce a new fully Bayesian scale-mixture shrinkage
prior. Our goal is twofold. Having observed a vector X = (X1 , ..., Xn ) with
entries from (1), we would like to achieve: 1) robust estimation of θ, and
2) a robust testing rule for identifying true signals. To tackle both these
problems, we introduce a new scale-mixture shrinkage prior known as the
Inverse Gamma-Gamma (IGG) prior.
The IGG prior has a number of attractive theoretical properties. Under
extremely mild conditions and with specification of appropriate hyperparameters, the IGG posterior density is able to attain the (near) minimax
contraction rate. Our work differs from the existing literature in several
notable ways. First, the IGG is a special case of the TPBN prior density
(4). However, we show that we can achieve (near) minimax posterior contraction by simply specifying sample-size dependent hyperparameters a and
4
b, rather than by tuning or estimating a shared global parameter τ . Our
prior therefore does not fall under the global-local framework and our theoretical results differ from many existing results based on global-local priors.
Moreover, our prior is an example of a scale-mixture shrinkage prior that
does not necessarily satisfy the conditions for optimal contraction given by
van der Pas et al. [25], thus proving that these conditions are not necessary
for minimax-optimal posterior contraction. Finally, we justify the use of the
IGG by showing that the posterior concentrates at a rate faster than any
known Bayes estimator (including the horseshoe and horseshoe+ densities)
in the Kullback-Leibler sense.
In addition, we show that our testing rule for classifying signals asymptotically achieves the optimal Bayes risk exactly. While previously, Ghosh
and Chakrabarti [16] demonstrated that testing rules based on global-local
priors could asymptotically attain the optimal Bayes risk exactly, their result required tuning or estimating a global parameter τ . The IGG prior
avoids this by placing appropriate values (dependent upon sample size) as
its hyperparameters instead.
The organization of this paper is as follows. In Section 2, we introduce the IGG prior. We show that it mimics traditional shrinkage priors
by placing heavy mass around zero. We also establish various concentration
properties of the IGG prior that characterize its tail behavior and that are
crucial for establishing our theoretical results. In Section 3, we discuss the
behavior of the posterior under the IGG prior. We show that for a class
of sparse normal mean vectors, the posterior distribution under the IGG
prior contracts around the true θ at (near) minimax rate under mild conditions, and moreover, that the posterior concentrates at a faster rate than
any known Bayes estimator. In Section 4, we introduce our thresholding rule
based on the posterior mean and demonstrate that it asymptotically attains
the ABOS risk exactly. In Section 5, we present simulation results which
demonstrate that the IGG prior has excellent performance for both estimation and classification in finite samples. Finally, in Section 6, we utilize the
IGG prior to analyze a prostate cancer data set.
1.3
Notation
We use the following notations for the rest of the paper. Let {an } and {bn }
be two non-negative sequences of real numbers indexed by n, where bn 6= 0
an
for sufficiently large n. We write an bn to denote 0 < lim inf
≤
n→∞ bn
an
lim sup
< ∞, and an - bn to denote that there exists a constant C > 0
n→∞ bn
5
independent of n such that an ≤ Cbn provided n is sufficiently large. If
limn→∞
an
bn
= 1, we write it as an ∼ bn . Moreover, if
an
bn
≤ M for all
sufficiently large n where M > 0 is a positive constant independent of n,
then we write an = O(bn ). If limn→∞ abnn = 0, we write an = o(bn ). Thus,
an = o(1) if limn→∞ an = 0.
Throughout the paper, we also use Z to denote a standard normal
N (0, 1) random variable having cumulative distribution function and probability density function Φ(·) and φ(·), respectively.
2
The Inverse Gamma-Gamma (IGG) Prior
Suppose we have observed X ∼ N (θ, In ), and our task is to estimate the
n-dimensional vector, θ. Consider putting a scale-mixture prior on each
θi , i = 1, ..., n of the form
θi |σi2
σi2
ind
∼
i.i.d.
∼
N (0, σi2 ), i = 1, ..., n,
β 0 (a, b), i = 1, ..., n,
(5)
where β 0 (a, b) denotes the beta prime density (4). The scale mixture prior
(5) is a special case of the TPBN family of priors with the global parameter
τ fixed at τ = 1. From (5), one easily sees that the posterior mean of θi
under (5) is given by
E{E(θi |Xi , σ 2 )} = {E(1 − κi )|Xi } Xi ,
(6)
1
where κi = 1+σ
2 . Using a simple transformation of variables, we also see
i
that the posterior density of the shrinkage factor κi is proportional to
κi Xi2
π(κi |Xi ) exp −
2
!
a−1/2
κi
(1 − κi )b−1 , κi ∈ (0, 1).
(7)
From (6), it is clear that the amount of shrinkage is controlled by the
shrinkage factor κi . With appropriately chosen a and b, one can obtain
sparse estimates of the θi ’s. For example, with a = b = 0.5, we obtain the
standard half-Cauchy density C + (0, 1).
To distinguish our work from previous results, we note that that the
beta prime density (4) can be rewritten as a product of independent inverse
gamma and gamma densities. We reparametrize (5) as follows:
6
Figure 1: Marginal density of the IGG prior (8) with hyperparameters a =
0.6, b = 0.4, in comparison to other shrinkage priors. The DL1/2 prior is
the marginal density for the Dirichlet-Laplace density with Dir(1/2, ...., 1/2)
specified as a prior in the Bayesian hierarchy.
ind
θi |λi , ξi ∼ N (0, λi ξi ), i = 1, ..., n,
i.i.d.
λi ∼ G(a, 1), i = 1, ..., n,
(8)
i.i.d.
ξi ∼ IG(b, 1), i = 1, ..., n,
where a, b > 0. It should be noted that the rate parameter 1 in (8) could
be replaced by any positive constant. Representation (8) gives us some
important intuition into the behavior of the IGG prior. Namely, for small
values of b, IG(b, 1) places more mass around zero. As Proposition 1 shows,
for any 0 < b ≤ 12 , the marginal distribution for a single θ under the IGG
prior has a singularity at zero.
Proposition 1. If θ is endowed with the IGG prior (8), then the marginal
distribution of θ is unbounded with a singularity at zero for any 0 < b ≤ 1/2.
Proof. See Appendix A.
7
Proposition 1 gives us some insight into how we should choose the hyperparameters in (8). Namely, we see that for small values of b, the IGG prior can
induce sparse estimates of the θi ’s by shrinking most observations to zero.
As we will illustrate in Section 2.1, the tails of the IGG prior are still heavy
enough to identify signals that are significantly far away from zero.
Figure 1 gives a plot of the marginal density π(θi ) for the IGG prior (8),
with a = 0.6 and b = 0.4. Figure 1 shows that with a small value for b, the
IGG has a singularity at zero. The IGG prior also appears to have slightly
heavier mass around zero than other well-known scale-mixture shrinkage
priors, but maintaining the same tail robustness. In Section 3, we provide a
theoretical argument that shows that the shrinkage profile near zero under
the IGG is indeed more aggressive than that of previous known Bayesian
estimators.
2.1
Concentration Properties of the IGG Prior
Consider the IGG prior given in (8), but now we allow the hyperparameter
bn is allowed to vary with n as n → ∞. Namely, we allow 0 < bn < 1 for all
n, but bn → 0 as n → ∞ so that even more mass is placed around zero as
n → ∞. We also fix a to lie in the interval ( 12 , ∞). To emphasize that the
hyperparameter bn depends on n, we rewrite the prior (8) as
ind
θi |λi , ξi ∼ N (0, λi ξi ), i = 1, ..., n,
i.i.d.
(9)
λi ∼ IG(a, 1), i = 1, ..., n,
i.i.d.
ξi ∼ G(bn , 1), i = 1, ..., n,
where bn ∈ (0, 1) = o(1) and a ∈ ( 12 , ∞). For the rest of the paper, we label
this particular variant of the IGG prior as the IGGn prior.
As described in Section 2, the shrinkage factor κi = 1+λ1 i ξi plays a critical
role in the amount of shrinkage of each observation Xi . In this section, we
further characterize the tail properties of the posterior distribution π(κi |Xi ),
which demonstrates that the IGGn prior (9) shrinks most estimates of θi ’s to
zero but still has heavy enough tails to identify true signals. In the following
results, we assume the IGGn prior on θi for Xi ∼ N (θi , 1).
Theorem 1. For any a, bn ∈ (0, ∞),
2
E(1 − κi |Xi ) ≤ eXi /2
Proof. See Appendix A.
8
bn
.
a + bn + 1/2
Corollary 1.1. If a is fixed and bn → 0 as n → ∞, then E(1 − κi |Xi ) → 0
as n → ∞.
Theorem 2. Fix ∈ (0, 1). For any a ∈ ( 21 , ∞), bn ∈ (0, 1),
2
P (κi < |Xi ) ≤ eXi /2
bn
.
(a + 1/2) (1 − )
Proof. See Appendix A.
Corollary 2.1. If a ∈ ( 12 , ∞) is fixed and bn → 0 as n → ∞, then by
Theorem 2, P (κi ≥ |Xi ) → 1 for any fixed ∈ (0, 1).
Theorem 3. Fix η ∈ (0, 1), δ ∈ (0, 1). Then for any a ∈ ( 12 , ∞) and
bn ∈ (0, 1),
P (κi > η|Xi ) ≤
a+
1
2
(1 − η)bn
1
bn (ηδ)a+ 2
η(1 − δ) 2
exp −
Xi .
2
Proof. See Appendix A.
Corollary 3.1. For any fixed n where a ∈ ( 21 , ∞), bn ∈ (0, 1), and for every
fixed η ∈ (0, 1), P (κi ≤ η|Xi ) → 1 as Xi → ∞.
Corollary 3.2. For any fixed n where a ∈ ( 12 , ∞), bn ∈ (0, 1), and for every
fixed η ∈ (0, 1), E(1 − κi |Xi ) → 1 as Xi → ∞.
Since E(θi |Xi ) = {E(1 − κi )|Xi } Xi , Corollaries 1.1 and 2.1 illustrate that
all observations will be shrunk towards the origin under the IGGn prior (9).
However, Corollaries 3.1 and 3.2 demonstrate that if Xi is big enough, then
the posterior mean {E(1 − κi )|Xi } Xi ≈ Xi . This assures us that the tails
of the IGG prior are still sufficiently heavy to detect true signals.
We will use the concentration properties established in Theorem 1 and
3 to provide sufficient conditions for which the posterior mean and posterior distribution under the IGGn prior (9) contract around the true θ0 at
minimax or near-minimax rate in Section 3. These concentration properties
will also help us to construct the multiple testing procedure based on κi in
Section 4.
9
3
3.1
Posterior Behavior Under the IGG Prior
Sparse Normal Vectors in the Nearly Black Sense
Suppose that we observe X = (X1 , ..., Xn ) ∈ Rn from (1). Let `0 [qn ] denote
the subset of Rn given by
`0 [qn ] = {θ ∈ Rn : #(1 ≤ j ≤ n : θj 6= 0) ≤ qn }.
(10)
If θ ∈ `0 [qn ] with qn = o(n) as n → ∞, we say that θ is sparse in the
“nearly black sense.” Let θ0 = (θ01 , ..., θ0n ) be the true mean vector. In
their seminal work, Donoho et al. [13] showed that for any estimator of θ,
b the corresponding minimax risk with respect to the l2 - norm
denoted by θ,
is given by
inf
sup
b θ0 ∈`0 [qn ]
θ
Eθ0 ||θb − θ0 ||2 = 2qn log
n
(1 + o(1)), as n → ∞.
qn
(11)
In (11) and throughout the paper, Eθ0 denotes expectation with respect to
the N (θ0 , In ) distribution. (11) effectively states that in the presence of
sparsity, a minimax-optimal estimator only loses as a logarithmic factor (in
the ambient dimension) as a penalty for not knowing the true locations of
the zeroes. Moreover, (11) implies that we only need a number of replicates
in the order of the true sparsity level qn to consistently estimate θ0 . In order
for the performance of Bayesian estimators to be compared with frequentist
ones, we say that a Bayesian point estimator θbB attains the minimax risk
(in the order of a constant) if
n
sup Eθ0 ||θ̂ − θ0 || qn log
.
qn
θ0 ∈`0 [qn ]
B
2
(12)
Examples of potential choices for θbB include the posterior median or the
posterior mean (as in Johnstone and Silverman [20]), or the posterior mode
(as in Ročková [22]). (12) pertains only to a particular point estimate.
For a fully Bayesian interpretation, we say that the posterior distribution
contracts around the true θ0 at a rate at least as fast as the minimax l2 risk
if
n
2
sup Eθ0 Π θ : ||θ − θ0 || > Mn qn log
X → 0,
(13)
qn
θ0 ∈`0 [qn ]
for every Mn → ∞ as n → ∞. On the other hand, in another seminal paper, Ghosal et al. [15] showed that the posterior distribution cannot contract
10
faster than the minimax rate of qn log qnn around the truth. Hence, the optimal rate of contraction of a posterior distribution around the true θ0 must
be the minimax optimal rate in (11), up to some multiplicative constant. In
other words, if we use a fully Bayesian model to estimate a “nearly black”
normal mean vector, the minimax optimal rate should be our benchmark,
and the posterior distribution
should capture the true θ0 in a ball of squared
n
radius at most qn log qn (up to a multiplicative constant) as n → ∞.
In the subsequent section, we first prove that the IGGn prior does not
satisfy the conditions for posterior contraction given by van der Pas et al. [25]
for scale-mixture shrinkage priors (2). We then provide sufficient conditions
on the rate of decay of bn in (9) such that, under the IGGn prior, the
posterior mean attains the minimax risk, while the posterior distribution
contracts at the minimax rate. In order for (12) and (13) to hold for the
IGGn prior, the true sparsity level qn must be known. However, if qn is
unknown, then the IGGn prior can still attain near-minimax concentration
rates.
3.2
Minimax Posterior Contraction Under the IGG Prior
For scale-mixture shrinkage priors (2) where the priors on σi2 ’s, i = 1, ..., n,
are a priori independent, van der Pas et al. [25] gave sufficient conditions
under which the posterior contracts at the minimax rate given in (13). A
variety of densities π(σi2 ) are known to satisfy these conditions, including the
horseshoe and horseshoe+ priors, the spike-and-slab LASSO, the normalgamma prior, and the inverse Gaussian prior. We first demonstrate that
the IGG prior (9) prior can fail to satisfy these conditions before proceeding
to provide conditions under which it does achieve the minimax posterior
contraction rate. We first restate Theorem 2.1 from van der Pas et al. [25]
below:
Proposition 2 (Van der Pas et al., 2016). Suppose that X = (X1 , ..., Xn ) ∼
N (θ0 , In ) is observed, and assume that the prior on θ is of form (2). Suppose further that qn = o(n) and let Mn be an arbitrary positive sequence
tending to +∞. Suppose that the following conditions on the scale prior π
hold:
1. For some b ≥ 0, we can write u 7→ π(u) = Ln (u)e−bu , where Ln
is a function that is uniformly regular varying. That is, there exist
constants R, u0 ≥ 1 (which do not depend on n), such that
1
L(au)
≤
≤ R, for all a ∈ [1, 2], and all u ≥ u0 .
R
L(u)
11
Suppose further that there are constants C 0 , K, b0 ≥ 0 and u∗ ≥ 1 such
that
K
qn
0
0
C π(u) ≥
e−b u for all u ≥ u∗ .
n
R1
2. Suppose that there is a constant c > 0 such that
3. Let sn = qnn log qnn , and let dn =
a constant C such that
Z ∞
sn
!
p
0
π(u)du ≥ c.
log(n/qn ). Assume that there is
d3
u ∧ √n π(u)du + dn
u
Z d2
n π(u)
1
√ du ≤ Csn .
u
(14)
Then, under Conditions 1-3,
sup
θ0 ∈`0 [qn ]
Eθ0 Π θ : ||θ − θ0 ||2 > Mn qn log
n
qn
X
→ 0,
i.e. the posterior distribution under prior (2) contracts at the minimax contraction rate.
To briefly summarize, Condition 1 of Proposition 2 assumes that the posterior recovers nonzero means at optimal rate by ensuring that the tails decay
no faster than at exponential rate. Condition 2 ensures that π puts some
finite mass on values between [0, 1]. Finally, Condition 3 describes the decay
of π away from a neighborhood of zero. One easily checks that the IGG prior
satisfies both the first two conditions for any a ∈ ( 21 , ∞) and appropriately
chosen b, but it does not necessarily satisfy the third one, as we show in the
below lemma.
Lemma 1. Suppose that we observe X = (X1 , ..., Xn ) ∼ N (θ0 , In ). Suppose that a ∈ ( 21 , ∞) and b ∈ (0, 1) is a sequence such that b = o(1). Then
the prior density on the scale term in the IGG prior with hyperparameters
(a, b) fails to satisfy Condition 3 in Proposition 2.
Proof. See Appendix B.
Lemma 1 shows that if a ∈ ( 12 , ∞), then for any b ∈ (0, 1) where b = o(1),
the IGGn prior does not satisfy the conditions given by van der Pas et al.
[25]. However, as we show in Theorem 6, for a ∈ ( 12 , ∞), the IGG posterior
does in fact contract at minimax rate, provided that an appropriate rate of
decay is placed on b. Therefore, we have shown that while the conditions
12
given in Proposition 2 are sufficient for minimax posterior contraction for
the sparse normal means problem, they are not necessary.
We next study the mean square error (MSE) and the posterior variance
of the IGG prior and provide an upper bound on both. For all our results,
we assume that the true θ0 belongs to the set of nearly black vectors defined
by (10). With a suitably chosen rate for bn in (9), these upper bounds are
equal, up to a multiplicative constant, to the minimax risk. Utilizing these
bounds, we also show that the posterior distribution under the IGGn prior
(9) is able to contract around θ0 at minimax-optimal rates.
Since the priors (9) are independently placed on each θi , i = 1, ..., n, we
denote the resulting vector of posterior means (E(θ1 |X1 ), ..., E(θn |Xn )) by
T (X) and the ith individual posterior mean by T (Xi ). Therefore, T (X) is
the Bayes estimate of θ under squared error loss. Theorem 4 gives an upper
bound on the mean squared error for T (X).
Theorem 4. Suppose X ∼ N (θ0 , In ), where θ0 ∈ `0 [qn ]. Let T (X) denote
the posterior mean vector under (9). If a ∈ ( 21 , ∞), bn ∈ (0, 1) with bn → 0
as n → ∞, the MSE satisfies
1
sup Eθ0 ||T (X) − θ0 || - qn log
bn
θ0 ∈`0 [qn ]
2
s
+ (n − qn )bn log
1
,
bn
provided that qn → ∞ and qn = o(n) as n → ∞.
Proof. See Appendix B.
By the minimax result in Donoho et al. [13], we also have the lower
bound,
sup
θ0 ∈`0 [qn ]
Eθ0 ||T (X) − θ0 ||2 ≥ 2qn log
n
(1 + o(1)),
qn
as n, qn → ∞ and qn = o(n). The choice of bn =
qn α
n , for
n
with
qn
α ≥ 1, therefore
leads to an upper bound MSE of order qn log
a multiplicative
constant of at most 2α. Based on these observations, we immediately have
the following corollary.
Corollary 4.1. Suppose that qn is known, and that we set bn =
where α ≥ 1. Then under the conditions of Theorem 4,
n
sup Eθ0 ||T (X) − θ0 || qn log
.
qn
θ0 ∈`0 [qn ]
2
13
qn α
,
n
Corollary 4.1 shows that the posterior mean under the IGG prior performs
well as a point estimator for θ0 , as it is able to attain the minimax risk (possibly up to a multiplicative constant of at most 2 for α = 1). Although the
IGG prior does not include a point mass at zero, Proposition 1 and Corollary 4.1 together show that the pole at zero for the IGG prior mimics the
point mass well enough, while the heavy tails ensure that large observations
are not over-shrunk.
The next theorem gives an upper bound for the total posterior variance
corresponding to the IGGn (9) prior.
Theorem 5. Suppose X ∼ N (θ0 , In ), where θ0 ∈ `0 [qn ]. Under prior (7)
and the conditions of Theorem 4, the total posterior variance satisfies
n
X
1
V ar(θ0i |Xi ) - qn log
sup Eθ0
bn
θ0 ∈`0 [qn ]
i=1
s
+ (n − qn )bn log
1
,
bn
provided that qn → ∞ and qn = o(n) as n → ∞.
Proof. See Appendix B.
Having proven Theorems 4 and 5, we are now ready to state our main
theorem concerning optimal posterior contraction. Theorem 6 shows that
the IGG is competitive with other popular heavy-tailed priors like the globallocal shrinkage priors considered in Ghosh and Chakrabarti [16] or the
Dirichlet-Laplace prior considered by Bhattacharya et al. [6]. As before,
we denote the posterior mean vector under (9) as T (X).
Theorem 6. Suppose X ∼ N (θ0 , In ), where θ0 ∈ `0 [qn ]. Suppose that the
true sparsity level qn is known, with qn → ∞, and qn = o(n) as n → ∞.
α
Under prior (9), with a ∈ ( 21 , ∞) and bn = qnn , α ≥ 1,
sup
θ0 ∈`0 [qn ]
Eθ0 Π θ : ||θ − θ0 ||2 > Mn qn log
n
qn
n
qn
X
→ 0,
(15)
→ 0,
(16)
and
sup
θ0 ∈`0 [qn ]
Eθ0 Π θ : ||θ − T (X)||2 > Mn qn log
X
for every Mn → ∞ as n → ∞.
Proof. A straightforward application of Markov’s inequality combined with
the results of Theorems 4 and 5 leads to (15), while (16) follows from
Markov’s inequality combined with only the result of Theorem 5.
14
Theorem 6 shows that under mild regularity conditions, the posterior
distribution under the IGG prior contracts around both the true mean vector
and the corresponding Bayes estimates at least as fast as the minimax l2
risk in (11). Since the posterior distribution
cannot contract around the
n
truth faster than the rate of qn log qn (by Ghosal et al. [15]), the posterior
distribution for the IGG prior under the conditions of Theorem 6 must
contract around the true θ0 at the minimax optimal rate in (11) up to some
multiplicative constant.
We remark that the conditions needed to attain the minimax rate of
posterior contraction are quite mild. Namely, we only require that qn =
o(n), and we do not need to make any assumptions on the size of the true
signal size or the sparsity level. For comparison, Castillo and van der Vaart
[10] showed that the spike-and-slab prior
with a Gaussian slab contracts
2
at sub-optimal rate if ||θ0 || % qn log qnn . Bhattacharya et al. [6] showed
that given the Dir(a, ..., a) prior in the Dirichlet-Laplace prior, the posterior
contracts around θ0 at the minimax rate, provided that ||θ0 ||22 ≤ qn log4 n
if a = n−(1+β) , or provided that qn % log n if a = n1 . The IGGn prior (9)
removes these restrictions on θ0 and qn . Moreover, our minimax contraction
result does not rely on tuning or estimating a global tuning parameter τ , as
many previous authors have done, but instead, on appropriate selection of
hyperparameters a and b in the Bayesian hierarchy for the product density
of an IG(a, 1) and G(b, 1).
In reality, the true sparsity level of qn is rarely known, so the best that
we can do is to obtain the near-minimax contraction rate of qn log n. A
suitable modification of Theorem 6 leads to the following corollary.
Corollary 6.1. Suppose X ∼ N (θ0 , In ), where θ0 ∈ `0 [qn ]. Suppose that
the true sparsity level qn is unknown, but that qn → ∞, and qn = o(n) as
n → ∞. Under prior (9), with a ∈ ( 12 , ∞) and bn = n1α , α ≥ 1, then
sup
θ0 ∈`0 [qn ]
2
Eθ0 Π θ : ||θ − θ0 || > Mn qn log n X
→ 0,
(17)
and
sup
θ0 ∈`0 [qn ]
2
Eθ0 Π θ : ||θ − T (X)|| > Mn qn log n X
→ 0,
(18)
for every Mn → ∞ as n → ∞.
Having shown that the posterior mean under (9) attains the minimax risk
up to a multiplicative constant, and that its posterior density captures the
15
true θ0 in a ball of squared radius at most qn log n up to some multiplicative
constant, we now quantify its shrinkage profile around zero in terms of
Kullback-Leibler risk bounds. We show that this risk bound is in fact sharper
than other known shrinkage priors.
3.3
Kullback-Leibler Risk Bounds
In Section 3.2, we established that the choice of bn = n1 allows the IGGn
posterior to contract at near minimax rate, provided that a ∈ ( 12 , ∞). Figure
1 suggests that the shrinkage around zero is more aggressive for the IGGn
prior than it is for other known shrinkage priors when a and b are both set
to small values. In this section, we provide a theoretical justification for this
behavior near zero.
Carvalho et al. [9] and Bhadra et al. [5] showed that when the true data
generating model is N (0, In ), the Bayes estimate for the sampling density
of the horseshoe and the horseshoe+ estimators converge to the true model
at a super-efficient rate in terms of the Kullback-Leibler (K-L) distance
between the true model and the posterior density. They argue that as a
result, the horseshoe and horseshoe+ estimators squelch noise better than
other shrinkage estimators. However, in this section, we show that the IGGn
prior is able to shrink noise even more aggressively with appropriate chosen
bn .
Let θ0 be the true parameter value and f (y|θ) be the sampling model.
Further, let K(q1 , q2 ) = Eq1 log(q1 /q2 ) denote the K-L divergence of the
density q2 from q1 . The proof utilizes the following result by Clarke and
Barron [11].
Proposition 3. (Clarke and Barron, 1990). Let νn (dθ|y1 , ..., yn ) be the posterior distribution corresponding to some prior ν(dθ) after observing data
y(n) = (y1 , ..., yn ) according to theR sampling model f (y|θ). Define the posterior predictive density q̂n (y) = f (y|θ)νn (dθ|y1 , ..., yn ). Assume further
that ν(A ) > 0 for all > 0. Then the Cesàro-average risk of the Bayes
P
estimator, define as Rn ≡ n−1 nj=1 K(qθ0 , q̂j ), satisfies
Rn ≤ −
1
log ν(A ),
n
where ν(A ) denotes the measure of the set {θ : K(qθ0 , qθ ) ≤ }.
Using the above proposition, it is shown in Carvalho et al. [9] and Bhadra
et al. [5] that when the global parameter τ is fixed at τ = 1 and the true
16
parameter θ0 = 0, the horseshoe and the horseshoe+ both have Cesàroaverage risk which satisfies
Rn = O
1
n
log
n
(log n)d
,
(19)
where d is a positive constant. This rate is super-efficient, in the sense that
the risk is lower than that of the maximum likelihood estimator (MLE),
which has the rate O(log n/n) when θ0 = 0. The next theorem establishes
that the IGG prior can achieve an even faster rate of convergence O(n−1 )
in the K-L sense, with appropriate choices of a and b.
Theorem 7. Suppose that the true sampling model pθ0 is yj ∼ N (θ0 , 1).
Then for q̂n under the IGG prior with any a > 0 and bn = n1 , the optimal
rate of convergence of Rn when θ0 = 0 satisfies the inequality,
√
1
1
Rn ≤
2 + log( π) + (a + 2) log(2) + log a +
n
2
+
2 log n
,
n2
(20)
Proof. See Appendix C.
n
Since log
= o(n−1 ), we see from Theorem 7 that the IGGn posterior density
n2
with hyperparameters a > 0 and bn = n1 has an optimal convergence rate
of O(n−1 ). This convergence rate is faster than that of the horseshoe or
horseshoe+, both of which converge at the rate of O{n−1 (log n−d log log n)}
when θ0 = 0. To our knowledge, this is the sharpest known bound on
Cesàro-average risk for any Bayes estimator. Our result provides a rigorous
explanation for the observation that the IGG seems to shrink noise more
aggressively than other scale-mixture shrinkage priors.
Theorem 7 not only justifies the use of bn = n1 as a choice for hyperparameter b in the IGG prior, but it also provides insight into how we
should choose the hyperparameter a. (20) shows that the constant C in
Rn ≤ Cn−1 + o(n−1 ) can be large if a is set to be large. This theorem thus
implies that in order to minimize the K-L distance between N (θ0 , In ) and
the IGG posterior density, we should pick a to be small. Since we require
a ∈ ( 21 , ∞) in order to achieve the near-minimax contraction rate, our theoretical results suggest that we should set a ∈ ( 12 , 12 + δ] for small δ > 0 for
optimal posterior concentration.
17
4
4.1
Multiple Testing with the IGG Prior
Asymptotic Bayes Optimality Under Sparsity
Suppose we observe X = (X1 , ..., Xn ), such that Xi ∼ N (θi , 1), for i =
1, ..., n. To identify the true signals in X, we conduct n simultaneous tests:
H0i : θi = 0 against H1i : θi 6= 0, for i = 1, ..., n. For each i, θi is assumed to
be generated by a true data-generating model,
i.i.d.
θi ∼ (1 − p)δ{0} + pN (0, ψ 2 ), i = 1, ..., n,
(21)
where ψ 2 > 0 represents a diffuse “slab” density. This point mass mixture
model is often considered a theoretical ideal for generating a sparse vector
θ in the statistical literature. Indeed, Carvalho et al. [8] referred to model
(21) as a “gold standard” for sparse problems.
Model (21) is equivalent to assuming that for each i, θi follows a random variable whose distribution is determined by the latent binary random
variable νi , where νi = 0 denotes the event that H0i is true, while νi = 1
corresponds to the event that H0i is false. Here νi ’s are assumed to be
i.i.d. Bernoulli(p) random variables, for some p in (0, 1). Under H0i , i.e.
θi ∼ δ{0} , the distribution having a mass 1 at 0, while under H1i , θi 6= 0 and
is assumed to follow an N (0, ψ 2 ) distribution with ψ 2 > 0. The marginal
distributions of the Xi ’s are then given by the following two-groups model:
i.i.d.
Xi ∼ (1 − p)N (0, 1) + pN (0, 1 + ψ 2 ), i = 1, ..., n.
(22)
Our testing problem is now equivalent to testing simultaneously
H0i : νi = 0 versus H1i : νi = 1 for i = 1, ..., n.
(23)
We consider a symmetric 0-1 loss for each individual test and the total loss
of a multiple testing procedure is assumed to be the sum of the individual
losses incurred in each test. Letting t1i and t2i denote the probabilities of
type I and type II errors of the ith test respectively, the Bayes risk of a
multiple testing procedure under the two-groups model (1) is given by
R=
m
X
{(1 − p)t1i + pt2i }.
(24)
i=1
Bogdan et al. [7] showed that the rule which minimizes the Bayes risk in
(24) is the test which, for each i = 1, ..., n, rejects H0i if
f (xi |νi = 1)
1−p
>
, i.e. Xi2 > c2 ,
f (xi |νi = 0)
p
18
(25)
where f (xi |νi = 1) denotes the marginal density of Xi under H1i , while
2
f (xi |νi = 0) denotes that under H0i and c2 ≡ c2ψ,f = 1+ψ
(log(1 + ψ 2 ) +
ψ2
2 log(f )), with f = 1−p
p . The above rule is known as the Bayes Oracle,
because it makes use of unknown parameters ψ and p, and hence, it is not
attainable in finite samples. By reparametrizing as u = ψ 2 and v = uf 2 ,
the above threshold becomes
c2 ≡ c2u,v = 1 +
1
u
log v + log 1 +
1
u
.
Bogdan et al. [7] considered the following asymptotic scheme.
Assumption 1
The sequences of vectors (ψn , pn ) satisfies the following conditions:
1. pn → 0 as n → ∞.
2. un = ψn2 → ∞ as n → ∞.
3. vn = un f 2 = ψn2
4.
log vn
un
1−pn 2
pn
→ ∞ as n → ∞.
→ C ∈ (0, ∞) as n → ∞.
Bogdan et al. [7] provided detailed insight on the threshold C. Summarizing
briefly, if C = 0, then both the Type I and Type II errors are zero, and for
C = ∞, the inference is essentially no better than tossing a coin. Under
Assumption 1, Bogdan et al. [7] showed that the corresponding asymptotic
optimal Bayes risk has a particularly simple form, which is given by
√
BO
ROpt
= n((1 − p)tBO
+ ptBO
(26)
1
2 ) = np(2Φ( C) − 1)(1 + o(1)),
where the o(1) terms tend to zero as n → ∞. A testing procedure with risk
R is said to be asymptotically Bayes optimal under sparsity (ABOS) if
R
→ 1 as n → ∞.
BO
ROpt
4.2
(27)
An Optimal Testing Rule Based on the IGG Estimator
As noted earlier, the posterior mean depends heavily on the shrinkage factor,
κi = λi ξ1i +1 . Because of the concentration properties of the IGG prior proven
in Sections 2.1 and 3, a sensible thresholding rule classifies observations as
19
signals or as noise based on the posterior distribution of this shrinkage factor.
Consider the following testing rule for the ith observation Xi :
1
Reject H0i if E(1 − κi |Xi ) > ,
2
(28)
where κi is the shrinkage factor based on the IGGn prior (9). Within the
context of multiple testing, a good benchmark for our test procedure (28)
should be whether it is ABOS, i.e. whether its optimal risk is asymptotically
equal to that of the Bayes Oracle risk. Adopting the framework of Bogdan
et al. [7], we let RIGG denote the asymptotic Bayes risk of testing rule (28),
and we compare it to the ABOS risk defined in (26).
The next theorem illustrates that in the presence of sparsity, rule (28) is
in fact ABOS.
Theorem 8. Suppose that X1 , ..., Xn are i.i.d. observations having distribution (22) where the sequence of vectors (ψ 2 , p) satisfies Assumption 1.
Suppose we wish to test (23) using the classification rule (28). Suppose further that a ∈ ( 21 , ∞) and bn ∈ (0, 1), with bn → 0 as n → ∞ in such a way
1/4
that limn→∞
bn
pn
∈ (0, ∞). Then
lim
n→∞
RIGG
= 1,
BO
ROpt
(29)
i.e. rule (28) based on the IGGn prior (9) is ABOS.
Proof. See Appendix D.
We have shown that our thresholding rule based on the IGGn prior asymptotically attains the ABOS risk exactly, provided that bn decays to zero at a
certain rate relative to the sparsity level p. For example, if the prior mixing
proportion pn is known, we can set the hyperparameter bn = p4n . Then the
conditions for classification rule (28) to be ABOS are satisfied.
Our work ultimately moves the testing problem beyond the global-local
framework. Previously, Datta and Ghosh [12], Ghosh et al. [17], Bhadra
et al. [5], and Ghosh and Chakrabarti [16]) have shown that horseshoe or
horseshoe-like priors asymptotically attain the Bayes Oracle risk (possibly
up to a multiplicative constant) either by specifying a rate for the global
parameter τ in (3) or by estimating it with an empirical Bayes plug-in estimator. In the case with the IGG prior, we prove that our thresholding
rule based on the posterior mean is ABOS without utilizing a shared global
tuning parameter.
20
5
5.1
Simulation Studies
Computation and Selection of Hyperparameters
Letting κi =
1
1+λi ξi ,
the full conditional distributions for (8) are
κi , i = 1, ..., n,
θi rest ∼ N (1
− κi )Xi , 1 −
2
1 θi
λi rest ∼ IG a + 2 , 2ξi + 1 , i = 1, ..., n,
ξi rest ∼ giG
θi2
λi , 2, b
−
1
2
(30)
, i = 1, ..., n,
where giG(a, b, p) denotes a generalized inverse Gaussian (giG) density with
f (x; a, b, p) ∝ x(p−1) e−(a/x+bx)/2 . Therefore, the IGG model (8) can be implemented straightforwardly with Gibbs sampling, utilizing the full conditionals in (30).
For all our simulations, we set a = 21 + n1 and b = n1 , in light of Theorems
6 and 7. These choices of a and b ensure that the IGG posterior will contract
around the true θ0 at least at near-minimax rate, while keeping a ∈ ( 21 , ∞)
small. We denote our IGG prior with hyperparameters (a, b) = 21 + n1 , n1
as IGG1/n . For both of the simulation studies described below, we run
10,000 iterations of a Gibbs sampler, discarding the first 5000 as burn-in.
5.2
Simulation Study for Sparse Estimation
To illustrate finite-sample performance of the IGG1/n prior, we use the set-up
in Bhadra et al. [5] where we specify sparsity levels of q/n = 0.05, 0.10, 0.20,
and 0.30, and set the signals all equal to values of either A = 7 or 8, for
a total of eight simulation settings. With n = 200, we randomly generate
n-dimensional vectors under these settings and compute the average squared
error loss corresponding to the posterior median across 100 replicates.
We compare our results for IGG1/n to the average squared error loss of
the posterior median under the Dirichlet-Laplace (DL), the horseshoe (HS),
and the horseshoe+ (HS+) estimators, since these are global-local shrinkage
priors (3) with singularities at zero. For the HS and HS+ priors, we use a
fully Bayesian approach, with τ ∼ C + (0, 1), as in Ghosh et al. [17]. For the
DL prior, we specify a = n1 in the Dir(a, ..., a) prior of the scale component,
along with τ ∼ G(na, 1/2), as in Bhattacharya et al. [6]. Our results are
presented in Table 1.
Table 1 shows that under these various sparsity and signal strength settings, the IGG1/n ’s posterior median has the lowest (estimated) squared error loss in nearly all of the simulation settings. It performs better than the
21
q/n
0.05
0.10
0.20
0.30
A
7
8
7
8
7
8
7
8
IGG
13.88
13.34
27.21
25.95
49.78
47.24
74.42
70.83
DL
14.30
13.27
29.91
27.67
56.40
52.22
85.72
79.03
HS
18.11
17.71
35.91
34.77
71.18
69.81
104.67
104.02
HS+
14.41
13.96
30.18
29.36
58.25
57.11
86.00
84.70
Table 1: Comparison of average squared error loss for the posterior median
estimate of θ across 100 replications. Results are reported for the IGG1/n ,
DL (Dirichlet-Laplace), HS (horseshoe), and the HS+ (horseshoe-plus).
horseshoe and the horseshoe+ in all settings. Our empirical results confirm
the theoretical properties that were proven in Section 3 and illustrate that
for finite samples, the IGG prior often outperforms other popular shrinkage priors. Our empirical results also lend strong support to the use of the
inverted beta prior β 0 (a, b) as the scale density in scale-mixture shrinkage
priors (2). However, our results suggest that we can obtain better estimation if we allow the a and b to vary with sample size, rather than keeping
them fixed (as the horseshoe priors do, with a = b = 0.5).
5.3
Simulation Study for Multiple Testing
For the multiple testing rule (28), we adopt the simulation framework of
Datta and Ghosh [12] and Ghosh et al. [17] and fix sparsity levels at p ∈
{0.01, 0.05, 0.10, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5} for a total of 11 simulation settings. For sample size n = 200
√ and each p, we generate data from
the two-groups model (21), with ψ = 2 log n = 3.26. We then apply the
thresholding rule (28) using IGG1/n to classify θi ’s in our model as either
signals (θi 6= 0) or noise (θi = 0). We estimate the average misclassification
probability (MP) for the thresholding rule (28) from 100 replicates.
Taking p = 0.10, we plot in Figure 2 the theoretical posterior inclusion
probabilities ωi (Xi ) = P (νi = 1|Xi ) for the two-groups model (21) given by
(
ωi (Xi ) = π(νi = 1|Xi ) =
)−1
X 2 ψ2
i
1−p q
−
1 + ψ 2 e 2 1+ψ2 + 1
,
p
along with the shrinkage weights 1 − E(κi |Xi ) corresponding to the IGG1/n
22
1.0
●
●
● ●
●
●
●
● ●
0.8
0.6
●
0.4
●
●
●
●
●
0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
0.0
Posterior Inclusion Probability
●
−10
−5
0
5
10
X
Figure 2: Comparison between the posterior inclusion probabilities and the
posterior shrinkage weights 1 − E(κi |Xi ) when p = 0.10.
prior. The circles in the figure denote the theoretical posterior inclusion
probabilities, while the triangles correspond to the shrinkage weights 1 −
E(κi |Xi ). The figure clearly shows that for small values of the sparsity level
p, the shrinkage weights are in close proximity to the posterior inclusion
probabilities. This and the theoretical results established in Section 4 justify
the use of using 1 − E(κi |Xi ) as an approximation to the corresponding
posterior inclusion probabilities ωi (Xi ) in sparse situations. Therefore, this
motivates the use of the IGG1/n prior (9) and its corresponding decision rule
(28) for identifying signals in noisy data.
Figure 3 shows the estimated misclassification probabilities (MP) for
decision rule (28) for the IGG1/n prior, along with the estimated MP’s for the
Bayes Oracle (BO), the Benjamini-Hochberg procedure (BH), the DirichletLaplace (DL), the horseshoe (HS), and the horseshoe+ (HS+). The Bayes
Oracle rule, defined in (25), is the decision rule that minimizes the expected
number of misclassified signals (24) when (p, ψ) are known. The Bayes
Oracle therefore serves as the lower bound to the MP, whereas the line
M P = p corresponds to the situation where we reject all null hypotheses
without looking into the data. For the Benjamini-Hochberg rule, we use
αn = 1/ log n = 0.1887. Bogdan et al. [7] theoretically established the
ABOS property of the BH procedure for this choice of αn . For the DL, HS,
23
0.5
0.4
0.3
0.2
0.1
0.0
Misclassification Probability
MP=p
Oracle
BH
IGG
DL
HS
HS+
0.0
0.1
0.2
0.3
0.4
0.5
Sparsity
Figure 3: Estimated misclassification probabilities. Thresholding rule (28)
based on the IGG posterior mean is nearly as good as the Bayes Oracle (25).
and HS+ priors, we use the classification rule
1
Reject H0i if E(1 − κi |X1 ..., Xn ) > ,
2
where κi =
1
1+σi2
(31)
and σi2 is the scale parameter in the scale-mixture shrink-
age model (2). For the horseshoe and horseshoe+ priors, we specify a halfCauchy prior on the global parameter τ ∼ C + (0, 1). Since τ is a shared
global parameter, the posterior for κi depends on all the data. Carvalho
et al. [9] first introduced thresholding rule (31) for the horseshoe. Ghosh
et al. [17] later extended rule (31) for a general class of global-local shrinkage priors (3), which includes the Strawderman-Berger, normal-exponentialgamma, and generalized double Pareto priors. Based on Ghosh et al. [17]’s
simulation results, the horseshoe performs similarly as or better than these
other aforementioned priors, so we do not include these other priors in our
comparison study.
Our results provide strong support to our theoretical findings in Section
4 and strong justification for the use of (28) to classify signals. As Figure 3
illustrates, the misclassification probability for the IGG prior with (a, b) =
( 21 + n1 , n1 ) is practically indistinguishable from the Bayes Oracle, which gives
the lowest possible MP. Thresholding rule (31) based on the horseshoe+
24
p
0.30
0.35
0.40
0.45
0.50
BO
0.08
0.08
0.09
0.10
0.10
BH
0.13
0.12
0.11
0.10
0.09
IGG
0.005
0.004
0.004
0.003
0.003
DL
0.08
0.09
0.10
0.12
0.13
HS
0.14
0.22
0.31
0.39
0.43
HS+
0.09
0.10
0.10
0.11
0.11
Table 2: Comparison of false discovery rate (FDR) for different classification
methods under dense settings. The IGG1/n has the lowest FDR of all the
different methods.
prior and the Dirichlet-Laplace priors also appears to be quite competitive
compared to the Bayes Oracle. Bhadra et al. [5] proved that the horseshoe+
prior asymptotically matches the Bayes Oracle risk up to a multiplicative
constant if τ is treated as a tuning parameter, but did not prove this for the
case where τ is endowed with a prior. There also does not appear to be any
theoretical justification for thresholding rule (31) under the DL prior in the
literature. On the other hand, Theorem 8 provides theoretical support for
the use of (28) under the IGG prior, which is confirmed by our empirical
study.
Figure 3 also shows that the performance for rule (31) under the horseshoe degrades considerably as θ = (θ1 , ..., θn ) becomes more dense. With
sparsity level p = 0.5, the horseshoe’s misclassification rate is close to 0.4,
only marginally better than rejecting all the null hypotheses without looking at the data. This phenomenon was also observed by Datta and Ghosh
[12] and Ghosh et al. [17]. This appears to be because in the dense setting,
there are many noisy entries that are “moderately” far from zero, and the
horseshoe prior does not shrink these aggressively enough towards zero in
order for testing rule (31) to classify these as true noise. The horseshoe+
prior seems to alleviate this by adding an additional half-Cauchy C + (0, 1)
prior to the Bayes hierarchy. In Table 2, we report the false discovery rate
(FDR) under dense settings for the different methods. We see that the FDR
is quite a bit larger for the horseshoe than for the other methods. Table
2 also shows that the IGG1/n prior has very tight control over the FDR in
dense settings. Although the IGG prior is not constructed to specifically
control FDR, we see that in practice, it does provide excellent control of
false positives.
Finally, we demonstrate the shrinkage properties corresponding to the
IGG1/n prior along with the horseshoe, the horseshoe+, and the Dirichlet-
25
10
5
0
−5
−10
Posterior Expectation
Flat
IGG
HS
HS+
DL
−10
−5
0
5
10
X
Figure 4: Posterior Mean E(θ|X) vs. X plot for p = 0.25.
Laplace priors. In Figure 4, we plot the posterior expectations E(θi |Xi )
for the IGG1/n prior and the posterior expectations E(θi |X1 , ..., Xn ) for the
HS, HS+, and DL priors and the posterior expectations. The amount of
posterior shrinkage can be observed in terms of distance between the 45◦
line and the posterior expectation. Figure 4 clearly shows that near zero,
the noisy entries are more aggressively shrunk towards zero for the IGG1/n
prior than for the other priors with poles at zero. This confirms our findings
in Theorem 7 which proved that the shrinkage profile near zero is more
aggressive for the IGG1/n prior in the Kullback-Leibler sense than for the
HS or HS+ priors. Meanwhile, Figure 4 also shows that the signals are left
mostly unshrunk, confirming that the IGG shares the same tail robustness
as the other priors. The more aggressive shrinkage of noise explains why
the IGG performs better estimation, as we demonstrated in Section 5.2.
6
Analysis of a Prostate Cancer Data Set
We demonstrate practical application of the IGG prior using a popular
prostate cancer data set introduced by Singh et al. [23]. In this data set,
there are gene expression values for n = 6033 genes for m = 102 subjects,
with m1 = 50 normal control subjects and m2 = 52 prostate cancer patients.
We aim to identify genes that are significantly different between control
26
and cancer patients. This problem can be reformulated as normal means
problem (1) by first conducting a two-sample t-test for each gene and then
transforming the test statistics (t1 , ..., tn ) to z-scores using the inverse normal cumulative distribution function (CDF) transform Φ−1 (Ft100 (ti )), where
Ft100 denotes the CDF for the Student’s t distribution with 100 degrees of
freedom. With z-scores (z1 , ..., zn ), our model is now
zi = θi + i ,
i = 1, ..., n,
(32)
where i ∼ N (0, 1). This allows us to implement the IGG prior on the
z-scores to conduct simultaneous testing H0i : θi = 0 vs. H1i : θi 6= 0,
i = 1, ..., n, to identify genes that are significantly associated with prostate
cancer. Additionally, we can also estimate θ = (θ1 , ..., θn ). As argued by
Efron [14], |θi | can be interpreted as the effect size of the ith gene for prostate
cancer. Efron [14] first analyzed model (32) for this particular data set by
obtaining empirical Bayes estimates θbiEf ron , i = 1, ..., n, based on the twogroups model (21). In our analysis, we use the posterior means θbi , i = 1, ..., n,
to estimate the strength of association.
With our z-scores, we implement the IGG1/n model with (a, b) = ( 21 +
1 1
n , n ) on model (32) and use classification rule (28) to identify significant
genes. For comparison, we also fit this model for the DL, HS, and HS+
priors, and benchmark it to the Benjamini-Hochberg (BH) procedure with
FDR α set to 0.10. The IGG1/n selects 85 genes as significant, in comparison to 60 genes under the BH procedure. The HS prior selects 62 genes
as significant. The HS+ and DL priors select 41 and 42 genes respectively,
indicating more conservative estimates. All 60 of the genes flagged as significant by the BH procedure are included in the 85 genes that the IGG
prior classifies as significant. On the other hand, the HS prior’s conclusions
diverge from the BH procedure. Seven genes (genes 11, 377, 637, 805, 1588,
3269, and 4040) are deemed significant by the HS, but not by BH.
Table 3 shows the top 10 genes selected by Efron [14] and their estimated
effect size on prostate cancer. We compare Efron [14]’s empirical Bayes
posterior mean estimates with the posterior mean estimates under the IGG,
DL, HS, and HS+ priors. Our results confirm the tail robustness of the IGG
prior. All of the scale-mixture shrinkage priors shrink the estimated effect
size for significant genes less aggressively than Efron’s procedure. Table 3
also shows that for large signals, the IGG posterior has slightly less shrinkage
for large signals than the DL posterior and roughly the same amount as the
HS posterior. The HS+ posterior shrinks the test statistics the least for
large signals, but the IGG’s estimates are still quite similar to those of the
HS+.
27
Gene
610
1720
332
364
914
3940
4546
1068
579
4331
z-score
5.29
4.83
4.47
-4.42
4.40
-4.33
-4.29
4.25
4.19
-4.14
θbiIGG
4.85
4.33
3.78
-3.78
3.71
-3.70
-3.59
3.49
3.31
-3.41
θbiDL
4.52
3.94
3.40
-3.10
3.11
-3.06
-3.09
3.09
2.98
-2.87
θbiHS
4.85
4.33
3.78
-3.78
3.71
-3.70
-3.59
3.49
3.31
-3.41
θbiHS+
4.91
4.35
3.99
-3.85
3.86
-3.80
-3.62
3.46
3.01
-3.43
θbiEf ron
4.11
3.65
3.24
-3.57
3.16
-3.52
-3.47
2.99
2.92
-3.30
Table 3: The z-scores and the effect size estimates for the top 10 genes
selected by Efron [14] by the IGG, DL, HS, and HS+ models and the twogroups empirical Bayes model by Efron [14].
7
Concluding Remarks
In this paper, we have introduced a new scale-mixture shrinkage prior called
the Inverse Gamma-Gamma prior for estimating sparse normal mean vectors. This prior has been shown to have a number of good theoretical properties, including heavy probability mass around zero and heavy tails. This
enables the IGG prior to perform selective shrinkage and to attain (near)
minimax contraction around the true θ in (1). The IGG posterior also converges to the true model at a faster rate than the horseshoe and horseshoe+
posterior densities in the Kullback-Leibler sense. The IGG, HS, and HS+
all fall under the class of priors which utilize a beta prime density (4) as a
prior on the scale component for model (2). However, our results suggest
that there is added flexibility in allowing the parameters (a, b) in the density
(4) to vary with sample size rather than keeping them fixed. This added
flexibility leads to excellent empirical performance and obviates the need to
estimate a global tuning parameter τ .
Moreover, by thresholding the posterior mean, the IGG can be used to
identify signals in θ. We have investigated the asymptotic risk properties
of this classification rule within the decision theoretic framework of Bogdan
et al. [7] and established asymptotically optimal theoretical properties for
multiple testing. Since we do not specify or estimate a global parameter τ ,
our paper appears to be the first article to establish the ABOS property for
a scale-mixture shrinkage prior which does not fall under the global-local
28
(3) framework.
Our simulation studies demonstrate the IGG’s strong finite sample performance for obtaining sparse estimates of θ and for correctly classifying
entries in θ as either signals or noise. By setting the hyperparameters
(a, b) = ( 21 + n1 , n1 ), the IGG prior outperforms other popular shrinkage
estimators. Finally, we demonstrated practical application of the IGG prior
with a prostate cancer data set.
In recent years, Bayesian scale-mixture shrinkage priors have gained a
great amount of attention because of their computational efficiency and
their ability to mimic point-mass mixtures in obtaining sparse estimates.
Our paper contributes to this large body of methodological and theoretical
work. There are a few possible future directions for research. For example,
the IGG prior can be adapted to other statistical problems such as sparse
covariance estimation, variable selection with covariates, and many others.
We conjecture that the IGG would satisfy many optimality properties (e.g.
model selection consistency, optimal posterior contraction, etc.) if it were
utilized in these other contexts.
Despite the absence of a data-dependent global parameter τ , the IGG
model adapts well to sparsity, performing well under both sparse and dense
settings. This seems to be in stark contrast to remarks made by authors
like Carvalho et al. [8] who have argued that scale-mixture shrinkage priors
which do not contain shared global parameters do not enjoy the benefits of
adaptivity. Nevertheless, we could investigate if theoretical and empirical
performance can be improved even further by incorporating a global parameter into the IGG framework. We leave these as interesting problems for
future research.
8
Acknowledgments
The authors would like to thank Dr. Anirban Bhattacharya and Dr. Xueying Tang for sharing their codes, which were modified to generate Figures
1-4.
A
Proofs for Section 2
Proof of Proposition 1. The joint distribution of the prior is proportional to
−1/2
π(θ, ξ, λ) ∝ (λξ)
θ2
exp −
2λξ
!
29
λ−a−1 exp −
1 b−1
ξ
exp (−ξ)
λ
!
∝ξ
b−3/2
−a−3/2
exp(−ξ)λ
!
θ2
1
exp −
+1
.
2ξ
λ
Thus,
π(θ, ξ) ∝ ξ
b−3/2
!
Z ∞
λ
exp(−ξ)
−a−3/2
λ=0
θ2
+1
2ξ
∝
!
θ2
1
exp −
+1
dλ
2ξ
λ
!−(a+1/2)
ξ b−3/2 e−ξ ,
and thus, the marginal density of θ is proportional to
π(θ) ∝
θ2
+1
2ξ
Z ∞
0
!−(a+1/2)
ξ b−3/2 e−ξ dξ.
(33)
As |θ| → 0, the expression in (33) is bounded below by
Z ∞
C
ξ b−3/2 e−ξ dξ,
(34)
0
where C is a constant that depends on a and b. The integral expression in
(34) clearly diverges to ∞ for any 0 < b ≤ 1/2. Therefore, (33) diverges to
infinity as |θ| → 0, by the monotone convergence theorem.
Proof of Theorem 1. From (7), the posterior distribution of κi under IGGn
is proportional to
κi Xi2
π(κi |Xi ) exp −
2
Since exp
κ X2
− i2 i
!
a−1/2
κi
(1 − κi )bn −1 , κi ∈ (0, 1).
is strictly decreasing in κi on (0, 1), we have
Z 1
a−1/2
κi
0
E(1 − κi |Xi ) = Z
1
0
(1 − κi )
a−1/2
κi
(1
2
eXi /2
≤
bn
κi
Z 1
a−1/2
0
κi
!
κi Xi2
− κi )bn −1 exp −
2
Z 1
a−1/2
0
κi Xi2
exp −
2
(1 − κi )bn dκi
(1 − κi )bn −1 dκi
30
dκi
!
dκi
(35)
Γ(a + 1/2)Γ(bn + 1)
Γ(a + bn + 1/2)
×
Γ(a + bn + 3/2)
Γ(a + 1/2)Γ(bn )
b
2
n
= eXi /2
.
a + bn + 1/2
2
= eXi /2
a−1/2
Proof of Theorem 2. Note that since a ∈ ( 12 , ∞), κi
is increasing in κi
on (0, 1). Additionally, since bn ∈ (0, 1), (1 − κi )bn −1 is increasing in κi on
(0, 1). Using these facts, we have
Z
P (κi < |Xi ) ≤ Z
0
κi Xi2
exp −
2
!
κi Xi2
exp −
2
!
1
Z
a−1/2
2
eXi /2
0
≤
κi
Z 1
a−1/2
κi
a−1/2
(1 − κi )bn −1 dκi
a−1/2
(1 − κi )bn −1 dκi
κi
κi
(1 − κi )bn −1 dκi
(1 − κi )bn −1 dκi
2
eXi /2 (1 − )bn −1
Z
a−1/2
κi
0
≤
a−1/2
Z 1
dκi
(1 − κi )bn −1 dκi
2
eXi /2 (1 − )bn −1 a +
=
1
2
−1
a+1/2
a−1/2 (1 − )b
b−1
n
bn
2
= eXi /2
.
(a + 1/2) (1 − )
Proof of Theorem 3. First, note that since bn ∈ (0, 1), (1−κi )bn −1 is increasing in κi on (0, 1). Therefore, letting C denote the normalizing constant that
depends on Xi , we have
Z η
π(κi |Xi )dκi = C
0
Z η
0
≥C
κi Xi2
exp −
2
Z ηδ
0
!
κi Xi2
exp −
2
31
a−1/2
κi
(1 − κi )bn −1 dκi
!
a−1/2
κi
(1 − κi )bn −1 dκi
ηδ
ηδ
a−1/2
≥ C exp − Xi2
κi
dκi
2
0
1
ηδ
1 −1
(ηδ)a+ 2 .
= C exp − Xi2
a+
2
2
Z
a−1/2
Also, since a ∈ ( 12 , ∞), κi
Z 1
π(κi |Xi )dκi = C
Z 1
η
η
(36)
is increasing in κi on (0, 1).
κi Xi2
exp −
2
ηX 2
≤ C exp − i
2
!Z
ηX 2
≤ C exp − i
2
!Z
ηX 2
= C exp − i
2
!
!
a−1/2
κi
1
a−1/2
κi
η
1
(1 − κi )bn −1 dκi
(1 − κi )bn −1 dκi
(1 − κi )bn −1 dκi
η
bn
b−1
n (1 − η) .
(37)
Combining (36) and (37), we have
Z 1
P (κi > η|Xi ) ≤ Z
π(κi |Xi )dκi
η
η
π(κi |Xi )dκi
0
a+
≤
B
1
2
(1 − η)bn
1
bn (ηδ)a+ 2
η(1 − δ) 2
exp −
Xi .
2
Proofs for Section 3.2
Proof of Lemma 1. It is enough to show that the second term on the lefthand side in (14) satisfies
dn
Z d2
n π(u)
1
√ du % sn .
u
(38)
We utilize the prior formulation in (5), so that the prior on the scale term
is
π(u) ∝ ua−1 (1 + u)−(a+b) .
32
Then for any a ∈ ( 12 , ∞), we have
dn
Z d2
n π(u)
√ du = dn
u
1
Z d2
n
ua−3/2 (1 + u)−(a+b) du
1
% (1 + d2n )−(a+b) d2a
n
= (1 +
Since d2n = log
d2n
n
qn
d2n )−(1+b)
d2n
1 + d2n
!a−1
d2n
(39)
→ ∞ as n → ∞ and b = o(1) by assumption, we have
a−1
∼ 1 and (1+d2n )−(1+b) ∼ (1+d2n )−1 for large n. Additionally, from
√
√
the inequality, log(1 + x) ≤ x for x ≥ 0, we have (1 + x)−1 ≥ exp (− x)
for x ≥ 0. Combining these with (39), we have
1+d2n
dn
Z d2
n π(u)
1
√ du % d2n (1 + d2n )−1
u
≥ d2n exp (−dn )
≥ d2n exp(−d2n )
n qn
= log
qn n
= sn ,
(for large n)
and thus (38) holds for any a ∈ ( 21 , ∞) and b = o(1).
Before proving Theorems 4 and 5, we first state two lemmas. For both
Lemmas 2 and 3, we denote T (x) = {E(1 − κ)|x}x as the posterior mean
1
under (9) for a single observation x, where κ = 1+λξ
. Our arguments follow
closely those of van der Pas et al. [27], Datta and Ghosh [12], and Ghosh and
Chakrabarti [16], except that their arguments rely on controlling the rate
of decay of tuning parameter τ or an empirical Bayes estimator τb. In our
case, since we are dealing with a fully Bayesian model, the degree of posterior
contraction is instead controlled by the positive sequence of hyperparameters
bn in (9).
Lemma 2. Let T (x) be the posterior mean under (9) for a single observation
x drawn from N (θ, 1). Suppose we have constants η ∈ (0, 12 ), δ ∈ (0, 1),
a ∈ ( 12 , ∞), and bn ∈ (0, 1), where bn → 0 as n → ∞. Then for any d > 2
and fixed n, |T (x) − x| can be bounded above by a real-valued function hn (x),
depending on d and satisfying the following:
33
For any ρ > d, hn (·) satisfies
lim
n→∞
|x|>
hn (x) = 0.
qsup
ρ log
1
bn
(40)
Proof of Lemma 2. Fix η ∈ (0, 1), δ ∈ (0, 1). First observe that
|T (x) − x| = |xE(κ|x)|
≤ |xE(κ1{κ < η}| + |xE(κ1{κ > η}|.
(41)
We consider the two terms in (41) separately. From (7) and the fact that
(1 − κ)bn −1 is increasing in κ ∈ (0, 1) when bn ∈ (0, 1), we have
2
Rη
κ · κa−1/2 (1 − κ)bn −1 e−κx /2 dκ
|xE(κ1{κ < η}| = x R 1
a−1/2 (1 − κ)bn −1 e−κx2 /2 dκ
0 κ
0
≤ x
= (1
R η a+1/2 −κx2 /2
e
dκ
bn −1 0 κ
(1 − η)
R1
2
a−1/2 e−κx /2 dκ
0 κ
R ηx2 t a+1/2 −t/2
e
dt
0
x2
− η)bn −1 R a−1/2
x2
t
e−t/2 dt
0
x2
= (1 − η)
bn −1
x
2
1 0ηx ta+1/2 e−t/2 dt
x
R
x2 x2 ta−1/2 e−t/2 dt
R
0
R ∞ a+1/2 −t/2
e
dt
−1
bn −1 0 t
x
≤ (1 − η)
R x2
0
ta−1/2 e−t/2 dt
#−1
" Z 2
x
a−1/2 −t/2
t
= C(n)
e
x
dt
−1
0
= h1 (x) (say),
(42)
where we use a change of variables t = κx2 in the second equality, and
C(n) = (1 − η)bn −1 Γ(1)Γ(a+3/2)
= (1 − η)bn −1 a +
Γ(a+5/2)
Next, we observe that since κ ∈ (0, 1),
3
2
−1
.
|xE(κ1{κ > η}|x| ≤ |xP (κ > η|x)|
≤
a+
1
2
(1 − η)bn
bn (ηδ)a+1/2
= h2 (x) (say),
34
η(1 − δ) 2
|x| exp −
x
2
(43)
where we use Theorem 3 for the second inequality.
Let hn (x) = h1 (x) + h2 (x). Combining (41)-(43), we have that for every
x ∈ R and fixed n,
|T (x) − x)| ≤ hn (x),
(44)
Observe from (42) that for fixed n, h1 (x) is strictly decreasing in |x|. Therefore, we have that for any fixed n and ρ > 0,
"s
|x|>
qsup
h1 (x) ≤ C(n)
ρ log
1
bn
1
ρ log
bn
Z ρ log
1
bn
#−1
t
a−1/2 −t/2
e
,
dt
0
and since bn → 0 as n → ∞, this implies that
lim
n→∞
|x|>
h1 (x) = 0.
qsup
1
bn
ρ log
(45)
Next, observe that from (43) that for fixed n, h2 (x) is eventually decreasing
in |x| with a maximum when |x| = √ 1
. Therefore, for sufficiently large
η(1−δ)
n, we have
s
h2 (x) ≤ h2
qsup
|x|>
1
bn
ρ log
Letting K ≡ K(a, η, δ) =
0 < bn < 1 for all n that
s
lim h2
n→∞
1
ρ log
bn
!
(a+ 12 )
,
(ηδ)a+1/2
1
ρ log
bn
1
≤ K lim
n→∞ bn
s
s
ρ log
√
= K ρ lim (bn )
1 − η(1−δ)
ρ log
2
e
ρ log
bn
1
bn
η(1−δ)
ρ
1
e 2 log(bn )
bn
η(1−δ)
2
n→∞
=
.
we have from (43) and the fact that
(1 − η)bn
= K lim
n→∞
bn
(
!
2
ρ− η(1−δ)
s
1
log
bn
2
0 if ρ > η(1−δ)
,
∞ otherwise,
from which it follows that
(
lim
n→∞
|x|>
qsup
ρ log
h2 (x) =
1
bn
35
2
0 if ρ > η(1−δ)
,
∞ otherwise.
(46)
Combining (45) and (46), we have for hn (x) = h1 (x) + h2 (x) that
(
lim
n→∞
|x|>
qsup
hn (x) =
ρ log
1
bn
2
0 if ρ > η(1−δ)
,
∞ otherwise.
(47)
Since η ∈ (0, 1), δ ∈ (0, 1), it is clear that any real number lager than 2 can
2
be expressed in the form η(1−δ)
. For example, taking η = 56 and δ = 15 , we
2
= 3. Hence, given any d > 2, choose 0 < η, δ < 1 such that
obtain η(1−δ)
2
c = η(1−δ) . Clearly, hn (·) depends on d. Following (44) and (47), we see
that |T (x) − x| is uniformly bounded above by hn (x) for all n and d > 2
and that condition (40) is also satisfied when d > 2. This completes the
proof.
Remark: Under the conditions of Lemma 2, we see that for any fixed n,
lim |T (x) − x| = 0.
(48)
|x|→∞
Equation (48) shows that for the IGG prior, large observations almost remain unshrunk no matter what the sample size n is. This is critical to its
ability to properly identify signals in our data. We now present our second
lemma which bounds the posterior variance.
Lemma 3. Let T (xi ) be the posterior mean under (9). Then for a single
observation x ∼ N (θ, 1), the posterior variance V ar(θ|x) can be bounded
above by
V ar(θ|x) ≤ 1 − (T (x) − x)2 + x2 ,
(49)
and
V ar(θ|x) ≤ 1 − T (x)2 + x2 .
(50)
Proof of Lemma 3. We first prove (49). By the law of the iterated variance
and the fact that θ|κ, x ∼ N ((1 − κ)x, 1 − κ), we have
V ar(θ|x) = E[V ar(θ|κ, x)] + V ar[E(θ|κ, x)]
= E(1 − κ|x) + V ar[(1 − κ)x|x]
= E(1 − κ|x) + x2 V ar(κ|x)
= E[(1 − κ|x) + x2 E(κ2 |x) − x2 [E(κ|x)]2 .
Since x − T (x) = xE(κ|x), we rewrite the above as
2
1 a+3/2
κ
(1 − κ)bn −1 e−κx /2 dx
T (x)
V ar(θ|x) =
− (T (x) − x)2 + x2 R01
a+1/2 (1 − κ)bn −1 e−κx2 /2 dx
x
0 κ
R
36
≤ 1 − (T (x) − x)2 + x2 ,
3
1
where the inequality follows from the fact that κ ∈ (0, 1), so κa+ 2 ≤ κa+ 2
for all a ∈ R.
Next, we show that (50) holds. We may alternatively represent V ar(θ|x)
as
V ar(θ|x) = E(1 − κ|x) + x2 E[(1 − κ)2 |x] − x2 E2 [(1 − κ)|x]
2
1 a+1/2
κ
(1 − κ)bn +1 e−κx /2 dx
T (x)
=
− T (x)2 + x2 R01
a+1/2 (1 − κ)bn −1 e−κx2 /2 dx
x
0 κ
R
≤ 1 − T (x)2 + x2 ,
where the final inequality holds because (1 − κ)bn +1 ≤ (1 − κ)bn −1 for all
bn ∈ R when κ ∈ (0, 1).
Lemmas 2 and 3 are crucial in proving Theorems 4 and 5, which provide
asymptotic upper bounds on the mean squared error (MSE) for the posterior
mean under the IGGn prior (9) and the posterior variance under (9). These
theorems will ultimately allow us to provide sufficient conditions under which
the posterior mean and posterior distribution under the IGGn prior contract
at minimax rates.
Proof of Theorem 4. Define qen = #{i : θ0i 6= 0}. We split the MSE,
Eθ0 ||T (X) − θ0 ||2 =
n
X
Eθ0i (T (Xi ) − θ0i )2
i=1
as
n
X
i=1
Eθ0i (T (Xi ) − θ0i )2 =
X
Eθ0i (T (Xi ) − θ0i )2 +
i:θ0i 6=0
X
Eθ0i (T (Xi ) − θ0i )2 .
i:θ0i =0
(51)
We consider the nonzero means and the zero means separately.
Nonzero means: For θ0i 6= 0, using the Cauchy-Schwartz inequality and the
fact that E0i (Xi − θ0i )2 = 1, we get
Eθ0i (T (Xi ) − θi )2 = Eθ0i (T (Xi ) − Xi + Xi − θi )2
= Eθ0i (T (Xi ) − Xi )2 + Eθ0i (Xi − θ0i )2 + 2Eθ0i (T (Xi ) − Xi )(Xi − θ0i )
q
q
≤ Eθ0i (T (Xi ) − Xi )2 + 1 + 2 Eθ0i (T (Xi ) − Xi )2 Eθ0i (Xi − θ0i )2
37
=
q
Eθ0i (T (Xi ) − Xi )2 + 1
2
.
(52)
We now define
s
1
2 log
.
bn
ζn =
(53)
Let us fix any d > 2 and choose any ρ > 2. Then, using Lemma 2, there
exists a non-negative real-valued function hn (·), depending on d such that
|Tn (x) − x| ≤ hn (x) for all x ∈ R,
(54)
and
lim
sup hn (x) = 0.
(55)
n→∞ |x|>ρζ
n
Using the fact that (T (Xi ) − Xi )2 ≤ Xi2 , together with (55), we obtain
Eθ0i (T (Xi ) − Xi )2 = Eθ0i [(T (Xi ) − Xi )2 1{|Xi | ≤ ρζn }]
+ Eθ0i [T (Xi − Xi )2 1{|Xi | > ρζn }]
!2
≤
ρ2 ζn2
+
sup hn (x)
.
(56)
|x|>ρζn
Using (55) and the fact that ζn → ∞ as n → ∞ by (53), it follows that
!2
= o(ζn2 ) as n → ∞.
sup hn (x)
|x|>ρζn
(57)
By combining (56) and (57), we get
Eθ0i (T (Xi ) − Xi )2 ≤ ρ2 ζn2 (1 + o(1)) as n → ∞.
(58)
Noting that (58) holds uniformly for any i such that θ0i 6= 0, we combine
(52), (53), and (58) to conclude that
X
2
Eθ0i (T (Xi ) − θ0i ) - qen log
i:θ0i 6=0
1
, as n → ∞,
bn
(59)
Zero means: For θ0i = 0, the corresponding MSE can be split as follows:
E0 T (Xi )2 = E0 [T (Xi )2 1{|Xi | ≤ ζn }] + E0 [T (Xi )2 1{|Xi | > ζn }],
38
(60)
where ζn is as in (53). Using Theorem 1, we have
2
E0 [T (Xi ) 1{|Xi | ≤ ζn }] ≤
bn
a + bn + 1/2
2 Z ζn
x2 e x
2 /2
dx
−ζn
b2n ζn 2 x2 /2
x e
dx
a2 −ζn
Z
2b2n ζn 2 x2 /2
x e
dx
= 2
a 0
2b2
2
≤ n (ζn eζn /2 )
as
1
,
- bn log
bn
Z
≤
(61)
where we use the integration by parts for the third inequality.
Now, using the fact that |T (x)| ≤ |x| for all x ∈ R,
2
E0 [T (Xi ) 1{|Xi | > ζn }] ≤ 2
Z ∞
x2 φ(x)dx
ζn
= 2[ζn φ(ζn ) + 1 − Φ(ζn )]
2φ(ζn )
≤ 2ζn φ(ζn ) +
ζn
r
2
2
=
ζn (e−ζn /2 + o(1))
π
s
1
- bn log
,
bn
(62)
d
where we used the identity x2 φ(x) = φ(x) − dx
[xφ(x)] for the first equality
φ(x)
and Mill’s ratio, 1 − Φ(x) ≤ x for all x > 0, in the second inequality.
Combining (61)-(62), we have that
s
X
2
Eθ0i T (Xi ) - (n − qen )bn log
i:θ0i =0
1
.
bn
(63)
From (51), (59) and (63), it immediately follows that
E||T (X) − θ0 ||2 =
n
X
Eθ0i (T (Xi ) − θ0i )2
i=1
1
- qen log
bn
39
s
+ (n − qen )bn log
1
.
bn
The required result now follows by observing that qen ≤ qn and qn = o(n)
and then taking the supremum over all θ0 ∈ `0 [qn ]. This completes the proof
of Theorem 4.
Proof of Theorem 5. Define qen = #{i : θ0i 6= 0}. We decompose the total
variance as
Eθ0
n
X
i=1
V ar(θ0i |Xi ) =
X
Eθ0i V ar(θ0i |Xi ) +
i:θ0i 6=0
X
Eθ0i V ar(θ0i |Xi ), (64)
i:θ0i =0
and consider the nonzero means and zero means separately.
Nonzero means: For θ0i 6= 0, we have
Eθ0i V ar(θ0i |Xi ) ≤ Eθ0i [1 − (T (Xi ) − Xi )2 + Xi2 ]
≤ 2 + Eθ0i (T (Xi ) − Xi )2
1
- log
,
bn
(65)
where the first inequality follows from (49) in Lemma 3 and the last one
follows from (58).
Zero means: For θ0i = 0, we have
E0 V ar(θ0i |Xi ) ≤ E0 [1 − T (Xi )2 + Xi2 ]
≤ 2 + E0 T (Xi )2
s
- bn log
1
,
bn
(66)
where the first inequality follows from (50) in Lemma 3 and the last one
follows from (60)-(62). Combining (64)-(66), we get
1
Eθ0i V ar(θ0i |Xi ) - qen log
bn
s
+ (n − qen )bn
1
log
.
bn
The required result now follows by observing that qen ≤ qn and qn = o(n)
and then taking the supremum over all θ0 ∈ `0 [qn ]. This completes the proof
of Theorem 5.
40
C
Proofs for Section 3.3
Proof of Theorem 7. Using the beta prime representation of the IGG prior,
we have
1
Z ∞
(2π) 2 B(a, b)
0
π(θ) =
1
!
3
θ2
exp −
ub− 2 (1 + u)−a−b du,
2u
where B(a, b) denotes the beta function. Under the transformation of variθ2
ables, z = 2u
, we have
1
2a+ 2
π(θ) =
1
(θ2 )b− 2
1
2
(2π) B(a, b)
Z ∞
1
exp(−z)z a− 2 (θ2 + 2z)−a−b dz.
(67)
0
Now define the set A = {θ : |θ| ≤ }. Then from (67), and for 0 < < 1,
we have
ν(A ) = P (|θ| ≤ )
1
2a+ 2
=
1
R∞
exp(−z)z a− 2
0
R
1
2 2b− 2 (θ 2 + 2z)−b dθ dz
|θ|≤ (θ )
1
(2π) 2 B(a, b)
1
2a+ 2
≥
1
R∞
exp(−z)z a− 2 (2z + 1)−a−b
0
R
|θ|≤ (θ
2 )b− 12 dθ
dz
1
(2π) 2 B(a, b)
1
2a+ 2 2−a−b
≥
R∞
0
1
exp(−z)z a− 2 (1 + z)−a−b
R
2 b− 12
dθ
0 (θ )
dz
1
2
(2π) B(a, b)
=
2b
Z ∞
2b bB(a, b)π 1/2
0
1
exp(−z)z a− 2 (1 + z)−a−b dz.
(68)
To bound the integral term in (68), note that
Z ∞
1
exp(−z)z a− 2 (1 + z)−a−b dz ≥
Z 1
0
1
exp(−z)z a− 2 (1 + z)−a−b dz
0
−1 −a−b
≥e
2
1
a+
2
−1
Therefore, combining (68) and (69), we have
2b
1
e−1 2−a−b a +
1/2
b
2
2 bB(a, b)π
ν(A ) ≥
41
−1
.
(69)
2b Γ(a + b)
e−1 2−a−b
2b Γ(a)Γ(b + 1)π 1/2
2b Γ(a)
−1 −a−2b
e 2
a+
≥
Γ(a)Γ(2)Γ( 12 )
=
b
≥ 2
(π)−1/2 e−1 2−a−2 a +
1
2
a+
1
2
1
2
−1
−1
−1
,
(70)
where we use the fact that 0 < b < 1 for the last two inequalities.
Following Clarke and Barron [11], the optimal rate of convergence comes
from setting n = 1/n, which reflects the ideal case of independent samples
y1 , ..., yn . We therefore apply Proposition 3, substituting in = 1/n and b =
1/n and invoking the lower bound for ν(A ) found in (70). This ultimately
gives us an upper bound on the Cesàro-average risk as
Rn ≤
1
n
−
=
1
n
h
1
n
log
2
1
n
n
π −1/2 e−1 2−a−2
a+
1
2
−1
i
√
2 + log( π) + (a + 2) log(2) + log a + 12 +
2 log n
,
n2
when θ0 = 0.
D
Proofs for Section 4
To establish conditions for (29) to hold, we must first find lower and upper
bounds on the Type I and Type II error probabilities, t1i and t2i respectively,
for rule (28). These error probabilities are given respectively by
t1i = P E(1 − κi |Xi ) >
t2i = P E(1 − κi |Xi ) ≤
1
2
H0i is true ,
1
2
H1i is true .
(71)
To this end, we first prove the following lemmas, Lemma 4-7 which give
upper and lower bounds for t1i and t2i in (71). Our proof methods follow
those of Datta and Ghosh [12], [17], and Ghosh and Chakrabarti [16], except
our arguments rely on control of the sequence of hyperparameters bn , rather
than on specifying a rate or an estimate for a global parameter τ , as in the
global-local (3) framework.
42
Lemma 4. Suppose that X1 , ..., Xn are i.i.d. observations having distribution (22) where the sequence of vectors (ψ 2 , p) satisfies Assumption 1.
Suppose we wish to test (23) using the classification rule (28). Then for all
n, an upper bound for the probability of a Type I error for the ith test is
given by
t1i ≤ √
2bn
a + bn + 1/2
log
π(a + bn + 1/2)
2bn
Proof of Lemma 4. By Theorem 1, the event
the event
−1/2
.
n
E(1 − κi |Xi ) >
1
2
o
implies
bn
1
>
a + bn + 1/2
2
a + bn + 1/2
2
.
⇔ Xi > 2 log
2bn
2
eXi /2
Therefore, noting that under H0i , Xi ∼ N (0, 1) and using Mill’s ratio, i.e.
P (|Z| > x) ≤ 2φ(x)
x , we have
t1i ≤ P
Xi2
a + bn + 1/2
> 2 log
2bn
s
= P |Z| >
r
2φ
≤
2 log
r
2 log
=√
2 log
H0i is true
a + bn + 1/2
2bn
a+bn +1/2
2bn
a+bn +1/2
2bn
a + bn + 1/2
2bn
log
π(a + bn + 1/2)
2bn
−1/2
.
Lemma 5. Suppose that X1 , ..., Xn are i.i.d. observations having distribution (22) where the sequence of vectors (ψ 2 , p) satisfies Assumption 1.
Suppose we wish to test (23) using the classification rule (28). Suppose further that a ∈ ( 12 , ∞) and bn ∈ (0, 1), with bn → 0 as n → ∞. Then for
43
any η ∈ (0, 12 ), δ ∈ (0, 1), and sufficiently large n, a lower bound for the
probability of a Type I error for the ith test is given by
v
u
u
u
t1i ≥ 1 − Φ t
1
2
η)bn
a+
(1 −
2
log
1
η(1 − δ)
bn (ηδ)a+ 2
.
Proof of Lemma 5. By definition, the probability of a Type I error for the
ith decision is given by
1
t1i = P E(1 − κi |Xi ) > H0i is true .
2
We have by Theorem 3 that
a+
E(κi |Xi ) ≤ η +
1
2
(1 − η)bn
1
bn (ηδ)a+ 2
η(1 − δ) 2
exp −
Xi ,
2
η(1 − δ) 2
1
Xi < − η .
2
2
and so it follows that
E(1 − κi |Xi ) >
1
2
⊇
a + 1 (1 − η)bn
2
a+ 21
exp −
bn (ηδ)
Thus, using the definition of t1i and the above and noting that under H0i ,
Xi ∼ N (0, 1), as n → ∞,
a+
t1i ≥ P
1
2
(1 − η)bn
1
bn (ηδ)a+ 2
1
η(1 − δ) 2
Xi < − η H0i is true
exp −
2
2
a + 12 (1 − η)bn
2
2
log
= P Xi >
1
η(1 − δ)
bn (ηδ)a+ 2 1 − η
2
v
u
u
u
= 2P Z > t
1
2
a+
1
2
a+
(1 − η)bn
2
log
1
η(1 − δ)
bn (ηδ)a+ 2 12 − η
v
u
u
u
= 2 1 − Φ t
,
(1 − η)bn
2
log
1
η(1 − δ)
bn (ηδ)a+ 2 12 − η
where for the last inequality, we used the fact that bn → 0 as n → ∞, and
the fact that η, ηδ ∈ (0, 21 ), so that the log(·) term in final equality is greater
than zero for sufficiently large n.
44
Lemma 6. Suppose we have the same set-up as Lemma 4. Assume further
1/4
that bn → 0 in such a way that limn→∞ bpnn ∈ (0, ∞). Then for any η ∈
(0, 12 ), δ ∈ (0, 1), and sufficiently large n, an upper bound for the probability
of a Type II error for the ith test is given by
s
"
t2i ≤ 2Φ
C
2η(1 − δ)
!
#
− 1 (1 + o(1)),
where the o(1) terms tend to zero as n → ∞.
Proof of Lemma 6. By definition, the probability of a Type II error is given
by
1
t2i = P E(1 − κi ) ≤ H1i is true .
2
Fix η ∈ (0, 12 ) and δ ∈ (0, 1). Using the inequality,
κi ≤ 1 {η < κi ≤ 1} + η,
we obtain
E(κi |Xi ) ≤ P (κi > η|Xi ) + η.
Coupled with Theorem 3, we obtain that for sufficiently large n,
1
E(κi |Xi ) >
2
⊆
a + 1 (1 − η)bn
2
1
bn (ηδ)a+ 2
η(1 − δ) 2
1
exp −
Xi > − η .
2
2
Therefore,
1
t2i = P E(κi |Xi ) > H1i is true
2
a+
≤P
1
2
(1 − η)bn
1
b(ηδ)an + 2
Xi2
=P
2
<
η(1 − δ)
log
=P
Xi2
1
2
η(1 − δ) 2
1
exp −
Xi > − η H1i is true
2
2
(
log
a + 12
bn (ηδ)a
− η (ηδ)1/2
(1 − η)bn
!
−
H is true
1i
a
2
<
log
(1 + o(1)) H1i is true ,
η(1 − δ)
bn (ηδ)a
45
(72)
where in the final equality, we used the fact that bn → 0 as n → ∞, so the
second log(·) term in the second to last equality is a bounded quantity.
Note that under H1i , Xi ∼ N (0, 1 + ψ 2 ). Therefore, by (72) and the fact
ψn2
that lim
= 1 (by the second condition of Assumption 1), we have
n→∞ 1 + ψ 2
n
s
t2i ≤ P |Z| <
2
η(1 − δ)
s
log(a(ηδ)−a b−1
n )
(1 + o(1)) as n → ∞. (73)
2
ψ
1/4
7/8
By assumption, limn→∞ bpnn ∈ (0, ∞). This then implies that limn→∞ bpn2 =
n
0. Therefore, by the fourth condition of Assumption 1 and the fact that
ψ 2 → ∞ as n → ∞, we have
log(a(ηδ)−a b−1
log(a(ηδ)−a ) + log(b−1
n )
n )
=
2
2
ψ
ψ
−1/8
=
log(bn
ψ2
)
−7/8
log(bn
+
ψ2
)
!
(1 + o(1))
−1/2
log(bn )
(1 + o(1))
4ψ 2
C
→
as n → ∞.
4
=
(74)
Thus, using (73) and (74), we have
s
t2i ≤ P
|Z| <
s
|Z| <
=P
"
=2 Φ
s
!
C
(1 + o(1))
2η(1 − δ)
as n → ∞
!
C
(1 + o(1)) as n → ∞
2η(1 − δ)
C
2η(1 − δ)
!
#
− 1 (1 + o(1)) as n → ∞.
Lemma 7. Suppose we have the same set-up as Lemma 4. Then a lower
bound for the probability of a Type II error for the ith test is given by
h
i
√
t2i ≥ 2Φ( C) − 1 (1 + o(1)) as n → ∞,
where the o(1) terms tend to zero as n → ∞.
46
Proof of Lemma 7. By definition, the probability of a Type II error for the
ith decision is given by
t2i = P E(1 − κi ) ≤
1
H1i is true
2
.
For any n, we have by Theorem 1 that
2
eXi /2
bn
a + bn + 1/2
≤
1
2
1
.
2
⊆ E(1 − κi |Xi ) ≤
Therefore,
1
t2i = P E(1 − κi |Xi ) ≤ H1i is true
2
bn
1
2
≥ P eXi /2
≤ H1i is true
a + bn + 1/2
2
a + bn + 1/2
2
= P Xi ≤ 2 log
H1i is true .
2bn
(75)
Since Xi ∼ N (0, 1 + ψ 2 ) under H1i , we have by the second condition in
ψn2
Assumption 1 that lim
→ 1. From (75) and the facts that a ∈
n→∞ 1 + ψ 2
n
−1/2
( 12 , ∞) and bn ∈ (0, 1) for all n (so b−1
for all n), we have for
n ≥ bn
sufficiently large n,
t2i ≥ P |Z| ≤
≥ P |Z| ≤
≥ P |Z| ≤
v
u
u 2 log a+bn +1/2
t
2bn
ψ2
v
u
u log 1
t
2bn
ψ2
(1 + o(1)) as n → ∞
(1 + o(1)) as n → ∞
v
u
u log(b−1/2 ) + log(1/2)
n
t
ψ2
(1 + o(1)) as n → ∞
√
= P (|Z| ≤ C)(1 + o(1)) as n → ∞
√
= 2[Φ( C) − 1](1 + o(1)) as n → ∞,
1/4
where in the second to last equality, we used the assumption that limn→∞
(0, ∞) and the second and fourth conditions from Assumption 1.
47
bn
pn
∈
Proof of Theorem 8. Since the κi ’s, i = 1, ..., n are a posteriori independent,
the Type I and Type II error probabilities t1i and t2i are the same for every
test i, i = 1, ..., n. By Lemmas 4 and 5, for large enough n,
v
u
u
u
2 1 − Φ t
≤√
2
η(1 − δ)
a+
(1 − η)
log
≤ t1i
a+ 1 1
1
2
bn (ηδ)
bn
2
2
2bn
a + bn + 1/2
log
π(a + bn + 1/2)
2bn
−η
−1/2
.
Taking the limit as n → ∞ of all the terms above and using the sandwich
theorem, we have
lim t1i = 0
(76)
n→∞
for the ith test, under the assumptions on the hyperparameters a and bn .
By Lemmas 6 and 7, for any η ∈ (0, 21 ) and δ ∈ (0, 1),
h
i
√
2Φ( C) − 1 (1 + o(1)) ≤ t2i ≤ 2Φ
"
s
C
2η(1 − δ)
!
#
− 1 (1 + o(1)). (77)
Therefore, we have by (76) and (77) that as n → ∞, the asymptotic risk
(24) of the classification rule (28), RIGG , can be bounded as follows:
√
np(2Φ( C) − 1)(1 + o(1)) ≤ RIGG ≤ np 2Φ
s
C
2η(1 − δ)
!
!
− 1 (1 + o(1).
(78)
Therefore, from (26) and (78), we have as n → ∞,
RIGG
RIGG
≤ lim sup BO ≤
n→∞ RBO
n→∞ ROpt
Opt
2Φ
1 ≤ lim inf
q
C
2η(1−δ)
−1
√
.
2Φ( C) − 1
(79)
Now, the supremum of η(1 − δ) over the grid (η, δ) ∈ (0, 21 ) × (0, 1) is clearly
1
2 , and so the√infimum of the numerator in the right-most term in (79) is
therefore 2Φ( C) − 1. Thus, one has that
1 ≤ lim inf
n→∞
RIGG
RIGG
≤ lim sup BO ≤ 1,
BO
ROpt
n→∞ ROpt
so classification rule (28) is ABOS, i.e.
RIGG
→ 1 as n → ∞.
BO
ROpt
48
References
[1] Armagan, A., Clyde, M., and Dunson, D. B. (2011). Generalized beta
mixtures of gaussians. In Shawe-taylor, J., Zemel, R., Bartlett, P., Pereira,
F., and Weinberger, K., editors, Advances in Neural Information Processing Systems 24, pages 523–531.
[2] Armagan, A., Dunson, D. B., and Lee, J. (2013). Generalized double
pareto shrinkage. Statistica Sinica, 23 1:119–143.
[3] Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery
rate: A practical and powerful approach to multiple testing. Journal of
the Royal Statistical Society. Series B (Methodological), 57(1):289–300.
[4] Berger, J. (1980). A robust generalized bayes estimator and confidence
region for a multivariate normal mean. Ann. Statist., 8(4):716–761.
[5] Bhadra, A., Datta, J., Polson, N. G., and Willard, B. (2017). The
horseshoe+ estimator of ultra-sparse signals. Bayesian Anal., 12(4):1105–
1131.
[6] Bhattacharya, A., Pati, D., Pillai, N. S., and Dunson, D. B. (2015).
Dirichlet–laplace priors for optimal shrinkage. Journal of the American
Statistical Association, 110(512):1479–1490. PMID: 27019543.
[7] Bogdan, M., Chakrabarti, A., Frommlet, F., and Ghosh, J. K. (2011).
Asymptotic bayes-optimality under sparsity of some multiple testing procedures. Ann. Statist., 39(3):1551–1579.
[8] Carvalho, C. M., Polson, N. G., and Scott, J. G. (2009). Handling
sparsity via the horseshoe. In van Dyk, D. and Welling, M., editors, Proceedings of the Twelth International Conference on Artificial Intelligence
and Statistics, volume 5 of Proceedings of Machine Learning Research,
pages 73–80, Hilton Clearwater Beach Resort, Clearwater Beach, Florida
USA. PMLR.
[9] Carvalho, C. M., Polson, N. G., and Scott, J. G. (2010). The horseshoe
estimator for sparse signals. Biometrika, 97(2):465–480.
[10] Castillo, I. and van der Vaart, A. (2012). Needles and straw in a
haystack: Posterior concentration for possibly sparse sequences. Ann.
Statist., 40(4):2069–2101.
49
[11] Clarke, B. S. and Barron, A. R. (1990). Information-theoretic asymptotics of bayes methods. IEEE Transactions on Information Theory,
36:453–471.
[12] Datta, J. and Ghosh, J. K. (2013). Asymptotic properties of bayes risk
for the horseshoe prior. Bayesian Anal., 8(1):111–132.
[13] Donoho, D. L., Johnstone, I. M., Hoch, J. C., and Stern, A. S. (1992).
Maximum entropy and the nearly black object. Journal of the Royal
Statistical Society. Series B (Methodological), 54(1):41–81.
[14] Efron, B. (2010).
25(2):145–157.
The future of indirect evidence.
Statist. Sci.,
[15] Ghosal, S., Ghosh, J. K., and van der Vaart, A. W. (2000). Convergence
rates of posterior distributions. Ann. Statist., 28(2):500–531.
[16] Ghosh, P. and Chakrabarti, A. (2017). Asymptotic optimality of onegroup shrinkage priors in sparse high-dimensional problems. Bayesian
Anal., 12(4):1133–1161.
[17] Ghosh, P., Tang, X., Ghosh, M., and Chakrabarti, A. (2016). Asymptotic properties of bayes risk of a general class of shrinkage priors in multiple hypothesis testing under sparsity. Bayesian Anal., 11(3):753–796.
[18] Griffin, J. E. and Brown, P. J. (2010). Inference with normal-gamma
prior distributions in regression problems. Bayesian Anal., 5(1):171–188.
[19] Griffin, J. E. and Brown, P. J. (2013). Some priors for sparse regression
modelling. Bayesian Anal., 8(3):691–702.
[20] Johnstone, I. M. and Silverman, B. W. (2004). Needles and straw in
haystacks: Empirical bayes estimates of possibly sparse sequences. Ann.
Statist., 32(4):1594–1649.
[21] Park, T. and Casella, G. (2008). The bayesian lasso. Journal of the
American Statistical Association, 103(482):681–686.
[22] Ročková, V. (2017). Bayesian estimation of sparse signals with a continuous spike-and-slab prior. Ann. Statist. To appear.
[23] Singh, D., Febbo, P. G., Ross, K., Jackson, D. G., Manola, J., Ladd,
C., Tamayo, P., Renshaw, A. A., D’Amico, A. V., Richie, J. P., Lander,
E. S., Loda, M., Kantoff, P. W., Golub, T. R., and Sellers, W. R. (2002).
50
Gene expression correlates of clinical prostate cancer behavior. Cancer
Cell, 1(2):203 – 209.
[24] Strawderman, W. E. (1971). Proper bayes minimax estimators of the
multivariate normal mean. Ann. Math. Statist., 42(1):385–388.
[25] van der Pas, S., Salomond, J.-B., and Schmidt-Hieber, J. (2016). Conditions for posterior contraction in the sparse normal means problem.
Electron. J. Statist., 10(1):976–1000.
[26] van der Pas, S., Szabó, B., and van der Vaart, A. (2017). Adaptive posterior contraction rates for the horseshoe. Electron. J. Statist., 11(2):3196–
3225.
[27] van der Pas, S. L., Kleijn, B. J. K., and van der Vaart, A. W. (2014).
The horseshoe estimator: Posterior concentration around nearly black
vectors. Electron. J. Statist., 8(2):2585–2618.
[28] Wellcome Trust (2007). Genome-wide association study of 14,000 cases
of seven common diseases and 3000 shared controls. Nature, 447:661–678.
51
| 10math.ST
|
arXiv:1707.07012v3 [cs.CV] 1 Dec 2017
Learning Transferable Architectures for Scalable Image Recognition
Barret Zoph
Google Brain
Vijay Vasudevan
Google Brain
Jonathon Shlens
Google Brain
Quoc V. Le
Google Brain
[email protected]
[email protected]
[email protected]
[email protected]
Abstract
cation represents one of the most important breakthroughs
in deep learning. Successive advancements on this benchmark based on convolutional neural networks (CNNs) have
achieved impressive results through significant architecture
engineering [52, 58, 20, 59, 57, 67].
Developing neural network image classification models
often requires significant architecture engineering. In this
paper, we attempt to automate this engineering process by
learning the model architectures directly on the dataset of
interest. As this approach is expensive when the dataset is
large, we propose to search for an architectural building
block on a small dataset and then transfer the block to a
larger dataset. Our key contribution is the design of a new
search space which enables transferability. In our experiments, we search for the best convolutional layer (or “cell”)
on the CIFAR-10 dataset and then apply this cell to the
ImageNet dataset by stacking together more copies of this
cell, each with their own parameters. Although the cell is
not searched for directly on ImageNet, an architecture constructed from the best cell achieves, among the published
works, state-of-the-art accuracy of 82.7% top-1 and 96.2%
top-5 on ImageNet. Our model is 1.2% better in top-1 accuracy than the best human-invented architectures while having 9 billion fewer FLOPS – a reduction of 28% in computational demand from the previous state-of-the-art model.
When evaluated at different levels of computational cost,
accuracies of our models exceed those of the state-of-theart human-designed models. For instance, a smaller network constructed from the best cell also achieves 74% top1 accuracy, which is 3.1% better than equivalently-sized,
state-of-the-art models for mobile platforms. On CIFAR10, an architecture constructed from the best cell achieves
2.4% error rate, which is also state-of-the-art. Finally, the
image features learned from image classification can also
be transferred to other computer vision problems. On the
task of object detection, the learned features used with the
Faster-RCNN framework surpass state-of-the-art by 4.0%
achieving 43.1% mAP on the COCO dataset.
In this paper, we consider learning the convolutional architectures directly from data with application to ImageNet
classification. In addition to being an difficult and important benchmark in computer vision, features derived from
ImageNet classifiers are of great importance to many other
computer vision tasks. For example, features from networks
that perform well on ImageNet classification provide stateof-the-art performance when transferred to other computer
vision tasks where labeled data is limited [13].
1. Introduction
Our approach is inspired by the recently proposed Neural Architecture Search (NAS) framework [70], which uses
a policy gradient algorithm to optimize architecture configurations. Even though NAS is attractive as a method to
search for good convolutional network architectures, applying it directly to the ImageNet dataset is computationally
expensive given the size of the dataset. We therefore propose to search for a good architecture on the far smaller
CIFAR-10 dataset, and automatically transfer the learned
architecture to ImageNet. We achieve this transferrability by designing a search space so that the complexity of
the architecture is independent of the depth of the network
and the size of input images. More concretely, all convolutional networks in our search space are composed of convolutional layers (or “cells”) with identical structure but different weights. Searching for the best convolutional architectures is therefore reduced to searching for the best cell
structure. Searching for the best cell structure has two main
benefits: it is much faster than searching for an entire network architecture and the cell itself is more likely to generalize to other problems. In our experiments, this approach
significantly accelerates the search for the best architectures
using CIFAR-10 by a factor of 7× and learns architectures
that successfully transfer to ImageNet.
ImageNet classification [11] is an important benchmark
in computer vision. The seminal work of [32] on using
convolutional architectures [17, 33] for ImageNet classifi-
Our main result is that the best architecture found on
CIFAR-10 achieves state-of-the-art accuracy when transferred to ImageNet classification without much modifica1
tion. On ImageNet, an architecture constructed from the
best cell achieves, among the published works, state-ofthe-art accuracy of 82.7% top-1 and 96.2% top-5. This
result amounts to a 1.2% improvement in top-1 accuracy
than the best human-invented architectures while having 9
billion fewer FLOPS. On CIFAR-10 itself, the architecture
achieves 2.4% error rate, which is also state-of-the-art.
Additionally, by simply varying the number of the convolutional cells and number of filters in the convolutional
cells, we can create convolutional architectures with different computational demands. Thanks to this property of the
cells, we can generate a family of models that achieve accuracies superior to all human-invented models at equivalent or smaller computational budgets [59, 29]. Notably, the
smallest version of the learned model achieves 74.0% top1 accuracy on ImageNet, which is 3.1% better than previously engineered architectures targeted towards mobile and
embedded vision tasks [24, 69].
Finally, we show that the image features learned from
image classification are generically useful and transfer to
other computer vision problems. In our experiments, the
features learned from ImageNet classification can be combined with the Faster-RCNN framework [46] to achieve
state-of-the-art on COCO object detection task for both the
largest as well as mobile-optimized models. Our largest
model achieves 43.1% mAP, which is 4% better than previous state-of-the-art.
2. Method
Our work makes use of search methods to find good convolutional architectures on a dataset of interest. The main
search method we use in this work is the Neural Architecture Search (NAS) framework proposed by [70]. In NAS,
a controller recurrent neural network (RNN) samples child
networks with different architectures. The child networks
are trained to convergence to obtain some accuracy on a
held-out validation set. The resulting accuracies are used
to update the controller so that the controller will generate
better architectures over time. The controller weights are
updated with policy gradient (see Figure 1).
The main contribution of this work is the design of a
novel search space, such that the best architecture found
on the CIFAR-10 dataset would scale to larger, higherresolution image datasets across a range of computational
settings. One inspiration for this search space is the recognition that architecture engineering with CNNs often identifies repeated motifs consisting of combinations of convolutional filter banks, nonlinearities and a prudent selection of
connections to achieve state-of-the-art results (such as the
repeated modules present in the Inception and ResNet models [58, 20, 59, 57]). These observations suggest that it may
be possible for the controller RNN to predict a generic convolutional cell expressed in terms of these motifs. This cell
Sample architecture A!
with probability p
Train a child network!
with architecture A to !
convergence to get !
validation accuracy R
The controller (RNN)
Scale gradient of p by R!
to update the controller
Figure 1. Overview of Neural Architecture Search [70]. A controller RNN predicts architecture A from a search space with probability p. A child network with architecture A is trained to convergence achieving accuracy R. Scale the gradients of p by R to
update the RNN controller.
can then be stacked in series to handle inputs of arbitrary
spatial dimensions and filter depth.
In our approach, the overall architectures of the convolutional nets are manually predetermined. They are composed of convolutional cells repeated many times where
each convolutional cell has the same architecture, but different weights. To easily build scalable architectures for
images of any size, we need two types of convolutional cells
to serve two main functions when taking in a feature map
as input: (1) convolutional cells that return a feature map of
the same dimension, and (2) convolutional cells that return
a feature map where the feature map height and width is reduced by a factor of two. We name the first type and second
type of convolutional cells Normal Cell and Reduction Cell
respectively. For the Reduction Cell, we make the initial
operation applied to the cell’s inputs have a stride of two to
reduce the height and width. All of our operations that we
consider for building our convolutional cells have an option
of striding.
Figure 2 shows our placement of Normal and Reduction
Cells for CIFAR-10 and ImageNet. Note on ImageNet we
have more Reduction Cells, since the incoming image size
is 299x299 compared to 32x32 for CIFAR. The Reduction
and Normal Cell could have the same architecture, but we
empirically found it beneficial to learn two separate architectures. We use a common heuristic to double the number
of filters in the output whenever the spatial activation size is
reduced in order to maintain roughly constant hidden state
dimension [32, 52]. Importantly, much like Inception and
ResNet models [58, 20, 59, 57], we consider the number of
motif repetitions N and the number of initial convolutional
filters as free parameters that we tailor to the scale of an
image classification problem.
What varies in the convolutional nets is the structures of
the Normal and Reduction Cells, which are searched by the
controller RNN. The structures of the cells can be searched
provides good results, although we have not exhaustively
searched this space due to computational limitations.
In steps 3 and 4, the controller RNN selects an operation
to apply to the hidden states. We collected the following set
of operations based on their prevalence in the CNN literature:
•
•
•
•
•
•
•
Figure 2. Scalable architectures for image classification consist of
two repeated motifs termed Normal Cell and Reduction Cell. This
diagram highlights the model architecture for CIFAR-10 and ImageNet. The choice for the number of times the Normal Cells that
gets stacked between reduction cells, N , can vary in our experiments.
within a search space defined as follows. In our search
space, each cell receives as input two initial hidden states hi
and hi−1 which are the outputs of two cells in previous two
lower layers or the input image. The controller RNN recursively predicts the rest of the structure of the convolutional
cell, given these two initial hidden states (Figure 3). The
predictions of the controller for each cell are grouped into
B blocks, where each block has 5 prediction steps made
by 5 distinct softmax classifiers corresponding to discrete
choices of the elements of a block:
Step 1. Select a hidden state from hi , hi−1 or from the set of hidden
states created in previous blocks.
Step 2. Select a second hidden state from the same options as in Step 1.
Step 3. Select an operation to apply to the hidden state selected in Step 1.
Step 4. Select an operation to apply to the hidden state selected in Step 2.
Step 5. Select a method to combine the outputs of Step 3 and 4 to create
a new hidden state.
The algorithm appends the newly-created hidden state to
the set of existing hidden states as a potential input in subsequent blocks. The controller RNN repeats the above 5
prediction steps B times corresponding to the B blocks in
a convolutional cell. In our experiments, selecting B = 5
identity
1x7 then 7x1 convolution
3x3 average pooling
5x5 max pooling
1x1 convolution
3x3 depthwise-separable conv
7x7 depthwise-separable conv
•
•
•
•
•
•
1x3 then 3x1 convolution
3x3 dilated convolution
3x3 max pooling
7x7 max pooling
3x3 convolution
5x5 depthwise-seperable conv
In step 5 the controller RNN selects a method to combine
the two hidden states, either (1) element-wise addition between two hidden states or (2) concatenation between two
hidden states along the filter dimension. Finally, all of the
unused hidden states generated in the convolutional cell are
concatenated together in depth to provide the final cell output.
To allow the controller RNN to predict both Normal Cell
and Reduction Cell, we simply make the controller have
2 × 5B predictions in total, where the first 5B predictions
are for the Normal Cell and the second 5B predictions are
for the Reduction Cell.
Finally, our work makes use of the reinforcement learning proposal in NAS [70] intensively; however, it is also
possible to use random search to search for models in our
search space. In random search, instead of sampling the decisions from the softmax classifiers in the controller RNN,
we can sample the decisions from the uniform distribution.
In our experiments, we find that random search is worse
than reinforcement learning on the CIFAR-10 dataset. Although there is value in using reinforcement learning, the
gap is smaller than what is found in the original work
of [70]. This result suggests that the new search space is
well designed such that random search can perform reasonably well. We will compare reinforcement learning against
random search in Section 3.4.
3. Experiments and Results
In this section, we describe our experiments with our
method described above to learn convolutional cells. In
summary, all architecture searches are performed using the
CIFAR-10 classification task [31]. The controller RNN was
trained using Proximal Policy Optimization (PPO) [50] by
employing a global workqueue system for generating a pool
of child networks controlled by the RNN. In our experiments, the pool of workers in the workqueue consisted of
500 GPUs. Please see Appendix A for complete details of
the architecture learning algorithm and the controller system.
softmax!
layer
Select second!
hidden state
Select operation for !
first hidden state
Select operation for!
second hidden state
controller!
hidden layer
Select one!
hidden state
Select method to!
combine hidden state
new hidden layer
add
repeat B times
3 x 3 conv
2 x 2 maxpool
hidden layer A
hidden layer B
Figure 3. Controller model architecture for recursively constructing one block of a convolutional cell. Each block requires selecting 5
discrete parameters, each of which corresponds to the output of a softmax layer. Example constructed block shown on right. A convolutional cell contains B blocks, hence the controller contains 5B softmax layers for predicting the architecture of a convolutional cell. In our
experiments, the number of blocks B is 5.
The result of this search process over 4 days yields several candidate convolutional cells. We note that this search
procedure is almost 7× faster than previous approaches [70]
that took 28 days.1 Additionally, we demonstrate below that
the resulting architecture is superior in accuracy.
Figure 4 shows a diagram of the top performing Normal
Cell and Reduction Cell. Note the prevalence of separable
convolutions and the number of branches compared with
competing architectures [52, 58, 20, 59, 57]. Subsequent
experiments focus on this convolutional cell architecture,
although we examine the efficacy of other, top-ranked convolutional cells in ImageNet experiments (described in Appendix B) and report their results as well. We call the three
networks constructed from the best three searches NASNetA, NASNet-B and NASNet-C.
We demonstrate the utility of the convolutional cells by
employing this learned architecture on CIFAR-10 and a
family of ImageNet classification tasks. The latter family of
tasks is explored across a few orders of magnitude in computational budget. After having learned the convolutional
cells, several hyper-parameters may be explored to build a
final network for a given task: (1) the number of cell repeats
N and (2) the number of filters in the initial convolutional
cell. After selecting the number of initial filters, we use a
common heuristic to double the number of filters whenever
the stride is 2. Finally, we define a simple notation, e.g.,
4 @ 64, to indicate these two parameters in all networks,
where 4 and 64 indicate the number of cell repeats and the
number of filters in the penultimate layer of the network,
respectively.
1 In particular, we note that previous architecture search [70] used 800
GPUs for 28 days resulting in 22,400 GPU-hours. The method in this paper uses 500 GPUs across 4 days resulting in 2,000 GPU-hours. The former effort used Nvidia K40 GPUs, whereas the current efforts used faster
NVidia P100s. Discounting the fact that the we use faster hardware, we
estimate that the current procedure is roughly about 7× more efficient.
3.1. Results on CIFAR-10 Image Classification
For the task of image classification with CIFAR-10, we
set N = 4 or 6 (Figure 2). The test accuracies of the
best architectures are reported in Table 1 along with other
state-of-the-art models. As can be seen from the Table, a
large NASNet-A model with cutout data augmentation [12]
achieves a state-of-the-art error rate of 2.40% (averaged
across 5 runs), which is slightly better than the previous
best record of 2.56% by [12]. The best single run from our
model achieves 2.19% error rate.
3.2. Results on ImageNet Image Classification
We performed several sets of experiments on ImageNet
with the best convolutional cells learned from CIFAR-10.
We emphasize that we merely transfer the architectures
from CIFAR-10 but train all ImageNet models weights from
scratch.
Results are summarized in Table 2 and 3 and Figure 5.
In the first set of experiments, we train several image classification systems operating on 299x299 or 331x331 resolution images with different experiments scaled in computational demand to create models that are roughly on par
in computational cost with Inception-v2 [29], Inception-v3
[59] and PolyNet [68]. We show that this family of models achieve state-of-the-art performance with fewer floating
point operations and parameters than comparable architectures. Second, we demonstrate that by adjusting the scale
of the model we can achieve state-of-the-art performance
at smaller computational budgets, exceeding streamlined
CNNs hand-designed for this operating regime [24, 69].
Note we do not have residual connections between convolutional cells as the models learn skip connections on
their own. We empirically found manually inserting residual connections between cells to not help performance. Our
training setup on ImageNet is similar to [59], but please see
Appendix A for details.
Table 2 shows that the convolutional cells discovered with CIFAR-10 generalize well to ImageNet prob-
hi+1
concat
hi+1
add
add
concat
max!
3x3
add
sep!
3x3
add
iden!
tity
sep!
3x3
add
sep!
5x5
iden!
tity
avg!
3x3
add
avg!
3x3
sep!
3x3
add
avg!
3x3
sep!
5x5
avg!
3x3
iden!
tity
add
sep!
3x3
sep!
7x7
add
sep!
5x5
max!
3x3
hi
hi
...
...
hi-1
hi-1
Normal Cell
sep!
7x7
add
avg!
3x3
sep!
5x5
Reduction Cell
Figure 4. Architecture of the best convolutional cells (NASNet-A) with B = 5 blocks identified with CIFAR-10 . The input (white) is the
hidden state from previous activations (or input image). The output (pink) is the result of a concatenation operation across all resulting
branches. Each convolutional cell is the result of B blocks. A single block is corresponds to two primitive operations (yellow) and a
combination operation (green). Note that colors correspond to operations in Figure 3.
model
depth
# params
error rate (%)
DenseNet (L = 40, k = 12) [26]
DenseNet(L = 100, k = 12) [26]
DenseNet (L = 100, k = 24) [26]
DenseNet-BC (L = 100, k = 40) [26]
40
100
100
190
1.0M
7.0M
27.2M
25.6M
5.24
4.10
3.74
3.46
Shake-Shake 26 2x32d [18]
Shake-Shake 26 2x96d [18]
Shake-Shake 26 2x96d + cutout [12]
26
26
26
2.9M
26.2M
26.2M
3.55
2.86
2.56
NAS v3 [70]
NAS v3 [70]
39
39
7.1M
37.4M
4.47
3.65
-
3.3M
3.3M
27.6M
27.6M
2.6M
3.1M
3.41
2.65
2.97
2.40
3.73
3.59
NASNet-A
NASNet-A
NASNet-A
NASNet-A
NASNet-B
NASNet-C
(6 @ 768)
(6 @ 768) + cutout
(7 @ 2304)
(7 @ 2304) + cutout
(4 @ 1152)
(4 @ 640)
Table 1. Performance of Neural Architecture Search and other state-of-the-art models on CIFAR-10. All results for NASNet are the mean
accuracy across 5 runs.
lems. In particular, each model based on the convolutional cells exceeds the predictive performance of the corresponding hand-designed model. Importantly, the largest
model achieves a new state-of-the-art performance for ImageNet (82.7%) based on single, non-ensembled predictions,
surpassing previous best published result by ∼1.2% [8].
Among the unpublished works, our model is on par with
the best reported result of 82.7% [25], while having significantly fewer floating point operations. Figure 5 shows a
complete summary of our results in comparison with other
published results. Note the family of models based on convolutional cells provides an envelope over a broad class of
human-invented architectures.
Finally, we test how well the best convolutional cells
may perform in a resource-constrained setting, e.g., mobile
devices (Table 3). In these settings, the number of floating point operations is severely constrained and predictive
performance must be weighed against latency requirements
on a device with limited computational resources. MobileNet [24] and ShuffleNet [69] provide state-of-the-art re-
85
85
NASNet-A (6 @ 4032)
NASNet-A (6 @ 4032)
80
PolyNet
Inception-ResNet-v2
NASNet-A (5 @ 1538)
Inception-v4
ResNet-152
Xception
DPN-131
ResNeXt-101
NASNet-A (7 @ 1920)
SENet
accuracy (precision @1)
accuracy (precision @1)
NASNet-A (7 @ 1920)
Inception-v3
75
Inception-v2
NASNet-A (4 @ 1056)
VGG-16
ShuffleNet
MobileNet
Inception-v1
70
80
DPN-131
Inception-ResNet-v2
Inception-v3
SENet
PolyNet
ResNeXt-101
NASNet-A (5 @ 1538) Inception-v4
Xception
ResNet-152
Inception-v2
75
NASNet-A (4 @ 1056)
VGG-16
ShuffleNet
MobileNet
70
Inception-v1
65
65
0
10000
20000
30000
0
40000
40
20
# Mult-Add operations (millions)
100
80
60
120
140
# parameters (millions)
Figure 5. Accuracy versus computational demand (left) and number of parameters (right) across top performing published CNN architectures on ImageNet 2012 ILSVRC challenge prediction task. Computational demand is measured in the number of floating-point multiplyadd operations to process a single image. Black circles indicate previously published results and red squares highlight our proposed
models.
Model
image size
# parameters
Mult-Adds
Top 1 Acc. (%)
Top 5 Acc. (%)
Inception V2 [29]
NASNet-A (5 @ 1538)
224×224
299×299
11.2 M
10.9 M
1.94 B
2.35 B
74.8
78.6
92.2
94.2
Inception V3 [59]
Xception [9]
Inception ResNet V2 [57]
NASNet-A (7 @ 1920)
299×299
299×299
299×299
299×299
23.8 M
22.8 M
55.8 M
22.6 M
5.72 B
8.38 B
13.2 B
4.93 B
78.0
79.0
80.4
80.8
93.9
94.5
95.3
95.3
ResNeXt-101 (64 x 4d) [67]
PolyNet [68]
DPN-131 [8]
SENet [25]
NASNet-A (6 @ 4032)
320×320
331×331
320×320
320×320
331×331
83.6 M
92 M
79.5 M
145.8 M
88.9 M
31.5 B
34.7 B
32.0 B
42.3 B
23.8 B
80.9
81.3
81.5
82.7
82.7
95.6
95.8
95.8
96.2
96.2
Table 2. Performance of architecture search and other published state-of-the-art models on ImageNet classification. Mult-Adds indicate
the number of composite multiply-accumulate operations for a single image. Note that the composite multiple-accumulate operations are
calculated for the image size reported in the table. Model size for [25] calculated from open-source implementation.
Model
# parameters
Mult-Adds
Top 1 Acc. (%)
Top 5 Acc. (%)
Inception V1 [58]
MobileNet-224 [24]
ShuffleNet (2x) [69]
6.6M
4.2 M
∼ 5M
1,448 M
569 M
524 M
69.8
70.6
70.9
89.9
89.5
89.8
NASNet-A (4 @ 1056)
NASNet-B (4 @ 1536)
NASNet-C (3 @ 960)
5.3 M
5.3M
4.9M
564 M
488 M
558 M
74.0
72.8
72.5
91.6
91.3
91.0
Table 3. Performance on ImageNet classification on a subset of models operating in a constrained computational setting, i.e., < 1.5 B
multiply-accumulate operations per image. All models use 224x224 images.
sults obtaining 70.6% and 70.9% accuracy, respectively on
224x224 images using ∼550M multliply-add operations.
An architecture constructed from the best convolutional
cells achieves superior predictive performance (74.0% ac-
curacy) surpassing previous models but with comparable
computational demand. In summary, we find that the
learned convolutional cells are flexible across model scales
achieving state-of-the-art performance across almost 2 or-
ders of magnitude in computational budget.
3.4. Efficiency of architecture search methods
3.3. Improved features for object detection
2 A primary advance in the best reported object detection system is the
introduction of a novel loss [36]. Pairing this loss with NASNet-A image
featurization may lead to even further performance gains. Additionally,
performance gains are achievable through ensembling multiple inferences
across multiple model instances and image crops (e.g., [28]).
0.930
0.925
Accuracy at 20 Epochs
Image classification networks provide generic image features that may be transferred to other computer vision problems [13]. One of the most important problems is the spatial localization of objects within an image. To further
validate the performance of the family of NASNet-A networks, we test whether object detection systems derived
from NASNet-A lead to improvements in object detection
[28].
To address this question, we plug in the family of
NASNet-A networks pretrained on ImageNet into the
Faster-RCNN object detection pipeline [46] using an opensource software platform [28]. We retrain the resulting object detection pipeline on the combined COCO training plus
validation dataset excluding 8,000 mini-validation images.
We perform single model evaluation using 300-500 RPN
proposals per image. In other words, we only pass a single image through a single network. We evaluate the model
on the COCO mini-val [28] and test-dev dataset and report
the mean average precision (mAP) as computed with the
standard COCO metric library [37]. We perform a simple
search over learning rate schedules to identify the best possible model. Finally, we examine the behavior of two object
detection systems employing the best performing NASNetA image featurization (NASNet-A, 6 @ 4032) as well as
the image featurization geared towards mobile platforms
(NASNet-A, 4 @ 1056).
For the mobile-optimized network, our resulting system
achieves a mAP of 29.6% – exceeding previous mobileoptimized networks that employ Faster-RCNN by over
5.0% (Table 4). For the best NASNet network, our resulting
network operating on images of the same spatial resolution
(800 × 800) achieves mAP = 40.7%, exceeding equivalent
object detection systems based off lesser performing image
featurization (i.e. Inception-ResNet-v2) by 4.0% [28, 51]
(see Appendix for example detections on images and sideby-side comparisons). Finally, increasing the spatial resolution of the input image results in the best reported, single
model result for object detection of 43.1%, surpassing the
best previous best by over 4.0% [36].2 These results provide
further evidence that NASNet provides superior, generic
image features that may be transferred across other computer vision tasks. Figure 9 and Figure 10 in Appendix C
show four examples of object detection results produced by
NASNet-A with the Faster-RCNN framework.
0.920
0.915
0.910
0.905
0.900
0.895
0.890
0
RL Top 1 Unique Models
RL Top 5 Unique Models
RL Top 25 Unique Models
RS Top 1 Unique Models
RS Top 5 Unique Models
RS Top 25 Unique Models
10000
20000
30000
40000
Number of Models Sampled
50000
Figure 6. Measuring the efficiency of random search (RS) to reinforcement learning (RL) for learning neural architectures. The
x-axis measures the total number of model architectures sampled,
and the y-axis is the validation performance after 20 epochs on a
proxy CIFAR-10 training task. We emphasize that the absolute
performance on the proxy task is not important (see text) but only
the relative gain from the initial state. Each pair of curves measures the mean accuracy across top ranking models identified by
each algorithm.
An open question in this proposed method is the training efficiency of the architecture search algorithm. In this
section, we demonstrate the effectiveness of reinforcement
learning for architecture search on the CIFAR-10 image
classification problem and compare it to brute-force random
search (considered to be a very strong baseline for blackbox optimization [5]) given an equivalent amount of computational resources.
We define the effectiveness of an architecture search algorithm as the increase in model performance from the initial architecture identified with the search method. Importantly, we emphasize that the absolute value of the model
performance on the proxy task is less important as this artificially reflects irrelevant factors employed in the architecture search process, e.g., the number of training epochs
and the specific model construction. Thus, we employ the
increase in the model performance as a proxy for judging
convergence of the architecture search algorithm.
Figure 6 shows the performance of reinforcement learning (RL) and random search (RS) as more model architectures are sampled. Note that the best model identified with
RL is significantly better than the best model found by RS
by over 1% as measured by the proxy classification task on
CIFAR-10. Additionally, RL finds an entire range of models that are of superior quality to random search. We observe this in the mean performance of the top-5 and top-25
Model
resolution
mAP (mini-val)
mAP (test-dev)
MobileNet-224 [24]
ShuffleNet (2x) [69]
NASNet-A (4 @ 1056)
600 × 600
600 × 600
600 × 600
19.8%
24.5%†
29.6%
-
ResNet-101-FPN [35]
Inception-ResNet-v2 (G-RMI) [28]
Inception-ResNet-v2 (TDM) [51]
NASNet-A (6 @ 4032)
NASNet-A (6 @ 4032)
800 (short side)
600 × 600
600 × 1000
800 × 800
1200 × 1200
35.7%
37.3%
41.3%
43.2%
36.2%
35.6%
36.8%
40.7%
43.1%
ResNet-101-FPN (RetinaNet) [36]
800 (short side)
-
39.1%
Table 4. Object detection performance on COCO on mini-val and test-dev datasets across a variety of image featurizations. All results
are with the Faster-RCNN object detection framework [46] from a single crop of an image. Top rows highlight mobile-optimized image
featurizations, while bottom rows indicate computationally heavy image featurizations geared towards achieving best results. All mini-val
results employ the same 8K subset of validation images in [28].
models identified in RL versus RS. We take these results
to indicate that although RS may provide a viable search
strategy, RL does significantly improve our ability to learn
neural architectures.
4. Related Work
The proposed method is related to previous work in hyperparameter optimization [43, 4, 5, 53, 54, 6, 39] – especially recent approaches in designing architectures such
as Neural Fabrics [47], DiffRNN [40], MetaQNN [3] and
DeepArchitect [42]. A more flexible class of methods for
designing architecture is evolutionary algorithms [64, 16,
56, 30, 45, 41, 66], yet they have not had as much success
at large scale. Xie and Yuille [66] also transferred learned
architectures from CIFAR-10 to ImageNet but performance
of these models (top-1 accuracy 72.1%) are notably below
previous state-of-the-art (Table 2).
The concept of having one neural network interact with a
second neural network to aid the learning process, or learning to learn or meta-learning [23, 48] has attracted much
attention in recent years [1, 61, 14, 19, 34, 44, 15]. Most
of these approaches have not been scaled to large problems
like ImageNet. An exception is the recent work focused
on learning an optimizer for ImageNet classification that
achieved notable improvements [63].
The design of our search space took much inspiration from LSTMs [22], and Neural Architecture Search
Cell [70]. The modular structure of the convolutional cell is
also related to previous methods on ImageNet such as VGG
[52], Inception [58, 59, 57], ResNet/ResNext [20, 67], and
Xception/MobileNet [9, 24].
5. Conclusion
In this work, we demonstrate how to learn scalable, convolutional cells from data that transfer to multiple image
classification tasks. The learned architecture is quite flexible as it may be scaled in terms of computational cost
and parameters to easily address a variety of problems. In
all cases, the accuracy of the resulting model exceeds all
human-designed models – ranging from models designed
for mobile applications to computationally-heavy models
designed to achieve the most accurate results.
The key insight in our approach is to design a search
space that decouples the complexity of an architecture from
the depth of a network. This resulting search space permits identifying good architectures on a small dataset (i.e.,
CIFAR-10) and transferring the learned architecture to image classifications across a range of data and computational
scales.
The resulting architectures approach or exceed stateof-the-art performance in both CIFAR-10 and ImageNet
datasets with less computational demand than humandesigned architectures [59, 29, 68]. The ImageNet results are particularly important because many state-of-theart computer vision problems (e.g., object detection [28],
face detection [49], image localization [62]) derive image features or architectures from ImageNet classification
models. For instance, we find that image features obtained from ImageNet used in combination with the FasterRCNN framework achieves state-of-the-art object detection
results. Finally, we demonstrate that we can use the resulting learned architecture to perform ImageNet classification with reduced computational budgets that outperform
streamlined architectures targeted to mobile and embedded
platforms [24, 69].
References
[1] M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman,
D. Pfau, T. Schaul, and N. de Freitas. Learning to learn by
gradient descent by gradient descent. In Advances in Neural
Information Processing Systems, pages 3981–3989, 2016.
[2] J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization.
arXiv preprint arXiv:1607.06450, 2016.
[3] B. Baker, O. Gupta, N. Naik, and R. Raskar. Designing neural network architectures using reinforcement learning. In International Conference on Learning Representations, 2016.
[4] J. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl. Algorithms for hyper-parameter optimization. In Neural Information Processing Systems, 2011.
[5] J. Bergstra and Y. Bengio. Random search for hyperparameter optimization. Journal of Machine Learning Research, 2012.
[6] J. Bergstra, D. Yamins, and D. D. Cox. Making a science
of model search: Hyperparameter optimization in hundreds
of dimensions for vision architectures. International Conference on Machine Learning, 2013.
[7] J. Chen, R. Monga, S. Bengio, and R. Jozefowicz. Revisiting
distributed synchronous sgd. In International Conference on
Learning Representations Workshop Track, 2016.
[8] Y. Chen, J. Li, H. Xiao, X. Jin, S. Yan, and J. Feng. Dual
path networks. arXiv preprint arXiv:1707.01083, 2017.
[9] F. Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2017.
[10] D.-A. Clevert, T. Unterthiner, and S. Hochreiter. Fast and
accurate deep network learning by exponential linear units
(elus). In International Conference on Learning Representations, 2016.
[11] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database. In
IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2009.
[12] T. DeVries and G. W. Taylor. Improved regularization of
convolutional neural networks with cutout. arXiv preprint
arXiv:1708.04552, 2017.
[13] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In International Conference on Machine Learning, volume 32, pages
647–655, 2014.
[14] Y. Duan, J. Schulman, X. Chen, P. L. Bartlett, I. Sutskever,
and P. Abbeel. RL2 : Fast reinforcement learning via slow
reinforcement learning. arXiv preprint arXiv:1611.02779,
2016.
[15] C. Finn, P. Abbeel, and S. Levine. Model-agnostic metalearning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017.
[16] D. Floreano, P. Dürr, and C. Mattiussi. Neuroevolution: from
architectures to learning. Evolutionary Intelligence, 2008.
[17] K. Fukushima. A self-organizing neural network model for a
mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, page 93202, 1980.
[18] X. Gastaldi. Shake-shake regularization of 3-branch residual
networks. In International Conference on Learning Representations Workshop Track, 2017.
[19] D. Ha, A. Dai, and Q. V. Le. Hypernetworks. In International Conference on Learning Representations, 2017.
[20] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In IEEE Conference on Computer
Vision and Pattern Recognition, 2016.
[21] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in
deep residual networks. In European Conference on Computer Vision, 2016.
[22] S. Hochreiter and J. Schmidhuber. Long short-term memory.
Neural Computation, 1997.
[23] S. Hochreiter, A. Younger, and P. Conwell. Learning to learn
using gradient descent. Artificial Neural Networks, pages
87–94, 2001.
[24] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang,
T. Weyand, M. Andreetto, and H. Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
[25] J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation networks. arXiv preprint arXiv:1709.01507, 2017.
[26] G. Huang, Z. Liu, and K. Q. Weinberger. Densely connected
convolutional networks. In IEEE Conference on Computer
Vision and Pattern Recognition, 2017.
[27] G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Weinberger. Deep
networks with stochastic depth. In European Conference on
Computer Vision, 2016.
[28] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara,
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, et al.
Speed/accuracy trade-offs for modern convolutional object
detectors. In IEEE Conference on Computer Vision and Pattern Recognition, 2017.
[29] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift.
In International Conference on Learning Representations,
2015.
[30] R. Jozefowicz, W. Zaremba, and I. Sutskever. An empirical
exploration of recurrent network architectures. In International Conference on Learning Representations, 2015.
[31] A. Krizhevsky. Learning multiple layers of features from
tiny images. Technical report, University of Toronto, 2009.
[32] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing System, 2012.
[33] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 1998.
[34] K. Li and J. Malik. Learning to optimize neural nets. arXiv
preprint arXiv:1703.00441, 2017.
[35] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and
S. Belongie. Feature pyramid networks for object detection.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2017.
[36] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár.
Focal loss for dense object detection. arXiv preprint
arXiv:1708.02002, 2017.
[37] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer, 2014.
[38] I. Loshchilov and F. Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on
Learning Representations, 2017.
[39] H. Mendoza, A. Klein, M. Feurer, J. T. Springenberg, and
F. Hutter. Towards automatically-tuned neural networks. In
Proceedings of the 2016 Workshop on Automatic Machine
Learning, pages 58–65, 2016.
[40] T. Miconi. Neural networks with differentiable structure.
arXiv preprint arXiv:1606.06216, 2016.
[41] R. Miikkulainen, J. Liang, E. Meyerson, A. Rawal, D. Fink,
O. Francon, B. Raju, A. Navruzyan, N. Duffy, and B. Hodjat.
Evolving deep neural networks.
arXiv preprint
arXiv:1703.00548, 2017.
[42] R. Negrinho and G. Gordon. DeepArchitect: Automatically
designing and training deep architectures. arXiv preprint
arXiv:1704.08792, 2017.
[43] N. Pinto, D. Doukhan, J. J. DiCarlo, and D. D. Cox. A highthroughput screening approach to discovering good forms of
biologically inspired visual representation. PLoS Computational Biology, 5(11):e1000579, 2009.
[44] S. Ravi and H. Larochelle. Optimization as a model for fewshot learning. In International Conference on Learning Representations, 2017.
[45] E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu,
Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning,
2017.
[46] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, pages 91–99, 2015.
[47] S. Saxena and J. Verbeek. Convolutional neural fabrics. In
Advances in Neural Information Processing Systems, 2016.
[48] T. Schaul and J. Schmidhuber. Metalearning. Scholarpedia,
2010.
[49] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 815–823, 2015.
[50] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and
O. Klimov. Proximal policy optimization algorithms. arXiv
preprint arXiv:1707.06347, 2017.
[51] A. Shrivastava, R. Sukthankar, J. Malik, and A. Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv preprint arXiv:1612.06851, 2016.
[52] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. In International
Conference on Learning Representations, 2015.
[53] J. Snoek, H. Larochelle, and R. P. Adams. Practical Bayesian
optimization of machine learning algorithms. In Neural Information Processing Systems, 2012.
[54] J. Snoek, O. Rippel, K. Swersky, R. Kiros, N. Satish, N. Sundaram, M. Patwary, M. Ali, R. P. Adams, et al. Scalable
Bayesian optimization using deep neural networks. In International Conference on Machine Learning, 2015.
[55] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov. Dropout: a simple way to prevent neu-
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
ral networks from overfitting. Journal of Machine Learning
Research, 15(1):1929–1958, 2014.
K. O. Stanley, D. B. D’Ambrosio, and J. Gauci. A
hypercube-based encoding for evolving large-scale neural
networks. Artificial Life, 2009.
C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi. Inceptionv4, Inception-Resnet and the impact of residual connections
on learning. In International Conference on Learning Representations Workshop Track, 2016.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In IEEE Conference on
Computer Vision and Pattern Recognition, 2015.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna.
Rethinking the Inception architecture for computer vision. In
IEEE Conference on Computer Vision and Pattern Recognition, 2016.
D. Ulyanov, A. Vedaldi, and V. Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv
preprint arXiv:1607.08022, 2016.
J. X. Wang, Z. Kurth-Nelson, D. Tirumala, H. Soyer,
J. Z. Leibo, R. Munos, C. Blundell, D. Kumaran, and
M. Botvinick. Learning to reinforcement learn. arXiv
preprint arXiv:1611.05763, 2016.
T. Weyand, I. Kostrikov, and J. Philbin. Planet-photo geolocation with convolutional neural networks. In European
Conference on Computer Vision, 2016.
O. Wichrowska, N. Maheswaranathan, M. W. Hoffman, S. G.
Colmenarejo, M. Denil, N. de Freitas, and J. Sohl-Dickstein.
Learned optimizers that scale and generalize. arXiv preprint
arXiv:1703.04813, 2017.
D. Wierstra, F. J. Gomez, and J. Schmidhuber. Modeling
systems with internal state using evolino. In The Genetic
and Evolutionary Computation Conference, 2005.
R. J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. In Machine
Learning, 1992.
L. Xie and A. Yuille. Genetic CNN. arXiv preprint
arXiv:1703.01513, 2017.
S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He. Aggregated
residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2017.
X. Zhang, Z. Li, C. C. Loy, and D. Lin. Polynet: A pursuit
of structural diversity in very deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2017.
X. Zhang, X. Zhou, L. Mengxiao, and J. Sun. Shufflenet: An
extremely efficient convolutional neural network for mobile
devices. arXiv preprint arXiv:1707.01083, 2017.
B. Zoph and Q. V. Le. Neural architecture search with reinforcement learning. In International Conference on Learning
Representations, 2017.
Appendix
A. Experimental Details
A.1. Dataset for Architecture Search
The CIFAR-10 dataset [31] consists of 60,000 32x32
RGB images across 10 classes (50,000 train and 10,000
test images). We partition a random subset of 5,000 images
from the training set to use as a validation set for the controller RNN. All images are whitened and then undergone
several data augmentation steps: we randomly crop 32x32
patches from upsampled images of size 40x40 and apply
random horizontal flips. This data augmentation procedure
is common among related work.
A.2. Controller architecture
The controller RNN is a one-layer LSTM [22] with 100
hidden units at each layer and 2 × 5B softmax predictions
for the two convolutional cells (where B is typically 5) associated with each architecture decision. Each of the 10B
predictions of the controller RNN is associated with a probability. The joint probability of a child network is the product of all probabilities at these 10B softmaxes. This joint
probability is used to compute the gradient for the controller
RNN. The gradient is scaled by the validation accuracy of
the child network to update the controller RNN such that the
controller assigns low probabilities for bad child networks
and high probabilities for good child networks.
Unlike [70], who used the REINFORCE rule [65] to update the controller, we employ Proximal Policy Optimization (PPO) [50] with learning rate 0.00035 because training with PPO is faster and more stable. To encourage exploration we also use an entropy penalty with a weight of
0.00001. In our implementation, the baseline function is
an exponential moving average of previous rewards with a
weight of 0.95. The weights of the controller are initialized
uniformly between -0.1 and 0.1.
A.3. Training of the Controller
For distributed training, we use a workqueue system
where all the samples generated from the controller RNN
are added to a global workqueue. A free “child” worker in
a distributed worker pool asks the controller for new work
from the global workqueue. Once the training of the child
network is complete, the accuracy on a held-out validation set is computed and reported to the controller RNN.
In our experiments we use a child worker pool size of 450,
which means there are 450 networks being trained on 450
GPUs concurrently at any time. Upon receiving enough
child model training results, the controller RNN will perform a gradient update on its weights using PPO and then
sample another batch of architectures that go into the global
workqueue. This process continues until a predetermined
number of architectures have been sampled. In our experiments, this predetermined number of architectures is 20,000
which means the search process is terminated after 20,000
child models have been trained. Additionally, we update the
controller RNN with minibatches of 20 architectures. Once
the search is over, the top 250 architectures are then chosen
to train until convergence on CIFAR-10 to determine the
very best architecture.
A.4. Details of architecture search space
We performed preliminary experiments to identify a flexible, expressive search space for neural architectures that
learn effectively. Generally, our strategy for preliminary experiments involved small-scale explorations to identify how
to run large-scale architecture search.
• All convolutions employ ReLU nonlinearity. Experiments with ELU nonlinearity [10] showed minimal
benefit.
• To ensure that the shapes always match in convolutional cells, 1x1 convolutions are inserted as necessary.
• Unlike [24], all depthwise separable convolution do
not employ Batch Normalization and/or a ReLU between the depthwise and pointwise operations.
• All convolutions followed an ordering of ReLU, convolution operation and Batch Normalization following
[21].
• Whenever a separable convolution is selected as an operation by the model architecture, the separable convolution is applied twice to the hidden state. We found
this empirically to improve overall performance.
A.5. Training with stochastic regularization
We performed several experiments with various stochastic regularization methods. Naively applying dropout [55]
across convolutional filters degraded performance. However, when training NASNet models, we found that stochastically dropping out each path (i.e., edge with a yellow box
in Figure 4) in the cell with some fixed probability to be an
effective regularizer. This is similar to [27] and [68] where
they dropout full parts of their model during training and
then at test time scale the path by the probability of keeping that path during training. Interestingly we found that
linearly increasing the probability of dropping out a path
over the course of training to significantly improve the final
performance for both CIFAR and ImageNet experiments.
A.6. Training of CIFAR models
All of our CIFAR models use a single period cosine decay as in [38, 18]. All models use the momentum optimizer
with momentum rate set to 0.9. All models also use L2
weight decay. Each architecture is trained for a fixed 20
epochs on CIFAR-10 during the architecture search process.
Additionally, we found it beneficial to use the cosine learning rate decay during the 20 epochs the CIFAR models were
trained as this helped to further differentiate good architectures. We also found that having the CIFAR models use a
small N = 2 during the architecture search process allowed
for models to train quite quickly, while still finding cells
that work well once more were stacked.
A.7. Training of ImageNet models
We use ImageNet 2012 ILSVRC challenge data for large
scale image classification. The dataset consists of ∼ 1.2M
images labeled across 1,000 classes [11]. Overall our training and testing procedures are almost identical to [59]. ImageNet models are trained and evaluated on 299x299 or
331x331 images using the same data augmentation procedures as described previously [59]. We use distributed synchronous SGD to train the ImageNet model with 50 workers (and 3 backup workers) each with a Tesla K40 GPU [7].
We use RMSProp with a decay of 0.9 and epsilon of 1.0.
Evaluations are calculated using with a running average of
parameters over time with a decay rate of 0.9999. We use
label smoothing with a value of 0.1 for all ImageNet models as done in [59]. Additionally, all models use an auxiliary
classifier located at 2/3 of the way up the network. The loss
of the auxiliary classifier is weighted by 0.4 as done in [59].
We empirically found our network to be insensitive to the
number of parameters associated with this auxiliary classifier along with the weight associated with the loss. All
models also use L2 regularization. The learning rate decay scheme is the exponential decay scheme used in [59].
Dropout is applied to the final softmax matrix with probability 0.5.
B. Additional Experiments
We now present two additional cells that performed well
on CIFAR and ImageNet. The search spaces used for these
cells are slightly different than what was used for NASNetA. For the NASNet-B model in Figure 7 we do not concatenate all of the unused hidden states generated in the convolutional cell. Instead all of the hiddenstates created within
the convolutional cell, even if they are currently used, are
fed into the next layer. Note that B = 4 and there are 4 hiddenstates as input to the cell as these numbers must match
for this cell to be valid. We also allow addition followed by
layer normalization [2] or instance normalization [60] to be
predicted as two of the combination operations within the
cell, along with addition or concatenation.
For NASNet-C (Figure 8), we concatenate all of the unused hidden states generated in the convolutional cell like
in NASNet-A, but now we allow the prediction of addition
Figure 7. Architecture of NASNet-B convolutional cell with B =
4 blocks identified with CIFAR-10. The input (white) is the hidden
state from previous activations (or input image). Each convolutional cell is the result of B blocks. A single block is corresponds
to two primitive operations (yellow) and a combination operation
(green). As do we not concatenate the output hidden states, each
output hidden state is used as a hidden state in the future layers.
Each cell takes in 4 hidden states and thus needs to also create 4
output hidden states. Each output hidden state is therefore labeled
with 0, 1, 2, 3 to represent the next four layers in that order.
followed by layer normalization or instance normalization
like in NASNet-B.
C. Example object detection results
Finally, we will present examples of object detection results on the COCO dataset in Figure 9 and Figure 10. As can
be seen from the figures, NASNet-A featurization works
well with Faster-RCNN and gives accurate localization of
objects.
Figure 9. Example detections showing improvements of object detection over previous state-of-the-art model for Faster-RCNN with
Inception-ResNet-v2 featurization [28] (top) and NASNet-A featurization (bottom).
Figure 8. Architecture of NASNet-C convolutional cell with B =
4 blocks identified with CIFAR-10. The input (white) is the hidden state from previous activations (or input image). The output
(pink) is the result of a concatenation operation across all resulting branches. Each convolutional cell is the result of B blocks. A
single block corresponds to two primitive operations (yellow) and
a combination operation (green).
Figure 10. Example detections of best performing NASNet-A featurization with Faster-RCNN trained on COCO dataset. Top and
middle images courtesy of http://wikipedia.org. Bottom
image courtesy of Jonathan Huang
| 1cs.CV
|
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT
SOLVERS
arXiv:1510.00523v1 [cs.DS] 2 Oct 2015
TAKAHISA TODA AND TAKEHIDE SOH
Abstract. All solutions SAT (AllSAT for short) is a variant of
propositional satisfiability problem. Despite its significance, AllSAT has been relatively unexplored compared to other variants.
We thus survey and discuss major techniques of AllSAT solvers.
We faithfully implement them and conduct comprehensive experiments using a large number of instances and various types of solvers
including one of the few public softwares. The experiments reveal
solver’s characteristics. Our implemented solvers are made publicly
available so that other researchers can easily develop their solver
by modifying our codes and compare it with existing methods.
1. Introduction
Propositional satisfiability (SAT for short) is to decide if a Boolean
formula is satisfiable. SAT is ubiquitous in computer science. Because
of its significance, it has attracted the attention of many researchers
from theory to practice. Many applications have motivated empirical
studies, in particular the development of SAT solvers, softwares to solve
satisfiability. A fundamental task of SAT solvers is to solve as many
instances as possible in a realistic amount of time. To this end, various
practical algorithms and elegant implementation techniques have been
developed [56] [9] [40].
There are many variants of SAT. All solutions SAT (AllSAT for
short) or model enumeration is studied in the paper. It is, given a
CNF formula, to generate partial satisfying assignments such that they
form a logically equivalent DNF formula. Compared to neighboring
areas, AllSAT has been relatively unexplored. This is mentioned in
the literature [29] and also supported by the fact that there are only a
few recent papers, almost no software1 is publicly available, and it has
not even been taken up in major handbooks related to satisfiability.
A recent application of AllSAT is data mining. A fundamental task
in data mining is to generate all interesting patterns from a given database [26]. Examples include frequent itemsets, maximal frequent itemsets, or closed itemsets in transaction databases. Although algorithms
1A
few exceptions are clasp [21], picosat [10], and relsat [6]. Although they
support solution generation, they are positioned as an answer set solver, a single
solution SAT solver, and a #SAT solver rather than AllSAT solvers, respectively.
1
2
T. TODA AND T. SOH
for generating various patterns have been proposed, they are basically
specialized for their target patterns. This means that different patterns
require new algorithms. For this reason, a framework based on declarative paradigm has recently been proposed [24]. A basic flow is that
constrains of patterns to be generated are formulated as logical formulae and solved with a generic solver. Hence, all that users do is simply to
model their problems, not to design algorithms. Among much related
work, an approach on which problems are encoded into CNF formulae
and solved with AllSAT solvers has been studied [31]. An advantage of
declarative paradigm is its ability to handle new patterns in a flexible
manner. There is no need to see details of algorithms on which solvers
are based, thereby it is opened to wider users. It is instead inferior in
efficiency to problem-specific approaches. In practice it is necessary to
balance between efficiency and flexibility. Therefore, improving solver’s
performance is essential in the declarative framework.
Besides the data mining, there have been many studies on the application of ALLSAT, in particular to formal verification, such as network
verification [58] [38] [49], predicate abstraction [14], backbone computation [41], image and preimage computation in unbounded model
checking [25] [43] [51] [37] [20] [23].
Considering the above, we find it important to clarify state-of-theart techniques of AllSAT solvers and to improve them on a firm basis.
However, there are the following issues in the existing researches of
ALLSAT.
• Several methods are proposed but they are not globally compared. It is thus difficult to decide which method is effective for
which kinds of ALLSAT instances.
• Experiments are not carried out on comprehensive benchmarks.
• There is few public ALLSAT solver, which makes it difficult to
compare existing techniques.
We thus would like to survey major techniques of AllSAT solvers and
try to complement past references by gathering and organizing existing techniques. We further add some novel techniques. To evaluate
solvers, we conduct experimental comparisons, including clasp, one of
the few off-the-shelf softwares with solution generation support. Our
implemented solvers are made publicly available with expectation that
further improvement on solvers and their evaluation are easily done
and the AllSAT research is stimulated.
The paper is organized as follows. Section 2 provides related work of
AllSAT. Section 3 provides necessary notions, terminology, and results.
Section 4 surveys major techniques of AllSAT solvers, where those including our original ideas are indicated by adding asterisks to their
titles. Section 5 provides experimental results. Section 6 concludes the
paper.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
3
2. Related Work
Another variant of SAT is the dualization of Boolean functions: given
a DNF formula of a Boolean function f , it is to compute the complete
DNF formula of the dual function f d . Since a CNF formula of f d
can be easily obtained by interchanging logical disjunction with logical conjunction as well as the constants 0 with 1, a main part is to
convert CNF to the complete DNF. Hence, an essential difference from
AllSAT is that the resulting DNF formula must be complete. Dualization has been well-studied in terms of complexity [18] [15], while there
seems no recent empirical study with a few exceptions [33] [11] [30] [54].
Practical algorithms for a restricted form of dualization have been presented [53] [46] and some implementations are available2, though they
are not for arbitrary Boolean functions.
Another variant is the problem of counting the number of total satisfying assignments, called propositional model counting or #SAT. It
has been well-studied because of good applications such as probabilistic
inference problems and hard combinatorial problems, and some solvers
are available [9]. Although #SAT is apparently similar to AllSAT,
techniques such as connected components and component caching are
inherent in counting, and they are not applicable to AllSAT as is [44].
3. Preliminaries
Necessary notions, terminology, and results concerning Boolean functions, satisfiability solvers, and binary decision diagrams are presented
in this section.
3.1. Boolean Basics. A literal is a Boolean variable or its negation. A
clause is a finite disjunction of literals, and a term is a finite conjunction
of literals. A propositional formula is in conjunctive normal form (CNF
for short) if it is a finite conjunction of clauses and in a disjunction normal form (DNF for short) if it is a finite disjunction of terms. We identify clauses with sets of literals and CNF formulae with sets of clauses.
The same applies to terms and DNFs. The dual of a Boolean function
f is the function f d defined by f d (x1 , . . . , xn ) = ¬f (¬x1 , . . . , ¬xn ). An
implicant of a Boolean function f is a term t with t ≤ f , where t is
considered as a Boolean function and the order of Boolean functions is
introduced as t ≤ f if t(v) ≤ f (v) for all v ∈ {0, 1}n . An implicant is
prime if the removal of any literal results in a non-implicant. A DNF
formula is complete if it consists of all prime implicants.
2Hypergraph
Dualization Repository, by Keisuke Murakami and Takeaki Uno,
http://research.nii.ac.jp/~uno/dualization.html, accessed on 19th Jan.,
2013. HTC-DD: Hypergraph Transversal Computation with Binary Decision Diagrams, by Takahisa Toda at ERATO MINATO Discrete Structure Manipulation
System Project, Japan Science and Technology Agency, at Hokkaido University,
http://www.sd.is.uec.ac.jp/toda/htcbdd.html, accessed on 10th Sept., 2015.
4
T. TODA AND T. SOH
An assignment to a set V of Boolean variables is a partial function
from V to {0, 1}. A satisfying assignment for a CNF formula is an assignment ν such that the CNF formula evaluates to 1. An assignment
to V is total (or complete) if it is a total function, that is, all variables in
V are assigned values. A Boolean formula is satisfiable if it has a satisfying assignment. For simplicity, we say that a literal is assigned a value
v if the assignment to the underlying variable makes the literal evaluate
to v. If there is no fear of confusion, we identify an assignment function
ν over V with the set of the form {(x, v) ∈ V × {0, 1} | ν(x) = v}. We
further identify the assignment x 7→ v with x if v = 1 and with ¬x if
v = 0. In this way, assignments and literals are used interchangeably
throughout the paper.
Example 1. Consider the sequence of literals x1 , ¬x5 , x3 . This means
that x1 , x5 , x3 are selected in this order and the values 1, 0, 1 are assigned to them, respectively.
3.2. Satisfiability Solvers. Propositional satisfiability problem (SAT
for short) is the problem of deciding if there exists a satisfying assignment for a CNF formula. Algorithm 1 shows a basic framework on
which modern SAT solvers are based. For simplicity, other techniques
such as lazy data structures, variable selection heuristics, restarting,
deletion policy of learnt clauses are omitted. See for details [56] and [9].
A basic behavior of Algorithm 1 is to search a satisfying assignment
in such a way that a solver finds a candidate assignment by assigning
values to variables and if the assignment turns out to be unsatisfying, the solver proceeds to the next candidate by backtracking. The
extension of an assignment is triggered at the decide stage, where an
unassigned variable x is selected and a value v ∈ {0, 1} is assigned to x,
and it is then spread at the deduce stage, where assignments to other
variables are deduced from the most recent decision (x, v).
Decision assignments are those given at the decide stage and decision
variables are those assigned values there. Consider a decision tree
that branches at each decision assignment. A decision level is the
depth of that decision tree, which is maintained by the variable dl in
Algorithm 1. The decision level of a variable x, denoted by δ(x), is one
at which x was assigned a value. For a literal l, the notation δ(l) is
defined as that of the underlying variable, and l@d denotes that l, seen
as an assignment, was given at level d, i.e. δ(l) = d.
The deduce stage is described below. A clause is unit if all but one
of literals are assigned the value 0 and the remaining one is unassigned.
The remaining literal is called a unit literal. Unit clauses are important
in the deduce stage because assignments to the underlying variables of
unit literals are necessarily determined so that unit literals evaluate to
1. After some assignments are determined by unit clauses, non-unit
clauses may become unit. Hence, all implications are deduced until an
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
5
ALGORITHM 1: DPLL procedure with conflict driven clause learning,
where δ(x) denotes the decision level of a variable x.
Input: a CNF formula ψ, an empty assignment ν.
Output: SAT if ψ is satisfied; UNSAT, otherwise.
dl ← 0;
/* Decision level */
while true do
ν ← propagate (ψ, ν);
/* Deduce stage */
if conflict happens then
if dl ≤ 0 then return UNSAT;
bl ← analyze (ψ, ν);
/* Diagnose stage */
ν ← {(x, v) ∈ ν | δ(x) ≤ bl};
dl ← bl;
else
if all variables are assigned values then
report ν;
return SAT;
else
dl ← dl + 1;
/* Decide stage */
select an unassigned variable x and a value v;
ν ← ν ∪ {(x, v)};
end
end
end
unsatisfied clause exists or no unit clause exists. This process is called
unit propagation. The function propagate performs unit propagation.
Implied assignments are those given at this stage and implied variables
are those assigned values there. The decision level of an implied variable
x, the notations δ(x) and l@d, where l is a literal representing an
implied assignment, are defined in the same way.
Example 2. Consider the CNF formula ψ that consists of the following
clauses.
C1
C2
C3
C4
C5
=
=
=
=
=
x1 ∨ ¬x3
x2 ∨ x3 ∨ x5
¬x1 ∨ ¬x3 ∨ x4
x4 ∨ ¬x5 ∨ x6
x5 ∨ ¬x6
Assume the decision assignment ¬x5 @1. The implied assignment ¬x6 @1
is obtained from C5 . Assume the decision assignment x3 @2. The implied assignments x1 @2 and x4 @2 are obtained from C1 and C3 in this
6
T. TODA AND T. SOH
¬
¬
¬
¬
¬
Figure 1. A conflict graph, where arcs are labeled with
antecedents of their target vertices.
order. Assume the decision assignment x2 @3, and the CNF formula is
satisfied.
It should be noted that in the middle of unit propagation, we may
encounter with an unsatisfied clause. This case is called conflict. As
soon as conflict happens, unit propagation halts, even though unit
clauses still remain. In conflict case, if all assigned variables are those
assigned prior to any decision (i.e. dl = 0), it means that there are
no other assignments to be examined, thereby a CNF formula must be
unsatisfiable. In that case, a solver halts, reporting UNSAT. If decision
has been made at least once, we enter into the diagnose stage to resolve
conflict.
At the diagnose stage, a ”cause” of the conflict we have just met is
analyzed, a new clause is learnt as a result, and it is added to a CNF
formula, by which a solver is guided not to fall into the conflict again
(and other conflicts related to it). To do this efficiently, modern solvers
maintain an implication graph during search, which represents an implication relation between assignments over unit propagation. Specifically, an implication graph is a directed acyclic graph G = (V, A) such
that
• vertices in V correspond to literals l representing assignments
to their variables;
• arcs in A correspond to implications so that if a unit clause C
with unit literal l yields in unit propagation, then arcs from all
assignments to underlying variables in C \ {l} to the implied
assignment l are added;
• if an unsatisfied clause exists, then arcs from all assignments to
underlying variables in that clause to the special vertex κ are
added.
An implication graph might be implemented so that whenever a variable x is implied, it is associated with the clause that determined the
assignment to x as a unit clause. This clause is called the antecedent
of x.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
7
Example 3. Consider the CNF formula ψ given in Example 2. Assume in turn the decision assignments ¬x4 @1, ¬x6 @2, and ¬x2 @3 in
this order. The resulting implication graph is shown in Fig. 1. This
case results in conflict because C1 becomes unsatisfied.
A conflict graph is a subgraph H of an implication graph G obtained
by restricting G so that all vertices have paths to κ. For a subset U
of vertices in a conflict graph H = (V, A), the arc-cut (hereafter cut)
corresponding to U is the set of arcs that connect vertices in U with
those in V \ U . Examples are illustrated by dotted curves in Fig. 1.
We are now ready to describe a clause learning scheme, which is
performed by the function analyze. Consider cuts such that all decision
assignments are on one side, called the reason side, and the special
vertex κ is on the other side, called the conflict side. Take negation of
literals on reason side that are incident to arcs in a cut, which form a
conflict-driven clause (or conflict clause). This clause is considered as
a ”cause” of conflict. Indeed,if variables are assigned values following
literals on reason side that are incident to arcs in a cut, then the same
implications on conflict side are derived and the same conflict must
take place. Therefore, to avoid the conflict, it is necessary for variables
to be assigned values so that at least one of those literals is negated.
This condition is formulated as the conflict clause obtained above.
As illustrated in Fig. 1, there are many choices for cuts that induce
conflict clauses. Among them, conflict clauses that contain exactly one
literal from the current decision level are known to be effective. A
unique implication point (UIP for short) is a vertex in a conflict graph
such that every path from the decision at the current decision level to
κ passes through it. Note that at least one UIP exists, because the
decision at the current decision level is a UIP. The first UIP scheme is
to find the UIP that is the closest to κ.
Example 4. Consider the conflict graph given in Fig. 1. The middle
curve gives the first UIP x3 . Hence, a conflict clause is x4 ∨ ¬x3 .
The first UIP scheme can be efficiently performed by traversing a
conflict graph from κ in the reverse order of implications, based on
the implementation of an implication graph stated above. During the
traversal, it is easy to decide if the current literal is a UIP. Consider
the cut that induces a UIP. For all literals on the reason side, their
assignments are determined prior to those on the conflict side. Since
the traversal is in the reverse order of implications, the UIP can then
be located by keeping track of the number of unvisited vertices that
are immediate neighbors of visited vertices.
All that remains is to decide a backtrack level bl, cancel all assignments above bl, and set bl to the current decision level. To decide a
backtracking level, there are two choices. Suppose that the decision p
at level i resulted in conflict and there is no solution that extends the
8
T. TODA AND T. SOH
current assignment. Chronological backtracking cancels all assignments
of level i including the decision p and attempts to find a solution by
extending the assignment from level i − 1 with a conflict clause. In
this way, chronological backtracking undoes assignments from a higher
to a lower level. A drawback is that if there is no solution in higher
levels, it is hard to get out of those levels. On the other hand, nonchronological backtracking jumps at once to a lower level j by canceling
all assignments above j and attempts to find a solution upward by extending assignment from level j. The backtracking level j is commonly
determined as the largest level of a variable in a conflict clause below
the current level i. For each canceled assignment between the levels j
and i, the possibility for becoming a solution is left in general, though
this does not mean that a solver loses an opportunity to find solutions.
Example 5. Consider the conflict graph given in Fig. 1. Since a conflict clause is x4 ∨¬x3 , the non-chronological backtrack level is 1. Hence,
the assignments ¬x6 @2, ¬x5 @2 are canceled and search restarts from
¬x4 @1, however a solution could be obtained by backtracking to the level
2.
Remark 1. When a conflict clause is learnt, an assignment to its
unit literal is implied. Hence, the function analyze adds the implied
assignment to the current assignment function ν, by which its effect to
other assignments is taken at a subsequent propagation.
3.3. Binary Decision Diagrams. A binary decision diagram (BDD
for short) is a graphical representation of Boolean functions in a compressed form [36] [1] [12]. We follow the notation and terminology in
Knuth’s book [35].
Figure 2 shows an example of a BDD. Exactly one node has indegree
0, which is called the root. Each branch node f has a label and two
children. Node labels are taken from variable indices, and the children
consists of the LO child and the HI child. The arc to a LO child is
called LO arc, illustrated by a dotted arrow, and the LO arc of f means
assigning the value 0 to the variable of f . Similarly, the arc to a HI
child is called HI arc, illustrated by a solid arrow, and the HI arc of f
in turn means assigning the value 1 to its variable. There are two sink
nodes, denoted by > and ⊥. Paths from the root to > and ⊥ mean
satisfying and unsatisfying assignments, respectively.
BDDs are called ordered if for any node u with a branch node v as
its child, the index of u is less than that of v. BDDs are called reduced
if the following reduction operations can not be applied further.
(1) If there is a branch node u whose arcs both point to v, then
redirect all the incoming arcs of u to v, and then eliminate u
(Fig. 3(a)).
(2) If there are two branch nodes u and v such that the subgraphs
rooted by them are equivalent, then merge them (Fig. 3(b)).
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
⊥
1
2
3
9
>
Figure 2. The BDD representation for the function:
f (x1 , x2 , x3 ) = x1 ∧ ¬x2 ∨ x3 .
i
"
j
|
j
+3
i
jv
(a) Node elimination
i
(
k
iu
j
+3
(
k
(b) Node merging
Figure 3. The reduction rules
In this paper, ordered reduced BDDs are simply called BDDs. Ordered
BDDs that need not be fully reduced are distinguished from ordinary
BDDs by calling OBDDs. Note that each node in a BDD (or an OBDD)
is conventionally identified with the subgraph rooted by it, which also
forms BDD (or OBDD).
4. Techniques of All Solutions SAT Solvers
In order to implement an efficient AllSAT solver, we have to carefully
determine an appropriate suite of techniques by considering various
factors and their characteristics. Some details have been mentioned
only partly and scattered in past references. We survey major existing
techniques of AllSAT solvers and try to complement past references
by gathering and organizing existing techniques. We further add some
novel techniques.
This section is organized as follows. Three major types of solvers
are presented in their own subsections. Each subsection starts with an
overview, then provides specific techniques, and ends with the configuration of our implemented solvers. We added an asterisk to the title
of each specific technique that contains our original ideas.
4.1. Blocking Solvers.
4.1.1. Overview. One of the easiest ways of implementing an AllSAT
solver is to repeatedly run an ordinary SAT solver as a black box and
find satisfying assignments one by one. A specific procedure is as follows.
10
T. TODA AND T. SOH
(1)
(2)
(3)
(4)
Run a SAT solver with a CNF formula ψ.
If ψ is unsatisfiable, halt.
Report a found total satisfying assignment ν.
Compute the clause C of the form {x | ν(x) = 0}∪{¬x | ν(x) = 1}
in a set notation.
(5) Add C to ψ and go to Step 1.
The clause obtained at Step 4 is called a blocking clause [43]. Since
the blocking clause C is the complement of the term corresponding
to ν, the extended CNF formula ψ ∪ {C} is not satisfied by ν in a
later search, thereby a new solution will be found in each repetition.
Furthermore, since ν is total, no assignment other than ν is blocked.
Example 6. Execute the procedure with the CNF formula:
ψ = (x1 ∨ ¬x2 ) ∧ (x2 ∨ ¬x3 ) ∧ (x3 ∨ ¬x1 ).
Suppose a solver returns the satisfying assignment ¬x1 , ¬x2 , ¬x3 . The
blocking clause x1 ∨ x2 ∨ x3 is added to ψ:
ψ = (x1 ∨ ¬x2 ) ∧ (x2 ∨ ¬x3 ) ∧ (x3 ∨ ¬x1 ) ∧ (x1 ∨ x2 ∨ x3 ).
The solver then returns the satisfying assignment x1 , x2 , x3 . The blocking clause ¬x1 ∨ ¬x2 ∨ ¬x3 is added to ψ:
ψ = (x1 ∨¬x2 )∧(x2 ∨¬x3 )∧(x3 ∨¬x1 )∧(x1 ∨x2 ∨x3 )∧(¬x1 ∨¬x2 ∨¬x3 ).
This time a solver returns UNSAT, which means all satisfying assignments are found.
Since a blocking clause has size equal to the number of variables,
unit propagation is likely to slow down. Hence, it is arguably better
to consider blocking clauses that consist only of decisions, which we
call a decision-based blocking clause for convenience. Since decisions
determine the other assignments, decision-based blocking clauses have
the same effect as those of all literals. However, to know which literals
are decisions, we need to modify a solver code. Algorithm 2 is a pseudo
code obtained by modifying Algorithm 1. For convenience, we simply
call this blocking procedure to distinguish it from the other procedures
presented later. Only lines 5,12-16 are changed. At line 5 and 12, a
solver halts because dl ≤ 0 means that all variables are implied without
any decision and all solutions are found. At line 15, all assignments
except for those determined without any decision are canceled, and at
line 16 a solver backtracks to the root level.
A decision-based blocking clause only blocks a single assignment and
there are as many number of blocking clauses as total satisfying assignments. Since they are stored, it is likely to result in space explosion and
slow down of unit propagation. This is considered as a serious issue
since unit propagation in modern SAT solvers occupies the majority
of a whole processing time. Another disadvantage is that whenever a
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
11
ALGORITHM 2: Blocking procedure, where δ(x) denotes the decision
level of x.
Input: a CNF formula ψ, an empty assignment ν.
Output: all satisfying assignments.
dl ← 0;
/* Decision level */
while true do
ν ← propagate (ψ, ν);
/* Deduce stage */
if conflict happens then
if dl ≤ 0 then halt;
bl ← analyze (ψ, ν);
/* Diagnose stage */
ν ← {(x, v) ∈ ν | δ(x) ≤ bl};
dl ← bl;
else
if all variables are assigned values then
report ν;
if dl ≤ 0 then halt;
compute a blocking clause C from ν;
ψ ← ψ ∪ {C};
ν ← {(x, v) ∈ ν | δ(x) = 0};
dl ← 0;
else
dl ← dl + 1;
/* Decide stage */
select an unassigned variable x and a value v;
ν ← ν ∪ {(x, v)};
end
end
end
solution is found, a solver is enforced to restart from scrach with an
extended CNF formula, not resuming search.
On the other hand, the blocking clause-based implementation might
be considered as a good choice for such instances as even one solution
is hard to find or for instances with a small number of solutions. It is
easily implementable, because the blocking clause mechanism can be
realized outside a solver or with a small modification on a solver code as
demonstrated in Algorithm 2. We can benefit from powerful techniques
of modern SAT solvers such as conflict-driven clause learning, nonchronological backtracking, and so on.
Remark 2. Once blocking clauses are added to a CNF formula, they
are not deleted afterward and must not be treated in the same way
as conflict clauses. Otherwise, solutions would be rediscovered many
times, which is not allowed in the paper.
12
T. TODA AND T. SOH
4.1.2. Simplifying Satisfying Assignments. Simplification of satisfying
assignments is to obtain from a satisfying assignment ν to a CNF formula ψ a smaller assignment3 ν 0 that still makes ψ evaluate to 1. This
is done by canceling assignments to redundant variables in ν, where
redundant means either value is assigned without effect on the value
of ψ. A simplified assignment is partial in general and it represents a
set of total assignments, including the original assignment and possibly other satisfying assignments. Variable lifting refers to a number of
such simplification techniques [50]. Since this topic is well-summarized
in the literature [44], we do not go into details. The interested readers
are referred to it, as well as the references therein. For recent results,
see also the literature [57].
Simplification allows us to obtain from a single solution possibly exponentially many solutions in a compact form, i.e., as a partial assignment. It is desirable if we can obtain a partial assignment of minimum
size, however minimization is known to be computationally hard, and
in practice we have to compromise with near-minimum assignment by
means of approximation4. To combine simplification with the blocking
mechanism of Algorithm 2, it suffices to perform simplification just before line 13 and then to take complement of decisions in a simplified
assignment. It should be noted that if a simplified blocking clause is
empty, that is, all variables except for implied ones turn out to be redundant, then it means that all remaining solutions are covered. Thus,
in that case, a solver must be halted. Thanks to simplification, the
number of blocking clauses may be largely reduced, which leads to a
good effect on unit propagation at the cost of performing simplification.
Example 7. Consider the CNF formula ψ given in Example 2. The
assignments ¬x5 @1, ¬x6 @1, x3 @2, x1 @2, x4 @2, x2 @3 were given in this
order. The last decision is redundant. By removing it, we obtain the
partial assignment ¬x5 @1, ¬x6 @1, x3 @2, x1 @2, x4 @2. Consider the decision assignments ¬x5 @1 and x3 @2, by which the other variables are
implied or can be assigned either value. By flipping those assignments,
we obtain the simplified blocking clause C6 = x5 ∨ ¬x3 .
4.1.3. Continuing Search. In Algorithm 2, whenever a solution is found,
a solver is enforced to backtrack to the root level. After that, due to
a variable selection heuristic, a different assignment will be examined
and a region of the search space in which a solution was just found
remains incomplete, which may give rise to unnecessary propagations
and conflicts in a later search for that region. Restart is, however,
essential in AllSAT solving in particular when simplification is used.
Indeed after simplification is performed and a new clause is added, the
3We
say that an assignment ν 0 is smaller than ν if all variables x assigned values
in ν are also assigned values in ν and their values coincide, i.e., ν(x) = ν 0 (x).
4Even computing a minimal satisfying assignment requires quadratic time, which
is still expensive.
0
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
13
state of implications such as which literals are decisions becomes inconsistent, and it is necessary to deduce implications again. It is not
straightforward to answer how to continue search [32].
The problem of over-canceling due to backtracking was addressed [47],
and a simple technique, called progress saving, that stores recent canceled decisions in an array and simulates them after backtracking was
proposed. Specifically, any time a solver enters into the decide stage, it
checks if an assignment to a selected variable is stored, and it simulates
the previous decision if exists; otherwise, it follows a default heuristic.
Although this technique was proposed in the context of SAT, it is also
applicable to AllSAT.
Example 8. Continuing Example 7, suppose that we added the blocking
clause C6 to ψ and backtracked to the root level with progress saving
enabled. At this point, all assignments above the root level are canceled,
yet the previous decisions (x5 , 0), (x3 , 1) are stored in an array. If x5
or x3 is selected, the previous decision is made again.
4.1.4. Implementation. We implemented 4 programs based on blocking procedure according to whether simplification and continuation
techniques are selected or not.
For a simplification technique, we used a method related to set covering model [44] and decision-based minimal satisfying cube [57]. For the
sake of efficiency, near-minimal satisfying assignments are computed.
A basic idea is that given a total assignment, we select as a small number of decision variables as make a CNF formula evaluate to 1. This is
done in the following way:
(1) select all decision variables that related to implications of at
least one variable;
(2) for each clause that is not satisfied by selected variables, select arbitrary decision variable that makes the current clause
satisfied.
A simplified satisfying assignment then consists of assignments to the
selected decision variables and all implied variables. By flipping the assignments to the selected decisions, we obtain a blocking clause, which
blocks all total assignments represented by the simplified assignment.
We implemented a continuation technique so that before backtracking at line 15, all decisions are stored in an array and after backtracking,
we simulate these decisions whenever possible in the order of their decision levels. Clearly, not all decisions are assumed due to blocking
clause, and conflict or contradiction to the previous decision will happen. It is this point where a solver continues search, and we will then
enter into the conflict resolution or the decide stage.
4.2. Non-blocking Solvers.
4.2.1. Overview. We give a basic idea for an AllSAT procedure without the aid of blocking clauses. Like blocking procedure, we modify
14
1
2
3
4
5
T. TODA AND T. SOH
ALGORITHM 3: Chronological backtracking in non-blocking procedure,
where δ(y) denotes the decision level of y.
Input: an assignment ν, the current decision level dl.
Output: updated objects ν, dl.
(x, v) ← the decision assignment of level dl;
ν ← {(y, w) ∈ ν | δ(y) ≤ dl − 1};
dl ← dl − 1;
ν ← ν ∪ {(x, v̄)}, where v̄ is the opposite value from v;
return ν, dl;
−1∗
−20∗ 19 9
−15
−8
12
−18∗ 2
−5 −17
−10 11 −14 13
−6
−3 7
4
16∗
(a) Solution Found
−1∗
−20∗
−8
−16∗
−7
19
12
−6
13
9
−15
18∗
11 2
−5 −3
(b) Conflict
−1∗
−20∗ 19 9
−15
−8
12 18∗ 6
−16 −3
(c) Conflict
Figure 4. Snapshots of a non-blocking solver’s state in
the three different cases (a),(b), and (c). Assignments
are given from left to right, top to bottom, separated by
line for each level. Integers i specify assignments such
that the variable x|i| is assigned the value 0 if i < 0
and 1 if i > 0. Decision assignments have an asterisk
as a superscript. Non-decision assignments with NULL
antecedent have an asterisk as a subscript.
Algorithm 1. A main feature is to employ chronological backtracking instead of non-chronological backtracking. The chronological backtracking used here is a bit different from the ordinary one described
in Section 3.2. As shown in Fig. 3, only differences are to insert the
flipped decision (x, v̄) in ν and register it to an implication graph so
that it has no incomming arc. This is because there is no reason that
implied flipped decisions because of the absence of blocking clauses.
We say that a literal, seen as an assignment, has NULL antecedent
if it has no reason that implied it and there is no incomming arc in
an implication graph. The chronological backtracking given above is
hereafter abbreviated as BT for convenience, and it is performed by
the function backtrack. We collectively call a number of procedures
for AllSAT solving based on BT non-blocking procedure in contrast
to blocking procedure presented in Section 4.1. If there is no fear of
confusion, chronological backtracking always means BT.
An important point in this approach is how we make BT compatible
with conflict-driven clause learning. Consider a conflict graph in clause
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
15
learning phase. Due to BT, a conflict graph may contain several roots,
i.e. assignments with NULL antecedent, in the same decision level. See
for example the literals ¬x20 and x18 in Fig. 5. Since an ordinary first
UIP scheme commonly assumes a unique root in the same decision level,
implementations based on that assumption get stuck in non-decision
literals with NULL antecedent. To resolve this problem, two techniques
are presented later.
Example 9. Look at Fig. 4. (a) All variables are assigned values without conflict, which means a solution is found. (b) Following BT, all
assignments of level 3 are canceled and the flipped decision x18 is inserted as a non-decision assignment at level 2. Since no propagation
takes place, a new decision ¬x16 is made and a subsequent propagation results in conflict. The decision is the only one that has NULL
antecedent in the current decision level, and ordinary first UIP scheme
suffices in this case. (c) From the conflict we have just met, the conflict
clause x6 ∨ x1 ∨ ¬x18 ∨ ¬x12 is learnt, and a solver backtracks to level
2. 5 The assignment x6 is implied by the conflict clause. A subsequent
propagation results in conflict again. This time, there is a non-decision
assignment with NULL antecedent in the same decision level.
A major advantage of non-blocking approach is that no matter how
many solutions exist, a performance of unit propagation does not deteriorate thanks to the absence of blocking clauses. Instead, it has to
find total satisfying assignments one by one. Hence, there is a limit in
the number of solutions to be generated in a realistic mount of time.
4.2.2. Sublevel-based First UIP Scheme. Grumberg et al. introduced
the notion of sublevels and presented a sublevel-based first UIP scheme
that is compatible with non-blocking approach [23]. A basic idea is to
divide a single decision level into sublevels. Specifically, a new sublevel
is defined whenever BT is performed, and sublevels are undefined as
their decision levels are undefined. An ordinary first UIP scheme can
then be applied in the current sublevel.
A conflict clause obtained by this approach may contain many literals that are below the current sublevel yet in the same decision level.
Among literals in that conflict clause, those with NULL antecedent are
necessary for avoiding rediscovery of solutions and can not be removed
if exist, however it is expected that other literals are reduced further.
4.2.3. Decision Level-based First UIP Scheme*. We present an alternative first UIP scheme that need not require sublevels. Our scheme
can be realized with a small modification: it is simply not to stop at
literals with NULL antecedent and attempt to find the first UIP in
5In
this case, there is no oppotunity to rediscover solutions by ordinary chronological backtracking, and hence the flipped decision x16 is not inserted. This technique is explained in more detail when non-chronological backtracking with level
limit is introduced.
16
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
T. TODA AND T. SOH
ALGORITHM 4: Non-blocking procedure with decision level-based first
UIP scheme.
Input: a CNF formula ψ, an empty variable assignment ν.
Output: all satisfying assignments.
dl ← 0;
/* Decision level */
lim ← 0;
/* Limit level */
while true do
ν ← propagate (ψ, ν);
/* Deduce stage */
if conflict happens then
if dl ≤ 0 then halt;
(ν, dl, lim) ← resolve (ψ, ν, dl, lim);
/* Resolve stage */
else
if all variables are assigned values then
report ν;
if dl ≤ 0 then halt;
(ν, dl) ← backtrack (ν, dl);
lim ← dl;
else
dl ← dl + 1;
/* Decide stage */
select an unassigned variable x and a value v;
ν ← ν ∪ {(x, v)};
end
end
end
the current decision level. A specific procedure is to traverse a conflict graph from κ in the reverse order of implications and construct a
conflict clause C, repeating the following procedure until the first UIP
appears:
(1) if the current literal is below the current decision level, add the
negated literal to C and do not go up through the incomming
arcs of the current literal;
(2) if the current literal has NULL antecedent, add the negated
literal to C;
(3) for the other case, go up through the incoming arcs of the current literal if their source vertices have not yet been visited.
After the first UIP is found, add the negated literal to C.
Compared to the sublevel-based first UIP scheme, the decision levelbased scheme might be considered better. First of all, it is simple. Secondly, a conflict clause contains unique literal from the current decision
level except for those with NULL antecedent. However, it should be
noted that conflict clauses obtained by the decision level-based scheme
are not necessarily smaller, because the unique implication point is further from κ and thus a conflict clause may contain more literals below
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
17
¬
¬
¬
¬
¬
¬
Figure 5. An implication graph for the conflict case (c)
in Fig. 4.
the current decision level. A pseudo code for the modified scheme is
omitted, because modification would be straightforward.
Example 10. Continuing Example 9, consider the conflict case (c)
in Fig. 4. Figure 5 illustrates the difference of the two schemes: the
sublevel-based scheme finds x6 as a first UIP and learns ¬x6 ∨ ¬x9 as
a conflict clause, while the decision level-based scheme finds ¬x20 as a
first UIP and learns x20 ∨ x1 ∨ ¬x18 as a conflict clause. Note that the
conflict clause in either case does not become unit after backtracking,
though it does not menace algorithmic correctness because of the flipped
decision x20 .
Remark 3. Recall that when the function analyze is performed, an
assignment to its unit literal is set to the current assignment function
ν. In non-blocking procedure, a conflict clause is not necessarily unit as
seen in Example 10. Hence, the function analyze in non-blocking procedure only adds a conflict clause, and it does not consider an induced
assignment.
Algorithm 4 is a pseudo code for the non-blocking approach using
clause learning with the decision level-based first UIP scheme. Recall
that the function backtrack is to backtrack chronologically, following
Algorithm 3. There are some choices for it, where a generic function,
named resolve, is called. A simple way of realizing the resolve stage is
to perform clause learning, based on either first UIP scheme and then
perform BT. More elaborate methods are introduced later, with which
resolve can also be replaced. Since the variable lim is used in one of
those methods, it is introduced when needed.
4.2.4. Conflict Directed Backjumping. Grumberg et al. augmented
non-blocking approach with conflict resolution by means of a restricted
non-chronological backtracking [23]. Their backtracking method can be
considered as a form of conflict directed backjumping (CBJ for short).
18
T. TODA AND T. SOH
CBJ has been studied as one of tree search algorithms for constraint
satisfaction problem [48] [13] [17].
A basic idea is described below. Consider the scenario where conflict
happens at decision level i+1, and after backtracking to level i, conflict
happens again. In this case, obtain a conflict clause cl1 from the former
conflict, while from the latter conflict, obtain a conflict clause cl2 so that
its UIP is ¬p, where p is the first UIP in the former analysis. Perform
resolution of cl1 and cl2 and obtain a resulting clause cl3 . After that,
backtrack to the level preceding the highest level in cl3 .
A pseudo code is given in Algorithm 5, which is almost faithfully
rephrased in our setting from the code given in the literature [23]. The
call of propagate at the end of the while loop was not explicitly written
in the original code, which we consider necessary. At this call, unit
propagation considers the effect of the recent flipped decision inserted
as a result of backtrack. Note that the effect of an assignment implied
by the recent conflict clause is not considered here, because our nonblocking procedure assumes that the function analysis simply records
a conflict clause and does not insert an implied assignment, which
is separately inserted at line 12. The halt means that non-blocking
procedure is halted too.
4.2.5. Non-chronological backtracking with Level Limit And Its Combination with CBJ*. We present an alternative conflict resolution by
means of non-chronological backtracking with backtrack level limit. To
our knowledge, this method was first presented by [22], though it was in
the context of answer set programming. We thus would like to import
their idea to our non-blocking procedure.
Since non-blocking procedure does not record blocking clauses, it
must not backtrack to arbitrary level, even though a backtracking level
is one that is legitimately derived from a conflict clause. However, we
can do this if there is no opportunity of rediscovering found solutions
from the derived level. We use the variable lim that holds a ”safe”
level to be backtracked, that is, the first level at which the current
assignment and the previous satisfying assignment differ.
A pseudo code is given in Algorithm 6. We call this approach nonchronological backtracking with level limit, denoted by BJ, where the
underline means a backtrack level is limited. Since lim is always less
than or equal to dl, BT is performed if and only if lim = dl. If lim < dl,
backtracking does not entail inserting a flipped decision (see Example 9). This is because lim < dl implies no solution yet found, and
hence no opportunity to rediscover solutions by backtracking.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
19
ALGORITHM 5: Conflict resolution based on conflict-directed backjumping
Input: a CNF formula ψ, an assignment ν, the current decision level dl.
Output: updated objects ν, dl.
stack ← an empty stack;
while true do
if conflict happens then
if dl ≤ 0 then halt;
analyze (ψ, ν);
push the learnt conflict clause into stack;
(ν, dl) ← backtrack (ν, dl);
else if stack is not empty then
cl1 ← the clause popped from stack;
if cl1 is a unit clause then
unit ← the unit literal in cl1 ;
add unit, seen as an assignment, to ν so that it has antecedent
cl1 ;
ν ← propagate (ψ, ν);
if conflict happens then
if dl ≤ 0 then halt;
cl2 ← the conflict clause from the recent conflict with unit
as UIP;
cl3 ← the resolution of cl1 and cl2 ;
push cl3 into stack;
bl ← the highest level in cl3 ;
(ν, dl) ← backtrack (ν, bl);
end
end
else
break;
end
ν ← propagate (ψ, ν);
end
return ν, dl;
We furthermore present a combination6 of BJ and CBJ. This is obtained by replacing the else part in BJ, in which BT is performed, with
CBJ: that is,
(1) if lim is less than the current decision level, then perform BJ;
(2) otherwise, perform CBJ.
6It
is mentioned [22] that a combination of conflict-directed backjumping and
non-chronological backtracking is proposed. However, their pseudo code almost
corresponds to Algorithm 6 and it is different from that stated in this paper.
20
1
2
3
4
5
6
7
8
9
10
T. TODA AND T. SOH
ALGORITHM 6: Conflict resolution based on non-chronological backtracking with level limit
Input: a CNF formula ψ, an assignment ν, the current decision level dl, a
level limit lim.
Output: updated objects ν, dl, lim.
bl ← analyze (ψ, ν);
/* Diagnose stage */
if lim < dl then
if bl < lim then bl = lim;
ν ← {(x, v) ∈ ν | δ(x) ≤ bl};
dl ← bl;
else
(ν, dl) ← backtrack (ν, dl);
lim ← dl;
end
return ν, dl, lim;
Step 2 is selected only when lim equals the current decision level:
in other words, when conflict happens just after the current assignment diverged from the previous satisfying assignment. Hence, BJ is
preferentially applied when one or more decisions from the diverging
point are made. This is designed with expectation that more decisions
are made from the diverging point, more effectively BJ prunes search
space. Since BJ is likely to be more frequently applied, this approach
is denoted by BJ+CBJ.
4.2.6. Implementation. We implemented 8 programs based on nonblocking procedure according to which of the two first UIP schemes
is selected and which of the conflict resolution methods, i.e. BT, BJ,
CBJ, or BJ+CBJ, is selected, where BT means performing BT after
clause learning.
4.3. Formula-BDD Caching Solvers.
4.3.1. Overview. Formula caching refers to a number of techniques
to memorize formulae to avoid recomputation of subproblems [7]. Examples include a caching technique in probabilistic planning [39], conflict clauses in SAT [6] [42], component caching and other cachings in
#SAT [5] [4], and blocking clauses in AllSAT [43].
Another type of formula caching in which formulae are associated
with propositional languages such as FBDD, OBDD, and a subset of
d-DNNF has been studied in the context of knowledge compilation [28].
Their work revealed a correspondence between exhaustive DPLL search
and propositional languages. They also proposed speeding up compilation by exploiting techniques of modern SAT solvers through the
correspondence. Although exhaustive DPLL search is simply used for
efficiency in their compilation approach, compilation in turn can contribute to speeding up exhaustive DPLL search. Actually, if a CNF
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
21
formula is compiled into a BDD, all satisfying assignments can be generated simply by traversing all possible paths from the root to the
sink node >. This seems like taking a long way around to AllSAT
solving, however thanks to the caching mechanism, recomputation of
many subproblems can be saved. A connection to AllSAT was mentioned, however their primary concern is on the compilation to suitable
languages for required queries, not restricted to AllSAT. To our knowledge, comparisons have been conducted only between various compilers. An application to an AllSAT solver itself was more explicitly
mentioned in the literature [55] and a compiler-based AllSAT solver
is released. However, comparisons with other AllSAT solvers have not
been conducted yet and its power remains unknown. Similar caching
techniques appear in other areas such as preimage computation in unbounded model checking [51] [37] [34], satisfiability [45], and discrete
optimization [3] [8].
The paper only deals with the caching method that records pairs of
formulae and OBDDs, which we call formula-BDD caching. A formulaBDD caching can be embedded in either blocking procedure or nonblocking procedure7. This is done without almost any loss of optimizations employed in an underlying procedure. An exception is that
variables must be selected in a fixed order at the decide stage. This
effect is far from negligible in terms of efficiency, as is well-recognized
in a single solution SAT. It is, however, confirmed in experiments that
formula-BDD caching solvers exhibit a quite good performance on the
whole, and it provides an efficient solution method for instances that
have a huge number of solutions and can not possibly be solved by
other means.
4.3.2. Caching Mechanism. We give a basic idea of formula-BDD caching,
using a simple BDD construction method with formula-BDD caching.
This method is elaborated later by implementing on top of a SAT
solver. We first introduce terminology. A subinstance of ψ in an assignment ν is the CNF formula derived from ψ by applying all assignments
defined in ν to ψ. The current subinstance refers to the subinstance
induced by the current assignment.
Consider the following procedure with a CNF formula ψ and an
empty variable assignment ν as initial arguments:
(1) if an unsatisfied clause exists in the current subinstance, then
return ⊥;
(2) if all variables are assigned values, then return >;
(3) i ← the smallest index of an unassigned variable;
(4) f0 ← the result obtained by a recursive call with ψ and ν ∪
{(xi , 0)};
7Only
the combination with blocking procedure has been presented in the past
work [27] [55].
22
T. TODA AND T. SOH
(5) f1 ← the result obtained by a recursive call with ψ and ν ∪
{(xi , 1)};
(6) return a node with the label i, the references to LO child f0 and
to HI child f1 ;
Since different assignments can yield subinstances that are logically
equivalent, we want to speed up the procedure by applying dynamic
programming. To do this, we need to quickly decide if the current
subinstance is solved. If it is unsolved, we compute a BDD for all
solutions of the instance and memorize it, associating with the instance.
Otherwise, the result is obtained in a form of BDD and recomputation
is avoided.
However, this approach involves the equivalence test of CNF formulae, which is computationally intractable, as it includes satisfiability
testing. Hence, we consider a weaker equivalence test. That is, we
encode subinstances into formulae so that if two subinstances are not
logically equivalent, then the encoded formulae are not identical. To decide if the current subinstance is solved, it suffices to search the encoded
formula in the set of registered formula-BDD pairs. All requirements
for formula-BDD caching to work is simply the sentence in italic above,
and any encoding that meets it will do. It should be noted that our
test is sound in that acceptance always is a correct decision, however
if we prioritize efficiency of encoding excessively, logically equivalent
subinstances are very likely to result in non-identical formulae, i.e., a
wrong decision.
Examples of formula-BDD cachings include those induced by cutsets
and separators [27], defined below, and a variant of cutsets [55].
Definition 1. The i-th cutset of a CNF formula ψ is the set of clauses
C in ψ such that C has literals with their underlying variables xj and
xk satisfying j ≤ i < k. The cutwidth of ψ is the maximum size of a
cutset of ψ.
Definition 2. The i-th separator of a CNF formula ψ is the set of
variables xj such that some clause in the i-th cutset has a literal with
its underlying variable xj satisfying j ≤ i. The pathwidth of ψ is the
maximum size of a separtor of ψ.
Example 11. Look at the CNF formula ψ illustrated in Fig. 6. The
3-rd cutset of ψ consists of C2 and C3 , while the 3-rd separator of ψ
consists of x1 , x2 , and x3 . The cutwidth and the pathwidth of ψ are
both 3.
The following proposition states that clauses (and variables) in cutsets (and separators) meet the requirement of formula-BDD caching,
respectively. Proof is omitted. See [27].
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
23
Figure 6. A CNF formula (left )and an OBDD for the
CNF (right), where arcs to the sink node ⊥ is omitted. Cutsets are associated with arcs, where underlined
clauses mean they are satisfied.
Proposition 1. Let ψ be a CNF formula, where variables are ordered
according to their indices. Let ν and µ be assignments with the i-th or
less variables assigned values and other variables unassigned.
(1) If satisfied clauses in the i-th cutset of ψ in ν coincide with
those in µ, then the subinstance in ν is logically equivalent to
that in µ.
(2) If variables assigned the value 1 in the i-th separator of ψ in ν
coincide with those in µ, then the subinstance in ν is logically
equivalent to that in µ.
Example 12. Figure 6 illustrates an OBDD constructed using cutsets
as formula-BDD caching. Cutsets are associated with arcs, and satisfied
clauses in them are underlined. If two arcs have the same set of satisfied
clauses, Proposition 1 implies that their target vertices can be merged
safely.
It should be noted that our weaker equivalence test may reject logically equivalent subintances; the subgraphs in a constructed BDD that
correspond to those subinstances are not merged. This means that
a constructed BDD is not fully reduced, i.e., an OBDD, though the
OBDD give in Fig. 6 happens to be fully reduced.
An important point of the formula-BDD caching approach is how
to balance quality with efficiency in our weaker equivalence test. The
quality here refers to how many correct decisions are made. Theoretically, it holds that a correct decision of the separator approach always
implies a correct decision of the cutset approach. On the other hand,
the efficiency refers to how much time is taken to create formulae from
24
T. TODA AND T. SOH
subinstances, which substantially amounts to evaluating clauses and
variables in cutsets and separators, respectively. In terms of efficiency,
evaluating clauses in a cutset would require time linear to the total size
of clauses due to lazy evaluation mechanism if it is implemented on top
of modern SAT solvers. On the other hand, evaluating variables in a
separator requires time linear to the number of variables.
From the argument above, we can say that for instances with small
cutwidth, evaluation cost of cutset is negligible compared to separator
and hence cutset is a better choice, and for instances with many clauses,
separator should be used instead.
4.3.3. Embedding Formula-BDD Caching in AllSAT Procedure*. We
demonstrate how to embed formula-BDD caching in a concrete AllSAT
procedure. We take non-blocking procedure as an example. To our
knowledge, this is the first time that the combination is presented. The
other combination, i.e. blocking procedure with formula-BDD caching,
is omitted.
We have assumed so far that conflict clauses (and blocking clauses
in blocking procedure) are added to a CNF ψ, but from now on we
will assume that they are separately maintained from ψ. This makes
ψ unchanged throughout the execution of our AllSAT procedure. Accordingly, the cutset and the separator in each level are unchanged
too.
Algorithm 7 is a pseudo code of non-blocking procedure with formulaBDD caching embedded. Formula-BDD caching mechanism consists of
the encode stage, the extend stage, and the enroll stage.
At the encode stage, the function makef ormula receives a CNF ψ,
the current assignment ν, and the index i − 1 of the largest assigned
variable. It computes a formula for the current subinstance. No specific
encoding is presented here. See Section 4.3.2 for examples of encoding.
Note that if all variables are assigned values, then i must be ∞, and
hence we have i − 1 = ∞. In this case, let the function makef ormula
return 1, which is the special formula representing true.
At line 15, we search an entry with the key (i − 1, φ) in S, where
S holds registered formula-BDD pairs. If it exists, the current subinstance is already solved, and the result is the OBDD g associated with
the key, which appears as a subgraph of f . Hence, at the extend stage,
the function extendobdd augments an OBDD f by adding a path from
the root of f to the root of g, following the current assignment ν. It
returns the pair of the extended OBDD and the added path. Since this
stage is straightforward, we omit a pseudo code of it.
At the enroll stage, we associate formulae for solved subinstances
with the corresponding OBDDs and insert these formula-BDD pairs to
S. To do this, important points are how to identify when subinstances
are solved and how to find their OBDDs. Let I be a subinstance
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
25
with the smallest unassigned variable xi . Since unit propagation is
performed at the beginning of each repetition, without loss of generality
we assume that xi is a decision variable. This means that a formula
ζ has been made from I, thereby (i − 1, ζ) is in T . Let νI and dlI be
the assignment and the decision level at this point. Without loss of
generality we assume that there exist one or more solutions extending
νI , because otherwise, the OBDD for I is not created. After adding at
least one path to f , we have the node that corresponds to the decision
at variable xi . Clearly, this node is reachable from the root of f through
the path following νI , and it is the root of the OBDD for I. When all
solutions for I are found, to go out of the exhausted search space,
backtracking to a lower level than dlI is performed, which is directly
triggered by an occurrence of conflict implying no solution left or by
the discovery of the last solution. Backtracking to a lower level can also
be performed without exhausting all solutions, however this case only
happens when no solution of I is yet found. Hence, we can distinguish
them. Summarizing the above, if a backtracking level is less than dlI
and at least one solution for I is found, then I is solved, and the root
of the OBDD for I is located at the end of the path following νI , which
is a part of the assignment just before backtracking.
The function associate is in charge of the enroll stage. With the
observation above, it is called whenever backtracking is performed. A
procedure for it is given as follows. We scan all nodes h in the most
recently added path π until the assignment at h taken in π contradicts
to the current assignment. Note that scanned nodes h correspond to
subinstances Ih such that Ih turns out to be satisfiable and the assignment that induced Ih is a part of the current assignment. For each
scanned node h, we test if a backtracking level is less than the decision
level of h and a formula ζ was made8 from Ih . If the test is passed, the
pair of (j − 1, ζ) and h is inserted into S, where j is the label of h.
The function resolveplus behaves in the same way as reslove except
that S and T are updated each time backtracking is performed, as in
lines 19-20.
Example 13. Figure 7 illustrates how Algorithm 7 constructs an OBDD
for the CNF formula ψ given in Fig. 6.
4.3.4. Refreshing OBDDs*. A constructed OBDD may become too
large to be stored on memory, though it would be in practice much
better in size than the list representation of all solutions and in the
worst case equal to it except for constant factor. We present a simple
technique to resolve this problem. Let n be the number of variables.
We introduce a threshold θ of an OBDD size, where θ > n. Insert
the following procedure after an BDD is extended and backtracking is
8
This is equivalent to finding an entry (j − 1, ζ) in T , where j is the label of h.
26
T. TODA AND T. SOH
Figure 7. Progress of OBDD construction. As paths
are added one by one, an OBDD is augmented from left
to right. In each step, thick arcs represent a path that
is about to be added, and gray nodes mean that the
subintances corresponding to them were solved.
performed. If the size of an OBDD is larger than or equal to θ − n,
then the current OBDD f is dumped to a file in a secondary storage,
and all objects f , S, T , and π of formula-BDD caching mechanism are
refreshed with initial states. Since formula-BDD caching is almost independent of the underlying non-blocking procedure, after refreshing,
the procedure simply attempts to examine unprocessed assignments
with formula-BDD caching empty.
4.3.5. Implementation. We implemented 2 programs on each nonblocking procedure and 2 programs on each blocking procedure according to which formula-BDD caching is selected, cutset or separator.
5. Experiments
5.1. Implementation And Environment. All solvers are implemented
in C on top of MiniSat-C v1.14.1 [19]. Clasp 3.1.2 was taken from
Potassco project9. As far as we are aware, clasp [21], picosat [10], and
relsat [6] are only SAT solvers which support the enumeration of all
solutions. Among them, we used clasp for the comparison, because it
achieved better performance than picosat, and relsat does not support
quiet mode in solution generation and generated solutions may be too
large to be stored.
All experiments were performed on 2.13GHz Xeon R E7- 2830 with
512GB RAM, running Red Hat Enterprise Linux 6.3 with gcc compiler
version 4.4.7. In the execution of each AllSAT solver, time limit and
memory limit were set to 600 seconds and 50GB, respectively. If either
limit is exceeded, the solver is enforced to halt. All solvers simply touch
found solutions and never output them.
9Potassco,
the Potsdam Answer Set Solving Collection, bundles tools for Answer Set Programming developed at the University of Potsdam,http://potassco.
sourceforge.net/, accessed on 13rd Sept., 2015.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
27
ALGORITHM 7: Non-blocking procedure with formula-BDD caching,
where δ(x) denotes the decision level of a variable x.
Input: a CNF formula ψ, an empty variable assignment ν.
Output: an OBDD f for all satisfying assignments.
dl ← 0;
/* Decision level */
lim ← 0;
/* Limit level */
f ← ⊥;
/* OBDD */
S ← {((∞, 1), >)};
/* Set of formula-BDD pairs */
T ← ∅;
/* Set of formulae with indices */
π ← ;
/* Sequence of BDD nodes */
while true do
ν ← propagate (ψ, ν);
/* Deduce stage */
if conflict happens then
if dl ≤ 0 then return f ;
(ν, dl, lim, S, T ) ← resolveplus (ψ, ν, dl, lim, S, T ); /* Resolve and
enroll stage */
else
i ← min {j | xj is not assigned value};
φ ← makeformula (ψ, ν, i − 1);
/* Encode stage */
if an entry with key (i − 1, φ) exists in S then
g ← the OBDD node associated with (i − 1, φ) in S;
(f, π) ← extendobdd (f, g, ν);
/* Extend stage */
if dl ≤ 0 then return f ;
S ← associate (π, ν, S, T, dl − 1);
/* Enroll stage */
T ← {(j − 1, ζ) ∈ T | δ(xj ) ≤ dl − 1};
(ν, dl) ← backtrack (ν, dl);
lim ← dl;
else
T ← T ∪ {(i − 1, φ)};
dl ← dl + 1;
/* Decide stage */
select a value v;
ν ← ν ∪ {(xi , v)};
end
end
end
The types of compared solvers are a blocking solver, a non-blocking
solver, a formula-BDD caching solver, and clasp. The first three types
have some variations according to which techniques are used (see the
end of each subsection in Section 4). Among solvers of the same type,
we selected a solver with the most solved instances. The selected
solvers, called representative solvers, are as follows.
• Blocking NoSimple Cont: the blocking solver with simplification unselected and continuation selected.
28
T. TODA AND T. SOH
• NonBlocking DLevel BJ: the non-blocking solver with decision level first UIP scheme and non-chronological backtracking
with level limit both selected.
• BDD Cut NonBlocking DLevel BJ: the formula-BDD caching
solver with cutset caching selected and it is implemented on top
of NonBlocking DLevel BJ.
Throughout the section, if there is no fear of confusion, they are abbreviated as Blocking, NonBLocking, and BDD. Notation for solvers
with other configurations is introduced in the same way.
It is known that variable orderings significantly affect the performance of BDD compilation. Hence, we used the software MINCE
version 1.0 [2] to decide a static variable order before the execution
of formula-BDD caching solvers. The execution of MINCE failed for
some instances, and for that case, we used the original order. The
time required for deciding a variable order is included. Although for
some instances time limit exceeded in preprocessing, it was negligible
for many instances.
5.2. Problem Instances. We used total 2867 CNF instances (all satisfiable), which are classified as follows.
• satlib: SATLIB benchmark problems (2707 instances), taken
from SATLIB website10.
• sc14: SAT Competition 2014 benchmarks, application track
(56 instances) and crafted track (65 instances), taken from SAT
Competition 2014 website11.
• iscas: ISCAS85 and 89 circuit benchmarks in DIMACS CNF
format (39 instances), taken from TG-Pro website12.
Among instances released in each repository, we selected all instances
such that satisfiability could be decided in 600 seconds by either one
of the SAT solvers clasp 3.1.2, glucose4, minisat 2.2, and minisat 1.3,
and its result was satisfiable. For satlib and sc14, random instances
are excluded.
5.3. Comparison of Running Time. Figure 8 shows a cactus plot of
representative solvers. Solved instances are ranked with respect to the
times required to solve them. Each point represents a solved instance
with its rank (the horizontal coordinate) and the required time (the
vertical coordinate). Since one wants to solve as many instances as
10SATLIB
- The Satisfiability Library, by Holger H Hoos and Thomas Stützle
at Darmstadt University of Technology, http://www.cs.ubc.ca/~hoos/SATLIB/
benchm.html, accessed on 16th May, 2014.
11SAT
Competition
2014,
http://www.satcompetition.org/2014/
description.shtml, accessed on 8th Sept., 2015.
12TG-Pro - A SAT-based ATPG System, by Huan Chen and Joao MarquesSilva,
http://logos.ucd.ie/web/lib/exe/fetch.php?media=benchmarks:
circuit-cnf-tgpro-20110317.tgz, accessed on 16th May, 2014
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
29
600
clasp
NonBlocking_DLevel_BJ
Blocking_NoSimple_Cont
BDD_Cut_NonBlocking_DLevel_BJ
500
Time (sec)
400
300
200
100
0
1000
1200
1400
1600
1800
2000
2200
2400
#Solved
Figure 8. Cactus plot of representative solvers with respect to running time.
possible in a given amount of time, it is thought that gentler the slope
of plotted points is, more efficient a solver is. The formula-BDD caching
solver clearly outperforms the other solvers. It is then followed by the
non-blocking solver, clasp, and the blocking solver in this order.
Figures 9, 10, and 11 depict differences between solvers of the same
types. From Figure 9, we can observe that continuation of search is
effective yet simplification degrades performance. For some instances,
simplification enables a solver to find a large number of solutions, however such instances are limited and the current implementation is not
powerful enough to make it possible to solve instances that can not be
handled without simplification. Figure 10 has a narrower horizontal
range than the other figures. This is because non-blocking solvers exhibit quite similar performance and they can not be distinguished otherwise. It is surprising that BT is almost as efficient as the other elaborated backtracking methods. Decision level-based scheme is equal to or
more efficient than sublevel-based scheme. Figure 11 shows that nonblocking procedure is clearly better as an underlying solver in which
caching mechanism is embedded, while there is almost no difference
between caching methods.
5.4. Comparison of Maximum Memory Usage. Figure 12 shows
a cactus plot of the maximum memory usage. Solved instances are in
turn ranked with respect to the maximum memory usage. Each point
then represents a solved instance with its rank (the horizontal coordinate) and the required memory (the vertical coordinate). In terms
of memory consumption, the formula-BDD caching solver is the worst,
while the non-blocking solver and clasp exhibit a stable performance.
30
T. TODA AND T. SOH
600
Blocking_NoSimple_Cont
Blocking_NoSimple_NoCont
Blocking_Simple_Cont
Blocking_Simple_NoCont
500
Time (sec)
400
300
200
100
0
1000
1200
1400
1600
1800
2000
2200
2400
#Solved
Figure 9. Cactus plot of blocking solvers.
600
500
NonBlocking_DLevel_BT
NonBlocking_DLevel_BJ
NonBlocking_DLevel_CBJ
NonBlocking_DLevel_BJCBJ
NonBlocking_SubLevel_BT
NonBlocking_SubLevel_BJ
NonBlocking_SubLevel_CBJ
NonBlocking_SubLevel_BJCBJ
Time (sec)
400
300
200
100
0
1860
1880
1900
1920
1940
1960
#Solved
Figure 10. Cactus plot of non-blocking solvers, where
the horizontal scale is narrowed to make the difference
clear.
The rapid increase in the curves of the non-blocking solver and clasp is
due to large CNF formulae. Although the formula-BDD caching solver
consumes much memory, these days it is not unusual that even laptop computers have several giga bytes of RAM, and advantage of the
formula-BDD caching solver is not impaired so much.
5.5. Comparison of Scalability in The Number of Solutions.
As shown in Table 1, each representative solver has the following limit
in the number of solutions within 600 seconds time limit .
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
31
600
BDD_Cut_Blocking
BDD_Cut_NonBlocking_DLevel_BJ
BDD_Sep_Blocking
BDD_Sep_NonBlocking_DLevel_BJ
500
Time (sec)
400
300
200
100
0
1000
1200
1400
1600
1800
2000
2200
2400
#Solved
Figure 11. Cactus plot of formula-BDD caching solvers.
Maximum Memory Usage (KB)
1e+08
clasp
NonBlocking_DLevel_BJ
Blocking_NoSimple_Cont
BDD_Sep_NonBlocking_DLevel_BJ
1e+07
1e+06
100000
10000
1000
100
1000
1200
1400
1600
1800
2000
2200
2400
#Solved
Figure 12. Cactus plot of representative solvers with
respect to the maximum memory usage, where the vertical axis is in logarithmic scale.
•
•
•
•
Blocking: one million solutions.
Clasp: one hundred million solutions.
Non-Blocking: ten billion solutions.
BDD: more than one quadrillion solutions.
5.6. Distribution of Solved Instances over Instance Series. Table 2 shows the distribution of solved instances over all instance series.
32
T. TODA AND T. SOH
Table 1. Distribution of solved instances with respect
to the number of solutions.
(0, 101 )
[101 , 102 )
[102 , 103 )
[103 , 104 )
[104 , 105 )
[105 , 106 )
[106 , 107 )
[107 , 108 )
[108 , 109 )
[109 , 1010 )
[1010 , 1011 )
[1011 , 1012 )
[1012 , 1013 )
[1013 , 1014 )
[1014 , ∞)
total
Blocking clasp NonBlocking BDD
38
36
39
37
11
11
11
11
417
417
417
416
682
682
682
682
408
408
408
408
82
134
133
134
0
86
82
86
0
56
87
91
0
0
68
83
0
0
23
92
0
0
0
83
0
0
0
44
0
0
0
29
0
0
0
19
0
0
0
8
1,638 1,830
1,950 2,223
In almost all series, differences in the number of solved instances between solvers can be explained by their scalability.
Exceptions are the sc14 instances. They are clearly harder instances.
All solvers were unable to find even one solution in many instances. For
sc14-crafted series, solved instances are those of less than 10 solutions
with a few exceptions of BDD. Table 3 shows the distribution of all
sc14 instances including unsolved instances. Although clasp could find
relatively many solutions, the other solvers could only find less than
10 solutions for a majority of instances. In terms of the ability to find
solutions for sc14 instances, we could say that the best solver is clasp;
the blocking solvers and the non-blocking solvers are a tie, both ranked
at the second; the formula-BDD caching solver is no match for those
instances, which is due to fixed variable ordering.
The favorite ranges of instances for representative solvers are illustrated in Fig. 13 according to the two factors: hardness of instances
and the numbers of solutions instances have. Each solver is placed in
such a way that the vertical position corresponds to the ability to find
solutions for sc14 instances, where the blocking solver is above the
non-blocking solver because of more benefit from a SAT solver, and
the horizontal position corresponds to the scalability in the number of
solutions. Hence, each indicated range refers to instances leftward or
downward from it as well as those within it. It should be noted that
shrinking the vertical axis would be more suitable for real performance,
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
33
Table 2. Distribution of solved instances over instance
series, where the number of instances in each series is
enclosed in parenthesis.
ais
(4)
bmc
(13)
bw
(7)
Flat125-301
(100)
Flat150-360
(101)
Flat175-417
(100)
Flat200-479
(100)
Flat75-180
(100)
flat
(1,199)
gcp
(1)
hanoi
(2)
inductive
(41)
logistics
(4)
parity
(20)
qg
(10)
ssa
(4)
SW100-8-0
(100)
SW100-8-1
(100)
SW100-8-2
(100)
SW100-8-3
(100)
SW100-8-4
(100)
SW100-8-5
(100)
SW100-8-6
(100)
SW100-8-7
(100)
SW100-8-8
(100)
SW100-8-p0
(1)
sc14-app
(56)
sc14-crafted
(65)
iscas
(39)
total
(2,867)
Blocking clasp NonBlocking BDD
4
4
4
4
0
0
0
1
7
7
7
6
4
51
94
100
0
12
61
101
0
2
15
98
0
0
3
77
87
100
100
100
1,136 1,195
1,199 1,199
0
0
0
0
2
2
2
2
1
5
9
8
0
0
0
1
20
20
20
20
10
10
10
10
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
5
68
55
100
95
100
100
100
100
100
100
100
100
100
100
100
100
100
1
1
1
1
0
0
0
0
5
3
6
5
6
17
19
22
1,638 1,830
1,950 2,223
because in reality all solvers, on the whole, exhibits a poor performance
over hard instances.
5.7. Comparisons to Publicly Available #SAT Solvers. We also
have a comparison with publicly available #SAT solvers, sharpSAT
version 1.1 [52], c2d version 2.20 [16] and relsat version 2.20 [6], to
check the performance of our developed ALLSAT solvers for #SAT.
In the result for the same 2867 instances, relsat, sharpSAT, c2d and
BDD Cut NonBlocking DLevel BJ solved 2191, 2196, 2196 and 2223
34
T. TODA AND T. SOH
Table 3. Distribution of all sc14 instances including unsolved instances according to the number of found solutions within time limit.
[0, 0]
(0, 101 )
[101 , 102 )
[102 , 103 )
[103 , 104 )
[104 , 105 )
[105 , 106 )
[106 , 107 )
[107 , 108 )
[108 , 109 )
[109 , 1010 )
[1010 , 1011 )
[1011 , 1012 )
[1012 , 1013 )
[1013 , 1014 )
[1014 , ∞)
total
Blocking clasp NonBlocking BDD
54
22
58
89
25
24
22
19
8
0
0
0
12
1
0
0
3
5
1
0
4
20
1
0
15
27
2
0
0
21
0
2
0
1
0
0
0
0
14
0
0
0
23
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
9
121 121
121
121
Figure 13. Favorite ranges of instances for representative solvers.
respectively. Since relsat supports solution count, it was used as a
solution counter in this comparison. So, even for #SAT, the formulaBDD caching solver shows better performance.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
35
6. Conclusion
We surveyed and discussed major techniques of existing AllSAT
solvers. We classified the types of solvers into a blocking solver, a nonblocking solver, and a formula-BDD caching solver. We faithfully implemented and released these solvers publicly so that other researchers
can easily develop their solver by modifying our codes and compare
it with existing methods. We conducted comprehensive experiments
with total 2867 instances taken from SATLIB, SAT competition 2014,
and ISCAS benchmarks. Apart from our implemented solvers, we used
clasp, one of the few off-the-shelf softwares with solution generation
support. The experiments revealed the following solver’s characteristics (600 seconds time limit). See also Fig. 13.
• The formula-BDD caching solver is the most powerful. It has
the most solved instances, including instances with more than
one quadrillion solutions. The maximum memory usage amounts
to several tens giga bytes in the worst case, though it is controllable by refreshing caches at the cost of a low cache hit rate.
They are bad at hard instances due to fixed variable ordering.
• The non-blocking solver is ranked at the next best, followed by
clasp. The non-blocking solver and clasp can handle instances
with ten billion solutions and one hundred million solutions
with a low maximum memory usage (a few mega bytes to several
tens of mega bytes), respectively. Although both solvers exhibit
relatively a similar performance, a difference is that clasp is able
to find a moderately many number of solutions from even hard
instances, though it is not powerful enough to make it possible
to solve instances that can not be handled by other means.
• The blocking solver is limited to instances with one million solutions as blocking clauses deteriorate the performance of unit
propagation. However, it can benefit from state-of-the-art techniques of SAT solvers as they are, thereby it is suitable for
finding a small number of solutions for hard instances.
From the above, we conclude that the formula-BDD caching solver is
the most superior in terms of exact AllSAT solving over various kinds
of instances. However, since not all solutions are necessary in some
practical applications and duplicated solutions may be allowed, it is
recommended to select an appropriate solver in accordance with types
of instances and applications.
References
[1] S. B. Akers. Binary decision diagrams. IEEE Trans. Comput.,
27(6):509–516, June 1978.
36
T. TODA AND T. SOH
[2] Fadi A. Aloul, Igor L. Markov, and Karem A. Sakallah. MINCE: A
static global variable-ordering heuristic for SAT search and BDD
manipulation. J. UCS, 10(12):1562–1596, 2004.
[3] H.R. Andersen, T. Hadzic, J.N. Hooker, and P. Tiedemann. A constraint store based on multivalued decision diagrams. In Christian
Bessiére, editor, Principles and Practice of Constraint Programming, volume 4741 of Lecture Notes in Computer Science, pages
118–132. Springer Berlin Heidelberg, 2007.
[4] F. Bacchus, S. Dalmao, and T. Pitassi. Algorithms and complexity
results for #sat and bayesian inference. In Foundations of Computer Science, 2003. Proceedings. 44th Annual IEEE Symposium
on, pages 340–351, Oct 2003.
[5] Roberto J. Bayardo and J. D. Pehoushek. Counting models using
connected components. In In Proceedings of the AAAI National
Conference, pages 157–162, 2000.
[6] Roberto J. Bayardo, Jr. and Robert C. Schrag. Using csp lookback techniques to solve real-world sat instances. In Proceedings of
the Fourteenth National Conference on Artificial Intelligence and
Ninth Conference on Innovative Applications of Artificial Intelligence, AAAI’97/IAAI’97, pages 203–208. AAAI Press, 1997.
[7] Paul Beame, Russell Impagliazzo, Toniann Pitassi, and Nathan
Segerlind. Formula caching in dpll. ACM Trans. Comput. Theory,
1(3):9:1–9:33, March 2010.
[8] D. Bergman, A. A. Cire, W.-J. van Hoeve, and J. N. Hooker.
Discrete optimization with decision diagrams. INFORMS Journal
on Computing. To appear.
[9] A. Biere, A. Biere, M. Heule, H. van Maaren, and T. Walsh. Handbook of Satisfiability: Volume 185 Frontiers in Artificial Intelligence and Applications. IOS Press, Amsterdam, The Netherlands,
The Netherlands, 2009.
[10] Armin Biere. Picosat essentials. JSAT, 4(2-4):75–97, 2008.
[11] Jrg Brauer, Andy King, and Jael Kriener. Existential quantification as incremental sat. In Ganesh Gopalakrishnan and Shaz
Qadeer, editors, Computer Aided Verification, volume 6806 of Lecture Notes in Computer Science, pages 191–207. Springer Berlin
Heidelberg, 2011.
[12] R.E. Bryant. Graph-based algorithms for boolean function manipulation. Computers, IEEE Transactions on, C-35(8):677–691,
Aug 1986.
[13] Xinguang Chen and Peter van Beek. Conflict-directed backjumping revisited. J. Artif. Int. Res., 14(1):53–81, March 2001.
[14] Edmund Clarke, Daniel Kroening, Natasha Sharygina, and Karen
Yorav. Predicate abstraction of ansi-c programs using sat. Formal
Methods in System Design, 25(2-3):105–127, 2004.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
37
[15] Y. Crama and P.L. Hammer. Boolean Functions: Theory, Algorithms, and Applications. Encyclopedia of Mathematics and its
Applications. Cambridge University Press, 2011.
[16] Adnan Darwiche. New advances in compiling CNF into decomposable negation normal form. In Proceedings of the 16th Eureopean
Conference on Artificial Intelligence, ECAI’2004, including Prestigious Applicants of Intelligent Systems, PAIS 2004, Valencia,
Spain, August 22-27, 2004, pages 328–332, 2004.
[17] Rina Dechter and Daniel Frost. Backjump-based backtracking for
constraint satisfaction problems. Artificial Intelligence, 136(2):147
– 188, 2002.
[18] Thomas Eiter, Kazuhisa Makino, and Georg Gottlob. Computational aspects of monotone dualization: A brief survey. Discrete
Applied Mathematics, 156(11):2035 – 2049, 2008. In Memory of
Leonid Khachiyan (1952 - 2005 ).
[19] Niklas En and Niklas Srensson. An extensible sat-solver. In Enrico Giunchiglia and Armando Tacchella, editors, Theory and Applications of Satisfiability Testing, volume 2919 of Lecture Notes
in Computer Science, pages 502–518. Springer Berlin Heidelberg,
2004.
[20] M.K. Ganai, A. Gupta, and P. Ashar. Efficient sat-based unbounded symbolic model checking using circuit cofactoring. In
Computer Aided Design, 2004. ICCAD-2004. IEEE/ACM International Conference on, pages 510–517, Nov 2004.
[21] Martin Gebser, Benjamin Kaufmann, André Neumann, and
Torsten Schaub. clasp: A conflict-driven answer set solver. In
Chitta Baral, Gerhard Brewka, and John Schlipf, editors, Logic
Programming and Nonmonotonic Reasoning, volume 4483 of Lecture Notes in Computer Science, pages 260–265. Springer Berlin
Heidelberg, 2007.
[22] Martin Gebser, Benjamin Kaufmann, André Neumann, and
Torsten Schaub. Conflict-driven answer set enumeration. In Chitta
Baral, Gerhard Brewka, and John Schlipf, editors, Logic Programming and Nonmonotonic Reasoning, volume 4483 of Lecture Notes
in Computer Science, pages 136–148. Springer Berlin Heidelberg,
2007.
[23] Orna Grumberg, Assaf Schuster, and Avi Yadgar. Memory efficient all-solutions sat solver and its application for reachability
analysis. In AlanJ. Hu and AndrewK. Martin, editors, Formal
Methods in Computer-Aided Design, volume 3312 of Lecture Notes
in Computer Science, pages 275–289. Springer Berlin Heidelberg,
2004.
[24] Tias Guns, Siegfried Nijssen, and Luc De Raedt. Itemset mining: A constraint programming perspective. Artificial Intelligence,
175(1213):1951 – 1983, 2011.
38
T. TODA AND T. SOH
[25] Aarti Gupta, Zijiang Yang, Pranav Ashar, and Anubhav Gupta.
Sat-based image computation with application in reachability
analysis. In Jr. Hunt, WarrenA. and StevenD. Johnson, editors,
Formal Methods in Computer-Aided Design, volume 1954 of Lecture Notes in Computer Science, pages 391–408. Springer Berlin
Heidelberg, 2000.
[26] Jiawei Han, Hong Cheng, Dong Xin, and Xifeng Yan. Frequent
pattern mining: current status and future directions. Data Mining
and Knowledge Discovery, 15(1):55–86, 2007.
[27] Jinbo Huang and Adnan Darwiche. Using dpll for efficient obdd
construction. In HolgerH. Hoos and DavidG. Mitchell, editors,
Theory and Applications of Satisfiability Testing, volume 3542
of Lecture Notes in Computer Science, pages 157–172. Springer
Berlin Heidelberg, 2005.
[28] Jinbo Huang and Adnan Darwiche. The language of search. J.
Artif. Intell. Res. (JAIR), 29:191–219, 2007.
[29] Saı̈d Jabbour, Jerry Lonlac, Lakhdar Sais, and Yakoub Salhi. Extending modern SAT solvers for models enumeration. In Proceedings of the 15th IEEE International Conference on Information
Reuse and Integration, IRI 2014, Redwood City, CA, USA, August 13-15, 2014, pages 803–810, 2014.
[30] Said Jabbour, Joao Marques-Silva, Lakhdar Sais, and Yakoub
Salhi. Enumerating prime implicants of propositional formulae
in conjunctive normal form. In Eduardo Fermé and João Leite,
editors, Logics in Artificial Intelligence, volume 8761 of Lecture
Notes in Computer Science, pages 152–165. Springer International
Publishing, 2014.
[31] Said Jabbour, Lakhdar Sais, and Yakoub Salhi. Boolean satisfiability for sequence mining. In Proceedings of the 22Nd ACM International Conference on Conference on Information; Knowledge
Management, CIKM ’13, pages 649–658, New York, NY, USA,
2013. ACM.
[32] HoonSang Jin, HyoJung Han, and Fabio Somenzi. Efficient conflict analysis for finding all satisfying assignments of a boolean
circuit. In Nicolas Halbwachs and LenoreD. Zuck, editors, Tools
and Algorithms for the Construction and Analysis of Systems, volume 3440 of Lecture Notes in Computer Science, pages 287–300.
Springer Berlin Heidelberg, 2005.
[33] HoonSang Jin and Fabio Somenzi. Prime clauses for fast enumeration of satisfying assignments to boolean circuits. In Proceedings of
the 42Nd Annual Design Automation Conference, DAC ’05, pages
750–753, New York, NY, USA, 2005. ACM.
[34] Hyeong-Ju Kang and In-Cheol Park. Sat-based unbounded symbolic model checking. Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on, 24(2):129–140, Feb
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
39
2005.
Donald E. Knuth. The Art of Computer Programming, Volume
4, Fascicle 1: Bitwise Tricks & Techniques; Binary Decision Diagrams. Addison-Wesley Professional, 12th edition, 2009.
C. Y. Lee. Representation of switching circuits by binary-decision
programs. Bell System Technical Journal, 38(4):985–999, 1959.
Bin Li, Michael S. Hsiao, and Shuo Sheng. A novel SAT allsolutions solver for efficient preimage computation. In 2004 Design, Automation and Test in Europe Conference and Exposition
(DATE 2004), 16-20 February 2004, Paris, France, pages 272–279,
2004.
Nuno Lopes, Nikolaj Bjorner, Patrice Godefroid, and George
Varghese. Network verification in the light of program verification. Technical report, September 2013.
Stephen M. Majercik and Michael L. Littman. Using caching to
solve larger probabilistic planning problems. In Proceedings of
the Fifteenth National Conference on Artificial Intelligence and
Tenth Innovative Applications of Artificial Intelligence Conference, AAAI 98, IAAI 98, July 26-30, 1998, Madison, Wisconsin,
USA., pages 954–959, 1998.
Sharad Malik and Lintao Zhang. Boolean satisfiability from theoretical hardness to practical success. Commun. ACM, 52(8):76–82,
August 2009.
Joao Marques-Silva, Mikoláš Janota, and Inês Lynce. On computing backbones of propositional theories. In Proceedings of the 2010
Conference on ECAI 2010: 19th European Conference on Artificial Intelligence, pages 15–20, Amsterdam, The Netherlands, The
Netherlands, 2010. IOS Press.
J.P. Marques-Silva and K.A. Sakallah. Grasp: a search algorithm
for propositional satisfiability. Computers, IEEE Transactions on,
48(5):506–521, May 1999.
KenL. McMillan. Applying sat methods in unbounded symbolic
model checking. In Ed Brinksma and KimGuldstrand Larsen, editors, Computer Aided Verification, volume 2404 of Lecture Notes
in Computer Science, pages 250–264. Springer Berlin Heidelberg,
2002.
A. Morgado and J. Marques-Silva. Good learning and implicit
model enumeration. In Tools with Artificial Intelligence, 2005.
ICTAI 05. 17th IEEE International Conference on, pages 6 pp.–
136, Nov 2005.
DoRonB. Motter and IgorL. Markov. A compressed breadth-first
search for satisfiability. In DavidM. Mount and Clifford Stein,
editors, Algorithm Engineering and Experiments, volume 2409 of
Lecture Notes in Computer Science, pages 29–42. Springer Berlin
Heidelberg, 2002.
40
T. TODA AND T. SOH
[46] Keisuke Murakami and Takeaki Uno. Efficient algorithms for dualizing large-scale hypergraphs. Discrete Applied Mathematics,
170:83 – 94, 2014.
[47] Knot Pipatsrisawat and Adnan Darwiche. A lightweight component caching scheme for satisfiability solvers. In Theory and
Applications of Satisfiability Testing - SAT 2007, 10th International Conference, Lisbon, Portugal, May 28-31, 2007, Proceedings, pages 294–299, 2007.
[48] Patrick Prosser. Hybrid algorithms for the constraint satisfaction
problem. Computational Intelligence, 9:268–299, 1993.
[49] J. Qadir and O. Hasan. Applying formal methods to networking:
Theory, techniques, and applications. Communications Surveys
Tutorials, IEEE, 17(1):256–291, Firstquarter 2015.
[50] Kavita Ravi and Fabio Somenzi. Minimal assignments for bounded
model checking. In Kurt Jensen and Andreas Podelski, editors,
Tools and Algorithms for the Construction and Analysis of Systems, volume 2988 of Lecture Notes in Computer Science, pages
31–45. Springer Berlin Heidelberg, 2004.
[51] Shuo Sheng and Michael Hsiao. Efficient preimage computation
using a novel success-driven atpg. In Proceedings of the Conference
on Design, Automation and Test in Europe - Volume 1, DATE
’03, pages 10822–, Washington, DC, USA, 2003. IEEE Computer
Society.
[52] Marc Thurley. sharpsat counting models with advanced component caching and implicit bcp. In Armin Biere and CarlaP.
Gomes, editors, Theory and Applications of Satisfiability Testing
- SAT 2006, volume 4121 of Lecture Notes in Computer Science,
pages 424–429. Springer Berlin Heidelberg, 2006.
[53] Takahisa Toda. Hypergraph transversal computation with binary
decision diagrams. In Vincenzo Bonifaci, Camil Demetrescu, and
Alberto Marchetti-Spaccamela, editors, Experimental Algorithms,
volume 7933 of Lecture Notes in Computer Science, pages 91–102.
Springer Berlin Heidelberg, 2013.
[54] Takahisa Toda. Dualization of boolean functions using ternary
decision diagrams. In International Symposium on Artificial Intelligence and Mathematics, ISAIM 2014, Fort Lauderdale, FL,
USA, January 6-8, 2014, 2014.
[55] Takahisa Toda and Koji Tsuda. Bdd construction for all solutions
sat and efficient caching mechanism. In Proceedings of the 30th
Annual ACM Symposium on Applied Computing, SAC ’15, pages
1880–1886, New York, NY, USA, 2015. ACM.
[56] Frank van Harmelen, Frank van Harmelen, Vladimir Lifschitz, and
Bruce Porter. Handbook of Knowledge Representation. Elsevier
Science, San Diego, USA, 2007.
IMPLEMENTING EFFICIENT ALL SOLUTIONS SAT SOLVERS
41
[57] Yinlei Yu, P. Subramanyan, N. Tsiskaridze, and S. Malik. Allsat using minimal blocking clauses. In VLSI Design and 2014
13th International Conference on Embedded Systems, 2014 27th
International Conference on, pages 86–91, Jan 2014.
[58] Shuyuan Zhang, Sharad Malik, and Rick McGeer. Verification of computer switching networks: An overview. In Supratik
Chakraborty and Madhavan Mukund, editors, Automated Technology for Verification and Analysis, Lecture Notes in Computer
Science, pages 1–16. Springer Berlin Heidelberg, 2012.
(T. Toda) Graduate School of Information Systems, University of
Electro-Communications, 1-5-1 Chofugaoka, Chofu, Tokyo 182-8585,
Japan.
E-mail address: [email protected]
(T. Soh) Information Science and Technology Center, Kobe University, 1-1 Rokkodai, Nada Kobe 657-8501, Japan.
E-mail address: [email protected]
| 8cs.DS
|
arXiv:1703.01640v1 [cs.CG] 5 Mar 2017
Approximation algorithms for TSP with neighborhoods
in the plane
Adrian Dumitrescu ∗
University of Wisconsin–Milwaukee
Milwaukee, WI 53201-0784
[email protected]
Joseph S. B. Mitchell †
Stony Brook University
Stony Brook, NY 11794-3600
[email protected]
August 9, 2014
Abstract
In the Euclidean TSP with neighborhoods (TSPN), we are given a collection of n regions
(neighborhoods) and we seek a shortest tour that visits each region. As a generalization of the
classical Euclidean TSP, TSPN is also NP-hard. In this paper, we present new approximation
results for the TSPN, including (1) a constant-factor approximation algorithm for the case of
arbitrary connected neighborhoods having comparable diameters; and (2) a PTAS for the important special case of disjoint unit disk neighborhoods (or nearly disjoint, nearly-unit disks).
Our methods also yield improved approximation ratios for various special classes of neighborhoods, which have previously been studied. Further, we give a linear-time O(1)-approximation
algorithm for the case of neighborhoods that are (infinite) straight lines.
1
Introduction
A salesman wants to meet a set of n potential buyers. Each buyer specifies a connected region
in the plane, his neighborhood, within which he is willing to meet the salesman. For example, the
neighborhoods may be disks centered at buyers’ locations, and the radius of each disk specifies the
maximum distance a buyer is willing to travel to the meeting place. The salesman wants to find
a tour of shortest length that visits all of buyers’ neighborhoods and finally returns to its initial
departure point. A variant of the problem, which we will address in this paper, is that in which no
departure point is specified, and only a tour of the neighborhoods is to be found. This problem,
which is known as the TSP with neighborhoods (TSPN), is a generalization of the classic Euclidean
Traveling Salesman Problem (TSP), in which the regions, or “neighborhoods,” are single points,
and consequently is NP-hard [9, 21].
Related Work.
The TSP has a long and rich history of research in combinatorial optimization.
It has been studied extensively in many forms, including geometric instances; see [4, 13, 14, 20, 24].
The problem is known to be NP-hard, even for points in the Euclidean plane [9, 21]. It has
∗
This work was done while the author was a visiting faculty member at Stony Brook University.
Partially supported by grants from HRL Laboratories, NASA Ames, the National Science Foundation (CCR9732220), Northrop-Grumman Corporation, Sandia National Labs, and Sun Microsystems.
†
1
recently been shown that the geometric instances of the TSP, including the Euclidean TSP, have
a polynomial-time approximation scheme, as developed by Arora [2] and Mitchell [17], and later
improved by Rao and Smith [23].
Arkin and Hassin [1] were the first to study approximation algorithms for the geometric TSPN.
They gave O(1)-approximation algorithms for several special cases, including parallel segments of
equal length, translates of a convex region, translates of a connected region, and more generally,
for regions which have diameter segments that are parallel to a common direction, and the ratio
between the longest and the shortest diameter is bounded by a constant.
For the general case of connected polygonal regions, Mata and Mitchell [15] obtained an
O(log n)-approximation algorithm, based on “guillotine rectangular subdivisions”, with time bound
O(N 5 ), where N is the total complexity of the n regions. Gudmundsson and Levcopoulos [10] have
recently obtained a faster method, which, for any fixed ǫ > 0, is guaranteed to perform at least
one of the following tasks1 : (1) it outputs a tour of length at most O(log n) times optimum in time
O(n log n + N ); (2) it outputs a tour of length at most (1 + ǫ) times optimum in time O(N 3 ). So
far, no polynomial-time approximation algorithm is known for general connected regions.
Recently it was shown that TSPN is APX-hard and cannot be approximated within a factor of
1.000374 unless P=NP [5, 11]. In fact, the inapproximability factor for the vertex cover problem in
graphs with degree bounded by 5, as stated in [7], and on which the result in [11] is based, implies
a factor larger than that.
In the time since this paper first appeared, it has been shown by de Berg et. al [5] that the TSPN
has an O(1)-approximation algorithm in the case that the regions are the regions are connected,
disjoint, convex and fat. Also, Schwartz and Safra [25] have improved the lower bounds of [5, 11]
on the hardness of approximation
of several variants of the TSPN problem, and Jonsson [12] has
√
given an O(n) time 2-approximation algorithm for the case of regions that are lines in the plane;
our previous O(n) time π/2-approximation algorithm appears in Section 5.
Summary of Our Results.
geometric TSPN, including:
In this paper, we obtain several approximation results on the
(1) We extend the approaches initiated in [1] and obtain the first O(1)-approximation algorithm
for the TSPN having connected regions of the same or similar diameter. This solves among
others, the open problem posed in [1], to provide a constant-factor approximation algorithm
for TSPN on segments of the same length and arbitrary orientation.
(2) We give a polynomial-time approximation scheme (PTAS) for the case of disjoint unit disks or
the case of nearly disjoint disks of nearly the same size. The algorithm is based on applying
the m-guillotine method with a new area-based charging scheme. The fact that there is a
PTAS for the case in which the neighborhoods are “nice,” with no point lying in more than
a constant number of neighborhoods, should be contrasted with the fact that the TSPN
on arbitrary regions is APX-hard. The construction in the proof of [11] utilizes “skinny”
neighborhoods, which intersect each other extensively.
(3) We also give modest improvements on earlier approximation bounds in [1] for the cases of
parallel segments of equal length, translates of a convex region, and translates of a connected
region.
1
One does not know in advance which one will be accomplished.
2
(4) We present simple algorithms which achieve a constant-factor guarantee for the case of equal
disks and for the case of infinite straight lines.
Preliminaries.
The input to our algorithms will be a set R of n regions, each of which is a
simply-connected, closed subset of the plane, ℜ2 , bounded by a finite union of arcs of constantdegree algebraic curves (degenerate regions are simply points). Since the regions are assumed to
be closed, they include the points that lie on the curves that form their boundary. The assumption
that regions are simply connected means that each region has no “holes.”
Ideally, each region is a subset of the plane lying inside a simple, closed, continuous curve, together with the curve itself. However, we can only deal with regions that are computer-representable,
hence the above definition.
Examples of allowable regions include simple polygons, whose boundaries are unions of a finite
number of straight line segments, circular disks, regions bounded by straight segments and circular
arcs, infinite straight lines, etc. We let N denote the total number of arcs specifying all n regions
in R, i.e. the total combinatorial complexity of the input.
A tour (or circuit) T , is a closed continuous curve that visits each region of R. The length of
tour T , denoted |T |, is the Euclidean length of the curve T . To avoid ambiguity, the size of a finite
set X is denote by #X. In the TSP with neighborhoods (TSPN) problem, our goal is to compute
a tour whose length is guaranteed to be close to the shortest possible length of a tour. We let T ∗
denote any optimal tour and let L∗ = |T ∗ | denote its length. An algorithm that outputs a tour
whose length is guaranteed to be at most c·L∗ is said to be a c-approximation algorithm and to have
an approximation ratio of c. A family of (1 + ǫ)-approximation algorithms, parameterized by ǫ > 0,
and each running in polynomial time for fixed ǫ, is said to be a polynomial-time approximation
scheme (PTAS).
Outline of the Paper.
In Section 2 we use some simple packing arguments to yield approximation algorithms for the TSPN for equal-size disks. Section 3 presents a PTAS for the TSPN
for equal-size disks. In Section 4 we give an approximation algorithm for the TSPN for regions
having the same diameter. Finally, in Section 5, we give an approximation algorithm for the case
of regions that are infinite straight lines. We conclude with a short list of open problems for future
research.
2
Equal-Size Disks
We begin by giving some simple arguments and corresponding algorithms that achieve a constant
approximation ratio for TSPN on a set R of n disks of the same size. Without loss of generality,
we assume that all disks have unit radius. Our results carry over naturally to disks of nearly the
same size, with corresponding changes in the approximation factor.
First, we consider the case of disjoint unit disks. The algorithm is simple and natural: using
known PTAS results for TSP on points [2, 17, 23], compute, in time O(n log n), a (1+ǫ)-approximate
tour, T = TC , of the center points of the n disks. We refer to it as the center tour. Clearly TC is a
valid region tour. We claim that
Proposition 1 Given set of n disjoint unit disks, one can compute a tour, T , whose length satisfies
|T | ≤ ((1 +
8
)|T ∗ | + 8)(1 + ǫ),
π
3
where T ∗ is an optimal tour. The running time is dominated by that of computing a (1 + ǫ)approximate tour of n points.
Proof. Put L∗ = |T ∗ |. Since T ∗ visits all disks, the area AL∗ swept by a disk of radius 2, whose
center moves along T ∗ , covers all of the unit disks R. This area is bounded as follows
πn ≤ AL∗ ≤ 4L∗ + 4π.
Thus, n ≤ 4 +
4L∗
π .
A center tour of length at most
L∗ + 2n ≤ L∗ + 2(4 +
4L∗
8
) = (1 + )L∗ + 8
π
π
can be obtained by going along T ∗ and making a detour of length at most 2 to visit the center of a
disk at each point where T ∗ first visits a disk. Hence, the length of the computed tour is bounded
as claimed.
✷
For large n, the approximation ratio above is (1 + π8 )(1 + ǫ) ≤ 3.55; for small constant values
of n, the problem can be solved exactly using brute force. We note that for any algorithm that
outputs a tour on the center points, we cannot expect an approximation ratio smaller than 2. To
see this, consider a large square, and place almost touching unit disks along its perimeter, both on
its inside and on its outside. All of the disks touch the perimeter, which is also an optimal disk
tour except at the four corners of the square. The length of the disk center tour is roughly two
times the perimeter of the square.
When the disks are nearly of the same size, so that the ratio between the maximum and
minimum radius is bounded by a constant k > 1, for large n, the approximation ratio is about
2
1 + 8kπ .
Next we consider the case in which the disks can overlap. First, we compute a maximal independent (pairwise-disjoint) set I of disks. Next, we compute TI , a (1 + ǫ)-approximate tour of the
center points of disks in I. Finally, we output a tour T obtained by following the tour TI , taking
detours around the boundaries of each of the disks in I, as illustrated in Figure 1. More specifically,
we select an arbitrary disk D0 and one of the intersection points, s, between D0 and TI . We start
at point s and go clockwise along TI . Whenever the boundary of a disk of I is encountered, we
follow clockwise around the boundary of the disk until we encounter again the tour TI . When we
finally reach s, we continue clockwise around D0 to TI and then continue counterclockwise around
TI , again taking detours clockwise around the disks of I that we encounter along the way. The
tour T finally ends when we return the second time to s. In this way, our tour T traverses the
boundary of each disk of I exactly once and therefore visits all of the disks that are not in I as well.
We remark that this method of constructing a feasible tour of the disks results in a slightly better
worst-case ratio than another natural strategy for extending TI to a full disk tour: While traversing
TI , each time one encounters the boundary of a disk D ∈ I, traverse the entire circumference of D
exactly once, and then go directly to the point where the tour TI exits D and continue along TI .
Denote by T ∗ an optimal disk tour, of length L∗ , and by TI∗ an optimal disk tour of the subset
I ⊆ R of disks. A constant approximation ratio can be derived from the following three inequalities:
|T | ≤ π|TI | + 2π,
8
|TI∗ | + 8)(1 + ǫ),
|TI | ≤ (1 +
π
4
(1)
(2)
a1
s
D0
a2
Figure 1: Construction of the tour T from the tour TI on the center points of a maximal independent
set, I, of disks. Starting from s, we follow the thick solid tour, back to s, and then follow the thick
dashed tour back around to s again.
|TI∗ | ≤ L∗ .
(3)
The third inequality follows from the fact that I ⊆ R, and the second from the case of disjoint
unit disks considered above. To check the first inequality, decompose TI into #I parts, assuming
#I ≥ 2, one for each disk in I, by cutting each segment between two consecutive centers in the
middle. Let x + y = d ≥ 2 be the length of one of these parts of tour TI , corresponding to a disk
D, where x, y are the lengths of the two segments of TI adjacent to the center of D. Write a1 , a2
for the arc lengths of D when its boundary is traversed by T ; we have a1 + a2 = 2π. Writing the
ratio of the length of the corresponding part of T to the length, d, of this part of TI , we get
2(d + π − 2)
(x − 1 + a1 + y − 1) + (x − 1 + a2 + y − 1)
=
≤ π,
d
d
the maximum being attained when d = 2. When #I = 1, we have |TI | = 0, |T | = 2π. Thus (1) is
satisfied in all cases. Putting (1), (2), (3) together, we get
Proposition 2 Given set of n unit disks, possibly overlapping, one can compute a tour, T , whose
length satisfies
|T | ≤ ((π + 8)|T ∗ | + 8π)(1 + ǫ) + 2π,
where T ∗ is an optimal tour. The running time is dominated by that of computing a (1 + ǫ)approximate tour of n points.
For large n, the approximation ratio is (π + 8)(1 + ǫ) ≤ 11.15. We note that the approximation
ratio we have obtained with this
√ approach for disjoint unit disks, 3.55, (resp., 11.15 for unit disks)
is better (resp., weaker) than 32 + 72 ≈ 7.62, the approximation ratio that will be given at the
end of Section 4 for translates of a convex region, which applies, of course, to the case of unit disks.
5
3
A PTAS for Disjoint Equal Disks
In this section we present a polynomial-time approximation scheme for the TSPN in the case of
regions that are equal-size disks or nearly equal-size disks.
Given the powerful methods that have been developed to obtain PTAS’s for various geometric
optimization problems, such as the Euclidean TSP, it is natural to suspect that these same techniques may apply to the TSPN. Indeed, one may expect that TSPN should have a PTAS based on
applying existing methods. However, we know now, from the recent APX-hardness result of [11],
that this cannot be. What goes wrong?
The basic issue we must address in order to apply these techniques is to be able to write a
recursion to solve an appropriate “succinct” subproblem with dynamic programming. What is the
subproblem “responsible” for solving? For problems involving points, the subproblem can be made
responsible for constructing some kind of inexpensive network on the points inside the subproblem
defined by a rectangle, and to interconnect this network with the boundary in some nicely controlled
way, e.g., with only a constant complexity of connection, in the case of m-guillotine methods. The
problem with regions is that they can cross subproblem boundaries. Then, we do not know if the
subproblem is responsible to visit the region, or if the region is visited outside the subproblem. We
cannot afford to enumerate the subset of regions that cross the boundary for which the subproblem
is responsible – there are too many such subsets, leading to too many subproblems. Thus, we need
a new idea.
Our approach is to employ a new type of structural result, based on the general method of
m-guillotine subdivisions. In particular, we show how to transform an optimal tour into one of a
special class of tours that recursively has a special m-guillotine structure, permitting us to have
a succinct, constant-size specification of the subset of regions, crossing the subproblem boundary,
for which the subproblem is “responsible” in that it must visit these regions on its interior. In
order to bound the increase in tour length in performing this transformation, we must “charge off”
the added tour length to some small fraction of the length of the optimal tour, just as is done
in proving the bounds for the m-guillotine PTAS method for TSP. In order to do this charging,
we must assume some special structure on the class of neighborhoods in the TSPN, e.g., that
the regions D = {D1 , . . . , Dn } are pairwise-disjoint, equal-size disks, or have a similar structure
allowing us to relate tour length to area.
Here, we show how the approach applies to disjoint disks having equal radii, δ; generalizations
are readily made to the case of nearly equal radii, with a constant upper bound on the ratio of
radii, and to the case of “modestly overlapping,” in which the disks become pairwise-disjoint if all
of them are decreased in size by a constant factor while keeping the center points the same.
We begin with some definitions, largely following the notation of [17]. Let G be an embedding
of a planar graph, and let L denote the total Euclidean length of its edges, E. We can assume
without loss of generality that G is restricted to the unit square, B; i.e., E ⊂ int(B).
Consider an axis-aligned rectangle W , a window, with W ⊆ B. Rectangle W will correspond
to a subproblem in a dynamic programming algorithm. Let ℓ be an axis-parallel line intersecting
W . We refer to ℓ as a cut and assume, without loss of generality, that ℓ is vertical.
The intersection, ℓ ∩ (E ∩ int(W )), of a cut ℓ with E ∩ int(W ) consists of a possibly empty
set of subsegments of ℓ. These subsegments are possibly singleton points. Let ξ be the number
of endpoints of such subsegments along ℓ, and let the points be denoted by p1 , . . . , pξ , in order of
decreasing y-coordinate along ℓ. For a positive integer m, we define the m-span, σm (ℓ), of ℓ with
respect to W as follows. If ξ ≤ 2(m − 1), then σm (ℓ) = ∅; otherwise, σm (ℓ) is defined to be the line
6
segment, pm pξ−m+1 , joining the mth endpoint, pm , with the mth-from-the-last endpoints, pξ−m+1 .
Refer to Figure 2. Note that the segment pm pξ−m+1 may be of zero length in case pm = pξ−m+1 .
B
W
p1
p3
p2
σ3(ℓ)
p9
ℓ
Figure 2: Definition of m-span: Here, the 3-span, σ3 (ℓ) of ℓ with respect to the window W ⊂ B
is highlighted with a thick shaded vertical segment.
The intersection, ℓ ∩ D ∩ W of ℓ with the disks that intersect W consists of a possibly empty set
of ξD ≤ |D ∩ W | subsegments of ℓ, one for each disk that is intersected by ℓ in W . Let these disks
be D1 , D2 , . . . , DξD , in order of decreasing y-coordinate. For a positive integer m, we define the
m-disk-span, σm,D (ℓ), of ℓ with respect to W as follows. If ξD ≤ 2m, then σm,D (ℓ) = ∅; otherwise,
σm,D (ℓ) is defined to be the possibly zero-length line segment joining the bottom endpoint of ℓ∩Dm
with the top endpoint of ℓ ∩ DξD −m+1 . Refer to Figure 3.
Line ℓ is an m-good cut with respect to W if σm (ℓ) ⊆ E and σm,D (ℓ) ⊆ E. In particular, if
ξ ≤ 2(m − 1) and ξD ≤ 2m, then ℓ is trivially an m-good cut, since both the m-span and the
m-disk-span are empty in this case.
We now say that E satisfies the m-guillotine property with respect to window W if either (1) W
does not fully contain any disk; or (2) there exists an m-good cut, ℓ, that splits W into W1 and
W2 , and, recursively, E satisfies the m-guillotine property with respect to both W1 and W2 .
7
W
D1
D2
D3
σ3,D (ℓ)
ℓ
Figure 3: Definition of m-disk-span: Here, the 3-disk-span, σ3,D (ℓ) of ℓ with respect to the window
W ⊂ B is highlighted with a thick shaded vertical segment.
Theorem 1 Let G be an embedded connected planar graph, with edge set E, of total length L, and
let D be a given set of pairwise-disjoint equal-radius disks (of radius δ) each of which intersects E.
Assume that E and D are contained in the unit square B. Then, for any positive integer m, there
exists a planar graph G′ that satisfies the m-guillotine property with respect to B and has an edge
set E ′ ⊇ E of length
!
√
16δ
2 + 16/π
′
L+
.
L ≤ 1+
m
m
Proof. We convert G into a √
new graph G′ by adding to E a new set of horizontal/vertical edges
whose total length is at most 2+16/π
L + 16δ
m
m . The construction is recursive: at each stage, we show
that there exists a cut, ℓ, with respect to the current window W (which initially is the unit square
B), such that we can afford to add both the m-span and the m-disk-span to E.
We say that a point p on a cut ℓ is m-dark with respect to ℓ and W if, along ℓ⊥ ∩ int(W ), there
are at least m edges of E intersected by ℓ⊥ on each side of p, where ℓ⊥ is the line through p and
perpendicular to ℓ.2 We say that a subsegment of ℓ is m-dark (with respect to W ) if all points of
the segment are m-dark with respect to ℓ and W .
The important property of m-dark points along ℓ is the following: Assume, without loss of
generality, that ℓ is horizontal. We consider any line segment that lies along an edge of E to have
a top side and a bottom side; the top is the side that can be seen from above, from a point with
y = +∞. Then, if all points on subsegment pq of ℓ are m-dark, we can charge the length of pq off
to the bottoms of the first m subsegments, E + ⊆ E, of edges that lie above pq, and the tops of the
first m subsegments, E − ⊆ E, of edges that lie below pq, since we know that there are at least m
edges “blocking” pq from the top/bottom of W . We charge pq’s length half to E + , charging each
1
units of charge, and half to E − , charging each of the
of the m levels of E + from below, with 2m
2
We can think of the edges E as being “walls” that are not very effective at blocking light — light can go through
m − 1 walls, but is stopped when it hits the mth wall; then, p on a line ℓ is m-dark if p is not illuminated when light
is shone in from the boundary of W , along the direction of ℓ⊥ .
8
1
units of charge. We refer to this type of charge as the “red”
m levels of E − from above, with 2m
charge.
We say that a point p on a cut ℓ is m-disk-dark with respect to ℓ and W if, along ℓ⊥ ∩ int(W ),
there are at least m disks of D that have a nonempty intersection with ℓ⊥ ∩ W on each side of
p. Here, again, ℓ⊥ is the line through p and perpendicular to ℓ. We say that a subsegment of ℓ
is m-disk-dark with respect to W if all points of the segment are m-disk-dark with respect to ℓ
and W . The chargeable length within W of a cut ℓ is defined to be the sum of the lengths of its
m-dark portion and its m-disk-dark portion. Refer to Figure 4.
W
p
ℓ
ℓ⊥
Figure 4: Definition of m-disk-dark: Here, the points that are 2-disk-dark with respect to ℓ are
(2,D)
those four subsegments of ℓ that lie within the shaded regions, which comprise the set Rx
of
points of W that are 2-disk-dark with respect to horizontal cuts.
The important property of m-disk-dark points along horizontal ℓ is the following: If all points
on subsegment pq of ℓ are m-disk-dark, then we can charge the length of pq off to the bottoms of
the first m disks that lie above pq, and the tops of the first m disks that lie below pq, since we know
that there are at least m disks “blocking” pq from the top/bottom of W . We charge pq’s length
1
units of charge, and
half upwards, charging the bottoms of each of the m “levels” of disks with 2m
1
half downwards, charging the tops of each of the m “levels” of disks with 2m units of charge. We
refer to this type of charge as the “blue” charge.
A cut ℓ is favorable if its chargeable length within W is at least as large as the sum of the
lengths of the m-span and the m-disk-span. We do not attempt to take advantage of the fact that
there may be overlap among the m-span and the m-disk-span, nor that there may be portions of
these two segments that are already part of E; an argument taking advantage of these facts may
improve slightly the constants in some of our bounds.
The existence of a favorable cut is guaranteed by the following key lemma, whose proof is similar
to that of the key lemma in [17]:
Lemma 1 For any G and any window W , there is a favorable cut.
Proof. We show that there must be a favorable cut that is either horizontal or vertical.
9
Let f (x) denote the “cost” of the vertical line, ℓx , passing through the point (x, 0), where “cost”
means the sum of the lengths of the m-span and the m-disk-span for ℓx . Then,
f (x) = |σm (ℓx )| + |σm,D (ℓx )|.
Thus,
Ax =
Z
0
(m)
1
f (x)dx =
Ax(m)
+ A(m,D)
x
=
Z
1
0
|σm (ℓx )|dx +
Z
1
0
|σm,D (ℓx )|dx,
(m)
R1
|σm (ℓx )|dx is the area of the x-monotone region Rx of points of B that are mR1
(m,D)
dark with respect to horizontal cuts, and Ax
= 0 |σm,D (ℓx )|dx is the area of the x-monotone
where Ax
=
(m,D)
0
region Rx
of points of B that are m-disk-dark with respect to horizontal cuts. Refer
R 1 to Figure 4.
Similarly, define g(y) to be the cost of the horizontal line through y, and let Ay = 0 g(y)dy.
Assume, without loss of generality, that Ax ≥ Ay . We claim that there exists a horizontal
favorable cut; i.e., we claim that there exists a horizontal cut, ℓ, such that its chargeable length is
at least as large as the cost of ℓ, meaning that the length of its m-dark portion plus its m-disk-dark
portion is at least |σm (ℓ)| + |σm,D (ℓ)|. To see this, note that Ax can be computed by switching the
(m,D)
(m)
horizontally, rather than vertically; i.e.,
order Rof integration,
Rx and Rx
R 1“slicing” theR regions
1
1
Ax = 0 h(y)dy = 0 hm (y)dy + 0 hm,D (y)dy, where hm (y) is the m-dark length of the horizontal
(m,D)
line through y, hm,D (y) is the length of the intersection of Rx
with a horizontal line through
y, and h(y) is the chargeable length of the horizontal line through y. In other words, hm (y) (resp.,
hm,D (y)) is the length of the m-dark (resp., m-disk-dark) portion of the horizontal line through y.
R1
R1
Thus, since Ax ≥ Ay , we get that 0 h(y)dy ≥ 0 g(y)dy ≥ 0. Thus, it cannot be that for all values
of y ∈ [0, 1], h(y) < g(y), so there exists a y = y ∗ for which h(y ∗ ) ≥ g(y ∗ ). The horizontal line
through this y ∗ is a cut satisfying the claim of the lemma.
If, instead, we had Ax ≤ Ay , then we would get a vertical cut satisfying the claim.
✷
Now that we know there must be a favorable cut, ℓ, we can charge off the cost of the m-span
and the m-disk-span of ℓ, making “red” charge on the bottoms (resp., tops) of segments of E that
lie above (resp., below) m-dark points of ℓ, and making “blue” charge on the bottoms (resp., tops)
of disks that lie above (resp., below) m-disk-dark points of ℓ. We then recurse on each side of the
cut, in the two new windows.
After a portion of E has been charged red on one side, due to a cut ℓ, it will be within m levels
of the boundary of the windows on either side of ℓ, and, hence, within m levels of the boundary of
any future windows, found deeper in the recursion, that contain the portion. Thus, no portion of
E will ever be charged red more than once from each side,
√ in each of the two directions, horizontal
or vertical, so no portion of E will ever pay more than 2/m times its length in red charge. We
1
per unit length of the perimeter of the segment’s axis-aligned bounding
charge at the rate of 2m
box,
and
the
worst
case
is
achieved for a segment of slope ±1. Thus, the total red charge is at most
√
2
m L.
Similarly, no disk will ever have its boundary charged blue more than once from each of the
1
per unit length of its
two directions, horizontal or vertical. Since it is charged at the rate of 2m
axis-aligned bounding box, whose perimeter is 8δ, we get a total blue charge of at most 4δ
m n. We
now appeal to the lower bound from the previous section, which was based on an area argument.
Here, for radius δ disks, that argument shows that πδ2 n ≤ AL ≤ 4δL + (2δ)2 π, from which we get
10
4
L + 4. Note that this area argument uses the fact that E is connected. Thus, the total blue
n ≤ πδ
16
charge is at most πm
L + 16δ
m .
It is also important to note that we are always charging red portions of the original edge set
E: the new edges added are never themselves charged, since they lie on window boundaries and
cannot therefore serve to make a portion of some future cut m-dark.
Overall, then, the total increase in length caused by adding the m-spans and m-disk-spans along
favorable cuts is bounded by
√
16δ
16 + π 2
L+
.
πm
m
Our goal in adding the m-disk-span is to obtain a succinct representation of which disks that
straddle the boundary of a window are visited within the window and which are visited outside
the window. The m-disk-span segment visits all but a constant, O(m), number of the disks on
the corresponding side of the window. There is, however, one remaining issue with respect to the
m-disk-span segments: We need to argue that we can “afford,” within our charging scheme, to
connect the m-disk-span to the input edge set, E. This is because, in our dynamic programming
optimization, we will find a shortest possible planar graph with the m-guillotine property that obeys
certain connectivity constraints, as well as other properties that guarantee that the graph has an
Eulerian subgraph spanning all disks. The optimal graph that we compute uses the m-disk-span
segments to visit the corresponding disks on the boundaries of windows that define subproblems.
Remark. (Clarification added after journal publication.) We had previously phrased the above as
“find a shortest possible connected planar graph with the m-guillotine property”; we have rephrased
to be“find a shortest possible planar graph with the m-guillotine property that obeys certain connectivity constraints”.
In particular, we add connections to E from the endpoints of the m-disk-span to the point of
E in the corresponding disk that is closest to the endpoint. We know that this connection is of
length at most 2δ per endpoint of the m-disk-span, since the disks are of diameter 2δ. In total,
this adds only 4δ to the length of the m-disk-span. Assuming the m-disk-span stabs at least three
or more disks, its length is at least Ω(δ), implying that we can charge off this extra 4δ for the
connections in the same way that we charge off the m-disk-spans themselves. If, on the other
hand, the m-disk-span stabs only one or two disks, then we can afford to skip the addition of the
m-disk-span altogether, and just keep track in the dynamic program of the necessary information
for these couple extra disks, specifying whether they are to be visited within the window or not.
Remark. (Clarification added after journal publication; thanks to Sophie Spirkl for her inquiry
and input.) We argue above that we can afford to add connections to E, since the length added is
proportional to the bridge length. Our dynamic programming algorithm computes a minimum-length
planar network with the m-guillotine property that obeys certain connectivity constraints, so that
all of the network is connected except (possibly) the m-disk-spans, which may not be connected to
the rest of the (connected) network. The objective function in the dynamic program requires that we
minimize the total length of the network, counting the lengths of the bridges that serve as m-diskspans a constant number of times. Then, since we know we can afford to add length proportional
to the total m-disk-span lengths, we know we can add the connections mentioned above, resulting
in an overall connected network (and can make it Eulerian with appropriate doubling of the bridge
segments, in the usual way, as mentioned again below), appropriately close to the optimal length. ✷
11
Corollary 1 The TSPN for a set of disjoint equal-size disks has a PTAS. The same is true for
a set of nearly disjoint, nearly equal-size disks, for which there is a constant upper bound on the
ratio of largest to smallest radius and there is a constant factor such that the disks become disjoint
if their radii are each multiplied by the factor while keeping the center points the same.
Proof. We consider only the case of disjoint disks each of radius δ; the generalization to nearly
disjoint, nearly equal-size disks is straightforward.
We impose a regular m-by-m grid on each disk; let G denote the resulting set of O(m2 n) grid
points.
Consider an optimal tour, OP T , of length L∗ . Now, OP T is a simple polygon. We can
perturb the vertices of OP T so that each lies at a grid point in G, resulting in a new tour, OP T ′ ,
visiting every disk, whose length is at most L∗ + O(nδ/m). Using the fact, from the previous
4 ∗
L + 4, and assuming that n ≥ 8, we get that the length of OP T ′ is at most
section, that n ≤ πδ
4
8
∗
L∗ + O( πm
L∗ + 4δ
m ) ≤ L (1 + O( πm )). If n < 8, then we can solve the problem in constant time by
brute force.
Theorem 1 implies that we can convert OP T ′ , which consists of some set E of edges, into a
planar graph, OP T ′′ , obeying the m-guillotine property, while not increasing the total length by
too much. In particular, the length of OP T ′′ is at most
√
8
16δ
2 + 16/π ∗
)L (1 + O(
)) +
≤ (1 + O(1/m))L∗ .
(1 +
m
πm
m
We now apply a dynamic programming algorithm, running in O(nO(m) ) time, to compute a
minimum-length planar graph having a prescribed set of properties: (1) it satisfies the m-guillotine
property, which is necessary for the dynamic program to have the claimed efficiency; (2) it visits at
least one grid point of G in each region Di ; and (3) it contains an Eulerian subgraph that spans the
disks. This third condition that allows us to extract a tour in the end. We only outline here the
dynamic programming algorithm; the details are very similar to those of [17], with the modification
to account for the m-disk-span.
A subproblem is defined by a rectangle W whose coordinates are among those of the grid
points G, together with a constant amount (O(m)) of information about how the solution to the
subproblem interacts across the boundary of W with the solution outside of W . This information
includes the following:
(a) For each of the four sides of W , we specify a “bridge” segment and at most 2m other segments
with endpoints among G that cross the side; this is done exactly as in the case of the Euclidean
TSP on points, as in [17].
(b) For each of the four sides of W , we specify a “disk bridge” segment corresponding to the
m-disk-span, and, for each of at most 2m disks that are not intersected by the disk bridge
segment, we specify in a single bit whether the disk is to be visited within the subproblem or
not; if not, it is visited outside the window W .
(c) We specify a required “connection pattern” within W . In particular, we indicate which subsets
of the O(m) specified edges crossing the boundary of W are required to be connected within
W . This, again, is done exactly as is detailed for the Euclidean TSP on point sets in [17].
In order to end up with a graph having an Eulerian subgraph spanning the disks, we use the
same trick as done in [17]: we “double” the bridge segments and the disk bridge segments, and then
12
require that the number of connections on each side of a bridge segment satisfy a parity condition.
Exactly as in [17], this allows us to extract a tour from the planar graph that results from the
dynamic programming algorithm, which gives a shortest possible graph that obeys the specified
conditions.
The result is that in polynomial time (O(nO(m) )) one can compute a shortest possible graph,
from a special class of such graphs, and this graph spans the regions D. Theorem 1 guarantees
that the length of the resulting graph is very close, within factor 1 + O(1/m), to the length of an
optimal solution to the TSPN. (We also know from the remarks previously, that we can afford to add
connections to assure that the m-disk-spans are connected to the rest of the (connected) network.)
Thus, once we extract a tour from the Eulerian subgraph, we have the desired (1+ǫ)-approximation
solution, where ǫ = O(1/m).
Finally, we mention that the running time of O(nO(m) ) can be improved to O(nC ), for a constant
C independent of m, using the notion of grid-rounded guillotine subdivisions, as in [18, 19]; here, the
dependence on the constant m is exponential in the multiplicative constant concealed by the big-Oh
of O(nC ). We suspect that the techniques of Rao and Smith [23], based on Arora’s method [2],
can be used to improve the time bound to O(n log n), while possibly also addressing the problem
in higher dimensions; we leave this for future work.
✷
4
Connected Regions of the Same Diameter
In this section we give a constant-factor approximation algorithm for the TSPN problem that
applies in the case that all regions have the same diameter or nearly the same diameter.
The diameter of a region, δ, is the distance between two points in the region that are farthest
apart. Without loss of generality, we assume that all regions have unit diameter, δ = 1. The general
method we use is to carefully select a representative point in each region, and then compute an
almost optimal tour on these representative points. This approach was initiated in [1]. We also
employ a specialized version of the Combination Lemma in [1]. Now we describe the algorithm,
which is similar in many aspects to the one in [1].
In each region, compute a unit-length diameter segment; in case of multiple such segments,
select one arbitrarily. Computing a diameter segment of a region can be done efficiently, in time
linear in the complexity of the region. Classify the regions into two types: (1) those for which
the selected diameter is almost horizontal, by which we mean its slope is between −45◦ and 45◦ ;
(2) those for which the selected diameter is almost vertical, by which we mean all the others. Use
algorithm A (below) for each of these two region types. We will prove that a constant ratio is
achievable for each class, after a suitable transformation is applied to regions of type (2). We then
apply the Combination Lemma to obtain a constant-factor approximation for all regions.
Lemma 2 (Combination Lemma) Given regions that can be partitioned into two types — those
having almost horizontal unit diameters, and those having almost vertical unit diameters respectively
— and constants c1 , c2 bounding the error ratios with which we can approximate optimal tours on
regions of types 1 and 2, then we can approximate the optimal tour on all regions with an error
ratio bounded by c1 + c2 + 2.
We omit the proof, which is a simplified version of the argument in [1]. The bound on the
approximation ratio is still the same. We remark that the Combination Lemma in [1] implicitly
13
assumes that for each of the two types, the diameters of the regions are parallel to some direction,
an assumption which does not hold in our case.
Algorithm A gives a constant-factor approximation of the optimal tour for regions of type (1).
Regions of type (2) are readily handled by rotating them by 90◦ to obtain type (1) regions. When
the diameters are nearly the same, so that the ratio of the largest to the smallest is bounded by
a constant, we can still get an approximation algorithm with a constant ratio; however since this
ratio is rather large even for same diameters, we omit the calculations.
A set of lines is a cover of a set of regions if each region is intersected by at least one line from
the covering set. We refer to such a set of lines as covering lines.
A
F
E
D
B
C
Figure 5: Illustration of Step 1 in Algorithm A
Algorithm A. Input: a set of n regions of type (1).
Step 1. Construct a cover of the regions by a minimum number of vertical lines. This procedure
works in a greedy fashion, namely the leftmost line is as far right as possible, so that it is a right
tangent of some region. To obtain this cover, the intervals of projection on the x-axis of all regions
are computed and a greedy cover of this set of intervals is found. After removing all intervals covered
by a previous line, and there are still uncovered intervals, another covering line is repeatedly added
to the cover. At the same time, a representative point for each region is arbitrarily selected on the
corresponding covering line and inside the region, e.g., the topmost boundary point of intersection
between the region and its covering line. In this way n representative points, one per region, are
selected. An illustration of this procedure appears in Figure 5. An important remark is that the
representative points are not necessarily on the selected diameters, since a diameter may not be
entirely contained in its region.
Assuming that we have only regions of type (1), the greedy cover has the effect of obtaining
14
a large enough horizontal distance between any two consecutive covering lines. We remark that
the greedy covering algorithm for a set of closed intervals on a line is known to output a cover of
minimum size, a property which carries over to our vertical line cover.
Step 2. Proceed according to the following three cases.
Case 1: The greedy cover contains one covering line.
Compute a smallest perimeter axis-aligned rectangle Q that intersects all regions, where Q is
considered a two-dimensional domain. Let w and h denote Q’s width and height respectively.
Consider Q to be a graph with four vertices and four edges (its four sides). Let e1 and e2
be the vertical segments of height h that partition Q into three equal-width parts. Add to
the graph Q a double edge corresponding to e1 and a doubled edge corresponding to e2 . The
resulting graph G is an Eulerian multigraph, since all node degrees are even, with 8 vertices
and 12 edges. Output any Euler tour T of G. While such a tour is not, in general, the shortest
possible tour visiting the regions, it suffices for our purposes of approximation.
Case 2: The greedy cover contains two covering lines.
Move (if possible) the rightmost vertical covering line to the left as much as possible (while
still covering all regions). Recompute the representative points obtained in this way. Set D
to be the distance between the two covering lines (clearly D > 0).
Case 2.1: D ≥ 3.
Compute an axis-aligned rectangle Q of width w = D, with its vertical sides along the
two covering lines, and of minimal height, which includes all representative points (on
the two covering lines). Let h denote Q’s height. Output the tour T that is the perimeter
of Q.
Case 2.2: D ≤ 3 (similar to Case 1).
Compute a smallest perimeter axis-aligned rectangle Q that touches (intersects) all regions, where Q is considered to be a two-dimensional domain. Let w and h denote Q’s
width and height respectively. Note that w ≤ 5. Let e1 , . . . , e7 be the vertical segments
of height h that partition Q into eight equal-width parts. Consider the edges of Q,
together with doubled copies of the edges e1 , . . . , e7 , to define an Eulerian multigraph,
G, having 18 vertices and 32 edges. Output any Euler tour T of G.
Case 3: The greedy cover contains at least three covering lines.
Compute T , a (1 + ǫ)-approximate tour of the representative points as the output tour.
If the regions are simple polygons, then a minimum-perimeter touching rectangle Q is determined by at most four contact points with region boundary arcs. A brute force procedure examining
all possible k-tuples, k ≤ 4, of such arcs computes Q in O(N 5 ) time. The total running time of
the approximation algorithm is either bounded by the complexity of the above step or by the complexity of computing (1 + ǫ)-approximate tours on n points, depending on the size of the greedy
cover.
Theorem 2 Given a set R of n connected regions of the same diameter in the plane, a O(1)approximation of an optimal tour can be computed in polynomial time.
15
Proof. Let OP T be an optimal region tour. We address each of the cases we distinguished in the
previous algorithm. We will use repeatedly the following simple fact (see [1]): For positive a, b, w, h
the following inequality holds
p
p
(4)
aw + bh ≤ a2 + b2 w2 + h2 .
Case 1. Write diag(Q) for the diagonal of the rectangle Q. We first argue that T visits all regions.
Since all regions are covered by a unique covering line, they lie in a vertical strip of width
√≤ 2δ = 2.
So w ≤ 2. The horizontal projection of each type (1) region (on the x-axis) is at least 1/ 2. Hence,
each region is intersected either by the boundary of Q or by one of the two vertical segments inside
√
Q, since these segments partition Q into three subrectangles each of width at most 2/3 < 1/ 2.
(Each region that is not intersected by the perimeter of Q lies entirely inside Q). Consequently, T
is a valid region tour. We first give a lower bound on |OP T |. Since the length of a tour touching
all four sides of a rectangle is at least twice the length of the diagonal of the rectangle (see [1]),
p
|OP T | ≥ 2diag(Q) = 2 w2 + h2 .
The length of T is
√ p
√
|T | = 2w + 6h = 2(w + 3h) ≤ 2 10 w2 + h2 ≤ 10|OP T |.
Case 2.1: D ≥ 3. We distinguish two sub-cases.
Case 2.1.a: h ≤ 2. Recall that w = D ≥ 3.
|OP T | ≥ 2(w − 2),
|T | = 2w + 2h ≤ 2w + 4 ≤ 10(w − 2) ≤ 5|OP T |.
Case 2.1.b: h ≥ 2. Since each region has unit diameter, the optimal tour may lie inside
Q at distance at most one from its boundary, so it must touch all four sides of some rectangle Q′
having width ≥ w − 2 and height ≥ h − 2. Hence
p
|OP T | ≥ 2 (w − 2)2 + (h − 2)2 .
Since 8w + 8h ≥ 40, we have
√
√ p
|T | = 2w + 2h ≤ 10((w − 2) + (h − 2)) ≤ 10 2 (w − 2)2 + (h − 2)2 ≤ 5 2|OP T |.
√
Case 2.2: D ≤ 3. The horizontal projection of each region is at least 1/ 2 and the rectangle Q,
of width w ≤ 5, is partitioned
√ into 8 subrectangles, each of width ≤ 5/8, by the vertical segments
e1 , . . . , e7 . Thus, since 1/ 2 > 5/8, T visits all regions. (As in Case 1, each region that is not
intersected by the perimeter of Q lies entirely inside Q; hence, it is intersected by one of the seven
vertical segments.) A similar calculation yields
p
|OP T | ≥ 2diag(Q) = 2 w2 + h2 .
p
p
p
|T | = 2w + 16h ≤ 2 12 + 82 w2 + h2 ≤ 2(8.1) w2 + h2 ≤ 8.1|OP T |.
Case 3. Partition the optimal tour OP T into blocks OP Ti , with i ≥ 1. OP T1 starts at an arbitrary
point of intersection of OP T with the leftmost covering line, and ends at the last intersection of
16
OP T with the second from the left covering line, before OP T intersects a different covering line.
Notice that OP T does not cross to the left of the leftmost covering line. In general, the blocks
OP Ti are determined by the last point of intersection of OP T with a covering line, before OP T
crosses a different covering line. In Figure 5, the hypothetical optimum tour is partitioned into six
blocks AB, BC, CD, DE, EF, F A. For example, OP T crosses the second vertical line twice before
crossing the third vertical line, and B is the last point of intersection with the second line. Consider
the bounding box Q of OP Ti , the smallest perimeter axis-aligned rectangle which includes OP Ti .
Write w for its width and h for its height. There are two cases to consider.
Case 3.1: OP Ti intersects regions stabbed by two consecutive covering lines only, l1 , l2 say, at
distance w1 . This implies that OP Ti lies between these two covering lines, so w = w1 . Without
loss of generality, OP Ti touches the lower side of Q, at some point C, before it touches
√ the upper
side of Q,
√ at some point D. Since the horizontal projection of each region is at least 1/ 2, we have
w1 ≥ 1/ 2. Let 0 ≤ a, b ≤ h specify the starting and ending points A and B of OP Ti ; see Figure 6.
Namely, a is the vertical distance between A, the start point of OP Ti , and the lower horizontal side
of Q, and b is the vertical distance between B, the end point of OP Ti , and the upper horizontal
side of Q. By considering the reflections of OP Ti with respect to the two horizontal sides of Q, we
get that |OP Ti | is bounded from below by the length of the polygonal line A′ CDB ′ :
p
(5)
|OP Ti | ≥ |A′ CDB ′ | ≥ (h + a + b)2 + w1 2 .
B’
b
b
B
D
E
B
h
A
1 F
1
h
A
a
a
C
w = w1
D 1
A’
l1
l2
w = w1
l1
1
G
l2
Figure 6: Case 3.1: 2 lines.
We show that there exists a partial tour Ti (a path) of the representative points of all the regions
OP Ti visits, of length bounded by c|OP Ti | for some positive constant c, where Ti starts at A and
ends at B. Take Ti = ADEF GB, where the points D, E, F, G are on the lines l1 , l2 at unit distance
from the corners of Q; see Figure 6.
|Ti | ≤ (a + 1) + (1 + h + 1) + (w1 ) + (1 + h + 1) + (1 + h − b) ≤ 3(h + a + b) + (w1 + 6)
q
p
√
√
≤ 3(h + a + b) + (6 2 + 1)w1 ≤ 32 + (6 2 + 1)2 (h + a + b)2 + w1 2 < 9.95|OP Ti |. (6)
Put Ti together to get a tour T of the representative points, having length smaller than 9.95|OP T |.
Then, for ǫ ≤ 0.05, the (1 + ǫ) approximate tour of representatives, which the algorithm actually
computes, has length at most 10|OP T |.
17
Case 3.2: OP Ti intersects regions stabbed by three consecutive covering lines only, l1 , l2 and l3 .
Denote by w1 (resp., w2 ) the horizontal distance between l1 and l2 (resp., l2 and l3 ). This implies
that OP Ti lies between these three covering lines, but it does not touch the rightmost line l3 . We
can assume that OP Ti starts at A and ends at B, and as in the previous case, that it touches the
′
lower side of Q, at some point C, before it touches the upper side of Q, at some point D. Let l
′
′
be the supporting line
side of Q, and w be the horizontal distance between l2 and l .
√ of the right
′
We have w1 , w2 ≥ 1/ 2 and w ≥ max(0, w2 − 1). The case when OP Ti touches the upper side of
′
Q before it touches l is shown is Figure 7; the other case is similar and we get the same bound on
|OP Ti |. Let a, b, A, B be as before. We distinguish two sub-cases:
Case 3.2.a: w2 ≤ 1. The lower bound on OP Ti we have used earlier is still valid.
p
|OP Ti | ≥ (h + a + b)2 + w1 2 .
We show a partial tour Ti of the representative points of all the regions OP Ti visits. Take Ti =
ADEF BGHIF B; see Figure 7. Points E, F and I (resp., D, G and H) are on the lines l1 , l2 and
l3 at unit vertical distance above the upper side (resp., below the lower side) of Q.
|Ti | ≤ (a + 1) + (1 + h + 1) + w1 + (1 + h + 1) + w2 + (1 + h + 1) + w2 + (1 + b)
√
≤ 3(h + a + b) + (w1 + 2w2 + 8) ≤ 3(h + a + b) + (10 2 + 1)w1
q
p
√
≤ 32 + (10 2 + 1)2 (h + a + b)2 + w1 2 < 15.45|OP Ti |.
(7)
Put Ti together to get a tour T of the representative points, having length smaller than 15.45|OP T |.
Then, for ǫ ≤ 0.05, the (1 + ǫ)-approximate tour of representatives has length at most 15.5|OP T |.
w
′
h
b
D
b
B
A
a
a
l1
w
l2
I
B
h
C w1
F
E
′
D
l
h
A
′
l1
w1
w2
G
l2
H
l3
Figure 7: Case 3.2: 3 lines.
Case 3.2.b: w2 ≥ 1. We use a different lower bound on OP Ti .
p
|OP Ti | ≥ (h + a + b)2 + (w1 + 2w2 − 2)2 ,
(8)
which we get by considering the reflections of OP Ti with respect to the two horizontal sides of Q
′
and with respect to l ; see Figure 7. Take Ti = ADEF BGHIF B as in Case 3.2.a.
√
|Ti | ≤ 3(h + a + b) + (w1 + 2w2 + 8) ≤ 3(h + a + b) + (10 2 + 1)(w1 + 2w2 − 2)
18
≤
q
p
√
32 + (10 2 + 1)2 (h + a + b)2 + (w1 + 2w2 − 2)2 < 15.45|OP Ti |.
(9)
Put Ti together to get a tour T of the representative points, having length smaller than 15.45|OP T |.
Then, for ǫ ≤ 0.05, the (1 + ǫ)-approximate tour of representatives has length at most 15.5|OP T |.
The overall approximation ratio of the algorithm, derived from the Combination Lemma is 15.5 +
15.5 + 2 = 33.
✷
Comparison with [1]. We point out some of the similarities and differences between the techniques of Arkin and Hassin [1] and our algorithm and its analysis, which is based on theirs. First,
in [1], three different algorithms are presented, for parallel equal-length segments, translates of a
convex region, and translates of an arbitrary connected region. The second two algorithms are
refinements of the first, and representative points are chosen differently in each case. We presented
here a single algorithm that works for all regions of type (1). Our consideration of cases 1 and 2
are slightly different from the corresponding cases of [1], which allows us to handle a larger class
of inputs, namely regions of type (1); e.g. in case 2, [1] distinguishes between the subcases D ≤ 1
and D ≥ 1, while we distinguish between D ≤ 3 and D ≥ 3, and the cases are treated slightly
differently. Finally, the analysis of the algorithms is based on similar ideas, e.g. to divide the
optimal tours into blocks; our analysis differs in being able to address a single algorithm and a
larger class of inputs.
Some Special Cases
We note that our calculations of the approximation ratio for connected regions of a same diameter, give improved bounds for three cases addressed in [1]:
√
√
1. parallel equal segments, from 3 2 + 1, to 3 2,
√
√
2. translates of a convex region, from 32 + 72 + 1, to 32 + 72 ,
√
√
3. translates of a connected region, from 32 + 112 + 1, to 32 + 112 .
The reason why these cases can be improved, and the new approximation ratios have similar
expressions with the old ones, is that our algorithm A is similar to the algorithms in [1] for those
cases.
We exemplify here the case of parallel equal segments and omit details for the rest. The
algorithm computes a greedy cover of the segments, assumed to be of unit length, using vertical
lines. Then it proceeds
√ according to the cardinality of the cover. Cases 1 and 2 are treated in
[1], and the ratio is 2; this is not the bottleneck case. In Case 1 (one covering line), an optimal
tour is easy to obtain. In Case 2 (two covering lines), a smallest aligned rectangle which touches
all segments is the output tour. In Case 3 (three or more covering lines) the algorithm computes
an almost optimal tour of the representative points, as algorithm A does. Its analysis is divided
into two sub-cases, as in the proof of Theorem 2. In the first sub-case (OP Ti intersects segments
covered by two consecutive covering lines only), the lower bound in Equation (5) on |OP Ti | is valid.
The upper bound in Equation (6) on |Ti | is adjusted by dropping the constant term equal to +6.
Then
p
√
|Ti | ≤ 3(h + a + b) + w1 ≤ 32 + 12 |OP Ti | = 10|OP Ti |.
19
In the second sub-case (OP Ti intersects segments covered by three consecutive covering lines only),
the lower bound in Equation (8) on |OP Ti | is valid. The upper bound in Equation (9) on |Ti | is
adjusted by dropping the constant term equal to +8. We also have that w1 , w2 > 1. Then
√
|Ti | < 3(h + a + b) + w1 + 2w2 ≤ 3(h + a + b) + 3(w1 + 2w2 − 2) ≤ 3 2|OP Ti |.
√
The overall approximation ratio obtained for parallel equal segments is 3 2.
5
Lines
We consider now the case in which the n regions defining the TSPN instance are infinite straight
lines in the plane. It is interesting that this case allows for an exact solution in polynomial time:
Proposition 3 Given a set L of n infinite straight lines in the plane, a shortest tour that visits L
can be computed in polynomial time.
Proof. We convert the problem to an instance of the watchman route problem in a simple polygon.
A watchman route in a polygon is a tour inside the polygon, such that every point in the polygon
is visible from some point along the tour. The watchman route problem asks for a watchman route
of minimum length [3]. The problem is known to have an O(n5 ) algorithm; see [26], as well as [8].
Let B be a rectangle that contains all of the vertices of the arrangement of L. At any one of the
two points of intersection between a line li ∈ L and the boundary of B, we extend a very narrow
“spike” outward from that point, along li , for some fixed distance. Let P be the simple polygon
having 3n + 4 vertices that is the union of B and these n spikes, as illustrated in Figure 8. We
make the observation that a tour T visits all of the lines in L if and only if it sees all of the polygon
P , as required by the watchman route problem. Consequently, we can solve the TSPN on the set
of lines L by solving the watchman route problem on polygon P .
✷
Figure 8: Proof of Proposition 3.
Given the high running time of the watchman route algorithms, it is of interest to consider
more efficient algorithms that may approximate the optimal solution. To this end, we now present
a linear-time constant-factor approximation algorithm.
20
Let L = {l1 , . . . , ln } be the input set of n lines. A minimum touching circle (disk) is a circle of
minimum radius which intersects all of the lines in L. The algorithm computes and outputs CL , a
minimum touching circle for L. We will show that this provides a tour of length at most π2 |OP T |,
where, as usual, OP T denotes an optimal tour. First we argue about the approximation ratio, and
leave for later the presentation of the algorithm. To start, we assume for simplicity that no two
lines are parallel, though this assumption will be later removed.
Observation 1 The optimal tour of a set of lines is a (possibly degenerate) convex polygon P .
Proof. It is easy to see that an optimal tour must be polygonal, consisting of a finite union of
straight line segments. If an optimal tour were a non-convex polygon P , we obtain a contradiction
to its optimality since the boundary of the convex hull of P , which is shorter than P , also visits all
of the lines L, since P does.
✷
Observation 2 CL is determined by 3 lines in L, i.e. it is the inscribed circle in the triangle ∆
formed by these 3 lines.
We distinguish two cases:
Case 1. ∆ is an acute triangle. It is well known that for an acute triangle, the minimum perimeter
inscribed triangle (having a vertex on each side of the triangle) is its pedal triangle, whose vertices
are the feet of the altitudes of the given triangle (see e.g. [22]). So in this case, the optimal tour
OP T∆ , which visits the 3 lines of ∆ is its pedal triangle; we denote by y its perimeter. Clearly y is
a lower bound on |OP T |: |OP T | ≥ |OP T∆ | = y. Denote by s the semi-perimeter of ∆, by R the
radius of its circumscribed circle, and by r the radius of its inscribed circle.
Fact 1 For an acute triangle ∆, s > 2R.
Proof. If A, B, C are the angles of ∆, this is equivalent to
R(sin A + sin B + sin C) > 2R.
After simplification with R, this is a well known inequality in the geometry of an acute triangle
([6], page 18).
✷
Fact 2 For an acute triangle ∆, y =
2rs
R .
A proof of this equality can be found in [6], page 86.
Claim 1 For an acute triangle ∆, r < y4 .
Proof. Putting the above together, we get r =
Ry
2s
<
The length of the output tour is bounded as follows
21
Ry
4R
= y4 .
✷
y
π
π
= y ≤ |OP T |.
4
2
2
Case 2. ∆ is an obtuse triangle. In this case, |OP T∆ | = 2h, where h is the length of the altitude
corresponding to the obtuse angle, say A. Clearly 2h is a lower bound on |OP T |: |OP T | ≥
|OP T∆ | = 2h.
|CL | = 2πr < 2π
Fact 3 For any triangle ∆, h > 2r.
Proof. Let S denote the area of the triangle with side lengths a, b, and c. Then, we know that
S = ah/2, by our definition of the altitude h. We also know from elementary geometry that
S = (a + b + c)r/2, recalling that r is the radius of the inscribed circle of the triangle. Thus,
ah = (a + b + c)r, from which we obtain
a+b+c
a+a
h
=
>
= 2,
r
a
a
using the triangle inequality.
✷
Using this inequality, we have |OP T∆ | = 2h > 4r. The length of the output tour is bounded as
follows
π
π
π
|CL | = 2πr = (4r) < |OP T∆ | ≤ |OP T |.
2
2
2
Thus, in both cases, the approximation ratio is π2 ≤ 1.58.
When CL is determined by 3 lines in L of which two are parallel at distance h, we say that they
form a generalized triangle ∆. Clearly 2h is a lower bound on |OP T |: |OP T | ≥ |OP T∆ | = 2h. We
also have r = h/2, where r is the radius of CL , thus
π
|CL | = 2πr = πh ≤ |OP T |.
2
We now describe the algorithm for computing CL . The distance d(p, l) from a point p of
coordinates (x0 , y0 ) to a line l of equation ax + by + c = 0 is
d(p, l) =
|ax0 + by0 + c|
√
.
a2 + b2
Let the lines in L have the equations
(li :)
ai x + bi y + ci = 0, i = 1, . . . , n.
Finding a minimum touching circle amounts to finding the center coordinates (x, y), and a minimum
radius z, s.t.
|ai x + bi y + ci |
q
≤ z, i = 1, . . . , n.
a2i + b2i
This is equivalent to solving the following 3-dimensional linear program
a x + b y + c
−ai x − bi y − ci
i
i
q i
q
≤ z,
≤ z, i = 1, . . . , n ,
min z subject to
a 2 + b2
a2 + b2
i
i
i
i
which takes O(n) time [16]. Consequently, we have proved
Theorem 3 Given a set L of n (infinite straight) lines in the plane, a
visits L can be computed in O(n) time.
22
π
2 -approximate
tour that
6
Conclusion
Several open problems remain, including
(1) Is there a constant-factor approximation algorithm for arbitrary connected regions in the plane?
What if the regions are disconnected? (giving us a geometric version of a “one-of-a-set TSP”)
(2) What approximation bounds can be obtained in higher dimensions? Our packing arguments
for disjoint disks lift to higher dimensions, but our other methods do not readily generalize.
A particularly intriguing special case is the generalization of the case of infinite straight lines:
What can be said in 3-space for the TSPN on a set of lines or of planes?
(3) Is there a PTAS for general pairwise-disjoint regions in the plane?
Acknowledgements
We thank Estie Arkin and Michael Bender for several useful discussions on the TSPN problem. We
thank the anonymous referees for their detailed comments and suggestions, which greatly improved
the paper.
References
[1] E. M. Arkin and R. Hassin, Approximation algorithms for the geometric covering salesman
problem, Discrete Appl. Math., 55:197–218, 1994.
[2] S. Arora, Nearly linear time approximation schemes for Euclidean TSP and other geometric
problems, J. of the ACM, 45(5):1–30, 1998.
[3] T. Asano, S. K. Ghosh, and T. C. Shermer, Visibility in the plane, In Handbook of Computational Geometry (J.-R. Sack and J. Urrutia, editors), pages 829–876. Elsevier Science
Publishers B.V. North-Holland, Amsterdam, 2000.
[4] J. L. Bentley, Fast algorithms for geometric traveling salesman problems, ORSA J. on Comput.,
4(4):387–411, 1992.
[5] M. de Berg, J. Gudmundsson, M. J. Katz, C. Levcopoulos, M. H. Overmars, and A. F. van der
Stappen, TSP with neighborhoods of varying size, Proc. 10th Annual European Symposium on
Algorithms, to appear, September, 2002.
[6] O. Bottema, R. Ž. Djordjević, R. Janić, D. S. Mitrinović and P. M. Vasić, Geometric Inequalities, Wolters-Noordhoff, Groningen, 1969.
[7] P. Berman and M. Karpinski, On some tighter inapproximability results, Technical Report
TR98-065, ECCC, 1998.
[8] S. Carlsson, H. Jonsson and B. J. Nilsson, Finding the shortest watchman route in a simple
polygon, Discrete Comput. Geom. 22(3):377–402, 1999.
[9] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory of
NP-Completeness, W. H. Freeman, New York, NY, 1979.
23
[10] J. Gudmundsson and C. Levcopoulos, A fast approximation algorithm for TSP with neighborhoods, Nordic J. of Comput., 6:469–488, 1999.
[11] J. Gudmundsson and C. Levcopoulos, Hardness result for TSP with neighborhoods, Technical Report LU-CS-TR:2000-216, Department of Computer Science, Lund University, Sweden,
2000.
[12] H. Jonsson, The traveling salesman problem for lines in the plane, Inform. Process. Lett.,
82(3):137–142, 2002.
[13] M. Jünger, G. Reinelt and G. Rinaldi, The traveling salesman problem, In Network Models,
Handbook of Operations Research/Management Science (M. O. Ball, T. L. Magnanti, C. L.
Monma, and G. L. Nemhauser, editors), pages 225–330, Elsevier Science, Amsterdam, 1995.
[14] E. L. Lawler, J. K. Lenstra, A. H. G. Rinnooy Kan and D. B. Shmoys, editors, The Traveling
Salesman Problem. John Wiley & Sons, New York, NY, 1985.
[15] C. Mata and J. S. B. Mitchell, Approximation algorithms for geometric tour and network
design problems, Proc. 11th Annu. ACM Sympos. Comput. Geom., pages 360–369, 1995.
[16] N. Megiddo, Linear programming in linear time when the dimension is fixed, J. of ACM,
31:114–127, 1984.
[17] J. S. B. Mitchell, Guillotine subdivisions approximate polygonal subdivisions: A simple
polynomial-time approximation scheme for geometric TSP, k-MST, and related problems,
SIAM J. on Comput., 28(4):1298–1309, 1999.
[18] J. S. B. Mitchell. Guillotine subdivisions approximate polygonal subdivisions: Part III – Faster
polynomial-time approximation schemes for geometric network optimization. Manuscript, University at Stony Brook, 1997.
[19] J. S. B. Mitchell. Approximation algorithms for geometric optimization problems. In Proc.
Ninth Canadian Conference on Computational Geometry, Queen’s University, Kingston,
Canada, August 11-14, 1997, pp. 229–232.
[20] J. S. B. Mitchell, Geometric shortest paths and network optimization, In Handbook of Computational Geometry (J.-R. Sack and J. Urrutia, editors), pages 633–701. Elsevier Science
Publishers B.V. North-Holland, Amsterdam, 2000.
[21] C. H. Papadimitriou, The Euclidean traveling salesman problem is NP-complete, Theoret.
Comput. Sci., 4:237–244, 1977.
[22] H. Rademacher and O. Toeplitz, The Enjoyment of Mathematics, Princeton University Press,
1957; a translation from Von Zahlen und Figuren, Springer, Berlin, 1933.
[23] S. B. Rao and W. D. Smith, Approximating geometrical graphs via “spanners” and “banyans”,
Proc. 30th Annu. ACM Sympos. Theory Comput., 540–550, 1998.
[24] G. Reinelt, Fast heuristics for large geometric traveling salesman problems, ORSA J. Comput.,
4:206–217, 1992.
24
[25] O. Schwartz and S. Safra, On the complexity of approximating TSP with neighborhoods and
related problems, Manuscript (submitted), July, 2002.
[26] X. Tan, Fast computation of shortest watchman routes in simple polygons, Inform. Process.
Lett., 77(1):27–33, 2001.
25
| 8cs.DS
|
Using Trusted Data to Train Deep Networks on
Labels Corrupted by Severe Noise
Dan Hendrycks * 1 Mantas Mazeika * 1 Duncan Wilson 2 Kevin Gimpel 3
arXiv:1802.05300v1 [cs.LG] 14 Feb 2018
Abstract
The growing importance of massive datasets with
the advent of deep learning makes robustness to
label noise a critical property for classifiers to
have. Sources of label noise include automatic labeling for large datasets, non-expert labeling, and
label corruption by data poisoning adversaries. In
the latter case, corruptions may be arbitrarily bad,
even so bad that a classifier predicts the wrong labels with high confidence. To protect against such
sources of noise, we leverage the fact that a small
set of clean labels is often easy to procure. We
demonstrate that robustness to label noise up to
severe strengths can be achieved by using a set of
trusted data with clean labels, and propose a loss
correction that utilizes trusted examples in a dataefficient manner to mitigate the effects of label
noise on deep neural network classifiers. Across
vision and natural language processing tasks, we
experiment with various label noises at several
strengths, and show that our method significantly
outperforms existing methods.
1. Introduction
Robustness to label noise is set to become an increasingly
important property of supervised learning models. With
the advent of deep learning, the need for more labeled data
makes it inevitable that not all examples will have highquality labels. This is especially true of data sources that
admit automatic label extraction, such as web crawling for
images, and tasks for which high-quality labels are expensive to produce, such as semantic segmentation or parsing.
Additionally, label corruption may arise in data poisoning
(Li et al., 2016; Steinhardt et al., 2017). Both natural and
malicious label corruption are known to sharply degrade the
performance of classification systems (Zhu & Wu, 2004).
We consider the scenario where we have access to a large set
of examples with potentially corrupted labels and determine
*
Equal contribution 1 University of Chicago 2 Foundational Research Institute 3 Toyota Technological Institute at Chicago. Correspondence to: Mantas Mazeika <[email protected]>.
True
GLC (Ours)
1.0
0.8
0.6
Forward
Confusion Matrix
0.4
0.2
0.0
Figure 1: A label corruption matrix C (top left) and three
matrix estimates for a corrupted CIFAR-10 dataset. Entry
Cij is the probability that a label of class i is corrupted to
class j, or Cij = p(ỹ = j|y = i). Our estimate matches
the true corruption matrix closer than the confusion matrix
and the Forward method. Further comparisons and method
descriptions are in Section 4.3.
how much can be gained from access to a small set of
examples where labels are considered gold standard. This
scenario is realistic, as it is usually the case that a number of
trusted examples have been gathered in the validation and
test sets, and that more could be gathered if necessary.
To leverage the additional information from trusted labels,
we propose a new loss correction and empirically verify it on
a number of vision and natural language datasets with label
corruption. Specifically, we demonstrate recovery from
extremely high levels of label noise, including the dire case
when the untrusted data has a majority of its labels corrupted.
Such severe corruption can occur in adversarial situations
like data poisoning, or when the number of classes is large.
In comparison to loss corrections that do not employ trusted
data (Patrini et al., 2016), our method is significantly more
Using Trusted Data to Train Deep Networks
accurate in problem settings with moderate to severe label
noise. Relative to a recent method which also uses trusted
data (Li et al., 2017), our method is far more data-efficient
and generally more accurate. These results demonstrate that
systems can weather label corruption with access only to a
small number of gold standard labels. The code is available
at https://github.com/mmazeika/glc.
2. Related Work
The performance of machine learning systems reliant on
labeled data has been shown to degrade noticeably in the
presence of label noise (Nettleton et al., 2010; Pechenizkiy
et al., 2006). In the case of adversarial label noise, this
degredation can be even worse (Reed et al., 2014). Accordingly, modeling, correcting, and learning with noisy labels
has been well studied (Natarajan et al., 2013; Biggio et al.,
2011; Frénay & Verleysen, 2014).
The methods of (Mnih & Hinton, 2012), (Larsen et al.,
1998), (Patrini et al., 2016) and (Sukhbaatar et al., 2014)
allow for label noise robustness by modifying the model’s
architecture or by implementing a loss correction. Unlike
(Mnih & Hinton, 2012) who focus on binary classification
of aerial images and (Larsen et al., 1998) who assume the
labels are symmetric (i.e., that the noise and the labels are
independent), (Patrini et al., 2016) and (Sukhbaatar et al.,
2014) consider label noise in the multi-class problem setting
with asymmetric labels.
In (Sukhbaatar et al., 2014), the authors introduce a stochastic matrix measuring label corruption, note its inability to be
calculated without access to the true labels, and propose a
method of forward loss correction. Forward loss correction
adds a linear layer to the end of the model and the loss is
adjusted accordingly to incorporate learning about the label
noise. In the work of (Patrini et al., 2016), they also make
use of the forward loss correction mechanism, and propose
an estimate of the label corruption estimation matrix which
relies on strong assumptions and no clean labels.
Contra (Sukhbaatar et al., 2014; Patrini et al., 2016), we
make the assumption that during training the model has access to a small set of clean labels and use this to create our
label noise correction. This assumption has been leveraged
by others for the purpose of label noise robustness, most notably (Veit et al., 2017; Li et al., 2017; Xiao et al., 2015), and
tenuously relates our work to the field of semi-supervised
learning (Zhu, 2005; Chapelle et al., 2010). In (Veit et al.,
2017), human-verified labels are used to train a label cleaning network by estimating the residuals between the noisy
and clean labels in a multi-label classification setting. In the
multi-class setting that we focus on in this work, (Li et al.,
2017) propose distilling the predictions of a model trained
on clean labels into a second network trained on the noisy
labels and the predictions of the first. Our work differs from
these two in that we do not train neural networks on the
clean labels alone.
3. Gold Loss Correction
e of u examples (x, ỹ),
We are given an untrusted dataset D
and we assume that these examples are potentially corrupted
examples from the true data distribution p(x, y) with K
classes. Corruption is according to a label noise distribution
p(ỹ | y, x). We are also given a trusted dataset D of t
examples drawn from p(x, y), where t/u 1. We refer
to t/u as the trusted fraction. Concretely, a web scraper
labeling images from metadata may produce an untrusted
set, while expert-annotated examples would form a trusted
dataset and be a gold standard.
In leveraging trusted data, we focus our investigation on the
stochastic matrix correction approach used by (Sukhbaatar
et al., 2014; Patrini et al., 2016). In this approach, a stochastic matrix is applied to the softmax output of a classifier,
and the resulting new softmax output is trained to match the
noisy labeling. If the stochastic matrix is engineered so as to
approximate the label noising procedure, this approach can
bring the original output close to the distribution of clean
labels, under moderate assumptions.
We explore two avenues of utilizing the trusted dataset to
improve this approach. The first involves directly using the
trusted data while training the final classifier. As this could
be applied to existing stochastic matrix correction methods,
we run ablation studies to demonstrate its effect. The second
avenue involves using the additional information conferred
by the clean labels to obtain a better matrix to use with
the approach. As a first approximation, one could use a
normalized confusion matrix of a classifier trained on the
untrusted dataset and evaluated on the trusted dataset. We
demonstrate, however, that this does not work as well as the
estimate used by our method, which we now describe.
Our method makes use of D to estimate the K × K matrix of corruption probabilities Cij = p(ỹ = j | y = i).
Once this estimate is obtained, we use it to train a modified
classifier from which we recover an estimate of the desired
conditional distribution p(y | x). We call this method the
Gold Loss Correction (GLC), so named because we make
use of trusted or gold standard labels.
3.1. Estimating The Corruption Matrix
To estimate the probabilities p(ỹ | y), we make use of the
identity
p(ỹ | x) =
K
X
y=1
p(ỹ | y, x)p(y | x).
Using Trusted Data to Train Deep Networks
The left hand side of the equality can be approximated by
e Let θ̃ be the parameters
training a neural network on D.
of this network, and let p̂(ỹ | x; θ̃) be its softmax output
vector.
Given an example x and its true one-hot label y, the term
on the right reduces to p(ỹ | y, x) = p(ỹ | x). In the
case where ỹ is conditionally independent of x given y, this
further reduces to p(ỹ | x) = p(ỹ | y). In the case where ỹ
is not conditionally independent of x given y, we can still
approximate p(ỹ | y). We know
p(ỹ, y, x)
p(x | ỹ, y)p(ỹ | y)p(y)
p(ỹ | y, x) =
=
p(y, x)
p(x | y)p(y)
p(x | ỹ, y)
= p(ỹ | y)
.
p(x | y)
This forces p(ỹ | y, x)p(x | y) = p(ỹ | y)p(x | ỹ, y).
Integrating over all x gives us
Z
Z
p(ỹ | y, x)p(x | y) dx = p(ỹ | y) p(x | ỹ, y) dx
= p(ỹ | y).
We approximate the integral on the left with the expectation
of p(ỹ | y, x) over the empirical distribution of x given y.
More explicitly, let Ai be the subset of x in D with label i.
b We have
Denote our estimate of C by C.
X
bij = 1
C
p̂(ỹ = j | x)
|Ai |
x∈Ai
1 X
=
p̂(ỹ = j | y = i, x)
|Ai |
x∈Ai
≈ p(ỹ = j | y = i).
This is how we estimate our corruption matrix for GLC. The
second equality comes from noting that if y is known, the
preceding discussion implies p(ỹ | x) = p(ỹ | y, x). This
approximation relies on p̂(ỹ | x) being a good estimate of
p(ỹ | x) and on the number of trusted examples of each
class.
3.2. Training a Corrected Classifier
b we follow the method of (Sukhbaatar et al.,
Now with C,
2014; Patrini et al., 2016) to train a corrected classifier.
Given the K × 1 softmax output s of our classifier, we
b We then reinitialize θ
define the new outputs as s̃ := Cs.
and train the model p̂(s̃ | x; θ) on the noisy labels with crossentropy loss. If ỹ is conditionally independent of x given y,
and if C is nonsingular, then
it follows from the invertibility
e = argminθ L s, D given that
of C that argminθ L s̃, D
b is a perfect estimate of C.
C
We find using s̃ to work well in practice, even for some
singular corruption matrices. We can further improve on
this method by using the data in the trusted set to train
the corrected classifier. On examples from the trusted set
b to the
encountered during training, we temporarily set C
identity matrix to turn off the correction. This has the effect of allowing our label correction to handle a degree of
instance-dependency in the label noise (Menon et al., 2016).
A summary of our method is in the algorithm below.
Algorithm G OLD L OSS C ORRECTION (GLC)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
e loss `
Input: Trusted data D, untrusted data D,
e with loss `
Train network f (x) = p̂(e
y |x; θ) on D
b ∈ RK×K with zeros, K the number of classes
Fill C
for k = 1, . . . , K do
num_examples = 0
for (xi , yi ) ∈ D such that yi = k do
num_examples += 1
b k += f (xi ) {add f (xi ) to kth column}
C
end for
b k /= num_examples
C
end for
Initialize new model g(x) = p̂(y|x; θ)
e
b
Train g(x) with `(g(x), y) on D and `(Cg(x),
ye) on D
Output: Model p̂(y|x; θ)
•
•
4. Experiments
We empirically demonstrate GLC on a variety of datasets
and architectures under several types of label noise.
4.1. Description
Generating Corrupted Labels. Suppose our dataset has
t + u examples. We sample a set of t datapoints D, and the
e which we probabilistically
remaining u examples form D,
corrupt according to a true corruption matrix C. Note that
we do not have knowledge of which of our u untrusted examples are corrupted. We only know that they are potentially
corrupted.
e
To generate the untrusted labels from the true labels in D,
we first obtain a corruption matrix C. Then, for an example
with true label i, we sample the corrupted label from the
categorical distribution parameterized by the ith row of C.
Comparing Loss Correction Methods. The GLC differs
from previous loss corrections for label noise in that it reasonably assumes access to a high-quality annotation source.
Therefore, to compare to other loss correction methods,
we ask how each method performs when starting from the
same dataset with the same label noise. In other words, the
only additional information our method uses is knowledge
of which examples are trusted, and which are potentially
Using Trusted Data to Train Deep Networks
0.4
0.2
1.0
CIFAR-10, Uniform, 10% trusted
0.8
0.8
0.6
0.6
0.6
0.4
0.0
0.0
1.0
0.4
0.2
0.2 0.4 0.6 0.8
Corruption Strength
1.0
IMDB, Flip, 5% trusted
0.2
0.0
0.0
1.0
0.2 0.4 0.6 0.8
Corruption Strength
1.0
CIFAR-100, Hier., 10% trusted
0.0
0.0
0.6
0.6
0.6
0.6
0.2
0.0
0.0
0.2
0.2 0.4 0.6 0.8
Corruption Strength
1.0
0.0
0.0
Test Error
0.8
0.4
1.0
0.0
0.0
0.4
0.2
0.2
0.2 0.4 0.6 0.8
Corruption Strength
1.0
CIFAR-100, Uniform, 10% trusted
0.8
0.4
0.2 0.4 0.6 0.8
Corruption Strength
1.0
0.8
0.4
CIFAR-100, Flip, 10% trusted
0.4
0.8
Test Error
Test Error
1.0
0.2 0.4 0.6 0.8
Corruption Strength
1.0
0.8
0.2
0.0
0.0
SST, Flip, 1% trusted
1.0
Test Error
0.6
MNIST, Flip, 1% trusted
Test Error
Test Error
0.8
1.0
Test Error
GLC (Ours)
Distillation
Forward Gold
Forward
No Correction
Test Error
CIFAR-10, Flip, 10% trusted
1.0
0.2 0.4 0.6 0.8
Corruption Strength
1.0
0.0
0.0
0.2 0.4 0.6 0.8
Corruption Strength
1.0
Figure 2: Error curves for the compared methods across a range of corruption strengths on different datasets.
corrupted.
4.2. Datasets, Architectures, and Noise Corrections
MNIST. The MNIST dataset contains 28 × 28 grayscale
images of the digits 0-9. The training set has 50,000 images
and the test set has 10,000 images. For preprocessing, we
rescale the pixels to a unit range.
We train a 2-layer fully connected network each with dimension 256. The network is again optimized with Adam for
10 epochs, all while using batches of size 32 and a learning
rate of 0.001. For regularization, we use `2 weight decay on
all layers with λ = 1 × 10−6 .
CIFAR. The two CIFAR datasets contain 32×32×3 color
images. CIFAR-10 has ten classes, and CIFAR-100 has 100
classes. CIFAR-100 has 20 “superclasses” which partition
its 100 classes into 20 semantically similar sets. We use
these superclasses for hierarchical noise. Both datasets have
50,000 training images and 10,000 testing images. For both
datasets, we train a Wide Residual Network (Zagoruyko
& Komodakis, 2016) of depth 40. We train for 75 epochs
using a widening factor of 2 and stochastic gradient descent
with restarts (Loshchilov & Hutter, 2016).
IMDB. The IMDB Large Movie Reviews dataset (Maas
et al., 2011) contains 50,000 highly polarized movie reviews
from the Internet Movie Database, split evenly into train
and test sets. We pad and clip reviews to a length of 200
tokens, and learn 50-dimensional word vectors from scratch
for a vocab size of 5,000.
We train an LSTM with 64 hidden dimensions on this data.
We train using the Adam optimizer (Kingma & Ba, 2014)
for 3 epochs with batch size 64 and the suggested learning
rate of 0.001. For regularization, we use dropout (Srivastava
et al., 2014) on the linear output layer with a keep probability
of 0.8.
Twitter. The Twitter Part of Speech dataset (Gimpel et al.,
2011) contains 1,827 tweets annotated with 25 POS tags.
The training set has 1,000 tweets, and the test set has 500.
We use pretrained 50-dimensional word vectors, and for
each token, we concatenate word vectors in a fixed window
centered on the token. These form our training and test
set. We use a window size of 3, and train a 1-layer fully
connected network with hidden size 256, and use the nonlinearity from (Hendrycks & Gimpel, 2016). We train using
the Adam optimizer for 15 epochs with batch size 64 and
learning rate 0.001. For regularization, we use `2 weight
decay with λ = 5 × 10−5 on all but the linear output layer.
SST. The Stanford Sentiment Treebank dataset consists of
single sentence movie reviews. There are 8,544 reviews in
the training set and 2,210 in the test set. We use binarized labels for sentiment classification. Moreover, we pad and clip
reviews to a length of 200 tokens and learn 100-dimensional
word vectors from scratch for a vocab size of 10,000.
Our classifier is a word-averaging model with an affine
output layer. We use the Adam optimizer for 5 epochs with
batch size 50 and learning rate 0.001. For regularization, we
CIFAR-100
CIFAR-10
MNIST
Using Trusted Data to Train Deep Networks
Corruption
Type
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Hierarchical
Hierarchical
Hierarchical
Mean
Percent
Trusted
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
Trusted
Only
37.6
12.9
6.6
37.6
12.9
6.6
19.0
39.6
31.3
17.4
39.6
31.3
17.4
29.4
82.4
67.3
52.2
82.4
67.3
52.2
82.4
67.3
52.2
67.3
No
Forward
Correction Correction
12.9
14.5
12.3
13.9
9.3
11.8
50.1
51.7
51.1
48.8
47.7
50.2
30.6
31.8
31.9
9.1
31.9
8.6
32.7
7.7
53.3
38.6
53.2
36.5
52.7
37.6
42.6
23.0
48.8
47.7
48.4
47.2
45.4
43.6
62.1
61.6
61.9
61.0
59.6
57.5
50.9
51.0
51.9
50.5
54.3
47.0
53.7
51.9
Forward
Gold
13.5
12.3
9.2
41.4
36.4
37.1
25.0
27.8
20.6
27.1
47.8
51.0
49.5
37.3
49.6
48.9
46.0
62.6
62.2
61.4
52.4
52.1
51.1
54.0
Distillation Confusion
Correction Matrix
42.1
21.8
9.2
15.1
5.8
11.0
46.5
11.7
32.4
5.6
28.2
3.8
27.4
11.5
29.7
22.4
18.3
22.7
11.6
16.7
29.7
8.1
18.1
8.2
11.8
7.1
19.9
14.2
87.5
53.6
61.2
49.7
39.8
39.6
87.1
28.6
61.8
26.9
40.0
25.1
87.1
45.8
61.7
38.8
39.7
29.7
62.9
37.5
GLC
(Ours)
10.3
6.3
4.7
3.4
2.9
2.6
5.0
9.0
6.9
6.4
6.6
6.2
6.1
6.9
42.4
33.9
27.3
27.1
25.8
24.7
34.8
30.2
25.4
30.2
Table 1: Vision dataset results. Percent trusted is the trusted fraction multiplied by 100. Unless otherwise indicated, all
values are percentages representing the area under the error curve computed at 11 test points. The best mean result is shown
in bold.
use `2 weight decay with λ = 1 × 10−4 on the output layer.
Forward Loss Correction. The forward correction
b by trainmethod from (Patrini et al., 2016) also obtains C
ing a classifier on the noisy labels, and using the resulting
softmax probabilities. However, this method does not make
use of a trusted fraction of the training data. Instead, it uses
the argmax at the 97th percentile of softmax probabilities
for a given class as a heuristic for detecting an example that
is truly a member of said class. As in the original paper, we
replace this with the argmax over all softmax probabilities
for a given class on CIFAR-100 experiments. The estimate
b is then used to train a corrected classifier in the same way
C
as GLC.
Forward Gold. To examine the effect of training on
trusted labels as done by GLC, we augment the Forward
b estimate with the identity on
method by replacing its C
trusted examples. We refer to the resulting method as Forward Gold, which can be seen as an intermediate method
between Forward and GLC.
Distillation. The distillation method of (Li et al., 2017)
involves training a neural network on a large trusted dataset
and using this network to provide soft targets for the untrusted data. In this way, labels are “distilled” from a neural
network. If the classifier’s decisions for untrusted inputs
are less reliable than the original noisy labels, then the network’s utility is limited. Thus, to obtain a reliable neural
network, a large trusted dataset is necessary. A new classifier is trained using labels that are a convex combination of
the soft targets and the original untrusted labels.
4.3. Uniform, Flip, and Hierarchical Corruption
Corruption-Generating Matrices. We consider three
types of corruption matrices: corrupting uniformly to all
classes, i.e. Cij = 1/K, flipping a label to a different class,
and corrupting uniformly to classes which are semantically
similar. In order to create a uniform corruption at different
strengths, we take a convex combination of an identity matrix and the matrix 11T /K. We refer to the coefficient of
11T /K as the corruption strength for a “uniform” corruption. A “flip” corruption at strength m involves, for each
row, giving an off-diagonal column probability mass m and
the entries along the diagonal probability mass 1 − m. Fi-
Twitter
IMDB
SST
Using Trusted Data to Train Deep Networks
Corruption
Type
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Percent
Trusted
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
Trusted
Only
45.4
35.2
26.1
45.4
35.2
26.1
35.6
36.9
26.2
22.2
36.9
26.2
22.2
28.5
35.9
23.6
16.3
35.9
23.6
16.3
25.3
No
Correction
27.5
27.2
26.5
50.2
49.9
48.7
38.3
26.7
25.8
21.4
49.2
47.8
39.4
35.0
37.1
33.5
25.5
56.2
53.8
43.0
41.5
Forward
26.5
26.2
25.3
50.3
50.1
49.0
37.9
27.9
27.2
23.0
49.2
48.3
39.6
35.9
51.7
49.5
40.6
61.6
59.0
52.5
52.5
Forward
Gold
26.6
25.9
24.6
50.3
49.9
47.3
37.4
27.6
26.1
20.1
49.2
47.5
36.6
34.5
44.1
40.2
26.4
54.8
48.9
36.7
41.9
Distillation Confusion
Matrix
43.4
26.1
33.3
25.0
25.0
22.4
48.8
26.0
42.1
24.6
31.8
22.4
37.4
24.4
35.5
25.4
24.9
23.3
21.0
18.9
41.8
25.8
28.0
22.1
23.5
19.2
29.1
22.5
32.0
41.5
22.2
33.6
16.6
20.0
36.4
23.4
26.1
15.9
20.5
13.3
25.7
24.6
GLC
(Ours)
24.2
23.5
21.7
24.9
23.5
21.7
23.3
25.0
22.3
18.7
25.2
22.0
18.5
22.0
31.0
22.3
15.5
15.8
12.9
12.8
18.4
Table 2: NLP dataset results. Percent trusted is the trusted fraction multiplied by 100. Unless otherwise indicated, all values
are percentages representing the area under the error curve computed at 11 test points. The best mean result is bolded.
nally, a more realistic corruption is hierarchical corruption.
For this corruption, we apply uniform corruption only to
semantically similar classes; for example, “bed” may be
corrupted to “couch” but not “beaver” in CIFAR-100. For
CIFAR-100, examples are deemed semantically similar if
they share the same “superclass” or coarse label specified
by the dataset creators.
Experiments and Analysis of Results. We train the models described in Section 4.2 under uniform, label-flipping,
and hierarchical label corruptions at various fractions of
trusted data in the dataset. To assess the performance of
GLC, we compare it to other loss correction methods and
two baselines: one where we train a network only on trusted
data without any label corrections, and one where the network trains on all data without any label corrections. Additionally, we report results on a variant of GLC that uses
normalized confusion matrices, which we elaborate on in
the discussion. We record errors on the test sets at the corruption strengths {0, 0.1, . . . , 1.0}. Since we compute the
model’s accuracy at numerous corruption strengths, CIFAR
experiments involves training over 500 Wide Residual Networks. In Tables 4 and 5, we report the area under the
error curves across corruption strengths {0, 0.1, . . . , 1.0}
for all baselines and corrections. A sample of error curves
are displayed in Figure 2.
Across all experiments, GLC obtains better area under the
error curve than the Forward and Distillation methods. The
rankings of the other methods and baselines are mixed. On
MNIST, training on the trusted data alone outperforms all
methods save for GLC and Confusion Matrix, but performs
significantly worse on CIFAR-100, even with large trusted
fractions. Interestingly, Forward Gold performs worse than
Forward on several datasets. We did not observe the same
behavior when turning off the corresponding component
of GLC, and believe it may be due to variance introduced
during training by the difference in signal provided by the
Forward method’s C estimate and the clean labels. The
GLC provides a superior C estimate, and thus may be better
able to leverage training on the clean labels. Additional
results on SVHN, are in the supplementary materials.
4.4. Weak Classifier Labels
Our next benchmark for GLC is to use noisy labels obtained
from a weak classifier. This models the scenario of label
noise arising from a classification system weaker than one’s
own, but with access to information about the true labels that
one wishes to transfer to one’s own system. For example,
scraping image labels from surrounding text on web pages
provides a valuable signal, but these labels would train a
sub-par classifier without correcting the label noise.
Using Trusted Data to Train Deep Networks
Percent
Trusted
1
CIFAR-10 5
10
Mean
5
CIFAR-100 10
25
Mean
Trusted
Only
62.9
39.6
31.3
44.6
82.4
67.3
52.2
67.3
No Correction
28.3
27.1
25.9
27.1
71.1
66
56.9
64.7
Forward
28.1
26.6
25.1
26.6
73.9
68.2
56.9
66.3
Forward
Gold
30.9
25.5
22.9
26.4
73.6
66.1
51.4
63.7
Distillation
60.4
28.1
17.8
35.44
88.3
62.5
39.7
63.5
Confusion
Matrix
31.9
27
24.2
27.7
74.1
63.8
50.8
62.9
GLC
(Ours)
26.9
21.9
19.2
22.7
68.7
56.6
40.8
55.4
Table 3: Results when obtaining noisy labels by sampling from the softmax distribution of a weak classifier. Percent trusted
is the trusted fraction multiplied by 100. Unless otherwise indicated, all values are the percent error attained under the
indicated correction. The best average result for each dataset is shown in bold.
Weak Classifier Label Generation. To obtain the labels,
we train 40-layer Wide Residual Networks on CIFAR-10
and CIFAR-100 with clean labels for ten epochs each. Then,
we sample from their softmax distributions with a temperature of 5, and fix the resulting labels. This results in noisy
labels which we use in place of the labels obtained through
the uniform, flip, and hierarchical corruption methods. The
weak classifiers obtain accuracies of 40% on CIFAR-10
and 7% on CIFAR-100. Despite the presence of highly
corrupted labels, we are able to significantly recover performance with the use of a trusted set. Note that unlike the
previous corruption methods, weak classifier labels have
only one corruption strength. Thus, performance is measured in percent error rather than area under the error curve.
Results are displayed in Table 6.
Analysis of Results. Overall, GLC outperforms all other
methods in the weak classifier label experiments. The Distillation method performs better than GLC by a small margin
at the highest trusted fraction, but performs worse at lower
trusted fractions, indicating that GLC enjoys superior data
efficiency. This is highlighted by GLC attaining a 26.94%
error rate on CIFAR-10 with a trusted fraction of 1%, down
from the original error rate of 60%. It should be noted, however, that training with no correction attains 28.32% error
on this experiment, suggesting that the weak classifier labels
have low bias. The improvement conferred by GLC is more
significant at higher trusted fractions.
5. Discussion and Future Directions
Confusion Matrices. An intuitively reasonable alternative to GLC is to estimate C by a confusion matrix. To do
this, one would train a classifier on the untrusted examples,
obtain its confusion matrix on the trusted examples, rownormalize the matrix, and then train a corrected classifier
as in GLC. However, GLC is a far more data-efficient and
lower-variance method of estimating C. In particular, for
K classes, a confusion matrix requires at least K 2 trusted
examples to estimate all entries of C, whereas GLC requires
only K trusted examples.
Another problem with using confusion matrices is that normalized confusion matrices give a biased estimate of C in
the limit, due to using an argmax over class scores rather
than randomly sampling a class. This leads to vastly overestimating the value in the dominant entry of each row, as can
be seen in Figure 1. Correspondingly, we found GLC outperforms confusion matrices by a significant margin across
nearly all experiments, with a smaller gap in performance
on datasets where K, the number of classes, is smaller. Results are displayed in the main tables. We also found that
smoothing the normalized confusion matrices was necessary
to stabilize training on CIFAR-100.
Data Efficiency. We have seen that GLC works for small
trusted fractions, and we further corroborate its data efficiency by turning to the Clothing1M dataset (Xiao et al.,
2015). Clothing1M is a massive dataset with both humanannotated and noisy labels, which we use to compare the
data efficiency of GLC to that of Distillation when very few
trusted labels are present. The Clothing1M dataset consists
of 1 million noisily labeled clothing images obtained by
crawling online marketplaces. 50,000 images have humanannotated examples, from which we take subsamples as our
trusted set.
For both GLC and Distillation, we first fine-tune a pretrained 34-layer ResNet on untrusted training examples for
four epochs, and use this to estimate our corruption matrix. Thereafter, we fine-tune the network for four more
epochs on the combined trusted and untrusted sets using the
respective method. During fine tuning, we freeze the first
seven layers, and train using gradient descent with Nesterov
momentum and a cosine learning rate schedule. For preprocessing, we randomly crop to a resolution of 224 × 224, and
use mirroring. We also upsample the trusted dataset, finding
this to give better performance for both methods.
Percent Accuracy
Using Trusted Data to Train Deep Networks
70
Distillation
GLC (Ours)
60
50
40
30
100
500
1000
Number of Trusted Examples
Figure 3: Data efficiency of our method compared to Distillation on Clothing1M.
As shown in Figure 3, GLC outperforms Distillation by
a large margin, especially at lower numbers of trusted examples. This is because Distillation requires fine-tuning a
classifier on the trusted data alone, which generalizes poorly
with very few examples. By contrast, estimating the C
matrix can be done with very few examples. Correspondingly, we find that our advantage decreases as the number
of trusted examples increases.
With more trusted labels, performance on Clothing1M saturates as evident in Figure 3. We consider the extreme and
train on the entire trusted set for Clothing1M. We fine-tune
a pre-trained 50-layer ResNeXt (Xie et al., 2016) on untrusted training examples to estimate our corruption matrix.
Then, we fine-tune the ResNeXt on all training examples.
During fine-tuning, we use gradient descent with Nesterov
momentum. During the first two epochs, we tune only the
output layer with a learning rate of 10−2 . Thereafter, we
tune the whole network at a learning rate of 10−3 for two
epochs, and for another two epochs at 10−4 . Then we apply
our loss correction. Now, we fine-tune the entire network
at a learning rate of 10−3 for two epochs, continue training
at 10−4 , and early-stop based upon the validation set. In
a previous work, (Xiao et al., 2015) obtain 78.24% in this
setting. However, our method obtains a state-of-the-art accuracy of 80.67%, while with this procedure the Forward
method only obtains 79.03% accuracy.
b Estimation. For some datasets, the classifier
Improving C
p̂(ỹ | x) may be a poor estimate of p(ỹ | x), presenting a
b for GLC. To see the extent
bottleneck in the estimation of C
to which this could impact performance, and whether simple
methods for improving p̂(ỹ | x) could help, we ran several
variants of GLC experiment on CIFAR-100 under the label
flipping corruption at a trusted fraction of 5/100 which we
now describe. For all variants, we averaged the area under
the error curve over five random initializations.
b
1. In the first variant, we replaced GLC estimate of C
with C, the true corruption matrix used for generating
the noisy labels.
2. As demonstrated by (Guo et al., 2017), modern deep
neural network classifiers tend to have overconfident
softmax distributions. We found this to be the case
with our p̂(ỹ | x) estimate, despite the higher entropy
of the noisy labels, and used the temperature scaling
confidence calibration method proposed in the paper
to calibrate p̂(ỹ | x).
3. Suppose we know the base rates of corrupted labels b̃,
where b̃i = p(ỹ = i), and the base rate of true labels
b0 corrupted the
b of the trusted set. If we posit that C
Tb
labels, then we should have b C0 = b̃T . Thus, we
may obtain a superior estimate of the corruption matrix
b = argmin b kbT C
b0 −
by computing a new estimate C
C
b−C
b0 k2 subject to C1
b = 1.
b̃T k + λkC
2
b
We found that using the true corruption matrix as our C
provides a benefit of 0.96 percentage points in area under
the error curve, but neither the confidence calibration nor the
base rate incorporation was able to change the performance
from the original GLC. This indicates that GLC is robust to
the use of uncalibrated networks for estimating C, and that
improving its performance may be difficult without directly
improving the performance of the neural network used to
estimate p̂(y | x).
Better Performance for Worst-Case Corruption. The
uniform corruption that we use in experiments is an example of worst-case corruption in the sense that the mutual
information between ỹ and y is zero when the corruption
strength equals 1.0. We found that training on the trusted
dataset only resulted in superior performance at this corruption setting, especially on Twitter. This indicates that
it may be possible to devise a re-weighting of the loss on
trusted and untrusted examples using information theoretic
b that would improve performance
measures obtained from C
in worst-case regimes.
6. Conclusion
In this work we have shown the impact of having a small
set of trusted examples on classifier label robustness. We
proposed the Gold Loss Correction (GLC), a method for
handling label noise. This method leverages the assumption
that the model has access to a small set of correct labels
to yield accurate estimates of the noise distribution. In
our experiments, GLC surpasses previous label robustness
methods across various natural language processing and
vision domains which we showed by considering several
Using Trusted Data to Train Deep Networks
corruptions and numerous strengths. Consequently, GLC is
a powerful, data-efficient label corruption correction.
References
Biggio, B, Nelson, B, and Laskov, P. Support vector machines under adversarial label noise. ACML, 2011.
Chapelle, Olivier, Schlkopf, Bernhard, and Zien, Alexander.
Semi-Supervised Learning. The MIT Press, 1st edition,
2010.
Frénay, Benoît and Verleysen, Michel. Classification in the
presence of label noise: a survey. IEEE Trans Neural
Netw Learn Syst, 25(5):845–869, May 2014.
Gimpel, Kevin, Schneider, Nathan, O’Connor, Brendan,
Das, Dipanjan, Mills, Daniel, Eisenstein, Jacob, Heilman,
Michael, Yogatama, Dani, Flanigan, Jeffrey, and Smith,
Noah A. Part-of-speech tagging for twitter: Annotation,
features, and experiments. In Proceedings of the 49th
Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers
- Volume 2, HLT ’11, pp. 42–47, Stroudsburg, PA, USA,
2011. Association for Computational Linguistics.
Guo, Chuan, Pleiss, Geoff, Sun, Yu, and Weinberger, Kilian Q. On calibration of modern neural networks. CoRR,
abs/1706.04599, 2017. URL http://arxiv.org/
abs/1706.04599.
Hendrycks, Dan and Gimpel, Kevin. Bridging nonlinearities
and stochastic regularizers with gaussian error linear units.
27 June 2016.
Kingma, Diederik P. and Ba, Jimmy. Adam: A method
for stochastic optimization. CoRR, abs/1412.6980, 2014.
URL http://arxiv.org/abs/1412.6980.
Larsen, J, Nonboe, L, Hintz-Madsen, M, and Hansen, L K.
Design of robust neural network classifiers. In Acoustics,
Speech and Signal Processing, 1998. Proceedings of the
1998 IEEE International Conference on, volume 2, pp.
1205–1208 vol.2, May 1998.
Li, Bo, Wang, Yining, Singh, Aarti, and Vorobeychik, Yevgeniy. Data poisoning attacks on factorization-based collaborative filtering. CoRR, abs/1608.08182, 2016. URL
http://arxiv.org/abs/1608.08182.
Li, Yuncheng, Yang, Jianchao, Song, Yale, Cao, Liangliang,
Luo, Jiebo, and Li, Jia. Learning from noisy labels with
distillation. CoRR, abs/1703.02391, 2017. URL http:
//arxiv.org/abs/1703.02391.
Loshchilov, Ilya and Hutter, Frank. SGDR: stochastic gradient descent with restarts. CoRR, abs/1608.03983, 2016.
URL http://arxiv.org/abs/1608.03983.
Maas, Andrew L., Daly, Raymond E., Pham, Peter T.,
Huang, Dan, Ng, Andrew Y., and Potts, Christopher.
Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association
for Computational Linguistics: Human Language Technologies, pp. 142–150, 2011.
Menon, Aditya Krishna, van Rooyen, Brendan, and Natarajan, Nagarajan. Learning from binary labels with instancedependent corruption. CoRR, abs/1605.00751, 2016.
URL http://arxiv.org/abs/1605.00751.
Mnih, Volodymyr and Hinton, Geoffrey E. Learning to
label aerial images from noisy data. In Proceedings of
the 29th International Conference on Machine Learning
(ICML-12), pp. 567–574, 2012.
Natarajan, Nagarajan, Dhillon, Inderjit S, Ravikumar,
Pradeep K, and Tewari, Ambuj. Learning with noisy
labels. In Burges, C J C, Bottou, L, Welling, M, Ghahramani, Z, and Weinberger, K Q (eds.), Advances in Neural
Information Processing Systems 26, pp. 1196–1204. Curran Associates, Inc., 2013.
Nettleton, David F, Orriols-Puig, Albert, and Fornells, Albert. A study of the effect of different types of noise
on the precision of supervised learning techniques. Artif
Intell Rev, 33(4):275–306, 1 April 2010.
Patrini, Giorgio, Rozza, Alessandro, Menon, Aditya, Nock,
Richard, and Qu, Lizhen. Making deep neural networks robust to label noise: a loss correction approach.
13 September 2016.
Pechenizkiy, M, Tsymbal, A, Puuronen, S, and Pechenizkiy,
O. Class noise and supervised learning in medical domains: The effect of feature extraction. In 19th IEEE Symposium on Computer-Based Medical Systems (CBMS’06),
pp. 708–713, 2006.
Reed, Scott, Lee, Honglak, Anguelov, Dragomir, Szegedy,
Christian, Erhan, Dumitru, and Rabinovich, Andrew.
Training deep neural networks on noisy labels with bootstrapping. 20 December 2014.
Srivastava, Nitish, Hinton, Geoffrey, Krizhevsky, Alex,
Sutskever, Ilya, and Salakhutdinov, Ruslan. Dropout: A
Simple Way to Prevent Neural Networks from Overfitting.
Journal of Machine Learning Research, 15:1929–1958,
2014.
Steinhardt, Jacob, Koh, Pang Wei, and Liang, Percy. Certified defenses for data poisoning attacks. In NIPS, 2017.
Sukhbaatar, Sainbayar, Bruna, Joan, Paluri, Manohar, Bourdev, Lubomir, and Fergus, Rob. Training convolutional
networks with noisy labels. 9 June 2014.
Using Trusted Data to Train Deep Networks
Veit, Andreas, Alldrin, Neil, Chechik, Gal, Krasin, Ivan,
Gupta, Abhinav, and Belongie, Serge J. Learning
from noisy large-scale datasets with minimal supervision.
CoRR, abs/1701.01619, 2017. URL http://arxiv.
org/abs/1701.01619.
Xiao, Tong, Xia, Tian, Yang, Yi, Huang, Chang, and Wang,
Xiaogang. Learning from massive noisy labeled data
for image classification. In 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), pp.
2691–2699, June 2015.
Xie, Saining, Girshick, Ross, Dollár, Piotr, Tu, Zhuowen,
and He, Kaiming.
Aggregated residual transformations for deep neural networks. arXiv preprint
arXiv:1611.05431, 2016.
Zagoruyko, Sergey and Komodakis, Nikos. Wide residual
networks. 23 May 2016.
Zhu, X. Semi-supervised learning literature survey. 2005.
Zhu, Xingquan and Wu, Xindong. Class noise vs. attribute
noise: A quantitative study. Artificial Intelligence Review,
22(3):177–210, 1 November 2004.
Using Trusted Data to Train Deep Networks
CIFAR-100
CIFAR-10
SVHN
MNIST
A. Additional Results and Figures
Corruption
Type
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Hierarchical
Hierarchical
Hierarchical
Mean
Percent
Trusted
5
10
25
5
10
25
0.1
1
5
0.1
1
5
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
Trusted
Only
37.56
12.93
6.62
37.56
12.93
6.62
19.04
80.42
79.66
24.28
80.42
79.66
24.28
61.45
39.6
31.26
17.43
39.6
31.26
17.43
29.43
82.37
67.28
52.21
82.37
67.28
52.21
82.37
67.28
52.21
67.29
No
Forward
Correction Correction
12.89
14.47
12.3
13.85
9.25
11.8
50.14
51.66
51.1
48.84
47.69
50.24
30.56
31.81
25.45
26.2
25.5
24.22
25.52
14.99
50.97
51.04
50.97
43.92
51.01
43.19
38.24
33.93
31.87
9.05
31.92
8.61
32.72
7.71
53.27
38.59
53.18
36.53
52.66
37.63
42.6
23.02
48.82
47.67
48.37
47.15
45.42
43.59
62.11
61.62
61.92
60.96
59.64
57.5
50.94
50.99
51.93
50.49
54.34
46.97
53.72
51.88
Forward
Gold
13.5
12.3
9.2
41.41
36.37
37.14
24.99
26.76
24.92
15.65
50.93
49.46
49.02
36.12
27.85
20.59
27.1
47.78
51.02
49.47
37.3
49.59
48.88
45.98
62.62
62.19
61.41
52.36
52.07
51.13
54.02
Distillation Confusion
Correction Matrix
42.08
21.81
9.21
15.09
5.78
11.01
46.55
11.65
32.37
5.63
28.21
3.77
27.37
11.49
80.93
25.71
80.39
28.23
24.06
2.67
89.11
19.81
86.27
17.81
17.56
2.22
63.05
16.07
29.73
22.42
18.34
22.69
11.62
16.65
29.66
8.13
18.1
8.16
11.77
7.06
19.87
14.18
87.48
53.64
61.2
49.7
39.82
39.61
87.11
28.58
61.85
26.93
39.96
25.14
87.12
45.75
61.71
38.84
39.73
29.7
62.89
37.54
GLC
(Ours)
10.31
6.33
4.67
3.36
2.88
2.57
5.02
24.36
28.13
2.82
19.35
21.7
2.21
16.43
9.02
6.92
6.36
6.62
6.2
6.1
6.87
42.39
33.91
27.34
27.13
25.83
24.69
34.78
30.16
25.41
30.18
Table 4: Vision dataset results. Percent trusted is the trusted fraction multiplied by 100. Unless otherwise indicated, all
values are percentages representing the area under the error curve computed at 11 test points. The best mean result is shown
in bold.
Twitter
IMDB
SST
Using Trusted Data to Train Deep Networks
Corruption
Type
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Uniform
Uniform
Uniform
Flip
Flip
Flip
Mean
Percent
Trusted
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
5
10
25
Trusted
Only
45.36
35.2
26.14
45.36
35.2
26.14
35.57
36.94
26.23
22.2
36.94
26.23
22.2
28.46
35.85
23.64
16.28
35.85
23.64
16.28
25.26
No
Correction
27.48
27.21
26.53
50.19
49.9
48.69
38.33
26.68
25.76
21.44
49.16
47.8
39.42
35.04
37.1
33.45
25.51
56.23
53.83
42.99
41.52
Forward
26.55
26.19
25.33
50.33
50.08
48.99
37.91
27.94
27.19
23.0
49.2
48.26
39.57
35.86
51.68
49.52
40.6
61.59
59.01
52.49
52.48
Forward
Gold
26.55
25.9
24.6
50.33
49.9
47.28
37.43
27.57
26.1
20.11
49.18
47.5
36.64
34.52
44.09
40.21
26.43
54.81
48.91
36.7
41.86
Distillation Confusion
Matrix
43.37
26.11
33.34
25.03
24.96
22.39
48.77
26.01
42.11
24.58
31.84
22.39
37.4
24.42
35.47
25.45
24.89
23.33
21.01
18.93
41.84
25.8
28.03
22.1
23.48
19.2
29.12
22.47
32.01
41.48
22.22
33.62
16.59
20.04
36.44
23.38
26.15
15.93
20.53
13.25
25.66
24.62
GLC
(Ours)
24.23
23.53
21.74
24.89
23.52
21.75
23.27
25.03
22.34
18.71
25.23
21.98
18.54
21.97
30.96
22.25
15.5
15.83
12.94
12.85
18.39
Table 5: NLP dataset results. Percent trusted is the trusted fraction multiplied by 100. Unless otherwise indicated, all values
are percentages representing the area under the error curve computed at 11 test points. The best mean result is bolded.
Percent
Trusted
1
CIFAR-10 5
10
Mean
5
CIFAR-100 10
25
Mean
Trusted
Only
62.89
39.6
31.26
44.58
82.37
67.28
52.21
67.29
No Correction
28.32
27.12
25.9
27.11
71.07
66.01
56.87
64.65
Forward
28.07
26.6
25.13
26.6
73.92
68.23
56.86
66.34
Forward
Gold
30.86
25.51
22.91
26.43
73.57
66.08
51.44
63.7
Distillation
60.42
28.12
17.79
35.44
88.29
62.51
39.69
63.50
Confusion
Matrix
31.92
26.98
24.16
27.69
74.12
63.75
50.75
62.87
GLC
(Ours)
26.94
21.89
19.15
22.66
68.69
56.65
40.82
55.39
Table 6: Results when obtaining noisy labels by sampling from the softmax distribution of a weak classifier. Percent trusted
is the trusted fraction multiplied by 100. Unless otherwise indicated, all values are the percent error attained under the
indicated correction. The best average result for each dataset is shown in bold.
Using Trusted Data to Train Deep Networks
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-100, Uniform, 10% trusted
0.2
0.4
0.6
0.8
Corruption Strength
CIFAR-100, Flip, 10% trusted
0.0
0.0
1.0
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-100, Uniform, 25% trusted
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
0.0
0.0
1.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-100, Flip, 25% trusted
0.4
0.6
0.8
Corruption Strength
1.0
Test Error
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
1.0
1.0
0.6
0.6
0.4
0.2
0.4
0.6
0.8
Corruption Strength
CIFAR-10, Flip, 5% trusted
1.0
0.8
0.8
0.6
0.6
0.4
0.0
0.0
1.0
CIFAR-10, Uniform, 10% trusted
1.0
0.2
0.4
0.6
0.8
Corruption Strength
CIFAR-100, Hierarchical, 25% trusted
0.4
1.0
0.0
0.0
0.2
0.4
0.6
0.8
Corruption Strength
MNIST, Uniform, 0.1% trusted
1.0
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
0.6
0.8
Corruption Strength
0.4
0.0
0.0
1.0
CIFAR-10, Flip, 10% trusted
1.0
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
MNIST, Uniform, 1% trusted
1.0
0.6
0.6
0.6
0.6
0.8
0.6
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-10, Uniform, 25% trusted
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
1.0
0.2
0.0
0.0
0.0
0.0
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-10, Flip, 25% trusted
0.4
0.6
0.8
Corruption Strength
1.0
0.0
0.0
1.0
0.8
0.8
0.6
0.6
0.4
0.2
0.2
0.4
0.0
0.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
MNIST, Uniform, 5% trusted
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
0.6
0.8
Corruption Strength
1.0
0.4
0.0
0.0
0.4
0.6
0.8
Corruption Strength
1.0
MNIST, Flip, 1% trusted
0.4
0.0
0.0
1.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
MNIST, Flip, 5% trusted
0.8
0.2
0.2
0.2
0.2
Test Error
1.0
0.2
Test Error
0.0
0.0
0.2
Test Error
0.2
Test Error
0.8
Test Error
0.8
Test Error
0.8
0.4
MNIST, Flip, 0.1% trusted
0.6
0.8
0.4
1.0
0.8
0.2
0.2
1.0
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
CIFAR-100, Hierarchical, 10% trusted
0.0
0.0
0.8
0.0
0.0
0.4
0.6
0.8
Corruption Strength
0.4
0.8
Test Error
0.2
0.2
0.2
0.2
CIFAR-10, Uniform, 5% trusted
0.2
1.0
Test Error
0.4
Test Error
Test Error
Test Error
0.2
0.6
0.2
Test Error
0.0
0.0
CIFAR-100, Hierarchical, 5% trusted
0.4
0.6
0.0
0.0
Test Error
0.4
0.6
0.2
1.0
0.6
0.8
0.8
0.0
0.0
0.6
0.8
1.0
0.6
0.8
0.8
0.0
0.0
0.8
0.8
1.0
0.2
1.0
1.0
0.2
Test Error
Test Error
1.0
CIFAR-100, Flip, 5% trusted
Test Error
1.0
Test Error
0.6
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
Test Error
Test Error
0.8
CIFAR-100, Uniform, 5% trusted
Test Error
1.0
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
0.0
0.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
Using Trusted Data to Train Deep Networks
1.0
0.0
0.0
1.0
0.4
0.6
0.8
Corruption Strength
SVHN, Uniform, 1% trusted
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.2
1.0
0.6
0.2
0.4
0.6
0.8
Corruption Strength
SVHN, Uniform, 5% trusted
0.4
0.6
0.8
Corruption Strength
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
SVHN, Flip, 1% trusted
1.0
Twitter, Uniform, 1% trusted
0.2
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
IMDB, Uniform, 5% trusted
1.0
0.8
0.6
0.6
0.6
0.4
0.6
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
SVHN, Flip, 5% trusted
1.0
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.4
1.0
0.4
0.2
1.0
0.4
0.6
0.8
Corruption Strength
Twitter, Flip, 1% trusted
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
IMDB, Uniform, 25% trusted
1.0
0.8
0.6
0.6
0.4
1.0
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
SST, Uniform, 0.1% trusted
1.0
0.6
0.6
0.6
1.0
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
Twitter, Uniform, 5% trusted
1.0
Test Error
0.6
Test Error
0.8
0.2
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
Twitter, Flip, 5% trusted
1.0
0.2
0.4
0.6
0.8
Corruption Strength
0.0
0.0
1.0
SST, Uniform, 1% trusted
1.0
0.6
0.6
0.6
0.6
0.0
0.0
1.0
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
Twitter, Uniform, 25% trusted
0.0
0.0
1.0
Test Error
0.8
Test Error
0.8
0.2
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
Twitter, Flip, 25% trusted
0.0
0.0
1.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
SST, Uniform, 5% trusted
0.0
0.0
1.0
0.6
0.6
0.6
0.0
0.0
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
0.0
0.0
Test Error
0.6
Test Error
0.8
Test Error
0.8
0.2
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
0.0
0.0
IMDB, Flip, 25% trusted
0.2
0.4
0.6
0.8
Corruption Strength
1.0
SST, Flip, 0.1% trusted
0.2
0.4
0.6
0.8
Corruption Strength
1.0
SST, Flip, 1% trusted
0.2
0.8
0.4
1.0
0.4
0.8
0.4
0.4
0.6
0.8
Corruption Strength
0.2
0.8
0.4
0.2
0.4
0.8
0.4
IMDB, Flip, 5% trusted
0.2
0.8
0.4
1.0
0.4
0.8
0.4
0.4
0.6
0.8
Corruption Strength
0.4
0.8
0.0
0.0
1.0
0.2
0.2
0.2
0.2
IMDB, Flip, 1% trusted
0.4
0.8
0.0
0.0
1.0
Test Error
0.2
0.2
Test Error
Test Error
0.2
0.4
0.8
0.0
0.0
Test Error
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.8
0.8
0.2
Test Error
0.4
1.0
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
0.2
1.0
0.6
0.0
0.0
1.0
0.4
0.0
0.0
0.6
Test Error
Test Error
0.8
0.6
0.2
Test Error
0.0
0.0
0.8
1.0
Test Error
Test Error
0.4
0.8
0.0
0.0
1.0
0.8
0.6
0.8
0.2
0.2
1.0
Test Error
0.2
IMDB, Uniform, 1% trusted
Test Error
0.4
1.0
Test Error
0.6
Ours
Distillation
Forward Method
Forward Gold
Confusion
All Data
Test Error
Test Error
0.8
SVHN, Flip, 0% trusted
Test Error
SVHN, Uniform, 0% trusted
Test Error
1.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
SST, Flip, 5% trusted
0.4
0.2
0.2
0.4
0.6
0.8
Corruption Strength
1.0
0.0
0.0
0.2
0.4
0.6
0.8
Corruption Strength
1.0
| 9cs.NE
|
arXiv:1205.4273v4 [math.CV] 12 Feb 2013
AN ALGEBRAIC APPROACH TO THE OPENNESS
CONJECTURE OF DEMAILLY AND KOLLÁR
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Abstract. We reduce the Openness Conjecture of Demailly and Kollár on
the singularities of plurisubharmonic functions to a purely algebraic statement.
Contents
1. Introduction
2. Background
3. Plurisubharmonic functions
4. Proof of the main results
References
1
4
13
21
25
1. Introduction
In this paper we study singularities of plurisubharmonic (psh) functions. These
are important in complex analytic geometry, see, for instance, [Lel69, Sko72,
Siu74, Dem87, Kis87, Kis94, DK01]. Specifically, we study the openness conjecture of Demailly and Kollár, and reduce this conjecture to a purely algebraic
statement.
Let ϕ be the germ of a psh function at a point x on a complex manifold. It
is easy to see that the set of real numbers c > 0 such that exp(−cϕ) is locally
integrable at x is an interval. It is nonempty by a result of Skoda [Sko72]. The
openness conjecture, see [DK01, Remark 5.3], asserts that this interval is open.
If we define the complex singularity exponent of ϕ at x by
cx (ϕ) = sup{c > 0 | exp(−2cϕ) is locally integrable at x},
then the conjecture can be stated as follows:
Conjecture A. If cx (ϕ) < ∞, then the function exp(−2cx (ϕ)ϕ) is not locally
integrable at x.
Date: April 9, 2018.
Key words and phrases. Plurisubharmonic function, graded sequence, log canonical threshold,
valuation.
2010 Mathematics Subject Classification. Primary 32U05; Secondary 32U25, 14F18, 12J20,
14B05.
The first author was partially supported by NSF grant DMS-1001740 and the second author
was partially supported by NSF grant DMS-1068190.
1
2
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
In fact, Demailly and Kollár made the following slightly more precise conjecture, which easily implies Conjecture A, see [DK01, Remark 4.4].
Conjecture B. If cx (ϕ) < ∞, then for every open neighborhood U of x on which
ϕ is defined, we have the estimate
Vol{y ∈ U | cx (ϕ)ϕ(y) < log r} & r2
as r → 0.
Here and throughout the paper we write f (r) & r2 as r → 0 if there exists
c > 0 such that f (r) ≥ cr2 for all sufficiently small r.
Demailly and Kollár also proved that Conjecture A implies a stronger openness
statement, namely, that the local integrability of exp(−2ϕ) is an open condition
with respect to the L1loc topology, see [DK01, Conjecture 5.4].
Conjectures A and B are easily verified in dimension one. A proof in the twodimensional case was given in [FJ05b]. In higher dimensions, they are open. In
this paper we reduce Conjecture B to a purely algebraic conjecture:
Conjecture C. Let k be an algebraically closed field of characteristic zero. Let
a• be a graded sequence of ideals in the polynomial ring R = k[x1 , . . . , xm ] such
that a1 is m-primary, where m is a maximal ideal in R. Then there exists a
quasimonomial valuation v on R which computes lct(a• ).
Let us briefly explain the meaning of the terms here; see §2.1 and [JM12] for
more details. The log canonical threshold lct(a) of an ideal a ⊆ R is the algebrogeometric analogue of the complex singularity exponent. A graded sequence a• =
(aj )∞
j=1 of ideals in R is a sequence such that ai · aj ⊆ ai+j for all i, j ≥ 1 (we
always assume that some ai is nonzero). We then define
lct(a• ) := sup j lct(aj ) = lim j lct(aj ),
j→∞
j
where the limit is over those j for which aj is nonzero. Similarly, if v : R → R≥0
is a valuation, then we set
v(a• ) := sup j −1 v(aj ) = lim j −1 v(aj ),
j→∞
j
where again the limit is over the j for which aj is nonzero. One can show that
lct(a• ) = inf
v
A(v)
,
v(a• )
(?)
where the infimum is over quasimonomial valuations v, that is, valuations that
are monomial in suitable coordinates on a suitable blowup of Spec R. Here A(v)
is the log discrepancy of v. Finally, we say that a quasimonomial valuation v
computes lct(a• ) if the infimum in (?) is achieved by v.
Conjecture C holds in dimension two, see [JM12, §9]. In higher dimensions, it
is open. Our main result is
Theorem D. If Conjecture C holds for any dimension m ≤ n and any algebraically closed field k of characteristic zero, then Conjecture B holds in dimension n.
OPENNESS CONJECTURE
3
In fact, we prove a slightly more general result, with the log canonical threshold replaced by more general jumping numbers in the sense of [ELSV04], see
Theorem D’ in §3.1. This result has the following consequence. Let ϕ be a psh
function on a complex manifold U . Recall that the multiplier ideal J (ϕ) of ϕ
is the analytic ideal sheaf on U whose stalk at a point x is given by the set of
holomorphic germs f ∈ Ox such that |f |2 e−2ϕ is locally integrable at x. This is
a coherent ideal sheaf. Now define J + (ϕ) as the (increasing, locally stationary)
limit of J ((1 + ε)ϕ) as ε & 0. We then show that a suitable generalization of
Conjecture C implies that J + (ϕ) = J (ϕ), see Remark 3.10.
We note that one can formulate a version of Conjecture C in a more general
setting, dealing with arbitrary graded sequences on regular excellent connected
schemes over Q. It was shown in [JM12] that this more general conjecture follows
from the special case in Conjecture C above.
One can also formulate a similar conjecture for subadditive sequences. Recall
that a sequence b• = (bj )∞
j=1 of nonzero ideals in R is subadditive if bi · bj ⊇ bi+j
for all i, j ≥ 1. As in the case of graded sequences, we can define lct(b• ) and
v(b• ) when v : R → R≥0 is a valuation, and we can consider whether v computes
lct(b• ). We say that b• has controlled growth if
1
1
v(bj ) ≤ v(b• ) ≤ (v(bj ) + A(v))
j
j
for all j ≥ 1 and all quasimonomial valuations v on R (with the left inequality
being obvious). Subadditive systems usually arise as multiplier ideals and then
are of controlled growth, see Proposition 3.12 and also [JM12, Proposition 2.13].
We show that Conjecture C implies (in fact, it is equivalent to) the following
statement.
Conjecture E. Let b• be a subadditive sequence of ideals in an excellent regular
domain R of equicharacteristic zero. If b• is of controlled growth and there is a
maximal ideal m in R and a positive integer p such that mpj ⊆ bj for all j, then
there exists a quasimonomial valuation v on R such that v computes lct(b• ).
It is this form of the conjecture that we will use in the proof of Theorem D.
Let us now indicate the strategy of this proof. Suppose ϕ is a psh germ at a
point x on a complex manifold with λ := cx (ϕ) < ∞. To ϕ we associate a
sequence b• = (bj )j≥1 of ideals by letting bj be the analytic multiplier ideal of
the psh function jϕ. It follows from [DEL00] that b• is subadditive. Further,
using techniques due to Demailly [Dem92, Dem93], one can show that b• has
controlled growth and that the singularities of bj closely approximate those of ϕ.
The complex singularity exponent cy (ϕ) is a lower semicontinuous function of
the point y, so we can define V as the “log canonical locus of λϕ”, that is, the
germ at x of the analytic set defined by cy (ϕ) ≤ λ. Assume that V is smooth
at x (this is in fact no restriction) and let Ox,V be the localization of the ring of
holomorphic germs at x, at the prime ideal defined by V . Since the latter ring is
excellent [Mat80, Theorem 102], so is Ox,V .
We now have the subadditive sequence b• · Ox,V of ideals in the excellent,
regular local ring Ox,V . By applying Conjecture E, we conclude that there exists
4
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
a quasimonomial valuation of Ox,V computing lct(b• · Ox,V ). This valuation is
monomial in suitable algebraic coordinates on a regular scheme X admitting a
projective birational map to Spec Ox,V . We can “analytify” the latter map and
interpret the quasimonomial valuation as an analytic invariant, the Kiselman
number of ϕ. Using basic properties of psh functions and Kiselman numbers we
then obtain the desired volume estimates in Conjecture B.
As already mentioned, Conjecture C holds in dimension two, so we obtain a
new proof of the openness conjecture in dimension two. In fact, this proof is quite
similar to the one in [FJ05b]. The strategy in loc. cit. is to consider a subspace
V of semivaluations v : Ox → [0, +∞] satisfying v(mx ) = 1, where mx is the
maximal ideal. One can equip V with a natural topology in which it is compact
Hausdorff; it also has the structure of a tree and is studied in detail in [FJ04]
(see also [Jon12]). To a psh germ at x one can associate a lower semicontinuous
function on V whose minimum is equal to cx (ϕ). It turns out that the minimum
must occur for a semivaluation that is either quasimonomial or associated to the
germ of an analytic curve at x. In both cases one can deduce the volume estimate
in Conjecture B using a simplified version of the arguments in §4.2.
In higher dimensions, the analogue of the space V was studied in [BFJ08],
where it was shown that cx (ϕ) can be computed using quasimonomial valuations.
However, when ϕ does not have an isolated singularity at the origin, it seems
difficult to define a suitable lower semicontinuous functional directly on V, having
minimum equal to cx (ϕ). The idea is to instead work at a generic point of the log
canonical locus of λϕ. This does not quite make sense in the analytic category,
and for this reason we pass to algebraic arguments using the subadditive sequence
b• . In the algebraic category, localization arguments work quite well and were
extensively used in [JM12].
The idea of studing psh functions using valuations was systematically developed in [FJ05a, FJ05b, BFJ08] but appears already in the work of Lelong [Lel69]
and Kiselman [Kis87, Kis94]. For some recent work on the singularities of psh
functions, see also [Ras06, Ber06, Lag10, Gue10].
The paper is organized as follows. In §2 we review facts about sequences of
ideals and log canonical thresholds in an algebraic setting and adapt some of
the statements to the setting of complex analytic manifolds. We also prove the
equivalence of Conjectures C and E above. In §3 we discuss plurisubharmonic
functions, Kiselman numbers, multiplier ideal sheaves and the Demailly approximation procedure. Finally, the main results are proved in §4.
Acknowledgment. We thank A. Rashkovskii for spotting a mistake in the
proof of Theorem D’ in an earlier version of the paper. We also thank the referee
for a careful reading and several useful remarks.
2. Background
2.1. Algebraic setting. We start by recalling some basic algebraic facts. For
more details we refer to [JM12] even though much of what follows is standard
material. Let R be an excellent, regular domain of equicharacteristic zero. In the
OPENNESS CONJECTURE
5
cases we will consider, R will be the localization at a prime ideal of the ring of
germs of holomorphic functions at a point in a complex manifold.
2.1.1. Quasimonomial valuations. By a valuation on R we mean a rank 1 valuation v : R \ {0} → R≥0 . A valuation is divisorial if there exists a projective
birational morphism π : X → Spec R, with X regular, a prime divisor E on X
and a positive number α > 0 such that v = α ordE , where ordE denotes the order
of vanishing along E.
More generally, consider a projective birational morphism πP: X → Spec R,
with X regular, a reduced
T simple normal crossing divisor E = i∈ITDi on X, a
subset J ⊆ I such that i∈J Di 6= ∅, an irreducible component Z of i∈J Di and
nonnegative numbers αi ≥ 0, i ∈ J, not all zero. Then there is a unique valuation
v on R such that the following holds: if (uj )j∈J are local coordinates at the generic
bX,ξ as
point ξ of Z such that Dj = {uj = 0} and we write f ∈ R ⊆ OX,ξ ⊆ O
P
β
bX,ξ and, for each β, either cβ = 0 or cβ (ξ) 6= 0, then
f = β cβ u , with cβ ∈ O
X
v(f ) = min{
αi βi | cβ 6= 0}.
i
We call such a valuation quasimonomial 1 and we say that the morphism π : X →
Spec R is adapted to v. In general, π is not unique. On the other hand, we can
always choose it so that αi > 0 for all i and that the αi are rationally independent,
see [JM12, Lemma 3.6]. Finally note that quasimonomial valuations are also
known as Abhyankar valuations, see [ELS03] and [JM12, §3.2].
2.1.2. Log discrepancy. Using the notation above we define the log discrepancy
A(v) of a quasimonomial valuation v by
X
A(v) =
αj (1 + ordDj (K)),
j∈J
where K = KX/ Spec R is the relative canonical divisor. In particular, A(ordDj ) =
1 + ordDj (K). One can show that the log discrepancy of a quasimonomial valuation does not depend on any choices made, see [JM12, §5.1]. Furthermore, one
can extend the definition of log discrepancy to arbitrary valuations on R; in this
case the log discrepancy can be infinite, see [JM12, §5.2].
2.1.3. Log canonical thresholds and jumping numbers. If a ⊆ R is a proper
nonzero ideal, then we define the log canonical threshold of a as
lct(a) = inf
v
A(v)
,
v(a)
(2.1)
where the infimum is over all nonzero valuations on R (it is enough to only
consider quasimonomial or even divisorial valuations). The quantity Arn(a) =
1As opposed to the convention in [JM12], we do not consider the trivial valuation, which is
identically zero on R \ {0}, to be quasimonomial.
6
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
lct(a)−1 is called the Arnold multiplicity of a. More generally, if q ⊆ R is a
nonzero ideal, then we define
lctq (a) = inf
v
A(v) + v(q)
v(a)
and
Arnq (a) = lctq (a)−1 .
(2.2)
Then lctq (a) is a jumping number of a in the sense of [ELSV04], and all jumping
numbers appear in this way. Note that lct(a) = lctR (a). The infimum in (2.1)
(resp. (2.2)) is attained at some divisorial valuation associated to a prime divisor
on some log resolution of a (resp. a · q). We make the convention that if a = 0 or
a = R, then lctq (a) = 0 or ∞, respectively.
2.1.4. Graded sequences. We now recall the definitions of the asymptotic invariants for graded sequences of ideals. For proofs and details we refer to [JM12], see
also [Mus02]. A sequence of ideals a• = (aj )j≥1 is a graded sequence if ai ·aj ⊆ ai+j
for all i, j ≥ 1. For example, if v is a valuation on R and α is a positive real
number, then by putting aj := {f ∈ R | v(f ) ≥ jα}, we obtain a graded sequence
in R. We refer to [Laz, §10.1] for other examples of graded sequences of ideals.
We assume that all graded sequences are nonzero in the sense that some aj is
nonzero.
It follows from the definition that if a• is a graded sequence of ideals in R and
v is a valuation on R, then v(ai+j ) ≤ v(ai ) + v(aj ) for all i, j ≥ 1. By Fekete’s
Lemma, this subadditivity property implies that
v(a• ) := inf
j≥1
v(aj )
v(aj )
= lim
,
j→∞
j
j
where the limit is over those j such that aj is nonzero. Similarly, if q is a nonzero
ideal in R, we have
lctq (a• ) := sup j · lctq (aj ) = lim j · lctq (aj ),
j→∞
j≥1
where the limit is over those j such that aj is nonzero. We also put
Arnq (a• ) := lctq (a• )−1 .
The jumping number lctq (a• ) is positive, but may be infinite. One can show (see
[JM12, Corolloray 6.9]) that as in the case of one ideal, we have
lctq (a• ) = inf
v
A(v) + v(q)
,
v(a• )
(2.3)
where the infimum is over all nonzero valuations of R (it is enough, in fact, to
only consider quasimonomial or even divisorial valuations).
2.1.5. Subadditive sequences. Let us now review the corresponding notions for
the case of subadditive sequences, referring for details to [JM12]. A sequence
b• = (bj )j≥0 of nonzero ideals in R is called subadditive if bi+j ⊆ bi · bj for all
i, j ≥ 0. This implies that v(bi+j ) ≥ v(bi ) + v(bj ) for all valuations v on R and
hence that
v(bj )
v(bj )
v(b• ) := sup
= lim
∈ R≥0 ∪ {+∞}.
j→∞
j
j
j
OPENNESS CONJECTURE
7
A subadditive sequence b• has controlled growth if
v(bj ) A(v)
v(b• ) ≤
+
(2.4)
j
j
for all j ≥ 1 and all quasimonomial valuation v (in fact, it is enough to only
impose this condition for divisorial valuations). In particular, for such b• we
have v(b• ) < ∞ for every quasimonomial valuation v.
For every subadditive system b• and every nonzero ideal q ⊆ R, we define
lctq (b• ) := inf j · lctq (bj ) = lim j · lctq (bj ).
j→∞
j≥1
q
q
We also put Arn (b• ) = lct (b•
)−1 .
For every subadditive sequence we have
lctq (b• ) = inf
v
A(v) + v(q)
,
v(b• )
(2.5)
where the infimum is over all valuations v of R with A(v) < ∞ (see [JM12,
Corollary 6.8]). It is clear from the definition that lctq (b• ) < ∞ unless bj = R
for all j. Moreover, if b• has controlled growth, then lctq (b• ) > 0. Indeed, one
can easily see that
1
1
1
Arnq (bj ) ≤ Arnq (b• ) ≤ Arnq (bj ) +
j
j
j
q
for all j ≥ 1, so that Arn (b• ) < ∞.
Subadditive sequences arise algebraically as asymptotic multiplier ideals. If a•
is a graded sequence of ideals in R and if bj = J (aj• ) is the asymptotic multiplier
ideal of a• of exponent j, then b• is a subadditive sequence of controlled growth
(see [JM12, Proposition 2.13]). Furthermore, we have lctq (a• ) = lctq (b• ) for
every nonzero ideal q, and v(a• ) = v(b• ) for every valuation v with A(v) < ∞
(see [JM12, Proposition 2.14, 6.2]).
2.1.6. Computing jumping numbers of graded sequences. If a• is a graded sequence of ideals in R and q is a nonzero ideal, then we say that a nonzero valuation
v computes lctq (a• ) if v achieves the infimum in (2.3), that is, lctq (a• ) = A(v)+v(q)
v(a• ) .
q
Note that if lct (a• ) = ∞, then v(a• ) = 0 for every v; hence every v computes
lctq (a• ). In what follows we will focus on the case lctq (a• ) < ∞; then every
valuation v that computes lctq (a• ) must satisfy A(v) < ∞.
It was shown in [JM12, Theorem 7.3] that for every graded sequence a• and
every nonzero ideal q, there is a valuation v on R that computes lctq (a• ). One
should contrast this with
Conjecture C’. For every excellent regular domain R of equicharacteristic zero,
every graded sequence of ideals a• , and every nonzero ideal q in R, there is a
quasimonomial valuation v on R that computes lctq (a• ).
One can also consider the following special case of this conjecture.
Conjecture C”. If k is an algebraically closed field of characteristic zero and
R = k[x1 , . . . , xm ], then for every nonzero ideal q and every graded sequence a•
of ideals in R such that a1 ⊇ mp for some maximal ideal m in R and some p ≥ 1,
there is a quasimonomial valuation v on R that computes lctq (a• ).
8
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Note that when the ideal q is equal to R, Conjecture C” specializes to Conjecture C in the Introduction. It is shown2 in [JM12, Theorem 7.6] that Conjecture C’ holds for rings of dimension ≤ n if and only if Conjecture C” holds
for rings of dimension ≤ n. We note that both conjectures are trivially true in
dimension one. They are also true in dimension two. A proof, modeled on ideas
in [FJ05b] is given in [JM12, §9].
2.1.7. Computing jumping numbers of subadditive sequences. We now turn to the
analogous considerations for subadditive sequences. If b• is such a sequence, then
a valuation v with A(v) < ∞ computes lctq (b• ) if v achieves the infimum in (2.5).
The following conjecture extends Conjecture E from the Introduction to the case
of arbitrary jumping numbers.
Conjecture E’. Let q be a nonzero ideal and b• a subadditive sequence of ideals
in an excellent regular domain R of equicharacteristic zero. If b• is of controlled
growth and there is a maximal ideal m in R and a positive integer p such that
mpj ⊆ bj for all j, then there exists a quasimonomial valuation v on R such that
v computes lctq (b• ).
The key requirement in the above conjecture is that the valuation v be quasimonomial. The next proposition shows that if we drop this requirement, we can
find a valuation computing the log canonical threshold. This is the analogue of
the corresponding result for graded sequences that we have mentioned above.
Proposition 2.1. Under the assumptions in Conjecture E’, there is a nonzero
valuation v on R with A(v) < ∞ which computes lctq (b• ).
Remark 2.2. A similar result appears in [Hu12a], see also [Hu12b].
Proof. The argument follows verbatim the proof of [JM12, Thm 7.3], which
treated the case of graded sequences.3 If lctq (b• ) = ∞, then the assertion is trivial: we may take v to be any quasimonomial valuation such that v(m) = 0, since
in this case v(b• ) = 0. Hence, from now on, we may assume that lctq (b• ) < ∞.
By the assumption on b• , if v(m) = 0, then v(b• ) = 0. Therefore we only need
to focus on valuations v with v(m) > 0, and, after normalizing, we may assume
v(b• )
that v(m) = 1. Let us fix with 0 < < Arnq (b• ) and suppose that A(v)+v(q)
> .
pj
For every j ≥ 1 we have m ⊆ bj , hence v(bj ) ≤ jp, and therefore v(b• ) ≤ p.
This implies that A(v) ≤ A(v) + v(q) ≤ M , where M = p/. We thus have
lctq (b• ) = inf
v∈VM
A(v) + v(q)
,
v(b• )
where VM is the set of all valuations v with v(m) = 1 and A(v) ≤ M .
The space of all valuations carries a natural topology, and the subspace VM
is compact by [JM12, Proposition 5.9]. Moreover, A is a lower semicontinuous
2For the comparison with [JM12], note that a ring R has equicharacteristic zero iff Spec R is
a scheme over Q.
3The proof of [JM12, Thm 7.3] involves an extra step, the reduction to the case when all
ideals are m-primary, for some maximal ideal m in R; in our case, we do not have to worry about
this step since this is part of the hypothesis.
OPENNESS CONJECTURE
9
function on VM , while the functions v 7→ v(q) and v 7→ v(b• ) are continuous on
VM by [JM12, Proposition 5.7, Corollary 6.6] (for the last assertion we make use
of the hypothesis that b• has controlled growth). It follows that the function
v 7→ A(v)+v(q)
is lower semicontinuous on VM ; hence it achieves its infimum at
v(b• )
some point v ∈ VM .
2.1.8. Equivalence of conjectures. Our main result is that Conjecture C’ implies
the openness conjecture, see Theorem D’ in §3.1. As a first step, we show that
Conjectures C’, C” and E’ are equivalent.
Proposition 2.3. If one of Conjectures C’,C”, and E’ holds for all rings of
dimension ≤ n, then the other two conjectures hold for such rings.
Proof. As we have already mentioned, [JM12, Theorem 7.6] gives the equivalence
of Conjectures C’ and C”. On the other hand, it is easy to see that if Conjecture E’
holds in dimension ≤ n, then so does Conjecture C”. Indeed, let q and a• be as
in Conjecture C”, and let bj = J (aj• ). In this case b• is a subadditive sequence
of controlled growth, and for every j ≥ 1 we have
mpj ⊆ aj1 ⊆ aj ⊆ bj .
Furthermore, if v is a quasimonomial valuation of R which computes lctq (b• ),
then, since v(a• ) = v(b• ) and lctq (a• ) = lctq (b• ), it follows that v computes
lctq (a• ). Therefore Conjecture C” holds in dimension ≤ n.
We now assume that Conjecture C’ holds in dimension ≤ n, and consider a
nonzero q and a subadditive sequence b• as in Conjecture E’, with dim(R) ≤ n.
We may assume that lctq (b• ) < ∞, since otherwise the assertion to be proved
is trivial (note also that lctq (b• ) > 0 since b• has controlled growth). It follows
from Proposition 2.1 that there is a nonzero valuation w of R with A(w) < ∞
which computes lctq (b• ). In particular, w(b• ) is finite and positive. If we put
aj = {f ∈ R | w(f ) ≥ j}, then a• is a graded sequence of ideals, and by
Conjecture C’ there is a quasimonomial valuation v on R which computes lctq (a• ).
It is enough to show that in this case v also computes lctq (b• ).
It follows easily from the definition of a• that w(a• ) = 1 and if v(a• ) = γ, then
γ = inf{v(f )/w(f ) | f ∈ R, w(f ) > 0}
(see for example [JM12, Lemma 2.4]). We first deduce that lctq (a• ) ≤ A(w) +
w(q) < ∞, hence γ > 0. Furthermore, we have γ −1 v ≥ w. Since v computes
lctq (a• ), we have
γ −1 (A(v) + v(q)) =
A(w) + w(q)
A(v) + v(q)
≤
= A(w) + w(q).
v(a• )
w(a• )
(2.6)
On the other hand, using the fact that γ −1 v ≥ w, we obtain γ −1 v(b• ) ≥ w(b• ),
and therefore (2.6) gives
A(w) + w(q)
A(v) + v(q)
≤
.
v(b• )
w(b• )
(2.7)
By assumption, w computes lctq (b• ), hence we have equality in (2.7), and v also
computes lctq (b• ).
10
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Remark 2.4. By running the argument in the proof of Proposition 2.3 with
q = R, we see that Conjectures C and E in the Introduction are equivalent, in
the sense that one holds for rings of dimension ≤ n if and only if the other one
does.
2.1.9. A converse to the conjectures. As a partial converse to the (equivalent)
conjectures C’, C” and E’ above we show that any quasimonomial valuation
computes some jumping number. This result will not be used in the sequel. In
its formulation and proof we freely use terminology from [JM12].
Proposition 2.5. Let X be an excellent, regular, connected, separated scheme
over Q and let v be a quasimonomial valuation on X. Then there exists a nonzero
ideal q on X and a graded sequence a• on X such that v computes lctq (a• ).
Proof. By [JM12, Thm 7.8] it suffices to find a nonzero ideal q such that the
following statement holds: for every valuation w ∈ ValX such that w ≥ v (in the
sense that w(a) ≥ v(a) for all ideals a on X) we have A(w) + w(q) ≥ A(v) + v(q).
Here A = AX is the log discrepancy with respect to X.
After replacing X by an open neighborhood of the center cX (v) of v on X
we may assume that X = Spec R is affine. Since v is quasimonomial, there
exists a proper birational morphism π : Y → X, with Y regular, and algebraic
local coordinates y1 , . . . , yn at cY (v) with respect to which v is monomial. Let
Ei = {yi = 0}, 1 ≤ i ≤ n, be the associated prime divisors on Y and pick N
large enough so that N ≥ A(ordEi ) for all i. Also write yi = ai /bi with ai , bi ∈ R
nonzero.
We claim that the principal ideal q = (b1 · . . . · bn )N does the job. Indeed,
suppose w ∈ ValX satisfies w ≥ v. In particular, we then have
w(ai ) ≥ v(ai )
and w(bi ) ≥ v(bi )
for all i.
(2.8)
Since v is monomial in coordinates y1 , . . . , yn we have
A(v) =
n
X
v(yi )A(ordEi ).
(2.9)
i=1
By the definition of A(w) we also have
A(w) ≥
n
X
w(yi )A(ordEi ).
(2.10)
i=1
Equations (2.9) and (2.10) and the definition of q now imply
A(w) − A(v) + w(q) − v(q) ≥
n
X
A(ordEi )(w(yi ) − v(yi )) +
i=1
=
n
X
i=1
A(ordEi )(w(ai ) − v(ai )) +
n
X
N (w(bi ) − v(bi ))
i=1
n
X
(N − A(ordEi ))(w(bi ) − v(bi )) ≥ 0,
i=1
where the last inequality follows from (2.8) and the choice of N . This completes
the proof.
OPENNESS CONJECTURE
11
Remark 2.6. It follows from [JM12, Thm 7.8] that with the choice of q above, v
also computes lctq (b• ) for some subadditive sequence b• as well as lctq (a0• ), where
a0• is the graded sequence defined by a0j = {v ≥ j} for j ≥ 1.
2.2. Analytic setting. Let U be a complex manifold. When talking about open
sets in U we always refer to the classical topology unless mentioned otherwise.
By an ideal on U we will always mean a coherent analytic ideal sheaf on U . For
a point x ∈ U , Ox denotes the ring of germs of holomorphic functions at x. Note
that Ox is isomorphic to the ring of convergent power series in n variables over
C, where n = dim(U ); hence Ox is an excellent regular local ring, see [Mat80,
Thm 102].
We denote by mx the maximal ideal in Ox . By a valuation at x we mean a
valuation on Ox in the sense of §2.1.1.
A subadditive sequence of ideals on U is a sequence b• = (bj )∞
j=1 of everywhere
nonzero ideals on U such that bi · bj ⊇ bi+j . If x ∈ U , then we write b• · Ox for
the corresponding subadditive sequence inside Ox . We say that b• has controlled
growth if b• · Ox has controlled growth for all x ∈ U .
If q is an everywhere nonzero ideal on U , b• is a subadditive sequence of ideals
on U and x ∈ U , then we define lctqx (b• ) := lctq·Ox (b• · Ox ) and Arnqx (b• ) :=
lctqx (b• )−1 . Thus we have
Arnqx (b• ) = sup
v
v(b• )
,
A(v) + v(q)
where the supremum is over all (quasimonomial) valuations at x (note that we
simply write v(b• ) and v(q) for v(b• · Ox ) and v(q · Ox ), respectively).
More generally, we shall consider the following situation. Let V be a germ
of a complex submanifold at a point x in a complex manifold. Let Ox,V be the
localization of Ox along the ideal IV . This is an excellent regular local ring with
maximal ideal mx,V . Consider a subadditive system of ideals b• defined near x
and a nonzero ideal q ⊆ Ox . We set
lctqx,V (b• ) := lctq·Ox,V (b• · Ox,V )
and
Arnqx,V (b• ) := lctqx,V (b• )−1
Then
v(b• )
,
A(v)
+ v(q)
v
where the supremum is over all (quasimonomial) valuations of Ox,V .
Note that if V = {x}, then we recover the previous situation.
Arnqx,V (b• ) = sup
2.2.1. Analytification of birational morphisms. Let x, V be as above. Consider
a projective birational morphism π : X → Spec Ox,V of schemes over C, with X
regular. We can then analytify π as follows.4 Since π is projective, there exists a
closed embedding
X ,→ Spec Ox,V ×Spec C PN
(2.11)
C
such that π is the restriction of the projection of the right hand side onto
Spec Ox,V . Thus X is cut out by finitely many homogeneous equations with
4The analytification procedure here is ad hoc and not functorial, but nevertheless related to
the construction of a complex manifold associated to a smooth complex projective variety.
12
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
coefficients in Ox,V . These coefficients can be written as fi /g, where fi ∈ Ox and
g ∈ Ox \ IV · Ox . Let U be a neighborhood of x on which g and the fi are defined.
Set W := {g = 0} ⊆ U . This is a (possibly empty) analytic subset of U that
does not contain V . After shrinking U we may assume that W is either empty
or contains x.
We now define a complex manifold
X an ,→ (U \ W ) × PN (C)
as the analytic subset cut out by the same equations as in (2.11). Then X an is
a complex manifold and the induced projection π an : X an → U \ W is a proper
modification. We shall have more to say about this construction later.
Given a point y ∈ U \ W we define in the same way a projective birational
morphism πy : Xy → Spec Oy . Namely, Xy ⊆ Spec Oy ×Spec C PN
C is defined by
the same homogeneous polynomials as above. After shrinking U and increasing
W (but keeping x ∈ U and W 6⊇ V ) we may further obtain that if E ⊆ πy−1 (V )
is any prime divisor, then the image πy (E) contains V . Further, suppose b is an
ideal on U and that π : X → Spec Ox,V as above is a log resolution of b · Ox,V .
Then, we may assume that the birational morphism πy : Xy → Spec Oy is a log
resolution of b · Oy for all y ∈ U \ W . Further, there is a bijection between the
set of prime divisors E of Xy for which ordE (b · Oy ) > 0 and the set of prime
divisors E of X such that ordE (b · Ox,V ) > 0.
2.2.2. Log canonical locus. The key to the proof of Theorem D is to localize at
the locus where the log canonical threshold is as small as possible. Let x and V
be as above and let b• (resp. q) be a subadditive system of ideals (resp. a nonzero
ideal) defined on some neighborhood U of x. Assume that U is small enough that
V is a submanifold of U .
Lemma 2.7. Assume that b• has controlled growth and that
lctqy (b• ) ≤ λ = lctqx (b• )
for all y ∈ V , where λ ≥ 0. Then lctqx,V (b• ) = λ.
Proof. Let us first prove that lctqx,V (b• ) ≥ λ. For this, fix m ≥ 1 and pick a log
resolution πm : Xm → Spec Ox of the ideal (qbm ) · Ox . After a base change by
Spec Ox,V → Spec Ox , πm induces a log resolution of the ideal (qbm ) · Ox,V . We
have
ordE (bm )
,
Arnqx (bm ) = max
E A(ordE ) + ordE (q)
where the maximum is over the set of prime divisors E ⊆ Xm for which ordE (bm ) >
0. On the other hand Arnqx,V (bm ) is given by the same expression, but where the
maximum is only over the subset of prime divisors E for which πm (E) contains
V . It is then clear that Arnqx,V (bm ) ≤ Arnqx (bm ). Dividing by m and letting
m → ∞ yields Arnqx,V (b• ) ≤ Arnqx (b• ) and hence lctqx,V (b• ) ≥ lctqx (b• ) = λ.
Now we prove the reverse inequality. Pick any m ≥ 1. Consider a log resolution
πm : Xm → Spec Ox,V of the ideal qbm · Ox,V . As in §2.2.1 this gives rise to
an open neighborhood Um of x, an analytic subset Wm ⊆ Um not containing
OPENNESS CONJECTURE
13
V ∩ Um , and for each y ∈ V ∩ (Um \ Wm ), a log resolution πm,y : Xm,y → Spec Oy
of qbm · Oy . Further, there is a bijection between the set of prime divisors E ⊆
Xm,y such that ordE (bm · Oy ) > 0 and the set of prime divisors E ⊂ Xm for
which ordE (bm · Ox,V ) > 0. This implies that Arnqy (bm ) = Arnqx,V (bm ) for any
y ∈ (Um \ Wm ) ∩ V since both quantities are calculated using the same divisors.
Thus we have
Arnqy (bm ) = Arnqx,V (bm ) ≤ m Arnqx,V (b• ),
the inequality being definitional.
Now, for any y ∈ (Um \ Wm ) ∩ V and any j ≥ 1 there exists a quasimonomial
(or even divisorial) valuation vj at y such that
vj (b• )
1
≥ Arnqy (b• ) − .
A(vj ) + vj (q)
j
This gives
vj (b• )
vj (bm ) + A(vj )
1
1
+ ≤
+
A(vj ) + vj (q) j
m(A(vj ) + vj (q)) j
q
Arny (bm )
vj (bm )
1
1
1
1
1
1
+
+ ≤
+
+ ≤ Arnqx,V (b• ) +
+ ,
≤
m(A(vj ) + vj (q)) m j
m
m j
m j
λ−1 ≤ Arnqy (b• ) ≤
where the second inequality follows from the assumption that b• ·Oy has controlled
growth. Letting first j → ∞ and then m → ∞ we obtain Arnqx,V (b• ) ≥ λ−1 , and
hence lctqx,V (b• ) ≤ λ. This completes the proof.
3. Plurisubharmonic functions
Let U be a complex manifold. A function ϕ : U → R ∪ {−∞} is plurisubharmonic (psh) if ϕ 6≡ −∞ on any connected component of U , ϕ is upper semicontinuous, and τ ∗ ϕ is subharmonic or ≡ −∞ for every holomorphic map τ : D → U ,
where D ⊆ C is the unit disc. A germ of a psh function at a point is defined in
the obvious way.
A basic example of a psh function is ϕ = log maxi |fi |, where f1 , . . . , fm are
holomorphic functions on U . If x ∈ U and q ⊆ Ox is an ideal, then we can define
a psh germ log |q| at x by setting
log |q| := log max |fi |
i
(3.1)
where fi ∈ Ox are generators of q. The choice of generators only affects log |q|
up to a bounded additive term. If a and b are two ideals in Ox , then
log |a · b| = log |a| + log |b| + O(1).
For further facts about psh functions, see e.g. [Hor94, Dem].
3.1. Jumping numbers and singularity exponents. Let ϕ be a psh germ at
a point x on a complex manifold and let q ⊆ Ox be a nonzero ideal. Define
cqx (ϕ) = sup{c > 0 | |q|2 exp(−2cϕ) is locally integrable at x}.
This definition does not depend on the choice of generators used to define |q|.
x
We also write cx (ϕ) := cO
x (ϕ). In [DK01], cx (ϕ) is called the complex singularity
14
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
exponent of ϕ at x, whereas cqx (ϕ) is a jumping number in the sense of [ELSV04].
If a ⊆ Ox is a nonzero ideal, then we have
cqx (log |a|) = lctqx (a),
where the right hand side is defined as in §2.2, see [DK01, Proposition 1.7].5 We
have the following generalizations of Conjectures A and B.
Conjecture A’. If cqx (ϕ) < ∞, then the function |q|2 exp(−2cqx (ϕ)ϕ) is not
locally integrable at x.
This conjecture, which is also due to Demailly and Kollár, can be paraphrased
as a semicontinuity statement for multiplier ideals, see Remark 3.10.
Conjecture B’. If cqx (ϕ) < ∞, then for any open neighborhood U of x on which
ϕ and q are defined, we have
Vol{y ∈ U | cx (ϕ)ϕ(y) − log |q| < log r} & r2
(3.2)
as r → 0.
It is clear that Conjecture B’ implies Conjecture A’ and that neither conjecture
depends on the choice of generators of q. The following result is a variation of
Theorem D from the introduction. We shall prove both these theorems in §4.
Theorem D’. If Conjecture C” in §2.1.6 holds for all m ≤ n and all algebraically
closed fields k of characteristic zero, then Conjecture B’ holds on complex manifolds of dimension n.
3.2. Kiselman numbers. We now recall an analytic version of monomial valuations due to Kiselman [Kis87, Kis94]. It is a special case of the generalized
Lelong numbers introduced by Demailly [Dem87]. As our setting differs slightly
from the above references, we give some details for the convenience of the reader.
Let Ω be a complex manifold of dimension n, Z ⊆ Ω a connected submanifold
of codimension m ≥ T
1 and D1 , . . . , Dm distinct, smooth, connected hypersurfaces
in Ω such that Z = i Di and such that the Di meet transversely along Z. Also
suppose we are given positive real numbers αi > 0, 1 ≤ i ≤ m. In this situation
we will associate to any psh function ϕ on U its Kiselman number τZ,D,α (ϕ) ≥ 0.
In preparation for the definition, pick a point z ∈ Z and local analytic coordinates (u1 , . . . , un ) at z such that Di = {ui = 0} for 1 ≤ i ≤ m locally at z. For
t ∈ Rn≤0 with ti 0 let Du (t) ⊆ Ω be the polydisc with radius et and Su (t) its
distinguished boundary, that is
n
n
\
\
Du (t) :=
{|ui | ≤ eti } and Su (t) :=
{|ui | = eti }.
(3.3)
i=1
i=1
We also write Du (s) = Du (s, s, . . . , s) for s ∈ R≤0 .
Let ϕ be a psh germ at z and pick ε > 0 small enough that ϕ is defined in an
open neighborhood of the polydisc Du (log ε). For t ∈ Rn≤log ε set
H(t) := sup ϕ = sup ϕ.
Du (t)
Su (t)
5In loc. cit. only the case q = O is treated but the same proof works in the general case.
x
OPENNESS CONJECTURE
15
Clearly H is increasing in each argument and it is finite-valued since ϕ is upper
semicontinuous. Less obvious is the fact that H is convex; for this see [Kis94,
p.12]. Note that H is continuous on the closed set Rn≤log ε since it is defined and
convex on an open neighborhood of this set.
We now define a new function
h = hϕ,Z,D,z,u,ε : Rn≥0 → R
by setting
H(log ε + sα)
.
s→−∞
s
The limit is well defined by the convexity of H.
h(α) := lim
Lemma 3.1. The function h = hϕ,Z,D,z,u,ε has the following properties:
(i) h is nonnegative, continuous, concave, 1-homogeneous and increasing in
each argument;
(ii) h does not depend on the choice of ε as long as ϕ is defined in an open
neighborhood of D(log ε).
If, further, αi = 0 for i > m, then
(iii) h does not depend on the choice of local coordinates (u1 , . . . , un ) at z, as
long as Di = {ui = 0} for i ≤ m;
(iv) h does not depend on the choice of point z ∈ Z as long as ϕ is defined in
a neighborhood of z.
Proof. To alleviate notation, we shall only write out the relevant part of the
subscripts of h = hϕ,Z,D,z,u,ε .
The fact that h is nonnegative, continuous, concave and increasing follows from
H being continuous, convex and increasing. That h is 1-homogeneous is clear.
This proves (i).
As for (ii), suppose 0 < ε0 < ε. It is clear that hε0 ≥ hε since H is increasing.
To prove the reverse inequality, first suppose that αi > 0 for 1 ≤ i ≤ n and set
δ = mini αi . Then
Du (log ε + sα) ⊆ Du (log ε0 + (s +
ε
1
log 0 )α),
δ
ε
so that H(log ε + sα) ≤ H(log ε0 + (s + 1δ log εε0 )α) for any s ≤ 0. This implies
hε (α) ≥ hε (α). By continuity of hε and hε0 we get hε ≥ hε0 and hence hε = hε0
on Rn≥0 .
Now we turn to (iii) and (iv) so suppose αi = 0 for i > m.
Let (u01 , . . . , u0n ) be another set of local analytic coordinates at z such that
Di = {u0i = 0} for 1 ≤ i ≤ m. We can write ui = u0i gi for 1 ≤ i ≤ m, where
gi ∈ Oz and gi (z) 6= 0. It is easy to see that if ε > 0 is small enough, then there
exists ε0 > 0 such that
Du0 (log ε0 + sα) ⊆ Du (log ε + sα)
for all s ≤ 0. This gives hu0 (α) = hu0 ,ε0 (α) ≥ hu,ε (α) = hu (α) and the reverse
inequality follows by symmetry. Thus (iii) holds.
16
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Finally we prove (iv). Thus pick a point z ∈ Z, and a set of local coordinates
u at z. Pick 0 < ε 1 and z 0 ∈ Du (log ε) ∩ Z. Then u0 := u − u(z 0 ) defines local
coordinates at z 0 and for any α as above and any s ≤ 0 we have
Du0 (log ε + sα) ⊆ Du (log 2ε + sα)
and Du (log ε + sα) ⊆ Du0 (log 2ε + sα).
This implies that hz 0 = hz 0 ,u0 ,ε ≥ hz,u,2ε = hz and, similarly, hz ≥ hz 0 . Thus z →
7
hz is locally constant on Z, which completes the proof since Z is connected.
Now assume αi = 0 for i > m and αi > 0 for 1 ≤ i ≤ m. The number
τZ,D,α (ϕ) := hϕ,Z,D (α)
is called the Kiselman number 6 of ϕ along Z with weight αi along Di . As explained in Lemma 3.1, it does not depend on the choice of coordinates ui defining
the hypersurfaces Di . However, given such coordinates, it follows from the convexity of H that we have the estimate
1
log |ui | + O(1),
(3.4)
ϕ ≤ τZ,D,α (ϕ) max
i≤m αi
near z. From this inequality we easily deduce
Lemma 3.2. Suppose ϕ, ψ are psh functions defined near some Zariski general
point z ∈ Z. Write τ = τZ,D,α . Then:
(i) if ϕ ≤ ψ + O(1) near z, then τ (ϕ) ≥ τ (ψ);
(ii) τ (max{ϕ, ψ}) = min{τ (ϕ), τ (ϕ)}.
Proof. The inequality in (i) follows immediately from the definition. As for (ii),
note that (i) implies τ (max{ϕ, ψ}) ≤ min{τ (ϕ), τ (ϕ)}. The reverse inequality
follows from (3.4).
Remark 3.3. It is also true that τ (ϕ + ψ) = τ (ϕ) + τ (ψ), but we do not need
this fact.
Remark 3.4. Using the same construction, we can define τZ,D,α when Z and
the Di are germs of complex submanifolds at a point in a complex manifold.
Remark 3.5. When αi = 1 for 1 ≤ i ≤ m, the choice of hypersurfaces Di play
no role; in this case the Kiselman number is equal to the Lelong number along
Z [Lel69].
3.3. Kiselman numbers and quasimonomial valuations. Let U be a complex manifold, x ∈ U a point and V the germ at x of a complex submanifold of
U . We allow for the case V = x but assume that V has codimension ≥ 1. As
in §2.2.1 let Ox,V be the localization of Ox at the ideal IV · Ox and let mx,V be
the maximal ideal of Ox,V . Let v be a quasimonomial valuation of Ox,V . We
want to associate to v a Kiselman number on a suitable modification.
Consider a projective birational morphism π : X → Spec Ox,V that is adapted
to v in
Pthe sense of §2.1.1. Thus there exist prime divisors D1 , . . . , Dm on X such
that Di has simple normal crossing singularities, and an irreducible component
6In [Kis87, Kis94], the Kiselman number is called a refined Lelong number whereas Demailly [Dem87] calls it a directional Lelong number.
OPENNESS CONJECTURE
17
T
Z of m
i=1 Di such that v is monomial with weight αi > 0 on Di for 1 ≤ i ≤ m.
The assumption that v ≥ 0 on Ox,V implies that π(Z) ⊂ V . Let ξ be the generic
point of Z and pick functions ui ∈ OX,ξ , 1 ≤ i ≤ m, such that Di = (ui = 0).
Thus the functions ui are regular on a Zariski open subset of Z.
Using the construction and conventions of §2.2.1, after shrinking U a little,
the projective birational morphism π : X → Spec Ox,V gives rise to a complex
manifold X an and a proper modification π an : X an → U \ W , where W ⊆ U is a
complex subvariety not containing V .
Further, after again shrinking U and increasing W if necessary, there exists an
open subset Ω of X an on which the functions ui , 1 ≤ i ≤ m, are holomorphic and
such that the following properties hold: the sets Dian := (ui = 0) are complex
submanifolds of Ω ofTcodimension one, meeting transversely along the connected
submanifold Z an := 1≤i≤m Dian . Further π an (Z an ) ⊇ V \ W .
Let τ = τZ an ,Dan ,α denote the Kiselman number with respect to the data above,
see §3.2.
Definition 3.6. If ϕ is the germ of a psh function at x, then we define
v(ϕ) := τ (ϕ ◦ π an ).
(3.5)
Note that this definition a priori depends on a lot of choices made above.
However, we have:
Proposition 3.7. The definition of v(ϕ) does not depend on any choices made
as long as the birational morphism π : X → Spec Ox,V is adapted to v.
We shall prove this result in §3.4 using multiplier ideals, see Remark 3.13. For
now, we only treat the following special case.
Lemma 3.8. If b ⊆ Ox is a nonzero ideal, then
v(b · Ox,V ) = τ ((log |b|) ◦ π an ).
(3.6)
Proof. Note that both sides of (3.6) depend continuously on the weight α ∈
Rm
>0 . Hence we may assume that the αi are rationally independent. In view of
Lemma 3.2 we may also assume that b is generated by a single element f ∈ Ox .
We must prove that τ (log |f | ◦ π an ) = v(f ).
Consider a Zariski general closed point z ∈ Z and pick functions um+1 , . . . , un ∈
OX,z such that u := (u1 , . . . , un ) define local algebraic coordinates on X at z.
Write u0 := (u1 , . . . , um ) and u00 := (um+1 , . . . , un ). Consider the expansion of
00
0
[
f ◦ π as a formal power series in O
X,ξ ' C[[u]] = C[[u ]][[u ]]:
X
X
f ◦π =
aβγ (u0 )β (u00 )γ =
aβ (u00 )(u0 )β ,
(3.7)
n−m
β∈Zm
≥0 ,γ∈Z≥0
β∈Zm
≥0
where aβ,γ ∈ C and
aβ (u00 ) =
X
γ∈Zn−m
≥0
aβγ (u00 )γ ∈ C[[u00 ]] ⊆ C[[u]].
(3.8)
18
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Since the αi are rationally independent, there exists a unique β̄ minimizing β·α :=
β1 α1 + · · · + βm αm over all β for which aβ 6≡ 0. By definition, we then have
v(f ) = β̄ · α.
Since the point z ∈ Z was generically chosen, it corresponds to a point, also
denoted z, on the complex manifold Z an . We may assume that such that f
is holomorphic near π an (z) ∈ U \ W . Pick 0 < ε 1 such that f ◦ π an is
holomorphic on the open polydisk |ui | < ε, 1 ≤ i ≤ n. The first series in (3.7)
is then the Taylor series of the holomorphic function f ◦ π an at z in the analytic
coordinates u and this series converges locally uniformly in the polydisk kuk < ε.
Further, for every β, the series in (3.8) converges locally uniformly for ku00 k < ε
to a holomorphic function aβ (u00 ) and the second series in (3.7) converges locally
uniformly for kuk < ε.
By assumption, the holomorphic function aβ̄ is not constantly equal to zero.
After moving z (but keeping z ∈ Z an ) a little and translating the coordinates
um+1 , . . . , un accordingly, we may assume that aβ̄ (0) 6= 0. Let us use the notation
of §3.2. For 0 < ε 1 we have
log |f | ◦ π an = log |f ◦ π an | ∼ log |aβ̄ (0)(u0 )β̄ | ∼ sβ̄ · α
on the set Su (log ε + sα), as s → −∞. This implies that
τ (log |f | ◦ π an ) = β̄ · α = v(f )
as was to be shown.
3.4. Multiplier ideal sheaves and Demailly approximation. To a psh function ϕ on a complex manifold U is associated a multiplier ideal sheaf J (ϕ). This
is an ideal sheaf on U whose stalk at a point x is the set of holomorphic germs
f ∈ Ox such that |f |2 e−2ϕ is locally integrable at x. The coherence of J (ϕ)
is a nontrivial result due to Nadel [Nad89, Nad90], which can be proved using
Hörmander’s L2 -estimates, see [DK01, Thm 4.1].
Recall from §3.1 the definition of the jumping number cqx (ϕ) of ϕ at x relative
to an ideal q on U . Given µ > 0 consider the colon ideal aµ = (J (µϕ) : q) on U .
This is an ideal sheaf on U whose stalk at a point x ∈ U is given by
aµ · Ox := {h ∈ Ox | |h|2 |q|2 e−2µϕ is locally integrable at x}.
(3.9)
Since J (µϕ) and q are coherent, so is aµ .
Lemma 3.9. We have cqx (ϕ) < µ iff aµ ·Ox 6= Ox . As a consequence, the function
x 7→ cqx is lower semicontinuous in the analytic Zariski topology on U .
Proof. The first statement is clear. Hence, for µ > 0, the set
Vµ− := {x ∈ U | cqx (ϕ) < µ}
(3.10)
is equal to the support of the coherent sheaf OU /aµ and in particular an analytic
subset of U . It follows that for λ > 0, the set
\
Vλ := {x ∈ U | cqx (ϕ) ≤ λ} =
Vµ−
(3.11)
µ>λ
is also an analytic subset of U . This concludes the proof.
OPENNESS CONJECTURE
19
Remark 3.10. Conjecture A’ in §3.1 is equivalent to a semicontinuity statement
about multiplier ideals. Indeed, define J + (ϕ) as the increasing (locally stationary) limit of J ((1 + ε)ϕ) as ε & 0. Then Conjecture A’ precisely says that
J + (ϕ) = J (ϕ).
If f1 , . . . , fm are holomorphic functions on U , generating an ideal sheaf q and
if log |q| is the corresponding psh function on U defined by (3.1), then we have
J (log |q|) = J (q),
(3.12)
where the right-hand side is defined as in §2.1, see [DK01, Proposition 1.7].
Lemma 3.11. If ϕ ≥ p log |q| + O(1) for some integer p ≥ 1, then J (ϕ) ⊇ qp .
Proof. In view of the assumption and (3.12) we have
J (ϕ) ⊇ J (p log |q|) = J (qp ) ⊇ qp .
Here the last inclusion holds since J (a) ⊇ a for any ideal a.
Now fix a psh function ϕ on U . For j ≥ 1 set
bj := J (jϕ).
It follows from [DEL00] that b• = (bj )∞
j=1 is a subadditive sequence of ideals
on U . The following result (which was known in the case V = x, see [DK01,
Theorem 4.2] and [BFJ08, Theorem 5.5]) allows us to understand the singularities
of ϕ in terms of those of b• .
Proposition 3.12. Let x be any point in U and let V be the germ at x of a proper
complex submanifold. Define Ox,V as in §2.2. Then the following properties hold:
(i) for every nonzero ideal q ⊆ Ox we have cqx (ϕ) = lctqx (b• );
(ii) the subadditive sequence b• · Ox,V has controlled growth;
(iii) for every quasimonomial valuation v on Ox,V we have v(ϕ) = v(b• ·Ox,V ).
Remark 3.13. In (iii) we compute v(ϕ) as a Kiselman number of the pullback
of ϕ under a suitable proper modification, the latter being the analytification of
a blowup of Spec Ox,V , see §3.3. Since the quantity v(b• · Ox,V ) does not depend
on any choices made, we see that v(ϕ) is uniquely defined. Thus we obtain a
proof of Proposition 3.7.
The proof of Proposition 3.12 relies on a fundamental approximation procedure
due to Demailly [Dem92, Dem93]. We refer to [DK01, §4] for details on what
follows.
Let ϕ be a psh function defined in some pseudoconvex domain B ⊆ U containing x. For p ≥ 1 consider the Hilbert space
Z
Hp := {f ∈ O(B) |
|f |2 e−2pϕ < ∞},
B
with the natural inner product. It is a fact that for every y ∈ B, the elements of
Hp generate the stalk at y of the multiplier ideal sheaf bp := J (pϕ). Define
Z
1
ϕp := sup{log |f | |
|f |2 e−2pϕ ≤ 1}.
p
B
20
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
Then ϕp is psh on B. It follows from the Ohsawa-Takegoshi Theorem that
ϕ ≤ ϕp +
C
p
(3.13)
on B, for some constant C not depending on ϕ or p. For any y ∈ B and any
nonzero ideal q ⊆ Oy we also have
(p lctqy (bp ))−1 = cqy (ϕp )−1 ≤ cqy (ϕ)−1
≤ cqy (ϕp )−1 +
1
1
= (p lctqy (bp ))−1 + . (3.14)
p
p
Here the two equalities follow from (3.12) whereas the first inequality results
from (3.13). The second inequality is proved in [DK01, Thm 4.2 (3)] in the case
q = OU and the same proof works in the general case.
Proof of Proposition 3.12. Clearly (i) follows from (3.14) with y = x by letting
p → ∞. It remains to prove (ii) and (iii).
We use the notation of §3.3. Write τ = τZ,D,α . It follows from (3.13) and from
Proposition 3.7 that
1
τ (ϕ ◦ π an ) ≥ τ (ϕp ◦ π an ) = v(bp · Ox,V )
p
(3.15)
for any p ≥ 1.
We will show that if f ∈ Hp , then
v(f ) + A(v) ≥ pτ (ϕ ◦ π an ).
(3.16)
Grant (3.16) for the moment. We then have
1
1
1
v(bp · Ox,V ) ≤ τ (ϕ ◦ π an ) ≤ v(bp · Ox,V ) + A(v).
p
p
p
(3.17)
Letting p tend to infinity we see that τ (ϕ ◦ π an ) = v(b• · Ox,V ), proving (iii). In
particular, v(ϕ) = τ (ϕ ◦ π an ) is well defined independently of any choices made
so we have established Proposition 3.7. Since v was an arbitrary quasimonomial
valuation on Ox,V we also see that b• · Ox,V has controlled growth, proving (ii).
It only remains to prove (3.16). Since both sides of (3.16) depend continuously
on the weight α, we may assume that α1 , . . . , αm are rationally independent. We
now argue as in the proof of Lemma 3.8, recycling the notation from that proof.
Thus we have the expansion (3.7) and we have aβ̄ (0) 6= 0 for the unique β̄ ∈ Zm
≥0
for which β̄ · α = v(f ).
Now fix K 1. Define a sequence of disjoint open subsets (Ωk )k≥0 of Ω by
Ωk :=
m
\
{−(K + k) <
i=1
n
\
log |ui |
< 1 − (K + k)} ∩
{−K < log |ui | < 1 − K}.
αi
i=m+1
(3.18)
For large k we then have the following estimates on Ωk :
log |f ◦ π an | ≥ −kv(f ) + O(1)
and ϕ ◦ π an ≤ −kτ (ϕ ◦ π an ) + O(1).
OPENNESS CONJECTURE
21
Here the second estimate follows from (3.4). Let η be a nonvanishing holomorphic
volume form near x and write (π an )∗ η = Jπ an · ηu near z, where ηu := du1 ∧ · · · ∧
dun . We then have
log |Jπ an | =
m
X
(Ai − 1) log |ui | + O(1)
(3.19)
i=1
near z, where Ai ∈ Z>0 . Further, we have
A(v) =
m
X
α i Ai .
(3.20)
i=1
As k → ∞, we then have
an
log |Jπ | ∼ −k(A(v) −
m
X
αi ) + O(1)
(3.21)
i=1
on Ωk . Moreover, the volume of Ωk can be estimated by
Z
m
X
√
αi + O(1)
log
( −1)n ηu ∧ ηu = −2k
Ωk
(3.22)
i=1
as k → ∞.
Note that if K is large enough, then π an is biholomorphic on Ωk and π an (Ωk )
is contained in B for all k ≥ 0. Thus we get
Z
+∞>
2 −2pϕ
|f | e
∞
X
√
( −1)n η ∧ η ≥
B
k=0
=
∞ Z
X
k=0
&
∞
X
Z
√
|f |2 e−2pϕ ( −1)n η ∧ η
π an (Ωk )
√
an
|f ◦ π an |2 e−2pϕ◦π |Jπ an |2 ( −1)n ηu ∧ ηu
Ωk
exp −2kv(f ) + 2kpτ (ϕ ◦ π an ) − k A(v) −
m
X
!! Z
αi
i=1
k=0
&
∞
X
√
( −1)n ηu ∧ ηu
Ωk
exp (−2k(v(f ) − pτ (ϕ ◦ π an ) + A(v))) ,
k=0
which yields (3.16) (with strict inequality).
4. Proof of the main results
We are now ready to prove Theorem D from the introduction and its variant
Theorem D’ from §3.1. Consider a germ of a psh function ϕ at a point x in a
complex manifold of dimension n and let q ⊆ Ox be a nonzero ideal such that
cqx (ϕ) < ∞. Let U be a small open neighborhood of x such that ϕ and q are
defined on an open neighborhood of U . Also fix a nonvanishing holomorphic
volume form η in a neighborhood
of U and compute all volumes with respect to
√
the positive measure ( −1)n η ∧ η.
22
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
4.1. Analytic reduction. As in §3.4 set
Vµ := {y ∈ U | cqy (ϕ) ≤ µ}
for µ ≥ λ := cqx (ϕ) and
V := Vλ = {y ∈ U | cqy (ϕ) ≤ λ}.
Note that x ∈ V . By the lower semicontinuity of y 7→ cqy (see Lemma 3.9), Vµ is
a proper analytic subset of U for any µ ≥ λ and V is the decreasing intersection
of Vµ for all µ > λ. Using the fact that ϕ and q are defined in a neighborhood of
U we deduce the existence of µ > λ such that V = Vµ .
Lemma 4.1. In order to prove Theorem D’, it suffices to assume that V is smooth
at x and that ϕ ≥ p log |IV | + O(1) near x for some integer p ≥ 0.
Proof. We can replace x by a Zariski general point in V . Indeed, we have cqy (ϕ) =
λ for a Zariski general point y ∈ V , and if the estimate
Vol{y 0 ∈ Uy | λϕ(y 0 ) − log |q| < log r} & r2
(4.1)
holds for every neighborhood Uy of any point y in a dense subset of V , then it
also holds for every neighborhood of x. In particular, we may assume that V is
smooth at x.
Pick generators of IV · Ox . After shrinking U , we may assume these generators
are defined on U and that the associated psh function log |IV |, defined as in (3.1),
is negative on U . For an integer p > 0 define
ϕ̃ := max{ϕ, p log |IV |}.
cqx (ϕ̃)
cqx (ϕ)
We claim that
=
for p 0. This will allow us to replace ϕ by ϕ̃ and
complete the proof. Indeed, we have ϕ ≤ ϕ̃ so if the estimate (3.2) holds with ϕ
replaced by ϕ̃, then it must also hold for ϕ.
To prove the claim, pick µ > λ such that Vµ = V . Consider the colon ideal
aµ = (J (µϕ) : q) on U . This is a coherent ideal sheaf on U whose stalk at y ∈ U
is given by
aµ · Oy := {h ∈ Oy | |h|2 |q|2 e−2µϕ is locally integrable at y}.
The fact that Vµ = V implies that the zero locus of aµ is equal to V . Hence the
Nullstellensatz implies that there exists N ≥ 1 such that IVN ⊆ aµ . Now pick the
integer p > 0 large enough so that
p>
N
.
µ−λ
Pick any λ0 ∈ (λ, µ) such that p > N/(µ − λ0 ). For 0 < r 1 define Borel
subsets Ur , Ũr and Ur0 of U by
Ur := {λ0 ϕ − log |q| < log r}
Ũr := {λ0 ϕ̃ − log |q| < log r}
Ur0 := {µϕ − log |q| − N log |IV | < log r}.
OPENNESS CONJECTURE
23
It follows from the choice of p that Ur ⊆ Ũr ∪ Ur0 . The inclusion IVN ⊆ aµ
guarantees that, after possibly shrinking U , we have
Z ∞
dr
Vol(Ur0 ) 3 < ∞.
(4.2)
r
0
Indeed, if we set F := exp(N log |IV | + log |q| − µϕ), then, after shrinking U ,
Z ∞
Z ∞
Z
dr
2
Vol(Ur0 ) 3 ,
Vol(U ∩ {F > t})t dt = 2
F =2
∞>
r
0
0
U
where the last equality follows from setting t = 1/r.
On the other hand, the fact that λ0 > λ = cqx (ϕ) implies that
Z ∞
dr
Vol(Ur ) 3 = ∞.
r
0
The inclusion Ur ⊆ Ũr ∪ Ur0 then gives
Z ∞
dr
Vol(Ũr ) 3 = ∞,
r
0
so that cqx (ϕ̃) ≤ λ0 . Letting λ0 → λ we get cqx (ϕ̃) ≤ λ. But ϕ̃ ≥ ϕ, so we must
have cqx (ϕ̃) ≥ cqx (ϕ) = λ and hence cqx (ϕ̃) = cqx (ϕ), establishing the claim and
completing the proof of the lemma.
Remark 4.2. The proof of Lemma 4.1 can be viewed as an analytic analogue of
the arguments in [JM12, §7.4].
4.2. End of proof. Let x and V be as above. In particular, V is smooth at x.
Let Ox,V be the localization of Ox at the ideal IV · Ox . Then Ox,V is a regular
local ring with maximal ideal mx,V = IV · Ox,V . Its dimension is equal to the
codimension of V and hence bounded by n. It is also an excellent ring. Indeed,
Ox is isomorphic to the ring of convergent power series in n variables, hence
excellent, see [Mat80, Theorem 102], and excellence is preserved by localization.
Set bj = J (jϕ) for j ≥ 0. Then b• · Ox,V is a subadditive system of ideals
having controlled growth, see Proposition 3.12. By Lemma 4.1 we may assume
ϕ ≥ p log |IV | + O(1); hence Lemma 3.11 implies bj · Ox,V ⊇ mpj
x,V for all j ≥ 1.
From the definition of V = Vλ and from Proposition 3.12 we see that
lctqy (b• ) = cqy (ϕ) = λ
for every y ∈ V . Lemma 2.7 then shows that
lctq·Ox,V (b• · Ox,V ) = λ.
Recall that we assume that Conjecture C’ holds in rings of dimension at most
n. Proposition 2.3 implies that Conjecture E’ also holds in rings of dimension at
most n. We can thus find a quasimonomial valuation v on Ox,V such that
A(v) + v(q · Ox,V )
= λ.
v(b• · Ox,V )
(4.3)
Consider a projective birational morphism π : X → Spec Ox,V such that π
defines a log resolution of q and such that X is adapted to v. Thus v is given by
data Z, D, α as in §2.1.1.
24
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
We analytify π following §2.2.1 and §3.3. Let τ denote the Kiselman number
with respect to the data Z an , Dan , α, see §3.2. We know from Proposition 3.12 (iii)
and Remark 3.13 that
τ (ϕ ◦ π an ) = v(b• · Ox,V ).
(4.4)
λτ (ϕ ◦ π an ) = A(v) + v(q).
(4.5)
Thus (4.3) yields
We use the notation from §3.3. Pick a Zariski general point z ∈ Z an . Then
log |q| ◦ π an =
m
X
ci log |ui | + O(1)
i=1
near z, where ci = ordDi (q) ≥ 0; see the end of §2.2.1. We also have
m
X
log |Jπ | =
(Ai − 1) log |ui | + O(1),
an
i=1
where Ai ∈ Z>0 , see (3.19). Finally, recall from (3.4) that
1
log |ui | + O(1).
1≤i≤m αi
ϕ ◦ π an ≤ τ (ϕ ◦ π an ) max
Fix K 1 and define disjoint open subsets Ωk , k ≥ 0, of Ω as in (3.18). As
k → ∞, we then have the following estimates on Ωk :
ϕ ◦ π an ≤ −kτ (ϕ ◦ π an ) + O(1),
log |q| ◦ π
an
≥ −k
m
X
ci αi + O(1) = −kv(q) + O(1).
i=1
Using (4.5) these estimates imply that
λϕ − log |q| ≤ −k(λτ (ϕ ◦ π an ) − v(q)) + O(1) = −kA(v) + O(1)
(4.6)
on the open set π an (Ωk ) ⊂ U . For 1 ≤ i ≤ m set
Ωk,i := {−(K + k) <
log |ui |
< 1 − (K + k)}.
αi
Then we can estimate the volume of π an (Ωk ) as follows:
Z
Z
√
√
Vol π an (Ωk ) = ( −1)n
η ∧ η = ( −1)n
|Jπ an |2 ηu ∧ η u
π an (Ωk )
&
m
Y
√
i=1
Z
−1
Ωk,i
|ui |2Ai −2 dui ∧ dūi &
Ωk
m
Y
exp(−k(2Ai αi )) = exp(−2kA(v)).
i=1
This estimate together with (4.6) concludes the proof of Theorem D’. By choosing
q = OU throughout all the arguments (see also Remark 2.4), we also obtain a
proof of Theorem D.
OPENNESS CONJECTURE
25
References
[Ber06]
B. Berndtsson. Subharmonicity properties of the Bergman kernel and some other
functions associated to pseudoconvex domains. Ann. Inst. Fourier 56 (2006), 1633–
1662. 4
[BFJ08] S. Boucksom, C. Favre and M. Jonsson. Valuations and plurisubharmonic singularities. Publ. Res. Inst. Math. Sci. 44 (2008), 449–494. 4, 19
[Dem]
J.-P. Demailly. Complex analytic and algebraic geometry. Book available at
www-fourier.ujf-grenoble.fr/∼demailly. 13
[Dem87] J.-P. Demailly. Nombres de Lelong généralisés, théorèmes d’intégralité et d’analyticité.
Acta Math. 157 (1987), 153–169. 1, 14, 16
[Dem92] J.-P. Demailly. Regularization of closed positive currents and intersection theory. J.
Alg. Geom. 1 (1992), 361–409. 3, 19
[Dem93] J.-P. Demailly. A numerical criterion for very ample line bundles. J. Differential
Geom. 37 (1993), 323–374. 3, 19
[DEL00] J.-P. Demailly, L. Ein and R. Lazarsfeld. A subadditivity property of multiplier ideals.
Michigan Math. J. 48 (2000), 137–156. 3, 19
[DK01]
J.-P. Demailly and J. Kollár. Semicontinuity of complex singularity exponents and
Kähler-Einstein metrics on Fano orbifolds. Ann. Sci. École Norm. Sup. (4) 34 (2001),
525–556. 1, 2, 13, 14, 18, 19, 20
[ELS03] L. Ein, R. Lazarsfeld, and K. E. Smith, Uniform approximation of Abhyankar valuations in smooth function fields, Amer. J. Math. 125 (2003), 409–440. 5
[ELSV04] L. Ein, R. Lazarsfeld, K. E. Smith and D. Varolin. Jumping coefficients of multiplier
ideals. Duke Math. J. 123 (2004), 469–506. 3, 6, 14
[FJ04]
C. Favre and M. Jonsson. The valuative tree. Lecture Notes in Mathematics 1853.
Springer-Verlag, Berlin, 2004. 4
[FJ05a] C. Favre and M. Jonsson. Valuative analysis of planar plurisubharmonic functions.
Invent. Math. 162 (2005), 271–311. 4
[FJ05b] C. Favre and M. Jonsson. Valuations and multiplier ideals. J. Amer. Math. Soc. 18
(2005), 655–684. 2, 4, 8
[Gue10] H. Guenancia. Toric plurisubharmonic functions and analytic adjoint ideal sheaves.
arXiv:1011.3162. 4
[Hor94]
L. Hörmander. Notions of convexity. Progress in Mathematics, 127. Birkhäuser,
Boston, MA, 1994. 13
[Hu12a] Z. Hu. Valuative multiplier ideals. Preprint. 8
[Hu12b] Z. Hu. Valuations and log canonical thresholds. Preprint. 8
[Jon12]
M. Jonsson. Dynamics on Berkovich spaces in low dimensions. arXiv:1201.1944.
To appear in Berkovich spaces and applications, Séminaires et Congrès, Société
Mathématique de France. 4
[JM12]
M. Jonsson and M. Mustaţă. Valuations and asymptotic invariants for sequences of
ideals. Ann. Inst. Fourier 62 (2012), 2145–2209. 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 23
[Kis87]
C.-O. Kiselman. Un nombre de Lelong raffiné. In Séminaire d’analyse complexe et
géométrie 1985-1987. Faculté des sciences de Tunis et Faculté des Sciences et Techniques de Monastir, 1987, pp. 61–70. 1, 4, 14, 16
[Kis94]
C. O. Kiselman. Attenuating the singularities of plurisubharmonic functions. Ann.
Polon. Math. 60 (1994), 173–197. 1, 4, 14, 15, 16
[Lag10]
A. Lagerberg. A new generalization of the Lelong number. arXiv:1001.3562. 4
[Laz]
R. Lazarsfeld. Positivity in algebraic geometry II. Ergebnisse der Mathematik und
ihrer Grenzgebiete, 3. Folge, Vol. 49, Springer-Verlag, Berlin, 2004. 6
[Lel69]
P. Lelong. Plurisubharmonic functions and positive differential forms. Gordon and
Breach, New York, and Dunod, Paris, 1969. 1, 4, 16
[Mat80] H. Matsumura. Commutative algebra. Mathematics Lecture Note Series, 56. Benjamin/Cummings. Reading, Mass., 1980. 3, 11, 23
26
MATTIAS JONSSON AND MIRCEA MUSTAŢĂ
[Mus02]
[Nad89]
[Nad90]
[Ras06]
[Siu74]
[Sko72]
M. Mustaţă. On multiplicities of graded sequences of ideals. J. Algebra 256 (2002),
229–249. 6
A. M. Nadel. Multiplier ideal sheaves and existence of Kähler-Einstein metrics of
positive scalar curvature. Proc. Nat. Acad. Sci. USA 86 (1989), 7299–7300. 18
A. M. Nadel. Multiplier ideal sheaves and Kähler-Einstein metrics of positive scalar
curvature. Ann. of Math. 132 (1990), 549–596. 18
A. Rashkovskii. Relative types and extremal problems for plurisubharmonic functions.
Int. Math. Res. Not. 2006, Art. ID 76283. 4
Y.-T. Siu. Analyticity of sets associated to Lelong numbers and the extension of
positive closed currents. Inventiones. Math. 27 (1974), 53–156. 1
H. Skoda. Sous-ensembles analytiques d’ordre fini ou infini dans Cn . Bull. Soc. Math.
France 100 (1972), 353–408. 1
Dept of Mathematics, University of Michigan, Ann Arbor, MI 48109-1043, USA
E-mail address: [email protected], [email protected]
| 0math.AC
|
Presentations for cusped arithmetic hyperbolic lattices
Alice Mark, Julien Paupert
arXiv:1709.06691v2 [math.GR] 31 Oct 2017
April 6, 2018
Abstract
We present a general method to compute a presentation for any cusped hyperbolic lattice Γ, applying a
classical result of Macbeath to a suitable Γ-invariant horoball cover of the corresponding symmetric space.
As applications we compute presentations for the Picard modular groups PU(2, 1, Od ) for d = 1, 3, 7 and the
quaternionic lattice PU(2, 1, H) with entries in the Hurwitz integer ring H.
1
Introduction
Discrete subgroups and lattices in semisimple Lie groups form a rich and well-studied class of finitely generated groups acting on non-positively curved metric spaces. The case of real rank one, where the associated
symmetric space is negatively curved, is of special interest. There are essentially two main families of constructions of such lattices, arithmetic on one hand and geometric on the other. Arithmetic lattices are roughly
speaking obtained by taking matrices with entries lying in the integer ring of some number field; the general
definition is more complicated and we will not give it here, as the arithmetic lattices that we consider in this paper are of this simplest type. By Margulis’ celebrated superrigidity and arithmeticity theorems, all (irreducible)
lattices in G are of this arithmetic type when G is a semisimple Lie group of real rank at least 2.
The other family involves geometric constructions such as polyhedra, reflections or other types of involutions
or other finite-order isometries. A prototype of this type of construction is given by Coxeter groups in the
constant curvature geometries E n , S n and H n , which are generated by reflections across hyperplanes. These
groups are classical and were classified by Coxeter in the spaces E n and S n , whereas their hyperbolic counterparts
(studied by Vinberg and others) are still not completely understood. However by construction these groups
come equipped with data including a presentation (as an abtract Coxeter group) and a fundamental domain for
their action on the symmetric space.
Arithmetic lattices are given by a global description and their global structure is in some sense well understood by work of Siegel, Borel, Tits, Prasad and others. However concrete information such as a presentation
and a fundamental domain are not readily accessible from the arithmetic construction. One can obtain geometric information such as volume by Prasad’s celebrated volume formula ([P]) but computing the constants
appearing in this formula usually involves some non-trivial work (see for example [Be] and [Sto]).
Very few presentations of arithmetic lattices, and of lattices in general, are known. Presentations can
provide useful geometric and algebraic information about groups, such as explicit index of torsion-free subgroups
(effective Selberg lemma, as used for example in [Sto]), cohomology of the group Γ or quotient space X/Γ, see
for instance [Y] (for the Picard modular groups with d = 1, 3) and of course representations of Γ, for instance
if one is interested in deformations of Γ into a larger Lie group.
Presentations for SL(n, Z) with n > 3 were given by Steinberg ([Ste], following Magnus); the case of SL(2, Z)
is classical and possibly dates to Gauss; see also Siegel [Si]. In rank one, Swan
√ gave in [Sw] presentations for
the Bianchi groups PGL(2, Od ) (where Od denotes the ring of integers of Q[i d] for d a positive square-free
integer), following Bianchi’s original construction in [Bi]. These act as isometries of (real) hyperbolic 3-space,
as they are lattices in PGL(2, C) ' Isom+ (H3R ).
Presentations for the related Picard modular groups PU(2, 1, Od ) were found only recently in the simplest
cases of d = 3 ([FP]) and d = 1 ([FFP]). One of the reasons for this is that the associated symmetric space,
complex hyperbolic 2-space H2C , is more complicated and in particular has non-constant (pinched) negative
curvature. A particular feature of such spaces, the absence of totally geodesic real hypersurfaces, makes constructions of fundamental domains difficult as there are no obvious walls to use to bound such domains. The
1
presentations obtained for d = 1, 3 were in fact obtained by constructing fundamental domains and using the
Poincaré polyhedron theorem. This approach seems to become too complicated when considering more complicated groups, such as Picard modular groups with higher values of d, and no further such constructions
have appeared. Using a similar strategy, Zhao gave in [Zh] generating sets for the Picard modular groups with
d = 1, 2, 3, 7, 11 but he does not go as far as obtaining a presentation, finding a set whose translates covers the
space but without control over intersections and cycles. (We will in fact use a covering argument closely related
to the one he uses to cover a fundamental prism on the ideal boundary by isometric spheres, see Lemma 13).
In this paper we present a method for obtaining presentations for cusped hyperbolic lattices, i.e. noncocompact lattices in semisimple Lie groups of real rank one, based on a classical result of Macbeath (Theorem 1
below) which gives a presentation for a group Γ acting by homeomorphisms on a topological space X, given
an open subset V whose Γ-translates cover X. We apply this by finding a suitable horoball V based at a cusp
point of Γ whose Γ-translates cover X, then analyzing the triple intersections and associated cyles to obtain a
presentation for Γ. The main tools for this analysis come from the additional arithmetic structure that we get by
assuming that Γ is in fact an integral lattice in the sense that it is contained in GL(n + 1, OE ) for some number
field E (or finitely generated division algebra over Q). The crucial such tool that we use is the notion of level
between two E-rational boundary points in ∂∞ X (see Definition 1) which gives a notion of distance between such
points using only algebraic data. More importantly for us, levels measure the relative sizes of horospheres based
at the correpsonding boundary points, which allows us to control whether or not such horospheres intersect at
a given height (see Lemma 4).
As applications of the method we compute presentations for the Picard modular groups PU(2, 1, Od ) with
d = 3, 1, 7, given in (12), (13) and (14). We expect to treat the cases d = 2, 11 with the same method but
they are computationally more intensive and will be treated elsewhere. We also compute a presentation (15)
for the quaternion hyperbolic lattice which we call the Hurwitz modular group PU(2, 1, H), where H is the ring
] ⊂ H. This is a lattice in PU(2, 1, H) (sometimes denoted PSp(2, 1)),
of Hurwitz integers H = Z[i, j, k, 1+i+j+k
2
acting on the 8-dimensional symmetric space H2H . As far as we know this is the first presentation ever found
for a higher-dimensional quaternion hyperbolic space. (In dimension 1, H1H ' H4R and such groups have been
studied e.g. in [DVV], see also [A], [W] and [Ph]).
The paper is organized as follows. In Section 2 we discuss generalities about horoball coverings of hyperbolic
spaces, levels for cusp points of integral lattices, and outline how we apply Macbeath’s theorem in this context.
In Section 3 we discuss horosphere intersections in more detail, in particular the quantitative relation between
levels and heights of horospheres for integral lattices. In Sections 4 and 5 we apply this method to compute
presentations for the Picard and Hurwitz modular groups respectively.
We would like to thank Daniel Allcock for suggesting this method and for many helpful comments, and
Matthew Stover and Daniel Allcock for pointing out a mistake in an earlier version of the paper.
2
2.1
Horoball coverings and lattice presentations
Adapted horoball coverings and covering complex
Let X be a negatively curved symmetric space, i.e. a hyperbolic space HnK , with K = R, C, H or O (and n > 2
if K = R, n = 2 if K = O). We refer the reader to [CG] for general properties of these spaces and their isometry
groups. In particular isometries of such spaces are roughly classified into the following 3 types: elliptic (having
a fixed point in X), parabolic (having no fixed point in X and exactly one on ∂∞ X) or loxodromic (having no
fixed point in X and exactly two on ∂∞ X).
Let Γ be a lattice in Isom(X); the well-known Godement compactness criterion states that Γ contains
parabolic isometries if and only if it is non-cocompact, which we now assume. A cusp point of Γ is a point of
∂∞ X fixed by a parabolic element of Γ; a cusp group of Γ is a subgroup of the form StabΓ (p) where p ∈ ∂∞ X
is a cusp point of Γ.
Assume that we are given a Γ-invariant covering of X by (open) horoballs (see Definition 3), i.e. a collection
B of horoballs such that:
γB ∈ B for all γ ∈ Γ and B ∈ B
S
B∈B B = X
2
We will moreover assume that each horoball B ∈ B is based at a cusp point of Γ, and that each cusp point
of Γ is the basepoint of a unique horoball in B (giving a bijection between cusp points of Γ and horoballs in
B); we will call such a covering B a Γ-adapted horoball covering. Since the lattice Γ has only finitely many cusp
points modulo the action of Γ, it follows that such a horoball covering is a finite union of Γ-orbits of horoballs.
Given a Γ-adapted horoball covering B, the covering complex C(B) associated to B is the simplicial 2-complex
with vertex set B, with an edge connecting each pair of vertices B1 and B2 such that B1 ∩ B2 6= ∅, and a triangle
for each triple of vertices B1 , B2 , B3 such that B1 ∩ B2 ∩ B3 6= ∅. This the 2-skeleton of a simplicial complex
sometimes called the nerve of the covering. By the above remark the quotient of the covering complex by the
action of Γ is a finite simplicial 2-complex.
We will use the following classical result of Macbeath ([M]):
Theorem 1 ([M]) Let Γ be a group acting by homeomorphisms on a topological space X. Let V be an open
subset of X whose Γ-translates cover X.
(1) If X is connected then the set E(V ) = {γ ∈ Γ | V ∩ γV 6= ∅} generates Γ.
(2) If moreover X is simply-connected and V path-connected, then Γ admits a presentation with generating set
E(V ) and relations γ · γ 0 = γγ 0 for all γ, γ 0 ∈ E(V ) such that V ∩ γV ∩ γγ 0 V 6= ∅.
Now if as above Γ is a lattice in Isom(X) and B is a Γ-adapted horoball covering of X, we may as remarked above write B as a finite union of Γ-orbits of horoballs B1 , ..., Bk (say, minimally). One can then apply
Macbeath’s theorem with V = B1 ∪ ... ∪ Bk , after possibly enlarging each horoball Bi in order for this union
to be (path-)connected. For simplicity of exposition, we henceforth asume that Γ has a single cusp, so that the
Γ-adapted horoball covering consists of a single Γ-orbit of horoballs (this is the case in all examples considered
in this paper). In that case the process of obtaining a presentation fom the covering complex is closely related
to a complex of groups structure on the quotient of the covering complex, the only difference being that we
need to take into account non-trivial edge and face stabilizers.
2.2
Levels and proximal cusp complex
Recall that if X is a hyperbolic space HnK (with K = R, C, H or O) then X admits the following projective
model which we briefly recall.
Consider Kn,1 , the vector space Kn+1 endowed with a Hermitian form h· , ·i of signature (n, 1). (When K = H
or O we will use the
convention that scalars acton vectors on the right, whereas matrices act on vectors on the
left.) Let V − = Z ∈ Kn,1 |hZ, Zi < 0 , V 0 = Z ∈ Kn,1 |hZ, Zi = 0 and let π : Kn+1 − {0} −→ KPn denote
projectivization. One then defines HnK to be π(V − ) ⊂ KPn , endowed with the distance d (Bergman metric)
given by, for Z, W ∈ V − :
cosh2
d(π(Z), π(W ))
2
=
|hZ, W i|2
.
hZ, ZihW, W i
(1)
Note that the right-hand side is independent of the choice of lifts Z, W . Then Isom0 (X) = PU(n, 1, K), the
(projectivization of) the matrix group preserving the Hermitian form (see [CG]). Note that PU(n, 1, K) is
usually denoted PO(n, 1) when K = R, PU(n, 1) when K = C and PSp(n, 1) when K = H. The boundary at
infinity ∂∞ X is then identified with π(V 0 ) ⊂ KPn . We would like to measure distances between points of ∂∞ X
using the Hermitian form as in (1); one way to do this is to use integral lifts of vectors with rational coordinates
as follows.
We now assume that Γ is an integral lattice in the sense that it is contained in U(H, OE ) for some number
field E (or finite-degree division algebra over Q when K = H or O) with ring of integers OE , and Hermitian
n+1
form H = h· , ·i defined over E. We say that an integral vector P0 = (p1 , ..., pn+1 ) ∈ OE
is primitive if it has
n+1
−1
no integral submultiple in the following sense: if P0 λ ∈ OE for some λ ∈ OE then λ is a unit in OE .
If p is an E-rational point in KP n , i.e. the projective image of a vector P = (p1 , ..., pn+1 ) ∈ E n+1 , a primitive
n+1
integral lift of p is any lift P0 of p to OE
which is a primitive integral vector.
Lemma 1 If OE is a principal ideal domain then primitive integral lifts are unique up to multiplication by a
unit.
3
Lemma 2 (a) Any column-vector of a matrix A ∈ U(H, OE ) is a primitive integral vector. (b) If moreover E is
imaginary quadratic with OE a principal ideal domain and one of the standard basis vectors Bi is H-isotropic,
then for any H-isotropic primitive integral vector V and A ∈ U(H, OE ), AV is a primitive integral vector.
Proof. (a) Let A ∈ U(H, OE ) and V a column-vector of A. Then V is integral; assuming that it is not
primitive, there would exist a non-unit λ ∈ OE such that V λ−1 is also integral. But then the matrix A0 obtained
from A by replacing the column-vector V by V λ−1 would also be in GL(n + 1, OE ), with det A0 = det Aλ−1 , a
contradiction since the latter is not an integer, as det A is a unit and λ is not.
(b) Let V be an H-isotropic primitive integral vector and A ∈ U(H, OE ). If OE is a principal ideal domain
then PU(H, OE ) has a single cusp (see [Zi]), therefore there exists M ∈ U(H, OE ) mapping Bi to AV λ for some
λ ∈ E. Then as in (a) λ must be a unit, hence AV is a column-vector of M λ−1 ∈ U(H, OE ) and we conclude
by (a).
Definition 1 Given two E-rational points p, q ∈ ∂∞ X, the level between p and q, denoted lev(p, q), is |hP0 , Q0 i|2
for any two primitive integral lifts P0 , Q0 of p, q respectively. When we are given a preferred E-rational point
∞ ∈ ∂∞ X, the depth of an E-rational point p ∈ ∂∞ X is the level between p and ∞.
By Lemma 1 this is well-defined when OE is a principal ideal domain. The proximal cusp complex of level
n, denoted Cn (Γ), is the complex whose vertices are cusp points of Γ, with an edge connecting 2 vertices p, q
whenever lev(p, q) 6 n, and a triangle for each triple of distinct edges.
Levels give a convenient way to distinguish orbits of edges and triangles in the covering complex, by the
following observation which follows from Lemmas 1 and 2:
Lemma 3 If OE is a principal ideal domain, for any two E-rational points p, q ∈ ∂∞ X and γ ∈ U(H, OE ),
lev(γp, γq) = lev(p, q).
More importantly, levels allow us to find the optimal height u of a horosphere Hu = ∂Bu based at a preferred
E-rational point ∞ ∈ ∂∞ X such that the orbit ΓBu covers X. This relies on the following result, which follows
from Corollary 4.6 of [KP]:
Lemma 4 There exists a decreasing function u : N −→ R such that, for any E-rational point p ∈ ∂∞ X with
depth n, and any (integral) Ap ∈ Γ satisfying Ap (∞) = p, the set Hu ∩ Ap (Hu ) is empty if and only if u > u(n).
In fact we will see in Corollary 1 below that the function u is given by u(n) =
Definition 2 The covering depth of Γ is the unique n ∈ N such that
the maximal height such that ΓBucov covers X.
2.3
√2
n+1
√2 .
n
< ucov 6
√2 ,
n
where ucov denotes
Reduction modulo the vertex stabilizer Γ∞
We choose a preferred cusp point ∞ ∈ ∂∞ X of Γ (in general we will take ∞ = π([1, 0, ..., 0]T ) in the Siegel
model, see section 3), and consider the cusp stabilizer Γ∞ = StabΓ (∞). Since Γ is a lattice, it is well known
that Γ∞ acts cocompactly on all horospheres based at ∞. Let n denote the covering depth of Γ and ucov the
corresponding covering height, so that the Γ-translates of the horoball Bu cover X, and let D∞ ⊂ Hu be a
compact fundamental domain for the action of Γ∞ on Hu ' ∂∞ X \ {∞}. In practice we will choose D∞ to be
an affinely convex polytope in Heisenberg coordinates (see section 3).
Assume that we are given a finite presentation Γ∞ = hS∞ |R∞ i. Then we may reduce the procedure in
Macbeath’s theorem to finitely many additional generators and relations as follows. Let {p1 , ..., pk } denote the
E-rational points with depth at most n in D∞ , and assume for simplicity that they are ordered in such a way
that the first r of them form a system of representatives under the action of Γ∞ . Assume moreover that we
have found for each i = 1, ..., r an element Ai ∈ Γ such that Ai (∞) = pi (this is possible in principle since Γ is
assumed to have a single cusp).
Generators: The group Γ is generated by {S∞ , A1 , ..., Ar }. This follows easily from part (1) of Macbeath’s
theorem and Lemma 4, as any E-rational point of ∂∞ X with depth at most n is in the Γ∞ -orbit of one of
4
p1 , ..., pr . Note that with the notation from the Theorem, we are using the open set V = Bu = B to cover X,
1
2
1
2
and E(B) = {γ ∈ Γ | B ∩ γB 6= ∅} = Γ∞ {A1 , ..., Ar }Γ∞ = {γ∞
Ai γ ∞
| γ∞
, γ∞
∈ Γ∞ , i = 1, ..., r}. Indeed, by
Lemma 4, B ∩ γB 6= ∅ if and only if γ∞ is either ∞ or an E-rational point of depth at most n, which is a
Γ∞ -translate of one of p1 , ..., pr .
Relations: We now rephrase part (2) of Macbeath’s theorem in this context. Let γ, γ 0 ∈ E(B) satisfy
B ∩ γB ∩ γγ 0 B 6= ∅, and first assume that γ∞ =
6 ∞ and γγ 0 ∞ =
6 ∞. After conjugating by an element of Γ∞ we
0
1
2
0
3
4
1
4
may asume that γ = Ai , γ = γ∞ Aj γ∞ , γγ = γ∞
Ak γ ∞
for some i, j, k ∈ {1, ..., r} and γ∞
, ..., γ∞
∈ Γ∞ . The
0
0
1
2
3
4
corresponding relation γγ = γ · γ is then: Ai γ∞ Aj γ∞ = γ∞ Ak γ∞ . Taking the image of ∞ under both sides of
1
3
this relation gives: Ai (γ∞
pj ) = γ∞
pk . In practice this is how we will detect the relations, finding which points
of depth at most n are sent to points of depth at most n by the generators Ai . One then recovers the relation
1
3
1
3
as follows. For each triple (i, j, k) for which there exist γ∞
, γ∞
∈ Γ∞ such that Ai (γ∞
p j ) = γ∞
pk , we obtain a
−1 3 −1
1
relation Ri,j,k by identifying the element Ak (γ∞ ) Ai γ∞ Aj ∈ Γ∞ as a word in the generators S∞ .
Now assume that one of γ∞, γγ 0 ∞ is ∞ but not both (as the relations in Γ∞ have already been considered).
The corresponding relation can be obtained as above, using the point p∞ = ∞ with correpsonding group element
A∞ = Id.
Summarizing the above discussion gives:
Lemma 5 With the above notation, Γ admits the presentation Γ = hS∞ , A1 , ..., Ar |R∞ , Ri,j,k i.
2.4
The method in practice
We now give an outline of the method we use to apply Macbeath’s theorem:
(1) Find an explicit (affine) fundamental domain D∞ ⊂ ∂∞ X ' Hu for the action of Γ∞ = StabΓ (∞), and a
presentation Γ∞ = hS∞ |R∞ i.
(2) Find the covering depth n of Γ. Consider the corresponding covering complex C(ΓBucov ) ' Cn (Γ).
(3) Find all E-rational points {p1 , ..., pk } in D∞ with depth at most n (and denote p∞ = ∞).
(4) For each of the r Γ∞ -orbit of points pi , find an explicit Ai ∈ Γ such that Ai (∞) = pi .
1
3
1
3
(5) For each triple (i, j, k) for which there exist γ∞
, γ∞
∈ Γ∞ such that Ai (γ∞
pj ) = γ∞
pk , we obtain a
−1 3 −1
1
relation Ri,j,k by identifying the element Ak (γ∞ ) Ai γ∞ Aj ∈ Γ∞ as a word in the generators S∞ .
Then by Macbeath’s theorem and Lemma 5, Γ = hS∞ , A1 , ..., Ar |R∞ , Ri,j,k i.
In order to avoid tedious repetition of similar arguments or straightforward computations, we will only give
one detailed proof for each step for the Picard modular groups, but we will give detailed arguments for the
quaternionic lattice when they are substantially different. We will usually choose the most difficult case, or the
most instructive if the various cases are of similar difficulty. Steps (1) and (5) are routine and we just state the
results (except step (1) for the quaternionic lattice, which we cover in detail in Lemmas 15 and 16). We give a
detailed argument and proof for step (2) for the Picard modular group Γ(7) in Lemma 13, and for step (3) for
the same group and depth 2 in Lemma 14.
There seems to be no general strategy for step (4); we find all relevant matrices in this paper by combining
two tricks, which luckily cover all the cases we need. The first trick is to use stabilizers of vertical complex line
in the Heisenberg group: it is easy to find such a matrix when it stabilizes the vertical axis, then we carry over
to other vertical lines by conjugating by a horizontal translation. The second trick is to hit all relevant integral
points by the group elements that we already know, and see if we land in the Γ∞ -orbit of the point we are
trying to reach.
2.5
A toy example: Γ = PSL(2, Z)
In order to illustrate the method, we now go through its steps for Γ = PSL(2, Z) exactly as we will for the
more complicated Picard and Hurwitz modular groups. The results are either well-known or elementary and
we state them without proof.
5
Presentation and fundamental domain for the cusp stablilizer Γ∞ : The cusp stabilizer Γ∞ has
presentation hT i; a fundamental domain for its action on ∂H2R \ {∞} ' R is D∞ = [0, 1]. Concretely we use the
following generator for Γ∞ :
1 1
T =
0 1
Covering depth and Q-rational points in D∞ : The covering depth of PSL(2, Z) is 1. The Q-rational
points of depth 1 in D∞ are 0 and 1, both in the same Γ∞ -orbit. An integral lift of 0 is p0 = [0, 1]T ; we denote
p1 = T p0 .
Generators: The following element A0 ∈ Γ maps the point ∞ = [1, 0]T to p0 :
0 −1
A0 = I0 =
1 0
Relations: We list in Table 1 the relations obtained for PSL(2, Z) by applying generators to points of depth
at most 1 as described in part (5) of section 2.4. The second relation is obtained by following the corresponding
cycle of points, which gives I0 T I0 T I0 ∈ Γ∞ . The latter element is computed to be T −1 , giving the relation
(I0 T )3 = Id.
Image of vertex
Cycle of points
I0
I0 p0 = ∞
I0 p1 = T −1 p0
Relation
I0
I02 = Id
∞ −→ p0 −→ ∞
(T I0 )
I
−1
−1
(T I0 )
0
∞ −→
p0 −−−−−→ p1 −−−−−→ ∞
(I0 T )3 = Id
Table 1: Action of generators on vertices for PSL(2, Z)
3
Horosphere intersections
Our main reference for this section is [KP]. We will use the Siegel model of hyperbolic space HnK (with
K = R, C, H), which is the projective model (as described in Section 2.2) associated to the Hermitian form on
Kn+1 given by hZ, W i = W ∗ JZ with:
0
0
1
J = 0 In−1 0
1
0
0
Then hyperbolic space HnK can be parametrized by Kn−1 × Im K × R+ as follows, denoting as before by π the
projectivization map: HnK = {π(ψ(ζ, v, u) | ζ ∈ Kn−1 , v ∈ Im K, u ∈ R+ )}, where:
(−|ζ|2 − u + v)/2
ζ
ψ(ζ, v, u) =
1
(2)
With this parametrization the boundary at infinity ∂∞ HnK corresponds to the one-point compactification:
π(ψ(ζ, v, 0) | ζ ∈ Kn−1 , v ∈ Im K ∪ {∞}
where ∞ = π((1, 0, ..., 0)T ). The coordinates (ζ, v, u) ∈ Kn−1 ×Im K×R+ are called the horospherical coordinates
of the point π(ψ(ζ, v, u) ∈ HnK .
Definition 3 For a fixed u0 ∈ R+ , the level set Hu0 = {π(ψ(ζ, v, u0 ) | ζ ∈ Kn−1 , v ∈ Im K} is called the
horosphere at height u based at ∞, and Bu0 = {π(ψ(ζ, v, u) | ζ ∈ Kn−1 , v ∈ Im K, u > u0 } is called the horoball
at height u based at ∞.
6
The punctured boundary ∂∞ HnK \ {∞} is then naturally identified to the generalized Heisenberg group
Heis(K, n), defined as the set Kn−1 × Im K equipped with the group law:
(ζ1 , v1 )(ζ2 , v2 ) = (ζ1 + ζ2 , v1 + v2 + 2Im (ζ1 · ζ2 ))
where · denotes the usual Euclidean dot-product on Kn−1 . This is the classical 3-dimensional Heisenberg group
when K = C and n = 2. The identification of ∂∞ HnK \ {∞} with Heis(K, n) is given by the simply-transitive
action of Heis(K, n) on ∂∞ HnK \ {∞}, where the element (ζ1 , v1 ) ∈ Heis(K, n) acts on the vector ψ(ζ2 , v2 , 0) by
left-multiplication by the following Heisenberg translation matrix in U(n, 1, K):
1 −ζ1∗ (−|ζ1 |2 + v1 )/2
ζ1
T(ζ1 ,v1 ) = 0 In−1
(3)
0
0
1
Given an element U ∈ U(n, K), the Heisenberg rotation by
1 0
RU = 0 U
0 0
U is given by the following matrix:
0
0
1
(4)
There is an additional class of isometries fixing ∞ when K = H, coming from the action of diagonal matrices
which is non-trivial in the non-commutative case. Recall that our convention is that matrices act on vectors
on the left, and scalars act on vectors on the right. Then, for any unit quaternion q ∈ H, the diagonal matrix
Cq = q Id acts by the isometry of hyperbolic space given by conjugating horospherical coordinates (the result
of multiplying the vector form (2) by q on the left, then normalizing by q −1 on the right):
Cq : (ζ, v, u) 7−→ (qζq −1 , qvq −1 , u)
(5)
HnH
For this reason, when K = H the relevant projectivization of U(n, 1, H) acting on
is PU(n, 1, H) =
U(n, 1, H)/{±Id} rather than U(n, 1, H)/U(1).
Heisenberg translations and rotations, as well as conjugation by unit quaternions, preserve the following
distance function on Heis(K, n), called the Cygan metric, defined for (ζ1 , v1 ), (ζ2 , v2 ) ∈ Heis(K, n) by:
dC ((ζ1 , v1 ), (ζ2 , v2 ))
=
|ζ1 − ζ2 |4 + v1 − v2 − 2Im (ζ1 · ζ2 )
2 1/4
(6)
1/2
= |2hψ(ζ1 , v1 , 0), ψ(ζ2 , v2 , 0)i|
(7)
This in fact the restriction to ∂∞ HnK \{∞} of an incomplete distance function on HnK \{∞} called the extended
Cygan metric (see [KP]), defined for (ζ1 , v1 , u1 ), (ζ2 , v2 , u2 ) ∈ Kn−1 × Im K × R>0 ' HnK \ {∞} by:
dXC ((ζ1 , v1 , u1 ), (ζ2 , v2 , u2 ))
=
|ζ1 − ζ2 |2 + |u1 − u2 |
2
+ v1 − v2 − 2Im (ζ1 · ζ2 )
1/2
= |2hψ(ζ1 , v1 , u1 ), ψ(ζ2 , v2 , u2 )i|
2 1/4
(8)
(9)
We define Cygan spheres, Cygan balls, extended Cygan spheres and extended Cygan balls in the usual way
relative to these distance functions.
When we apply Macbeath’s theorem we argue that the images under Γ of the horoball Bu based at ∞ at
a certain height u > 0 cover X, or equivalently cover the horosphere Hu = ∂Bu . The following result, which
follows from Proposition 4.3 of [KP], allows us to control the traces on Hu of these images only in terms of
Cygan spheres depending only on arithmetic data.
Lemma 6 p
Let A = (ai,j ) ∈ U(n, 1, K) such that A(∞) 6= ∞, S the extended Cygan sphere with center A−1 (∞)
and radius 2/|an,1 |, and Hu the horosphere based at ∞ at height u > 0. Then Hu ∩ A(Hu ) = Hu ∩ S.
Corollary 1 Let E be a number field such that OE is a principal ideal domain, p ∈ ∂∞ X an E-rational point
with depth n > 1 and Ap ∈ U(H, OE ) satisfying Ap (∞) = p. Then Hu ∩ A−1
p (Hu ) = Hu ∩ S, where S is the
4 1/4
−1
Cygan sphere centered at p with radius n
. In particular: Hu ∩ Ap (Hu ) = ∅ ⇐⇒ u > u(n) = √2n .
7
Proof. Since Ap (∞) = p and e1 = (1, 0, ..., 0)T is a lift of ∞, the first column vector of Ap is a lift P0 of p,
and since Ap ∈ U(H, OE ) it is an integral lift. In fact by Lemma 2 it is a primitive lift, therefore the depth of
p is |hP0 , e1 i|2 = |an,1 |2 , denoting as above Ap = (ai,j ), and the result follows from Lemma 6. The second part
of the statement follows by using this radius in the formula for the extended Cygan metric, Equation 8.
We will also use the following observation, which is Lemma 1 of [FFP], in our covering arguments:
Lemma 7 Extended Cygan balls are affinely convex in horospherical coordinates.
Finally, when considering the action of a discrete subgroup Γ∞ of Isom(∂∞ X) (relative to the Cygan metric)
it is convenient to consider its vertical and horizontal components defined as follows (see [FP] for the case
K = C and n = 2). The homomorphism Π : Heis(K, n) → Kn−1 given by projection to the first factor in the
decomposition Heis(K, n) ' Kn−1 o Im K induces a short exact sequence:
Π∗
1 −→ Isom(Im K) −→ Isom(Heis(K, n)) −−→ Isom(Kn−1 ) −→ 1,
(10)
where the isometries of Im K and Kn−1 are relative to the Euclidean metric. Denoting Γv∞ = Γ∞ ∩ Isom(Im K)
and Γh∞ = Π∗ (Γ∞ ) this gives the short exact sequence:
Π∗
1 −→ Γv∞ −→ Γ∞ −−→ Γh∞ −→ 1.
4
(11)
Picard modular groups
In this section we use the method described in Section 2.4 to compute the following
presentations for the
√
Picard modular groups Γ(d) = PU(2, 1, Od ) with d = 1, 3, 7. We will denote τ = 1+i2 d when d ≡ 3 (mod 4), so
that Od = Z[τ ].
*
PU(2, 1, O3 ) =
T1 , Tτ , R, I
PU(2, 1, O1 ) =
D
+
[[T1 , Tτ ], T1 ] , [[T1 , Tτ ], Tτ ] , [[T1 , Tτ ], R] ,
6
−1
−1
−1
R Tτ R = T1 , R T1 R = T1 Tτ , R ,
.
I 2 , [R, I], ([T1 , Tτ ]I)6 , T1 Tτ−1 [T1−1 , I]Tτ T1−1 = I[T1 , Tτ ]
T2 , Tτ , Tv , R, I
(12)
[Tτ , T2 ] = Tv4 , [Tv , T2 ], [Tv , Tτ ], [Tv , R], R4 , I 2 , [R, I],
RT2 R−1 = Tτ2 T2−1 Tv4 , RTτ R−1 = Tτ T2−1 Tv2 ,
[I, T2 ]2 , (ITv )3 = R, [I, Tτ ] = Tτ IR2 , (Tv IR−1 Tv2 I)2 ,
ITv−1 Tτ IRT2−1 Tv−1 = T2 Tτ−1 ITτ R2 Tv I,E
(13)
(ITv−1 Tτ IRT2−1 Tv−1 )2 = R−1 T2−1 Tτ Tv−3
PU(2, 1, O7 ) =
D
T1 , Tτ , Tv , R, I0 , I1
[Tτ , T1 ] = Tv , [Tv , T1 ], [Tv , Tτ ], [Tv , R], R2 , (RTτ )2 ,
(RT1 )2 = Tv , I02 , I12 , [R, I0 ], [R, I1 I0 T1−1 Tτ ]2 ,
[R, I1 I0 T1−1 Tτ ] = Tv I0 I1 Tτ T1−1 I1 I0 Tτ T1−2 Tv ,
[R, I1 I0 T1−1 Tτ ] = Tv T1−1 I0 T1 I0 Tτ−1 I1 RI0 Tv−1 ,
[I0 , Tv−1 Tτ T1 ] = T1 I0 I1 I0 T1−1 ,
−1
R[R, I1 I0 T1−1 Tτ ] = T1 I0 Tv T1−2 I0 TE
1 Tv R,
(14)
I1 = T12 Tτ RT12 I0 T1−1 I0 T1 I0
The action of Γ∞ (d) = StabΓ(d) (∞) on ∂H2C is well understood for all d, see [FP] for d = 3, [FFP] for d = 1
and Section 5.3 of [PW] for all other values (using unpublished notes of Falbel-Francsics-Parker). We will refer
to these papers for presentations and fundamental domains for Γ∞ (d) which we state in Lemmas 8, 10 and 12.
8
Figure 1: Covering the prism D∞ (3) by Cygan balls of depth 1
4.1
The Eisenstein-Picard modular group Γ(3) = PU(2, 1, O3 )
Presentation and fundamental domain for the cusp stabilizer Γ∞ (3):
Lemma 8
1. The cusp stabilizer Γ∞ (3) admits the following presentation:
Γ∞ (3) =
T1 , Tτ , R
[[T1 , Tτ ], T1 ] , [[T1 , Tτ ], Tτ ] , [[T1 , Tτ ], R] ,
R−1 Tτ R = T1 , R−1 T1 R = T1 Tτ−1 , R6
.
2. Let D∞ (3) ⊂√∂H2C be the
convex
√ hull of the points with horospherical coordinates (0, 0), (1, 0),
√ affine
τ +1
,
0),
(0,
2
,
2
( τ +1
3),
(1,
2
3),
(
3). Then D∞ (3) is a fundamental domain for Γ∞ (3) acting on
3
3
∂H2C \ {∞}.
Concretely, we use the following generators for Γ∞ (3) (recall that τ =
1 −1 τ 2
1 −τ̄
1 Tτ = T(τ,√3) = 0 1
T1 = T(1,√3) = 0 1
0 0
1
0 0
√
1+i 3
)
2
τ2
1
τ R= 0
1
0
0
τ
0
0
0
1
√
Covering depth and Q[i 3]-rational points in D∞ (3):
We denote B ((z, t), r) the open extended Cygan ball centered at p = (z, t) ∈ ∂∞ H2C with radius r (see
Equation 8 for the definition of the extended Cygan metric). Recall that u(n) = √2n is the height at which balls
of depth n appear, in the sense of Corollary 1.
Lemma 9 Let u = u(5) + ε = 0.895 and Hu the horosphere of height u based at ∞. Then the prism D∞ (3)
√ ×
{u} is covered by the intersections with Hu of the following extended Cygan balls of depth 1: B (0, 0), 2 ,
√ √
√ √
B (0, 2 3), 2 and B (1, 3), 2 .
We omit the proof, which is similar to the proof of Lemma 13 but much simpler; see Figure 4.1.
9
Note that in order to cover D∞ (3) × {u} we only need balls of depth 1, in particular none of depths 3 or
4 even though they are present at the height u = u(5) + ε which we consider. It is however necessary to pass
to this height, as we have observed experimentally
that at height u = u(4) + ε balls of depth at most 3 do not
√
cover D∞ (3) × {u} (there are no Q[i 3]-rational points of depth 2).
Corollary 2 The covering depth of Γ(3) is at most 4.
√
By inspection, we see that the Q[i 3]-rational points in D∞ (3) with depth at most 4 are, in horospherical
coordinates:
√
√
• Depth 1: (0, 0), (0, 2 3) and (1, 3), all in the same Γ∞ (3)-orbit;
√
√
√
√
• Depth 3: (0, 23 3), (1, 53 3) in one Γ∞ (3)-orbit, and (0, 43 3), (1, 13 3) in the other;
√
√
• Depth 4: (0, 3), (1, 0) and (1, 2 3), all in the same Γ∞ (3)-orbit.
Integral lifts of representatives of Γ∞ (3)-orbits of these points are:
√
−1
−2
0
i 3
0 p3,2 = √
0 p4 = 0
p0 = 0 p3,1 = √
1
i 3
2
i 3
Generators: The following elements
above (for α = 0; 3, 1; 3, 2; 4) :
0
A0 = I0 = 0
1
Aα ∈ Γ(3) map the point ∞ = [1, 0, 0]T to the corresponding pα as
0
−1
0
−1 0
1
0
1
0 A3,1 = A3 = √
0
i 3 0
√
−i 3
A4 = 0
−2
A3,2 = A−1
3
0
−1
0
√
i 3
0
2
−2
0
√
i 3
Relations: We list in Table 2 the relations obtained for Γ(3) by applying generators to points of depth at
most 4 as described in part (5) of section 2.4. By successively eliminating A4 then A3 we obtain presentation (12).
The details are straightforward and left to the reader.
4.2
The Gauss-Picard modular group Γ(1) = PU(2, 1, O1 )
Presentation and fundamental domain for the cusp stabilizer Γ∞ (1):
Lemma 10
1. The cusp stabilizer Γ∞ (1) admits the following presentation:
Γ∞ (1) =
T2 , Tτ , Tv , R
[Tτ , T2 ] = Tv4 , [Tv , T2 ], [Tv , Tτ ], [Tv , R], R4 ,
RT2 R−1 = Tτ2 T2−1 Tv4 , RTτ R−1 = Tτ T2−1 Tv2
.
2. Let D∞ (1) ⊂ ∂H2C be the affine convex hull of the points with horospherical coordinates (0, 0), (1, 0), (τ, 0),
(0, 2), (1, 2), (τ, 2). Then D∞ (1) is a fundamental domain for Γ∞ (1) acting on ∂H2C \ {∞}.
Concretely, we use the following generators for Γ∞ (1) (denoting τ = 1 + i):
1 −2 −2
1 −τ̄
2 Tτ = T(τ,0) = 0 1
T2 = T(2,0) = 0 1
0 0
1
0 0
Tv = T(0,2)
1
= 0
0
0
1
0
i
0
1
1
R= 0
0
10
0
i
0
0
0
1
−1
τ
1
Image of vertex
Cycle of points
I0
A3
Rp3,2 = p3,2
√
I0 (0, 2 3) = Tv−1 p3,2
√
I0 (1, 3) = Tv−1 Tτ p0
A3 p 0 = p 4
A3 p4 = Tv p0
A3 p3,1 = p3,2
A4 p0 = p3,2
A4 p3,1 = Tv p0
I02 = Id
A24 = Id
I0
R
[R, I0 ] = Id
A−1
3
R
→ p3,1 −−−→ ∞
∞ −−→ p3,1 −
Rp3,1 = p3,1
I0 p4 = Tv−1 p3,1
A4
→ p0 −→ ∞
∞ −→ p0 −
Rp0 = p0
I0 p3,2 = Tv−1 p4
A4
Relation
∞ −−→ p4 −−→ ∞
A4 p4 = ∞
I0 p3,1 =
I0
∞ −→ p0 −→ ∞
I0 p0 = ∞
Tv−1 p0
I0
A−1
3
[R, A3 ] = Id
A
R
3
∞ −−−→ p3,2 −
→ p3,2 −−→
∞
√ Tv I 0
A3
∞ −−−→ (0, 2 3) −−−→ p3,2 −−→ ∞
√ T −1 Tv I0
T I0
I0
(1, 3) −−τ−−−−→ p0 −→
∞ −−1−→
∞
[R, A3 ] = Id
Tv I0
A3
Tv I0
I0
T I
A−1
4
A3 = R3 (Tv I0 )−2
I0 Tτ−1 Tv I0 T1 I0 = Tv−1 T1 Tτ−1
∞ −−→ p3,1 −−−→ p0 −→ ∞
A−1
3
0
∞ −−−→ p3,2 −−v−→
p4 −−−→ ∞
A−1
3
T I
A
0
4
p3,1 −−−→ ∞
p4 −−v−→
∞ −−→
I
A−1
4
A
0
3
p0 −−→
p4 −−−→ ∞
∞ −→
A3 = I0 Tv−1 I0 R3 Tv−1
3
A4 = Tv I0 A−1
3 R
A4 = I0 Tv−1 A3 R3
A4 = A3 I0 Tv−1 R3
A
Tv−1 A3
I
3
A4 = A−1
3 Tv I0 R
A3
A3
A3
A33 = R3
0
4
∞
∞ −−→
p4 −−−−→ p0 −→
∞ −−→ p3,1 −−→ p3,2 −−→ ∞
I0
A4
A3
∞ −→ p0 −−→ p3,2 −−→ ∞
A
Tv−1 A4
I
3
0
∞ −−→
p3,1 −−−−→ p0 −→
∞
3
A4 = A−1
3 R Tv I0
A4 = A3 R3 I0 Tv−1
Table 2: Action of generators on vertices for d = 3
Covering depth and Q[i]-rational points in D∞ (1):
We denote B ((z, t), r) the open extended Cygan ball centered at p = (z, t) ∈ ∂∞ H2C with radius r (see
Equation 8 for the definition of the extended Cygan metric). Recall that u(n) = √2n is the height at which balls
of depth n appear, in the sense of Corollary 1.
Lemma 11 Let u = u(5)+ε = 0.895 and Hu the horosphere of height u based at ∞. Then the prism D∞ (1)×{u}
is covered by the intersections with Hu of the following extended Cygan balls:
√
√
√
√
√ √
• (depth 1) B (0, 0), 2 , B (0, 2), 2 , B (τ, 0), 2 , B (τ, 2), 2 , B (1, 7), 2 ,
√
• (depth 2) B (1, 1), 4 2 .
We omit the proof, which is similar to the proof of Lemma 13 but simpler; see Figure 4.2.
Note that in order to cover D∞ (1) × {u} we only need balls of depth at most 2, in particular none of depth
4 even though they are present at the height u = u(5) + ε which we consider. It is however necessary to pass
to this height, as we have observed experimentally that at height u = u(4) + ε balls of depth at most 2 do not
cover D∞ (1) × {u} (there are no Q[i]-rational points of depth 3).
Corollary 3 The covering depth of Γ(1) is at most 4.
By inspection, we see that the Q[i]-rational points in D∞ (1) with depth at most 4 are, in horospherical
coordinates:
• Depth 1: (0, 0), (0, 2), (τ, 0) and (τ, 2), all in the same Γ∞ (1)-orbit;
• Depth 2: (1, 1)
11
Figure 2: Covering the prism D∞ (1) by Cygan balls of depth 1 and 2
• Depth 4: (0, 1), (τ, 1) in one Γ∞ (1)-orbit and (1, 0), (1, 2) in the other.
Integral lifts of representatives of Γ∞ (1)-orbits of these points are:
0
−1
i
−1
p0 = 0 p2 = 1 + i p4,1 = 0 p4,2 = 2
1
1+i
2
2
Generators: The following elements
above (for α = 0; 2; 4, 1; 4, 2) :
0
A0 = I 0 = 0
1
A4,1
i
= 0
2
Aα ∈ Γ(1) map the point ∞ = [1, 0, 0]T to the corresponding pα as
0
−1
0
0
−1
0
−1
1
0 A2 = 1 + i
1+i
0
1
0
−i
A4,2
−2
2+i
1+i
−1
= 2
2
−2
3
2
2+i
−1 − i
−i
2
−2
−1
Relations: We list in Table 3 the relations obtained for Γ(1) by applying generators to points of depth
at most 4 as described in part (5) of section 2.4. By successively eliminating A4,2 , A4,1 then A2 we obtain
presentation (13). The details are straightforward and left to the reader.
4.3
The Picard modular group Γ(7) = PU(2, 1, O7 )
Presentation and fundamental domain for the cusp stabilizer Γ∞ (7):
Lemma 12
1. The cusp stabilizer Γ∞ (7) admits the following presentation:
Γ∞ (7) =
D
T1 , Tτ , Tv , R
[Tτ , T1 ] = Tv , [Tv , T1 ], [Tv , Tτ ], [Tv , R], (RTτ )2 , (RT1 )2 = Tv , R2
E
√
2. Let D∞ (7)√⊂ ∂H2C be
√ the affine
√ convex hull of the points with horospherical coordinates (0, 0), 2(1, 7),
(τ, 0), (0, 2 7), (1, 3 7), (τ, 2 7). Then D∞ (7) is a fundamental domain for Γ∞ (7) acting on ∂HC \{∞}.
12
Image of vertex
Cycle of points
I0
Relation
I0
I02 = Id
∞ −→ p0 −→ ∞
I0 p 0 = ∞
A4,1
A4,1
A24,1 = Id
A4,2
A4,2
A24,2 = Id
A4,1 p4,1 = ∞
∞ −−−→ p4,1 −−−→ ∞
A4,2 p4,2 = ∞
∞ −−−→ p4,2 −−−→ ∞
Rp0 = p0
0
0
p0 −
→ p0 −→
∞
∞ −→
I
A4,1
I0 (τ, 2) = Tτ T2−1 p2
(Tτ I0 )
I0 Tτ T2−1
[I0 , Tτ ] = Tτ I0 R3
A
(Tv Tτ I0 )−1
A2 = T2 Tτ−1 I0 Tv Tτ I0 R2
2
p2 −−−−−−→ (τ, 2) −−−−−−−→ ∞
∞ −−→
A
Tτ−1 Tv I0
I
A2 = I0 Tv−1 Tτ I0 RT2−1 Tv−1
Tv2 I0
I
A4,1 = I0 Tv−2 I0 RTv−1
2
0
p2 −−−−−−→ p0 −→
∞
∞ −−→
A4,1
0
∞
∞ −−−→ p4,1 −−−→ p0 −→
A4,2
T2−1 I0
I
Tv2 I0
I0
0
∞ −−−→ p4,1 −−−−→ p0 −→
∞
A4,1
Tτ−1 T2 Tv4 p0
A4,2 = [I0 , T2 ]
A4,1 = I0 Tv−2 I0 RTv−1
∞ −−−→ p4,1 −−−→ p0 −→ ∞
A
A
A22 = R−1 T2−1 Tτ Tv−3
2
2
∞ −−→
p2 −−→
∞
A2 p 2 = ∞
A−1
2
Tv−4 T2−1 Tτ A2
A4,1
2 −1
4
−1
A4,1 = A−1
2 Tτ T2 Tv A2 R T2
∞ −−−→ p4,1 −−−−−−−−−→ p2 −−−→ ∞
A4,2
Tv−1 A2
A
Tv−1 A4,1
I
2
∞ −−−→ p4,2 −−−−→ p2 −−→
I
−1
−1
A4,2 = A−1
2 Tv A2 T2 Tτ R
∞
A4,1 = Tv I0 R−1 Tv2 I0
0
0
∞ −→
p0 −−−−−→ p0 −→
∞
A4,1 p0 = Tv p0
A−1
2
A4,1
A
A4,1 = A2 R2 Tτ−1 T2−1 Tv−2 A−1
2
2
∞ −−→
p2 −−−→ p2 −−−→ ∞
A4,1 p2 = p2
T2−1 A4,2
I
I
0
0
∞
∞ −→
p0 −−−−−→ p0 −→
A4,2 p0 = T2 p0
A4,2 p2 = Tv−1 (τ, 0)
(I0 Tv )3 = R
−1
I
T I
I0 p4,2 = T2 p0
A2 p4,2 = Tv p2
[R, A4,1 ] = Id
I0
0
0
(τ, 0) −−−−−−→ ∞
(τ, 0) −→
∞ −−τ−→
I0 p4,1 = Tv−2 p0
A2 p4,1 = Tτ−1 T2 Tv4 p2
Tv I0
∞ −−−→ (0, 2) −−−→ p0 −→ ∞
I0 p2 = Tv−1 Tτ p0
A2 p0 =
R
Tv I0
Tv−1 p0
I0 (τ, 0) = (τ, 0)
[R, I0 ] = Id
A−1
4,1
∞ −−−→ p4,1 −
→ p4,1 −−−→ ∞
Rp4,1 = p4,1
I0 (0, 2) =
I
R
A4,1 = [T2 , I0 ]
(T τ I0 )−1
Tv A4,2
A2
A4,2 = Tv−1 A2 R−1 Tτ T2−1 A−1
2
∞ −−→ p2 −−−−→ (τ, 0) −−−−−−→ ∞
Table 3: Action of generators on vertices for d = 1
√
Concretely, we use the following generators for Γ∞ (7) (denoting τ = 1+i2 7 ):
1 −τ̄
1 −1 τ − 1
1 Tτ = T(τ,0) = 0 1
T1 = T(1,√7) = 0 1
0 0
0 0
1
Tv = T(0,2√7)
1
= 0
0
0
1
0
√
i 7
0
1
1
R= 0
0
0
−1
0
−1
τ
1
0
0
1
√
Covering depth and Q[i 7]-rational points in D∞ (7):
We denote B ((z, t), r) the open extended Cygan ball centered at p = (z, t) ∈ ∂∞ H2C with radius r (see
Equation 8 for the definition of the extended Cygan metric). Recall that u(n) = √2n is the height at which balls
of depth n appear, in the sense of Corollary 1.
Lemma 13 Let u = u(8) + ε = 0.70711 and Hu the horosphere of height u based at ∞. Then the prism
D∞ (7) × {u} is covered by the intersections with Hu of the following extended Cygan balls:
√
√ √
√
√ √
√ √
• (depth 1) B (0, 0), 2 , B (0, 2 7), 2 , B (τ, 0), 2 , B (τ, 2 7), 2 , B (1, 7), 2 ,
√
√
√ √
• (depth 2) B (τ /2, 3 7/2), 4 2 , B ((τ + 1)/2, 7), 4 2
13
Figure 3: Covering the prism D∞ (7) by Cygan balls of depth 1, 2 and 4
√
• (depth 4) B (τ, 7), 1 .
Proof. Figure 4.3 shows the prism D∞ (7) and the relevant Cygan balls. We prove the result by dissecting
the prism D∞ (7) × {u} into affine polyhedra, each of which lies in one of the extended Cygan balls. This is
reminiscent of the proof of Proposition 5.2 of [Zh]. Consider the following points of ∂∞ H2C , in horospherical
coordinates (see Figure 4.3):
q1 = (0.65τ, 2.8)
q2 = (τ, 2)
q6 = (1, 4.3)
q7 = ((τ + 1)/2, 1.5)
√
q11 = (τ, 4)
q12 = ((τ + 1)/2, 2 7)
q16 = (0.3τ, 3)
q17 = (τ /2, 1)
q3 = (τ, 3.3) q4 = (0.35 + 0.65τ, 3.4) q5 = (0.7 + 0.3τ, 4.2)
q8 = (τ, 1.5)
q9 = ((τ + 1)/2,
q10 = (1, 1)
√ 0)
q13 = (τ /2, 4)
q14 = (τ /2, 2 7)
q15 = (0, 3.5)
q18 = (0, 1.7)
q19 = (τ /2, 0)
Denoting Hull(S) the affine hull (in horospherical coordinates) of a subset S ⊂ Hu ' ∂∞ H2C × {u}, we
claim that the following affinely convex pieces of D∞ (3) × {u} are each contained in the corresponding (open)
extended Cygan sphere:
√
• C1 = Hull ((0, 0), (1, 0), q9 , q10 , q17 , q18 , q19 ) ⊂ B (0, 0), 2
√
√
√ √
• C2 = Hull (0, 2 7), (1, 2 7), q6 , q12 , q13 , q14 , q15 ⊂ B (0, 2 7), 2
√
• C3 = Hull ((τ, 0), q7 , q8 , q9 , q17 , q19 ) ⊂ B (τ, 0), 2
√
√ √
• C4 = Hull (τ, 2 7), q11 , q12 , q13 , q14 ⊂ B (τ, 2 7), 2
√ √
√
• C5 = Hull (1, 7), q6 , q7 , q10 , q15 , q16 , q17 , q18 ⊂ B (1, 7), 2
√
√
• C6 = Hull (q1 , q3 , q4 , q5 , q11 , q12 , q13 , q15 , q16 ) ⊂ B (τ /2, 3 7/2), 4 2
√ √
• C7 = Hull (q1 , q2 , q4 , q5 , q6 , q7 , q8 , q16 , q17 ) ⊂ B ((τ + 1)/2, 7), 4 2
√
• C8 = Hull (q1 , q2 , q3 , q4 ) ⊂ B (τ, 7), 1
14
To verify each of these claims, we check numerically that each of the vertices indeed belongs to the ball in
question using Equation (8), then extend to √
the whole
√ affine covex hull√by Lemma
√ 7. For example,
√
the point
q1 = (0.65τ, 2.8) indeed belongs to B (τ /2, 3 7/2), 4 2 , B ((τ + 1)/2, 7), 4 2 and B (τ, 7), 1 because:
√
√
4
2 ' 1.189
dXC (0.65τ, 2.8, u), (τ /2, 3 7/2,
√ 0) ' 1.179 < √
dXC (0.65τ, 2.8, u), ((τ + 1)/2, √7, 0) ' 1.172 < 4 2 ' 1.189
dXC (0.65τ, 2.8, u), (τ, 7, 0) ' 0.982 < 1
The result then follows as the prism D∞ (7) × {u} is the union of the affinely convex pieces C1 , ..., C8 , see
Figure 4.3.
Corollary 4 The covering depth of Γ(7) is at most 7.
Note that in the above covering argument we have only needed balls of depth at most 4 (in particular none of
depth 7) even though they are present at the height u = u8 + ε which we consider. It is however necessary to
pass to this height, as we have observed √
experimentally that at height u = u7 + ε balls of depth at most 4 do
not cover D∞ (7) × {u} (there are no Q[i 7]-rational points of depths 5 or 6).
√
By inspection, we see that the Q[i 7]-rational points in D∞ (7) with depth at most 7 are the following, in
horospherical coordinates. We give in Lemma 14 below a detailed justifcation for the points of depth 2.
√
√
√
• Depth 1: (0, 0), (0, 2 7), (τ, 0), (τ, 2 7) and (1, 7), all in the same Γ∞ (7)-orbit;
√
√
• Depth 2: ( τ2 , 32 7) in one Γ∞ (7)-orbit and ( τ +1
2 , 7) in the other;
√
√
√
√
√
τ +1
• Depth 4: (0, 7), (τ, 7), (1, 0), (1, 2 7) in one Γ∞ (7)-orbit, ( τ2 , 12 7) in a second and ( τ +1
2 , 0), ( 2 , 2 7)
in the third;
√
√
7), (τ, 2k
7),
• Depth 7: for each k = 1, ..., 6, there is a Γ∞ (7)-orbit containing the 3 points (0, 2k
7
7
√
(1, 2k+7
7)
(with
2k
+
7
taken
mod
7).
7
Integral lifts of representatives of Γ∞ (7)-orbits of these points are:
0
p0 = 0
1
p2,1
p7,1
−1
0
= √
i 7
2
−τ − 2
= −1 p2,2 = 1 − τ
−τ
τ −1
p7,2
−2
0
= √
i 7
p7,3
−3
0
= √
i 7
p4,1
√
i 7
= 0
2
p4,2
p7,4
−4
0
= √
i 7
−1
τ −1
= τ p4,3 = τ + 1
2
2
p7,5
−5
0
= √
i 7
p7,6
−6
0
= √
i 7
√
√
√
Lemma 14 The Q[i 7]-rational points in D∞ (7) with depth 2 are exactly ( τ2 , 23 7) and ( τ +1
2 , 7).
Proof. We illustrate the general procedure for finding points of depth n, then specialize to the present case.
The depths at which there will be E-rational points are the natural numbers n such that |z|2 = n has a
solution z ∈ Od . For d = 7, the first few n’s are 1, 2, 4, 7, . . .. Begin by assuming we have found all points of
depth less than n. In this case those are just the points at depth 1.
1. Find all q = √
x + yτ ∈ Od (up to multiplication by a unit) with |q|2 = n. There are two possibilities: either
1+i 7
q = τ := 2 or q = τ − 1.
15
√
(1, 2 7)
q12
2
4
√
(1, 2 7)
√
(0, 2 7)
√
(τ, 2 7)
q14
4
4
2
√
(1, 2 7)
q12
2
2
q6
q6
q13
q15
q11
6
q4
q3
q16
q1
q5
6
8
7
8
5
5
q2
7
q18
5
q8
q10
q7
q10
q17
1
1
3
3
1
(1, 0)
q19
(0, 0)
(τ, 0)
q9
3
1
q9
(1, 0)
Figure 4: Affine cell decomposition of the prism D∞ (7)
16
(1, 0)
2. Consider the standard lift of a point of ∂HK2 \ {∞}:
P =
−|z|2 +it
2
z
1
√
Note that t = b 7 for some b ∈ Q. The calculation is more transparent if we rewrite the first coordinate:
−|z|2 + it
−|z|2 − b
=
+ bτ
2
2
For P to be the vector form of an E-rational point p of depth 2, it must satisfy the following:
(a) P does not have depth 1 (in general, P does not have depth less than n). In other words, the
coordinates of P are not in O7 .
(b) P q’s coordinates must be in O7 for some q from step (1).
Next, we do some calculations to make sure (b) is satisfied.
3. Find all z in the projection to C of D∞ such that zq ∈ O7 . If q = τ we can have z = 0, 1, τ, or 1+τ
2 . If
q = τ − 1, we can have z = 0, 1, τ, τ2 .
2
4. For each possible z, find |z|2 and compute −|z|2 −b + bτ q. Use this to list all b’s such that (z, b) ∈ D∞
2
and −|z|2 −b + bτ q ∈ O7 .
z
|z|2
q
0
0
1
1
τ
2
1+τ
2
τ
2
1
τ
τ −1
τ
τ −1
τ
τ −1
τ
1
2
τ −1
−|z|2 −b
2
+ bτ q
−2b + bτ
2
b
− 3b
2 − 2τ
2b + b−1
2 τ
1−3b
1+b
−
2
2 τ
−2b − b−2
2
2−3b
2+b
2 − 2 τ
2b + b−1
2 τ
1−6b
1+2b
4 − 4 τ
b’s
0, 2
0, 2
1
1
0, 2
0, 2
1
3
2
point(s) in horo. coords
√
(0, 0), (0, 2√7)
(0, 0), (0,
√ 2 7)
(1, √7)
(1, 7)√
(0, 0), (0, 2√7)
(0, 0), (0,√2 7)
1+τ
2 ,√ 7
τ 3 7
2, 2
5. Get rid ofnall the ones that areo
level 1 (or from any previous level). What you are left with are the level
√ τ 3√7
1+τ
2 points:
.
2 , 7 , 2, 2
Generators: The following elements Aα ∈ Γ(7) map the point ∞ = [1, 0, 0]T to the corresponding pα as
above (for α = 0; 2, 1; 2, 2; 4, 1; 4, 2; 4.3; 7, 1; ...; 7, 6) :
17
0
A0 = I 0 = 0
1
A4,1
√
i 7
= 0
2
0
−1
0
1
0
0
−5
−1
√
i 7
5
0√
−i 7
τ −1
τ
2
= τ
2
1−τ
2
1−τ
−τ
A2,1
0
4
1
0√
0 −i 7
−τ − 2
= −1
τ −1
A4,2
−1 0 0
0
1 0
= √
i 7 0 −1
A7,1
√
−4 0 3i 7
0
1
0
= √
5
i 7 0
A7,2
A7,5
√
−5 0 3i 7
0
1
0
= √
4
i 7 0
A2,2
−τ
τ −1
−2
τ
τ −1
2
2
= 1−τ
−τ
A4,3
−1 τ − 2
3
= τ +1
2
2−τ
A7,3
A7,4
√
i 7
0
3
−2 0
0
1
= √
i 7 0
−3
0
= √
i 7
A7,6
−6
0
= √
i 7
2
−τ − 1
−1
√
0 i 7
1
0
0
2
√
0 i 7
1
0
0
1
Relations: We list in Table 4 the relations obtained for Γ(7) by applying generators to points of depth at
most 7 as described in part (5) of section 2.4. By successively eliminating A2,1 , A7,1 , A7,4 , A7,6 , A7,2 , A7,3 ,
A7,5 , A2,2 , A4,3 then A4,1 , and setting I1 = A4,2 we obtain presentation (14). The details are straightforward
and left to the reader.
5
The Hurwitz quaternion modular group PU(2, 1, H)
In this section we use the method described in Section 2.4 to compute the following presentation for the
Hurwitz modular group Γ(H) = PU(2, 1, H), sometimes denoted PSp(2, 1, H). Recall that the Hurwitz integer
ring is H = Z[i, j, k, σ] ⊂ H, where we denote σ = 1+i+j+k
. The generators S∞ and relations R∞ for Γ∞ (H)
2
are defined in Lemma 15.
PU(2, 1, H) =
D
R∞ , I 2 , [I, Ri ], [I, Rσ ], [I, Ci ], [I, Cσ ], [ITv−2
IRi Tv−1
Ci , Rσ ],
i
i
S∞ , I
[Cσ−1 (ITv−1
Tv−1
Tv−1
)3 Ri2 Cσ , Ri ], [Cσ−1 (ITv−1
Tv−1
Tv−1
)3 Ri2 Cσ , Rσ ],
i
j
i
j
k
k
[Cσ−1 (ITv−1
Tv−1
Tv−1
)3 Ri2 Cσ , Cσ ], (ITv−2
IRi Tv−1
Ci )2 , (Tvi I)3 = Ri Ci ,
i
j
i
i
k
(Tvi Tvj I)4 = Ri2 , (Tvi Tvj Tvk I)6 , ITv−1
(ITvj )2 = (Tvj ITv−1
)2 Ri ,
i
i
Tv−1
Tv−1
Cσ−1 ,
Tv−1
)3 Ri2 Cσ ], IT1 Tσ−1 ITσ I = T1 Tv−1
[(Tvi Tvj I)2 , Cσ−1 (ITv−1
Tv−1
j
j
i
k
k
E i
I = Tσ ITvi Tvj IT1−1 Tσ Ci Tk
ITi Tk Tv−1
Tv−2
i
k
(15)
Presentation and coarse fundamental domain for Γ∞ (H):
In this section we study the action of Γ∞ (H) on ∂∞ H2H \ {∞} ' H × Im H. We use the notation of Equations (3), (4) and (5), namely T(ζ,v) , Rw , Cw respectively denote Heisenberg translation by (ζ, v), Heisenberg
rotation by w and conjugation by w. Note that for any unit quaternion w ∈ H and purely imaginary q ∈ H,
Γ∞ (H) contains the following Heisenberg translations:
Tw = T(w,i+j+k)
1
= 0
0
−w̄
1
0
−1+i+j+k
2
w
1
,
Tvq = T(0,q)
1
= 0
0
0
1
0
q
0
1
Lemma 15 Γ∞ (H) admits the following presentation:
Γ∞ (H) = hS∞ | R∞ i = Ri , Rσ , Ci , Cσ , T1 , Ti , Tj , Tk , Tσ , Tvi , Tvj , Tvk | A1 , A2 , A3 , A4 , A5
18
(16)
Image of vertex
Cycle of points
Relation
I0
I0
I02 = Id
∞ −→ p0 −→ ∞
I0 p0 = ∞
A4,1
A4,1
A24,1 = Id
A4,2
A4,2
A24,2 = Id
A4,3
A4,3
A24,3 = Id
A4,1 p4,1 = ∞
∞ −−−→ p4,1 −−−→ ∞
A4,2 p4,2 = ∞
∞ −−−→ p4,2 −−−→ ∞
A4,3 p4,3 = ∞
∞ −−−→ p4,3 −−−→ ∞
I0
→ p0 −→ ∞
∞ −→ p0 −
Rp0 = p0
I0 p2,2 = T1−1 p4,3
I0 p4,1 = Tv−1 p7,5
I0 p4,2 = T1−1 Tv−1 p2,1
I0 p4,3 = Tτ T1 Tv−1 p0
I0 p7,1 = Tv−1 p0
I0 p7,2 = Tv−1 p4,1
√
I0 (0, 2 7) = Tv−1 p7,6
√
I0 (1, 7) = Tv−1 p2,1
I0 (τ, 0) = (τ, 0)
A2,1 p0 = p7,5
A2,2 p0 = p4,2
A4,1 p0 = p7,4
A4,1 p7,3 = Tv p0
A4,2 p0 = p2,2
A4,2 p7,2 = Tτ T1−1 p2,2
I0
R
A2,2
[R, I0 ] = Id
T I
A−1
4,3
T I
A−1
7,5
0
p4,3 −−−→ ∞
∞ −−−→ p2,2 −−1−→
A4,1
0
p7,5 −−−→ ∞
∞ −−−→ p4,1 −−v−→
A4,2
A−1
2,1
T T I
v 1 0
−−→ p2,1 −−−→ ∞
∞ −−−→ p4,2 −−
A−1
4,3
I0 Tτ T1 Tv−1
I
0
p0 −−−−−−−→ p4,3 −−−→ ∞
∞ −→
A−1
7,1
I0 Tv−1
I
0
p0 −−−−→ p7,1 −−−→ ∞
∞ −→
A−1
7,2
I0 Tv−1
A4,1
∞ −−−→ p4,1 −−−−→ p7,2 −−−→ ∞
√ T I0
A−1
T I0
7,6
(0, 2 7) −−v−→
p7,6 −−−→ ∞
∞ −−v−→
√ T I0
A−1
T I0
2,1
∞ −−1−→
(1, 7) −−v−→
p2,1 −−−→ ∞
T I
(Tτ I0 )
I
A2,1
A−1
7,5
I
A2,2
A−1
4,2
I0
A4,1
A−1
7,4
0
∞ −→
p0 −−−→ p7,5 −−−→ ∞
0
∞ −→
p0 −−−→ p4,2 −−−→ ∞
∞ −→ p0 −−−→ p7,4 −−−→ ∞
Tv−1 A4,1
A7,3
I0
∞ −−−→ p7,3 −−−−−→ p0 −→ ∞
A−1
2,2
A4,2
I
0
∞ −→
p0 −−−→ p2,2 −−−→ ∞
A7,2
T1 Tτ−1 A4,2
I
T1−1 Tτ−1 Tv A4,3
A−1
2,2
∞ −−−→ p7,2 −−−−−−−→ p2,2 −−−→ ∞
I
A4,3 p0 = Tv−1 Tτ T1 p0
0
0
∞
∞ −→
p0 −−−−−−−−−−→ p0 −→
A7,1 p0 = p0
0
0
∞ −→
p0 −−−→ p0 −→
∞
A7,2 p4,1 = p0
A4,1
A7,3 p0 = p4,1
A7,3 p4,2 = T1−1 p2,1
A7,3 p7,1 = p7,4
A7,5 p2,2 = Tv T1−1 p4,3
A7,5 p4,1 = Tv p0
A7,5 p7,3 = p7,6
A7,5 p7,4 = ∞
A7,6 p0 = Tv p0
A7,6 p7,1 = ∞
A7,6 p7,2 = p7,5
A7,2
A7,3
I0
A7,2
A7,4
A−1
7,1
A7,2
∞ −−−→ p7,4 −−−→ p7,1 −−−→ ∞
I
A−1
4,1
A7,3
0
∞ −→
p0 −−−→ p4,1 −−−→ ∞
A4,2
A−1
2,1
T1 A7,3
∞ −−−→ p4,2 −−−−→ p2,1 −−−→ ∞
A7,1
A−1
7,4
A7,3
∞ −−−→ p7,1 −−−→ p7,4 −−−→ ∞
A7,2
A7,3
A−1
4,1
A7,4
0
∞ −→
p0 −−−→ p4,1 −−−→ ∞
A−1
4,3
T1 Tv−1 A7,5
A2,2
∞ −−−→ p2,2 −−−−−−−→ p4,3 −−−→ ∞
Tv−1 A7,5
A4,1
I
0
∞ −−−→ p4,1 −−−−−→ p0 −→
∞
A7,3
A−1
7,6
A7,5
∞ −−−→ p7,3 −−−→ p7,6 −−−→ ∞
A7,4
A7,5
∞ −−−→ p7,4 −−−→ ∞
I
Tv−1 A7,6
I
0
0
∞ −→
p0 −−−−−→ p0 −→
∞
A7,6
A7,1
∞ −−−→ p7,1 −−−→ ∞
A7,2
A4,3 = I0 Tτ T1 Tv−1 I0 Tτ−1 T1−1
A7,1 = I0 Tv−1 I0 R
A7,2 = I0 Tv−1 A4,1
A7,6 = (Tv I0 )2
A2,1 = Tv I0 T1 I0 Tv T1−2
A7,6
(I0 Tτ R)3 = Id
A7,5 = A2,1 I0 T1 Tv−1 R
A4,2 = A2,2 I0
A7,4 = A4,1 I0
A7,3 = A−1
4,1 Tv I0 R
A4,2 = A2,2 I0
−1
−2
A7,2 = A−1
4,2 Tτ T1 A2,2 Tτ T1 Tv
A4,3 = Tv−1 Tτ T1 I0 T1−1 Tτ−1 Tv I0
A7,1 = I0 RTv−1 I0
A7,2 = I0 Tv−1 A−1
4,1
A7,4 = A−1
7,2 A7,1 R
A7,3 = A4,1 RTv I0
A7,3 = T1−1 A2,1 Tτ−1 T12 Tv A−1
4,2
A7,4 = A7,3 A7,1
A7,3 = A−1
7,2 R
∞ −−−→ p7,2 −−−→ ∞
I
−1
2
A−1
4,2 = T1 Tτ Tv RA2,1 Tv T1 I0
A7,3 = A−1
7,2 R
∞ −−−→ p7,3 −−−→ ∞
A7,3 p7,2 = ∞
A7,4 p0 = p4,1
I
∞ −−−→ p4,1 −−−→ p0 −→ ∞
A7,2 p7,3 = ∞
A7,2 p7,4 = p7,1
A7,1
I
A7,5 = RTv I0 A4,1
−1
0
0
∞ −−τ−→
(τ, 0) −→
(τ, 0) −−−−−−→ ∞
I
A4,3 = T1 I0 A2,2 T1−1
A−1
7,5
∞ −−−→ p7,2 −−−→ p7,5 −−−→ ∞
19
A7,4 = A4,1 I0
A7,5 = Tv T1−1 A4,3 RTτ−2 T1 A−1
2,2
A7,5 = Tv I0 RA−1
4,1
A7,6 = A7,5 A7,3
A7,5 = RTv A−1
7,4
A7,6 = (Tv I0 )2
A7,6 = RTv A−1
7,1
A7,6 = A7,5 RA−1
7,2
Table 4: Action of generators on vertices for d = 7
where the An are the following sets of relations:
A1 = {Ri4 , Ri2 Rσ−3 , (Ri Rσ )3 , (Ri−1 Rσ )2 Ri−1 Rσ−2 }
A2 = {Ci2 , Cσ3 , (Ci Cσ )3 }
A3 =
[Tvw , T ] , [T1 , Tw ] = Tv−2
, [T1 , Tσ ] = Tv−1
Tv−1
Tv−1
,
w
i
j
k
2
−1
[Tw , Tŵ ] = Tvwŵ , [Tw , Tσ ] = Tvw Tvŵ Tvwŵ
where T runs over T1 , Ti , Tj , Tk , Tσ , Tvi , Tvj , Tvk , w runs over i, j, k, and î = j, ĵ = k, k̂ = i.
A4 = Ci Ri Ci = Ri , Ci Rσ Ci = Ri Rσ4 Ri , Cσ Ri Cσ−1 = Rσ Ri Rσ−1 , Cσ Rσ Cσ = Rσ
A5 = {GT G−1 = EG,T },
where G runs over Ri .Rσ , Ci , Cσ , T runs over T1 , Ti , Tj , Tk , Tσ , Tvi , Tvj , Tvk , and EG,T is the entry in the G
column and T row of the table below:
T1
Ti
Tj
Tk
Tσ
Tvi
Tvj
Tvk
Ri
Ti
−1
T1 Tvi Tvj Tvk
Tk
Tj−1 Tvi Tvj Tvk
Tσ−1 Ti Tk Tvj Tv−1
k
Tvi
Tvj
Tvk
Rσ
Tσ
−1 −1
T1 Tk Tσ Tvj Tv2k
T1−1 Ti−1 Tσ Tv2i Tvk
T1−1 Tj−1 Tσ Tvi Tv2j
T1−1 Tσ
Tvi
Tvj
Tvk
Ci
Tv−1
T1 Tv−1
j
k
Ti Tv−1
Tv−1
j
k
Tj−1 Tvi
Tk−1 Tvi
−1 −1
Tj Tk Tσ Tvj
Tvi
Tv−1
j
Tv−1
k
Cσ
T1
Tj
Tk
Ti
Tσ
Tvj
Tvk
Tvi
Proof. To obtain a presentation for Γ∞ (H), we identify 3 of its subgroups, observe that some of them
normalize each other, and build up the presentation via a sequence of extensions using the following procedure.
Suppose that G is a group with subgroups N and K where N is normal in G and G is an extension of N by K.
Suppose also that we know presentations for N and K, N = hSN | RN i and K = hSK | RK i. Then G admits the
presentation hSN ∪ SK | RN ∪ RK ∪ Ri, where the set R consists of relations of the form knk −1 = n0 where k runs
over all elements of SK , n runs over the elements of SN , and n0 ∈ N is expressed as a word in the generators
SN .
The three subgroups of Γ∞ (H) we identify are the rotation, conjugation, and translation subgroups.
The rotation subgroup consists of all Heisenberg rotations Rw where w is a Hurwitz integral unit quaternion.
It is isomorphic to the binary tetrahedral group, which has order 24 (see e.g. [CS]). It admits the presentation
hRi , Rσ | A1 i
The conjugation subgroup consists of all conjugations by unit quaternions. Elements of this group also
correspond to Hurwitz unit integral quaternions, only Cw acts the same as C−w . Thus, this group is isomorphic
to the quotient of the binary tetrahedral group by −1, which is the tetrahedral group (or the alternating group
on 4 elements). It admits the presentation
hCi , Cσ | A2 i
The translation subgroup consists of all Heisenberg translations. It admits the presentation
T1 , Ti , Tj , Tk , Tσ , Tvi , Tvj , Tvk | A3
The rotation subgroup is normalized by the conjugation subgroup. The extension of the rotation subgroup
by the conjugation subgroup is a finite group of order 288. We obtain the four relations in A4 by conjugating
Ri and Rσ by Ci and Cσ .
20
The translation subgroup is normalized by the rotation-conjugation subgroup. The conjugates of the translation generators by the rotation and conjugation generators are listed in the table. The translations T are
along the left side, the rotations/conjugations G are along the top, and the table entry EG,T is the element
GT G−1 written as a word in the translation generators. A5 contains these relations.
The action of the translation subgroup of Γ∞ (H) is relatively straightforward, however the finite group
Γ0,∞ (H) fixing 0 and ∞ is more complicated. Namely, the group of units in H (which has order 24 and is
generated by say i and σ) acts on ∂∞ H2H faithfully by rotations, and with kernel {±1} by conjugation. Together
these actions produce the group Γ0,∞ (H) as a semi-direct product of order 288 as described above, and the
geometric compatibility between the action of this finite group and the translation subgroup is not clear. For
this reason, we do not make explicit a fundamental domain for the action of Γ∞ (H) on ∂∞ H2H \ {∞}, rather we
only use a larger subset D∞ (H) which is geometrically simpler and sufficient for our purposes in the sense that
its Γ∞ (H)-translates cover ∂∞ H2H \ {∞}. Such a set is sometimes called a coarse fundamental domain.
1+j 1+k −1+i −1+j −1+k i+j i+k j+k
h
h
Define D∞
(H) = Hull( 12 , − 12 , 1+i
2 , 2 , 2 , 2 ,
2 ,
2 , 2 , 2 , 2 ) ⊂ H, and D∞ (H) = D∞ (H) ×
3
2
[0, 1] ⊂ ∂∞ HH \ {∞} ' H × Im H.
Lemma 16 The Γ∞ (H)-translates of D∞ (H) cover ∂∞ H2H \ {∞}.
Proof. We proceed in several steps, first considering the horizontal projections (as defined in the short exact
sequence (11)) Γh∞ = Π∗ (Γ∞ ) of Γ∞ = Γ∞ (H) and of some of its subgroups.
h
• Claim 1: A fundamental domain for hT1 , Ti , Tj , Tk i in H ' R4 is [−1/2, 1/2]4 .
This is clear as Heisenberg translations act as ordinary translations on the horizontal factor.
h
• Claim 2: A fundamental domain for hT1 , Ti , Tj , Tk , Ri , Rj , Rk i in H is [0, 1/2]4 ∪ [−1/2, 0] × [0, 1/2]3 :
This can be seen by subdividing the cube [−1/2, 1/2]4 into 16 cubes of half side-length. The group
hRi , Rj , Rk i, isomorphic to the classical quaternion group of order 8, acts on these 16 subcubes with 2
orbits, and the claim follows from choosing 2 adjacent representatives of these orbits.
h
h
• Claim 3: A fundamental domain for hT1 , Ti , Tj , Tk , Ri , Rj , Rk , Tσ i in H is D∞
(H):
h
First note that D∞ (H) ⊂ H is the union of two half-cubes of edge-length 1/2, D+ and D− , where
1+j 1+k i+j i+k j+k
−
D+ = Hull(0, 21 , 2i , 2j , k2 , 1+i
is obtained from D+ by negating the first
2 , 2 , 2 , 2 , 2 , 2 ) and D
coordinate. Each of these is obtained by cutting a cube in 2 along a diagonal hyperplane, for example
D+ P
is obtained by taking half of the cube [0, 1/2]4 , cut along the diagonal hyperplane {(x1 , ..., x4 ) ∈
4
R |
xi = 1}.
Note that this hyperplane is the equidistant hyperplane between 0 and Tσ (0); likewise the hyperplane
{(x1 , ..., x4 ) ∈ R4 | − x1 + x2 + x3 + x4 = 1} is the equidistant hyperplane between 0 and Tσ T1−1 (0). Since
h
h
Tσ2 = T1 Ti Tj Tk , the subgroup hT1 , Ti , Tj , Tk , Ri , Rj , Rk i has index 2 in hT1 , Ti , Tj , Tk , Ri , Rj , Rk , Tσ i ,
and the claim follows from the previous claim.
• Claim 4: A fundamental domain for T1 , Ti , Tj , Tk , Ri , Rj , Rk , Tσ , Tvi , Tvj , Tvk in ∂∞ H2H \{∞} ' H×Im H
h
is D∞
(H) × [−1, 1]3 :
This follows from the previous claim as the additional generators Tvi , Tvj , Tvk are vertical translations by
2 units in Heisenberg coordinates.
h
• Claim 5: The translates of D∞
(H) × [0, 1]3 under T1 , Ti , Tj , Tk , Ri , Rj , Rk , Tσ , Tvi , Tvj , Tvk , Ci , Cj , Ck
2
cover ∂∞ HH \ {∞}:
Recall from (5) that for any unit quaternion w, the matrix Cw acts on ∂∞ H2H ' H × Im H by conjugation
by w on both the horizontal and vertical factors. Since each of i, j, k conjugate the two others to their
opposites, the claim follows by considering only the action on the vertical factor.
Covering depth and Q[i, j, k]-rational points in D∞ (H):
21
We denote as before B ((ζ, v), r) the open extended Cygan ball centered at p = (ζ, v) ∈ ∂∞ H2H with radius r
(see Equation 8 for the definition of the extended Cygan metric dXC ). Recall that u(n) = √2n is the height at
which balls of depth n appear, in the sense of Corollary 1.
Lemma 17 Let u = u(6)+ε = 0.8165 and Hu the horosphere of height u based at ∞. Then the prism D∞ (H)×
√
{u} is covered by the intersections with Hu of the following extended Cygan balls of depth 1: B (0, 0), 2 ,
√
√
B (σ, i + j + k), 2 and B (σ − 1, i + j + k), 2 .
h
h
Proof. Recall that D∞ (H) = D∞
(H) × [0, 1]3, where the base D∞
(H) ⊂ H is the union
of two half-cubes of
1 i j k 1+i 1+j 1+k i+j i+k j+k
+
−
+
edge-length 1/2, D and D where D = Hull 0, 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2
and D− is obtained
from D+ by negating the first coordinate.
√
√
Claim: The half-prism D+ × [0, 1]3 is contained in B (0, 0), 2 ∪ B (σ, i + j + k), 2 .
We separate D+ × [0, 1]3 into 2 pieces by the ”horizontal” hyperplane A = {(x1 + x2 i + x3 j + x4 k, t1 i + t2√j +
t3 k) ∈ H×Im H | t1 +t2 +t3 = 2.5}. The claim is then verified by showing that D+ ×[0, 1]3 ∩A− ⊂ B (0, 0), 2
√
and D+ ×[0, 1]3 ∩A+ ⊂ B (σ, i + j + k), 2 , where A± denote the 2 half-spaces bounded by A, with (0, 0) ∈ A−
and (σ, i + j + k) ∈ A+ . This is done as previously by checking the vertices numerically using Equation (8), then
extending the result to their convex hull by Lemma 7. For the vertices, note that for each of the 11 base vertices
of D+ there is a 3-cube above it in the vertical direction, spanned by 0, i, j, k, i+j, i+k, j +k, i+j +k. We √
first
check using Equation (8) that for each of the 11 base vertices, all vertical vertices above it are in B (0, 0), 2
except for the top vertex
√ (corresponding to i+j +k), and that likewise all 11 of these top vertices are in the other
ball B (σ, i + j + k), 2 . Finally, we check that all 33 midpoints between these top vertices and the lower
√
√
level (spanned by i + j, i + k and j + k) are contained in the intersection B (0, 0), 2 ∩ B (σ, i + j + k), 2 .
j+k
j+k
For example, the midpoint ( j+k
2 , i + j + k/2) between ( 2 , i + j) and ( 2 , i + j + k) satisfies:
√
j+k
, i + j + k/2, u), (0, 0, 0) ' 1.187 < 2,
dXC (
2
√
j+k
, i + j + k/2, u), (σ, i + j + k, 0) ' 1.4127 < 2.
dXC (
2
+
−
All other computations are similar. Likewise, replacing D√
by
gives that
√ coordinate
D by negating the first
−
3
the other half-prism D × [0, 1] is contained in B (0, 0), 2 ∪ B (σ − 1, i + j + k), 2 .
Corollary 5 The covering depth of Γ(H) is at most 5.
In fact we suspect that the covering depth of Γ(H) is 4, however the covering argument is much more delicate
at height u(5) + ε, in particular the 3 Cygan balls used above do not suffice.
By inspection, we see that the Q[i, j, k]-rational points in D∞ (H) with depth at most 5 are, in horospherical
coordinates:
• Depth 1: (0, 0) and its Γ∞ -translates;
• Depth 2: (0, i + j) and its Γ∞ -translates;
i+j+k
) and their Γ∞ -translates;
• Depth 4: (0, i), (0, i + j + k), ( 1+i
2 ,
2
• Depth 5: (0, 4i+2j
5 ) and its Γ∞ -translates.
Integral lifts of these points are:
22
0
p0 = 0
1
−1
p2 = 0
i+j
p4,1
i
= 0
2
p4,2
i+j+k
1
1
p4,3 = −(i + k) p5 =
0
0
=
2
−2σ
−2i − j
Generators: The following elements Aα ∈ Γ(1) map the point ∞ = [1, 0, 0]T to the corresponding pα as
above (for α = 0; 2; 4, 1; 4, 2) :
0
A0 = I0 = 0
1
A4,2
i+j+k
0
=
2
0
−1
0
0
−1
0
1
0
0
−1
A2 = 0
i+j
0 i+j
1
0
0
1
2
1
0
0
A4,3 = −(i + k)
0
1
0
−(i + j + k)
−2σ
−(i + k) 1
A4,1
i
= 0
2
1
0
A5 =
−2i − j
0
−1
0
0
1
0
1
0
−i
2i + j
0
6
Relations: We list in Tables 5 and 6 the relations obtained for Γ(H) by applying generators to points of
depth at most 5 as described in part (5) of section 2.4. By successively eliminating the generators A5 , A4,3 , A4,2 ,
A4,1 then A2 , we obtain the presentation (15). Most of the details are straightforward and left to the reader, but
we now give more details on how to obtain the relations (Tvi Tvj Tvk I)6 and [(Tvi Tvj I)2 , Cσ−1 (ITv−1
Tv−1
Tv−1
)3 Ri2 Cσ ]
i
j
k
as these are not entirely straightforward.
Substituting for A2 using (5) in (6) (both from Table 6), we obtain
A4,2 = ITv−1
Tv−1
ITv−1
Tv−1
Tvk Tvi Tvj ITvi Tvj ICσ−1 Ri Cσ Tvi Tvj Tv−1
Cσ−1 Ci Cσ
i
j
i
j
k
Use the relations Cσ Tvw Cσ−1 = Tvŵ and [Ri , Tvw ] = 1 from Γ∞ (H) and the relations (2) and (3) from Table 5
to get
A4,2 = Cσ−1 ITv−1
Tv−1
ITvi ITvj Tvk ITvj Tvk Tv−1
Ri Ci Cσ
j
i
k
Use the relation (ITvi )3 = Ri Ci (which comes from eliminating A5 using (7) and (8)) to make the subsitution
ITvi I = Tv−1
ITv−1
Ri Ci .
i
i
A4,2 = Cσ−1 ITv−1
Tv−1
Tv−1
ITv−1
Ri Ci Tvj Tvk ITvj Tvk Tv−1
Ri Ci Cσ
j
i
i
i
k
Use the relations [Ci , Tvi ] = 1 and Ci Tvw Ci−1 = Tv−1
if w 6= i from Γ∞ (H) to get
w
Tv−1
Ri2 Cσ
A4,2 = Cσ−1 ITv−1
Tv−1
Tv−1
ITv−1
Tv−1
Tv−1
ITv−1
Tv−1
i
i
j
j
i
j
k
k
k
−1
−1 −1 −1 3 2
= Cσ (ITvi Tvj Tvk ) Ri Cσ
The relation (Tvi Tvj Tvk I)6 comes from substituting this expression for A4,2 in the relation (1) from Table
5. The relation [(Tvi Tvj I)2 , Cσ−1 (ITv−1
Tv−1
Tv−1
)3 Ri2 Cσ ] comes from first observing that (4) in Table 5 and (6)
i
j
k
in Table 6 are conjugate to each other by A2 . Since both are equal to A4,2 , we obtain [A2 , A4,2 ] = 1. Then to
get the relation that appears in the presentation, we substitute for A2 using (5) from Table 6 and for A4,2 using
the expression obtained above.
References
[A]
D. Allcock; New complex- and quaternion-hyperbolic reflection groups. Duke Math J. 103 (2000), no.
2, 303–333.
[Be]
M. Belolipetsky; On the volumes of arithmetic quotients of SO(1,n). Ann. Scuola Norm. Sup. Pisa Cl.
Sci. (5) III (2004), 749–770.
23
Image of vertex
Cycle of points
Relation
I0
I0
I02 = Id
A2
A2
A22 = Ri2
∞ −→ p0 −→ ∞
I0 p0 = ∞
∞ −−→ p2 −−→ ∞
A2 p 2 = ∞
A4,1
A4,1
A4,2
A4,2
A4,1 p4,1 = ∞
∞ −−−→ p4,1 −−−→ ∞
A4,2 p4,2 = ∞
∞ −−−→ p4,2 −−−→ ∞
Ri p0 = p0
Rσ p0 = p0
Ci p0 = p0
Cσ p0 = p0
Ri p2 = p2
Rσ p2 = p2
Ri p4,1 = p4,1
Rσ p4,1 = p4,1
Ri p4,2 = p4,2
Rσ p4,2 = p4,2
Ri p5 = p5
Rσ p5 = p5
Ci p4,1 = p4,1
Cσ p4,2 = p4,2
A4,1 p0 = Tvi p0
A4,1 p2 = Tv−1
p2
j
A4,2 p2 = Tvk p2
I0
Ri
I0
I0
Rσ
I0
I0
Ci
I0
I0
Cσ
I0
A24,1 = Id
−→ p0 −→ ∞
∞ −→ p0 −
∞ −→ p0 −−→ p0 −→ ∞
Rσ
A2 −1
∞ −−→ p2 −−→ p2 −−→
A4,1
Ri
A−1
4,1
A4,1
Rσ
A−1
4,1
A4,2
Ri
A−1
4,2
R
A−1
4,2
∞
[Rσ , A2 ] = Id
[Ri , A4,1 ] = Id
∞ −−−→ p4,1 −−→ p4,1 −−−→ ∞
[Rσ , A4,1 ] = Id
∞ −−−→ p4,2 −
−→ p4,2 −−−→ ∞
[Ri , A4,2 ] = Id
σ
∞ −−−→ p4,2 −−→
p4,2 −−−→ ∞
A5
Ri
A−1
5
A5
Rσ
A−1
5
[Rσ , A4,2 ] = Id
∞ −−→ p5 −
−→ p5 −−−→ ∞
[Ri , A5 ] = Id
∞ −−→ p5 −−→ p5 −−−→ ∞
A4,1
C
A−1
4,1
C
A−1
4,2
[Rσ , A5 ] = Id
i
∞ −−−→ p4,1 −→
p4,1 −−−→ ∞
A4,2
[Ci , A4,1 ] = Id
σ
∞ −−−→ p4,2 −−→
p4,2 −−−→ ∞
I
Tv−1 A4,1
i
[Cσ , A4,2 ] = Id
I
A4,1 = Tvi I0 Ri−1 Tv2i I0 Ci
0
0
∞ −→
p0 −−−−−→ p0 −→
∞
A
Tvj A4,1
A−1
2
A
Tv−1 A4,2
k
A−1
2
A4,1 = Tv−1
A2 Ri Tv−1
Tvj Ci A−1
2
j
i
2
∞ −−→
p2 −−−−−→ p2 −−−→ ∞
2
∞ −−→
p2 −−−−−→ p2 −−−→ ∞
(3)
[Ri , A2 ] = Id
∞ −−−→ p4,1 −
−→ p4,1 −−−→ ∞
A4,2
(2)
[Cσ , I0 ] = Id
A−1
2
−→ p2 −−−→ ∞
∞ −−→ p2 −
A2
[Ri , I0 ] = Id
[Ci , I0 ] = Id
∞ −→ p0 −−→ p0 −→ ∞
Ri
(1)
[Rσ , I0 ] = Id
∞ −→ p0 −→ p0 −→ ∞
A2
A24,2 = Id
A4,2 = Tvk A2 Cσ−1 Ri Cσ Tvi Tvj Tv−1
Cσ−1 Ci Cσ A−1
2
k
(4)
Table 5: Action of generators on vertices for H, degenerate cycles
[Bi]
L. Bianchi; Sui gruppi di sostituzioni lineari con coefficienti appartenenti a corpi quadratici
immaginari. Math. Ann. 40 (1892), 332–412.
[C]
E. Cartan; Sur le groupe de la géométrie hypersphérique. Comment. Math. Helv. 4 (1932), 158–171.
[CS]
J.H. Conway, D.A. Smith; On quaternions and octonions: their geometry, arithmetic and symmetry.
AK Peters (2003).
[CG]
S. Chen, L. Greenberg; Hyperbolic spaces, in Contributions to Analysis. Academic Press, New York
(1974), 49–87.
[DVV] J.P. Diaz, A. Verjovsky, F. Vlacci; Quaternionic Kleinian modular groups and arithmetic hyperbolic
orbifolds over the quaternions. arXiv:1503.07214.
[FFP] E. Falbel, G. Francsics, J. R. Parker; The geometry of the Gauss-Picard modular group. Math. Ann.
349 (2011), no. 2, 459–508.
[FP]
E. Falbel, J. R. Parker; The geometry of the Eisenstein-Picard modular group. Duke Math. J. 131
(2006), no. 2, 249–289.
24
Image of vertex
I0 p2 = Tv−1
Tv−1
p0
i
j
I0 p4,1 = Tv−2
p0
i
I0 (σ, i + j + k) = Tσ T1−1 p0
I0 p4,3 = Ti Tk Tv−2
Tv−1
Tv−2
p0
i
j
k
A2 p0 = Tvi Tvj p0
Cycle of points
A
Relation
Tvj Tvi I0
I
A2 = (I0 Tv−1
Tv−1
)2 Ri2
i
j
Tv2 I0
i
I
A4,1 = I0 Tv−2
I0 Ri Tv−1
Ci
i
i
2
0
p2 −−−−−−→ p0 −→
∞
∞ −−→
A4,1
0
∞ −−−→ p4,1 −−−→ p0 −→
∞
T I
T1 Tσ−1 I0
I
I0 T1 Tσ−1 I0 Tσ I0 = T1 Tv−1
Tv−1
Tv−1
Cσ−1
i
j
k
A4,3
Tv2 Tvj Tv2 Tk−1 Ti−1 I0
i
k
I
A4,3 = I0 Ti Tk Tv−2
Tv−1
Tv−2
I0
i
j
k
0
0
(σ, i + j + k) −−−−−→ p0 −→
∞
∞ −−σ−→
0
∞ −−−→ p4,3 −−−−−−−−−−−−−→ p0 −→
∞
I
Tv−1 Tv−1 A2
j
i
I
0
0
p0 −−−−−−−→ p0 −→
∞ −→
∞
A4,1
Tv−1 A2
i
A−1
2
A4,2
Tv−1 A2
A−1
A4,3
T −1 A2
A−1
A
Tvj Tv2 I0
i
A2 p4,1 = Tvi p2
∞ −−−→ p4,1 −−−−→ p2 −−−→ ∞
A2 p4,2 = Tvk p2
k
∞ −−−→ p4,2 −−−
−→ p2 −−2−→ ∞
A2 p4,3 = Tσ p2
σ
∞ −−−→ p4,3 −−
−−→ p2 −−2−→ ∞
I0 p5 = Tv−2
Tv−1
p0
i
j
A2 = (Tvi Tvj I0 )2
(5)
A4,1 =
−1 −1
−1
A−1
2 Tvi A2 Cσ Ri Cσ Tvi Tvj Cσ Ci Cσ
A4,2 =
(6)
−1
−1 −1
A−1
2 Tvk A2 Cσ Ri Cσ Tvi Tvj Tvk Cσ Ci Cσ
−1
−1
A4,3 = A−1
2 Tσ A2 T1 Tσ Tk Tvi Ci
I
A5 = [I0 , Tv−2
Tv−1
]
i
j
5
0
p5 −−−−−−→ p0 −→
∞
∞ −−→
A
Tv−1 Tv−2 I0
j
i
I
A2 p5 = Tv2i Tvj p0
5
0
∞ −−→
p5 −−−−−−−→ p0 −→
∞
A4,1 p5 = Tvi Tv−1
p0
j
5
i
0
∞ −−→
p5 −−−
−−−−−→ p0 −→
∞
A
Tv−1 Tvj A4,1
I
−1 3 −1
2
A5 = A−1
(8)
2 Tvi Tvj I0 Ri Tvi Tvj Ci
A5 =
−2 −2 −1
−1
−1 −1
A−1
4,1 Tvj Tvi I0 Cσ Ri Cσ Tvi Tvj Cσ Ci Cσ
Table 6: Action of generators on vertices for H, nondegenerate cycles
[GR]
H. Garland, M.S. Raghunathan; Fundamental domains for lattices in (R-)rank 1 semisimple Lie
groups. Ann. of Math. (2) 92 (1970) 279–326.
[G]
W.M. Goldman; Complex Hyperbolic Geometry. Oxford Mathematical Monographs. Oxford University
Press (1999).
[H]
R.-P.Holzapfel; Invariants of arithmetic ball quotient surfaces. Math. Nachr. 103 (1981), 117–153.
[KP]
I. Kim, J.R. Parker; Geometry of quaternionic hyperbolic manifolds. Math. Proc. Camb. Phil. Soc.
135 (2003), 291–320.
[M]
A.M. Macbeath; Groups of homeomorphisms of a simply connected space. Ann. of Math. (2) 79,
473–488.
[Ph]
Z. Philippe; Invariants globaux des espaces hyperboliques quaternioniques. PhD thesis, Université de
Bordeaux, 2016.
[P]
G. Prasad; Volumes of S-arithmetic quotients of semi-simple groups. Inst. Hautes Etudes Sci. Publ.
Math. 69 (1989), 91–117.
[PW]
J. Paupert, P. Will; Real reflections, commutators and cross-ratios in complex hyperbolic space.
Groups Geom. Dyn. 11 (2017), 311–352.
[Si]
C.L. Siegel; Discontinuous groups. Ann. of Math. (2) 44 (1943), 674–689.
[Ste]
R. Steinberg; Some consequences of the elementary relations in SLn . Cont. Math. 45 (1985), 335–350.
[Sto]
M. Stover; Volumes of Picard modular surfaces. Proc. Amer. Math. Soc. 139 (2011), no. 9, 3045–3056.
[Sw]
R.G. Swan; Generators and relations for certain special linear groups. Advances in Math. 6 (1971),
1–77.
[Y]
D. Yasaki; Integral cohomology of certain Picard modular surfaces. J. Number Theory 134 (2014),
13–28.
25
(7)
[W]
J.M. Woodward; Integral lattices and hyperbolic manifolds. PhD thesis, University of York, 2006.
[Zh]
T. Zhao; Generators for the Euclidean Picard modular groups. Trans. Amer. Math. Soc. 364 (2012),
3241–3263.
[Zi]
T Zink; Über die Anzahl der Spitzen einiger arithmetischer Untergruppen unitärer Gruppen. Math.
Nachr. 89 (1979), 315–320.
Alice Mark, Julien Paupert
School of Mathematical and Statistical Sciences, Arizona State University
[email protected], [email protected]
26
| 4math.GR
|
A Digital Neuromorphic Architecture Efficiently
Facilitating Complex Synaptic Response Functions
Applied to Liquid State Machines
Michael R. Smith∗ , Aaron J. Hill∗ , Kristofor D. Carlson∗ , Craig M. Vineyard∗ , Jonathon Donaldson∗ ,
David R. Follett† , Pamela L. Follett†‡ , John H. Naegle∗ , Conrad D. James∗ and James B. Aimone∗
arXiv:1704.08306v1 [q-bio.NC] 21 Mar 2017
∗ Sandia
National Laboratories, Albuquerque, NM 87185 USA
Email:{msmith4, ajhill, kdcarls, cmviney, jwdonal, jhnaegl, cdjame, jbaimon}@sandia.gov
† Lewis Rhodes Labs, Concord, MA 01742 USA
Email:{drfollett, plfollett}@earthlink.net
‡ Tufts University, Medford, MA 02155 USA
Abstract—Information in neural networks is represented as
weighted connections, or synapses, between neurons. This poses
a problem as the primary computational bottleneck for neural
networks is the vector-matrix multiply when inputs are multiplied
by the neural network weights. Conventional processing architectures are not well suited for simulating neural networks, often requiring large amounts of energy and time. Additionally, synapses
in biological neural networks are not binary connections, but
exhibit a nonlinear response function as neurotransmitters are
emitted and diffuse between neurons. Inspired by neuroscience
principles, we present a digital neuromorphic architecture, the
Spiking Temporal Processing Unit (STPU), capable of modeling
arbitrary complex synaptic response functions without requiring
additional hardware components. We consider the paradigm of
spiking neurons with temporally coded information as opposed to
non-spiking rate coded neurons used in most neural networks. In
this paradigm we examine liquid state machines applied to speech
recognition and show how a liquid state machine with temporal
dynamics maps onto the STPU—demonstrating the flexibility and
efficiency of the STPU for instantiating neural algorithms.
I. I NTRODUCTION
Neural-inspired learning algorithms are achieving state of
the art performance in many application areas such as speech
recognition [1], image recognition [2], and natural language
processing [3]. Information and concepts, such as a dog or
a person in an image, are represented in the synapses, or
weighted connections, between the neurons. The success of a
neural network is dependent on training the weights between
the neurons in the network. However, training the weights in a
neural network is non-trivial and often has high computational
complexity with large data sets requiring long training times.
One of the contributing factors to the computational complexity of neural networks is the vector-matrix multiplications
This work was supported by Sandia National Laboratories Laboratory
Directed Research and Development (LDRD) Program under the Hardware
Acceleration of Adaptive Neural Algorithms (HAANA) Grand Challenge
project. Sandia National Laboratories is a multi-mission laboratory managed
and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed
Martin Corporation, for the U. S. Department of Energys National Nuclear
Security Administration under Contract DE-AC04-94AL85000.
(the input vector multiplied by the synapse or weight matrix).
Conventional computer processors are not designed to process
information in the manner that a neural algorithm requires
(such as the vector-matrix multiply). Recently, major advances
in neural networks and deep learning have coincided with
advances in processing power and data access. However, we
are reaching the limits of Moore’s law in terms of how much
more efficiency can be gained from conventional processing
architectures. In addition to reaching the limits of Moore’s
law, conventional processing architectures also incur the von
Neumann bottleneck [4] where the processing unit’s program
and data memory exist in a single memory with only one
shared data bus between them.
In contrast to conventional processing architectures which
consist of a powerful centralized processing unit(s) that operate(s) in a mostly serialized manner, the brain is composed of
many simple distributed processing units (neurons) that are
sparsely connected and operate in parallel. Communication
between neurons occurs at the synaptic connection which
operate independently of the other neurons that are not involved in the connection. Thus, vector-matrix multiplications
are implemented more efficiently facilitated by parallel operations. Additionally, the synaptic connections in the brain are
generally sparse and information is encoded in a combination
of the synaptic weights and the temporal latencies of a spike
on the synapse [5]. Biological synapses are not simply a
weighted binary connection but rather exhibit a non-linear
synaptic response function due to the release and dispersion
of neurotransmitters in the space between neurons.
Biological neurons communicate using simple “data packets,” that are generally accepted as binary spikes. This is in
contrast to the neuron models used in traditional artificial neural networks (ANN) which are commonly rate coded neurons.
Rate coded neurons encode information between neurons as
a real-valued magnitude of the output of a neuron—a larger
output represents a higher firing rate. The use of rate coded
neurons stems from the assumption that the firing rate of a
Fig. 1. High level overview of the STPU. The STPU is composed of a set of leaky integrate and fire neurons. Each neuron has an associated temporal buffer
such that inputs can be mapped to a neuron with a time delay. W(t) is the neuronal encoding transformation which addresses connectivity, efficacy and
temporal shift. The functionality of the STPU mimics the of functionality of biological neurons.
neuron is the most important piece of information, whereas
temporally coded neurons encode information based on when
a spike from one neuron arrives at another neuron. Temporally
coded information has been shown to be more powerful than
rate coded information and more biologically accurate [6] .
Based on these neuroscience principles, we present the Spiking Temporal Processing Unit (STPU), a novel neuromorphic
hardware architecture designed to mimic neuronal functionality and alleviate the computational restraints inherent in conventional processors. Other neuromorphic architectures have
shown very strong energy efficiency [7], powerful scalability
[8], and aggressive speed-up [9] by utilizing the principles observed in the brain. We build upon these efforts leveraging the
benefits of low energy consumption, scalability, and run time
speed ups and include an efficient implementation of arbitrarily
complex synaptic response functions in a digital architecture.
This is important as the synaptic response function has strong
implications in spiking recurrent neural networks [10].
We also examine liquid state machines (LSMs) [11] to show
how the constructs available in the STPU facilitate complex
dynamical neuronal systems. While we examine the STPU in
the context of LSMs, the STPU is a general neuromorphic
architecture. Other spiked-based algorithms have been implemented on the STPU [12], [13].
In Section II, we present the STPU. A high level comparison
with other neuromorphic architectures is presented in Section
III. We present LSMs in Section IV. In Section V, we examine
how LSMs map onto the STPU and show results from running
the LSM on the STPU. We conclude in Section VI.
II. T HE S PIKING T EMPORAL P ROCESSING U NIT
In this section, we describe the Spiking Temporal Processing
Unit (STPU) and how the components in the STPU map
to functionality in biological neurons. The design of the
STPU is based on the following three neuroscience principles
observed in the brain: 1) the brain is composed of simple
processing units (neurons) that operate in parallel and are
sparsely connected, 2) each neuron has its own local memory for maintaining temporal state, and 3) information is
encoded in the connectivity, efficacy, and signal propagation
characteristics between neurons. A high-level overview of a
biological neuron and how its components map onto the STPU
are shown in Figure 1. The STPU derives its dynamics from
the leaky integrate and fire (LIF) neuron model [14]. Each
LIF neuron j maintains a membrane potential state variable,
vj , that tracks its stimulation at each time step based on the
following differential equation [10]:
dvj
vj X X
=− +
wkj · s(t − tkl − ∆kl ).
(1)
dt
τj
k
l
The variable τj is the time constant of the first-order dynamics,
k is the index of the presynaptic neuron, wkj is the weight
connecting neuron j to neuron k, tkl is the time of the lth spike
from neuron k, ∆kl is the synaptic delay from neuron k on the
lth spike, and s(·) is the dynamic synaptic response function
to an input spike. In the LIF model, neuron j will fire if vj
exceeds a threshold θj . The synapses between input neurons
to destination neurons are defined in the weight matrix W(t)
for a given time t as the weights between inputs and neurons
can change over time.
Unique to the STPU, each LIF neuron has a local temporal
memory buffer R composed of D memory cells to model
synaptic delays. When a biological neuron fires, there is a
latency associated with the arrival of the spike at the soma of
the postsynaptic neuron due to the time required to propagate
down the axon of the presynaptic neuron and the time to
propagate from the dendrite to the soma of the postsynaptic
neuron (∆kl ). The temporal buffer represents different synaptic
junctions in the dendrites where a lower index value in the
temporal buffer constitutes a dendritic connection closer to the
soma and/or a shorter axon length than one with a larger index
value. Thus, synapses in the STPU are specified as a weight
wkjd from a source input neuron k, to a destination neuron j in
the dth cell of the temporal buffer, d ∈ {0, 1, 2, . . . , D − 1}.
This allows multiple connections between neurons with different synaptic delays. At each time step a summation of
the product of the inputs i(t) and synaptic weights W(t)
occurs and is added to the current value
P in that position of
the temporal buffer R̂d (t) = Rd (t) + k ik (t)wkjd (t) where
R̂(t) is a temporary state of the temporal buffer. The value
in each cell of the temporal buffer is then shifted down one
position, that is Rd (t+1) = R̂d−1 (t). The values at the bottom
of the buffer are fed into the LIF neuron.
In biological neurons, when a neuron fires a (near) binary
spike is propagated down the axon to the synapse, which
defines a connection between neurons. The purpose of the
synapse is to transfer the electric activity or information from
one neuron to another neuron. Direct electrical communication
does not take place, rather a chemical mediator is used. In the
presynaptic terminal, an action potential from the emitted spike
causes the release of neurotransmitters into the synaptic cleft
(space between the pre and postsynaptic neurons) from the
synaptic vescles. The neurotransmitters cross the synaptic cleft
and attach to receptors on the postsynaptic neuron injecting
a positive or negative current into the postsynaptic neuron.
Through a chemical reaction, the neurotransmitters are broken
down in receptors on the postsynaptic neuron and are released
back into the synaptic cleft where the presynaptic neuron
reabsorbs the broken down molecules to synthesize new neurotransmitters. In terms of electrical signals, the propagation
of activation potentials on the axon is a digital signal as shown
in Figure 2. However, the chemical reactions that occur at the
synapse to release and reabsorb neurotransmitters are modeled
as an analog signal.
The behavior of the synapse propagating spikes between
neurons has important ramifications on the dynamics of the
liquid. In Equation 1, the synaptic response function is represented by s(·). Following Zhang et al. [10], the Dirac delta
function δ(·) can be used as the synaptic response function
and is convenient for implementation on digital hardware.
However, the Dirac delta function exhibits static behavior.
Zhang et al. show that dynamical behavior can be modeled in
the synapse by using the first-order response to a presynaptic
spike:
1 − t−tkls−∆kl
τ
e
· H(t − tkl − ∆kl )
(2)
τs
where τ s is the time constant of the first-order response, H(·)
is the Heaviside step function, and 1/τ s normalizes the firstorder response function. The dynamical behavior can also be
implemented using a second-order dynamic model for s(·):
t−tkl −∆kl
t−tkl −∆kl
1
−
−
s
s
τ1
τ2
(e
−
e
) · H(t − tkl − ∆kl ) (3)
τ1s − τ2s
Fig. 2. Spike propagation along the axon and across the synapse. The spike
propagated on the axon is generally accepted as a binary spike. Upon arrival
at the synapse, the spike initiates a chemical reaction in the synaptic cleft
which stimulates the postsynaptic neuron. This chemical reaction produces
an analog response that is fed into the soma of the postsynaptic neuron. In
the STPU, arbitrary synaptic response functions are modeled efficiently using
the temporal buffer. The synaptic response function is discretely sampled and
encoded into the weights connecting one neuron to another and mapped to
the corresponding cells in the temporal buffer.
where τ1s and τ2s are the time constants for the second
order response and 1/(τ1s − τ2s ) normalizes the second-order
dynamical response function. Zhang et al. showed significant
improvements in accuracy and the dynamics of the liquid when
using these dynamical response functions.
Implementing exponential functions in hardware is expensive in terms of the resources needed to implement exponentiation. Considering that the STPU is composed of
individual parallel neuronal processing units, each neuron
would need its own exponentiation functionality. Including the
hardware mechanisms for each neuron to do exponentiation
would significantly reduce the number of neurons by orders
of magnitude as there are limited resources on an FPGA.
Rather than explicitly implement the exponential functions in
hardware, we use the temporal buffer associated with each
neuron. The exponential function is discretely sampled and
the value at each sample is assigned a connection weight
wkjd from the presynaptic neuron k to the corresponding
cell d in the temporal buffer of the postsynaptic neuron j.
Thus, a single weighted connection between two neurons is
expanded to multiple weighted connections between the same
two neurons. This is shown graphically in Figure 2. The use
of the temporal buffer allows for an efficient implementation
of the digital signal propagation down the axon of a neuron or
TABLE I
H IGH - LEVEL C OMPARISON OF THE STPU WITH T RUE N ORTH AND S PI NNAKER .
Platform:
STPU
TrueNorth
SpiNNaker
Interconnect:
Neuron Model:
Synapse Model:
3D mesh multicast1
LIF
Programmable4
2D mesh unicast
LIF2
Binary
2D mesh multicast
Programmable3
Programmable5
1 The 3D mesh is enabled due to the temporal buffer available for each neuron in STPU.
2 TrueNorth provides a highly programmable LIF to facilitate additional neural dynamics.
3 SpiNNaker provides flexibility for the neuron model, however more complex biological models
4 The synapse model
5 As with the neuron
are more computationally expensive.
is programmable in the STPU via the temporal buffer by discretely sampling an arbitrary synapse model.
model, SpiNNaker is optimized for simpler synaptic models. More complex synaptic models incur a cost in computational complexity.
the analog signal propagation between neurons at the synapse.
III. C OMPARISON WITH OTHER N EUROMORPHIC
A RCHITECTURES
The STPU is not the first neuromorphic architecture. Four
prominent neuromorphic architectures are IBM’s TrueNorth
chip [7], the Stanford Neurogrid [15], the Heidelberg BrainScaleS machine [16] and the Manchester Spiking Neural Network Architecture (SpiNNaker) [17]. The Stanford Neurogrid
and the Heidelberg BrainScaleS are analog circuits while
TrueNorth and SpiNNaker are digital circuits. As the STPU
is also a digital system, we will focus on a comparison with
TrueNorth and SpiNNaker.
The TrueNorth chip leverages a highly distributed crossbar based architecture designed for high energy-efficiency
composed of 4096 cores. The base-level neuron is a highly
parametrized LIF neuron. A TrueNorth core is a 256 × 256
binary crossbar where the existence of the synapse is encoded
at each junction, and individual neurons assign weights to
particular sets of input axons. The crossbar architecture allows
for efficient vector-matrix multiplication. TrueNorth only allows for point-to-point routing. Each of the 256 neurons on
a core is programmed with a spike destination addressed to a
single row on a particular core which could be the same core,
enabling recurrence, or a different core. The crossbar inputs
are coupled via delay buffers to insert axonal delays. A neuron
is not natively able to connect to multiple cores or to connect
to a single neuron with different temporal delays. As a work
around, a neuron is to be replicated within the same core and
mapped to the different cores. For multiple temporal delays
between two neurons (such as those in the STPU), there is no
obvious mechanism for an implementation [18].
SpiNNaker is a massively parallel digital computer composed of simple ARM cores with an emphasis on flexibility.
Unlike the STPU and TrueNorth, SpiNNaker is able to model
arbitrary neuron models via an instruction set that is provided
to the ARM core. SpiNNAker is designed for sending large
numbers of small data packages to many destination neurons.
While SpiNNaker was designed for modeling neural networks,
it could potentially be used more generally due to its flexibility.
The STPU architecture falls in between the TrueNorth
and SpiNNaker architectures. The STPU implements a less
parameterized LIF neuron than TrueNorth, however, its routing
of neural spikes is more flexible and allows a multicast similar
to SpiNNaker rather than the unicast used in TrueNorth. A
key distinguishing feature of the STPU is the temporal buffer
associated with each neuron, giving the STPU 3-dimensional
routing. A high-level summary of the comparison of STPU
with TrueNorth and SpiNNaker is shown in Table I.
IV. L IQUID S TATE M ACHINES
The liquid state machine (LSM) [11] is a neuro-inspired
algorithm that mimics the cortical columns in the brain.
It is conjectured that the cortical microcircuits nonlinearly
project input streams into a high-dimensional state space.
This high-dimension representation is then used as input to
other areas in the brain where learning can be achieved. The
cortical microcircuits have a sparse representation and fading
memory—the state of the microcircuit “forgets” over time.
While LSMs may be able to mimic certain functionality in
the brain, it should be noted that LSMs do not try to explain
how or why the brain operates as it does.
In machine learning, LSMs are a variation of recurrent neural networks that fall into the category of reservoir computing
(RC) [19] along with echo state networks [20]. LSMs differ
from echo state machines in the type of neuron model used.
LSMs use spiking neurons while echo state machines use rate
coded neurons with a non-linear transfer function.
LSMs operate on temporal data composed of multiple
related time steps. LSMs are composed of three general
components: 1) input neurons, 2) randomly connected leakyintegrate and fire spiking neurons called the liquid, and 3)
readout nodes that read the state of liquid. A diagram of an
LSM is shown in Figure 3. Input neurons are connected to
a random subset of the liquid neurons. The readout neurons
may be connected to all the neurons in the liquid or a subset
of them. Connections between neurons in the liquid are based
on probabilistic models of brain connectivity [11]:
Pconnection (N1 , N2 ) = q · e−
E(N1 ,N2 )
r2
(4)
where N1 and N2 represent two neurons and E(N1 , N2 ) is
the Euclidean distance between N1 and N2 . The variables q
and r are two chosen constants. In this paper, we use a 3dimensional grid to define the positions of neurons on the
liquid. The liquid functions as a temporal kernel, casting the
input data into a higher dimension. The LIF neurons allow
for temporal state to be carried from one time step to another.
LSMs avoid the problem of training recurrent neural models
V. M APPING THE LSM ONTO THE STPU
Fig. 3. A liquid state machine, composed of three components: 1) a set of
input neurons, 2) the liquid—a set of recurrent spiking neurons, and 3) a set
of readout neurons with plastic synapses that can read the state of the neurons
in the liquid.
TABLE II
PARAMETERS FOR THE SYNAPSES ( OR CONNECTIONS BETWEEN
NEURONS ) IN THE LIQUID .
Parameter
type
r from Equation 4
ALL
value
q from Equation 4
E→E
E→I
I→E
I→I
0.45
0.30
0.60
0.15
Synaptic weight
E→E
E→I
I→E
I→I
3
6
-2
-2
2
by only training the synaptic weights from the liquid to the
readout nodes, similar to extreme machine learning that use
a random non-recurrent neural network for non-temporal data
[21]. It is assumed that all temporal integration is encompassed
in the liquid. Thus, the liquid in an LSM acts similarly to
the kernel in a support vector machine on streaming data
by employing a temporal kernel. In general, the weights and
connections in the liquid do not change, although some studies
have looked at plasticity in the liquid [22].
The readout neurons are the only neurons that have plastic
synapses, allowing for synaptic weight updates via training.
Using each neurons firing state from the liquid, the temporal
aspect of learning on temporal data is transformed to a static
(non-temporal) learning problem. As all temporal integration
is done in the liquid, no additional mechanisms are needed to
train the readout neurons. Any classifier can be used, but often
a linear classifier is sufficient. Training of the readout neurons
can be done in a batch or on-line manner [10].
LSMs have been successfully applied to several applications
including speech recognition [10], vision [23], and cognitive
neuroscience [11], [24]. Practical applications suffer from the
fact that traditional LSMs take input in the form of spike trains.
Transforming numerical input data into spike data, such that
the non-temporal data is represented temporally, is nontrivial.
In this section, we implement the LSM on the STPU. There
have been previous implementations of LSMs on hardware,
however, in most cases an FPGA or VLSI chip has been
designed specifically for a hardware implementation of an
LSM. Roy et al. [25] and also Zhang et al. [10] present
a low-powered VLSI hardware implementation of an LSM.
Schrauwen et al. [26] implement an LSM on an FPGA chip.
In contrast to other work, the STPU has been developed to
be a general neuromorphic architecture. Other neuroscience
work and algorithms are being developed against the STPU
such as spike sorting and using spikes for median filtering
[13]. Currently, we have an STPU simulator implemented in
MATLAB as well as an implementation on an FGPA chip. The
MATLAB simulator has a one-to-one correspondence with the
hardware implementation.
Given the constructs provided by the STPU, the LSM with
a liquid composed of LIF neurons maps naturally onto the
STPU. We use the second-order synaptic response function of
Equation 3 that is based on the work of Zhang et al. [10].
They found that the second-order response function produced
more dynamics in the liquid allowing the neural signals to
persist longer after the input sequence had finished. This lead
to improved classification results. Following Zhang et al., the
synaptic properties of the liquid, including parameters for the
connection probabilities between the liquid neurons defined in
Equation 4 and the synaptic weights, are given in Table II.
There are two types of neurons: excitatory (E) and inhibitory
(I). As has been observed in the brain [11], the liquid is
made up of an 80/20 network where 80% of the nuerons
are excitatory and 20% of the neurons are inhibitory. The
probability of a synapse existing between two neurons and
the weights between the neurons is dependent on the types of
the considered neurons. E/I → E/I denotes the presynaptic
and postsynaptic neurons being connected by the synapse. For
example, E → I denotes the connection between an excitatory
presynaptic neuron with an inhibitory postsynaptic neuron.
Excitatory neurons increase the action potential at a target
neurons (positive synaptic weights) while the inhibitory neurons decrease the action potential (negative synaptic weights).
When the connections are generated between neurons in the
liquid, the neurons are randomly connected according to
Equation 4 with the parameters given in Table II. Each input
neuron is randomly connected to a subset of 30% of the
neurons in the liquid with a weight of 8 or -8 chosen uniformly
at random.
To implement the second-order synaptic response function,
Equation 3 is sampled at discrete time steps and multiplied by
the synaptic weight value between the neurons as specified in
Table II. The discretely sampled weights are then encoded via
multiple weights at corresponding cells in the temporal buffer
for the postsynaptic neuron. In this implementation, there is
no synaptic delay (∆kl = 0) and τ1s is set to 4 and τ2s is set
to 8 for excitatory neurons. For inhibitory neurons, τ1s and τ2s
are set to 4 and 2 respectively. For all neurons, τj is set to 32.
TABLE III
S EPARATION VALUES , AVERAGE SPIKING RATES , AND CLASSIFICATION
ACCURACY FROM DIFFERENT SYNAPTIC RESPONSE FUNCTIONS .
Synaptic Res
TrainSep
TrainRate
TestSep
TestRate
SVM
A. Experiments
Dirac Delta
First-Order
Second-Order
0.129
0.251
0.263
0.931
0.845
0.261
0.139
0.277
0.290
0.931
0.845
0.255
0.650
0.797
0.868
First-Order 30
First-Order 40
First-Order 50
0.352
0.293
0.129
0.689
0.314
0.138
0.389
0.337
0.134
0.688
0.314
0.138
0.811
0.817
0.725
To evaluate the effect of different parameters for the liquid
state machine, we use a data set for spoken digit recognition
of Arabic digits from 0 to 9 [27]. The dataset is composed of
the time series Mel-frequency cepstral coefficients (MFCCs)
of 8800 utterances of each digit from 88 speakers on with 10
repetitions per digit (10 × 10 × 88). The MFCCs were taken
from 44 male and 44 female native Arabic speakers between
the ages of 18 and 40. The dataset is partitioned into a training
set from 66 speakers and a test set from the other 22 speakers.
We scale all variables between 0 and 1.
To evaluate the performance of the LSM, we examine
the classification accuracy on the test set, and measure the
separation in the liquid from the training set. If there is good
separation within the liquid, then the state vectors from the
trajectories for each class should be distinguishable from each
other. To measure the separability of a liquid Ψ on a set of
state vectors O from the liquid perturbed by a given input
sequence, we follow the definition from Norton and Ventura
d
where cd is the inter-class distance
[22]: Sep(Ψ, O) = cvc+1
and cv is the intra-class variance. Separation is the ratio of the
distance between the classes divided by the class variance.
The inter-class difference is the mean difference of
the
mass for every pair of classes: cd =
Pn center
Pn ofkµ(O
l )−µ(Om )k2
where k·k2 is the L2 norm, n is
l=1
m=1
n2
the number of classes, and µ(Ol ) is the center of mass for each
class. For a given class, the intra-class variance is the mean
variance of the state vectors from thePinputs from the center
Pn
kµ(Ol )−ok k2
of mass for that class: cv = n1 l=1 ok ∈Ol |Ol |
.
We investigate various properties of the liquid state machine, namely the synaptic response function, the input encoding scheme, the liquid topology, and the readout training algorithm. We also consider the impact of θj on the
liquid. A neuron j will spike if vj exceeds θj . Thus, θj
can have a significant impact on the dynamics of the liquid. Beginning with a base value of 20 (as was used by
Zhang et al.) we consider the effects of decreasing values
of θj ∈ {20, 17.5, 15, 12.5, 11, 10, 9, 7.5, 5, 3, 2.5, 2, 1}. For
default parameters, we use a reservoir of size 3 × 3 × 15, we
feed the the magnitude of the inputs into the input neurons
(current injection) and a linear SVM to train the synapses for
the readout neurons.
1) Synaptic Response Functions: We first investigate the
effect of the synaptic response function using default parameters. Using θj = 20, the average separation values, average
spiking rates and the classification accuracy of a linear SVM
are given in Table III. As highlighted in bold, the second-order
synaptic function (Equation 3) achieves the largest separation
values for training and testing, the lowest average spike rate,
and the highest classification accuracy. The average spike rate
is significantly higher for the first-order response function
response val
The plastic readout neurons are connected to all of the
neurons in the liquid. Training is done off-line using a linear
classifier on the average firing rate of the neurons in the liquid.
We examine the effect of the various linear classifiers below.
0.2
First-Order
Second-Order
0.1
0
0
5
10
15
20
time
25
30
35
Fig. 4. Visualization of the first- and second-order response functions.
than for the second-order response function, which is counterintuitive since the second-order response function perpetuates
the signal through the liquid longer. However, examining the
first-order and second-order response functions, as shown in
Figure 4, shows that the first-order response function has a
larger initial magnitude and then quickly subsides. The secondorder response function has a lower initial magnitude, but is
slower to decay giving a more consistent propagation of the
spike through time.
Adjusting the value of θj to 30, 40, and 50 accommodate
this behavior for the first-order response function (the bottom
three rows int he Table III) shows that an improvement can be
made in the separation values, spiking rate, and classification
accuracy. Despite this improvement, the first-order response
function does not achieve a better performance than the
second-order response function for the classification accuracy.
The first-order response function does get a better separation
score, but this does not translate into better accuracy.
2) Input Encoding Schemes: In traditional LSMs, the input
is temporally encoded in the form of a spike train. Unfortunately, most datasets are not temporally encoded, but
rather are numerically encoded. A spike train input aligns
with neuroscience, but practically it is non-trivial to encode
all information temporally as the brain does. Therefore, we
examine three possible encoding schemes: 1) rate encoding
where the magnitude of the numeric value is converted into
a rate and a spike train at that rate is fed into the liquid 2)
bit encoding where the magnitude of the numeric value is
converted into its bit representation at a given precision, and
3) current injection. Rate encoding requires n time steps to
encode a single input by converting a magnitude to a rate.
This is similar to binning and has some information loss. Bit
encoding, only requires one time step, however, it requires
m inputs per standard input to convert the magnitude into its
TABLE IV
T HE SEPARATION VALUES , AVERAGE SPIKING RATES OF THE LIQUID , AND
CLASSIFICATION ACCURACY EXAMINING LIQUID TOPOLOGIES WITH
DIFFERENT INPUT ENCODING SCHEMES AND VALUES FOR θj . T HE
LARGEST SEPARATION VALUES AND ACCURACIES FOR EACH ENCODING
SCHEME ARE IN BOLD
θj
Encoding Scheme
20
15
10
5.5
3
TABLE V
T HE SEPARATION VALUES AND AVERAGE SPIKE RATE OF THE LIQUID
USING DIFFERENT LIQUID TOPOLOGIES . T HE LARGEST SEPARATION
VALUES AND ACCURACIES FOR EACH TOPOLOGY ARE IN BOLD
Topology/
Num Neurons
θj :
2
Current Injection
0.263
0.261
0.868
0.409
0.580
0.905
0.378
0.750
0.894
0.334
0.843
0.873
0.324
0.873
0.868
0.324
0.878
0.866
Bit Encoding
0.271
0.434
0.741
0.310
0.497
0.741
0.338
0.544
0.735
0.350
0.592
0.755
0.353
0.634
0.764
0.357
0.620
0.761
Rate Encoding
0.164
0.146
0.747
0.364
0.199
0.733
0.622
0.594
0.643
0.197
0.952
0.601
0.047
0.985
0.548
0.048
0.985
0.558
m bit-precise representation. We set m to 10. Compared to
current injection, the execution time increases linearly in the
number of time steps for rate encoding.
Table IV shows the separation values (first row for each
encoding scheme), average spiking rates (second row), and the
accuracy from a linear SVM on the test set (third row) for the
input encoding schemes with various values of θj . The average
spiking rate gives the percentage of neurons that were firing
in the liquid through the time series and provides insight into
how sparse the spikes are within the liquid. Table IV shows
a representative subset of the values that were used for θj .
The bold values represent the highest separation value and
classification accuracy for each encoding scheme.
The results show that the value of θj has a significant effect
on the separation of the liquid as well as the classification
accuracy from the SVM. This is expected as the dynamics of
the liquid are dictated by when neurons fire. A lower threshold
allows for more spikes as is indicated in the increasing values
of the average spiking rates as the values for θj decrease.
Overall, using rate encoding produces the greatest values in
separation. However, there is significant variability as the
values for θj change. For rate encoding, the greatest accuracy
from the SVM is achieved with a low separation value. In the
other encoding schemes, separation and classification accuracy
appear to be correlated. The greatest classification accuracy is
achieved from current injection.
3) Liquid Topology: The topology of the liquid in an LSM
determines the size of the liquid and influences the connections
within the liquid as the distance between the neurons impacts
the connections made between neurons. A more cubic liquid
(e.g. 5×5×5) should be more densely connected compared to
a column of liquid (e.g. 2×2×20). In this section, we examine
using liquids with grids of 3 × 3 × 15, 2 × 2 × 20, 4 × 5 × 20,
and 5 × 5 × 5. As before, we consider different values for θj .
The separation values, average spike rates, and the accuracy
from a linear SVM are given in Table V for the values of θj
that provided the largest separation values for each topology
configuration and encoding scheme combination.
3x3x15/135
θj :
5x5x5/125
θj :
4x5x10/200
θj :
2x2x20/80
Current Injection
Encoding Scheme
Bit Encoded
Rate Encoded
15
12.5
3
2
11
10
0.409
0.580
0.905
0.405
0.693
0.884
0.353
0.634
0.764
0.357
0.620
0.761
0.622
0.563
0.658
0.622
0.594
0.643
12.5
11
3
2
10
9
0.467
0.534
0.889
0.432
0.682
0.900
0.384
0.596
0.765
0.380
0.586
0.755
0.384
0.473
0.637
0.451
0.600
0.610
15
12.5
15
12.5
15
12.5
0.418
0.698
0.879
0.384
0.760
0.876
0.506
0.576
0.830
0.501
0.603
0.835
0.479
0.289
0.644
0.352
0.577
0.612
11
10
3
2
10
9
0.397
0.608
0.903
0.389
0.658
0.912
0.264
0.537
0.672
0.247
0.550
0.666
0.305
0.609
0.695
0.255
0.846
0.678
Again, the value for θj has a significant impact on the
separation of the liquid and the classification accuracy. The
greatest separation values and classification accuracies for each
topology is highlighted in bold. For all of the topologies,
current injection achieves the highest classification accuracy.
Interestingly, the separation values across encoding schemes
and topologies do not correlate with accuracies. Within the
same encoding scheme and topology, however, the accuracy
generally improves as the separation increases.
For current injection, the different topologies do not appear
to have a significant impact on the classification accuracy
except for the 4 × 5 × 10 topology which has a decrease in
accuracy. This may be due to increased number of liquid nodes
that are used as input to the SVM. The converse is true for
bit encoding as the 4 × 5 × 10 topology achieves the highest
accuracy possibly due to the increased number of inputs due
to the bit representation of the input.
4) Readout Training Algorithms: How the plastic synapses
are trained will have a significant effect on the performance of
the LSM. Traditionally, LSMs use a linear classifier based on
the assumption that the liquid has transformed the state space
such that the problem is linearly separable. Linear models are
represented as a set of weights and a threshold—which can
be implemented in neuromorphic hardware. By using a linear
model, the liquid and the classification can all be done on
the STPU avoiding the overhead of going off chip to make a
prediction. We consider four linear classifiers: 1) linear SVM,
2) linear discriminant analysis (LDA), 3) ridge regression, and
4) logistic regression. With each of these algorithms, we use
the default parameters as they are set in the Statistics and
Machine Learning Toolbox in MATLAB.
We examine the classification of each of the linear classifiers
TABLE VI
C LASSIFICATION ACCURACY ON THE TEST SET FROM DIFFERENT LINEAR
CLASSIFIERS . T HE GREATEST ACCURACY FOR EACH TOPOLOGY IS IN
BOLD
Linear Model:
Linear SVM
LDA
Ridge Regress
Logistic Regress
3 × 3 × 15
θj = 15
5×5×5
θj = 11
4 × 5 × 10
θj = 15
2 × 2 × 20
θj = 10
0.906
0.921
0.745
0.431
0.900
0.922
0.717
0.254
0.900
0.922
0.717
0.254
0.914
0.946
0.897
0.815
on the topologies and θj values that achieved the highest classification accuracy on the linear SVM in previous experiments.
We also limit ourselves to examining current injection for the
input scheme as current injection consistently achieved the
highest classification accuracy. The results are shown in Table
VI. LDA consistently achieves the highest classification accuracy of the considered classifiers. The highest classification
accuracy achieved is 0.946.
VI. C ONCLUSION AND F UTURE W ORK
In this paper, we presented the Spiking Temporal Processing
Unit or STPU—a novel neuromorphic processing architecture.
It is well suited for efficiently implementing neural networks
and synaptic response functions of arbitrary complexity. This
is facilitated by using the temporal buffers associated with
each neuron in the architecture. The capabilities of the STPU,
including complex synaptic response functions, were demonstrated through implementing the functional mapping and
implementation of an LSM onto the STPU architecture.
As neural algorithms grow in scale and conventional processing units reach the limits of Moore’s law, neuromorphic
computing architectures, such as the STPU, allow efficient
implementations of neural algorithms. However, neuromorphic
hardware is based on spiking neural networks to achieve low
energy. Thus, more research is needed to understand and
develop spiking-based algorithms.
R EFERENCES
[1] L. Deng, G. E. Hinton, and B. Kingsbury, “New types of deep neural
network learning for speech recognition and related applications: an
overview,” in Proceedings of International Conference on Acoustics,
Speech, and Signal Processing (ICASSP), 2013, pp. 8599–8603.
[2] D. Ciresan, U. Meier, and J. Schmidhuber, “Multi-column deep neural
networks for image classification,” in Proceedings of the 2012 IEEE
Conference on Computer Vision and Pattern Recognition.
IEEE
Computer Society, 2012, pp. 3642–3649.
[3] R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Ng,
and C. Potts, “Recursive deep models for semantic compositionality
over a sentiment treebank,” in Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Processing. Association for
Computational Linguistics, October 2013, pp. 1631–1642.
[4] J. Backus, “Can programming be liberated from the von neumann style?:
A functional style and its algebra of programs,” Communications of the
ACM, vol. 21, no. 8, pp. 613–641, Aug. 1978.
[5] P. L. Follett, C. Roth, D. Follett, and O. Dammann, “White matter
damage impairs adaptive recovery more than cortical damage in an in
silico model of activity-dependent plasticity,” Journal of Child Neurology, vol. 24, no. 9, pp. 1205–1211, 2009.
[6] T. J. Sejnowski, “Time for a new neural code?” Nature, vol. 376, no.
6535, pp. 21–22, Jul. 1995.
[7] P. Merolla, J. Arthur, R. Alvarez-Icaza, A. Cassidy, J. Sawada,
F. Akopyan, B. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo,
I. Vo, S. Esser, R. Appuswamy, B. Taba, A. Amir, M. Flickner, W. Risk,
R. Manohar, and D. Modha, “A million spiking-neuron integrated circuit
with a scalable communication network and interface,” Science, pp. 668–
673, August 2014.
[8] S. B. Furber, D. R. Lester, L. A. Plana, J. D. Garside, E. Painkras,
S. Temple, and A. D. Brown, “Overview of the spinnaker system
architecture,” IEEE Transactions on Computers, vol. 62, no. 12, pp.
2454–2467, 2013.
[9] J. Schemmel, J. Fieres, and K. Meier, “Wafer-scale integration of analog
neural networks,” in IEEE International Joint Conference on Neural
Networks, june 2008, pp. 431 –438.
[10] Y. Zhang, P. Li, Y. Jin, and Y. Choe, “A digital liquid state machine with
biologically inspired learning and its application to speech recognition,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 26,
no. 11, pp. 2635–2649, 2015.
[11] W. Maass, T. Natschläger, and H. Markram, “Real-time computing
without stable states: A new framework for neural computation based
on perturbations,” Neural Computation, vol. 14, no. 11, pp. 2531–2560,
Nov. 2002.
[12] S. J. Verzi, C. M. Vineyard, E. D. Vugrin, M. Galiardi, C. D. James, and
J. B. Aimone, “Optimization-based computation with spiking neurons,”
in Proceedings of the IEEE International Joint Conference on Neural
Network, 2017, p. Accepted.
[13] S. J. Verzi, F. Rothganger, O. D. Parekh, T.-T. Quach, N. E. Miner, C. D.
James, and J. B. Aimone, “Computing with spikes: The advantage of
fine-grained timing,” submitted.
[14] P. Dayan and L. F. Abbott, Theoretical neuroscience : computational
and mathematical modeling of neural systems, ser. Computational neuroscience. Cambridge (Mass.), London: MIT Press, 2001.
[15] B. V. Benjamin, P. Gao, E. McQuinn, S. Choudhary, A. Chandrasekaran,
J.-M. Bussat, R. Alvarez-Icaza, J. Arthur, P. Merolla, and K. Boahen,
“Neurogrid: A mixed-analog-digital multichip system for large-scale
neural simulations,” Proceedings of the IEEE, vol. 102, no. 5, pp. 699–
716, 2014.
[16] J. Schemmel, D. Briiderle, A. Griibl, M. Hock, K. Meier, and S. Millner,
“A wafer-scale neuromorphic hardware system for large-scale neural
modeling,” in Proceedings of 2010 IEEE International Symposium on
Circuits and Systems, May 2010, pp. 1947–1950.
[17] S. Furber, F. Galluppi, S. Temple, and L. A. Plana, “The spinnaker
project,” Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014.
[18] W. Severa, K. D. Carlson, O. Parekh, C. M. Vineyard, and J. B. Aimone,
“Can we be formal in assessing the strengths and weaknesses of neural
architectures? A case study using a spiking cross-correlation algorithm,”
in NIPS Workshop Computing with Spikes, 2016.
[19] M. Lukoševičius and H. Jaeger, “Reservoir computing approaches to
recurrent neural network training,” Computer Science Review, vol. 3,
no. 3, pp. 127–149, August 2009.
[20] H. Jaeger, “Adaptive nonlinear system identification with echo state
networks,” in Advances in Neural Information Processing Systems 15.
MIT Press, 2003, pp. 609–616.
[21] G. B. Huang, Q. Y. Zhu, and C. K. Siew, “Extreme learning machine:
theory and applications,” Neurocomputing, vol. 70, no. 1-3, pp. 489–501,
2006.
[22] D. Norton and D. Ventura, “Improving liquid state machines through
iterative refinement of the reservoir,” Neurocomputing, vol. 73, no. 1618, pp. 2893–2904, 2010.
[23] H. Burgsteiner, M. Kröll, A. Leopold, and G. Steinbauer, “Movement
prediction from real-world images using a liquid state machine,” Applied
Intelligence, vol. 26, no. 2, pp. 99–109, 2007.
[24] D. Buonomano and W. Maass, “State-dependent computations: Spatiotemporal processing in cortical networks,” Nature Reviews Neuroscience, vol. 10, no. 2, pp. 113–125, 2009.
[25] S. Roy, A. Banerjee, and A. Basu, “Liquid state machine with dendritically enhanced readout for low-power, neuromorphic VLSI implementations,” IEEE Transactions on Biomedical Circuits and Systems, vol. 8,
no. 5, 2014.
[26] B. Schrauwen, M. D‘Haene, D. Verstraeten, and D. Stroobandt, “Compact hardware liquid state machines on fpga for real-time speech
recognition,” Neural Networks, no. 21, pp. 511–523, 1 2008.
[27] N. Hammami and M. Bedda, “Improved tree model for arabic speech
recognition,” in Proceddings of the IEEE International Conference on
Computer Science and Information Technology, 2010.
| 9cs.NE
|
The Effect of Phasor Measurement Units on the
Accuracy of the Network Estimated Variables
H. Abdollahzadeh Sangrody
M.T. Ameli
Power and Water University of Technology
Tehran, Iran
[email protected]
Power and Water University of Technology
Tehran, Iran
[email protected]
M.R.Meshkatoddini
Power and Water University of Technology
Tehran, Iran
[email protected]
Abstract— The most commonly used weighted least square state
estimator in power industry is nonlinear and formulated by using
conventional measurements such as line flow and injection
measurements. PMUs (Phasor Measurement Units) are gradually
adding them to improve the state estimation process. In this
paper the way of corporation the PMU data to the conventional
measurements and a linear formulation of the state estimation
using only PMU measured data are investigated. Six cases are
tested while gradually increasing the number of PMUs which are
added to the measurement set and the effect of PMUs on the
accuracy of variables are illustrated and compared by applying
them on IEEE 14, 30 test systems.
Keywords-conventional state estimation; hybrid
estimation; linear formulation; phasor measurement unit
state
I.
INTRODUCTION
State estimation is a key element of the online security
analysis function in modern power system energy control
centers. The function of state estimation is to process a set of
redundant measurements to obtain the best estimate of the
current state of a power system. State estimation is traditionally
solved by the weighted least square algorithm with
conventional measurements such as voltage magnitude, real
and reactive power injection, real and reactive power flow [1].
Recently, synchronized phasor measurement techniques based
on a time signal of the GPS (Global Positioning System) are
introduced in the field of power systems. A PMU, when placed
at a bus, can measure the voltage phasor at the bus, as well as
the current phasors through the lines incident to the bus. It
samples the ac voltage and current waveforms while
synchronizing the sampling instants with a GPS clock. The
computed values of voltage and current phasors are then timestamped and transmitted by the PMUs to the local or remote
receiver [2]-[4]. The traditional state estimation is by nature a
nonlinear problem. The most commonly used approach is
Weighted Least Squares which converts the nonlinear
equations into the normal equations by using first-order Taylor
series. However, the state estimation equations for PMU
measurements are inherently linear equations. Some research
has been conducted to try to formulate the mixed set of
traditional and PMU measurements. The natural approach is to
treat PMU measurements as additional measurements to be
appended to traditional measurements, which causes the
additional computation burden of calculation. Another
approach is to use the distributed scheme for the mixed state
estimation [5-9]. The problem of finding optimal PMU
locations for power system state estimation is well investigated
in the literature [10]–[13]. This paper shows the effect of
PMUs on the accuracy of the estimated variables. Six cases are
tested by gradually increasing the PMU numbers and applying
them on IEEE 14, 30 test systems. In the first case, state
estimation without any PMU and in the sixth case, linear
formulation of the state estimation using only PMU measured
data is discussed. In the other four cases, hybrid state
estimation with different number of added PMUs to the
conventional measurement set are tested.
II.
WEIGHTED LEAST SQUARED STATE ESTIMATION
METHOD
As shown “(1)”, this method minimizes the weighted sum
of squares of the residuals
Nm
J ( x)
i 1
( z i h i ( x )) 2
i2
z h( x )T R 1 z h( x )
Where in this equation z is measurement vector, x state
vector, standard deviation and h is the nonlinear function
relating measurement i to the state vector x. R is measurement
covariance matrix is given by “(2)”.
2
R diag 12 , 22 , , Nm
Pi
Pij
Q
H i
Q
ij
0
At the minimum value of the objective function, the firstorder optimality conditions have to be satisfied. These can be
expressed in compact form as follows:
g ( x)
J ( x )
H T ( x) R 1 ( z h( x)) 0
x
(3)
Where
H ( x)
h ( x )
x
(4)
The nonlinear function g (x) can be expanded into its
Taylor series around the state vector x k neglecting the higher
order terms. An iterative solution scheme known as the GaussNewton method is used to solve “(3)”:
x k 1 x k G ( x k )
1
.g ( x k )
(5)
Where, k is the iteration index and x k is the solution vector
at iteration k. G( x k ) is called the gain matrix, and expressed by:
G( x k )
g ( x k )
H T ( x k ) R 1 H ( x k )
x
g ( x k ) H T ( x k ) R 1 z h( x k )
(6)
Pi
V
Pij
V
Qi
V
Qij
V
VMag
V
In this matrix and V are state variables, Pi and Qi are
real and reactive power injection at bus i. Pij , Qij are real and
reactive power flow from bus i to bus j. In this condition of
estimation, relation between measurement data and state
variables are nonlinear and its final solution depend on an
iterative solution scheme expressed “(5)”.
IV. HYBRID STATE ESTIMATION
One PMU can measure the voltage and the current phasors.
The equivalent model of the line connecting buses i and j
with assumption of a PMU connected to bus i is shown in Fig.
1.
If Yij g ij jbij is defined as the series admittance and
as the shunt admittance, current phasor
measurements can be written in rectangular coordinates as
shown in Fig. 1. The expressions for Cij and Dij are:
Ysi g si jb si
(7)
These iterations are going on until the maximum variable
Cij Vi Ysi cos( i si )
system having (N) buses; the state vector will have (2N-1)
components which are composed of (N) bus voltage
magnitudes and (N-1) phase angles.
III. CONVENTIONAL STATE ESTIMATION
There are three most commonly used measurement types in
conventional state estimation. They are bus power injections,
line power flows and bus voltage magnitudes. These
measurement equations can be expressed using the state
variables. Jacobian matrix H has rows at each measurement
and columns at each variable. Considering “(4)”, power
injection and power flow in Fig. 1, H matrix components
corresponding to these measurements are partial derivation of
each variable [14].
V j Yij cos( j ij ) Vi Yij cos( j ij )
difference satisfies the condition, ' Max x k '. Consider a
Dij Vi Ysi sin( i si )
V j Yij sin( j ij ) ViYij sin( j ij )
Iij (Cij jDij )
Yij Yij ij
Vi Ei jFi
_
Ysi Ysi si
Figure 1. Transmission Line Model
V j E j jF j
_
The entries of the measurement Jacobian H corresponding
to the real and reactive parts of the current phasors are:
C ij
Vi
Ysi cos( i si ) Yij cos( i ij )
Cij
Yij cos( j ij )
Vi Ysi sin( i si ) Vi Yij sin( i ij )
V j
Cij
i
Cij
V j Yij sin( j ij )
j
measurement data can be expressed in rectangular coordinate
system. As shown in Fig. 1 a PMU located at bus i measured
voltage Vi and line current I ij . The voltage measurement
( Vi Vi i ) can be expressed as ( Vi Ei jFi ), and the
current measurement can be expressed as ( I ij Cij jDij ). In
this condition of estimation, measurement vector z and sate
vector x are:
T
T
z Ei Cij Fi Di j
x Ei E j Fi F j
In Fig. 1, line current flow I ij can be expressed as a linear
function of voltages.
Cij jDij ( g ij jbij ) ( g si jbsi ) ( Ei jFi )
Dij
Vi
( g ij jbij ) ( E j jF j )
Ysi sin( i si ) Yij sin( i ij )
i
Cij
Yij sin( j ij )
Vi Ysi cos( i si ) Vi Yij cos( i ij )
V j
Dij
Ei
Cij
Cij
j
V j Yij cos( j ij )
E j
The measurement vector z contains , Cij and Dij as well
as the power injections, power flows and voltage magnitude
measurements.
T
T
T
Z Pinj
, Qinj
, P Tflow , QTflow , V , T , CijT , DijT
Jacobian matrix H components are expressed by
Dij
LINEAR FORMULATION OF STATE ESTIMATION USING
ONLY PMUS
If the measurement set is composed of only voltage and
current measured by PMUs, the state estimation can be
formulated as a linear problem. The state vector and
Fi
Cij
F j
Dij
Generally, those measurements received from PMUs are
more accurate with small variances compared to the variances
of conventional measurements. Therefore, including PMU
measurements is expected to produce more accurate estimates.
Cij
Ei
Dij
E j
g ij g si
g ij
bij bsi
bij
bij bsi
bij
g ij g si
V.
Dij
Fi
Dij
F j
g ij
Then, the estimated value xˆ Eˆ i jFˆi can be obtained by
solving the linear equation below:
xˆ ( H T R 1 H ) 1 H T R 1 zˆ
This is very simple and fast, because it doesn't need any
iteration. In addition covariance matrix R in “(31)” is very
smaller than covariance matrix of conventional measurement,
so the estimated variables are very accurate.
VI. SIMULATION RESULTS
To investigate the effect of PMUs on the accuracy of
estimated variables, several cases are tested with different
number of added PMUs to the conventional measurement set.
Two different IEEE test systems (IEEE 14, IEEE 30 bus
system) are tested with 6 different cases which are shown in
table I. Fig. 2 and Fig. 3 show the network diagrams for each
system. Arrow with circle at the bus means a pair of real and
reactive power injection measurements and any point on the
transmission line; it means a pair of real and reactive power
flow measurements.
TABLE I.
Case 1
Case 2
Case 3
Case 4
Case 5
Case 6
SIX D IFFERENT CASES BY ADDING PMUS
Conventional Measurements with No PMUs
Conventional Measurements with PMUs of (10% of bus number)
Conventional Measurements with PMUs of (20% of bus number)
Conventional Measurements with PMUs of (30% of bus number)
Conventional Measurements with PMUs of (40% of bus number)
Only Minimum PMUs
Figure 3. IEEE30 Bus System Diagram with Conventional Measurements
Each network has a voltage magnitude measurement
connected to bus 1. Table II has more detailed information
about the conventional measurement set installed in the
networks.
The setting of error standard deviations for power injection,
power flow and voltage magnitude are 0.01, 0.08 and 0.04
respectively. A PMU has much smaller error deviations than
other conventional measurements as 0.00001. PMUs are
located at buses 2, 6, 7 and 9 in IEEE 14 bus system, and at
buses 3, 5, 6, 9, 10, 12, 19, 23, 25 and 29 in IEEE 30 bus
system at case 6 [10]. One of the ways of representing the level
of state estimation accuracy is to refer the covariance of the
estimated variables. The covariance of the estimated variable
vector is obtained from the inverse diagonal elements of gain
matrix. The accuracy of two variables (voltage magnitude,
voltage angle) is investigated separately. Fig. 4 and Fig. 5 show
the accuracy of the estimated voltage magnitudes of each
system. Fig. 6 and Fig .7 show the accuracy of the estimated
voltage angles of each system.
TABLE II.
VARIABLE NUMBERS, MEASUREMENT TYPE AND NUMBERS
Measurements
Figure 2. IEEE14 Bus System Diagram with Conventional Measurements
Network
IEEE 14 bus
system
IEEE 30 bus
system
Variables
Power
injection
Power
flow
Voltage
magnitude
Total
27
18
24
1
43
59
38
56
1
95
No PMU
10% PMU
20% PMU
30% PMU
40% PMU
No PMU
Only PMU
20% PMU
30% PMU
40% PMU
Only PMU
0.006
Standard Diviation
0.005
Standard Diviation
10% PMU
0.007
0.006
0.004
0.003
0.005
0.004
0.003
0.002
0.002
0.001
0.001
0
0
1
2
3
4
5
6
7
8
Bus Number
9
10
11
12
13
Figure 4. Accuracy of |V| of IEEE14 Bus System with PMUs
No PMU
10% PMU
20% PMU
30% PMU
40% PMU
Only PMU
0.008
0.007
0.006
Standard Diviation
1
14
0.005
0.004
3
5
7
9
11
13
15
17
19
21
23
25
27
29
Bus Number
Figure 7. Voltage Angle Accuracy of IEEE30 Bus System with PMUs
These figures show the effect of the PMUs on the accuracy
of the estimated variable. Average valued S.D. (Standard
Deviation) of each variable and their percentage values are
shown in tables III and IV. The percentage values in these
tables mean that how the S.D. values at each cases are
decreased compared to the S.D. of ‘Case 1’ which is forced to
be set as 100%. In case of ‘Only PMUs’, it becomes nearly
zero. Average valued S.D. of voltage magnitude and voltage
angle are shown in Fig. 8 and Fig. 9 respectively.
0.003
0.002
TABLE III.
0.001
cases
0
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
Bus Number
Figure 5. Accuracy of |V| of IEEE30 Bus System with PMUs
No PMU
10% PMU
20% PMU
30% PMU
40% PMU
Only PMU
0.0035
No PMUs
10% PMUs
20% PMUs
30% PMUs
40% PMUs
Only PMUs
AVERAGE ERROR STANDARD DEVIATIONS OF THE VOLTAGE
MAGNITUDE
IEEE 14 Bus
IEEE 30 Bus
Average
Error S.D.
Percentage
Average
Error S.D.
Percentage
0.0043741
0.0008449
0.0001727
0.0001497
0.0000810
0.0000055
100%
19.31%
3.95%
3.42%
1.85%
0.13%
0.0046548
0.0010444
0.0004986
0.0003240
0.0003229
0.0000042
100%
22.43%
10.70%
6.96%
6.94%
0.09%
Standar Diviation
0.003
0.0025
TABLE IV.
0.002
cases
0.0015
0.001
0.0005
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
bus number
Figure 6. Voltage Angle Accuracy of IEEE14 Bus System with PMUs
No PMUs
10% PMUs
20% PMUs
30% PMUs
40% PMUs
Only PMUs
AVERAGE ERROR STANDARD DEVIATIONS OF THE VOLTAGE
ANGLE
IEEE 14 Bus
IEEE 30 Bus
Average
Error S.D.
Percentage
Average
Error S.D.
Percentage
0.0023332
0.0008298
0.0001555
0.0001311
0.0000618
0.0000027
100%
35.56%
6.66%
5.62%
2.65%
0.11%
0.0030143
0.0010244
0.0004918
0.0003182
0.000317
0.0000032
100%
33.98%
16.3%
10.56%
10.52%
0.11%
Average valued standard diviation of voltage
angle
IEEE 14 bus sys tem
IEEE 30 bus sys tem
[3]
0.005
[4]
0.004
[5]
0.003
[6]
0.002
0.001
[7]
0
No PMUs
10% PMUs
20% PMUs
30% PMUs
40% PMUs
Only PMUs
[8]
Figure 8. Average Voltage Magnitude Standard Deviation of Two Systems
[9]
IEEE 14 bus system
IEEE 30 bus system
[10]
Average valued standard diviation of voltage
magnitude
0.0035
[11]
0.003
0.0025
[12]
0.002
0.0015
0.001
[13]
0.0005
0
No PMUs
[14]
10% PMUs
20% PMUs
30% PMUs
40% PMUs
Only PMUs
A. G. Phadke and J. S. Thorp, Synchronized Phasor Measurements and
Their Applications, Springer, 2008, pp. 93-97.
R. Zivanovic and C. Cairns, “Implementation of PMU technology in
state estimation: an overview”, IEEE AFRICON, pp. 1006-1011, 2006.
Y.M. El-Fattah and M. Ribbens-Pavella, “Multi-level approach to state
estimation in electric power systems, ” Proc. of the IV IFAC Symposium
on Identification and System Parameter Estimation, Tbilisi, USSR, pp.
166-179, Sept. 1976.
K. A. Clements, O. J. Denison, and R. J. Ringlee, “A multi-area
approach to state estimation in power system networks,” in roc. IEEE
Power Eng. Society Summer Meeting, San Francisco, CA, 1972,
C72465-3.
T. Van Cutsem, J. L. Horward, and M. Ribbens-Pavella, “A two-level
static state estimator for electric power systems,” IEEE Trans. Power
Apparat. Syst., vol. PAS-100, no. 8, pp. 3722–3732, Aug. 1981.
L. Zhao and A. Abur, “Multiarea state estimation using synchronized
phasor measurements,” IEEE Trans. Power Syst., vol. 20, no. 2, pp.
611–617, May 2005.
Weiqing Jiang; Vittal, V.; Heydt, G.T, “A Distributed state estimator
utilizing synchronized phasor measurements,” IEEE Tran. Power Syst.,
vol. 22, pp. 563 – 571, May 2007.
B. Xu and A. Abur, “Optimal placement of phasor measurement units
for state estimation, ” PSERC, Oct. 2005, Fin. Proj. Rep.
B. Xu and A. Abur, “Observability analysis and measurement placement
for systems with PMUs,” in Proc. IEEE PES Power Systems Conf.
Expo., Oct. 2004, pp. 943–946.
C. Rakpenthai, S. Premrudeepreechacharn, S. Uatrongjit, and N. R.
Watson, “An optimal PMU placement method against measurement loss
and branch outage,” IEEE Trans. Power Del., vol. 22, no. 1, pp.101–107,
Jan. 2005.
S. Chakrabarti and E. Kyriakides, “Optimal placement of phasor
measurement units for power system observability,” IEEE Trans. Power
Syst., vol. 23, no. 3, pp. 1433–1440, Aug. 2008.
Y.J. Yoon, “Study of the utilization and benefits of phasor measurement
units for large scale power system state estimation”, Master of Science
Thesis.Texas Univ. December 2005.
Figure 9. Average Voltage Angle Standard Deviation of Two Systems
H. Abdollahzadeh Sangrody
VII. CONCULTION
In this paper the way of incorporating the PMU data to the
conventional measurements set is discussed. It is expected that
those PMU measured data improve the measurement
redundancy and accuracy, due to the small error standard
deviations of PMU. A linear formulation of the state estimation
is investigated using only PMU measured data. This linear
formulation of the PMU data can produce the estimation result
by a single calculation not requiring any iteration. Six cases are
tested while gradually increasing the number of PMUs which
are added to the measurement set by applying them on IEEE
14, 30 test systems. With the help of advanced accuracy of
PMU, it was seen that the estimated accuracy is also gradually
increase. One of the interesting thing is that the accuracy of
estimated variables improves most effectively when the
number of implemented PMUs are around ‘10%’ of the system
buses.
REFERENCES
[1]
[2]
A. Abur and A. G. Exposito, Power System State Estimation, Theory
and Implementation, MAECEL DEKKER,2005, pp. 9-27.
Real time dynamics monitoring system [Online]. Available: http://
www.phasor-rtdms.com.
received the
B.Sc. degree in Electrical Engineering from Zanjan
University, Iran 2006. He is currently pursuing Master
of Science in Electrical Engineering at Power and
Water University of Technology. His research interests
include Power System Observability and State
Estimation in Power Systems.
M.T. Ameli received B.Sc. degree in Electrical
Engineering from Technical College of Osnabrueck,
Germany in 1988 and M.Sc. & Ph.D. from Technical
University of Berlin in 1992 & 1997. Since then He
teaches and researches as associated Professor in
Electrical Engineering Dept of Power & Water
University of Technology in Teheran. Areas of research:
Power system Simulation, Operation, Planning &
Control of power system, usage of renewable Energy in Power system.
M.R. Meshkatoddini(SM'96) received his B.Sc. and
M.Sc. degrees, in 1980 and 1990 respectively, in
Electrical Engineering from Tehran Polytechnic
University, Iran, and the Ph.D. degree in Electrical
Engineering, from Paul Sabatier University of Toulouse,
France in 1996. Since 1992, he has been a faculty
member at PWUT. Dr. Meshkatoddini has more than 25
years of experience in the areas of power transformers, surge arresters, electric
materials and power network transients.
| 3cs.SY
|
Does quantification without adjustments work?
Dirk Tasche∗
arXiv:1602.08780v2 [stat.ML] 12 Aug 2016
August 12, 2016
Classification is the task of predicting the class labels of objects based on the observation
of their features. In contrast, quantification has been defined as the task of determining the
prevalences of the different sorts of class labels in a target dataset. The simplest approach
to quantification is Classify & Count where a classifier is optimised for classification on a
training set and applied to the target dataset for the prediction of class labels. In the case of
binary quantification, the number of predicted positive labels is then used as an estimate of
the prevalence of the positive class in the target dataset. Since the performance of Classify &
Count for quantification is known to be inferior its results typically are subject to adjustments.
However, some researchers recently have suggested that Classify & Count might actually work
without adjustments if it is based on a classifier that was specifically trained for quantification.
We discuss the theoretical foundation for this claim and explore its potential and limitations
with a numerical example based on the binormal model with equal variances. In order to
identify an optimal quantifier in the binormal setting, we introduce the concept of local Bayes
optimality. As a side remark, we present a complete proof of a theorem by Ye et al. (2012).
Keywords: Classification, quantification, confusion matrix method, Bayes error.
1. Introduction
The formal definition of quantification as a machine learning task is often credited to Forman (2008) who
wrote: “The quantification task for machine learning: given a limited training set with class labels, induce
a quantifier that takes an unlabeled test set as input and returns its best estimate of the number of cases
in each class. In other words, the quantification task is to accurately estimate the test class distribution
via machine learning, but without assuming a large training set that is sampled at random from the test
distribution. The input to a quantifier is a batch of cases, whereas a traditional classifier takes a single
case at a time and predicts its single class (or a distribution of classes reflecting its uncertainty about
that one case).”
At least since the 1960s (Gart and Buck, 1966), researchers and practitioners were aware of the need
to track changes of the prior probabilities (or prevalences) of classes between different datasets. In the
machine learning community, the topic received renewed attention after Saerens et al. (2002) suggested
a powerful alternative to the ‘confusion matrix method’ they considered the standard approach at the
time.
∗ E-mail:
[email protected]
The author currently works at the Swiss Financial Market Supervisory Authority (FINMA). Most of the research for
this paper was done while he was an employee of the Prudential Regulation Authority (a directorate of the Bank of
England) and on secondment to the Bank’s research hub. The opinions expressed in this paper are those of the author
and do not necessarily reflect views of FINMA or the Bank of England.
1
Fawcett and Flach (2005) marked another milestone in the discussion of how to deal with changed
prevalences when they noticed that as a consequence of different causalities there are different dataset
shift regimes that need to be tackled in different ways. Since then a number of papers has been published
with proposals of how to categorise different types of dataset shift (Storkey, 2009; Moreno-Torres et al.,
2012; Kull and Flach, 2014).
There are two types of dataset shift between training and target dataset that can be easily characterised.
Under ‘covariate shift’, the assumption is that the posterior (i.e. conditional) class probabilities are the
same on the training and the target datasets. However, the distribution of the covariates (or features)
may change. This change should be taken into account already when a classifier is learnt (Shimodaira,
2000; Sugiyama et al., 2007; Bickel et al., 2009). The other easily characterised dataset shift type is
‘prior probability shift’ where the feature distributions conditional on the classes remain the same when
datasets are switched. Typically, in this case, adjustments to the posterior class probabilities and decision
thresholds are recommended in the literature (Elkan, 2001; Forman, 2008; Xue and Weiss, 2009; Hopkins
and King, 2010; Bella et al., 2010). Another approach is the direct estimation of the changed prior
probabilities by minimising the distance of the feature distributions on the training and target datasets
(Saerens et al., 2002; Forman, 2008; González-Castro et al., 2013; Hofer and Krempl, 2013; Du Plessis
and Sugiyama, 2014; Kawakubo et al., 2016).
Other types of dataset shift are less easy to describe and to deal with. Tasche (2014) defined the ‘invariant
density ratio’ dataset shift which generalises the prior probability shift in such a way that only ratios
of feature densities but not the densities themselves are unchanged. A recent paper by Hofer (2015) is
outstanding by dealing with dataset shift under very weak assumptions on the structure of the shift.
Esuli et al. (2010) suggested that specially trained classifiers (called quantifiers) can be used for the socalled ‘Classify & Count’ quantification (Forman, 2008) without a need of adjustments to the estimates.
Classify & Count means that a classifier is optimised for classification on a training set and applied to
the target dataset for the prediction of class labels. The number of predicted positive labels (in the case
of binary quantification) is then used as an estimate of the prevalence of the positive class in the target
dataset. The practical implementation of this proposal for binary classification and quantification has
been explored in papers by Milli et al. (2013), Barranquero et al. (2015) and Esuli and Sebastiani (2015).
In all these three papers, emphasis is put on the need to have the quantifier properly calibrated on the
training dataset, in the sense that the number of objects predicted to be positive should be equal to the
true number of positive objects.
The experiments by Milli et al. (2013) suggest that after all Classify & Count with adjustments works
better than pure Classify & Count. In contrast, Barranquero et al. (2015) and Esuli and Sebastiani
(2015) report Classify & Count quantification performance of the specially trained quantifiers that is
at least comparable to the performance of classifiers with adjustments. Esuli and Sebastiani (2015) are
somewhat ambiguous with respect to clearly stating all of their optimisation criteria. This is different to
Barranquero et al. (2015) whose authors clearly say that the calibration criterion must be supplemented
by a condition enforcing good classification. This is an intuitive requirement because, in particular, the
complementarity of the two criteria of calibration and classification seems to guarantee uniqueness of the
optimal quantifier.
In this paper, we discuss the theoretical foundations and limitations of quantification without adjustments
and illustrate the insights with the classical example of the binormal model with equal variances (van
Trees, 1968). We focus on analysis of the ‘Q-measure’ approach by Barranquero et al. (2015) because it
has been well and intuitively documented. Our first finding is that quantification without adjustments
in principle may work if proper calibration on the training dataset is ensured and the positive class
prevalences on the training and target datasets are the same or nearly the same. However, a second finding
is that the Q-measure approach has to be deployed with care because the quantifiers resulting from it
may be miscalibrated. All in all, these findings suggest that the potential applications of quantification
2
without adjustments are rather limited.
This paper is organised as follows: In Section 2 and its subsections, the theory of binary classification
is revisited such that the theoretically best quantifiers can be identified for some specific optimisation
criteria. In particular, we introduce the concept of ‘local Bayes optimality’ and discuss its applications
to minimax tests of two simple hypotheses and the optimisation of F-measure in the context of binary
classification. As another application of the results from Section 2, in Section 3 the experiment of Barranquero et al. (2015) is revisited and reviewed in the fully controlled setting of the binormal model with
equal variances. Section 4 concludes the paper. As a side remark, we present in Appendix A a complete
proof of Theorem 4 of Ye et al. (2012) who published only the key argument of the proof but did not
mention other important steps.
2. Classifying for quantification
In order to be able to appropriately assess the merits and limitations of the proposal by Barranquero
et al. (2015) we adopt a precise mathematical formalism. Based on this formalism, we can characterise
‘locally optimal binary classifiers’ which are closely related to both optimal Bayes classifiers and optimal
Neyman-Pearson classifiers (van Trees, 1968, Section 2.2). The concept of local optimality allows us
to characterise minimax tests of two simple hypotheses in a way alternative to that by Scharf (1991,
Chapter 5) and to provide an alternative proof of Theorem 4 of Ye et al. (2012) on F-measure-optimal
classifiers. In Section 3, we use results of this section to explore the Barranquero et al. proposal in detail
by inspecting its implementation for the binormal model with equal variances.
2.1. Locally optimal binary classifiers
We discuss binary classification and the properties of classifiers in a probabilistic setting specified by
a probability space as it was done by many authors before (see, e.g. van Trees, 1968). The probability
space (Ω, A, P) describes the experiment of choosing an object at random1 . The object has a class label
and features. The features can be observed immediately while, depending on whether the probability
space is interpreted as a training sample or target sample (sometimes also called test sample), the label
is also observable at once or can be observed only with some delay. We interpret A as the σ-field (see,
e.g. Billingsley, 1995, Section 2) of all admissible events, including events that cannot yet be observed.
In addition, we have a σ-field H which is the family of the events that can be observed now. The event
A with A ∈ A but A ∈
/ H reveals the object’s class label. If A occurs the object has got class label 1
(positive). If Ac = Ω\A occurs the object’s label is −1 (negative).
Assumption 2.1
• (Ω, A, P) is a probability space. This space describes the experiment of selecting an object from a
population at random and observing its features and (typically with some delay) class label.
• A ∈ A is a fixed event with 0 < P[A] < 1. If A is observed, the object’s class label is 1, otherwise if
Ac = Ω\A is observed, the object’s class label is -1.
• H ⊂ A is a sub-σ-field of A such that A ∈
/ H. H is the σ-field of immediately observable events
and, in particular, features.
In a binary classification problem setting, typically there are random variables X : Ω → Rd for some
d ∈ N (vector of explanatory variables or scores) and Y : Ω → {−1, 1} (dependent or class variable) such
1 The
case of Ω being finite and P being the uniform distribution on Ω provides the sample-based setting which often is
assumed for machine learning papers.
3
that H = σ(X) and Y −1 ({1}) = A.
In the setting of Assumption 2.1, typically one wants to predict an object’s class label (i.e. predict whether
or not event A has occurred) based on the observable information captured in the events H ∈ H. Each
of these events H defines a binary ‘classifier’ in the following sense:
• If H occurs the object’s label is predicted as 1.
• If H c = Ω\H occurs the object’s label is predicted as -1.
This way binary classifiers are identified with elements of H. We therefore do not introduce extra notation
for classifiers. Note that the object’s class A does not define a classifier because by assumption we have
A∈
/ H.
Define the expected ‘misclassification cost’ La,b (H) for H ∈ H and fixed a, b ≥ 0 with a + b > 0 by
La,b (H) = a P[H c ∩ A] + b P[H ∩ Ac ].
(2.1)
According to (2.1), there is no cost for misclassification if the label is correctly predicted (i.e. if the event
(A ∩ H) ∪ (Ac ∩ H c ) occurs). If 1 is predicted for true label −1 (event H ∩ Ac , false positive), the cost
is b. If −1 is predicted for true label 1 (event H c ∩ A, false negative), the cost is a. The misclassification
cost La,b (H) is the expected cost for the classifier represented by event H.
From Section 2.2 of van Trees (1968) or Section 1.3 of Elkan (2001), we know the optimal choice H ∗
(‘Bayes classifier’) of H for minimising La,b (H):
def
H∗ =
P[A | H] >
b
a+b
=
ω ∈ Ω : P[A | H](ω) >
b
a+b
= arg min La,b (H),
H∈H
(2.2)
where P[A | H] denotes the conditional probability (or posterior probability) of A given H as defined in
standard text books on probability theory (e.g. Billingsley, 1995, Section 33). The following proposition
shows that in some sense all classifiers of the shape
def
Hq = {P[A | H] > q},
0 < q < 1,
(2.3)
are local minimisers of La,b (H).
Proposition 2.2 Under Assumption 2.1, let a, b ≥ 0 with a + b > 0 be fixed. Define La,b (H) and Hq by
(2.1) and (2.3) respectively. Let pq = P[Hq ]. Then the following two statements hold:
(i) q <
(ii) q >
b
a+b
⇒
b
a+b
⇒
Hq = arg
Hq = arg
min
H∈H, P[H]≥pq
min
H∈H, P[H]≤pq
La,b (H).
La,b (H).
Proof. Let H ∈ H be given. With some algebra2 , it can be shown that
La,b (H) = a P[A] + b − (a + b) q P[H] + (a + b) E (q − P[A | H]) 1H∩Hq
+ (a + b) E (q − P[A | H]) 1H∩Hqc
≥ a P[A] + b − (a + b) q P[H] + (a + b) E (q − P[A | H]) 1H∩Hq
≥ a P[A] + b − (a + b) q P[H] + (a + b) E (q − P[A | H]) 1Hq .
(2.4)
b
we have b − (a + b) q > 0. Then it holds that b − (a + b) q P[H] ≥ b − (a + b) q pq for
In case q < a+b
P[H] ≥ pq . By (2.4), this implies (i).
b
In case q > a+b
we have b − (a + b) q < 0. Then it holds that b − (a + b) q P[H] ≥ b − (a + b) q pq for
P[H] ≤ pq . From this observation and (2.4), statement (ii) follows.
2
21
S
denotes the indicator function of the set S, i.e. 1S (s) = 1 for s ∈ S and 1S (s) = 0 for s ∈
/ S.
4
Proposition 2.2 is about ’locally’ optimal classifiers in the sense that only classifiers with identical probability of predicting 1 are compared. We state this observation more precisely in item (i) of the following
remark:
Remark 2.3
b
(i) We have H ∗ = H b for H ∗ as defined in (2.2). Hence the case q = a+b
is not treated in Propoa+b
sition 2.2 because it is covered by (2.2). Nonetheless, it is worth noting that Proposition 2.2 and
(2.2) together imply that for all a, b ≥ 0 with a + b > 0 and 0 < q < 1 it holds that
Hq = arg
min
H∈H, P[H]=pq
La,b (H).
(ii) In the case a = 0, b = (1 − P[A])−1 , Proposition 2.2 (i) implies for all 0 < q < 1 that
Hq = arg
P[H ∩ Ac ]
= arg
min
P[H | Ac ].
P[Ac ]
H∈H, P[H]≥pq
H∈H, P[H]≥pq
min
P[H | Ac ] is called ‘false positive rate’ (FPR).
(iii) In the case a = P[A]−1 and b = 0, Proposition 2.2 (ii) implies for all 0 < q < 1 that
Hq = arg
max
H∈H, P[H]≤pq
P[H ∩ A]
= arg
max
P[H | A].
P[A]
H∈H, P[H]≤pq
P[H | A] is called ‘true positive rate’ (TPR).
As mentioned above, Proposition 2.2 may be interpreted as a result in between the characterisation of
optimal Bayes classifiers and the optimal classifier (test) from the Neyman-Pearson lemma. The following
theorem gives a precise statement of this observation.
Theorem 2.4 Let (Ω0 , M, µ) be a measure space. Assume that Q− and Q+ are probability measures on
(Ω0 , M) which are absolutely continuous with respect to µ. Assume furthermore that the densities f −
+
and f + of Q− and Q+ respectively are positive. Define the likelihood ratio λ by λ = ff − > 0. If the
distribution of λ is continuous under Q− and Q+ , i.e. Q− [λ = `] = 0 = Q+ [λ = `] for all ` > 0, then
there is a number `∗ > 0 with Q− [λ > `∗ ] = Q+ [λ ≤ `∗ ] such that
min max Q− [M ], Q+ [M c ] = Q− [λ > `∗ ].
(2.5)
M ∈M
Proof. Define the probability space (Ω, A, P) by
• Ω = Ω0 × {−1, 1} with projections X(ω, c) = ω and Y (ω, c) = c for (ω, c) ∈ Ω,
• A = M ⊗ P({−1, 1}) = σ(X, Y ),
• P[Y = −1] = 1/2 = P[Y = 1] as well as P[X ∈ M | Y = −1] = Q− [M ] and P[X ∈ M | Y = 1] =
Q+ [M ] for M ∈ M.
Then Assumption 2.1 is satisfied if A is chosen as A = {Y = 1} and H is chosen as H = σ(X) =
M × ∅, {−1, 1} . By construction of P, it follows that the probability of A = {Y = 1} conditional on H
is given by
f+ ◦ X
.
P Y = 1|H = +
f ◦ X + f− ◦ X
This implies for any 0 < q < 1 that
Q− [λ >
P P[Y = 1 | H] > q =
5
q
1−q ]
+ Q+ [λ >
2
q
1−q ]
.
(2.6)
Fix M ∈ M and let H = {X ∈ M }. The assumption on the continuity of the distribution of λ under Q−
and Q+ implies that
min max Q− [M ], Q+ [M c ] < 1.
M ∈M
For this proof, we therefore may assume without loss of generality that max Q− [M ], Q+ [M c ] < 1 and
hence 0 < P[H] < 1. Again by the assumption on the continuity of the distribution of λ under Q− and
Q+ , then (2.6) implies that there is q = q(H) such that P[H] = P P[Y = 1 | H] > q . From Remark 2.3
(ii) and (iii) now it follows that
max Q− [M ], Q+ [M c ] = max P[H | Y = −1], P[H c | Y = 1]
≥ max P P[Y = 1 | H] > q | Y = −1 , P {P[Y = 1 | H] > q}c | Y = 1
+
q
q
= max Q− λ > 1−q
, Q λ ≤ 1−q
.
Since this holds for all M ∈ M with max Q− [M ], Q+ [M c ] < 1, we can conclude that
min max Q− [M ], Q+ [M c ]
M ∈M
≥ min max Q− λ > ` , Q+ λ ≤ ` .
`>0
The intermediate value theorem implies that there is an `∗ such that Q− [λ > `∗ ] = Q+ [λ ≤ `∗ ]. Since
` 7→ Q− λ > ` is non-increasing and ` 7→ Q+ λ ≤ ` is non-decreasing, it follows that
min max Q− λ > ` , Q+ λ ≤ ` = Q− [λ > `∗ ].
2
`>0
Remark 2.5 One interpretation of Theorem 2.4 is as providing a minimax test for the decision between
two simple hypotheses (see Chapter 5 of Scharf, 1991). The test problem is to distinguish Q− and Q+ .
Tests are characterised by observable sets M ∈ M where ω ∈ M means ’accept Q+ ’ and ω ∈
/ M means
’reject Q+ in favour of Q− ’. However, in contrast to the setting of the Neyman-Pearson lemma, none of
the two hypotheses is considered more important than the other. Therefore, as expressed on the left-hand
side of (2.5), an optimal test is meant to minimise the probabilities of the type I and II errors at the
same time.
(2.5) shows that in a ’continuous’ setting there is an optimal test under the criterion given by the left-hand
side of (2.5) which is based on the likelihood ratio λ, i.e. the ratio of the densities of the tested probability
measures. Hence the structure of the optimal test is the same as for the optimal Neyman-Pearson test,
the cost-optimal Bayes test and the minimax optimal Bayes test (see Section 2.2 of van Trees, 1968, or
Chapter 5 of Scharf, 1991).
The concept of local Bayes optimality (i.e. Proposition 2.2) can also be applied to the question of how
to determine binary classifiers that are optimal with respect to the ‘F-measure’ criterion. Fβ -measure for
β > 0 was introducted by van Rijsbergen (1974) in order to avoid neglecting the minority class when
learning binary classifiers. For a given classifier, its Fβ -measure is defined as
Fβ =
1 + β2
β2
Recall
+
1
Precision
,
where the classifier’s ‘precision’ is the ratio of the number of positively predicted true positive objects
and the number of positively predicted objects while its ‘recall’ is the ratio of the number of positively
predicted true positive objects and the number of true positive objects.
In the setting of Assumption 2.1, we have
Precision = P[A | H]
and
6
Recall = P[H | A],
where A denotes ‘positive class’ and H stands for ‘classifier predicts positive’. Note that recall is identical
with ‘true positive rate’ as defined in Remark 2.3. We can rewrite the definition of Fβ as
β2 + 1
Fβ (H) =
β2
P[H | A]
+
1
P[A | H]
=
(1 + β 2 ) P[H | A]
,
β 2 + P[H]/P[A]
(2.7)
for H ∈ H and fixed β > 0. Ye et al. (2012, Theorem 4) observed that classifiers that are optimal in the
sense of maximising Fβ can be constructed by ‘thresholding’ the conditional class probability P[A | H].
In the notation of this paper, their observation can be precisely stated as follows:
sup Fβ (H) =
H∈H
sup max Fβ ({P[A | H] > q}), Fβ ({P[A | H] ≥ q}) .
(2.8)
0≤q≤1
Ye et al. (2012) published only the most important part of the proof3 of (2.8). In Appendix A of this
paper, we provide a complete proof of (2.8).
However, in the case where the conditional class probability P[A | H] has a continuous distribution, (2.8)
is an immediate consequence of Remark 2.3 (iii):
• By continuity of the distribution of P[A | H], for any classifier H ∈ H with 0 < P[H] < 1 there is a
number 0 < q < 1 such that P[H] = P[Hq ] with Hq defined as in (2.3).
• Therefore, Remark 2.3 (iii) implies that
Fβ (H) ≤
(1 + β 2 ) P[Hq | A]
(1 + β 2 ) P[Hq | A]
=
= Fβ (Hq ).
2
β + P[H]/P[A]
β 2 + P[Hq ]/P[A]
2.2. Application to quantification under prior probability shift
In contrast to Esuli and Sebastiani (2015) and Milli et al. (2013), Barranquero et al. (2015) specify the
dataset shift problem they are going to tackle with their proposal: It is prior probability shift. We modify
Assumption 2.1 accordingly.
Assumption 2.6 We extend the setting of Assumption 2.1 by assuming that there is a second probability
measure P1 on (Ω, A). P1 evolves from P by ‘prior probability shift’, i.e. the probabilities of sets H ∈ H
conditional on A and the probabilities of sets H ∈ H conditional on Ac are the same under P and P1 :
P[H | A] = P1 [H | A]
and
P[H | Ac ] = P1 [H | Ac ],
for all H ∈ H.
Under Assumption 2.6, we can describe for any classifier H ∈ H the probability P1 [H] as an affine
function of w = P1 [A]:
P1 [H] = w P[H | A] − P[H | Ac ] + P[H | Ac ].
(2.9)
In practice, the true positive rate (TPR) P[H | A] and the false positive rate (FPR) P[H | Ac ] can be
estimated (possibly with large potential bias) from the training set (in our setting: (Ω, A, P)) and the
probability P1 [H] of an object to be classified as positive (after prior probability shift) can be estimated
from the target set (in our setting: (Ω, A, P1 )). Then (2.9) can be solved for w = P1 [A] to obtain an
estimate of the new prior probability of the positive class:
P1 [A] =
P1 [H] − P[H | Ac ]
.
P[H | A] − P[H | Ac ]
3 The
proof is available in an appendix to the paper of Ye et al. which can be downloaded at
https://www.comp.nus.edu.sg/∼leews/publications/fscore-appendix.pdf
7
(2.10)
This approach is called ‘confusion matrix method’ (Saerens et al., 2002). It has also been described as
‘Adjusted Count’ approach (Forman, 2008) and has been deployed by practitioners at least since the
1960s (Gart and Buck, 1966).
In theory, for any classifier H, (2.10) provides the adjustment needed to obtain an accurate estimate of
the probability of the positive class from a potentially quite inaccurate estimate by P1 [H]. Experiments
by some research teams, however, have cast doubt on the appropriateness of this approach. Both good
(Xue and Weiss, 2009; Hopkins and King, 2010) and unsatisfactory performance (Saerens et al., 2002;
Forman, 2008) of the confusion matrix method have been reported. Other papers report mixed findings
(Bella et al., 2010; González-Castro et al., 2013; Hofer and Krempl, 2013; Du Plessis and Sugiyama,
2014).
As (2.10) is valid only under prior probability shift (Assumption 2.6), performance issues with the confusion matrix method should not be a surprise in circumstances when there is little evidence of prior
probability shift against other types of dataset shift (see Moreno-Torres et al., 2012, for a taxonomy of
dataset shift types). But most if not all of the above-mentioned reports on the performance of the confusion matrix method refer to controlled environments with prior probability shift. A number of reasons
have been identified to potentially negatively impact the confusion matrix method performance. Among
them are class imbalance in the training set (Forman, 2008) and issues with the accurate estimation of
TPR and FPR on the training set (Esuli and Sebastiani, 2015).
Although many other approaches to prior probability estimation have been proposed (Hofer and Krempl,
2013; Du Plessis and Sugiyama, 2014; Hofer, 2015; Kawakubo et al., 2016), no gold standard has yet
emerged because all approaches appear to suffer from numerical problems to some extent.
This observation has led some authors to suggest that so-called ‘quantifiers’ (classifiers, specifically developed for quantification) might be a viable solution (Esuli et al., 2010; Milli et al., 2013; Barranquero
et al., 2015; Esuli and Sebastiani, 2015). In the notation of this paper, both classifiers and quantifiers
are characterised by observable events H ∈ H which are interpreted as ’predict positive’. The difference
between the concepts of ’classifier’ and ’quantifier’ is the intended use, as explained in the quotation from
Forman (2008) in Section 1:
• Classifiers are deployed for predicting the class labels of single objects. Therefore, development of
a classifier typically involves minimising the expected loss of decisions about single objects (see, for
instance, the right-hand-side of (2.2)).
• Quantifiers are deployed for estimating the prevalence of a class in a sample or population. Barranquero et al. (2015) have argued that this different purpose should be reflected in a different
objective function for the development of a quantifier. They suggest that with an appropriate objective function, no adjustment like (2.10) would be needed.
• In their paper, Barranquero et al. (2015) suggest maximising a Q-measure criterion (see (2.14a)
below). Their experiments were conducted in a prior probability shift setting (see Assumption 2.6).
Similarly, Milli et al. (2013) report experimental results from a prior probability shift setting.
• As a different approach, Esuli et al. (2010) suggest minimising the Kullback-Leibler distance between
the observed class distribution and the predicted class distribution on the training set. Implicitly,
as they work with Support Vector Machines (Esuli and Sebastiani, 2015), they also apply some
classification optimisation criterion. Esuli and Sebastiani only use ’natural’ datasets such that their
datashift environment cannot be characterised as prior probability shift.
In the following, we focus on the analysis of the approach proposed by Barranquero et al. (2015) to deal
with prior probability shift. Analysis of the approach followed by Esuli and Sebastiani (2015) is harder
because Esuli and Sebastiani do not specify their dataset shift assumption and because the performance
of their approach seems to depend on their choice of the classifier development methodology (as support
8
0.25
Figure 1: Illustration of prediction error function (2.11).
0.15
0.10
0.00
0.05
Prediction error
0.20
Uniformly best
Locally best
Qbeta best
0.0
0.2
0.4
0.6
0.8
1.0
Test set prevalence of positive class
vector machines). Analysis and potential criticism of Esuli and Sebastiani (2015), therefore, is not undertaken in this paper. The results of Milli et al. (2013) are less controversial than those of Barranquero et al.
(2015) and Esuli and Sebastiani (2015) because Milli et al. report superior quantification performance
for adjusted classifiers.
For a fixed, unadjusted classifier specified by a set H ∈ H, the absolute prediction error |P1 [A] − P1 [H]|
of the prevalence of the positive class in the target dataset is the combination of a decreasing and
an increasing straight line if it is represented as a function of the true positive class prior probability
w = P1 [A]:
(
| Ac ]
w P[H | A] − P[H | Ac ] − 1 + P[H | Ac ], for w ≤ P[H | AP[H
c ]+1−P[H | A] ,
(2.11)
EH (w) =
w 1 − P[H | A] + P[H | Ac ] − P[H | Ac ], otherwise.
See Figure 1 for an illustration of the absolute prediction error concept. The rationale for the three
different curves is explained in Section 3 below. For the moment, if we ignore the question of how to
minimise the absolute prediction error, the figure tells us that every classifier is a perfect predictor of one
positive class prevalence in the target dataset. Unfortunately, it is not very helpful to know this because
it is perfection in the way a broken clock is perfectly right once a day. Hence it is worthwhile to try and
find out more about minimising the error as we do in the following.
9
(2.11) immediately implies the following result on error bounds for the prediction of the positive class
prevalence by a classifier:
Proposition 2.7 Under Assumption 2.6, the following inequality holds for any H ∈ H:
P1 [A] − P1 [H] ≤ max P[H | Ac ], P[H c | A] .
(2.12)
Proposition 2.7 shows that a classifier’s prediction error with regard to the positive class prevalence
is controlled by the classifier’s false positive rate (FPR) P[H | Ac ] and its false negative rate (FNR)
P[H c | A]. Theorem 2.4 (or Remark 2.3 (ii) and (iii)) makes it possible to identify the optimal (in this
case the minimax) classifier with regard to the prediction of the positive class prevalence:
Corollary 2.8 Under Assumption 2.6, define the classifiers Hq , 0 < q < 1 by (2.3). Assume in addition,
that the distribution of P[A | H] is continuous under both P[· | A] and P[· | Ac ]. Then there is a number
0 < q ∗ < 1 with P[Hq∗ | Ac ] = 1 − P[Hq∗ | A] such that
min max P[H | Ac ], P[H c | A] = P[Hq∗ | Ac ].
H∈H
Corollary 2.8 is nice in telling us which classifier minimises at the same time the probabilities of ‘false
negative’ and ‘false positive’ predictions on the target dataset for whatever value of the positive class
prevalence. Note the similarity between this classifier and the classifier serving as the basis for the ’method
X’ of Forman (2008). It is also interesting to see that ’Method Max’ of Forman (2008) (maximise TPR FPR) is a special case of (2.2) with a = 1/P[A], b = 1/(1 − P[A]).
But from (2.11) it follows that the prediction error EHq∗ (w) of Hq∗ is zero if and only if w = 1/2, i.e. if the
prior positive class probability on the target dataset is 50%. This may seem unsatisfactory in particular
if the positive class prevalence P[A] on the training set is very different from 1/2. It might be more
appropriate to have prediction error zero for w = P[A], i.e. if the positive class prevalence on the target
dataset is the same as on the training set. For this case the following result applies:
Corollary 2.9 Under Assumption 2.6, define the classifiers Hq , 0 < q < 1 by (2.3). Assume in addition,
that there is a number 0 < r < 1 such that P[Hr ] = P[A]. Then it holds that
min
max P[H | Ac ], P[H c | A] = max P[Hr | Ac ], P[Hrc | A] .
H∈H, P[H]=P[A]
It can be easily checked that for Hr from Corollary 2.9 we have EHr (P[A]) = 0, with E defined as in
(2.11).
2.3. The Q-measure approach
Barranquero et al. (2015) define the ‘Normalized Absolute Score’ (NAS) for measuring how well a classifier
(characterised by a set H ∈ H, as explained below Assumption 2.1) predicts the prior class probabilities
P[A] and 1 − P[A] in the binary classification setting as described by Assumption 2.1:
def
NAS(H) = 1 −
|P[H] − P[A]|
.
max(P[A], 1 − P[A])
(2.13a)
By definition, we have NAS(H) = 1 if and only if P[H] = P[A]. Otherwise, the range of NAS(H) depends
P[A]
on the value of P[A]. If P[A] ≤ 1 − P[A], then P[H] = 0 implies NAS(H) = 1−2
1−P[A] and P[H] = 1
implies NAS(H) = 0. If P[A] > 1 − P[A], then P[H] = 0 implies NAS(H) = 0 and and P[H] = 1 implies
NAS(H) = 2 P[A]−1
P[A] .
10
The dependence of the range of NAS on the value of P[A] is unsatisfactory because it makes comparison
of NAS values computed for different underlying values of P[A] incommensurable and, potentially, could
entail bias if NAS is used as an optimization criterion. The following alternative definition of NAS∗ avoids
these issues:
max(P[H] − P[A], 0) max(P[A] − P[H], 0)
def
NAS∗ (H) = 1 −
−
.
(2.13b)
1 − P[A]
P[A]
By this definition, we have NAS∗ (H) = 1 if and only if P[H] = P[A], NAS∗ (H) = 0 if and only if
P[H] ∈ {0, 1}, and 0 < NAS∗ (H) < 1 otherwise. In the following, we use NAS∗ instead of NAS in order
to make sure that the full potential of the approach by Barranquero et al. (2015) is realised.
Barranquero et al. (2015) suggest training ‘reliable’ classifiers in order to predict prior (unconditional)
class probabilities on the target dataset under prior probability shift. With reliability they mean that
the classifiers in question should perform well in terms of being good at classification and being good at
quantification at the same time.
In order to train an optimal quantifier for predicting the class probability P[A], Barranquero et al. (2015)
suggest to maximise for some β > 0 the ‘Qβ -measure’ over all possible classifiers (characterised by the
sets H ∈ H whose outcomes trigger the prediction of class 1):
(1 + β 2 ) P[H | A] NAS(H)
β 2 P[H | A] + NAS(H)
1 + β2
.
=
2
β
1
NAS(H) + P[H | A]
def
Qβ (H) =
(2.14a)
(2.14b)
By the definitions of P[H | A] and NAS(H) the denominator on the right-hand side of (2.14a) takes the
value 0 if and only if P[H] = 0. Representation (2.14b) of Qβ (H) implies limP[H]→0 Qβ (H) = 0. Therefore
we can define Qβ (H) by (2.14a) for H ∈ H with P[H] > 0 and Qβ (H) = 0 for H ∈ H with P[H] = 0.
Qβ (H) is a weighted harmonic mean of the true positive rate P[H | A] and the normalized absolute score
and, as such, is increasing in both P[H | A] and NAS(H). Barranquero et al. (2015) suggest that by
maximising Qβ (H) over H for fixed A on a training set, the resulting classifier should be able to provide
good estimates of P1 [A] on target datasets with possibly different prior class distributions.
Using the observation from Remark 2.3 (iii), in Section 3 we demonstrate by the standard example of a
binormal model that optimal classifiers with respect to Qβ in general are not best quantifiers. First note
that Remark 2.3 (iii) implies the following result.
Proposition 2.10 Under Assumption 2.1, define Qβ (H), H ∈ H by (2.14a) and Hq , q ∈ (0, 1) by (2.3).
For 0 < u < 1 denote by q(u) the u-quantile4 of P[A | H]. If the distribution of P[A | H] is continuous
then it holds that
sup Qβ (H) =
sup Qβ (Hq(1−u) ).
(2.15)
H∈H
Proof. Observe that sup Qβ (H) =
H∈H
P[A]≤u<1
sup
Qβ (H) since Qβ (H) = 0 for P[H] = 0 and P[H] = 1.
H∈H, 0<P[H]<1
Fix H ∈ H with 0 < P[H] < 1. By continuity of x 7→ P P[A | H] ≤ x , then we have P[Hq(u) ] = P[H] for
u = 1 − P[H]. Hence Remark 2.3 (iii) implies
Qβ (H) =
1 + β2
β2
NAS(Hq(u) )
≤
+
1
P[H | A]
1 + β2
β2
NAS(Hq(u) )
+
1
P[Hq(u) | A]
= Qβ (Hq(u) ).
4 For
a real random variable X and α ∈ (0, 1), the α-quantile qα (X) of X is defined by qα (X) = inf{x ∈ R : P[X ≤ x] ≥ α}.
11
Figure 2: Illustration of Proposition 2.10 and Corollary 2.11.
0.4
0.2
Qbeta
0.6
0.8
beta = 1
beta = 2
0.0
p = 0.25
0.0
0.2
0.4
0.6
0.8
1.0
Predicted prevalence of positive class
This implies
sup Qβ (H) =
H∈H
sup Qβ (Hq(1−u) ).
0<u<1
However, for 0 < u ≤ P[A] both P[Hq(1−u) | A] and NAS(Hq(1−u) ) are non-decreasing in u. Hence it
follows that Qβ (Hq(1−u) ) ≤ Qβ (Hq(1−P[A]) ) for 0 < u ≤ P[A] and therefore (2.15).
2
See Figure 2 for an illustration of Proposition 2.10. Unfortunately, in practice most of the time it is not
possible to accurately estimate general posterior probabilities like P[A | H]. This has led some authors to
propose workarounds like the one by Platt (2000) which do not necessarily deliver good results. However,
the following corollary describes a special setting in which the right-hand side of (2.15) can be considerably
simplified.
Corollary 2.11 Under Assumption 2.1, define Qβ (H), H ∈ H by (2.14a) and Hq , q ∈ R by (2.3). Make
these two additional assumptions:
(i) There is a real random variable X on (Ω, A, P) with continuous distribution such that H ⊃ σ(X),
i.e. X is H-measurable.
(ii) There is a continuous function f : R → [0, 1] such that
P[A | H] = f (X).
12
The function f is either strictly increasing with limx→−∞ f (x) = 0 and limx→∞ f (x) = 1 or strictly
decreasing with limx→−∞ f (x) = 1 and limx→∞ f (x) = 0.
Then if f is increasing it holds that
sup Qβ (H) =
H∈H
Qβ {X > q1−u (X)} .
sup
P[A]≤u≤1
Otherwise, if f is decreasing it holds that
sup Qβ (H) =
H∈H
Qβ {X < qu (X)} .
sup
P[A]≤u≤1
Proof. Follows from Proposition 2.10 because Hq(1−u) = {X > q1−u (X)} in the case of increasing f and
Hq(1−u) = {X < qu (X)} in the case of decreasing f , for P[A] ≤ u < 1.
2
Corollary 2.11 allows us in Section 3 to replicate the experiment by Barranquero et al. (2015) in a fully
controlled environment such that the merits and limitations of their approach can carefully be studied.
3. The binormal case with equal variances
We consider the ‘binormal model’ with equal variances as an example that fits into the setting of Assumption 2.1 and Corollary 2.11.
• Ω = R×{−1, 1}, A = B(R)⊗P({−1, 1}) where B(R) denotes the Borel-σ-field on R and P({−1, 1})
is the power set of {−1, 1}.
• On Ω, we define the projections X and Y , i.e. for ω = (x, y) ∈ Ω we let X(ω) = x and Y (ω) = y.
• A = {Y = 1} ∈
/ H = σ(X).
• P is defined by specifying the marginal distribution of Y with P[A] = p ∈ (0, 1), and defining the
conditional distribution of X given Y as normal distributions with equal variances:
P[X ∈ · | A] = N (ν, σ 2 ),
(3.1a)
P[X ∈ · | Ac ] = N (µ, σ 2 ).
In (3.1a), we assume that µ < ν and σ > 0. (3.1a) implies that the distribution of X is given by a
mixture of normal distributions5
x−ν
x−µ
P[X ≤ x] = p Φ
+ (1 − p) Φ
, x ∈ R.
(3.1b)
σ
σ
• The posterior probability P[A | H] in this setting is given by
1
,
1 + exp(a X + b)
P[A | H] =
with a =
µ−ν
σ2
< 0 and b =
ν 2 −µ2
2 σ2
+ log
1−p
p
(3.2)
.
We replicate the experiment of Barranquero et al. (2015) in this setting:
• We look at a ‘training set’ (Ω, A, P) and a ‘target dataset’ (Ω, A, P1 ). The probability measures P
and P1 are defined like P above but with (possibly) different values of P[A] = p and P1 [A] = p1
respectively.
5Φ
denotes the standard normal distribution function Φ(x) =
13
√1
2π
Rx
−∞
e− y
2
/2
dy.
• We “train” a classifier on the training set (Ω, A, P) by maximising Qβ as given by (2.14a). Classifiers
can be identified with sets H ∈ H in the sense that the prediction is the positive class A if H occurs
and the negative class Ac otherwise. “Training” in the binormal setting actually means to make use
of Corollary 2.11 in order to identify the optimal classifier H ∗ . The following formulae are used for
the optimization of Qβ :
q1−u (X) − ν
P[X > q1−u (X) | A] = 1 − Φ
,
(3.3a)
σ
max(u − p, 0) max(p − u, 0)
−
.
(3.3b)
NAS∗ ({X > q1−u (X)}) = 1 −
1−p
p
The quantile q1−u (X) (see footnote 4 for its definition) must be numerically determined by solving
for x the following equation:
x−ν
x−µ
1 − u = p0 Φ
+ (1 − p0 ) Φ
.
(3.4)
σ
σ
As the optimization problem has no closed-form solution, we apply the R-function ’optimize’ (R
Development Core Team, 2010) to find the optimal classifier H ∗ = {X > q1−u∗ (X)}.
• We evaluate the Qβ -optimal classifier H ∗ on the target dataset (Ω, A, P1 ) by calculating P1 [H ∗ ]
and compare its value to p1 = P1 [A] in order to check how good the Classify & Count approach
(Forman, 2008) based on H ∗ is. We compare the performance of H ∗ with two other classifiers: the
minimax classifier HMini from Corollary 2.8 and the locally best classifier Hloc from Corollary 2.9.
This means that we need to evaluate (2.11) for the three classifiers. For this purpose, the following
formulae are used:
q1−u∗ (X) − ν
q1−u∗ (X) − µ
∗
∗
c
,
P[H | A ] = Φ
;
(3.5a)
P[H | A] = Φ
σ
σ
!
!
µ+ν
µ+ν
2 −ν
2 −µ
,
P[HMini | Ac ] = Φ
;
(3.5b)
P[HMini | A] = Φ
σ
σ
q1−p (X) − µ
q1−p (X) − ν
,
P[Hloc | Ac ] = Φ
.
(3.5c)
P[Hloc | A] = Φ
σ
σ
• For the calculations, we have used the following parameters:
µ = 0,
ν = 2,
σ = 1,
P[A] = p = 25%.
(3.6)
The results of the calculations are shown in Figures 2 and 1. Figure 2 presents graphs of u 7→ Qβ {X >
q1−u (X)} in the binormal setting of this section, with parameters chosen as in (3.6). The Q-measure
Qβ is defined by (2.14b), but we use NAS∗ instead of NAS. The solid curve is for β = 1, where equal
weights are put on NAS∗ and the TPR P[H | A], and the dashed curve is for β = 2, where the weight put
on NAS∗ is four times the weight for the TPR.
The kinks in both graphs of Figure 2 are due to the fact that the mapping u 7→ NAS∗ ({X > q1−u (X)})
with NAS∗ ({X > q1−u (X)}) defined by the right-hand side of (3.3b) is not differentiable in u = p. The
function u 7→ NAS would not be differentiable in u = p either. Both curves have a unique maximum
which is at u = p = 0.25 for β = 2 and at some u > p = 0.25 for β = 1.
Hence the graph for β = 2 has its maximum at that u where NAS∗ ({X > q1−u (X)}) takes its maximum
value 1. As a consequence, in the case β = 2, the Qβ -optimal classifier is identical with the locally best
classifier according to Corollary 2.9.
In contrast, in the case β = 1, for u slightly greater than p = 0.25 the decline in value of NAS∗ ({X >
q1−u (X)}) is over-compensated by a rise in the value of the TPR such that the maximum is incurred at
14
some u > p. The consequence of this for the error (2.11) in the prediction of a target dataset positive
class prevalence is displayed in Figure 1.
Figure 1 shows that by incident the error-performance of the Qβ -optimal classifier for β = 1 is close to the
performance of the minimax classifier identified in Corollary 2.8. Nonetheless, its performance would be
deemed unsatisfactory if the true positive class prevalence in the target dataset were the same or nearly
the same as the assumed positive class prevalence of 25% in the training dataset. In that case, clearly the
locally best classifier as identified by the Qβ -measure for β = 2 or by Corollary 2.9 would perform much
better and even perfectly if the training and target prevalences were the same.
4. Conclusions
We have investigated a claim by Barranquero et al. (2015) and Esuli and Sebastiani (2015) that binary
class prevalences on target datasets can be estimated by classifiers without further adjustments if these
classifiers are developed as so-called quantifiers on the training datasets. The development of such quantifiers involves special optimisation criteria covering both calibration (i.e. the number of objects predicted
as positive equals the true number of positives) and classification power.
Barranquero et al. (2015) recommended the so-called Q-measure as the optimisation criterion for the
quantifier and tested their approach on some real-world datasets. It is not fully clear, however, from
Barranquero et al. (2015) which of their observations are fundamental and which are sample-driven
and incidental. In this paper, therefore, we have identified the theoretically correct way to determine
the best quantifiers according to the Q-measure criterion. We then have replicated the experiment of
Barranquero et al. (2015) in the fully controlled setting of the binormal model with equal variances. For
binary classification settings, we have found that
1) quantification without adjustments in principle may work if proper calibration on the training
dataset is ensured and the positive class prevalences on the training and target datasets are the
same or nearly the same, and
2) the Q-measure approach has to be deployed with care because the quantifiers resulting from it may
be miscalibrated.
All in all, these findings suggest that the potential applications of quantification without adjustments are
rather limited.
References
J. Barranquero, J. Dı́ez, and J.J. del Coz. Quantification-oriented learning based on reliable classifiers.
Pattern Recognition, 48(2):591–604, 2015.
A. Bella, C. Ferri, J. Hernandez-Orallo, and M.J. Ramı́rez-Quintana. Quantification via probability
estimators. In Data Mining (ICDM), 2010 IEEE 10th International Conference on, pages 737–742.
IEEE, 2010.
S. Bickel, M. Brückner, and T. Scheffer. Discriminative learning under covariate shift. The Journal of
Machine Learning Research, 10:2137–2155, 2009.
P. Billingsley. Probability and measure. John Wiley & Sons, third edition, 1995.
M.C. Du Plessis and M. Sugiyama. Semi-supervised learning of class balance under class-prior change by
distribution matching. Neural Networks, 50:110–119, 2014.
15
C. Elkan. The foundations of cost-sensitive learning. In B. Nebel, editor, Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence, IJCAI 2001, pages 973–978. Morgan
Kaufmann, 2001.
A. Esuli and F. Sebastiani. Optimizing Text Quantifiers for Multivariate Loss Functions. ACM Trans.
Knowl. Discov. Data, 9(4):1–27, June 2015. ISSN 1556-4681. doi: 10.1145/2700406. URL http:
//doi.acm.org/10.1145/2700406.
A. Esuli, F. Sebastiani, and A. Abbasi. Sentiment quantification. IEEE intelligent systems, 25(4):72–79,
2010.
T. Fawcett and P.A. Flach. A response to Webb and Ting’s On the Application of ROC Analysis to
Predict classification Performance under Varying Class Distributions. Machine Learning, 58(1):33–38,
2005.
G. Forman. Quantifying counts and costs via classification. Data Mining and Knowledge Discovery, 17
(2):164–206, 2008.
J.J. Gart and A.A. Buck. Comparison of a screening test and a reference test in epidemiologic studies.
II. A probabilistic model for the comparison of diagnostic tests. American Journal of Epidemiology,
83(3):593–602, 1966.
V. González-Castro, R. Alaiz-Rodrı́guez, and E. Alegre. Class distribution estimation based on the
Hellinger distance. Information Sciences, 218:146–164, 2013.
V. Hofer. Adapting a classification rule to local and global shift when only unlabelled data are available.
European Journal of Operational Research, 243(1):177–189, 2015.
V. Hofer and G. Krempl. Drift mining in data: A framework for addressing drift in classification. Computational Statistics & Data Analysis, 57(1):377–391, 2013.
D.J. Hopkins and G. King. A Method of Automated Nonparametric Content Analysis for Social Science.
American Journal of Political Science, 54(1):229–247, 2010.
H. Kawakubo, M.C. du Plessis, and M. Sugiyama. Computationally Efficient Class-Prior Estimation under
Class Balance Change Using Energy Distance. IEICE Transactions on Information and Systems, 99
(1):176–186, 2016.
M. Kull and P. Flach. Patterns of dataset shift. Working paper, 2014.
L. Milli, A. Monreale, G. Rossetti, F. Giannotti, D. Pedreschi, and F. Sebastiani. Quantification trees.
In 13th International Conference on Data Mining (ICDM), pages 528–536. IEEE, December 2013. doi:
10.1109/ICDM.2013.122.
J.G. Moreno-Torres, T. Raeder, R. Alaiz-Rodriguez, N.V. Chawla, and F. Herrera. A unifying view on
dataset shift in classification. Pattern Recognition, 45(1):521–530, 2012.
J.C. Platt. Probabilities for SV Machines. In P.J. Bartlett, B. Schölkopf, D. Schuurmans, and A.J. Smola,
editors, Advances in Large-Margin Classifiers, pages 61–74. MIT Press Cambridge, 2000.
R Development Core Team. R: A Language and Environment for Statistical Computing. R Foundation
for Statistical Computing, Vienna, Austria, 2010. URL http://www.R-project.org/.
M. Saerens, P. Latinne, and C. Decaestecker. Adjusting the Outputs of a Classifier to New a Priori
Probabilities: A Simple Procedure. Neural Computation, 14(1):21–41, 2002.
16
L.L. Scharf. Statistical signal processing: detection, estimation, and time series analysis. Addison-Wesley,
1991.
H. Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood
function. Journal of statistical planning and inference, 90(2):227–244, 2000.
A. Storkey. When Training and Test Sets Are Different: Characterizing Learning Transfer. In
J. Quiñonero-Candela, M. Sugiyama, A. Schwaighofer, and N.D. Lawrence, editors, Dataset Shift in
Machine Learning, pages 3–28. Cambridge, MA: MIT Press, 2009.
M. Sugiyama, M. Krauledat, and K.-R. Müller. Covariate shift adaptation by importance weighted cross
validation. The Journal of Machine Learning Research, 8:985–1005, 2007.
D. Tasche. Exact fit of simple finite mixture models. Journal of Risk and Financial Management, 7(4):
150–164, 2014.
C.J. van Rijsbergen. Foundation of evaluation. Journal of Documentation, 30(4):365–373, 1974.
H.L. van Trees. Detection, Estimation, and Modulation Theory, Part I. John Wiley & Sons, 1968.
J.C. Xue and G.M. Weiss. Quantification and Semi-supervised Classification Methods for Handling
Changes in Class Distribution. In Proceedings of the 15th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 897–906, New York, 2009.
N. Ye, K.M. Chai, W.S. Lee, and H.L. Chieu. Optimizing F-measure: A Tale of Two Approaches.
In J. Langford and J. Pineau, editors, Proceedings of the 29th International Conference on Machine
Learning (ICML-12), pages 289–296, New York, NY, USA, 2012. ACM. URL http://icml.cc/2012/
papers/175.pdf.
A. Appendix: Optimal classifiers for the F-measure
In the following, we give a proof of (2.8) which is based on the idea of Ye et al. (2012) and fills the gaps
left by them.
Proof of (2.8). The proof makes use of the four following lemmata.
Lemma A.1 Let (Ω, A, P), A ∈ A and H ⊂ A be as in Assumption 2.1.Then there are a probability
space (Ω̃, Ã, P̃), an event à ∈ à and a sub-σ-field H̃ of à with the following three properties:
(i) There is a measurable mapping X : (Ω̃, Ã) → (Ω, A) such that
P̃ ◦ X −1 = P,
à = X −1 (A),
H̃ ⊃ X −1 (H),
P̃[Ã | H̃] = P[A | H] ◦ X.
(ii) (Ω̃, Ã, P̃), Ã ∈ Ã and H̃ satisfy Assumption 2.1.
def
(iii) For any H ∈ H and α ∈ [0, 1] there exists H̃α ∈ H̃, H̃α ⊂ X −1 (H) = H̃ such that
P̃[H̃α ] = α P̃[H̃].
Proof of Lemma A.1. For (i) and (ii), define Ω̃ = Ω × (0, 1), Ã = A ⊗ B(0, 1), P̃ = P ⊗ U (0, 1), and let
X be the projection from Ω̃ to Ω, where B(0, 1) denotes the Borel-σ-field on (0, 1) and U (0, 1) stands for
the uniform distribution on (0, 1). In addition, let H̃ = H ⊗ B(0, 1). For (iii), define H̃α = H × (0, α). 2
Lemma A.2 Under Assumption 2.1, for any fixed H ∈ H there is a number q ∈ [0, 1] such that
P P[A | H] > q ≤ P[H] ≤ P P[A | H] ≥ q .
17
(A.1)
2
Proof. Choose q as an (1 − P[H])-quantile of P[A | H] or q ∈ {0, 1}.
Lemma A.3 Under Assumption 2.1, if for a fixed H ∈ H there are H ∗ ∈ H and 0 ≤ q ≤ 1 such that
{P[A | H] > q} ⊂ H ∗ ⊂ {P[A | H] ≥ q}
and
P[H] = P[H ∗ ],
then it holds that P[H ∩ A] ≤ P[H ∗ ∩ A].
Proof. By assumption we have H ∗ ⊂ {P[A | H] ≥ q} and (H ∗ )c ⊂ {P[A | H] ≤ q}. This implies
P[H ∗ ∩ A] − P[H ∩ A] = P[H ∗ ∩ H c ∩ A] − P[(H ∗ )c ∩ H ∩ A]
= E P[A | H] 1H ∗ ∩H c − E P[A | H] 1(H ∗ )c ∩H
≥ q P[H ∗ ∩ H c ] − q P[(H ∗ )c ∩ H]
= q P[H ∗ ] − P[H] = 0.
2
Lemma A.4 Under Assumption 2.1, for all β > 0 and q ∈ [0, 1] it holds that
Fβ (H) ≤ max Fβ (H ∩ {P[A | H] > q}), Fβ (H ∪ {P[A | H] = q}) .
(A.2)
Proof. The proof of this lemma is essentially identical with the proof given by Ye et al. (2012) for
Theorem 4. This suggests that the proof by Ye et al. (2012) is incomplete since some more steps are
needed to show that (A.2) implies (2.8).
Fix β > 0. We distinguish the two cases q = 1 and 0 ≤ q < 1. For the sake of a more concise notation,
we define
Z = P[A | H].
Case q = 1. We show that Fβ (H) ≤ Fβ (H ∪ {Z = 1}). Note that H ∪ {Z = 1} = H ∪ (H c ∩ {Z = 1})
and that
P[A ∩ H c ∩ {Z = 1}] = E Z 1H c ∩{Z=1} = P[H c ∩ {Z = 1}].
This implies
Fβ (H) ≤ Fβ (H ∪ {Z = 1})
(1 + β 2 ) P[A ∩ H] + P[H c ∩ {Z = 1}]
(1 + β 2 ) P[A ∩ H]
≤
⇔
β 2 P[A] + P[H]
β 2 P[A] + P[H] + P[H c ∩ {Z = 1}]
⇔ P[A ∩ H] P[H c ∩ {Z = 1}] ≤ P[H] P[H c ∩ {Z = 1}] + β 2 P[A] P[H c ∩ {Z = 1}].
The last row of this equation chain is true because P[A ∩ H] ≤ P[H]. This proves (A.2) in the case q = 1.
Case 0 ≤ q < 1. We show that Fβ (H) > Fβ (H ∩ {Z > q}) implies Fβ (H) ≤ Fβ (H ∪ {Z = q}). Observe
first that
⇔
Fβ (H) > Fβ (H ∩ {Z > q})
β P[A] + P H ∩ {Z > q} P A ∩ H ∩ {Z ≤ q} > P H ∩ {Z ≤ q} P A ∩ H ∩ {Z > q} . (A.3)
2
Secondly, we have
Fβ (H) ≤ Fβ (H ∪ {Z = q})
c
⇔ P[A ∩ H] P H ∩ {Z = q} ≤ β 2 P[A] + P[H] q P H c ∩ {Z = q} .
18
(A.4)
If P H c ∩ {Z = q} = 0 then by (A.4) there is nothing left to be proved. Hence assume P H c ∩ {Z =
q} > 0. Then from (A.4) we obtain
Fβ (H) ≤ Fβ (H ∪ {Z = q})
⇔ P[A ∩ H] ≤
β 2 P[A] + P[H] q.
(A.5)
(A.3) implies P H ∩ {Z ≤ q} > 0 because otherwise we would have 0 > 0. Therefore, (A.5) follows from
(A.3) because
P[A ∩ H] = P A ∩ H ∩ {Z ≤ q} + P A ∩ H ∩ {Z > q}
β 2 P[A] + P H ∩ {Z > q} P A ∩ H ∩ {Z ≤ q}
≤ P A ∩ H ∩ {Z ≤ q} +
P H ∩ {Z ≤ q}
!
β 2 P[A] + P H ∩ {Z > q}
= E Z 1H∩{Z≤q} 1 +
P H ∩ {Z ≤ q}
β 2 P[A] + P[H]
≤ P H ∩ {Z ≤ q} q
P H ∩ {Z ≤ q}
= β 2 P[A] + P[H] q.
Hence (A.5) is true if (A.3) holds. This completes the proof of (A.2).
2
Finishing the proof of (2.8). Fix H ∈ H. We need to show that there is a number 0 ≤ q ≤ 1 such that
Fβ (H) ≤ max Fβ ({P[A | H] > q}), Fβ ({P[A | H] ≥ q}) .
(A.6)
Since Fβ (H) = 0 if P[H] = 0, (A.6) is obvious is that case. Therefore, we may assume P[H] > 0 for the
remainder of the proof.
The notation Fβ (H) hides the fact that Fβ (H) does not only depend on the classifier H, but also on
the class event A and the probability measure P. This matters because the next step in the proof is to
replace (Ω, A, P), A and H by (Ω̃, Ã, P̃), Ã and H̃ as provided by Lemma A.1. In the following, when
using notation like Fβ (H̃) we implicitly assume that all ingredients for the calculation come from the
probability space (Ω, A, P).
Against this backdrop, define H̃ = X −1 (H) where X denotes the projection on Ω as defined in Lemma A.1.
Then it is easy to see that
Fβ (H) = Fβ (H̃), Fβ ({P[A | H] > q}) = Fβ ({P̃[Ã | H̃] > q}), Fβ ({P[A | H] ≥ q}) = Fβ ({P̃[Ã | H̃] ≥ q}).
Hence, if we prove
Fβ (H̃) ≤ max Fβ ({P̃[Ã | H̃] > q}), Fβ ({P̃[Ã | H̃] ≥ q}) ,
(A.7)
(A.6) immediately follows. Choose 0 ≤ q ≤ 1 according to Lemma A.2 such that
P̃ P̃[Ã | H̃] > q ≤ P̃[H̃] ≤ P̃ P̃[Ã | H̃] ≥ q .
(A.8)
By Lemma A.1 (iii) and (A.8), there exists an event H̃0 ∈ H̃ such that
H̃0 ⊂ {P̃[Ã | H̃] = q} and P̃[H̃0 ] = P̃[H̃] − P̃ P̃[Ã | H̃] > q .
Hence for H̃ ∗ = H̃0 ∪ {P̃[Ã | H̃] > q} we obtain
P̃[H̃ ∗ ] = P̃[H̃]
and {P̃[Ã | H̃] > q} ⊂ H̃ ∗ ⊂ {P̃[Ã | H̃] ≥ q}.
Now, Lemma A.3 implies that
Fβ (H̃) =
(1 + β 2 ) P̃[Ã ∩ H̃ ∗ ]
(1 + β 2 ) P̃[Ã ∩ H̃]
≤
= Fβ (H̃ ∗ ).
β 2 P̃[Ã] + P̃[H̃]
β 2 P̃[Ã] + P̃[H̃ ∗ ]
19
(A.9)
From Lemma A.4, it follows that
Fβ (H̃ ∗ ) ≤ max Fβ (H̃ ∗ ∩ {P̃[Ã | H̃] > q}), Fβ (H̃ ∗ ∪ {P̃[Ã | H̃] = q}) .
(A.10)
Note that by construction, we have got that
H̃ ∗ ∩ {P̃[Ã | H̃] > q} = {P̃[Ã | H̃] > q}
and H̃ ∗ ∪ {P̃[Ã | H̃] = q} = {P̃[Ã | H̃] ≥ q}.
2
Hence, (A.9) and (A.10) together imply (A.7).
20
| 10math.ST
|
UNIFORM SYMBOLIC TOPOLOGIES VIA MULTINOMIAL EXPANSIONS
arXiv:1703.04530v2 [math.AC] 15 Dec 2017
ROBERT M. WALKER
Abstract. When does a Noetherian commutative ring R have uniform symbolic topologies on
primes–read, when does there exist an integer D > 0 such that the symbolic power P (Dr) ⊆ P r for
all prime ideals P ⊆ R and all r > 0? Groundbreaking work of Ein-Lazarsfeld-Smith, as extended
by Hochster and Huneke, and by Ma and Schwede in turn, provides a beautiful answer in the setting
of finite-dimensional excellent regular rings. It is natural to then sleuth for analogues where the ring
R is non-regular, or where the above ideal containments can be improved using a linear function
whose growth rate is slower. This manuscript falls under the overlap of these research directions.
Working with a prescribed type of prime ideal Q inside of tensor products of domains of finite type
over an algebraically closed field F, we present binomial- and multinomial expansion criteria for
containments of type Q(Er) ⊆ Qr , or even better, of type Q(E(r−1)+1) ⊆ Qr for all r > 0. The final
section consolidates remarks on how often we can utilize these criteria, presenting an example.
1. Introduction and Conventions for the Paper
Given a Noetherian commutative ring R, when is there an integer D, depending only on R, such
that the symbolic power P (Dr) ⊆ P r for all prime ideals P ⊆ R and all positive integers r? In
short, when does R have uniform symbolic topologies on primes [6, Section 3] [16]?
The Ein-Lazarsfeld-Smith Theorem [8], as extended by Hochster and Huneke [13], says that if R
is a d-dimensional regular ring containing a field, then Q(Dr) ⊆ Qr for all radical ideals Q ⊆ R and
all r > 0, where D = max{1, d−1}.1 To what extent does this theme ring true for non-regular rings?
Under mild stipulations, a local domain R regular on the punctured spectrum has uniform symbolic
topologies on primes [15, Cor. 3.10]. This paper makes first strides in establishing affirmative cases
of the above questions for rings with non-isolated singularities. We postpone sojourning into that
wilderness until Section 3, once we have proven our main result–Theorem 1.2.
We now revisit the regular setting. Over an arbitrary field F, S = F[PN ] = F[x0 , x1 , . . . , xN ]
is a standard N-graded polynomial ring. The groundbreaking work of Ein-Lazarsfeld-Smith and
Hochster-Huneke [8, 13] implies that the symbolic power I (N r) ⊆ I r for all graded ideals 0 $ I $ S
and all integers r > 0. In particular, I (4) ⊆ I 2 holds for all graded ideals in F[P2 ], and Huneke
asked whether an improvement I (3) ⊆ I 2 holds for any radical ideal I defining a finite set of points
in P2 . Building on this, Harbourne proposed dropping the symbolic power from N r down to the
Harbourne-Huneke bound N r − (N − 1) = N (r − 1) + 1 when N ≥ 2 [3, Conj. 8.4.2]: i.e.,
I (N (r−1)+1) ⊆ I r for any graded ideal 0 $ I $ S, all r > 0, and all N ≥ 2.
(1.0.1)
There are several scenarios where these improved containments hold: for instance, they hold for all
monomial ideals in S over any field [3, Ex. 8.4.5]; see also recent work of Grifo-Huneke [10].
2010 Mathematics Subject Classification: 13H10, 14C20, 14M25.
Keywords: multinomial theorem, non-isolated singularities, symbolic powers, toric variety.
1This result has been extended to all excellent regular rings, even in mixed characteristic, by Ma-Schwede [17].
1
However, Dumnicki, Szemberg, and Tutaj-Gasińska showed in characteristic zero [7] that the
containment I (3) ⊆ I 2 can fail for a radical ideal defining a point configuration in P2 . HarbourneSeceleanu showed in odd positive characteristic [12] that (1.0.1) can fail for pairs (N, r) 6= (2, 2)
and ideals I defining a point configuration in PN . Akesseh [1] cooks up many new counterexamples
to (1.0.1) from these original constructions. No prime ideal counterexample has been found.
Our goal is to establish Harbourne-Huneke bounds on the growth of symbolic powers of certain
primes in non-regular Noetherian rings containing a field. This project first began with Theorems
1.1 and 3.1 in [22]: to clarify, normal affine semigroup rings are domains generated by Laurent
monomials that arise as the coordinate rings of normal affine toric varieties [5, 9].
Theorem 1.1 ([22, Thm. 1.1]). Let R1 , . . . , Rn be normal affine semigroup rings over a field
F, built, respectively, from full-dimensional strongly convex rational polyhedral cones σi ⊆ Rmi ,
1 ≤ i ≤ n. For each 1 ≤ i ≤ n, suppose there is an integer Di > 0 such that P (Di (r−1)+1) ⊆ P r for
all r > 0 and all monomial primes P ⊆ Ri . Set D := max{D1 , . . . , Dn }. Then Q(D(r−1)+1) ⊆ Qr
for all r > 0 and any monomial prime Q in the normal affine semigroup ring R = R1 ⊗F · · · ⊗F Rn .
To prove Theorem 1.1, we needed to know first that monomial primes
P in Ri expand to monomial
primes in R, that any Q as above can be expressed as a sum Q = ni=1 Pi R where each Pi ⊆ Ri is
a monomial prime, and that the symbolic powers of Q admit a multinomial expansion in terms of
symbolic powers of the Pi R. These ideas will resurge below, but in a more general setup.
One drawback of Theorem 1.1 is that it only covers a finite collection of prime ideals. What
follows is the main result of this paper, a more powerful variant of Theorem 1.1 that will typically
cover infinitely-many primes inside of tensor product domains; see Remark 2.11 for details.
Theorem 1.2. Let F be an algebraically closed field. Let R1 , . . . , Rn (n ≥ 2) be affine commutative
F-algebras which are domains. Suppose that for each 1 ≤ i ≤ n, there exists a positive integer Di
such that for all prime ideals P in Ri , either: (1) P (Di r) ⊆ P r for all r > 0 and for all i; or, even
stronger, (2) P (Di (r−1)+1) ⊆ P r for all r > 0 and for all i. Fix any
N nnprime ideals Pi in Ri , and
′
consider the expanded ideals Pi = Pi T in the affine domain T = ( F )i=1 Ri , along with their sum
P
Q = ni=1 Pi′ in T . When (1) holds, Q(Dr) ⊆ Qr for all r > 0, where D = D1 + · · · + Dn . When
(2) holds, this improves to Q(D(r−1)+1) ⊆ Qr for all r > 0, where D = max{D1 , . . . , Dn }.
The proof of this theorem leverages a multinomial formula for the symbolic powers of the prime
ideal Q in T (Theorem 2.8). Hà, Nguyen, Trung, and Trung recently announced a binomial theorem
for symbolic powers of ideal sums [11, Thm. 3.4], generalizing [4, Thm. 7.8], where one takes two
arbitrary ideals I ⊆ A, J ⊆ B inside of two Noetherian commutative algebras over a common field
k, whose tensor product R = A ⊗k B is Noetherian; see Remark 2.12 for details. However, we give
a proof of the Multinomial Theorem 2.8 which is more elementary and self-contained.
Conventions: All our rings are Noetherian and commutative with identity. Indeed, our rings will
typically be affine F-algebras, that is, of finite type over a fixed field F of arbitrary characteristic.
By algebraic variety, we will mean an integral scheme of finite type over the field F.
Acknowledgements: I thank my thesis adviser Karen E. Smith, while on her sabbatical, and my
surrogate adviser, Mel Hochster, for several patient and fruitful discussions during the Fall 2016
semester. I thank Huy Tài Hà for sharing a preliminary draft of Section 2 of [11] in November 2016. I
thank Eloı́sa Grifo Pires, Daniel Hernández, Jack Jeffries, Luis Núñez-Betancourt, and Felipe Pérez
for reading a draft of the paper. I thank an anonymous referee for comments improving exposition
in the paper. I acknowledge support from a NSF GRF (Grant No. PGF-031543), NSF RTG grant
DMS-0943832, and a 2017 Ford Foundation Dissertation Fellowship.
2
2. A Multinomial Theorem for Symbolic Powers of Primes
If P is any prime ideal in a Noetherian ring R, its a-th (a ∈ Z>0 ) symbolic power ideal
P (a) = P a RP ∩ R = {f ∈ R : uf ∈ P a for some u ∈ R − P }
is the P -primary component in any Lasker-Noether minimal primary decomposition of P a ; it is the
smallest P -primary ideal containing P a . We separately set P (0) = P 0 = R to be the unit ideal.
Note that P (1) = P , while the inclusion P (a) ⊇ P a for each a > 1 can be strict. Before proceeding,
we record a handy asymptotic conversion lemma.
Lemma 2.1 (Cf., [22, Lem. 3.3]). Given a prime ideal P in a Noetherian ring S, and E ∈ Z>0 ,
P (N ) ⊆ P ⌈N/E⌉ for all N ≥ 0 ⇐⇒ P (E(r−1)+1) ⊆ P r for all r > 0.
Torsion free modules over Noetherian Domains. A module M over a domain R is torsion
free if whenever rx = 0 for some x ∈ M and r ∈ R, then either r = 0 or x = 0. We first record
a lemma on torsion free modules to be used both here and in the next subsection (cf., Lemmas
15.6.7-8 from the Stacks Project page [20] on torsion free modules):
Lemma 2.2. Let R be a Noetherian domain. Let M be a nonzero finitely generated R-module.
Then the following assertions are equivalent:
(1) M is torsion free;
(2) M is a submodule of a finitely generated free module;
(3) (0) is the only associated prime of M, i.e., AssR (M ) = {(0)}.
Working over an arbitrary field F, we fix two affine F-algebras R and S which are domains. The
tensor product T = R ⊗F S will be an affine F-algebra. T is a domain when F is algebraically closed
(Milne [18, Prop. 4.15]). We note that when R and S are duly nice (e.g., polynomial, or normal
toric rings more generally), T is a domain over any field. We now record two additional lemmas.
Lemma 2.3. Suppose that all three of R, S, and T = R ⊗F S are affine domains over a field F. If
M and N are finitely generated torsion free modules over R and S, respectively, then M ⊗F N is a
finitely generated torsion free T -module.
Proof. Viewed as vector spaces, M ⊗F N = 0 if and only if M = 0 or N = 0, in which case torsion
freeness is vacuous. So we will assume all three of M, N, and M ⊗F N are nonzero. Per Lemma 2.2,
suppose we have embeddings M ⊆ Ra and N ⊆ S b . Apply the functor • ⊗F N to the first inclusion
to get M ⊗ N ⊆ Ra ⊗ N , which in turn is contained in Ra ⊗ S b by tensoring the inclusion N ⊆ S b
with Ra . Thus M ⊗ N ⊆ Ra ⊗ S b ∼
= (R ⊗ S)ab = T ab , where the isomorphism is easily checked
in the category of F-vector spaces since direct sum commutes with tensor product. Of course, this
inclusion holds in the category of T -modules, and all T -submodules of T ab are finitely generated
since T is Noetherian, so we are done by invoking Lemma 2.2 again.
Lemma 2.4. For any prime P in any Noetherian ring A, the finitely generated module P (a) /P (a+1)
is torsion free as an A/P -module for all integers a ≥ 0.
Proof. Say x ∈ (P (a) /P (a+1) ) is killed by r ∈ A/P . This means, lifting to A, that x ∈ P (a)
and rx ∈ P (a+1) . Localize at P . Then rx ∈ P (a+1) AP = P a+1 AP . If r 6∈ P , this means x ∈
P a+1 AP ∩ A = P (a+1) . That is, either r = 0 in A/P or otherwise, x = 0 in (P (a) /P (a+1) ). Ergo by
definition, (P (a) /P (a+1) ) is a torsion-free A/P -module.
3
Finally, we record a consequence of Lemma 2.3 that will be important in the next subsection.
The following proposition follows immediately from Lemmas 2.3 and 2.4
Proposition 2.5. Suppose that all three of R, S, and T = R ⊗F S are affine domains over a
field F. Fix two prime ideals P and Q in R and S respectively, such that the affine F-algebra
T ′ = (R/P ) ⊗F (S/Q) is a domain. Then (P (a) /P (a+1) ) ⊗F (Q(b) /Q(b+1) ) is finitely generated and
torsion free over T ′ for any pair of nonnegative integers a and b.
Proving the Multinomial Theorem. Working over an algebraically closed field F, we fix two
affine F-algebras R and S that are domains, and two prime ideals P ⊆ R, Q ⊆ S. Let
T = R ⊗ S ⊇ P ⊗ S + R ⊗ Q =: P T + QT, T ′ = (R/P ) ⊗ (S/Q) ∼
= T /(P T + QT ),
where all tensor products are over F. Both T and T ′ are affine domains over F. Because F is
algebraically closed, the extended ideals P T, QT are both prime, along with their sum P T + QT .
We cannot relax the assumption that F is algebraically closed to its merely being perfect. For
instance, R is perfect (being of characteristic zero), and along the ring extension
S :=
R[x] ∼
C[x] ∼
= C ֒→ T := 2
= C ⊗R S ∼
= C ⊗R C
(x2 + 1)
(x + 1)
the zero ideal of S (which is maximal) extends to a radical ideal which is not prime.
Relative to a flat map φ : A → B of Noetherian rings, we define the ideal JB := hφ(J)iB for
any ideal J in A. Then J r B = (JB)r for all r ≥ 0, since the two ideals share a generating set. We
define the set P(A) = {prime ideals P ⊆ A : P B is prime} to consist of prime ideals that extend
along φ to prime ideals of B. We now record without proof a handy proposition.
Proposition 2.6 (Cf., [22, Prop. 2.1]). Suppose φ : A → B is a faithfully flat map of Noetherian
rings. Then for each prime ideal P ∈ P(A) and all integer pairs (N, r) ∈ (Z≥0 )2 , we have
P (N ) B = (P B)(N ) ,
and
P (N )
⊆
Pr
if and only if
(P B)(N )
=
P (N ) B
⊆
P rB
=
(2.0.1)
(P B)r .
When B is a polynomial ring in finitely many variables over A and φ is inclusion, P(A) = Spec(A).
It is possible that P(A) 6= Spec(A) in Proposition 2.6, per the T ∼
= C ⊗R C example. Working
over a field F, we use Proposition 2.6 when B = A ⊗F C for two affine F-algebras, so B is an affine
F-algebra; when A and C are domains and F is algebraically closed, B is a domain, P(A) = Spec(A)
and P(C) = Spec(C).
We are now ready to prove a binomial theorem for the symbolic powers of P T + QT .
P
Theorem 2.7. For all n ≥ 1, the symbolic power (P T + QT )(n) = a+b=n (P T )(a) (QT )(b) .
Proof. We’ll drop the T ’s from the notation,
P and we will assume that both P, Q are nonzero to
justify the effort. For 0 ≤ c ≤ n, set Jc = ct=0 P (c−t) Q(t) , so Jc ⊆ Jc−1 for all 1 ≤ c ≤ n, since
P (c−t) ⊆ P (c−1−t) for t ≤ c − 1 and for t = c, Q(c) ⊆ Q(c−1) . Note that
X
X
(!)
(P + Q)n =
P a Qb ⊆ Jn =
P (a) Q(b) ⊆ (P + Q)(n) ,
a+b=n
a+b=n
and (!) is easy to verify term-by-term for each P (a) Q(b) . Indeed, P (a) Q(b) is generated by elements
of the form f g with f ∈ P (a) ⊂ R and g ∈ Q(b) ⊂ S (viewing them as elements of T ). We need
f g ∈ (P + Q)(a+b) . Per Proposition 2.6, there exist u ∈ R − P and v ∈ S − Q such that uf ∈ P a and
4
vg ∈ Qb . Viewing u and v as elements of the overring T , we have uv 6∈ (P + Q). Indeed, since P + Q
is prime, if uv ∈ P + Q, then either u or v is in P + Q, but (P + Q)T ∩ R = P and (P + Q)T ∩ S = Q,
contradicting that u 6∈ P and v 6∈ Q. Therefore, in T , (uf )(vg) = (uv)(f g) ∈ P a Qb ⊂ (P + Q)a+b ,
which means f g ∈ (P + Q)(a+b) . Thus (!) holds, and notably Jn is a proper ideal–read, Jn $ T .
Since Jn contains (P + Q)n , and (P + Q)(n) is the smallest (P + Q)-primary ideal containing
(P +Q)n , the opposite inclusion to (!) will follow once we show that Jn is (P +Q)-primary, i.e., that
the set of associated primes AssT (T /Jn ) = {P + Q}. We have short exact sequences of T -modules
0 → Jc−1 /Jc → T /Jc → T /Jc−1 → 0,
for all 1 ≤ c ≤ n.
Thus AssT (Jc−1 /Jc ) ⊆ AssT (T /Jc ) ⊆ AssT (Jc−1 /Jc ) ∪ AssT (T /Jc−1 ) for all 1 ≤ c ≤ n, using the
fact that given an inclusion of modules N ⊆ M ,
Ass(N ) ⊆ Ass(M ) ⊆ Ass(N ) ∪ Ass(M/N ).
Thus by iterative unwinding and using that J0 = T , i.e., AssT (T /J0 ) = ∅, we conclude that
∅ 6= AssT (T /Jn ) ⊆
n
[
AssT (Jc−1 /Jc ).
(2.0.2)
c=1
Taking all direct sums and tensor products over F, we have a series of vector space isomorphisms
M
[P (a) /P (a+1) ⊗ Q(b) /Q(b+1) ], 1 ≤ c ≤ n.
(2.0.3)
Jc−1 /Jc ∼
=
a+b=c−1
We prove this first, considering two chains of symbolic powers, where each ideal is expressed as a
direct sum of F-vector spaces:
P (c) = V0 ⊆ P (c−1) = V0 ⊕ V1 ⊆ . . . ⊆ P (0) = R = V0 ⊕ · · · ⊕ Vc ,
Q(c) = W0 ⊆ Q(c−1) = W0 ⊕ W1 ⊆ . . . ⊆ Q(0) = S = W0 ⊕ · · · ⊕ Wc .
In particular, for all pairs 0 ≤ a, b ≤ c − 1,
P (a) =
c−a
M
Vi ,
P (a+1) =
c−a−1
M
Q(b) =
Vi ,
Q(b+1) =
Wj ,
c−b−1
M
Wj .
j=0
j=0
i=0
i=0
c−b
M
For any pair a, b as above with a + b = c − 1, c − b = a + 1, and so
M
a+b=c−1
P (a)
Q(b) ∼
⊗
=
F
P (a+1)
Q(b+1)
M
Vc−a ⊗ Wc−b =
c−1
M
Vc−a ⊗ Wa+1 .
a=0
a+b=c−1
We now prove (2.0.3) by killing off a common vector space. First,
X
M
Jc−1 =
P (a) Q(b) =
Vi ⊗ W j
a+b=c−1
0≤a≤c−1
0≤i≤c−a,0≤j≤a+1
=
M
(Vi ⊗ Wj ) ⊕
0≤a≤c−1
or
M
0≤i<c−a
while
Jc =
X
a+b=c
P (a) Q(b) =
0≤a≤c
0≤j<a+1
0≤i≤c−a,0≤j≤a
5
Vi ⊗ W j .
c−1
M
a=0
Vc−a ⊗ Wa+1 ,
Identifying repeated copies of a Vi ⊗ Vj term with i + j ≤ c (we can do this since we are working
with vector subspaces of the ring T ), it is straightforward to check that the boxed sums are equal.
Thus for each 1 ≤ c ≤ n, we have canonical isomorphisms of F-vector spaces:
Jc−1 /Jc ∼
=
c−1
M
Vc−a ⊗ Wa+1 ∼
=
a=0
M
a+b=c−1
Q(b)
P (a)
⊗
.
F
P (a+1)
Q(b+1)
Therefore, since for each 1 ≤ c ≤ n there is a natural surjective T -module map (hence F-linear)
M
[P (a) /P (a+1) ⊗ Q(b) /Q(b+1) ] → Jc−1 /Jc ,
a+b=c−1
this map must be injective per isomorphism (2.0.3). Thus for all 1 ≤ c ≤ n,
[
AssT (Jc−1 /Jc ) =
AssT [P (a) /P (a+1) ⊗ Q(b) /Q(b+1) ].
a+b=c−1
For any 1 ≤ c ≤ n such that Jc−1 /Jc 6= 0, i.e., AssT (Jc−1 /Jc ) 6= ∅, in turn the above identity
implies that one of the modules P (a) /P (a+1) ⊗ Q(b) /Q(b+1) is nonzero, in which case
[
AssT (Jc−1 /Jc ) =
AssT [P (a) /P (a+1) ⊗ Q(b) /Q(b+1) ] = {P + Q}.
(2.0.4)
a+b=c−1
To explain the right-hand equality: for any pair (a, b) ∈ (Z≥0 )2 , Proposition 2.5 says that
Ma,b := P (a) /P (a+1) ⊗ Q(b) /Q(b+1)
is a finitely generated torsion-free module over T ′ = (R/P ) ⊗ (S/Q) ∼
= T /(P + Q); thus when
Ma,b 6= 0, we have AssT /(P +Q) (Ma,b ) = {(0)} by Lemma 2.2: that is, AssT (Ma,b ) = {P + Q}.
Finally, combining (2.0.4) with the inclusion
(2.0.2) for AssT (T /Jn ) 6= ∅–recall, Jn is a proper
S
ideal, we conclude that AssT (T /Jn ) = nc=1 AssT (Jc−1 /Jc ) = {P + Q}, that is, the ideal Jn is
(P + Q)-primary as was to be shown. Thus Jn ⊇ (P + Q)(n) , and indeed this is an equality.
We now deduce a multinomial theorem by induction on the number of tensor factors:
Theorem 2.8. Let F be an algebraically closed field. Let R1 , . . . , Rn (n ≥ 2) be affine commutative
F-algebras which are domains. Fix any
N n prime ideals Pi in Ri , and consider the expanded ideals
Pi′ = Pi T in the affine domain T = ( F )ni=1 Ri . Then the symbolic power
!(N )
n
n
Y
X
X
′
(Pi′ )(Ai ) for any N ≥ 0.
(2.0.5)
Pi
=
i=1
A1 +···+An =N i=1
Proof. Induce on the number n of tensor factors with base case n = 2 being Theorem 2.7. Now
suppose n ≥ 3, and assume the result for tensoring up to n − 1 factors. Suppose that R = R1 and
S = R2 ⊗F · · · ⊗F Rn , and that we have an expansion result in S of the form
!(N )
n
n
X
Y
X
(A )
Pi
Pi i
for all nonnegative integers N
(2.0.6)
=
i=2
A2 +...+An =N i=2
P
for n − 1 primes Pi ⊆ Ri (2 ≤ i ≤ n). The sum Q := ni=2 Pi is prime along with all extensions of
the Pi to S. Given a prime P = P1 in R, the sum P + Q is prime in T = R ⊗F S, together with all
6
extensions Pi T and QT being prime. The first equality below holds by Theorem 2.7, and applying
Proposition 2.6 to the extension φ : S ֒→ T , the second equality holds by (2.0.6):
n
N
Y
X
X
X
(A )
(A )
P i
P1 1
(P + Q)(N ) =
P (A1 ) Q(B) =
i
A2 +...+An =N −A1 i=2
A1 =0
A1 +B=N
n
Y
X
⊆
(Ai )
Pi
,
A1 +A2 +...+An =N i=1
using the fact that I(J + K) ⊆ IJ + IK whenever I, J, K are ideals in a commutative ring. This
proves the n-fold version of the hard inclusion in the proof of Theorem 2.7; deducing the opposite
inclusion is about as easy as before, hence the above inclusion is an equality.
Proving Theorem 1.2: We now use the Multinomial Theorem 2.8 to deduce a corollary. Note
that Theorem 1.2 is the version of this corollary where all tensor factors are assumed to satisfy
uniform symbolic topologies on primes.
Corollary 2.9. Let F be an algebraically closed field. Let R1 , . . . , Rn (n ≥ 2) be affine commutative
F-algebras which are domains.
n primes Pi P
⊆ Ri , and consider the expanded ideals Pi′ = Pi T
N Fix
in the affine domain T = ( F )ni=1 Ri ; set Q = ni=1 Pi′ . Suppose that for each 1 ≤ i ≤ n, there
exists a positive integer Di such that either:
(D r)
(1) Pi i ⊆ Pir for all r > 0 and for all i; or, even stronger,
(D (r−1)+1)
(2) Pi i
⊆ Pir for all r > 0 and for all i.
When (1) holds, Q(Dr) ⊆ Qr for all r > 0, where D = D1 + · · · + Dn . When (2) holds, this improves
to Q(D(r−1)+1) ⊆ Qr for all r > 0, where D = max{D1 , . . . , Dn }.
Proof. Assume (1) holds. Per Theorem 2.8 note that for D = D1 + D2 + · · · + Dn ,
Q
(Dr)
n
Y
(Pi′ )(Ai ) .
X
=
A1 +A2 +···+An =D1 r+D2 r+···+Dn r i=1
In each n-tuple of indices (A1 , . . . , An ), we must have that Aj ≥ Dj r for some j, otherwise
Pn
Pn
Qn
′ (Ai ) will lie in some (P ′ )r
j
i=1 Ai <
i=1 Di r, a contradiction. Thus each summand
i=1 (Pi )
r
applying (1) and Proposition 2.6, and hence also in Q . Since r > 0 was arbitrary, we win.
(D(r−1)+1)
If (2) holds, then Pi
⊆ Pir for all r > 0 and all i, where D = max1≤i≤n Di , so equivalently
per Lemma 2.1 and Proposition 2.6, for all n-tuples (A1 , . . . , An ) ∈ (Z≥0 )n , we have containments
(Pi′ )(Ai ) ⊆ (Pi′ )⌈Ai /D⌉ ⊆ Q⌈Ai /D⌉ . For all nonnegative integers N , per Theorem 2.8
Q
(N )
=
X
n
Y
(Pi′ )(Ai ) ⊆
A1 +···+An =N i=1
Pn
X
n
Y
A1 +···+An =N i=1
(Pi′ )⌈Ai /D⌉
⊆
X
n
Y
Q⌈Ai /D⌉ ⊆ Q⌈N/D⌉ ,
A1 +···+An =N i=1
P
since the integer i=1 ⌈Ai /D⌉ ≥ ⌈( ni=1 Ai )/D⌉ = ⌈N/D⌉ for all n-tuples (A1 , . . . , An ) ∈ (Z≥0 )n
Pn
with i=1 Ai = N . Thus equivalently, Q(D(r−1)+1) ⊆ Qr for all r > 0 by Lemma 2.1.
Remark 2.10. We get a much stronger conclusion in Corollary 2.9 when (2) holds. This is because
we can then give a proof using Lemma 2.1 as a workaround. It is less clear what the strongest
7
conclusion to shoot for is when (1) holds. We note that if (1) holds under Corollary 2.9, then
setting D = max Di , one can alternatively prove by contradiction that
Q(n(Dr−1)+1) ⊆ Qr for all r > 0.
In part (1) of the proof above, simply adjust the claim “A
Pj ≥ Dj r for some j” to “Aj ≥ Dr for
some j.” Otherwise, some tuple satisfies n(Dr − 1) + 1 = ni=1 Ai ≤ n(Dr − 1), a contradiction.
Remark 2.11. When all hypotheses are satisfied, Corollary 2.9 typically applies to an infinite set
of prime ideals in the tensor product T . If R is a Noetherian ring of dimension at least two, or a
Noetherian ring of dimension one which has infinitely many maximal ideals, then Spec(R) is infinite;
see [2, Exercises 21.11-21.12]. Now suppose R1 , . . . , Rn are F-affine domains, with n ≥ 2 and F
algebraically
closed, at least one of which is of dimension
one or more. Then in the domain T =
N
P
( F )ni=1 Ri , the following set QED (T ) := {Q = ni=1 Pi T ∈ Spec(T ) : ∀1 ≤ i ≤ n, Pi ∈ Spec(Ri )}
is infinite.
Remark 2.12. Let R = F[x1 , . . . , xm ], S = F[y1 , . . . , yn ], and T = R⊗F S ∼
= F[x1 , . . . , xm , y1 , . . . , yn ]
be polynomial rings over a field F. Our original inspiration for Theorem 2.8 was the following
Theorem (Thm. 7.8 of Bocci et al [4]). Let I ⊆ R and J ⊆ S be squarefree monomial ideals. Let
I ′ = IT and J ′ = JT be their expansions to T . Then for any N ≥ 0, the symbolic power
(I ′ + J ′ )(N ) =
N
X
(I ′ )(N −i) (J ′ )(i) =
i=0
X
(I ′ )(A) (J ′ )(B) .
(2.0.7)
A+B=N
Ha, Nguyen, Trung, and Trung [11, Thm. 3.4] recently extended the above theorem to the case of
two nonzero ideals I ⊆ R, J ⊆ S in two Noetherian commutative F-algebras such that T = R ⊗F S
is also Noetherian. A general multinomial theorem then follows by adapting the proof of Theorem
2.8, where one containment would require the n-fold version of [11, Lem. 2.1(i)]. Combining this
multinomial expansion with the general versions of Lemma 2.1 and Proposition 2.6 we proved in
[22, Prop. 2.1, Lem. 3.3, Prop. 3.4], one can extend Corollary 2.9 to a form allowing, for instance,
any proper ideals Ii ⊆ Ri . As a final note in passing, the proof of [22, Prop. 2.1] still works up to a
tweak of multiplicative system, for those who opt to define symbolic powers of proper ideals using
only minimal associated primes as in [11], rather than using all associated primes as in [22].
3. Finale: Sample Applications to Tensor Power Domains
We begin with two results on uniform linear bounds on asymptotic growth of symbolic powers
for equicharacteristic Noetherian domains that have nice structure, but need not be regular. The
first is due to Huneke, Katz, and Validashti, while the second is due to Ajinkya A. More.
Theorem 3.1 ([15, Cor. 3.10]). Let R be an equicharacteristic Noetherian local domain such that
R is an isolated singularity. Assume that R is either essentially of finite type over a field of
characteristic zero or R has positive characteristic, is F -finite and analytically irreducible. Then
there exists an E ≥ 1 such that P (Er) ⊆ P r for all r > 0 and all prime ideals P in R.
Theorem 3.2 ([19, Thm. 4.4, Cor. 4.5], see also [6, Thm. 3.25]). Suppose R ⊆ S is a finite
extension of equicharacteristic normal domains such that: (1) S is a regular ring generated as an
R-module by n elements, and n! is invertible in S; and (2) R is either essentially of finite type over
an excellent Noetherian local ring (or over Z), or is characteristic p > 0 and F -finite. Then there
exists an E ≥ 1 such that P (Er) ⊆ P r for all r > 0 and all prime ideals P in R.
8
Remark 3.3. Suppose that R is the coordinate ring of an affine variety over any perfect field F,
whose singular locus is zero dimensional. Then in tandem with the results of Ein-Lazarsfeld-Smith
and Hochster-Huneke [8, 13], Theorem 3.1 would yield a uniform slope E for all primes in R.
In particular, this covers F[x, y, z]/(y 2 − xz), F[x, y, z, w]/(xy − zw), and more generally when R
corresponds to the affine cone over any smooth projective variety.
Indeed, the class of rings R to which Theorems 3.1-3.2 apply is large. Applying Theorem 1.2 to any
collection of two or more rings under Remark 3.3, Remark 2.11 says that we can create an infinite
set as a vantage point for data suggestive of uniform symbolic topologies in the corresponding tensor
product domain. At present, since the domain we create has non-isolated singularities, there is no
theorem in the literature affirming that the domain has uniform symbolic topologies on all primes.
We illustrate how these matters occur together in an example below.
But first, we fix an algebraically
N closed field F. If R is an F-affine domain, we use the tensor
power notation T = R⊗N = ( F )N
i=1 Ri to denote the F-affine domain obtained by tensoring
together N copies of R over F, where Ri and Rj are presented as quotients of polynomial rings in
disjoint sets of variables
PN when i < j. From Remark 2.11, we recall the following set of prime ideals
QED (T ) := {Q = i=1 Pi T ∈ Spec(T ) : each Pi ∈ Spec(Ri )}.
Example 3.4. To start, we fix an algebraically closed field F. Given integers a and d both at least
two, consider an affine hypersurface domain R = F[z1 , . . . , za ]/(Fd (z̄)) where Fd is an irreducible
homogeneous polynomial of degree d, with isolated singularity at the origin. Consider the varieties
VR = Spec(R) ⊆ Fa and V = Spec(T ) ⊆ FaN where
T = R⊗N =
F[zi,1 , . . . , zi,a : 1 ≤ i ≤ N ]
.
(Fd (zi,1 , . . . , zi,a ) : 1 ≤ i ≤ N )
Per Remark 3.3, Theorem 1.2 implies that Q(N E·r) ⊆ Qr for all r > 0 and all primes Q ∈ QED (T ).
Meanwhile, in terms of n-factor Cartesian products, the singular locus
Sing(V ) = ({0} × VR × · · · × VR ) ∪ (VR × {0} × VR × · · · × VR ) ∪ · · · ∪ (VR × · · · × VR × {0})
is equidimensional of dimension (a − 1)(N − 1). In particular, while T is not an isolated singularity
when N ≥ 2, the set QED (T ) is infinite by Remark 2.11 and provides a vantage point for witnessing
uniform linear bounds lurking for the asymptotic growth of symbolic powers of primes in T .
Remark 3.5. For pointers to results where an explicit value E ≥ 1 as in Theorems 3.1-3.2 is given
for particular examples of domains R under Remark 3.3, we invite the reader to see the recent
survey paper [6, Thm. 3.29, Cor. 3.30], along with the main results featured in the introductions to
our papers [21, 22, 23]. Section 3 or 4 in each of the latter papers typically includes remarks about
when the designated value E can be considered optimal.
Closing Remarks. Launching from Theorem 1.1 in the introduction, we have deduced a more
powerful criterion for proliferating uniform linear bounds on the growth of symbolic powers of prime
ideals (e.g., Harbourne-Huneke bounds)–Theorem 1.2. In the setting of domains of finite type over
algebraically closed fields, this criterion contributes further evidence for Huneke’s philosophy in
[14] about uniform bounds lurking throughout commutative algebra. We close with a goalpost
question that exceeds our grasp at present: Given the role of tensor products in our manuscript, do
analogues of the above criteria hold for other product constructions in commutative algebra, such
as Segre products of N-graded rings, or fiber products of toric rings?
9
References
[1] S. Akesseh. Ideal Containments Under Flat Extensions. arXiv/1512.08053
[2] A. Altman and S. Kleiman. A Term of Commutative Algebra. Worldwide Center of Mathematics LLC, Cambridge,
MA, 2014.
[3] T. Bauer, S. Di Rocco, B. Harbourne, M. Kapustka, A.L. Knutsen, W. Syzdek, T. Szemberg. A primer on Seshadri
constants. Contemporary Mathematics 496 (2009), pp. 33-70. arXiv/0810.0728
[4] C. Bocci, S. Cooper, E. Guardo, B. Harbourne, M. Janssen, U. Nagel, A. Seceleanu, A. Van Tuyl, T. Vu. The
Waldschmidt constant for squarefree monomial ideals. J. Algebr. Comb. 44, no.4, pp. 875-904. arXiv/1508.00477
[5] D.A. Cox, J.B Little, and H.K. Schenck. Toric Varieties (2011), Graduate Studies in Mathematics 124.
American Mathematical Society, Providence, RI.
[6] H. Dao, A. De Stefani, E. Grifo, C. Huneke, L. Nuñez-Betancourt. Symbolic Powers of Ideals. To appear in
Advances in Singularities and Foliations: Geometry, Topology and Applications, Springer Proceedings in Mathematics & Statistics. arXiv/1708.0301
[7] M. Dumnicki, T. Szemberg, and H. Tutaj-Gasińska. Counterexamples to the I (3) ⊆ I 2 containment. J. Algebra
393 (2013) pp.24-29. arXiv/1301.7440
[8] L. Ein, R. Lazarsfeld, and K. Smith. Uniform bounds and symbolic powers on smooth varieties. Invent. Math. 144
(2001), pp. 241-252. arXiv/0005098
[9] W. Fulton. Introduction to Toric Varieties (1993), Annals of Math. Studies 131. Princeton University Press,
Princeton, NJ.
[10] E. Grifo and C. Huneke. Symbolic powers of ideals defining F-pure and strongly F-regular rings. International
Mathematics Research Notices, rnx213, https://doi.org/10.1093/imrn/rnx213 arXiv/1702.06876
[11] H.T. Hà, H.D. Nguyen, N.V. Trung, T.N. Trung. Symbolic Powers of Sums of Ideals. arXiv/1702.01766
[12] B. Harbourne and A. Seceleanu. Containment counterexamples for ideals of various configurations of points in
Pn . J. Pure Appl. Algebra 219 (2015), no.4, pp. 1062-1072. arXiv/1306.3668
[13] M. Hochster and C. Huneke. Comparison of ordinary and symbolic powers of ideals. Invent. Math. 147 (2002),
pp. 349-369. arXiv/0211174
[14] C. Huneke. Uniform bounds in Noetherian rings. Invent. Math. 107 (1992), pp. 203-223.
[15] C. Huneke, D. Katz, and J. Validashti. Uniform equivalence of symbolic and adic topologies. Illinois J. Math. 53
(2009), no. 1, pp. 325–338.
[16] C. Huneke, D. Katz, and J. Validashti. Uniform symbolic topologies and finite extensions. J. Pure Appl. Algebra
219 (2015), no. 3, pp. 543–550.
[17] L. Ma and K. Schwede. Perfectoid multiplier/test ideals in regular rings and bounds on symbolic powers.
arXiv/1705.02300
[18] J.S. Milne. Algebraic Geometry. Version 5.22, 2013. http://www.jmilne.org/math/CourseNotes/ag.html
[19] A.A. More. Uniform bounds on symbolic powers. J. Algebra 383 (2013), pp. 29-41.
[20] The Stacks Project authors. The Stacks Project, 2016.
[21] R.M. Walker. Rational Singularities and Uniform Symbolic Topologies. Illinois J. Math. Vol. 60, no. 2, Summer
2016, pp. 541-550. arXiv/1510.02993
[22] R.M. Walker. Uniform Harbourne-Huneke Bounds via Flat Extensions. arXiv/1608.02320
[23] R.M. Walker. Uniform Symbolic Topologies in Normal Toric Rings. arXiv/1706.06576
Department of Mathematics, University of Michigan, Ann Arbor, MI, 48109
E-mail address: [email protected]
10
| 0math.AC
|
MULTIVARIATE INTEGRAL PERTURBATION
TECHNIQUES - I (THEORY)
JAN W. DASH
J. Dash Consultants
[email protected]
Submitted for publication
September 2006
Abstract
We present a quasi-analytic perturbation expansion for multivariate N dimensional Gaussian integrals. The perturbation expansion is an infinite series
of lower-dimensional integrals (one-dimensional in the simplest approximation).
This perturbative idea can also be applied to multivariate Student-t integrals. We
evaluate the perturbation expansion explicitly through 2nd order, and discuss the
convergence, including enhancement using Padé approximants. Brief comments
on potential applications in finance are given, including options, models for
credit risk and derivatives, and correlation sensitivities.
1. Introduction
The evaluation of multivariate integrals is of substantial interest in various areas
of finance including options and credit derivatives, as well as in science and
engineering. While some techniques are available in special cases, a common
procedure is brute-force numerical integration, e.g. via Monte Carlo simulation.
In this paper, we present a technique that we believe is new and may prove to
be useful. It is quasi-analytic and is based on a perturbation expansion. The
perturbation expansion gives the N -dimensional multivariate Gaussian integral
©September 2006, Jan W. Dash
2
as an infinite series of low-dimensional integrals, which in the simplest case are
just one-dimensional integrals. The idea is applicable also to Student-t integrals,
and probably others as well. Possibly the expansion, with a clever choice of the
initial term, will turn out to provide an interesting and viable numerical approach.
This paper discusses theory. Subsequent papers will report on numerical aspects.
The initial “point” about which the expansion is performed is key, as it
always is for perturbation expansions. In this work, this “point” is constructed
from an approximation to the original correlation matrix involving a factorized
expression (one factor), or of sums of such factorized expressions (several
factors)i. In this paper, we focus on the one-factor approximation. The
perturbation expansion is then found in terms of expectation values of {powers of
[the difference of (the inverse of the original correlation matrix) and the (inverse
of the approximate correlation matrix)]}. These expectation values are with
respect to the approximate probability density function, and are readily obtained
analytically. We give explicit expressions for Gaussian multivariate integrals of
arbitrary dimension using one factor up to second order in the expansion. To 0th
order, there is a single one-dimensional integral. In first order, there are N 2
additional one-dimensional integrals. In second order, there are N 4 additional
one-dimensional integrals. All integrals can be grouped into classes of similar
appearance, so only a few integrals need to be programmed.
In Sect. 2, we treat the perturbation expansion of multivariate Gaussian
integrals for one factor. Sect. 3 has an outline of the procedure for multiple
factors. Sect. 4 contains the perturbation expansion of multivariate Student-t
integrals. Sect. 5 has some details of the formalism for Gaussian perturbation
theory. Sect. 6 deals with the 1st-order perturbation results, and Sect. 7 has the
2nd-order perturbation results. Sect. 8 presents the cluster decomposition
diagrammatic notation. Sect. 9 discusses the approximate correlation matrix
needed to start the perturbation analysis. Sections 10 and 11 briefly discuss
potential applications to options and to credit risk and derivatives. Sect. 12
discusses correlation sensitivity. Sect. 13 discusses convergence, and Sect. 14
discusses enhanced convergence using Padé approximants. Sect. 15 has a logical
“flow chart” of the steps needed for numerical investigations.
2. Perturbation Expansion of Multivariate Gaussian Integrals
Consider the N -dimensional Gaussian multivariate integral I N with positivedefinite correlation matrix with inverse matrix 1 , determinant :
x1max
;
I N x
max
i
1
2
xNmax
...
exp 12 x T 1 x D x
©September 2006, Jan W. Dash
(2.1)
3
Here, x xi is the vector of variables,
x with
max
i
upper limits assumed constant (i.e. independent of the
N
i 1...N are the
xi
variables), and
D x dxi / 2 . Matrix multiplication is understood in the exponent.
i 1
Now consider the one-factor approximation f to the correlation matrix. We
introduce numbers ci and write for the i j matrix elements of f :
f ij ci c j
(2.2)
It is easy to prove (see Curnow and Dunnett (Ref. i) and Sect. 5) that if we
were to replace by f , then I N would become a one-dimensional integral
;
I N x
max
i
f
f
d
2
exp 12 2
1
x1max
2
xNmax
...
exp 12 x T f 1 x D x
N
(2.3)
N
max
i
i 1
(2.4)
Here is an auxiliary variable. There is one such variable (i.e. one factor),
corresponding to the single factorized term ci c j for f ij . Also N is the
standard one-dimensional cumulative normal, and
imax
ximax ci
si
(2.5)
Here si 1 ci2 needs to be positive to reproduce I N 1 as xi .
The single integral in Eq. (2.4) can be done straightforwardly using good
uniform algebraic approximations for the normal integral N ii.
The perturbation expansion follows from the trivial identity
1 f 1 1 f 1
©September 2006, Jan W. Dash
(2.6)
4
inserted into the original integral. We keep the explicit f dependence in the
exponent as in Eq. (2.3) for I N ximax ; f , and we expand the rest of the
exponential in the integrand in a power series in the matrix , defined as
1 f 1
(2.7)
So we get
I N ximax ; I N( ) ximax ; f ;
0
(2.8)
where
I N( ) ximax ; f ;
1
x1max
xNmax
...
1
12 *
!
exp x f x
1
2
1
T
x
T
x
(2.9)
Dx
If we include all the terms in the sum, the result must only depend on the original
correlation matrix , with all dependence on f cancelling out.
Using standard functional techniques described in Sect. 5, we can evaluate
for given as a sum of one-dimensional integrals. For 0 we get the
I
( )
N
same result up to normalization as the result for I N ximax ; f , namely the
single one-dimensional integral
I
( 0)
N
ximax ; f ; J f , d exp 12 2
2
N
N
i 1
max
i
(2.10)
Here the normalization is J f ,
f
.
For 1 we get one-dimensional integrals of the form
©September 2006, Jan W. Dash
5
d
I N( 1) ximax ; f ; 12 J f ,
exp 12 2
2
N
i , j 1
ij
Gij
N
N
l i , j ; l 1
max
l
*
(2.11)
Here, the various Gij are functions of that are different according to
whether i j or i j (see Sect. 6). The perturbation matrix element is
ij 1 f 1 . Note that I N( 1) has an explicit minus sign. There are N 2
ij
of these two types of integrals.
For 2 we get one-dimensional integrals of the form (see Sect. 7)
I
( 2)
N
ximax ; f ; 18 J f , d exp 12 2
2
N
i , j , k ,l 1
*
(2.12)
ij kl G(ijkl ) in C
()
Here is a “class label” corresponding to the seven different classes C of
indices i, j , k , l that arise when the indices are all distinct, have one pair equal,
two pairs equal, three values equal, or four values equal. The classes can be
conveniently associated with diagrams, in a sort of “cluster decomposition” iii, 1.
)
The various functions G((ijkl
are categorized using the class label . There
) in C
are a total of N 4 of these integrals, of five different types. Because the indexing
)
is complicated, products of normal functions are lumped into G((ijkl
.
) in C
3. Multiple Factors for the Gaussian Perturbation Expansion
Using multiple-factors means we take approximated by f of the form
K
fij ci c j for i j
(3.1)
1
1
Cluster Decomposition: The cluster decomposition in the present paper is similar in
spirit but different in detail from the cluster decomposition for multivariate Gaussian
integrals that I envisioned in the late 1980’s (see p. 623 of Ref. iii).
©September 2006, Jan W. Dash
6
with 1...K . This leads to K -dimensional integrals using K auxiliary
variables that replace the one-dimensional integrals. Perturbation theory
is performed along the same lines. The tradeoff is between the extra complexity
of multidimensional integrals versus a better initial description of the correlation
matrix. If K N and with principal components to get f as explained in
Sect. 9), we restore the original problem, and nothing is accomplished.
We shall, however, only give results for perturbations about a one-factor
approximation in this paper.
4. Perturbation Expansion of Multivariate Student-t Integrals
We next outline perturbation theory for multivariate Student-t integrals. This
produces two-dimensional integrals instead of one-dimensional integrals, in the
one-factor approximation to f . The basic idea is to use an integral
representation in order to rewrite things in exponential form, so that we can again
use perturbation theory similarly to the Gaussian case.
The multivariate Student-t (“St”) cumulative probability distribution in N
dimensions with degrees of freedom isiv
I
( St )
N
; ;
x
max
i
K 1
x1max
1
2
with K 1 12 N 12
N
2
1 1 x T 1 x
N / 2
xNmax
...
1 1 x
T
1 x
N / 2
Dx
(4.1)
. We use the identity
1
2
1
K 2 y N 1 exp 12 1 1 x T 1 x y 2 dy (4.2)
0
1
with K 2 12 N 2 N 2 / 2 . The integral over y is an extra
complication for the Student-t distribution. We make the change of variables
ui yxi
and define uimax y yximax
. We obtain
©September 2006, Jan W. Dash
7
I
( St )
N
x
max
i
; ;
u1max y
max
uN
y
Here Du
...
du
N
i 1
i
1 12
2
1
2
1
2
1 2
1 2 y
y
e
*
0
(4.3)
exp 12 u T 1u Du dy
/ 2 . We now have a Gaussian form and thus can use the
same perturbation procedure in Sect. I. We expand in powers of 1 f 1
and perform the integrals over the u variables with respect to the measure with
exp 12 u T f 1u . Assuming a one factor approximation for f , this again
yields one-dimensional integrals over , which along with the integral over y ,
gives the perturbative result for I N( St ) ximax ; ; in terms of sums of two-
dimensional integrals. Explicitly,
I
( St )
N
1 2
yximax
1 2
1 2 y
ximax ; ; 2
y
e
I
; dy
N
1
2
0
1
(4.4)
yximax
; is the Gaussian multivariate Eq. (2.1) with changed
Here I N
upper limits for fixed y , namely ximax yximax
y ;
I N u
max
i
12
u1max y
max
uN
y
...
uimax y . We have
exp 12 u T 1u Du
(4.5)
We can now use the results from the previous analysis for multivariate
Gaussian perturbation theory, inserted into Eq. (4.4), to get the perturbation
expansion for the multivariate Student-t integral, including the y integral.
For large , it is well known that the Student-t approaches the Gaussian.
Using WKB around y* eliminates the y integral, and yields the result:
I N( St ) ximax ; ;
WKB
Large
I N ximax ;
©September 2006, Jan W. Dash
(4.6)
8
5. Formalism for Gaussian Perturbation Theory
This section discusses some details of the formalism for Gaussian perturbation
theory. We first show how the extra variable arises. Start with Eq. (2.3) and
note the identity
exp x f x
1
T
1
2
2
d
exp
2
1
2
2
xi ci 2
(5.1)
exp 12
si2
i 1
N
Here
N
2 1
l 1
cl2
sl2
(5.2)
This identity introduces the extra variable . Now we use Eq. (5.1) in Eq.(2.3),
and change variables, introducing i to eliminate the xi :
i
xi ci
or xi ci si i
si
(5.3)
The integral in Eq. (2.3) becomes
;
I N x
max
i
f
d
2
exp
1
2
2
i d
i
1 2 (5.4)
exp
2 2 i
i 1
max
N
where imax is in Eq. (2.5). This immediately yields Eq. (2.4).
Consider the average F
of any function F xi with respect to the
f
approximation measure with f in Eq. (2.3). This is:
F
f
f
1
2
x1max
xNmax
...
F xi exp 12 x T f 1 x D x
(5.5)
Such integrals can be handled using a standard functional technique. We
introduce “currents” J i and functional derivatives J i . Making the same
change of variables in Eq. (5.3), we get
©September 2006, Jan W. Dash
9
exp 12 2 F ci si
*
J i
2
F
f
d
(5.6)
i d
i
exp 12 i2 i J i J 0
i
2
i 1
max
N
upper limit parameters, we get
Restoring the original ximax
exp 12 2 F ci si
*
J i
2
N
ximax ci
2
1
exp(
J
)
N
J
i J 0
2 i
i
si
i 1
F
f
d
(5.7)
For the problem at hand, we set
1
12 x T x
0 !
F xi
(5.8)
Writing out the components of the matrices, we have
x
T
x
N
i , j 1
1...
xi i
1
j
x j
We make the replacement xi ci si J i , evaluate the
(5.9)
J i derivatives
and then set J i 0. Sects. 6 and 7 contain the results for 2 .
It is useful to have an analytic inverse for the one-factor matrix f . This isv
f
1
ii
1
f
ij
1
si2
ci2
1
2 2
si
ci c j
si2 s 2j 2
for i j
©September 2006, Jan W. Dash
(5.10)
(5.11)
10
where 2 is given in Eq. (5.2). For multiple factors, the inverse of f is
complicated, and we have been unable to find an analytic inverse.
6. Details of the 1 (1st order) terms
In this section, we give the 1 functions Gij in the text. First, define
i N imax
and
1
i
exp 12 imax
2
(6.1)
2
(6.2)
Also define
wi(1) ci i si i
(6.3)
and
wi(2) ci2 2 si2 i 2ci si si2 imax i
(6.4)
Then for equal indices i j
Gii wi(2)
(6.5)
and for unequal indices i j
Gij
i j
wi(1) w(1)
j
(6.6)
7. Details of the 2 (2nd order) terms
In this Section, we discuss the 2nd-order 2 terms. We first define wi(3)
and wi(4) :
©September 2006, Jan W. Dash
11
wi(3) ci3 3 3ci si2 i
2
3ci2 si 2 3ci si2 imax si3 imax 2 i
(7.1)
wi(4) ci4 4 6ci2 si2 2 3si4 i
4c 3 s 3 6c 2 s 2 2 max 4c s 3 max 2 2
i i i
i i
i
i i
i
2
4 max
max
si i i 3
(7.2)
The seven classes of terms are defined as follows corresponding to the indices
(i, j , k , l ) in the outer product of the two matrices with elements ij kl . Two
of them are degenerate, so there are really only five independent classes.
Class 1 : No two indices are equal. There are N ( N 1)( N 2)( N 3) such
terms for N 4 . An example is 12 34 . The corresponding function is
( 1)
G1234
w1(1) w2(1) w3(1) w4(1)
N
m . In general,
m 1,2,3,4
( 1)
Gijkl
wi(1) w(1)j wk(1) wl(1)
N
m i , j ,k ,l
m
(7.3)
Class 2 : One pair of indices is equal in one matrix; the others are unequal
to each other and unequal to the pair. There are 2 N ( N 1)( N 2) such terms for
N 3 . An example is 11 23 . The corresponding function is
( 2)
G1123
w1(2) w2(1) w3(1)
N
. In general,
m 1,2,3
( 2)
Giijk
wi(2) w(1)j wk(1)
m
N
m i , j ,k
m
(7.4)
Class 3 : One pair of indices is crosswise equal between the two matrices;
the others are unequal to each other and unequal to the pair. There are
©September 2006, Jan W. Dash
12
4 N ( N 1)( N 2)
such terms for
N 3 . An example is 12 13 . The
( 3)
( 2)
corresponding function is degenerate, G1213
G1123
. In general,
( 3)
Gijik
Giijk(2)
(7.5)
Class 4 : Each pair of indices on each matrix are equal, but distinct. There
are N ( N 1) such terms for N 2 . An examples is 11 22 . The corresponding
function is
N
( 4)
G1122
w1(2) w2(2) m , or in general
m 1,2
N
( 4)
Giijj
wi(2) w(2)
j m
(7.6)
m i , j
Class 5 : Two pairs of indices are crosswise equal, but distinct. There are
2 N ( N 1) such terms for N 2 . An example is 12 12 . The corresponding
( 5)
( 4)
function is degenerate, G1212
G1122
. In general,
( 5)
Gijij
Giijj(4)
(7.7)
Class 6 : Three indices are equal, different from the fourth. An example is
1112 . There are 4 N ( N 1) such terms for N 2 . An example is 1112 . The
( 6)
corresponding function is G1112
w1(3) w2(1)
N
. In general,
m 1,2
m
N
( 6)
Giiij
wi(3) w(1)j m
(7.8)
m i , j
Class 7 : All indices are equal. There are N terms. An example is 1111 .
( 7)
The corresponding function is G1111
w1(4)
N
. In general
m 1
m
N
( 7)
Giiii
wi(4) m
(7.9)
m i
The total number of terms from all classes is N 4 , and as is clear from the
above discussion there are five types of integrals with different parameters.
©September 2006, Jan W. Dash
13
8. Cluster Decomposition Diagrammatic Notation
The cluster decomposition is a useful device that is a visual mnemonic for the
various types of terms (cf. Ref. iii, p.700). Each index is represented by a line
going from left to right. Each pair of lines starting at the left and going from top
to bottom corresponds to indices i, j on one of the ij matrices. Consider Fig. 1
below. If two indices are equal, we make the lines enter and leave a “bubble”,
like the picture on the left. If two indices are unequal, they will not go into the
same “bubble”, like the picture on the right.
Indices
unequal
Indices
equal
Fig. 1: 1st-order
diagrams
The seven classes in the 2nd order perturbation term with two ij kl matrices
with four indices (so, four lines) can be drawn as in Fig. 2 below.
©September 2006, Jan W. Dash
14
Class 1
Class 2
Class 4
Class 3
Class 5
Class 6
Class 7
Fig. 2: 2nd-order
diagrams
9. Choice of the parameters in the matrix f
The choice of the ci parameters in the K -factor approximation matrix f is
arbitrary. Of course the perturbation results to some given order will depend on
the choice. One idea is to use a principal component (PC) decomposition of the
original correlation matrix , keep the first K of the PCs, and thereby identify
ci . The PC decomposition of the
NxN matrix is a sum of N factorized
terms, and is an identity. It reads
©September 2006, Jan W. Dash
15
N
ij i j
(9.1)
1
Here and are the -th eigenfunction and eigenvalue of . This
is already a sum of factorized terms in the matrix indices. Hence, truncating the
sum at the K th term yields a K - factor approximation f to . This is exactly
the procedure used in VAR calculations using SVD to get rid of N K negative
eigenvalues, setting them to zero. We need to renormalize the eigenfunctions so
that f will have unit diagonal elements (cf. p. 327 of Ref. iii). So we replace
i
by i
i
,
i
1
2
K
2
where i i , and then write
1
K
K
1
1
fij ci c j i j
(9.2)
Identifying terms with , we then can set ci
i
. If is
positive definite with N positive eigenvalues 0 , then the NxN matrix
f is positive semi-definite with K positive eigenvalues and N K zero
eigenvalues.
For one factor, K 1 , this procedure just produces fij 1 , which is not too
useful. However, we could really assume any constant fij 1 N 1 for
i j , maintaining f positive definiteness. To do this, we can look for a best fit
to for the constant approximation matrix f with fij c 2 for i j . Hence,
write
2
c
2
N
i j 1
2
ij
and set 2 / c 2 0 . We get
1
2
c 1
abs il
N ( N 1)
i ,l i
(9.3)
Here, “ abs ” means absolute value; we include this so that the expression always
makes sense.
©September 2006, Jan W. Dash
16
A generalization to a nonconstant matrix f can be done by pulling out the
sum over the index i in Eq. (9.3), and then identifying ci for given i as
1
2
ci 1 abs il sgn il
l i
l i
( N 1)
(9.4)
Here, “ sgn ” means sign, and is included for generality so that f can have
matrix elements with signs. While this procedure may seem arbitrary, again recall
that we are free to choose the approximate correlation matrix f any way we
want. With Eq. (9.4), we get the final result for the one factor approximate matrix
elements i j that we propose to use,
1
f ij
1
2
2
1
abs ik abs lj sgn ik sgn lj
( N 1)
k i
k i
l j
l j
(9.5)
Visually, this amounts to approximating the original correlation matrix element
ij by a geometric average of signed averages over the matrix elements of the
row and column whose intersection contains ij . Note that the approximation
can have either positive or negative sign. Since ci is obtained using an average,
the extent to which f approximates will depend on the internal differences
of matrix elements of , with relatively constant matrix elements being more
tractable.
A least-squares approach for getting the ci may also be useful. These
equations are nonlinear, and need to be solved numerically.
Regardless, the requirement that f is positive definite needs to be checked
explicitly.
10. Applications to Options
The perturbation expansion has potential application to options that are
dependent on several variables. For simplicity, we restrict the discussion to an
illustrative example.
©September 2006, Jan W. Dash
17
Consider the expectation of a “payoff” function C x of the variables xi
remain constants, and not interlocked , independent
2
such that the limits ximax
of the variables xi . We can use either the Gaussian or Student-t multivariate
distribution, depending on the correlation matrix . Carrying out the same
perturbation procedure as above (for which C x 1 ) leads to a perturbation
series for this expectation of
C x using the
f -dependent approximate
measure.
Recall the general nature of the expectation F
with respect to the f f
dependent approximate measure in Eq. (5.6). We now want to use
1
12 x T x
0 !
F xi C x
(10.1)
We go through the same procedure as before. Similarly to Sect. 6 and 7, the
functions for a given approximation at 0,1, 2 etc. need to be evaluated
including the dependence on the payoff function C x .
11. Application to Credit Risk and Credit Derivatives
The perturbation expansion developed in this paper can be potentially useful for
analysis for credit risk and credit derivativesvi.
For buy-and-hold portfolios, the determination of credit risk obtained using
historical correlations between asset returns in a “structural model” can be
appropriate. Analysis with these large correlation matrices could potentially be
numerically handled using the perturbation methods in this paper.
For trading portfolios, “reduced-form models” that calibrate to market CDS,
CDO prices and other data are appropriate. If the reduced-form model has several
factors, we could then use the one-factor perturbation expansion to provide a
potentially useful evaluation tool for credit derivatives.
2
Options: This assumption is true for some, but not all, options. Some options with
min/max conditions that interlock the limits through dependence on the {xi} can be
handled by changing variables in the integrals, before applying perturbation theory.
©September 2006, Jan W. Dash
18
12. Correlation Sensitivity
We get straightforward approximations for correlation sensitivity using the
perturbation expansion that may prove useful. Consider a function C of the
correlation matrix. Suppose the correlation matrix changes from 1 to
2 . The resultant change in the function, i.e. the correlation sensitivity, is
C C 2 C 1
(12.1)
Suppose further that we choose the approximate matrix f2 for 2 as
unchanged from the approximate matrix f1 of 1 , that is f2 f1 f .
Then subtracting the two perturbation expansions we get an approximation to the
correlation sensitivity. In first order 1 , the f dependence cancels out
except for the approximate measure.
13. Convergence of the Perturbation Series
In this Section we discuss convergence of the perturbation series, although we
have not constructed a formal proof. There are two main points. First, as a
!
1
function of the order , the exponential expansion contains a factor
,
which eventually overwhelms powers Q for any Q as a function of .
Second, rapid decrease of the integrand at large values of exists, which means
that the integrals are effectively finite, so the interchange of -summation and
-integration is allowed.
Some intuition can be gained by looking at the 0th-order approximation.
max
Define average parameters cavg and xavg
(the details will not matter). The
dependence of the integrand is essentially of the form exp where
12 2 N log N
max
xavg
cavg
savg
©September 2006, Jan W. Dash
(13.1)
19
is the dimension and N
Again, N
is the normal integral. We
differentiate Eqn. (13.1) and set ' * 0 at the point * where the
magnitude of the integrand is largest. The equation to determine * is
*
Ncavg
savg
max
xavg
cavg *
N '
N
savg
max
xavg
cavg *
savg
Here, the derivative of the normal integral is N '
1
(13.2)
. The 0th-order integrand has
a pronounced bump at * , falling off rapidly on either side. Similar remarks
hold for the 1st and 2nd order integrands, proportional to N and N ' , although
due to cancellations between terms there may be some additional structure in
besides a simple bump. As N increases, it would appear that * will increase
too, but the other factors can decrease to make * relatively constant in N .
Given the above bump structure, it is intriguing to speculate that a WKB
approximation might be useful as a rough guide, without doing any integrations.
It is also instructive to look at the dominant order * at fixed dimension N .
Arguments similar to the above yield a rough estimate for * as
* 12 N 2 avg
(13.3)
where avg is an average matrix element of the matrix. Finally, at a given
order , the most important dimension N * is roughly given by
N*
2
(13.4)
ln 1/ N avg
where N avg is a normal integral at * with average parameters as above.
Metrics for Convergence
The theory in this paper is exact if all orders in perturbation theory are utilized.
However, the whole idea is to stop at some manageable point, e.g. 2nd order.
Hence there will be some error. We need to characterize the error to understand
when the approximation can be useful. To this end, we define some metrics; the
numerical error will be a function of these metrics.
©September 2006, Jan W. Dash
20
Since we are using a one-factor approximation with a limited number of
variables ci , if the correlation matrix elements are approximately constant the
approximation will be good, and the amount to which the correlation matrix
elements are highly variable will limit the utility of the approximation. Hence, we
anticipate that the internal correlation variance 2 ,Int of will be a useful
metric, where
2 ,Int 12 N N 1 1
1
2
N
i j
ij
avg
(13.5)
Here, avg is the average of the off-diagonal matrix elements of . In particular,
since the approximate matrix f is obtained through averaging over elements of
, its internal variance 2 f ,Int will be smaller than 2,Int . We also anticipate
that 2,Int , the internal variance of the expansion matrix 1 f 1 , will
be useful.
If is close to being singular (i.e. has one or more very small eigenvalues),
then 1 will have large matrix elements, but f 1 will not. Therefore
1 f 1 will have large elements and convergence will be less rapid.
Cutoffs on the minimum eigenvalue may have to be applied in some cases,
Min N
(13.6)
That is, the correlation matrix may need to be regularized somewhat in order to
apply this method. Naturally, the correlations in the regularized matrix will be
changed somewhat, which is a drawback. On the other hand, it is known that
correlations in practice are highly unstable with large uncertaintiesiii, so the
change due to regularization may be small compared to these uncertainties. In
any case, this procedure must be consistent with the experimental uncertainties of
the correlation matrix elements. We anticipate that the cutoff will depend on the
dimension N .
A useful metric for the correlation matrix is the distance of the matrix from a
singular matrix, which we call R N . We suggest R N be defined as
R N
1
(13.7)
1
1
N
©September 2006, Jan W. Dash
21
Eqn. (13.7) is formally the same as the total resistance of N resistors in parallel,
with the resistances taken as . If some 0 , then R N 0 and
the matrix becomes singular. For example, if all correlation matrix elements are
equal to one, then all but one eigenvalue are equal to zero. A zero eigenvalue in
the resistor circuit analogy corresponds to a “short circuit”.
The maximum distance of the matrix from the singular case occurs when all
N
eigenvalues are equal. Since we must have
N , we get
1 , and
1
so Max R N . Note that the distance from a singular matrix decreases
as the dimension increases.
1
N
14. Enhancing the Convergence with Padé Approximants
In this Section we discuss enhancing the convergence of the perturbation series
using Padé approximants. This is a common procedure in scientific applications,
but it may not be so familiar in finance, and so we shall spend a little time giving
some background. We shall also use the idea in a somewhat extended fashion.
A Padé approximant is defined as a rational function that agrees with the
perturbation expansion of a function to a given order. Padé approximantsvii can
enhance numerical convergence. In some cases (including applications of this
theory), Padé approximants may be more than numerically helpful; they may be
necessary.
For small values of matrix elements ij of 1 f 1 in which the
perturbation is being carried out, the perturbation expansion will converge
rapidly, and no problems are anticipated. However, for larger values of ij , the
convergence will be less rapid, or even appear to diverge at some finite order. We
have calculated through 2nd order. We claim that the perturbation series
converges. Even if the perturbation series were to diverge, for example as an
asymptotic series, the Padé approximants provide a method of summing the
series.
Consider the simple case with equal matrix elements 0 (i.e. ij 0 for
i j ), so the perturbation expansion is in 0 . Simplifying the notation, we have
for the th order perturbative term with 0,1, 2 . The series
I O 0
( )
expansion of the Padé approximant I Pade
matches the sum of the perturbative
©September 2006, Jan W. Dash
22
( )
terms to order , viz I Pade
I O
0
1
0
. As
increases, the match to
the perturbation series becomes more exact. The expansion can be carried out at
small values of 0 , and then the Padé approximants can be analytically continued
to large values of 0 . The Padé approximants effectively provide approximations
to arbitrarily high-order perturbative terms without explicitly calculating them.
Again, the Padé results can be numerically better than the perturbation expansion.
Padé approximants are labeled by L, M or L / M , where L and M are the
orders of the numerator and denominator.
( )
So, we define successive Padé approximants I Pade
of 0th, 1st, and 2nd order as:
(0)
I Pade
I 0
I 1
(1)
I Pade
I 0 1 0
I
I
(2)
Pade (1,1)
(14.1)
1
(14.2)
I 0I 2 I 2
I 0 I 1 1 1 1
I
I
1
(14.3)
The above 2nd-order Padé approximant is of 1/1 type. Another 2nd-order
approximant, of 0 / 2 type, is defined as
I 1 I 2 I 1 2
0
(2)
I Pade (0,2) I 1 0 1 0
I
I
I
1
(14.4)
Often the diagonal approximants L / L are used, but off-diagonal approximants
(1)
can be useful. Note that I Pade
has type 0 /1.
We will use these same Padé approximant formulae for the real case of nonconstant ij matrix elements, dropping the restriction of constant 0 . This
procedure goes beyond the usual Padé theory. However, the assumption is simple
to write down, and it can be tested numerically against direct integration.
An additional acceleration device may turn out to be useful. We can try to
parametrically and approximately extrapolate the Padé approximants to ,
i.e. infinite order in , thereby obtaining I for the final approximation to the
original multivariate Gaussian integral, Eqn. (2.1). Consider the ansatz
©September 2006, Jan W. Dash
23
( )
I Pade
I 0 I e g I
(14.5)
Here is a parameter and g is some reasonable function, with g 0 1 so
(0)
that I Pade
I 0 . We might choose g cos [or g 1 ] for an
oscillating [non-oscillating] Padé sequence, which is known at 0,1, 2 . We
(1)
(2)
know I Pade
and I Pade
(see footnote3). Hence there are 2 equations and 2
unknowns, and hence we can get the final result I by determining .
15. Numerical Aspects, Logical Flow Chart
Numerical aspects of the theory presented in this paper are currently being
investigated, including error analysis as a function of parameters, and will be
reported separately viii. Preliminary results are encouraging, but also indicate that
care is required in applying the theory, including Padé approximants.
Here is a logical “flow chart” with explicit steps in order to obtain the
Gaussian I N ximax ; f in Eqn. (2.1), using the 1-factor approximation.
1.
Obtain the N -dimensional correlation matrix . Check that is
positive definite (PD). It may be necessary to impose an eigenvalue
cutoff Min N to make the 2nd-order approximation
“accurate enough”. Get the inverse 1 and determinant .
2. Get
ci via Eq. (9.4), or least squares, from
; write si 1 ci2 .
3. Get the one-factor approximation f to in Eqn. (9.5). Check that f
is PD. Calculate the inverse matrix f 1 in Eqns. (5.10) - (5.11), the
determinant f , and J f ,
f
.
4. Get 1 f 1 (subtract the matrices element by element).
5. Specify the upper limit values ximax .
Which 2nd-order Padé? It may be useful to take the maximum or the average of the
two 2nd-order Padé approximants.
3
©September 2006, Jan W. Dash
24
6. For each in the 1-factor integration , where is a
“sufficiently big number”, calculate for each i 1...N :
imax in Eqn. (2.5)
i N imax using a normal integral approximationii
i , wi(1) , wi(2) in Eqns. (6.2) - (6.4)
Gij in Eqns. (6.5) - (6.6) for the 1st order contribution
I N( 1) ximax ; f ;
, wi(4) in Eqns. (7.1) - (7.2)
w
)
in Eqns. (7.1) - (7.9) for the 2nd order contribution
G((ijkl
) in C
(3)
i
I N( 2) ximax ; f ;
7. Calculate 0th order contribution I N( 0) in Eqn. (2.10).
8. Calculate 1st order contribution I N( 1) in Eqn. (2.11) and obtain the 1storder approximation I N(1st Order Approx.) I N( 0) I N( 1) .
9. Calculate 2nd order contribution I N( 2) in Eqn. (2.12) and obtain the 2ndorder approximation I N(2 nd Order Approx.) I N( 0) I N( 1) I N( 2) .
( )
10. Calculate the 1, 2 Padé approximants I Pade
in Eqns. (14.2) - (14.4).
11. Obtain the extrapolation of the Padé approximants, I , as
using Eqn. (14.5). I is the suggested final result.
12. Use the internal correlation variance 2 ,Int and the distance from a
singular matrix R N from Sect. 13 as metrics for application analysis.
Acknowledgements
I thank Kris Kumar for discussions regarding numerical aspects and potential
applications of the theory.
References
©September 2006, Jan W. Dash
25
[i] Curnow, R. and Dunnett, C., The Numerical Evaluation of Certain Multivariate
Normal Integrals. Ann. Math. Statist., 1962, Pp. 571-579, and references therein.
[ii] Abramowitz, M., Stegun, I., Handbook of Mathematical Functions. National Bureau
of Standards, Applied Mathematics Series #55, 1964. See Sect. 26.2.17.
[iii] Dash, J. W., Quantitative Finance and Risk Management, A Physicist’s Approach,
World Scientific, 2nd printing 2005, ISBN 9812387129.
[iv] Hodoshima, J., The Effects of Nonnormality on the Market Model in the Class of
Elliptical Distributions, Nagoya City U. preprint, March 2001, p. 7.
[v] Chern, S. S. and Shen, Z., Riemann-Finsler Geometry, World Scientific, Nankai
Tracts in Mathematics, Vol. 6, 2005, Pp. 4,5.
[vi] Duffie, D. and Singleton, K. J., Credit Risk: Pricing, Measurement, and Management,
Princeton U. Press, 2003; Smithson, C. W., Credit Portfolio Management, Wiley
Finance, 2003
[vii ] Kleinert, H., Path Integrals in Quantum Mechanics, Statistics and Polymer Physics,
2nd Edition, World Scientific 1995, p 611; Honerkamp, J., Statistical Physics – An
Advanced Approach with Applications, Springer-Verlag, 1998, p 91; Wikipedia
http://en.wikipedia.org/wiki/Pade_approximant; Baker, G. et al, Padé Approximants,
Encyclopedia of Mathematics and its Applications, 2nd Ed, Cambridge U. Press, 1996.
[viii] Dash, J. W. and Kumar, K., Multivariate Integral Perturbative Techniques - II
(Numerical Aspects), to appear.
File =
DASH Multivariate Integral Perturbative Techniques I Theory Sept06 R2 posted.doc
Last accessed: 10/13/2006, 7:46 AM
©September 2006, Jan W. Dash
| 5cs.CE
|
arXiv:1509.02629v2 [quant-ph] 22 Sep 2016
Error Bounds for Finite-Dimensional Approximations
of Input-Output Open Quantum Systems by Subspace
Truncation and Adiabatic Elimination∗
Onvaree Techakesari and Hendra I. Nurdin
†
February 22, 2018
Abstract
An important class of physical systems that are of interest in practice are inputoutput open quantum systems that can be described by quantum stochastic differential equations and defined on an infinite-dimensional underlying Hilbert space.
Most commonly, these systems involve coupling to a quantum harmonic oscillator
as a system component. This paper is concerned with error bounds in the finitedimensional approximations of input-output open quantum systems defined on an
infinite-dimensional Hilbert space. We develop a framework for developing error
bounds between the time evolution of the state of a class of infinite-dimensional
quantum systems and its approximation on a finite-dimensional subspace of the original, when both are initialized in the latter subspace. This framework is then applied
to two approaches for obtaining finite-dimensional approximations: subspace truncation and adiabatic elimination. Applications of the bounds to some physical examples
drawn from the literature are provided to illustrate our results.
Keywords: Quantum stochastic differential equations, input-output open quantum systems,
finite-dimensional approximations, error bounds, approximation errors
1
Introduction
Quantum stochastic differential equations (QSDEs) developed independently by Hudson
and Parthasarathy [1] and Gardiner and Collett [2] (the latter in a less general form than
the former) have been widely used to describe the input-output models of physical open
markov quantum systems [3–5]. Such models describe the evolution of Markovian quantum
∗
Research supported by the Australian Research Council
O. Techakesari and H. I. Nurdin are with the school of Electrical Engineering and Telecommunications, UNSW Australia, Sydney NSW 2052, Australia. Email: [email protected] and
[email protected]
†
1
systems interacting with a propagating quantum field, such as a quantum optical field,
and are frequently encountered in quantum optics, optomechanics, and related fields. An
example in quantum optics would be a cavity QED (quantum electrodynamics) system
where a single atom is trapped inside an optical cavity that interacts with an external
coherent laser beam impinging on the optical cavity. These input-output models have
subsequently played an important role in the modern development of quantum filtering
and quantum feedback control theory [6, 7]. Many types of quantum feedback controllers
have been proposed in the literature on the basis of QSDEs, using both measurementbased quantum feedback control, e.g., [3,4,6,7], and coherent feedback control, e.g., [8–10].
Besides, QSDEs have also been applied in various developments in quantum information
processing, such as in quantum computation technology; e.g., see [11].
In various physical systems of interest, one often deals with input-output systems that
include coupling to a quantum harmonic oscillator. For instance, typical superconducting
circuits that are of interest for quantum information processing consist of artificial two-level
atoms coupled to a transmission line resonator. The former is typically described using a
finite-dimensional Hilbert space and the latter is a quantum harmonic oscillator with an
infinite-dimensional underlying Hilbert space (i.e., L2 (R), the space of square-integrable
complex-valued functions on the real line). Another example is a proposed photonic realization of classical logic based on Kerr nonlinear optical cavities in [12], which is built around
a quantum harmonic oscillator with a Kerr nonlinear medium inside it. If a mathematical model for such quantum devices is sufficiently simple, it is often possible to simulate
the dynamics of the system on a digital computer to assess the predicted performance of
the actual device, as carried out in [12]. The simulation carried out is typically that of a
stochastic master equation that simulates the stochastic dynamics of a quantum system
when one of its output is observed via laboratory procedures such as homodyne detection
or photon counting, see [3, 6, 13]. However, since it is not possible to faithfully simulate a
quantum system with an infinite-dimensional Hilbert space, often in simulations this space
is truncated to some finite-dimensional subspace. Two approaches that are often employed
to approximate a quantum system, defined on an infinite-dimensional space, are subspace
truncation approximation and adiabatic elimination (also known as singular perturbation).
Subspace truncation approximation is applied to eliminate higher dimensions of the original infinite-dimensional Hilbert space. An operator X on the infinite-dimensional space is
approximated by a truncated operator of the form P XP , where P denotes an orthogonal
projection projector onto the approximate finite-dimensional subspace. For instance, with
quantum harmonic oscillators, a commonly used finite-dimensional space is the span of a
finite number of Fock states |0i, |1i, . . . , |ni. On the other hand, adiabatic elimination is
often used to simplify quantum systems comprising components that evolve at multiple
well-separated time-scales. In this approach, the faster variables are eliminated from the
mathematical model description of the systems.
Despite the ubiquity of approximating infinite-dimensional Hilbert spaces of quantum
systems by finite-dimensional subspaces for simulations of input-output quantum systems,
to the best of the authors’ knowledge, there does not appear to be any work that has tried to
obtain some explicit bounds on the approximation error of the joint state of the system and
2
the quantum field it is coupled to. This work develops a framework for developing bounds
on the error between the quantum state of a quantum system described by the QSDE and
the quantum state of a finite-dimensional approximation described by another QSDE, when
both systems are initialized in a state in the finite-dimensional subspace. Central to the
framework is a contractive semigroup associated with the unitary QSDEs of input-output
Markov quantum systems. Error bounds are developed for both adiabatic elimination
and subspace truncation approximations. For illustration, our results are applied to some
physical examples drawn from the literature. Prelimary results of this work were announced
in the conference paper [14]. The results presented in the work go significantly beyond [14].
In particular, [14] only treates the subspace truncation approximation with some elements
of the proofs omitted, error bounds for adiabatic elimination had not been developed, and
computability of the error bounds were not considered.
The rest of this paper is structured as follows. In Section 2, we present the class of
open quantum systems and the associated QSDEs describing Markovian open quantum
systems. Explicit error bounds for the subspace truncation approximation of a Markovian
open quantum system are established in Section 3 and some examples are provided. We
then establish error bounds for adiabatic elimination approximation in Section 4 and some
examples are also provided. Finally, concluding remarks close the paper in Section 5.
2
2.1
Preliminaries
Notation
√
We use ı = −1 and let (·)∗ denote the adjoint of a linear operator on a Hilbert space as well
as the conjugate of a complex number, and (·)⊤ denote matrix transposition. We denote
by δij the Kronecker delta function. We define ℜ{A} = 12 (A + A∗ ) and ℑ{A} = 2ı1 (A − A∗ ).
For a linear operator A, we write ker(A) to denote the kernel of A and ran(A) the range
of A. We often write |·i to denote an element of a Hilbert space and denote by H⊗F the
algebraic tensor product of Hilbert spaces H and F . For a subspace H0 of a Hilbert space
H, we write PH0 to denote the orthogonal projection operator onto H0 . For a Hilbert space
H = H0 ⊕ H1 , we will write H ⊖ H0 to denote H1 . For a linear operator X on H, X|H0
denotes the restriction of X to H0 . We use B(H) to denote the algebra of all bounded
linear operators on H. We write [A, B] = AB − BA. The notation k · k will be used
to denote Hilbert space norms and operator norms, h·, ·i denotes an inner product on a
Hilbert space, linear in the right slot and antilinear in the left, and |·i h·| denotes an outer
product. Here, 1[0,t] : [0, t] → {0, 1} denotes the indicator function. Finally, Z+ denotes
the set of all positive integers.
2.2
Open quantum systems
Consider a separable Hilbert space H0 and the symmetric boson Fock space (of multiplicity m) F defined over the space L2 ([0, T ]; Cm ) = Cm ⊗ L2 ([0, T ]) with 0 < T < ∞;
3
see [15, Ch. 4-5] for more details. We will use e(f ) ∈ F , with f ∈ L2 ([0, T ]; Cm ), to
m
denote exponential vectors in F . Let S ⊂ L2 ([0, T ]; Cm ) ∩ L∞
loc ([0, T ]; C ) be an admissible subspace in the sense of Hudson-Parthasarathy [1] which contains at least all simple
m
functions, where L∞
loc ([0, T ]; C ) is the space of locally bounded vector-valued functions.
Here, we will consider a dense domain D0 ⊂ H0 and a dense domain of exponential vectors
E = span{e(f ) | f ∈ S} ⊂ F .
Consider an open Markov quantum system that can be described by a set of linear
operators defined on the Hilbert space H0 : (i) a self-adjoint Hamiltonian operator H, (ii) a
vector of coupling operators L with the j-th element, Lj : H0 → H0 for all j = 1, 2, . . . , m,
and (iii) a unitary scattering matrix S with the ij-th element, Sij : H0 → H0 for all i, j =
1, 2, . . . , m. Moreover, the operators Sij , Lj , H and their adjoints are assumed to have D0
as a common invariant dense domain. Under this description, we note that m corresponds
to the number of external bosonic input fields driving the system. Each bosonic input
∗
field can be described by annihilation and creation field operators, bit and bit , respectively,
∗
which satisfy the commutation relations [bit , bjs ] = δij δ(t − s) for all i, j = 1, 2, . . . , m and
∗
all t, s ≥ 0. We can then define the annihilation process Ait , the creation process Ait , and
the gauge process Λij
t as
Z t
Z t
Z t
∗
ij
i
i
i∗
i∗
At =
bs ds,
At =
bs ds,
Λt =
bis bjs ds.
0
0
0
Note that these processes are adapted quantum stochastic processes. In the vacuum repre∗
∗
∗
sentation, the products of their forward differentials dAit = Ait+dt −Ait , dAit = Ait+dt −Ait ,
ij
ij
and dΛij
t = Λt+dt − Λt satisfy the quantum Itō table
dAkt
0
0
0
0
×
dAit
∗
dAjt
dΛij
t
dt
∗
dAkt
δik dt
0
∗
δjk dAit
0
dΛkℓ
t
δik dAℓt
0
δjk dΛiℓ
t
0
dt
0
0 .
0
0
dAi
Here, bit = dtt can be interpreted as a vacuum quantum white noise, while Λiit can be
interpreted as the quantum realization of a Poisson process with zero intensity [1].
Following [16], the time evolution of a Markov open quantum system is given by an
adapted process Ut satisfying the left Hudson-Parthasarathy QSDE [1]:
( m
m
X
X
ij
∗
dUt = Ut
L∗i dAit
Sji − δij )dΛt +
i=1
i,j=1
−
m
X
i,j=1
i∗
∗
Sji
Lj dAt
# )
m
1X ∗
(L Li ) dt
+ ıH −
2 i=1 i
"
(1)
with U0 = I. The quantum stochastic integrals are defined relative to the domain D0 ⊗E.
With the left QSDE, the evolution of a state vector ψ ∈ H0 ⊗ F is given by Ut∗ ψ.
4
In this paper, we are interested in the problem of approximating the system with
operator parameters (S, L, H) by an open quantum system with linear operator parameters
(S (k) , L(k) , H (k) ) defined on a closed subspace H(k) ⊂ H0 , where S (k) is unitary and H (k) is
self-adjoint. Consider a dense domain D (k) ⊂ H(k) . Again, the operators (S (k) , L(k) , H (k) )
and their adjoints are assumed to have D (k) as a common invariant dense domain. Similar
(k)
to (1), the time evolution of the approximating system is given by an adapted process Ut
satisfying the left Hudson-Parthasarathy QSDE [1]:
( m
m
X
X (k)∗
(k)∗
(k)
(k)
i
+
L
dA
Sji − δij )dΛij
dUt = Ut
t
t
i
i=1
i,j=1
−
m
X
i,j=1
∗
(k)∗
Sji Lj dAit
# )
m
X
∗
1
(k)
(k)
(L Li ) dt ,
+ ıH (k) −
2 i=1 i
"
(2)
(k)
with U0 = I. Here, the quantum stochastic integrals in the above equation are defined
relative to the domain D (k) ⊗E. Similarly, the evolution of a state vector ψ ∈ H(k) ⊗ F is
(k) ∗
given by Ut ψ.
2.3
Associated semigroups
Let θt : L2 ([t, T ]; Cm ) → L2 ([0, T ]; Cm ) be the canonical shift θt f (s) = f (t + s). We also
let Θt : F[t → F denote the second quantization of θt , where F[t denotes the Fock space
over L2 ([t, ∞); Cm). Note that an adapted process Ut on H0 ⊗ F is called a contraction
(or unitary) cocycle if Ut is a contraction (or unitary) for all t ≥ 0, t 7→ Ut is strongly
continuous, and Us+t = Us (I ⊗ Θ∗s Ut Θs ).
Let us now impose an important condition on the open quantum systems under consideration, adopted from [16].
Condition 1 (Contraction cocycle solutions). For all t ≥ 0 and all k ∈ Z+ ,
(a) the QSDE (1) possesses a unique solution Ut which extends to a unitary cocycle on
H0 ⊗ F ,
(k)
(b) the QSDE (2) possesses a unique solution Ut
on H(k) ⊗ F .
(αβ)
Let us define an operator Tt
(αβ)
hu, Tt
1
which extends to a contraction cocycle
: H0 → H0 via the identity
2 +kβk2 )t
vi = e− 2 (kαk
u ⊗ e(α1[0,t] ), Ut v ⊗ e(β1[0,t] )
for all u, v ∈ H0 and all α, β ∈ Cm . From [16, Lemma 1], under Condition 1(a), the
(αβ)
operator Tt
∈ B(H0 ) is a strongly continuous contraction semigroup on H0 and its
5
generator L(αβ) satisfies Dom(L(αβ) ) ⊃ D0 such that
" m
m
m
X
X
X
∗
∗
L(αβ) u =
αi∗ Sji
βj −
αi∗ Sji
Lj +
L∗j βj
i,j=1
1
+ ıH −
2
i,j=1
m
X
L∗i Li
i=1
!
j=1
#
kαk2 + kβk2
−
u
2
(3)
for all u ∈ D0 . Here, we note that Dom(L(αβ) ) is dense in H0 . We likewise define an
(k;αβ)
(k)
operator Tt
: H(k) → H(k) by replacing Ut with Ut .
Condition 2 (Core for generators). For all t ≥ 0 and all k ∈ Z+ ,
(a) D0 is a core for L(αβ) ,
(b) D (k) is a core for L(k;αβ) .
Condition 2(a) ensures that the definition (3) completely determines L(αβ) for all α, β ∈
C , likewise 2(b) completely determines L(k;αβ) . In the sequel, we will make use of the above
semigroups associated with open quantum systems in establishing our model approximation
error bound. Several sufficient conditions are known to guarantee that a QSDE possesses
a unique solution that extends to a unitary cocycle when the Hilbert space is infinitedimensional and the operator coefficients of the QSDE are unbounded, see, e.g., [17], and a
related discussion in [16, Remark 4]. Throughout the paper, we will assume that Conditions
1 and 2 are fulfilled.
m
3
Error bounds for subspace truncation approximations
In this section, we consider the problem where the infinite-dimensional space H0 is truncated to a finite-dimensional subspace H(k) , and the original operators X on H0 is approximated by truncated operators of the form PH(k) XPH(k) . Here, the dimension of H(k)
(k)
increases with k ∈ Z+ and D (k) = H(k) . Moreover, Condition 1(b) (in fact, Ut is unitary [1]) and 2(b) hold immediately.
3.1
Assumptions and preliminary results
Assumption 1. For any k ∈ Z+ and any α, β ∈ Cm , H(k) ⊂ Dom(L(αβ) ).
Let M(k) = ran L(α,β) − L(k;α,β) H(k) . Supposing that Assumption 1 holds, we also
assume the following.
(αβ)
Assumption 2. For each k ∈ Z+ and each α, β ∈ Cm , there exists γk
0, and a non-trivial subspace {0} ⊂ K(k) ⊆ H(k) such that
6
(k;αβ)
, qL
(k;αβ)
, qa
(k;αβ)
, qe
>
(k;αβ)
L(αβ) − L(k;αβ) K(k) ≤ qL
.
(b) L(αβ) − L(k;αβ) H(k) ⊖K(k) = 0,
T
i.e., H(k) ⊖ K(k) ⊆ α,β∈Cm ker L(αβ) − L(k;αβ)
(a)
H(k)
(c) For any u ∈ H(k) ,
ℜ{hPK(k) L(k;αβ) u, PK(k) ui} = −g(k, α, β) kPK(k) uk2 + h(k, α, β, u)
(αβ)
for some g(k, α, β) ≥ γk
(k;αβ)
and some |h(k, α, β, u)| ≤ qa
kPK(k) uk kuk.
(d) For any u ∈ H(k) and any t ≥ 0,
(αβ)
ℜ{hTt
(α,β)
L(αβ) PM(k) u, Tt
(αβ)
for some ĝ(k, α, β) ≥ γk
(αβ)
PM(k) ui} = −ĝ(k, α, β) Tt
2
PM(k) u
(k;αβ)
and some ĥ(t, k, α, β, u) ≤ qe
+ ĥ(t, k, α, β, u)
(αβ)
Tt
PM(k) u kuk.
Assumption 3. There exists r, s ∈ Z+ such that, for all α, β ∈ Cm , we have that
(k;αβ)
(k;αβ)
lim qL
qe
(αβ)
k→∞
γk
!(1−2−r )
(k;αβ)
qa
(αβ)
γk
!(1−2−s )
=0
Moreover, for any i = 0, 1, 2, . . . , r, and j = 0, 1, 2, . . . , s, we have that
(k;αβ)
lim
k→∞
qL
(αβ)
γk
(αβ)
lim e−γk
k→∞
(k;αβ)
qe
(αβ)
γk
t (k;αβ)
qL
!(1−2−i )
(k;αβ)
qa
γk
!(1−2−ℓ )
(k;αβ) (k;αβ)
qa
(αβ) 2
(γk )
qe
(αβ)
!(1−2−j )
=0
= 0,
for any α, β ∈ Cm , any t > 0, and for ℓ = 0, 1, 2, . . . , min{r, s} − 1.
Let us present some useful lemmas.
Lemma 1. Suppose that Assumption 2(c) holds. Then for any k, r ∈ Z+ , any α, β ∈ Cm ,
any u ∈ H(k) , and any t ≥ 0, it holds that
−r
−j
!
!
r−1
(k;αβ) (1−2 )
(k;αβ) (1−2 )
X
(αβ)
qa
−j
qa
(k;αβ)
kuk (4)
PK(k) Tt
u ≤
+
cj e−(2 )γk t
(αβ)
(αβ)
γk
γk
j=0
q
where c0 = 1 and cj = cj−1 2j (2j − 1)−1 for j ≥ 1.
7
(αβ)
Proof. First note, from the definition of a strongly continuous semigroup, that T0
(αβ)
and dtd Tt u = L(αβ) u [18]. From Assumption 2(c), we have that
E
2
d D
d
(k;αβ)
(k;αβ)
(k;αβ)
P (k) T
u =
P (k) T
u, PK(k) Tt
u
dt K t
dt nDK t
Eo
(k;αβ)
(k;αβ)
= 2ℜ PK(k) L(k;αβ) Tt
u, PK(k) Tt
u
(k;αβ)
=−2g(k, α, β) PK(k) Tt
2
(k;αβ)
u + 2h(k, α, β, Tt
=I
u).
Solving the above ODE gives us that
2
(k;αβ)
PK(k) Tt
u
−2g(k,α,β)t
2
Z
t
h(k, α, β, Tτ(k;α,β)u)e−2g(k,α,β)(t−τ ) dτ
0
Z t
(αβ)
(αβ)
2
−2γk
t
(k;αβ)
≤e
kPK(k) uk + 2qa
e−2γk (t−τ ) PK(k) Tτ(k;αβ) u kukdτ (5)
0
"
#
(k;αβ)
(αβ)
qa
≤ e−2γk t + (αβ) kuk2.
γk
=e
kPK(k) uk + 2
Here, the second step follows from Assumption 2(c). The lastp
step follows because kPK(k) uk ≤
(k;αβ)
kuk and Tt
is a contraction semigroup. Noticing that |a|2 + |b|2 ≤ |a| + |b| for any
a, b ∈ R, we have that
! 21
(k;αβ)
(αβ)
qa
(k;αβ)
kuk.
(6)
PK(k) Tt
u ≤ e−γk t +
(αβ)
γk
Now, substituting (6) into the right-handed side of (5), we have that
(k;αβ)
PK(k) Tt
(αβ)
−2γk
≤e
t
2
kuk +
Z
+ 2qa(k;αβ)
(αβ)
≤ e−2γk
t
t
2qa(k;αβ)
Z
t
e
(αβ)
−γk
t
+ 2e
(αβ)
+2e−γk
t
(αβ)
e−γk
(2t−τ )
0
(αβ)
−2γk
(t−τ )
0
(αβ)
−2γk
t
= e
2
u
(k;αβ)
qa
(αβ)
γk
! 12
(αβ)
−γk
t
1−e
(k;αβ)
qa
(αβ)
γk
!
+
kuk2dτ
kuk2dτ
(k;αβ)
qa
(αβ)
γk
(k;αβ)
qa
(αβ)
γk
!
! 32
(αβ)
−2γk
t
+ 1−e
! 32
kuk2
kuk2.
Taking the square root on both sides of the equation, we get
!1
(k;αβ) 2
√
(αβ)
(αβ)
1
qa
(k;αβ)
+
PK(k) Tt
u ≤ e−γk t + 2e− 2 γk t
(αβ)
γk
8
(k;αβ)
qa
(αβ)
γk
(k;αβ)
qa
(αβ)
γk
! 34
kuk.
From repeat application of the above steps, we establish the lemma statement.
Lemma 2. Suppose that Assumption 2(d) holds. Then for any k, r ∈ Z+ , any α, β ∈ Cm ,
any u ∈ H(k) , and any t ≥ 0, it holds that
−r
!
!(1−2−j )
r−1
(k;αβ) (1−2 )
(k;αβ)
X
(αβ)
qe
qe
(αβ)
t
−(2−j )γk
kuk
Tt PM(k) u ≤
+
c
e
(7)
j
(αβ)
(αβ)
γk
γ
j=0
k
q
where c0 = 1 and cj = cj−1 2j (2j − 1)−1 for j ≥ 1.
Proof. Similar to Lemma 1, using Assumption 2(d), we have that
2
d
(αβ)
(αβ)
(αβ)
Tt PM(k) u = 2ℜ{hTt L(αβ) PM(k) u, Tt PM(k) ui}
dt
(αβ)
= −2ĝ(k, α, β) Tt
PM(k) u
2
+ 2ĥ(t, k, α, β, u).
Solving the ODE, we have that
(αβ)
Tt PM(k) u
2
(αβ)
−2γk
≤e
"
2
t
2qe(k;αβ)
kPM(k) uk +
#
(k;αβ)
(αβ)
q
e
≤ e−2γk t + (αβ) kuk2 .
γk
Z
t
(αβ)
e−2γk
(t−τ )
0
(αβ)
Tt
PM(k) u kukdτ
The lemma statement is then established by following similar arguments to Lemma 1.
3.2
Error bounds for finite-dimensional approximations
We begin by defining
k
zr,s
(t, α, β)
:=
(k;αβ)
qL
+
(αβ)
γk
s−1
X
2 i ci
(αβ)
i=1
γk
r−1 X
s−1
X
(αβ)
γk
r−1
X
2 i ci
i=0
+
t
!(1−2−r )
(k;αβ)
qe
(αβ)
−2−i γk
t
1−e
(αβ)
−2−i γk
t
1−e
(k;αβ)
qa
(αβ)
γk
!(1−2−s )
q (k;αβ)
e
(αβ)
γk
q (k;αβ)
a
(αβ)
γk
!(1−2−i )
!(1−2−i )
(k;αβ)
qa
(αβ)
γk
(k;αβ)
qe
(αβ)
γk
!(1−2−s )
!(1−2−r )
q (k;αβ)
−i (αβ)
(αβ)
e
−2 γk
t
−2−j γk
t
ci cj
+
e
−e
(αβ)
(αβ)
i
j
(2 − 2 ) γk
γk
i=0 j=0
j6=i
−i
!
min{r,s}−1
(k;αβ) (k;αβ) (1−2 )
X
(αβ)
qe
qa
−i
,
+t
c2i e−2 γk t
(αβ) 2
(γ
)
i=0
k
2(i+j)
9
!(1−2−i )
! −j
(k;αβ) (1−2 )
qa
(αβ)
γk
(8)
p
where c0 = 1, and cj = cj−1 2j (2j − 1)−1 for j ≥ 1. We now establish an error bound
between the two semigroups associated with the open quantum systems.
Lemma 3. Suppose Assumptions 1 and 2 hold. Then for any k, r, s ∈ Z+ , any α, β ∈ Cm ,
any u ∈ H(k) , and any t ≥ 0, it holds that
(αβ)
(k;αβ)
k
(t, α, β)kuk.
(9)
u ≤ zr,s
Tt
− Tt
Proof. First note, from the definition of a strongly continuous semigroup, that [18]
(αβ)
1. T0
(k;αβ)
= T0
2.
d (αβ)
T
u
dt t
3.
d (k;αβ)
T
u
dt t
= I for all k ∈ Z+ ,
= L(αβ) u for all u ∈ Dom(L(αβ) ),
= L(k;αβ) u for all u ∈ H(k) (since H(k) is finite-dimensional).
From the above properties and Assumption 1, we can write for all u ∈ H(k) and all t ≥ 0
that
d (αβ)
(k;αβ)
(αβ) (αβ)
(k;αβ) (k;αβ)
u
u = L Tt
−L
Tt
T
− Tt
dt t
(k;αβ)
(αβ)
(k;αβ)
u + L(αβ) − L(k;αβ) H(k) Tt
u (10)
− Tt
= L(αβ) Tt
(αβ)
(k;αβ)
u = 0. Note that by Assumption 2,
with T0 − T0
bounded operator. Since L
(αβ)
(due to Assumption 1), and
L(αβ) − L(k;αβ)
H(k)
is a
is a generator of a semigroup on a Hilbert space,
(αβ)
(k;αβ)
u ∈ Dom(L(αβ) ),
T0 − T0
L(αβ) − L(k;αβ)
(k;αβ)
H(k)
Tt
u ∈ C 1 ([0, t]; H0 )
(the class of continuously differentiable functions from [0, t] to H0 ), a unique solution of
(10) exists and is given by [18, Thm 3.1.3]
Z t
(αβ)
(αβ)
(k;αβ)
u=
Tt−τ L(αβ) − L(k;αβ) Tτ(k;αβ) udτ
Tt
− Tt
0
for all t ≥ 0 and all u ∈ H(k) . From Assumption 2(b) and the definition of M(k) , we then
have for all u ∈ H(k) and all t ≥ 0 that
Z t
(αβ)
(αβ)
(k;αβ)
Tt−τ L(αβ) − L(k;αβ) Tτ(k;αβ) u dτ
u ≤
Tt
− Tt
Z0 t
(αβ)
(11)
=
Tt−τ PM(k) L(αβ) − L(k;αβ) PK(k) Tτ(k;αβ) u dτ.
0
10
Now using the bounds (4) and (7) (established in Lemmas 1 and 2, respectively), and
applying Assumption 2(a), we have that
(αβ)
Tt−τ PM(k) L(αβ) − L(k;αβ) PK(k) Tτ(k;αβ) u
! −r
! −s
(k;αβ) (1−2 )
(k;αβ) (1−2 )
qa
(k;αβ) qe
≤ qL
(αβ)
(αβ)
γk
γk
! −s " r−1
! −i #
(k;αβ) (1−2 )
(k;αβ) (1−2 ) X
(αβ)
q
qa
−i
e
ci e−2 γk (t−τ )
+
(αβ)
(αβ)
γk
γk
i=0
−i #
! −r " s−1
!
(k;αβ) (1−2 ) X
(k;αβ) (1−2 )
(αβ)
qe
q
−i
a
+
ci e−2 γk τ
(αβ)
(αβ)
γk
γk
i=1
−i
−j #
!
!
" r−1 s−1
(k;αβ) (1−2 )
(k;αβ) (1−2 )
XX
(αβ)
qa
qe
−i
−i
−j
kuk.
+
ci cj eγk (−2 t+(2 −2 )τ )
(αβ)
(αβ)
γ
γ
i=0 j=0
k
k
The result (9) then follows from substitution of the above identity into (11) and integration.
This establishes the lemma statement.
Let S′ ⊂ L2 ([0, T ]; Cm) denote the dense set of all simple functions in L2 ([0, T ]; Cm ).
That is, for any t ∈ [0, T ] and f ∈ S′ , there
Pℓ exists 0 < ℓ < ∞ and a sequence 0 = t0 m<
t1 < · · · < tℓ < tℓ+1 = t such that f = i=0 α(i)1[ti ,ti+1 ) for some constants α(i) ∈ C ,
i = 0, 1, . . . , ℓ. Let U (k) = {u ⊗ e(f ) | u ∈ H(k) , f ∈ S′ }. We can now proceed to derive
error bounds for approximations by subspace truncation.
Lemma 4. Suppose Assumptions 1 and 2 hold. For any ψ1 = u1 ⊗ e(f1 ), ψ2 = u2 ⊗
e(f2 ) ∈ U (k) , let
P t0 = 0 < t1 < . . . < tℓ < tℓ+1 = t with 0 < t ≤ T be a sequence
such that fj = ℓi=0 αj (i)1[ti ,ti+1 ) for all i = 0, 1, . . . , ℓ, with αj (i) ∈ Cm for j = 1, 2 and
i = 0, 1, . . . , ℓ. Then for any k, r, s ∈ Z+ ,
ℓ
∗
E
D
X
(k)
k
ψ1 , ψ2 ≤
zr,s
((ti+1 − ti ), α1 (i), α2 (i)) kψ1 kkψ2 k.
Ut − Ut
(12)
i=0
Proof. First recall that our admissible subspace S contains S′ . Hence, U (k) ⊂ H(k) ⊗E
and the quantum stochastic integrals are well defined for all ψ ∈ U (k) . Also, recall that
kψk2 = hu ⊗ e(f ), u ⊗ e(f )i = kuk2 ke(f )k2 for any ψ ∈ H0 ⊗ F [15]. Using the cocycle
(αβ)
(k;αβ)
properties (Condition 1) as well as the definitions of Tt
and Tt
, we have the identity
[16]
(α (0)α2 (0))
1
hUt∗ ψ1 , ψ2 i = ke(f1 )k ke(f2 )k hu1 , Tt1 −t
0
(αβ)
and likewise when Ut and Tt
(α (ℓ)α2 (ℓ))
u2 i,
(k)
(k;αβ)
· · · Tt−t1ℓ
are respectively replaced by Ut
11
and Tt
(13)
, from which
we immediately obtain
∗
D
E
(k)
Ut − Ut
ψ1 , ψ2
(α1 (0)α2 (0))
(α1 (ℓ)α2 (ℓ))
(k;α1 (0)α2 (0))
(k;α1 (ℓ)α2 (ℓ))
≤ kψ1 k ke(f2 )k Tt1 −t
u2 .
·
·
·
T
−
T
·
·
·
T
t−tℓ
t1 −t0
t−tℓ
0
(14)
Now, note that for any u ∈ H(k) that
(α1 (0)α2 (0)) (α1 (1)α2 (1))
(α1 (ℓ)α2 (ℓ))
(k;α1 (0)α2 (0)) (k;α1 (1)α2 (1))
(k;α1 (ℓ)α2 (ℓ))
Tt1 −t
T
·
·
·
T
−
T
T
·
·
·
T
u
t2 −t1
t−tℓ
t1 −t0
t2 −t1
t−tℓ
0
h
(α1 (0)α2 (0))
(k;α (0)α (0))
(k;α (1)α (1)) (k;α (2)α (2))
(k;α (ℓ)α (ℓ))
= Tt1 −t
− Tt1 −t01 2
Tt2 −t11 2 Tt3 −t21 2 · · · Tt−tℓ 1 2
0
(α1 (0)α2 (0))
(α1 (1)α2 (1))
(k;α1 (1)α2 (1))
(k;α (1)α (1))
(k;α (ℓ)α (ℓ))
+ Tt1 −t
T
−
T
Tt2 −t11 2 · · · Tt−tℓ 1 2
t2 −t1
t2 −t1
0
i
(α1 (0)α2 (0))
(α1 (ℓ)α2 (ℓ))
(k;α1 (ℓ)α2 (ℓ))
(α1 (ℓ−1)α2 (ℓ−1))
u.
+ · · · + Tt1 −t
T
−
T
·
·
·
T
t−tℓ
t−tℓ
tℓ −tℓ−1
0
From (14), the bound (9) (established in Lemma 3), and that fact that the semigroups are
contractions, we have for any u ∈ H(k) that
(α1 (0)α2 (0)) (α1 (1)α2 (1))
(k;α1 (0)α2 (0)) (k;α1 (1)α2 (1))
(α1 (ℓ)α2 (ℓ))
(k;α1 (ℓ)α2 (ℓ))
Tt1 −t0
Tt2 −t1
− Tt1 −t0
· · · Tt−tℓ
u
Tt2 −t1
· · · Tt−tℓ
≤
ℓ
X
i=0
k
zr,s
((ti+1 − ti ), α1 (i), α2 (i)) kuk.
(15)
The bound (12) then follows by substituting (15) into (14). This establishes the theorem
statement.
Corollary 1. Suppose Assumptions 1, 2, and 3 hold. For any t ∈ [0, T ] with 0 < T < ∞,
any ψ1 = u1 ⊗ e(f1 ), ψ2 = u2 ⊗ e(f2 ) ∈ H(k) ⊗ F , we have that
∗
D
E
(k)
Ut − Ut
ψ1 , ψ2 ≤ 2 ku1kke(f1 ) − e(f1′ )kkψ2 k + ku2kke(f2 ) − e(f2′ )kkψ1 k
+
ℓ
X
i=0
′
k
′
′
zr,s
(ti+1 − ti ), f1,i
, f2,i
kψ1 kkψ2′ k.
(16)
P
′
for any ψj′ = uj ⊗ e(fj′ ) ∈ U (k) with fj′ = ℓi=0 fj,i
1[ti ,ti+1 ) for some ℓ ∈ Z+ , some sequence
′
t0 = 0 < t1 < . . . < tℓ < tℓ+1 = t, and some constants fj,i
∈ Cm for j = 1, 2 and
i = 0, 1, . . . , ℓ. Moreover, for any fixed positive integer p ∈ Z+ , and any ψ1 = u1 ⊗
e(f1 ), ψ2 = u2 ⊗ e(f2 ) ∈ H(p) ⊗ F ,
∗
D
E
(k)
Ut − Ut
ψ1 , ψ2
lim
= 0.
(17)
k→∞
(k)
Proof. Recall that Ut is unitary and Ut
is a contraction (Condition 1). From the triangle
12
inequality and Cauchy-Schwarz’s inequality, we note that
∗
∗
D
D
E
E
(k)
(k)
Ut − Ut
Ut − Ut
u1 ⊗ (e(f1 ) − e(f1′ )) , u2 ⊗ e(f2 )
ψ1 , ψ2 ≤
∗
E
D
(k)
′
′
u1 ⊗ e(f1 ), u2 ⊗ (e(f2 ) − e(f2 ))
+
Ut − Ut
∗
E
D
(k)
u1 ⊗ e(f1′ ), u2 ⊗ e(f2′ )
+
Ut − Ut
≤ 2ku1 k ke(f1 ) − e(f1′ )k kψ2 k + 2ku2 k ke(f2 ) − e(f2′ )k kψ1′ k
∗
D
E
(k)
Ut − Ut
u1 ⊗ e(f1′ ), u2 ⊗ e(f2′ ) .
+
The result (16) then follows from the bound (12) (established in Lemma 4).
To show (17), recall that S′ is dense in L2 ([0, T ]; Cm ). Therefore, for each ǫ > 0 and
each fj ∈ L2 ([0, T ]; Cm ), there exists fj′ ∈ S′ such that kfj −fj′ k < ǫ. Moreover, there exists
P
0 < ℓ < ∞ and 0 = t0 < t1 < · · · < tℓ < tℓ+1 = t, such that fj′ = ℓi=0 αj′ (i)1[ti ,ti+1 ) with
αj′ (0), αj′ (1), . . . , αj′ (ℓ) ∈ Cm . Suppose that u1 , u2 6= 0 (otherwise the corollary statement
becomes trivial), then for any ǫ > 0 we may choose f1′ , f2′ ∈ S′ (choosing f1′ first followed
by f2′ ) such that
ǫ
ke(f1 ) − e(f1′ )k <
6ku1 k ke(f2 )k
ǫ
ke(f2 ) − e(f2′ )k <
6ku2 k ke(f1′ )k
Finally, from Assumption 3 and the bound (12) (established in Lemma 4), we can find a
sufficiently large k ∈ Z+ , larger than p, such that
ℓ
X
i=0
k
zr,s
(((ti+1 − ti ), α1′ (i), α2′ (i))) <
ǫ
3kψ1′ kkψ2′ k
.
From (16) and the above choices, we then have that
∗
D
E
(k)
Ut − Ut
ψ1 , ψ2 < ǫ.
This establishes the corollary statement.
Theorem 1. Suppose that Assumptions 1,2, and 3 hold. Let 0 < T < ∞. For any
PL′t ′
′
′
ψj,t , with ψj,t
= u′j,t ⊗ e(gj,t
) ∈ U (k) and
t ∈ [0, T ], consider any L′t ∈ Z+ , any ψt′ = j=1
′
′
′
uj,t 6= 0. Also, consider any f ∈ S . Let ℓ be a positive integer and t0 = 0 < t1 < . . . <
P
P
′
′
1[ti ,ti+1 ) for
= ℓi=0 gj,t,i
tℓ < tℓ+1 = t be a sequence such that f ′ = ℓi=0 fi′ 1[ti ,ti+1 ) and gj,t
′
m
′
some constants fi , gj,t,i ∈ C for for j = 1, 2, . . . , Lt and i = 0, 1, . . . , ℓ. Let u ∈ H(k) with
kuk = 1, and |f i = e(f )/ke(f )k ∈ F (i.e., |f i is a coherent state with amplitude f ), and
ψ = u ⊗ |f i. Then,
∗ 2
(k)
Ut − Ut
ψ ≤ 4 k|f i − |f ′ik + kUt ∗ ψ − ψt′ k
′
+2
Lt
ℓ
X
X
j=1 i=1
13
′
k
′
zr,s
(ti+1 − ti ), fi′ , gj,t,i
kψj,t k.
(18)
(k)
If Ut
is unitary for each t then the following bound holds,
∗
(k)
Ut − Ut
ψ
2
′
≤ 4 k|f i − |f ik +
+2
(k) ∗
Ut ψ
ψt′
−
′
Lt
ℓ
X
X
j=1 i=1
′
k
′
zr,s
(ti+1 − ti ), fi′ , gj,t,i
kψj,t k.
Moreover, for any fixed p ∈ Z+ and any ψ = u ⊗ |f i ∈ H(p) ⊗ F ,
∗
(k)
Ut − Ut
ψ = 0.
lim
(19)
(20)
k→∞
for any t ∈ [0, T ] with 0 < T < ∞.
Remark 1. Note that a stronger
result of strong convergence uniformly over compact time
∗
(k)
ψ = 0, has been established in [16, Proposition
intervals, limk→∞ sup0≤t≤T Ut − Ut
20] employing a Trotter-Kato theorem. However, no error bound as in (18) for a finite
value of k has previously been established.
(k)
Proof. First note that since Ut is unitary and Ut
2
∗
(k)
ψ
Ut − Ut
a contraction,
(k)
≤ 2ℜ{h(Ut − Ut )∗ ψ, Ut∗ ψi}
(k)
≤ 2|h(Ut − Ut )∗ ψ, Ut∗ ψi|.
Also, we have that
∗
∗
E
E
D
D
(k)
(k)
∗
′
∗
ψ, Ut ψ ≤
u ⊗ (|f i − |f i) , Ut ψ
Ut − Ut
Ut − Ut
∗
D
E
(k)
+
Ut − Ut
u ⊗ |f ′ i, (Ut∗ ψ − ψt′ )
∗
D
E
(k)
′
′
Ut − Ut
+
u ⊗ |f i, ψt
≤ 2kukkf i − |f ′ ikkψk + 2kukk|f ′ik kUt∗ ψ − ψt′ k
L′t D
∗
E
X
(k)
′
u ⊗ |f ′ i, ψj,t
.
+
Ut − Ut
j=1
The result (18) then follows from the bound (12), and substituting kuk = 1, kψk = 1, and
(k)
k|f ′ ik = 1. If Ut is unitary then we have the bound,
∗
(k)
Ut − Ut
ψ
2
(k)
= 2ℜ{h(Ut
(k) ∗
− Ut )∗ ψ, Ut
(k)
(k) ∗
≤ 2|h(Ut − Ut )∗ ψ, Ut
14
ψi}
ψi|,
and following analogous calculations to the above yields,
∗
D
E
(k)
(k) ∗
(k) ∗
Ut − Ut
ψ, Ut ψ ≤ 2kukkf i − |f ′ikkψk + 2kukk|f ′ik Ut ψ − ψt′
+
L′t D
X
j=1
Ut −
(k)
Ut
∗
u ⊗ |f
′
′
i, ψj,t
E
,
leading to the alternative bound (19).
To show (20), let t ∈ [0, T ] be fixed. Suppose ψ 6= 0. Then for any ǫ > 0, we may
choose f ′ ∈ S′ such that
k|f i − |f ′ik <
ǫ2
.
12
′
Since T is finite, for k0 ∈ Z+ sufficiently large we can choose 0 < L′t < ∞ and ψj,t
∈ U (k0 )
for j = 1, 2, . . . , L′t such that
kUt∗ ψ − ψt′ k <
ǫ2
,
12
PL′t ′
′
with ψt′ =
j=0 ψj,t . Finally, from Corollary 1, since Lt is finite we can find k1 ∈ Z+
sufficiently large, with k1 > max{p, k0 }, such that
D
Ut −
(k)
Ut
∗
u ⊗ |f
′
i, ψt′
E
<
ǫ2
.
6
for all k > k1 . From the above choices and taking square roots on both sides of (18), we
have that
∗
(k)
Ut − Ut
ψ < ǫ,
for all k > k1 . Since the theorem statement holds trivially when ψ = 0, this completes the
proof.
A discussion of the error bounds presented in Theorem 1 is now in order, starting with
(18). The error bound on the right hand side of (18) is the sum of three terms. The first
term is a bound on the error committed by approximating e(f ) by e(f ′ ) for some simple
function f ′ . The second term bounds the error in approximating Ut∗ ψ by a finite sum of
terms in U (k) given in ψt′ . Finally, the last term gives an upper bound for the magnitude
(k)
of the inner product between the error term (Ut − Ut )∗ u ⊗ e(f ′ ) and ψt′ . The first and
15
final terms are computable. However, the second term is difficult. We note that
*
+1/2
L′t
L′t
L′t
X
X
X
′
′
′
,
= Ut∗ ψ −
ψj,t
, Ut∗ ψ −
ψj,t
Ut∗ ψ −
ψj,t
j=1
j=1
′
′
= kψk2 − 2
j=1
2
= kψk − 2
(f ′ g ′
Lt
X
′
ℜ{hUt∗ ψ, ψj,t
i} +
Lt
X
′
ke(gj,t
)kℜ
j
)
′
j
′
ℓ j,t,ℓ
′
Ttℓ+1
−tℓ uj,t
Eo
+
Lt
X
j=1
Lt
X
j=1
2 1/2
′
ψj,t
,
nD
′
′
(f0′ gj,t,0
) (f1′ gj,t,1
)
u, Tt1 −t
Tt2 −t1 · · ·
0
2 1/2
′
ψj,t
,
(21)
where in the last line, we again use the identity (13). However, this is difficult to compute
(αβ)
as it involves the semigroup Tt
which acts on the infinite dimensional space H0 . Thus
to alleviate this difficulty we now turn to the alternative bound (19).
The first and third terms of (19) are the same as for (18). However, for the second
term we have the identity
L′t
L′t
nD
X
X
∗
(k;f ′ g ′
) (k;f ′ g ′
)
(k)
′
′
2
Ut ψ −
ψj,t =
ke(gj,t)kℜ u, Tt1 −t00 j,t,0 Tt2 −t11 j,t,1 · · ·
kψk − 2
j=1
j
′
′
)
(ℓ;fℓ′ gj,t,ℓ
Ttℓ+1 −tℓ u′j,t
Eo
+
Lt
X
j=1
2 1/2
′
ψj,t
,
(22)
derived in the same manner as (21). However, unlike (21), (22) involves only the semigroup
T (k;αβ) which acts on the finite dimensional Hilbert space H(k) . Thus it is a quantity
that will be much easier to compute. All that remains now is to construct a suitable
(k) ∗
′
′
approximation ψj,t
to Ut ψ. One way to do this is to choose ψj,t
to locally minimize the
right hand side of (22) or, equivalently, the term under the square root. Unfortunately,
although this is in principle possible, it is in general a challenging and computationally
intensive non-convex optimization problem. We will demonstrate this optimization in an
example that will follow.
3.3
3.3.1
Subspace truncation examples
Kerr-nonlinear optical cavity:
Consider a single-mode Kerr-nonlinear optical cavity coupled to a single external coherent
field (m = 1), which is used in the construction of the photonic logic gates presented in [12].
16
P
2
Let H0 = ℓ2 (the space of infinite complex-valued sequences with ∞
n=1 |xn | < ∞) which
has an orthonormal Fock state basis {|ni}n≥0 . On this basis {|ni}n≥0 , the annihilation,
creation, and number operators of the cavity oscillator can be defined (see, e.g., [16])
satisfying
√
√
a∗ |ni = n + 1 |n + 1i ,
a∗ a |ni = n |ni ,
a |ni = n |n − 1i ,
respectively. Similar to Examples 14-15 of [16], we set D0 = span{|ni | n ∈ Z+ }. The
Kerr-nonlinear optical cavity can be described by
√
S = I,
L = λa,
H = ∆a∗ a + χa∗ a∗ aa,
where λ, ∆, χ > 0. We now show that Conditions 1(a) and 2(a) hold for the Kerr-nonlinear
cavity.
Consider H(k) = span {|ni | n = 0, 1, 2, . . . , k} and a system approximation of the form
S (k) = I,
H (k) = PH(k) HPH(k) .
L(k) = PH(k) LPH(k) ,
(23)
Conditions 1(b) and 2(b) hold immediately because D (k) = H(k) is finite dimensional. Note
that H(k) ⊂ D0 . Hence, Assumption 1 holds. Recall that M(k) = ran L(α,β) − L(k;α,β) H(k) .
From the (S, L, H), we see that M(k) = span{|k + 1i}. Now consider K(k) = span {|ki}
(αβ)
and γk = 21 (λk + |α − β|2). We will now show that Assumptions 2 and 3 hold for the
Kerr-nonlinear optical cavity and the approximation.
Assumption 2(a): Note that
L(αβ) − L(k;αβ)
√
= β λ a∗ |K(k) .
K(k)
(k;αβ)
Thus, we have that Assumption 2(a) holds with qL
=
p
λ(k + 1)|β|.
Assumption 2(b): This assumption follows for the defined K(k) because
ker L(αβ) − L(k;αβ) H(k) = span {|ni | n = 0, 1, 2, . . . , k − 1} for all α, β ∈ C.
∗
Assumption 2(c): First
note that L(k) PK(k) = 0, PK(k) L(k) PH(k) = 0, and
∗
∗
PK(k) 21 L(k) L(k) − ıH PH(k) ⊖K(k) = 0. Also, for any u ∈ H(k) , PK(k) L(k) L(k) PK(k) u =
λkPK(k) u. From these identities and the fact that H (k) is self-adjoint, we have for any
u ∈ H(k) that
ℜ{hPK(k) L(k;αβ) u, PK(k) ui} = ℜ{hPK(k) L(k;αβ) (PK(k) + PH(k) ⊖K(k) ) u, PK(k) ui}
1
= − λk + |α|2 + |β|2 − 2ℜ{α∗ β} kPK(k) uk2
2 nD
Eo
√
+ℜ
λβPK(k) a∗ PH(k) ⊖K(k) u, PK(k) u
= −g(k, α, β) kPK(k) uk2 + h(k, α, β, u).
17
Notice that |α|2 + |β|2 − 2ℜ{α∗ β} = |α − β|2. Also, note that
ℜ
nD√
λβPK(k) a∗ PH(k) ⊖K(k) u, PK(k) u
√
≤ |β| λ kPK(k) a∗ PH(k) ⊖K(k) uk kPK(k) uk
√
≤ |β| λkkuk kPK(k) uk .
Eo
(αβ)
Therefore, we have that Assumption 2(c) holds for the defined K(k) and the defined γk
√
(k;αβ)
with qa
= |β| λk.
Assumption 2(d): For any u ∈ H0 , HPM(k) u = (∆(k + 1) + χk(k + 1)) PM(k) u. This
(αβ)
(αβ)
implies that ℜ{hTt (ıH)PM(k) u, Tt PM(k) ui} = 0. Also, L∗ LPM(k) u = λ(k + 1)PM(k) u
for any u ∈ H0 . Therefore, we have that
(αβ)
(αβ)
ℜ{hTt L(αβ) PM(k) u, Tt PM(k) ui}
1
(αβ)
2
2
∗
= − λ(k + 1) + |α| + |β| − 2ℜ{α β} Tt PM(k) u
2
Eo
nD
√
(αβ)
(αβ)
λ (βa∗ − α∗ a) PM(k) u, Tt PM(k) u
+ ℜ Tt
(αβ)
= −ĝ(k, α, β) Tt
PM(k) u
2
2
+ ĥ(t, k, α, β, u).
(αβ)
Similarly to the previous derivation, using that Tt
is a contraction, we see that
p Assump(αβ)
(k;αβ)
(k)
λ(k + 1) +
tion
2(d)
holds
for
the
defined
K
and
the
defined
γ
with
q
=
|α|
e
k
p
|β| λ(k + 2).
(αβ)
(k;αβ)
(k;αβ)
(k;αβ)
Assumption 3: From the defined γk , qL
, qa
, qe
, we see that this assumption holds for any r, s ∈ Z+ such that r + s ≥ 3.
Finally, because Assumptions 1-3 hold, Lemma 4, Corollary 1, and Theorem 1 can be
applied to obtain error bounds on the finite-dimensional approximations.
Numerical example of the Kerr-nonlinear optical cavity: Consider λ = 25, ∆ =
50, and χ = −∆/60 (these parameters are used in [12]). We also consider an input field
with a constant amplitude of α = 0.1 for
t ∈ [0, T] with T = 5. Let ψ = |0i ⊗ |α1[0,t] i. We
(k)
∗
Ut − Ut
ψ for different values of k. From (18)
will now compute an error bound on
(established in Theorem 1) and the fact that α1[0,t] ∈ S′ is a simple function, we have that
(k)
Ut − Ut
∗
2
ψ
(k) ∗
≤ 4 Ut
ψ − ψt′
′
+2
Lt
ℓ
X
X
j=1 i=1
′
k
′
zr,s
(ti+1 − ti ), f ′ 1[ti ,ti+1 ) , gj,t
1[ti ,ti+1 ) kψj,t
k
18
Table 1: Numerical computation of error bounds on
nonlinear optical cavity
k Error Bound
19
0.2366
29
0.2115
39
0.1970
49
0.1872
59
0.1799
69
0.1742
79
0.1696
89
0.1658
99
0.1625
where f ′ = α1[0,t] , ψt′ =
we have
(k) ∗
Ut
ψ − ψt′
2
PL′t
′
j=1 ψj,t ,
∗
(k)
Ut − Ut
ψ
for the Kerr-
′
′
′
and ψj,t
= u′j,t ⊗ e(gj,t
) with gj,t
∈ S′. Now, using (22)
= kψk2 + kψt′ k2
′
−2
Lt
X
j=1
′
ke(gj,t
)kℜ
nD
0
(k;α,β ′ (0)) (k;α,β ′ (1))
Tt1 −t0 j Tt2 −t1 j
(α,β ′ (ℓ))
· · · Tt−tℓ j u′j,t
Eo
(24)
′
where βj′ (i) = gj,t
1[ti ,ti+1 ) (s) for any s ∈ [ti , ti+1 ). To find an appropriate ψt′ , we set up the
cost function
!
Lt
nD
Eo 1/2
X
′ (0))
′ (1))
′ (ℓ))
(k;α,β
(k;α,β
(k;α,β
′
ke(gj,t
)kℜ 0 Tt1 −t0 j Tt2 −t1 j · · · Tt−tℓ j u′j,t
Jk (ψt′ ) = 1 + kψt′ k2 − 2
,
j=1
(25)
PL′t ′
′
′
uj,t ⊗ e(gj,t
). We then choose u′j,t and gj,t
for j = 1, 2, . . . , L′t such that
for any ψt′ = j=1
it is a local minimizer of Jk .
For computational simplicity, let us fix Lt = 1 and take ti+1 − ti = 0.5 for all i =
0, 1, 2, . . . , ℓ. With t = tℓ+1 = 5, we then have that ℓ = 9. We take an initialize
′
guess at ψ1,t
= |0i ⊗ e(α1[0,T ]). Using the general purpose unconstrained optimiza′
tion function fminunc in Matlab, a local minimizer ψt′ = u′1,t ⊗ e(g1,t
) was found as
′
′
u1,t = 0.9999 |0i − (0.0024 − 0.0094ı) |1i − 0.0001 |2i and g1,t = (0.0866 + 0.0462ı)1[t0,t1 ) +
(0.0882 + 0.0471ı)1[t1,T ) . The
cost is J(ψt′ ) = 0.0096. Consider r = s = 2.
corresponding
(k)
∗
ψ for various values of k are shown in Table 1.
Using ψt′ , error bounds on
Ut − Ut
Recall that the dimension of the reduced subspace is k + 1.
3.3.2
Atom-cavity model:
Consider a three-level atom coupled to an optical cavity, which itself is coupled to a single
external coherent field (m = 1) [11]. Let H0 = C3 ⊗ ℓ2 . We will use |ei = (1, 0, 0)⊤ ,
19
|+i = (0, 1, 0)⊤ , and |−i = (0, 0, 1)⊤ to denote the canonical basis vectors in Cm . We also
consider |ni, denoting the normalized n-photon Fock state of the system, as the basis vector
of ℓ2 (as in the previous example). Let D0 = C3 ⊗ span{|ni | n ∈ Z+ }. This atom-cavity
system is then described by the following parameters:
√
H = ıχ (σ+ ⊗ a − σ− ⊗ a∗ ) ,
S = I ⊗ I,
L = I ⊗ λa,
where λ, χ > 0, σ+ = |ei h+|, and σ− = |+i he|. From Lemma 12 of [16], we have that
Condition 1(a) and 2(a) holds for this atom-cavity model.
Consider H(k) = C3 ⊗ span {|ni | n = 0, 1, 2, . . . , k} and a system approximation of
the form (23). That is, we are approximating the dynamics of the harmonic oscillator.
Recall that Conditions 1(b) and 2(b) hold immediately because D (k) = H(k) has finite
dimension. We also see that H(k) ⊂ D0 . Hence, Assumption 1 holds. Similar to the
previous example, we note that M(k) = C3 ⊗ |k + 1i. Now consider K(k) = C3 ⊗ |ki and
(αβ)
γk
= 12 (λk + |α − β|2 ). Using similar derivation to the previous example, we will now
show that Assumptions 2 and 3 hold for the atom-cavity model and the approximation.
Assumption 2(a): Note that
√
=
β
λ
+
χσ
a∗ |K(k) .
−
K(k)
√
√
(k;αβ)
Thus, we have that Assumption 2(a) holds with qL
= k + 1 |β| λ + χ .
L(αβ) − L(k;αβ)
Assumption 2(b): This assumption follows for the defined K(k) because
ker L(αβ) − L(k;αβ) H(k) = C3 ⊗ span {|ni | n = 0, 1, 2, . . . , k − 1} for all α, β ∈ C.
∗
Assumption 2(c): Note that L(k) PK(k) = 0, PK(k) L(k) PH(k) = 0,
∗
PK(k) L(k) L(k) PH(k) ⊖K(k) = 0, and PK(k) HPH(k) ⊖K(k) = −ıχPK(k) σ− a∗ PH(k) ⊖K(k) . Also, for any
∗
u ∈ H(k) , PK(k) L(k) L(k) PK(k) u = λkPK(k) u. From these identities and the fact that H (k) is
self-adjoint, we have for any u ∈ H(k) that
ℜ{hPK(k) L(k;αβ) u, PK(k) ui} = ℜ{hPK(k) L(k;αβ) (PK(k) + PH(k) ⊖K(k) ) u, PK(k) ui}
1
= − λk + |α|2 + |β|2 − ℜ{α∗ β} kPK(k) uk2
2 nD
Eo
√
+ ℜ PK(k) β λ + χσ− a∗ PH(k) ⊖K(k) u, PK(k) u
= −g(k, α, β) kPK(k) uk2 + h(k, α, β, u).
Noticing that |α|2 + |β|2 − 2ℜ{α∗ β} = hα, αi + hβ, βi − hα, βi − hα, βi = |α − β|2. Also,
note that
Eo
√
nD
∗
ℜ PK(k) β λ + χσ− a PH(k) ⊖K(k) u, PK(k) u
√
≤ PK(k) β λ + χσ− a∗ PH(k) ⊖K(k) u kPK(k) uk
√
√
≤ k χ + |β| λ kuk kPK(k) uk .
20
(αβ)
Therefore, we have that Assumption
2(c) holds for the defined K(k) and the defined γk
√
√
(k;αβ)
with qa
= k χ + |β| λ .
Assumption 2(d): For any u ∈ H0 , note that L∗ LPM(k) u = λ(k + 1)PM(k) u and
ıHPM(k) u = ı (∆(k + 1) + χk(k + 1)) PM(k) u. Also, Thus, we have that
(αβ)
(αβ)
ℜ{hTt L(αβ) PM(k) u, Tt PM(k) ui}
1
(αβ)
2
2
∗
= − λ(k + 1) + |α| + |β| − 2ℜ{α β} Tt PM(k) u
2
nD
Eo
√
(αβ)
(αβ)
∗
∗
+ ℜ Tt
λ (βa − α a) PM(k) u, Tt PM(k) u
nD
Eo
(αβ)
(αβ)
+ ℜ Tt χ (σ− a∗ − σ+ a) PM(k) u, Tt PM(k) u
(αβ)
= −ĝ(k, α, β) Tt
PM(k) u
2
2
+ ĥ(t, k, α, β, u).
(αβ)
Similar to the previous derivation, using that Tt
is a contraction, we
As see that
√
√
(αβ)
(k;αβ)
(k)
sumption 2(d) holds for the defined K and γk
with qe
= k + 1 χ + |α| λ +
√
√
k + 2 χ + |β| λ .
(αβ)
(k;αβ)
(k;αβ)
(k;αβ)
Assumption 3: From the defined γk , qL
, qa
, qe
, we see that this assumption holds for any r, s ∈ Z+ such that r + s ≥ 3.
Finally, because Assumptions 1-3 hold, Lemma 4, Corollary 1, and Theorem 1 can be
applied to obtain error bounds on the finite-dimensional approximations.
4
Error bounds for adiabatic elimination approximations
When an open quantum system comprises subsystems evolving at two well-separated
timescales, the system dynamics can be approximated by eliminating the fast variables
from the model description. This method is known as adiabatic elimination in the physics
literature and singular perturbation in the applied mathematics literature. In this section,
we will establish error bounds for this type of finite dimensional approximation of open
quantum systems, when the slow subsystem lives on a finite-dimensional subspace.
(k)
Let Ut satisfying (2) describe the time evolution of the original Markov open quantum
system to be approximated. We set H(k) = H and D (k) = D. We also let Ut satisfying
(1) describe the time evolution of the adiabatic elimination approximation defined on a
finite-dimensional subspace H0 ⊂ H. In this setting, D0 = H0 . We note that Conditions
1(a) and 2(a) hold immediately because H0 is finite dimensional. As in [16], we assume
the following.
21
Assumption 4 (Singular scaling). For i, j = 1, 2, . . . , m, there exists operators Y , Y ∗ , A,
A∗ , B, B ∗ , Fj , Fj∗ , Gj , G∗j , Wij , Wij∗ with the common invariant domain D such that
"
ıH
(k)
#
m
1 X (k) ∗ (k)
Li Li
= k 2 Y + kA + B,
−
2 i=1
(k) ∗
Lj
= kFj + Gj ,
(k) ∗
Sji
= Wij .
Assumption 5 (Structural requirements). The subspace H0 ⊂ H is a closed subspace such
that
1. D0 = PH0 D ⊂ D
2. Y PH0 = 0 on D
3. There exists Ỹ , Ỹ ∗ with the common invariant domain D so that Ỹ Y = Y Ỹ = PH⊥0 .
4. Fj∗ PH0 = 0 on D for all j = 1, 2, . . . , m
5. PH0 APH0 = 0 on D.
Assumption 6 (Limit coefficients). The approximating system operators (S, L, H) are
such that
∗
Sji
=
m
X
ℓ=1
∗
PH0 Wiℓ Fℓ Ỹ Fj + δℓj PH0 ,
L∗j = PH0 Gj − AỸ Fj PH0 ,
o
n
H = ℑ PH0 B − AỸ A PH0 .
Let us now re-state an important result.
Lemma 5 ( [16, Lemma 10]). Suppose Assumptions 4, 5, and 6 hold. The linear operators
S, L, H, defined in Assumption 6, have the common invariant domain D0 . The operator S
is unitary and H is self-adjoint.
For any α, β ∈ Cm and any v ∈ D, let us define
A(αβ) v :=
A+
m
X
j=1
B (αβ) v :=
Fj βj −
m
X
αi∗ Wij Fj∗
i,j=1
!
v,
!
m
m
X
X
|α|2 + |β|2
−
+B+
Gj βj +
αi∗ Wij βj − G∗j
v.
2
j=1
i,j=1
We now introduce an additional assumption required in obtaining our error bound results.
22
Assumption 7 (Boundedness of operators). For any k ∈ Z+ and any α, β ∈ Cm , we have
that
1 (αβ)
(k;α,β)
(αβ)
(αβ)
(αβ)
B
− A Ỹ A
,
−
M1
:= Ỹ PH⊥0 A
k
H0
1 (αβ)
(k;α,β)
(αβ)
(αβ)
(αβ)
−
M2
:= Ỹ PH⊥0 A
B
− A Ỹ A
L(αβ)
k
1
(αβ)
(αβ)
(αβ)
(αβ)
− A Ỹ A
Ỹ PH⊥0 B
+ B (αβ) Ỹ A(αβ) + A(αβ) − B
k
H0
(k;αβ)
are finite (i.e., 0 ≤ M1
(k;αβ)
, M2
< ∞).
Lemma 6. Suppose Assumptions 4, 5, and 6 hold. Then for any k ∈ Z+ , any α, β ∈ Cm ,
any u ∈ H0 , and any t ≥ 0, it holds that
1
(k;αβ)
(αβ)
(k;α,β)
(k;α,β)
u ≤
Tt
− Tt
kuk.
(26)
2M1
+ tM2
k
(αβ)
(k;αβ)
(αβ)
Moreover, if Tt
and Tt
are also norm continuous for each α, β, with k(I −Tt )uk ≤
(αβ)
(k;αβ)
(k;αβ)
N1 (t)kuk and k(I − Tt
)uk ≤ N2
(t)kuk for some continuous nonnegative func(αβ)
(k;αβ)
tions N1
and N2
, then
1 (k;α,β) (αβ)
(k;αβ)
(k;α,β)
(k;αβ)
(αβ)
M1
(N1 (t) + N2
(t)) + tM2
kuk.
(27)
u ≤
Tt
− Tt
k
Proof. First note, from the definition of a strongly continuous semigroup, that [18]
(αβ)
1. T0
(k;αβ)
= T0
2.
d (αβ)
T
u
dt t
3.
d (k;αβ)
T
u
dt t
= I for all k ∈ Z+ ,
= L(αβ) u for all u ∈ H0 (since H0 is finite-dimensional),
= L(k;αβ) u for all u ∈ Dom(L(k;αβ) ).
From the above properties and Assumption 5, we can write for all u ∈ H0 and all t ≥ 0
that
d (k;αβ)
(k;αβ)
(αβ)
(αβ)
u
u = L(k;αβ) Tt
− L(αβ) Tt
Tt
− Tt
dt
(αβ)
(k;αβ)
(αβ)
u + L(k;αβ) − L(αβ) H0 Tt u (28)
− Tt
= L(k;αβ) Tt
(αβ)
(k;αβ)
u = 0. Since L(k;αβ) is a generator of a semigroup on a Hilbert space,
with T0 − T0
(k;αβ)
(αβ)
(αβ)
T0
− T0
u ∈ Dom(L(k;αβ) ) (due to Assumption 5), and L(k;αβ) − L(αβ) H0 Tt u ∈
C 1 ([0, t]; H) (the class of continuously differentiable functions from [0, t] to H), a unique
solution of (28) exists and is given by [18, Thm 3.1.3]
Z t
(k;αβ)
(k;αβ)
(αβ)
(29)
L(k;αβ) − L(αβ) Tτ(αβ) udτ
u=
Tt−τ
Tt
− Tt
0
23
for all t ≥ 0 and all u ∈ H0 .
From Assumption 4, we now note that
L(k;αβ) u = k 2 Y + kA(αβ) + kB (αβ) u
for any u ∈ D and any α, β ∈ Cm . Using Assumptions 4 and 6, it has been shown
that [16, p. 3146]
L(αβ) u = PH0 B (αβ) − A(αβ) Ỹ A(αβ) u
for any u ∈ D0 and any α, β ∈ Cm .
From the above identities and Assumption 4, for any v ∈ H0 , we have that [16]
L(k;αβ) − L(αβ) v
1
1
= L(k;αβ) − L(αβ) − L(k;αβ) Ỹ A(αβ) + L(k;αβ) Ỹ A(αβ)
k
k
1
− 2 L(k;αβ) Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ)
k
1 (k;αβ)
(αβ)
(αβ)
(αβ)
v
− A Ỹ A
Ỹ PH⊥0 B
+ 2L
k
1 (k;αβ)
1 (αβ)
=
B
Ỹ A(αβ) + A(αβ) Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ)
L
Ỹ A(αβ) −
k
k
1
+ 2 L(k;αβ) Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ)
k
1 (αβ)
(αβ)
(αβ)
(αβ)
− 2B
v.
− A Ỹ A
Ỹ PH⊥0 B
k
We also note, using integration by parts (in a similar manner to [19, p. 898] and [20, Eq.
(2.2)]), that
Z t
(k;αβ)
(k;αβ)
(αβ)
u
Tt−τ L(k;αβ) Ỹ A(αβ) Tτ(αβ) udτ = Tt
Ỹ A(αβ) − Ỹ A(αβ) Tt
0
Z t
(k;αβ)
+
Tt−τ Ỹ A(αβ) L(αβ) Tτ(αβ) udτ
0
(k;αβ)
(k;αβ)
(αβ)
(αβ)
where we have used that − dτd Tt−τ u = Tt−τ L(k;αβ) u and dτd Ỹ A(αβ) Tτ u = Ỹ A(αβ) L(αβ) Tτ
[18]. Similarly, we have that
Z t
(k;αβ)
Tt−τ L(k;αβ) Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ) Tτ(αβ) udτ
0
(k;αβ)
(αβ)
u
= Tt
Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ) − Ỹ PH⊥0 B (αβ) − A(αβ) Ỹ A(αβ) Tt
Z t
(k;αβ)
(αβ)
(αβ)
(αβ)
+
Tt−τ Ỹ PH⊥0 B
L(αβ) Tτ(αβ) udτ.
− A Ỹ A
0
24
u
(αβ)
Now, because Tt u ∈ H0 for all t ≥ 0, by substituting the above three identities into
(29), we have that
(k;αβ)
(αβ)
u
Tt
− Tt
1
1 (αβ)
(k;αβ)
(αβ)
(αβ)
(αβ)
=
u
−
T
B
− A Ỹ A
Ỹ PH⊥0 A
k t
k
1 (αβ)
(αβ)
(αβ)
(αβ)
(αβ)
B
− A Ỹ A
Tt u
−
− Ỹ PH⊥0 A
k
Z t
1 αβ
(k;αβ)
(αβ)
(αβ)
(αβ)
+
Tt−τ
−
Ỹ PH⊥0 A
B − A Ỹ A
L(αβ) Tτ(αβ) udτ
k
0
Z t
1 (αβ)
(k;αβ)
(αβ)
(αβ)
(αβ)
(αβ)
(αβ)
(αβ)
− A Ỹ A
Ỹ PH⊥0 B
+
B
Ỹ A
+ A
− B
Tt−τ
k
0
×Tτ(αβ) udτ .
Here, we have used Assumption 5 that PH0 APH0 = 0. We then have that
1 h (k;αβ) (k;αβ)
(αβ)
(k;αβ)
(αβ)
M1
Tt
+ Tt
Tt
− Tt
≤
k
Z t
(k;αβ)
(k;αβ)
(αβ)
+M2
Tt
Tt
dτ kuk.
0
(k;αβ)
(k;αβ)
The result (26) then follows from the fact that Tt
and Tt
are contraction semi(αβ)
(k;αβ)
groups. Moreover, when Tt
and Tt
are also norm continuous and satisfy k(I −
(αβ)
(αβ)
(k;αβ)
(k;αβ)
Tt )uk ≤ N1 (t)kuk and k(I − Tt
)uk ≤ N2
(t)kuk as stipulated in the lemma,
we have
(k;αβ)
(αβ)
u
Tt
− Tt
1
1 (αβ)
(k;αβ)
(αβ)
(αβ)
(αβ)
(αβ)
=
T
Ỹ PH⊥0 A
−
B
− A Ỹ A
(I − Tt )u
k t
k
1 (αβ)
(k;αβ)
(αβ)
(αβ)
(αβ)
(αβ)
−
− (I − Tt
)Ỹ PH⊥0 A
Tt u
B
− A Ỹ A
k
Z t
1 αβ
(k;αβ)
(αβ)
(αβ)
(αβ)
L(αβ) Tτ(αβ) udτ
B − A Ỹ A
−
Ỹ PH⊥0 A
+
Tt−τ
k
0
Z t
1 (αβ)
(k;αβ)
(αβ)
(αβ)
(αβ)
(αβ)
(αβ)
(αβ)
− A Ỹ A
Ỹ PH⊥0 B
B
Ỹ A
+ A
− B
+
Tt−τ
k
0
×Tτ(αβ) udτ .
(αβ)
(k;αβ)
From this and the contractivity of Tt
and Tt
, it follows that
i
1 h (k;αβ) (αβ)
(k;αβ)
(k;αβ)
(k;αβ)
(αβ)
N1 (t) + N2
(t) + M2
t kuk.
Tt
M1
− Tt
≤
k
This establishes the lemma statement.
25
Recall that S′ ⊂ L2 ([0, T ]; Cm) is the set of all simple functions in L2 ([0, T ]; Cm), which
is dense in L2 ([0, T ]; Cm ). Similar to the previous section, let U0 = {u ⊗ e(f ) | u ∈ H0 , f ∈
S′ } and U (k) = {u ⊗ e(f ) | u ∈ H(k) , f ∈ S′ } for k > 0.
Lemma 7. Suppose Assumptions 4, 5, 6, and 7 hold. Then for any k ∈ Z+ , any ψ1 , ψ2 ∈
U0 , and any t ∈ [0, T ] with 0 ≤ T < ∞, we have that
ℓ
D
∗
E
1 X
(k;α1 (i)α2 (i))
(k;α1 (i)α2 (i))
(k)
kψ1 kkψ2 k
2M1
+ (ti+1 − ti )M2
Ut − Ut ψ1 , ψ2 ≤
k i=0
(30)
P
(αβ)
and
where t0 = 0, tℓ+1 = t, and fj = ℓi=0 αj (i)1[ti ,ti+1 ) for j = 1, 2. Moreover, if Tt
(k;αβ)
(αβ)
(αβ)
Tt
are also norm continuous for each α, β, with k(I − Tt )uk ≤ N1 (t)kuk and
(k;αβ)
(k;αβ)
(αβ)
k(I − Tt
)uk ≤ N2
(t)kuk for some continuous nonnegative functions N1
and
(k;αβ)
N2
, then
D
(k)
Ut
− Ut
∗
ψ1 , ψ2
E
ℓ
1 X (k;α1 (i)α2 (i)) (αβ)
(k;αβ)
≤
M1
N1 (ti+1 − ti ) + N2
(ti+1 − ti )
k i=0
(k;α1 (i)α2 (i))
+(ti+1 − ti )M2
kψ1 kkψ2 k
(31)
Proof. This proof follows similar arguments to the proof of Lemma 4 using the result (26)
established in Lemma 6.
Corollary 2. Suppose Assumptions 4, 5, 6, and 7 hold. For any t ∈ [0, T ] with 0 < T < ∞,
any ψ1 = u1 ⊗ e(f1 ), ψ2 = u2 ⊗ e(f2 ) ∈ H0 ⊗ F , we have that
D
∗
E
(k)
′
′
Ut − Ut ψ1 , ψ2 ≤ 2 ku1 kke(f1 ) − e(f1 )kkψ2 k + ku2 kke(f2 ) − e(f2 )kkψ1 k
ℓ
1 X
(k;α (i)α (i))
(k;α (i)α (i))
kψ1′ kkψ2′ k.
2M1 1 2 + (ti+1 − ti )M2 1 2
+
k i=0
(32)
for any ψj′ = uj ⊗ e(fj′ ) ∈ U (k) with for some ℓ ∈ Z+ and a sequence t0 = 0 < t1 < . . . <
P
P
tℓ < tℓ+1 = t such that f1′ = ℓi=0 α1 (i)1[ti ,ti+1 ) and f2′ = ℓi=0 α2 (i)1[ti ,ti+1 ) . If in addition,
(αβ)
(k;αβ)
(αβ)
(αβ)
Tt
and Tt
are also norm continuous for each α, β, with k(I −Tt )uk ≤ N1 (t)kuk
(k;αβ)
(k;αβ)
(αβ)
and k(I − Tt
)uk ≤ N2
(t)kuk for some continuous nonnegative functions N1
and
(k;αβ)
N2
, then
D
∗
E
(k)
Ut − Ut ψ1 , ψ2 ≤ 2 ku1kke(f1 ) − e(f1′ )kkψ2 k + ku2 kke(f2 ) − e(f2′ )kkψ1 k
ℓ
1 X (k;α1 (i)α2 (i)) (α(i)β(i))
M1
N1
(ti+1 − ti )
k i=0
(k;α(i)β(i))
(k;α (i)α (i))
+N2
(ti+1 − ti ) +(ti+1 − ti )M2 1 2
kψ1′ kkψ2′ k.
+
(33)
26
Moreover, it holds that
D
∗
E
(k)
Ut − Ut ψ1 , ψ2
lim
k→∞
= 0.
(34)
Proof. This proof follows similar arguments to the proof of Corollary 1 using (30) and (31)
established in Lemma 7.
Theorem 2. Suppose Assumptions 4, 5, 6, and 7 hold. Let 0 < T < ∞. For any t ∈ [0, T ],
PL′t ′
′
′
consider any L′t ∈ Z+ , any ψt′ = j=1
ψj,t , with ψj,t
= u′j,t ⊗ e(gj,t
) ∈ U (k) and u′j,t 6= 0.
Also, consider any f ′ ∈ S′ . P
Let ℓ be a positive integer and
Pℓ t0 < t1 < . . . < tℓ < tℓ+1 =(k)t
ℓ
′
′
be a sequence such that f = i=0 α(i)1[ti ,ti+1 ) and gj,t = i=0 βj,t (i)1[ti ,ti+1 ) . Let u ∈ H
with kuk = 1, and |f i = e(f )/ke(f )k ∈ F (i.e., |f i is a coherent state with amplitude f ),
and ψ = u ⊗ |f i. Then,
(k)
Ut
− Ut
∗
2
ψ
≤ 4 k|f i − |f ′ik + kUt ∗ ψ − ψt′ k
L′
ℓ
t X
2X
(k;α(i)βj,t (i))
(k;α(i)βj,t (i))
′
2M1
+ (ti+1 − ti )M2
kψj,t
k, (35)
+
k j=1 i=1
(αβ)
(k;αβ)
If in addition, Tt
and Tt
are also norm continuous for each α, β, with k(I −
(αβ)
(αβ)
(k;αβ)
(k;αβ)
Tt )uk ≤ N1 (t)kuk and k(I − Tt
)uk ≤ N2
(t)kuk for some continuous nonneg(k;αβ)
(αβ)
ative functions N1
and N2
, then
(k)
Ut
− Ut
∗
2
ψ
Moreover, it holds that
≤ 4 k|f i − |f ′ik + kUt ∗ ψ − ψt′ k
L′
ℓ
t X
2X
(k;α(i)βj,t (i))
(α(i)β(i))
M1
N1
(ti+1 − ti )
+
k j=1 i=1
(k;α(i)βj,t (i))
(k;α(i)β(i))
′
kψj,t
k.
+N2
(ti+1 − ti ) + (ti+1 − ti )M2
lim
k→∞
for any t ∈ [0, T ] with 0 < T < ∞.
(k)
Ut
− Ut
∗
ψ = 0.
(36)
(37)
Remark 2. As with Theorem 1, a stronger
of strong convergence uniformly over
result
∗
(k)
compact time intervals limk→∞ sup0≤t≤T Ut − Ut ψ = 0 has been established in [16,
Theorem 11] for adiabatic elimination based on a Trotter-Kato theorem, but without error
bounds for finite values of k.
Proof. This proof follows similar arguments to the proof of Theorem 1 using (32) and (33)
established in Corollary 2.
27
4.1
Adiabatic elimination examples
4.1.1
Elimination of a harmonic oscillator:
Consider a class of open quantum systems that comprises a finite-dimensional atomic
system coupled to a harmonic oscillator which is driven by m external coherent fields
(originally presented in [16]). Let H = H′ ⊗ ℓ2 , where H′ is a finite-dimensional
Hilbert
P∞
2
2
space and ℓ is the space of infinite complex-valued sequences with n=1 |xn | < ∞. Similar
to Example 3.3.1, let {|ni}n≥0 be an orthonormal Fock state basis of ℓ2 . On this basis, the
annihilation, creation, and number operators can be defined (see, e.g., [16]) satisfying
√
√
a∗ |ni = n + 1 |n + 1i ,
a∗ a |ni = n |ni ,
a |ni = n |n − 1i ,
respectively. Following [16], we choose the dense domain D = H′ ⊗span{|ni | n ∈ Z+ }. Let
P
(k) ∗ (k)
us define K (k) = ıH (k) − 12 m
Li ). Consider the system operators (S (k) , L(k) , H (k) )
i=1 (Li
defined such that
K (k) = k 2 E11 ⊗ a∗ a + k(E10 ⊗ a∗ + E01 ⊗ a) + E00 ⊗ I,
(k) ∗
Lj
(k) ∗
Sji
= kFj ⊗ a∗ + Gj ⊗ I,
= Wij ⊗ I
where E11 , E10 , E01 , E00 , Fj , Gj , Wij are bounded operators on the finite-dimensional space
H′ .
Now consider H0 = D0 = H′ ⊗ C |0i. That is, the harmonic oscillator is eliminated
from the model as it is forced into its ground state (i.e., |0i) in the limit as k → ∞. In
quantum optics, this process is the adiabatic elimination of an optical cavity in the strong
damping limit. Now consider an approximation system with the operators (S, L, H) which
are defined such that
∗
Sji
=
m
X
ℓ=1
L∗j
Wiℓ Fℓ∗ (E11 )−1 Fj + δℓj ⊗ I.
= Gj − E01 (E11 )−1 Fj ⊗ I,
H = ℑ E00 − E01 (E11 )−1 E10 ⊗ I.
Here, we stress that S is unitary and H is self-adjoint [16, Proposition 14].
It has been shown in [16] that Conditions 1 and 2 hold for the above systems. The
original system satisfies Assumption 4 with Y = E11 ⊗ a∗ a, A = E10 ⊗ a∗ + E01 ⊗ a and
B = E00 ⊗ I. Suppose E11 has a bounded inverse, then Assumption 5 is satisfied with
Ỹ which is defined such that Ỹ ψ ⊗ |ni = n1 (E11 )−1 ψ ⊗ |ni for n ≥ 1 and ψ ∈ H′ . The
operators (S, L, H) also satisfy Assumption 6. It now remains to show that Assumption 7
holds.
P
(αβ)
Note that PH⊥0 B (αβ) PH0 = 0. Now let Q(β) = E10 + m
=
j=1 Fj βj and let P
P
P
2
2
m
m
|α| +|β|
∗
(β)
(αβ)
∗
+ E00 + j=1 Gj βj + i,j=1 αi Wij βj − Gj . Note that Q
and P
are
− 2
28
bounded operators on the finite-dimensional Hilbert space H′ for any α, β ∈ Cm since E10 ,
E00 , Fj , Gj and Wij are bounded operator on H′ . For any ψ ∈ H′ , we have that
(Ỹ A(αβ) )ψ ⊗ |0i = (E11 )−1 Q(β) ψ ⊗ |1i ,
1
(Ỹ PH⊥0 A(αβ) Ỹ A(αβ) )ψ ⊗ |0i = √ (E11 )−1 Q(β) ψ ⊗ |2i ,
2
(αβ)
(αβ)
(αβ)
(B
Ỹ A )ψ ⊗ |0i = P
(E11 )−1 Q(β) ψ ⊗ |1i ,
r
3 (β)
(A(αβ) Ỹ PH⊥0 A(αβ) Ỹ A(αβ) )ψ ⊗ |0i =
Q (E11 )−1 Q(β) ψ ⊗ |3i
2
!
m
X
αi∗ Wij Fj∗ (E11 )−1 Q(β) ψ ⊗ |1i ,
+ E01 −
i,j=1
1
(B (αβ) Ỹ PH⊥0 A(αβ) Ỹ A(αβ) )ψ ⊗ |0i = √ P (αβ) (E11 )−1 Q(β) ψ ⊗ |2i .
2
From these identities, we see that Assumption 7 holds because E11 has a bounded inverse, E01 , Fj , Wij Q(β) , P (αβ) are bounded operators, and L(αβ) is defined on the finitedimensional subspace H0 (i.e., it is also a bounded operator). Thus, the conditions of
Lemma 6, Lemma 7, Corollary 2, and Theorem 2 have been verified.
4.1.2
Atom-cavity model [16, Example 15]:
Consider a system consisting of a three-level atom coupled to an optical cavity. The
cavity and the uncoupled leg of the atom is driven by an external coherent field (m = 1).
Let H = C3 ⊗ ℓ2 . As in the previous example, we consider the orthonormal Fock state
basis {|ni}n≥0 of ℓ2 . We will use |ei = (1, 0, 0)⊤, |+i = (0, 1, 0)⊤ , and |−i = (0, 0, 1)⊤
to denote the canonical basis vectors in C3 . In the basis {|ei , |+i , |−i}, let us define
(+)
(−)
(±)
(±) ∗
σ+ = |ei h+| and σ+ = |ei h−|. We also define σ− = σ+ and P− = |−i h−|. Here,
D = C3 ⊗ span{|ni | n ∈ Z+ }. Under the rotating wave approximation and in the rotating
frame of reference, the system is described by the following operators: [16]
S (k) = I,
√
L(k) = I ⊗ k γa,
(+)
(−)
(+)
(−) ∗
(k)
2
∗
H = ıgk σ+ ⊗ a − σ− ⊗ a + ık σ+ α − σ− α ⊗ I
where γ, g > 0. Here, α ∈ C is the amplitude of the external coherent field driving the
cavity and the uncoupled leg of the atom.
Now consider H0 = D0 = span{|+i ⊗ |0i , |−i ⊗ |0i}. That is, the cavity oscillator and
the excited state of the atom (i.e., |ei) is eliminated from the model in the limit as k → ∞.
Consider an approximating system described the operators (S, L, H) which are defined as
√
α γ (+) (−)
σ σ ⊗ I,
H = 0.
S = (I − 2P− ) ⊗ I,
L=−
g − +
29
It can be easily verified that S is unitary and H is self-adjoint.
Again, it has been shown in [16] the above
systems satisfy Conditions
1 and 2. We then
(+)
(−)
(−)
γ
+
∗
∗
see that Assumption 4 with Y = − 2 ⊗a a+g σ− ⊗ a − σ+ ⊗ a , A = σ− α∗ − σ+ α ⊗
√
I, B = 0, F = γ ⊗ a∗ , G = 0, and W = I. Let us define
Hj = span{|+i ⊗ |ji , |−i ⊗ |ji , |ei ⊗ |j − 1i},
for j ∈ Z+ .
Assumption 5 holds with Ỹ which is defined, with respect to the basis {|+i ⊗ |ji , |−i ⊗
|ji , |ei ⊗ |j − 1i}, as [16]
√
γ
(j − 1) 0 g j
2
1
j(j − 1)γ 2
2dj
,
0
0
Ỹ |Hj = −
d
=
+ jg 2 .
j
jγ
√
dj
4
jγ
−g j
0
2
It can also be seen that Assumption 6 holds for the defined (S, L, H). Thus, it remains to
show that Assumption 7 holds. Note that B (αβ) = − 12 |α − β|2 and thus, PH⊥0 B (αβ) PH0 = 0.
Then note that
Ỹ A(αβ) PH0 ⊂ H1 ,
Ỹ PH⊥0 A(αβ) Ỹ A(αβ) PH0 ⊂ H2 ,
B (αβ) Ỹ A(αβ) PH0 ⊂ H1 ,
A(αβ) Ỹ PH⊥0 A(αβ) Ỹ A(αβ) PH0 ⊂ H3 ⊕ H1 ,
B (αβ) Ỹ PH⊥0 A(αβ) Ỹ A(αβ) PH0 ⊂ H2 .
From the above relations and the fact L(αβ) is defined on the finite-dimensional subspace
H0 (i.e., it is a bounded operator), we have that Assumption 7 holds.
Numerical example of the atom-cavity model: Consider γ = 25, g = 5, and t ∈
[0, T ] with T = 1. Let
∗⊗ |0i) ⊗ |α1[0,T ]i with α = 0.1. We will now compute
ψ = (|−i
(k)
the error bound on
Ut − Ut
ψ for different values of k. From (35) (established in
Theorem 2) and the fact that α1[0,T ] ∈ S′ is a simple function, we have that
∗
(k)
Ut − Ut
ψ
2
≤ 4 (kUt ∗ ψ − ψt′ k)
L′
ℓ
t X
2X
(k;α(i)βj,t (i))
(k;α(i)βj,t (i))
′
2M1
+ (ti+1 − ti )M2
kψj,t
k.
+
k j=1 i=1
PL′t ′
′
′
′
where ψt′ = j=1
ψj,t , and ψj,t
= u′j,t ⊗ e(gj,t
) with gj,t
∈ S′ . Similar to Example 3.3.1,
kUt ∗ ψ − ψt′ k is bounded by (24). Again,
find an appropriate ψt′ , let us define the cost
PLto
t
′
′
′
) ∈ U (k) . We then find ψt′ that is a
function J(ψt ) as in (25) for any ψt = j=1 u′j,t ⊗ e(gj,t
local minimizer of J by numerical optimization.
30
Table 2: Numerical computation of error bounds on
elimination approximation in the atom-cavity example
k
Error bound
4
10
0.9347
5
10
0.2957
6
10
0.0309
107
0.0131
8
10
0.0096
Ut −
(k)
Ut
∗
ψ
for the adiabatic
For computational simplicity, we fix Lt = 5 and set ti+1 − ti = 10−3 for all i =
0, 1, 2, . . . , ℓ. With t = T , we then have that ℓ = 999. Note that simultaneously optimizing over 1000 time intervals is computationally intensive. Thus, we simplify the
computation further by optimizing sequentially over blocks of 10 time intervals at a time.
Thus optimization is done over 100 blocks. Optimization of the first block is initialized
′
with ψj,t
= (|−i ⊗ |0i) ⊗ e(α1[0,t10 ] ) for all j = 1, 2, . . . , Lt . The optimization result of
each block is then used to initialize the optimization of the next block in the sequence.
′
′
A local minimizer ψj,t
= u′j,t ⊗ e(gj,t
), for j = 1, 2, . . . , Lt , was found using Matlab as
′
before, given by u1,t = (0.1008 + 0.1549ı)(|+i ⊗ |0i) + (0.2311 + 0.1012ı)(|−i ⊗ |0i),
u′2,t = (0.0949 − 0.0031ı)(|+i ⊗ |0i) + (0.1500 + 0.0270ı)(|−i ⊗ |0i), u′3,t = (−0.0245 +
0020ı)(|+i⊗|0i)+(0.2012+0.0311ı)(|−i⊗|0i), u′4,t = (0.3266−0.1431ı)(|+i⊗|0i)+(0.2928−
0.0995ı)(|−i ⊗ |0i), u′5,t = −(0.4956 + 0.0101ı)(|+i ⊗ |0i) + (0.1232 − 0.0591ı)(|−i ⊗ |0i).
′
′
The
overall resulting optimized cost is J(ψt ) = 0.0046. Using ψt , the error bounds on
(k)
Ut − Ut
5
∗
ψ
using (32) for various values of k are shown in Table 2.
Conclusion
This work has developed a framework for developing error bounds for finite dimensional approximations of input-output quantum stochastic models defined on infinite-dimensional
underlying Hilbert spaces, with possibly unbounded coefficients in their QSDEs. The
framework exploits a contractive semigroup that can be associated with the QDESs. This
gives for the first time error bound expressions for two types of approximations that are often employed in the literature, subspace truncation and adiabatic elimination. The bounds
are in principle computable and vanish for each t in the limit as the parameter k, representing the dimension of the approximating subspace in the case of subspace truncation
approximation or a large scaling parameter in the case of adiabatic elimination, goes to ∞.
The theory developed was applied to some physical examples taken from the literature.
There are several directions for further investigation along the theme initiated in this
(k) ∗
paper.
Devising
a
more
efficient
method
for
computing
bounds
for
the
term
kU
ψ−
t
P ′
P ′
∗
j ψj,t k for subspace truncation and kUt ψ −
j ψj,t k for adiabatic elimination, beyond
the computationally intensive optimization based approach that was considered herein, will
be important. Tensor network methods that have recently met a lot of success for efficient
31
simulation of one dimensional many-body systems could potentially be important for this
purpose. There also remains the question of the conservatism in the error bounds and if
there could be tighter bounds that can be achieved by using a different set of assumptions.
In the numerical example of adiabatic elimination, the bound (35) was employed rather
than the potentially less conservative (36). This is because the latter requires determining
(k;αβ)
(k;αβ)
whether Tt
is a norm continuous semigroup and finding the bounding function N2
,
a non-trivial task in general that deserves further investigation. Moreover, it would be
interesting to see if there are exactly solvable QSDE models of a physical system with an
infinite-dimensional system Hilbert space for initial states that lie in a finite-dimensional
subspace of the original, against which the conservatism of error bounds can be assessed.
The authors are currently unaware of any such exactly solvable models.
Acknowledgements
The authors are grateful for the support of the Australian Research Council under Discovery Project DP130104191.
References
References
[1] R. L. Hudson and K. R. Parthasarathy, “Quantum Ito’s formula and stochastic evolution,”
Commun. Math. Phys., vol. 93, pp. 301–323, 1984.
[2] C. Gardiner and M. Collett, “Input and output in damped quantum systems: Quantum
stochastic differential equations and the master equation,” Phys. Rev. A, vol. 31, pp. 3761–
3774, 1985.
[3] H. M. Wiseman and G. J. Milburn, Quantum Measurement and Control.
versity Press, 2010.
Cambridge Uni-
[4] V. P. Belavkin and S. C. Edwards, “Quantum filtering and optimal control,” in Quantum
Stochastics and Information: Statistics, Filtering and Control (University of Nottingham,
UK, 15 - 22 July 2006), V. P. Belavkin and M. Guta, Eds. Singapore: World Scientific,
2008, pp. 143–205.
[5] H. I. Nurdin, M. R. James, and A. C. Doherty, “Network synthesis of linear dynamical
quantum stochastic systems,” SIAM J. Control Optim., vol. 48, no. 4, pp. 2686–2718, 2009.
[6] L. Bouten, R. van Handel, and M. R. James, “An introduction to quantum filtering,” SIAM
J. Control Optim., vol. 46, pp. 2199–2241, 2007.
[7] L. Bouten and R. van Handel, “On the separation principle of quantum control,” in Quantum
Stochastics and Information: Statistics, Filtering and Control (University of Nottingham,
32
UK, 15 - 22 July 2006), V. P. Belavkin and M. Guta, Eds.
2008, pp. 206–238.
Singapore: World Scientific,
[8] M. R. James, H. I. Nurdin, and I. R. Petersen, “H ∞ control of linear quantum stochastic
systems,” IEEE Trans. Autom. Control, vol. 53, no. 8, pp. 1787–1803, 2008.
[9] H. I. Nurdin, M. R. James, and I. R. Petersen, “Coherent quantum LQG control,” Automatica, vol. 45, pp. 1837–1846, 2009.
[10] J. Kerckhoff, H. I. Nurdin, D. Pavlichin, and H. Mabuchi, “Designing quantum memories
with embedded control: photonic circuits for autonomous quantum error correction,” Phys.
Rev. Lett., vol. 105, pp. 040 502–1–040 502–4, 2010.
[11] L. Duan and H. J. Kimble, “Scalable photonic quantum computation through cavity-assisted
interaction,” Phys. Rev. Lett., vol. 92, pp. 127 902–1–127 902–4, 2004.
[12] H. Mabuchi, “Nonlinear interferometry approach to photonic sequential logic,” Appl. Phys.
Lett., vol. 99, pp. 153 103–1–153 103–3, 2011.
[13] C. W. Gardiner and P. Zoller, Quantum Noise: A Handbook of Markovian and NonMarkovian Quantum Stochastic Methods with Applications to Quantum Optics, 3rd ed.
Berlin and New York: Springer-Verlag, 2004.
[14] O. Techakesari and H. I. Nurdin, “Error bounds on finite-dimensional approximations of
input-output open quantum systems,” in Proceedings of the 2015 IEEE Conference on Decision and Control (Osaka, 15-18 Dec. 2015). IEEE, 2015, pp. 5772–5777.
[15] P.-A. Meyer, Quantum Probability for Probabilists, 2nd ed.
Verlag, 1995.
Berlin-Heidelberg: Springer-
[16] L. Bouten, R. van Handel, and A. Silberfarb, “Approximation and limit theorems for quantum stochastic models with unbounded coefficients,” J. Funct. Anal., vol. 254, pp. 3123–3147,
2008.
[17] F. Fagnola, “On quantum stochastic differential equations with unbounded coefficients,”
Probab. Th. Rel. Fields, vol. 86, pp. 501–516, 1990.
[18] R. F. Curtain and H. J. Zwart, An Introduction to Infinite-Dimensional Linear Systems
Theory, ser. Text in Applied Mathematics. New York: Springer-Verlag, 1995, no. 21.
[19] I. Lasiecka and A. Manitius, “Differentiability and convergence rates of approximating semigroups for retarded function differential equations,” SIAM J. Numer. Anal., vol. 25, no. 4,
pp. 883–907, 1988.
[20] K. Ito and F. Kappel, “The Trotter-Kato theorem and approximation of PDEs,” Mathematics
of Computation, vol. 67, no. 221, pp. 21–44, 1998.
33
| 3cs.SY
|
Multipair Massive MIMO Two-Way Full-Duplex
Relay Systems with Hardware Impairments
arXiv:1708.00742v1 [cs.IT] 2 Aug 2017
Ying Liu, Xipeng Xue, Jiayi Zhang, Xu Li, Linglong Dai, and Shi Jin
Abstract—Hardware impairments, such as phase noise, quantization errors, non-linearities, and noise amplification, have
baneful effects on wireless communications. In this paper, we
investigate the effect of hardware impairments on multipair
massive multiple-input multiple-output (MIMO) two-way fullduplex relay systems with amplify-and-forward scheme. More
specifically, novel closed-form approximate expressions for the
spectral efficiency are derived to obtain some important insights
into the practical design of the considered system. When the
number of relay antennas N increases without bound, we propose
a hardware scaling law, which reveals that the level of hardware
impairments that can be tolerated is roughly proportional to
√
N . This new result inspires us to design low-cost and practical
multipair massive MIMO two-way full-duplex relay systems.
Moreover, the optimal number of relay antennas is derived to
maximize the energy efficiency. Finally, Motor-Carlo simulation
results are provided to validate our analytical results.
I. I NTRODUCTION
The two-way full-duplex (FD) relay system can ideally
achieve almost twice of the spectral efficiency (SE) achieved
by the traditional two-way half-duplex (HD) scheme, since
the relay can transmit and receive signals simultaneously.
However, the practical implementation of the two-way FD
relay is challenging due to the severe self-interference (SI)
caused by FD [1]–[3].
Recently, massive multiple-input multiple-output (MIMO)
has been proposed as an efficient approach to suppress the
SI of two-way FD relay systems in the spatial domain [4].
Different from most of existing works, which consider systems deploying high-cost ideal hardware components, in this
paper we consider a multipair massive MIMO two-way FD
relay system with low-cost non-ideal hardware that suffers
from hardware impairments. In practical systems, the cost
and power consumption increase with the number of radio
frequency (RF) chains. In order to achieve higher energy
efficiency (EE) and/or lower hardware cost, each RF chain
can use some cheap hardware components [5], [6]. However,
low-cost hardware is particularly prone to the impairments
in transceivers, such as quantization errors of low-resolution
This work was supported in part by the National Natural Science Foundation of China (Grant No. 61601020) and the Fundamental Research Funds
for the Central Universities (Grant Nos. 2016RC013, 2017JBM319, and
2016JBZ003). (Corresponding author: [email protected])
Y. Liu, X. Xue, J. Zhang and X. Li are with the School of Electronic and
Information Engineering, Beijing Jiaotong University, Beijing 100044, P. R.
China.
L. Dai is with the Department of Electronic Engineering, Tsinghua University, Beijing 100084, P. R. China.
S. Jin is with the National Mobile Communications Research Laboratory,
Southeast University, Nanjing 210096, P. R. China.
analog to digital converters (ADCs), I/Q-imbalance, phase
noise, and non-linearities [7]–[9].
Although the influence of hardware impairments can be
mitigated by some compensation algorithms, residual impairments still exist due to time-varying and random hardware
characteristics. The effect of hardware impairments on the
massive MIMO two-way FD relay system has only recently
been studied in [10], which focused on the decode-and-forward
(DF) scheme at the relay. However, the signal processing
complexity of the DF scheme is much higher than that of the
amplify-and-forward (AF) scheme for the implementation of
massive MIMO relay systems [11]. Therefore, the AF scheme
is more attractive for practical system design. To the best of
authors’ knowledge, the performance of AF based multipair
massive MIMO two-way FD relay systems with hardware
impairments has not been investigated in the literature, partially due to the difficulty of manipulating products of SI and
hardware impairments vectors.
Motivated by the aforementioned consideration, a natural
question is that, whether the low-cost non-iedeal hardware
can be deployed at the AF based multipair massive MIMO
two-way FD relay system without sacrificing the expected
performance gains? In this paper, we try to answer this
question with the following contributions:
• An analytical SE approximation of multipair massive
MIMO two-way FD relay systems with hardware impairments is derived in closed-form. The effect of both
the number of relay antennas and the level of transceiver
hardware impairments on the SE has been investigated.
• A hardware scaling law has been presented to show that
one can tolerate larger level of hardware impairments as
the number of antennas increases. This is an analytic
proof that the considered system can be deployed with
low-cost hardware components.
• The sum SEs of FD and HD systems have been compared with different levels of hardware impairments. It
is interesting to find that the FD system with hardware
impairments can achieve the same SE of the HD system
with larger loop interferences. Finally, we derive the
optimal number of relay antennas to maximize the EE.
II. S YSTEM M ODEL
We consider a massive MIMO two-way FD relay system
where K pairs of devices on two sides communicate with
each other through a single relay TR . The devices are denoted
as TAi and TBi , for i = 1, . . . , K, respectively. The devices
could be sensors that exchange a small amount of information
or small cell base stations which need high throughput links.
The relay is equipped with 2N antennas, where N antennas
are used for transmission, and the other N antennas are used
for reception. Each device is equipped with one receive and
one transmit antenna. In addition, the relay and devices are
assumed to work in FD mode, i.e., they can transmit and
receive signals at the same time. We further assume that there
is no direct communication link between each pair of devices
due to heavy shadowing and/or large path loss. The device is
interfered by other devices on the same side.
A. Channel Model
Block fading is considered in this paper. This means that
it is an ergodic process with a static channel realization
in a coherence block and the realizations in blocks are
∆
independent. Then, we define Gu = [gu1 , . . . , guK ] and
∆
Hu = [hu1 , . . . , huK ], where gui ∈ CN ×1 and hui ∈ CN ×1
(i = 1, . . . , K), which denote the uplink channels between
TAi and TR , TBi and TR , respectively. In addition, the
downlink channels between TAi and TR , TBi and TR are given
∆
∆
by Gd = [gd1 , . . . , gdK ] and Hd = [hd1 , . . . , hdK ], where
N ×1
N ×1
gdi ∈ C
and hdi ∈ C
(i = 1, . . . , K), respectively.
Furthermore, Gu , Hu , Hd and Hd are assumed to follow
the independent identically distributed (i.i.d.) Rayleigh fading,
i.e., the elements of guK , huK , gdK , and hdK are i.i.d.
CN (0, σg2ui ), CN (0, σh2 ui ), CN (0, σg2di ), and CN (0, σh2 di )
random variables [6]. Furthermore, Gu , Hu , Gd , and Hd
1/2
1/2
can be expressed as Gu = Sgu Dgu , Hu = Shu Dhu ,
1/2
1/2
Gd = Sgd Dgd , and Hd = Shd Dhd , respectively, where
Sgu , Shu , Sgd , and Shd stand for the small-scale fading and
their elements are all i.i.d. CN (0, 1) random variables. On the
other hand, Dgu , Dhu , Dgd , and Dhd are diagonal matrices
representing the large-scale fading, and the kth diagonal elements are denoted as σg2uk , σh2 uk , σg2dk , and σh2 dk , respectively.
Furthermore, let GRR ∈ CN ×N denote the SI matrix between the transmit and receive arrays of the relay due to the FD
mode. Each row of GRR such as GRRi denotes the channel
between ith receive antenna and all transmit antennas of the
relay. Ωk,i is the inter-device interference channel coefficient
from ith device to kth device. Note that Ωk,k denotes the SI
at the kth device. The elements of GRR and Ωk,i are random
variables following the i.i.d. complex Gaussian distribution,
2
2
), respectively [2].
e.g., CN (0, σLIr
) and CN (0, σk,i
B. Hardware Impairments
As shown in [12], the residual hardware impairments at
the transmitter and receiver can be modeled as additive
distortion noises that are proportional to the signal power.
Thus, the additive distortion term ηr describes the residual
impairments of receiver at the relay and is proportional to the
instantaneous power of received signals at the
relay antenna
as ηr ∼ CN 0, κ2r diag (W11 , . . . , WN N ) , where Wii is
the
matrix W =
Pcovariance
PK ith diagonalH elementH of Pthe
N
H
R
j=1 GRRj GRRj with PU
j=1 PU (huj huj +guj guj )+ N
being the power constraint of the device and PR being the
transmit power of relay [12]. Furthermore, the proportionality
coefficient κr describes the level of hardware impairments and
is related to the received error vector magnitude (EVM) [8].
Note that the EVM is a common quality indicator of the signal
distortion magnitude, and it can be defined as the ratio of the
signal distortion to the signal magnitude. For example, the
EVM at relay can be defined as [8, Eq. (5)]
v n
o s
u
u E kηr k2 |ℑ
tr (κ2r W)
u
o =
= κr ,
(1)
EV Mr = t n
2
tr (W)
E kxk |ℑ
where ℑ denotes the set of channel realizations (i.e., gu , hu ∈
ℑ). Furthermore, 3GPP LTE suggests that the EVM should be
smaller than 0.175 [8].
C. Signal Transmission
At the time instant n, all devices TAi and TBi
(i = 1, . . . , K) transmit their signals xAi (n) and xBi (n) to
the relay TR , respectively, and TR broadcasts its processed
previously received signal yt (n) to all devices.
First, we assume that xAi (n) and xBi (n) are Gaussian
distributed signals. Due to the FD mode, TR also receives the
signal, i.e., yt (n) which is broadcasted to all devices. Thus,
at the time instant n, the received signal at TR is given by
yr (n) = Ax (n) + GRR yt (n) + ηr + nR (n) ,
(2)
T
∆
∆
where A = [Gu , Hu ], x (n) =
xTA (n) , xTB (n)
∆
∆
with xA (n) = [xA1 (n), . . . , xAK (n)] and xB (n) =
2
IN ) denotes
[xB1 (n), . . . , xBK (n)], and nR (n) ∼ CN (0, σR
an additive white Gaussian noise (AWGN) vector at TR .
Then, we analyze the received signal at devices. At the time
instant n (n > 1), the relay using the simple AF protocol amplifies the previously received signal yr (n − 1) and broadcasts
it to the devices. Therefore, the transmit signal vector at the
relay is given by
′
yt (n) = ρFyr (n − 1) ,
(3)
N ×N
where F ∈ C
is the precoding matrix and ρ is the
′
amplification factor. Then TR broadcasts yt (n) to all devices.
However, due to the hardware impairments of RF chains at
the transmitter, TR actually broadcasts yt (n) to all devices as
′
yt (n) = yt (n) + ηt = ρFyr (n − 1) + ηt ,
(4)
where ηt ∼ CN 0, κ2t PNR IN with the proportionality parameters κt characterizing the level of hardware impairment at the
transmitter. Here we assume that each antenna has the same
power and TR can obtain perfect channel state information
(CSI) according to uplink pilots from the devices, and the
devices can then obtain CSI through channel reciprocity [13].
Due to the power constraint of the relay PR , ρ is normalized
by the instantaneous received signal power
s
PR
ρ=
.
2
2
2
2 kFk2
PU kFAk + PNR kFGRR k + kFηr k + σR
At the relay, we adopt the low-complexity maximum-ratio
(MR) scheme suitable for low-cost massive MIMO deployment [13]. Therefore, the precoding matrix F can be written
∆
∆
as F = B∗ AH , where B = [Hd , Gd ], Gd = [gd1 , . . . , gdK ]
∆
and Hd = [hd1 , . . . , hdK ]. To our best knowledge, it is very
challenging to analyze the residual loop interference power if
substituting (4) into (2) iteratively. However, the residual loop
interference can be modeled as additional Gaussian noise. This
is due to the fact that the loop interference can be significantly
degraded and the residual loop interference is too weak by
applying loop interference mitigation schemes [4].
Following similar steps in [4], yr (n) in (2) can be
approximated
by a Gaussian noise source ỹr (n) with
E ỹr (n) ỹrH (n) = PNR IN . Furthermore, TAi and TBi receive the combined signal as
X
T
Ωi,k xAk (n) + nAi (n) , (5)
yt (n) +
ZAi (n) = gdi
i,k∈UA
ZBi (n) = hTdi yt (n) +
X
Ωi,k xBk (n) + nBi (n) , (6)
i,k∈UB
where the noise nAi (n) and nBi (n) are AWGN with
2
2
nAi (n) ∼ CN (0, σA
) and nBi (n) ∼ CN (0, σB
), respeci
i
tively. In the following, we only discuss the analytical result
for TAi . The corresponding result of TBi can be obtained by
replacing TAi with TBi . Note that the relay can only receive
signal and the transmission part keeps silent at the first time
slot (n = 1), during which the received signals at the relay
and devices are respectively given by
yr (1) = Ax (1) + ηr + nR (1)
ZAi (1) =
T
gdi
yt
(7)
(1) + nAi (1) , i = 1, . . . , K.
(8)
For simplicity, the time label n is omitted in the following [6].
Substituting (2) and (4) into (5), the combined received signal
ZAi can be expressed as
K
X
T
T
T
Fhui xBi +ρ
ZAi = ρgdi
Fhuj xBj
gdi
Fguj xAj +gdi
{z
}
|
j=1,j6=i
desired signal
|
{z
}
inter-pair interference
T
T
+ ρgdi
Fgui xAi + ρgdi
FGRR ỹr +
{z
} |
{z
}
|
self-interference
Ωi,k xAk
i,k∈UA
loop interference
|
{z
}
inter-device interference by FD mode
T
T
T
+ ρgdi
Fηr + gdi
ηt + ρgdi
FnR + nAi ,
|
{z
} |
{z
}
hardware impairments
X
(9)
compound noise
where we use the set notation of UA = {1, 3, . . . , 2K − 1}
or UB = {2, 4, . . . , 2K} to represent the devices on bothsides
of relay. Note that one set of devices (UA ) can not exchange
information with the other set (UB ) directly. From (9), we
can find that ZAi is composed of seven terms: the signal that
TAi desires to receive, the inter-pair interference due to other
devices’ signal, the SI from the device, the loop interference
from the relay, the inter-device interference caused by other
devices due to FD mode, the distortion noise induced by
hardware impairments at the relay, and the compound noise.
With the power constraint of the relay and perfect CSI,
the FD relay can take advantage of massive antennas and
simple SI cancellation (SIC) schemes to eliminate the SI [4].
Furthermore, the interference and noise power can be obtained
by taking expectation with respect to interference and noise
within one coherence block of channel fading. As a result, the
SE of TAi is given by
RAi = E {log2 (1 + SINRAi )} ,
for i = 1, . . . , K,
(10)
where SINRAi denotes the signal-to-interference plus noise
ratio (SINR) of Ai and can be expressed as
SINRAi =
T
PU gdi
Fhui
2
1
T
T Fη
A′ +B ′ +C ′ +D′ +E ′ + gdi
r + ρ2 gdi ηt
∆
where A′ = PU
2
giT F
σR
P
PU
ρ2
i,k∈UA
2
2
∆
K
P
j=1,j6=i
σ2
giT Fgj
2
2
+ giT Fhj
2
∆
2,
∆
, B′ =
∆
′
, C ′ = ρAi
= giT FGRR ỹr , and E ′ =
2 , D
2
σi,k
, respectively.
III. P ERFORMANCE A NALYSIS
To the best of authors’ knowledge, the exact derivation of
(10) is really difficult [14]. Herein we consider the asymptotic
scenario when N → ∞, which is the large system limit. Utilizing the convexity of log2 (1 + 1/x) and Jensen’s inequality,
the lower bound of RAi in (10) can be written as
1
o .
(11)
RAi ≥ R̃Ai = log2 1 + n
−1
E [SINRAi ]
Based on (11) and considering devices at both sides, we can
obtain the sum SE of the multipair massive MIMO two-way
FD relay system as
Rsum =
K
X
R̃Ai + R̃Bi .
(12)
i=1
Note that, in the remainder of the paper, we only show
the analytical results for RAi since the formula of RBi is
symmetric with that of RAi . In the following, we present the
SE of Ai in Lemma 1.
Lemma 1. With hardware impairments and MR processing at
the relay, R̃Ai can be approximated as
N
R̃Ai −log2 1+
−−−−→ 0,
Ai +Bi +Ci +Di +Ei +Fi +Gi N →∞
(13)
4
2
4
2
2
K
σh σg
σg σh
σg2
σh
P
∆
uj
uj
uj
uj
dj
dj
,
+
+
+
where Ai =
σ2
σ4 σ2
σ2
σ4 σ2
hui
j=1,j6=i
2
∆
σR
=
Bi
=
, Ci
2
PU σh
ui
!
K
P
2
2
κ2r PU
+PR σLIr
σh
+σg2
hui
j=1
uj
ui
hui
di
uj
2
PU σh
gdi
,
Ei
∆
=
gdi
hui
2
σA
J
i
4
PR PU σg4 σh
∆
,
∆
=
Di
ui
κ2t J
4
PU σg2 σh
di
ui
,
2
∆ P σLIr
Fi = PR
,
2
U σh
Gi =
iu
J
∆
∆
=
P
2
σi,k
i,k∈UA
4
PR σg4 σh
di
ui
J
PU
have the identical scaling, we can finish the proof by fulfilling
1 − max(z, 2z − 1) ≥ 0 as 0 < z ≤ 1.
, and
K
P
σg4ui σh2 di +σg2di σh4 ui
+
Corollary 1 reveals that large level of hardware impair! ments can be compensated by increasing number of antennas
K
K
P
κ2r P
2
.at the relay in multipair massive MIMO two-way relaying sysσg2uj +σh2 uj +PR σLIr
PU
σg2ui σh2 di +σg2di σh2 ui
N
j=1
i=1
tems. Furthermore, the EVM at relay is defined as EVM = κ
[12]. Considering the condition of κ2 = κ20 N z in Corollary
Proof. Please refer to Appendix.
1 and z = 1, it is easy to have EVM2 = κ20 N , which
From Lemma 1, it is clear to see that the SE RAi increases means the EVM can be increased proportionally to N 1/2 .
with the number of antennas N . Further insights can be Thus, for the negligible SE loss, we can replace 8 high-quality
gained by investigating the terms Ai , Bi , Ci , Di , Ei , Fi , and antennas with EVM = 0.05 with 128 low-quality antennas
Gi in (13), respectively. First, we focus on the inter-device with EVM = 0.2. This encouraging result enable reducing the
interference term Ai caused by the broadcasting signal from power consumption and cost of the multipair massive MIMO
the relay. The SE RAi increases when we enlarge the values two-way FD relay system.
of σg2di and σh2 ui , which indicates that reducing the channel
In the following, we evaluate the EE of the multipair
fading of the ith device pair. However, RAi will decrease if we massive MIMO two-way FD relay system when the number of
enlarge σg2uj and σh2 uj , for j 6= i, which means that reducing relay antennas becomes large. The EE is defined as the ratio of
the channel fading of other device pairs except the ith device the sum SE to the total power consumption of the system [13].
pair. This finding is consistent with the result in [14].
Considering the classical architecture where each antenna is
Furthermore, Lemma 1 reveals that Bi consists of the connected to one RF chain. The total power consumption of
transmit power of TBi and the channel fading σh2 ui from TBi the system can be modeled as [4]
to TR . Therefore, we can increase the transmit power of TBi
Ptotal = (N + 2K) (Pt + Pr ) + P0 + (2KPU + PR ) /ϕ,
and/or decrease σh2 ui to increase RAi . Then, from Ci , we can
(15)
find that RAi increases when the transmit power for TR and
the transmit power of devices increase, but decreases when ρ where Pt and Pr are the power of RF chains at the transmitter
becomes large. Moreover, it is clear to see from (13) that the and receiver, respectively. Moreover, P0 denotes the power of
detrimental effect of hardware impairments in Di , Ei and Gi the static circuits, the term 2KPU + PR is the total power of
on the SE RAi . Finally, the loop interference due to the FD the power amplifiers at devices and the relay, and ϕ denotes
mode in Fi and Gi can also reduce the SE.
the efficiency of the power amplifier in each RF chain. Thus,
In order to show how fast the hardware impairments can the EE of the considered system is given by EE = Rsum /Ptotal .
increase with N while maintaining the constant rate, we
establish an important hardware scaling law in the following
IV. N UMERICAL R ESULTS
corollary.
In this section, the derived results of the multipair massive
Corollary 1. Suppose the hardware impairment parameters MIMO two-way FD relay system with hardware impairments
are replaced by κ2r = κ20r N z and κ2t = κ20t N z for an initial and AF schemes are validated through Monte-Carlo simulavalue κ0r ≥ 0, κ0t ≥ 0 and a given scaling exponent 0 < z ≤ tions by averaging over 104 independent channel samples.
1, the SE R̃Ai , under MR processing and N → ∞, converges Similar to previous works [4], [14], we set Pu = 10 W,
2
2
2
PR = 40 W and normalize σR
= σA
= σB
= 1 for i =
to a non-zero limit
i
i
1,
.
.
.
,
K.
Furthermore,
without
loss
of
generality,
we simply
2
4
1−z
σ
σ
N
R̃Ai − log2 1 + κ2 σ2gdiσ2huiξ+2Kκ2 µ̃ −−−−→ 0, 0 < z < 1 set the same values for both loop and inter-device interferences
0r gdi hui
0t
N →∞
2
2
as σ 2 = σLIr
= σk,i
= 1 (i ∈ UA ∪ UB , k = 1, . . . , 2K) and
4
σg2 σh
ui
di
R̃
−
log
1
+
κ0 = κ0r = κ0t , respectively [2], [6].
2
Ai
2
2
2
2
2
κ0r 2Kκ0t µ̄+σg σh
ξ+2Kκ0t µ̃
di
ui
The simulated and analytical asymptotic sum SE (13) are
−−−−→ 0, z = 1
plotted as a function of the half number of antennas N at the
N →∞
(14) relay in Fig. 1. The simulation results validate the tightness
of the derived large-scale approximations. Moreover, Fig. 1
∆
2
where
ξ
=
2Kµ
+
PR σLIr
/PU , validates the hardware scaling law established by Corollary 1.
P
∆
∆
K
2
2
µ
=
µ̃
= The SE grows with low levels of hardware impairments (z =
j=1 σhuj + σguj / (2K),
PK
∆
2
4
2
4
and
µ̄
= 0.5, 1). However, the SE curve asymptotically bend toward
i=1 σgui σhdi + σgdi σhui / (2K)
PK
zero when the scaling law is not satisfied (z = 1.5).
2
2
2
2
i=1 σgui σhdi + σgdi σhui / (2K).
Note that the analytical curves plotted in Fig. 1 are not
Proof. Substituting κ2r = κ20r N z and κ2t = κ20t N z into (13), always below the simulated curves. This is due to the reason
with N → ∞ and 0 < z ≤ 1, Ai , Bi , Ci , Fi , and Gi tend that we utilize the large number law to derive the SE. When
to zero. Moreover, Di behaves as O(N z ), while Ei behaves N is relatively small (e.g. N < 350), the low order term
as O(N z + N 2z−1 ). To make the numerator and denominator of N cannot be omitted. Thus, the analytical result is a little
i=1
70
0.04
0.038
60
50
0.034
EE (bps/J/Hz)
Sum SE (bps/Hz)
0.036
40
30
Large-scale Approximation (z=0.5)
Large-scale Approximation (z=1)
Large-scale Approximation (z=1.5)
Monte-Carlo Simulation
200
300
400
500
600
700
800
900
1000
1/2 Number of Antennas (N)
100
Full-Duplex Approximation(κ0=0.1)
Half-Duplex Approximation(κ0=0.1)
Full-Duplex Approximation(κ0=0)
Half-Duplex Approximation(κ0=0)
Asymptotic Limits
90
Sum SE (bps/Hz)
80
70
60
50
40
30
20
10
10 0
0.026
Large-scale Approximation(z=0)
Large-scale Approximation(z=0.75)
Large-scale Approximation(z=1)
Large-scale Approximation(z=1.25)
0.024
0.022
100
200
300
400
500
600
700
800
900
1000
1/2 Number of Antennas (N)
Fig. 1. Hardware scaling law of multipair massive MIMO two-way FD relay
systems against different number of antennas N at the relay (K = 10, κ0 =
0.0156).
0
10 -1
0.03
0.028
20
10
100
0.032
10 1
10 2
2
Level of Loop and Inter-device Interference (σ )
Fig. 2. Sum SE of multipair massive MIMO two-way FD relay systems
with hardware impairments against different levels of loop and inter-device
interference (N = 1000, z = 1, K = 10).
larger than the corresponding simulation result. However, even
if N is smaller, the curves of large-scale approximation and
Monte-Carlo simulations are close [14]. Compared with the
DF scheme in [10], the signal processing can achieve a smaller
sum SE. However, the complexity of AF based system at the
relay is much lower than the DF based system.
Fig. 2 shows the large-scale approximation (13) and the
asymptotic SE limit (14) against the levels of the loop and
inter-device interferences σ 2 . The SE of such system in HD
model is also plotted as a baseline for comparison. Since the
HD system utilizes two phases to transmit and receive signal,
the inherent loop and inter-device interference do not exist.
Therefore, the SE of HD systems is constant in Fig. 2. The
Fig. 3. EE of multipair massive MIMO two-way FD relay systems with
hardware impairments against different number of antennas N at the relay
(K = 10, κ0 = 0.0156).
first observation from Fig. 2 is that when N = 1000, the
asymptotic SE limit of FD systems outperforms the one of
HD systems for small and moderate levels of interference, e.g.,
σ 2 < 101.18 for ideal hardware (κ0 = 0) and σ 2 < 100.9 for
non-ideal hardware (κ0 = 0.1). This can be explained that only
the half time required in the FD mode compared with the HD
mode. Interestingly, the SE of the multipair massive MIMO
two-way FD relay system with hardware impairments is larger
than the one with ideal hardware for small and moderate levels
of loop and inter-device interference. However, large value of
loop and inter-device interference decreases the sum SE of FD
systems. Moreover, the gap of SE curves between FD and HD
systems increases with the level hardware impairments κ0 .
The large-scale approximation of EE as a function of the
number of antennas N at the relay is plotted in Fig. 3. Similar
to [15], we set Pt = 1 W, Pr = 0.3 W, P0 = 2 W and
ϕ = 0.35. It is clear to see that the EE decreases with z
due to the distortion noise caused by hardware impairments.
Moreover, there exists an optimal number of antennas Nopt
to reach the corresponding maximum EE. When N ≤ Nopt ,
the EE can be improved by increasing N . However, when
N > Nopt , increasing N will reduce the EE since the addition
power consumption of RF chains and static circuits dominate
the performance.
V. C ONCLUSIONS
In this paper, we investigate the SE and EE of AF-based
multipair massive MIMO two-way FD relay systems with
hardware impairments. The effect of N and κ on the SE
has been investigated by deriving a closed-form large-scale
approximate expression. In addition, the optimal number of
relay antennas has been derived to maximize the EE. We also
find that the SE of the massive MIMO two-way FD system
with hardware impairments outperforms that of the HD system
when the level of loop and inter-device interference is small
and moderate. Finally, an useful hardware scaling law has been
established to prove that low-cost hardware can be deployed
at the relay due to the huge degrees-of-freedom brought by
massive antennas.
A PPENDIX
o
n
−1
as
From (11), we can rewrite E [SINRAi ]
)
(
2
T
gdi
FGRR
PR
1
=
E
E
2
T Fh
SIN RAi
N PU
gdi
ui
(
(
)
)
2
2
T
T
2
gdi
F
gdi
ηt
σR
1
+
E
+ 2 E
2
2
T Fh
T Fh
PU
ρ PU
gdi
gdi
ui
ui
P
2
(
(
)
)
σi,k
2
σA
1
1
i,k∈UA
i
+
E
+ 2 E
2
2
T Fh
T Fh
ρ2
ρ PU
gdi
gdi
ui
ui
)
(
)!
(
2
2
K
T
T
X
gdi
Fhuj
gdi
Fguj
+E
+
E
2
2
T Fh
T Fh
gdi
gdi
ui
ui
j=1,j6=i
(
)
2
T
gdi
Fηr
1
E
.
(16)
+
2
PU
gT Fhui
Moreover, we have the following approximations
K
o
n
X
2
σg4ui σh2 di + σg2di σh4 ui −−−−→ 0,
E kFAk − N 3
N →∞
i=1
o
n
2
2
E kFGRR k − N 3 σLIr
K
X
i=1
σg2ui σh2 di + σg2di σh2 ui −−−−→ 0,
N →∞
K
n
o
X
2
E kFηr k −κ2r N 2 ξ
σg2ui σh2 di +σg2di σh2 ui /PU −−−−→ 0,
n
o
2
2
E kFk − σR
N2
i=1
K
X
i=1
N →∞
σg2ui σh2 di + σg2di σh2 ui −−−−→ 0,
N →∞
PR
E ρ2 − 3 −−−−→ 0.
(19)
N J N →∞
By substituting (18)-(19) into (16), we can complete the proof
after some simplifications.
R EFERENCES
[1] V. W. Wong, R. Schober, D. W. K. Ng, and L.-C. Wang, Key Technologies for 5G Wireless Systems. Cambridge university press, 2017.
[2] H. Q. Ngo, H. A. Suraweera, M. Matthaiou, and E. G. Larsson, “Multipair full-duplex relaying with massive arrays and linear processing,”
IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1721–1737, Sep. 2014.
[3] X. Xie, M. Peng, Y. Li, W. Wang, and H. V. Poor, “Channel estimation
for two-way relay networks in the presence of synchronization errors,”
di
IEEE Trans. Signal Process., vol. 62, no. 23, pp. 6235–6248, Dec. 2014.
According to the law of large numbers, we have
[4] Z. Zhang, Z. Chen, M. Shen, and B. Xia, “Spectral and energy efficiency
of multi-pair two-way full-duplex relay systems with massive MIMO,”
1 ∗ 2
1 T
2
IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 848–863, May 2016.
gdi Fhui − kgdi
k khui k −−−−→ 0,
N →∞
[5] J. Zhang, L. Dai, Z. He, S. Jin and X. Li, “Performance analysis of
N
N
mixed-ADC massive MIMO systems over rician fading channels,” IEEE
1 T
1 ∗ 2 H
T ∗
kgdi k hui huj +khuj k2 gdi
gdj −−−−→ 0,
gdi Fhuj −
J. Sel. Areas Commun., vol. 35, no. 6, pp. 1327–1338, Jun. 2017.
N →∞
N
N
[6] J. Zhang, L. Dai, S. Sun, and Z. Wang, “On the spectral efficiency of
1 T
1 ∗ 2 H
2 T ∗
massive MIMO systems with low-resolution ADCs,” IEEE Commun.
kgdi k hui guj +kguj k gdi hdj −−−−→ 0,
gdi Fguj −
Lett., vol. 20, no. 5, pp. 842–845, Feb. 2016.
N →∞
N
N
[7] J. Zhang, L. Dai, X. Zhang, E. Björnson, and Z. Wang, “Achievable
1 T
1
2
2
∗ 4
rate of Rician large-scale MIMO channels with transceiver hardware
k khdi k −−−−→ 0.
g F − kgui
N →∞
N di
N
impairments,” IEEE Trans. Veh. Technol., vol. 65, no. 10, pp. 8800–
8806, Oct. 2016.
When N → ∞, we further have
[8] E. Björnson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive
)
!
(
2 T ∗
2
4
2
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
H
1 σhuj σhuj σgdj
hui huj khuj k gdi gdj
and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112–
−
+
−
−
−
−
→
0,
+
E
2
2
∗ k2
7139, Nov. 2014.
N σh2 ui σh4 ui σg2di N →∞
khui k
khui k kgdi
[9] J. Zhang, L. Dai, M. Matthaiou, C. Masouros, and S. Jin, “On the
)
(
!
2 T ∗
2
4
2
spectral efficiency of space-constrained massive MIMO with linear
σ
σ
σ
kg
k
g
h
1
hH
g
uj
g
guj
di dj
uj h
ui uj
E
+ 4 2dj −−−−→ 0, receivers,” in Proc. IEEE ICC, 2016, pp. 1–6.
2+
2
2 −N
2
∗
N →∞ [10] X. Xia, D. Zhang, K. Xu, W. Ma, and Y. Xu, “Hardware impairments
σhui σhui σgdi
khui k
kg k khui k
(
) di
aware transceiver for full-duplex massive MIMO relaying,” IEEE Trans.
1
1
Signal Process., vol. 63, no. 24, pp. 6565–6580, Dec. 2015.
−−−−→ 0,
E
−
[11] X. Chen, L. Lei, H. Zhang, and C. Yuen, “Large-scale MIMO relaying
2
N σh2 ui N →∞
khui k
techniques for physical layer security: AF or DF?” IEEE Trans. Wireless
(
)
2
Commun., vol. 14, no. 9, pp. 5135–5146, Sep. 2015.
T
2
gdi Fηr
κr ξ
[12]
E.
Björnson, M. Matthaiou, and M. Debbah, “Massive MIMO with
E
−−−−→ 0.
(17)
2 − N P σ2
arbitrary non-ideal arrays: Hardware scaling laws and circuit-aware
T Fh
U hui N →∞
gdi
ui
design,” IEEE Trans. Wireless Commun., vol. 14, no. 8, pp. 4353–4368,
Aug. 2015.
)
(
2
[13] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effiT
2
gdi ηt
κ PR
ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
− 4 t2 4 −−−−→ 0,
E
2
vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
T
N
σgdi σhui N →∞
gdi Fhui
[14] S. Jin, X. Liang, K.-K. Wong, X. Gao, and Q. Zhu, “Ergodic rate analysis
(
)
2
T
2
for multipair massive MIMO two-way relay networks,” IEEE Trans.
gdi
FGRR
σLIr
Wireless Commun., vol. 14, no. 3, pp. 1480–1491, Mar. 2015.
E
−−−−→ 0,
−
2
2
T Fh
σhiu N →∞
[15] C. Li, J. Zhang, and K. B. Letaief, “Throughput and energy efficiency
gdi
ui
)
(
analysis of small cell networks with multi-antenna base stations,” IEEE
Trans. Wireless Commun., vol. 13, no. 5, pp. 2505–2517, May 2013.
1
1
E
T Fh
gdi
ui
2
−
N 4 σg4di σh4 ui
−−−−→ 0.
N →∞
(18)
| 7cs.IT
|
A New Functional-Logic Compiler for Curry: Sprite ⋆
Sergio Antoy and Andy Jost
arXiv:1608.04016v1 [cs.PL] 13 Aug 2016
Computer Science Dept., Portland State University, Oregon, U.S.A.
[email protected]
[email protected]
Abstract. We introduce a new native code compiler for Curry codenamed
Sprite. Sprite is based on the Fair Scheme, a compilation strategy that provides
instructions for transforming declarative, non-deterministic programs of a certain
class into imperative, deterministic code. We outline salient features of Sprite,
discuss its implementation of Curry programs, and present benchmarking results. Sprite is the first-to-date operationally complete implementation of Curry.
Preliminary results show that ensuring this property does not incur a significant
penalty.
Keywords: Functional logic programming, Compiler implementation, Operational completeness
1 Introduction
The functional-logic language Curry [16,18] is a syntactically small extension of the
popular functional language Haskell. Its seamless combination of functional and logic
programming concepts gives rise to hybrid features that encourage expressive, abstract,
and declarative programs [5,18].
One example of such a feature is a functional pattern [3], in which functions are
invoked in the left-hand sides of rules. This is an intuitive way to construct patterns
with syntactically-sugared high-level features that puts patterns on a more even footing
with expressions. In Curry, patterns can be composed and refactored like other code,
and encapsulation can be used to hide details. We illustrate this with function get,
defined below, which finds the values associated with a key in a list of key-value pairs.
with x = _ ++ [x] ++ _
get key (with (key, value)) = value
(1)
Operation with generates all lists containing x. The anonymous variables, indicated
by “_ ”, are place holders for expressions that are not used. Function “++” is the listappending operator. When used in a left-hand side, as in the rule for get, operation
with produces a pattern that matches any list containing x. Thus, the second argument
to get is a list — any list — containing the pair (key, value). The repeated variable,
⋆
This material is based upon work partially supported by the National Science Foundation under
Grant No. CCF-1317249.
key, implies a constraint that, in this case, ensures that only values associated with the
given key are selected.
By similar means, we may identify keys:
key_of (with (key, _ )) = key
(2)
This non-deterministically returns a key of the given list; for example:
> key_of [(’a’,0), (’b’,1), (’c’,2)]
’a’
’b’
’c’
(3)
This is just one of many features [5,18] that make Curry an appealing choice, particularly when the desired properties of a program result are easy to describe, but a set of
step-by-step instructions to obtain the result is more difficult to come by.
This paper describes work towards a new Curry compiler we call Sprite. Sprite
aims to be the first operationally complete Curry compiler, meaning it should produce
all values of a source program (within time and space constraints). Our compiler is
based on a compilation strategy named the Fair Scheme [7] that sets out rules for compiling a functional-logic program (in the form of a graph rewriting system) into abstract
deterministic procedures that easily map to the instructions of a low-level programming
language. Section 2 introduces Sprite at a high level, and describes the transformations
it performs. Section 3 describes the implementation of Curry programs as imperative
code. Section 4 contains benchmark results. Section 5 describes other Curry compilers.
Section 6 addresses future work, and Section 7 contains our concluding remarks.
2 The Sprite Curry Compiler
Sprite is a native code compiler for Curry. Like all compilers, Sprite subjects source
programs to a series of transformations. To begin, an external program is used to convert Curry source code into a desugared representation called FlatCurry [17], which
Sprite further transforms into a custom intermediate representation we call ICurry.
Then, following the steps laid out in the Fair Scheme, Sprite converts ICurry into a
graph rewriting system that implements the program. This system is realized in a lowlevel, machine-independent language provided by the open-source compiler infrastructure library LLVM [22]. That code is then optimized and lowered to native assembly,
ultimately producing an executable program. Sprite provides a convenience program,
scc, to coordinate the whole procedure.
2.1 ICurry
ICurry, where the “I” stands for “imperative,” is a form of Curry programs suitable
for translation into imperative code. ICurry is inspired by FlatCurry [17], a popular
representation of Curry programs that has been very successful for a variety of tasks including implementations in Prolog [19]. FlatCurry provides expressions that resemble
those of a functional program — e.g., they may include local declarations in the form of
let blocks and conditionals in the form of case constructs, all possibly nested. Although
the pattern-matching strategy is made explicit through case expressions, FlatCurry is
declarative. ICurry’s purpose is to represent the program in a more convenient imperative form — more convenient since Sprite will ultimately implement it in an imperative
language. In imperative languages, local declarations and conditionals take the form
of statements while expressions are limited to constants and/or calls to subroutines,
possibly nested. ICurry provides statements for local declarations and conditionals. It
provides expressions that avoid constructs that cannot be directly translated into the
expressions of an imperative language.
In ICurry all non-determinism — including the implicit non-determinism in highlevel features, such as functional patterns — is expressed through choices. A choice is
the archetypal non-deterministic function, indicated by the symbol “?” and defined by
the following rules:
x ? _ = x
_ ? y = y
(4)
The use of only choices is made possible, in part, by a duality between choices and free
variables [4,23]: any language feature expressed with choices can be implemented with
free variables and vice versa. Algorithms exists to convert one to the other, meaning we
are free to choose the most convenient representation in Sprite.
Finally, as in FlatCurry, the pattern-matching strategy in ICurry is made explicit and
guided by a definitional tree [1], a structure made up of stepwise case distinctions that
combines all rules of a function. We illustrate this for the zip function, defined as:
zip [] _ = []
zip ( _ : _ ) [] = []
zip (x:xs) (y:ys) = (x,y) : zip xs ys
(5)
The corresponding definitional tree is shown below as it might appear in ICurry.
zip = \a b -> case a of
[]
-> []
(x:xs) -> case b of
[]
-> []
(y:ys) -> (x,y) : zip xs ys
(6)
2.2 Evaluating ICurry
It is understood how to evaluate the right-hand side of (6) efficiently; the Spineless
Tagless G-machine (STG) [28], for instance, is up to the task. But the non-deterministic
properties of functional-logic programs complicate matters. To evaluate zip, its first
argument must be reduced to head-normal form. In a purely functional language, the
root node of a head-normal form is always a data constructor symbol (assuming partial
application is implemented by a data-like object), or else the computation fails. But for
functional-logic programs, two additional possibilities must be considered, leading to
an extended case distinction:
zip = \a b -> case a of
x ? y -> (pull-tab)
- - implied
⊥
-> ⊥
- - implied
[]
-> []
(x:xs) -> case b of ...
(7)
The infrastructure for executing this kind of pattern matching very efficiently by means
of dispatch tables will be described shortly, but for now we note two things. First, there
is no need for ICurry to spell out these extra cases, as they can be generated by the
compiler. Second, their presence calls for an expanded notion of the computation that
allows for additional node states. Because of this, Sprite hosts computations in a graph
whose nodes are taken from four classes: constructors, functions, choices, and failures.
Constructors and functions are provided by the source program; choices are built-in;
and failures, denoted “⊥”, arise from incompletely defined operations such as head,
the function that returns the head of a list. For example, head [] rewrites to “⊥”. A
simple replacement therefore propagates failure from needed arguments to roots.
Choices execute a special step called a pull-tab [2,9]. Pull-tab steps lift non-determinism out of needed positions, where they prevent completion of pattern matches.
The result is a choice between two more-definite expressions. A pull-tab step is shown
below:
zip (a ? b) c → zip a x ? zip b x where x = c
(8)
A pattern match cannot proceed while (a ? b) is the first argument to zip because
there is no matching rule in the function definition (one cannot exist because the choice
symbol is disallowed on left-hand sides). We do not want to choose between a and b
because such a choice would have to be reconsidered to avoid losing potential results.
The pull-tab transformation “pulls” the choice to an outermore position, producing two
new subexpressions, zip a c and zip b c, that can be evaluated further. The fact
that c is shared in the result illustrates a desirable property: that node duplication is
minimal and localized. Pull-tabbing involves some technicalities that we address later.
The complete details are in [2].
Due to the extra cases, additional node types, and, especially, the unusual mechanics of pull-tabbing steps, we chose to develop in Sprite a new evaluation machine from
scratch rather than augment an existing one such as STG. The property of pull-tabbing
that it “breaks-out” of recursively-descending evaluation into nested expressions fundamentally changes the computation so that existing functional strategies are difficult to
apply. In Sprite, we have implemented de novo an evaluation mechanism and runtime
system based on the Fair Scheme. These are the topic of the next section.
3 Implementation
In this section, we describe the implementation of Curry programs in imperative code.
Sprite generates LLVM code, but we assume most readers are not familiar with that.
So, rather than presenting the generated code, we describe the implemented programs
in terms of familiar concepts that appear directly in LLVM. In this way, the reader
can think in terms of an unspecified target language — one similar to assembly —
that implements those concepts. To facilitate the following description, we indicate in
parentheses where a similar feature exists in the C programming language.
In the target language, values are strongly typed, and the types include integers,
pointers, arrays, structures and functions. Programs are arranged into compilation units
called modules that contain symbols. Symbols are visible to other modules, and to control access to them each one is marked internal (static) or external (extern). Control
flow within functions is carried out by branch instructions. These include unconditional
branches (goto), conditional branches (if, for, while) and indirect branches (goto*).
The target of every branch instruction is a function-local address (label). A call stack is
provided, and it is manipulated by call and return instructions that enter and exit functions, respectively. Calls are normally executed in a fresh stack frame, but the target
language also supports explicit tail recursion, and Sprite puts it to good use.
3.1 Expression Representation
The expressions evaluated by a program are graphs consisting of labeled nodes having
zero or more successors. Each node belongs to one of four classes, as discussed in the
previous section. For constructors and functions, node labels are equivalent to symbols
defined in the source program. Failures and choices are labeled with reserved symbols.
Successors are references to other nodes. The number of successors, which equals the
arity of the corresponding symbol, is fixed at compile time. Partial applications are
written in eval/apply form [26].
Sprite implements graph nodes as heap objects. The layout of a heap object is
shown in Fig. 1. The label is implemented as a pointer to a static info table that will be
described later. Sprite emits exactly one table for each symbol in the Curry program.
Successors are implemented as pointers to other heap objects.
3.2 Evaluation
Evaluation in Sprite is the repeated execution of rewriting and pull-tabbing steps. Both
are implemented by two interleaved activities: replacement and pattern-matching. A
replacement produces a new graph from a previous one by replacing a subexpression
matching the left-hand side of a rule with the corresponding right-hand side. For instance, 1 + 1 might be replaced with 2. A replacement is implemented by overwriting
the heap object at the root of the subexpression being replaced. The key advantage of
this destructive update is that no pointer redirection [12, Def. 8][15] is required during a
rewrite step. Reusing a heap object also has the advantage of saving one memory allocation and deallocation per replacement, but requires that every heap object be capable of
storing any node, whatever its arity. Sprite meets this requirement by providing in heap
Heap Object
Info Pointer
Payload
Info Table
Step Function
.
.
.
Fig. 1: The heap object layout.
objects a fixed amount of space capable of holding a small number of successors. For
nodes with more successors than would fit in this space, the payload instead contains
a pointer to a larger array. This approach simplifies memory management for heap objects: since they are all the same size, a single memory pool suffices. Because arities are
known at compile time, no runtime checks are needed to determine whether successor
pointers reside in the heap object.
Pattern-matching consists of cascading case distinctions over the root symbol of the
expression being matched that culminate either in a replacement or in the patter match
of a subexpression. The Fair Scheme implements this according to a strategy guided by
the definitional trees encoded in ICurry. Case distinction as exemplified in (7) assumes
that an expression being matched is not rooted by a function symbol. Thus, when a node
needed to complete a match is labeled by a function symbol, the expression rooted by
that node is evaluated until it is labeled by a non-function symbol. A function-labeled
node, n, is evaluated by a target function called the step function that performs a pattern
match and replacement at n. Each Curry function gives rise to one target function, a
pointer to which is stored in the associated info table (see Fig. 1).
Operationally, pattern-matching amounts to evaluating nested case expressions similar to the one shown in (7). Sprite implements this through a mechanism we call tagged
dispatch. With this approach, the compiler assigns each symbol a tag at compile time.
Tags are sequential integers indicating which of the four classes discussed earlier the
node belongs to. The three lowest tags are reserved for functions, choices, and failures (all functions have the same tag). For constructors, the tag additionally indicates
which constructor of its type the symbol represents. To see how this works, consider the
following type definition:
data ABC = A | B | C
(9)
ABC comprises three constructors in a well-defined order (any fixed order would do).
To distinguish between them, Sprite tags these with sequential numbers starting at the
integer that follows the reserved tags. So, the tag of A is one less than the tag of B,
which is one less than the tag of C. These values are unique within the type, but not
throughout the program: the first constructor of each type, for instance, always has the
same tag. Following these rules, it is easy to see that every case discriminator is a node
tagged with one of 3 + N consecutive integers, where N is the number of constructors in
its type. To compile a case expression, Sprite emits a jump table that transfers control
to a code block appropriate for handling the discriminator tag. For example, the block
that handles failure rewrites to failure, and the block that handles choices executes a
pull-tab. This is shown schematically in Fig. 2. It is in general impossible to know at
compile time which constructors may be encountered when the program runs, so the
jump table must be complete. If a functional logic program does not define a branch
for some constructor — i.e., a function is not completely defined — the branch for that
constructor is a rewrite to failure.
To implement tagged dispatch, Sprite creates function-local code blocks as labels,
constructs a static jump table containing their addresses, and executes indirect branch
instructions — based on the discriminator tag — through the table. Figure 3 shows
a fragment of C code that approximates this. Case distinction occurs over a variable
of List type with two constructors, nil and cons. Five labeled code blocks handle
the five tags that may appear at the case discriminator. A static array of label address
implements the jump table. This example assumes the function, choice, failure, nil, and
cons tags take the values zero through four, respectively. The jump table contains one
extra case not depicted in (7). When the discriminator is a function, the step function of
the discriminator root label is applied as many times as necessary until the discriminator
class is no longer function.
discriminator
range 0 .. 4
0
function
1
choice
2
failure
3
nil
cons
4
Fig. 2: Schematic representation of the Sprite tagged dispatching mechanism for a distinction of a List type.
3.3 Completeness and Consistency
Sprite aims to be the first complete Curry compiler. Informally, complete means the
program produces all the intended results of the source program. More precisely, and
especially for infinite computations, an arbitrary value will eventually be produced,
given enough resources. This is a difficult problem because a non-terminating computation for obtaining one result could block progress of some other computation that
would obtain another result. For example, the following program has a result, 1, that
can be obtained in only a couple of steps, but existing Curry compilers fail to produce
it:
loop = loop
main = loop ? (1 ? loop)
(10)
The Fair Scheme defines a complete evaluation strategy. It creates a work queue containing all of the expressions that might produce a result. At all times, the expression at
the head of the queue is active, meaning it is being evaluated. Initially, the work queue
contains only the goal expression. Whenever pull-tabbing places a choice at the root
of an expression, that expression forks. It is removed from the queue, and its two alternatives are added. Whenever an expression produces a value, it is removed from the
queue. To avoid endlessly working on an infinite computation, the program rotates the
active computation to the end of the work queue every so often. In so doing, Sprite
guarantees that no expression is ignored forever, hence no potential result is lost.
A proof of correctness of compiled programs is provided in [7] for the abstract formulation of the compiler, the Fair Scheme. In this domain, correctness is the property
that an executable program produces all and only the values intended by the corresponding source program. A delicate point is raised by pull-tabbing. A pull-tab step may duplicate or clone a choice, as the following example shows. Cloned choices should be
static void* jump_table[5] = {
&function_tag, &choice_tag, &failure_tag, &nil_tag, &cons_tag
};
entry:
function_tag:
choice_tag:
failure_tag:
nil_tag:
cons_tag:
goto* jump_table[discriminator.tag];
call_step_function(discriminator);
goto* jump_table[discriminator.tag];
/*execute a pull tab*/
/*rewrite to failure*/
/*rewrite to []*/
/*process the nested case expression*/
Fig. 3: An illustrative implementation in C of the case expression shown in (7).
This code fragment would appear in the body of the step function for zip. Variable
discriminator refers to the case discriminator. Label entry indicates the entry point
into this case expression.
seen as a single choice. Thus when a computation reduces a choice to its right alternative, it should also reduce any other clone of the same choice to the right alternative,
and likewise for the left alternative. Computations obeying this condition are called
consistent.
xor x x where x = T ? F
→ pull−tab (xor T x) ? (xor F x) where x = T ? F
(11)
In the example above, a pull-tab step applied to the choice in x leads to its duplication.
Now, when evaluating either alternative of the topmost choice, a consistent strategy
must recognize that the remaining choice (in x) is already made. For instance, when
evaluating xor T x, the value of x can only be T, the left alternative, because the left
alternative of x has already been selected to obtain xor T x. To keep track of clones,
the Fair Scheme annotates choices with identifiers. Two choice nodes with identical
identifiers represent the same choice. Fresh identifiers are assigned when new choices
arise from a replacement; pull-tab steps copy existing identifiers. Every expression in
the work queue owns a fingerprint, which is a mapping from choice identifiers to values
in the set {left,right,either}. The fingerprint is used to detect and remove inconsistent
computations from the work queue.
It is possible to syntactically pre-compute pull-tab steps: that is, a case statement
such as the one in (7) could implement pull-tabbing by defining an appropriate righthand side rule for the choice branch. In fact, a major competing implementation of
Curry does exactly that [8]. A disadvantage of that approach is that choice identifers
must appear as first-class citizens of the program and be propagated through pull-tab
steps using additional rules not encoded in the source program. We believe it is more efficient to embed choice identifiers in choice nodes as an implementation detail and process pull-tab steps dynamically. Section 4.2 compares these two approaches in greater
detail.
4 Performance
In this section we present a set of benchmark results. These programs were previous
used to compare three implementations of Curry [8]: Mcc, Pakcs, and KiCS2. We shall
use KiCS2 to perform direct comparisons with Sprite1 , since it compares favorably to
the others, and mention the relative performance of the others. KiCS2 compiles Curry
to Haskell and then uses the Glasgow Haskell Compiler (GHC) [13] to produce executables. GHC has been shown to produce very efficient code [21,20,27]. Like Sprite,
KiCS2 uses a pull-tabbing evaluation strategy, but unlike Sprite, it does not form a work
queue; hence, is incomplete when faced with programs such as (10). Instead, it builds
a tree containing all values of the program and executes (lazily and with interleaved
steps) a user-selected search algorithm.
A major highlight of KiCS2 is that purely functional programs compile to “straight”
Haskell, thus incurring no overhead due to the presence of unused logic capabilities.
1
Available at https://github.com/andyjost/Sprite-3
Program
PaliFunPats
LastFunPats
Last
PermSortPeano
PermSort
ExpVarFunPats
Half
Reverse
ReverseUser
ReverseBuiltin
ReverseHO
Primes
ShareNonDet
PrimesBuiltin
PrimesPeano
QueensUser
Queens
TakPeano
Tak
Type
FL
FL
FL
FL
FL
FL
FL
F
F
F
F
F
FL
F
F
F
F
F
F
KiCS2
0.64
1.85
1.90
44.04
42.72
5.92
42.31
0.36
0.34
0.40
0.36
0.29
0.28
0.73
0.41
0.87
0.80
0.84
0.32
Sprite
0.09
0.30
0.31
8.14
8.15
1.29
9.55
0.21
0.21
0.39
0.39
0.32
0.33
1.10
0.66
1.83
1.81
2.08
0.92
∆
-7.1
-6.2
-6.1
-5.4
-5.3
-4.6
-4.4
-1.7
-1.6
-1.0
1.1
1.1
1.2
1.5
1.6
2.1
2.3
2.5
2.9
Fig. 4: Execution times for a set of functional (F) and functional-logic (FL) programs
taken from the KiCS2 benchmark suite. Times are in seconds. The final column (∆) reports the speed-up (negative) or slow-down (positive) factor of Sprite relative to KiCS2.
System configuration: Intel i5-3470 CPU at 3.20GHz, Ubuntu Linux 14.04.
Sprite, too, enjoys this zero-overhead property, but there is little room to improve upon
GHC for functional programs, as it is the beneficiary of exponentially more effort. Our
goal for functional programs, therefore, is simply to measure and minimize the penalty
of running Sprite. For programs that utilize logic features KiCS2 emits Haskell code
that simulates non-determinism. In these cases, there is more room for improvement
since, for example, Sprite can avoid simulation overhead by more directly implementing logic features.
4.1 Functional Programs
The execution times for a set of programs taken from the KiCS2 benchmark suite2 are
shown in Fig. 4. The results are arranged in order from greatest improvement to greatest
degradation in execution time. The most striking feature is the clear division between
the functional (deterministic) and functional-logic (non-deterministic) subsets, which is
consistent with our above-stated expectations. On average, Sprite produces relatively
slower code for functional programs and relatively faster code for functional-logic ones.
We calculate averages as the geometric mean, since that method is not strongly influenced by extreme results in either direction. The functional subset runs, on average,
1.4x slower in Sprite as compared to KiCS2. Figures published by Braßel et al. [8,
2
Downloaded from https://www-ps.informatik.uni-kiel.de/kics2/benchmarks.
Fig.2, Fig.3] indicate that Pakcs and Mcc run 148x and 9x slower than KiCS2, respectively, for these programs. We take these results as an indication that the functional parts
of Sprite — i.e., those parts responsible for pattern-matching, rewriting, memory management, and optimization — although not as finely-tuned as their GHC counterparts,
still compare favorably to most mainstream Curry compilers.
We note that Sprite currently does not perform optimizations such as deforestation
[14] or unboxing [21]. These, and other optimizations of ICurry, e.g., [6], could potentially impact the benchmark results. Inspecting the output of GHC reveals that the
tak program (incidentally, the worse-case for Sprite) is optimized by GHC to a fullyunboxed computation. To see how LLVM stacks up, we rewrote the program in C and
converted it to LLVM using Clang [11], a C language front-end for LLVM. When we
compiled this to native code and measured the execution time, we found that it was
identical3 to the KiCS2 (and GHC) time. We therefore see no fundamental barrier to reducing the Sprite “penalty” to zero for this program, and perhaps others, too. We have
reason to be optimistic that implementing more optimizations at the source and ICurry
levels, without fundamentally changing the core of Sprite, will yield substantive improvements to Sprite.
4.2 Functional-Logic Programs
For the functional-logic subset, Fig. 4 shows that Sprite produces relatively faster code:
4.4x faster, on average. Published comparisons [8, Fig.4] indicate that, compared to
KiCS2, Pakcs is 5.5x slower and Mcc is 3.5x faster for these programs. Our first thought
after seeing this result was that Sprite might enjoy a better algorithmic complexity. We
had just completed work to reduce Sprite’s complexity when processing choices, so
perhaps, we thought, in doing that work we had surpassed KiCS2. We set out to test
this by selecting a program dominated by choice generation and running it for different input sizes, with and without the recent modifications to Sprite. The results are
shown in Fig. 5. Contrary to our expectation, Sprite and KiCS2 exhibit strikingly similar complexity: both fit an exponential curve with r2 in excess of 0.999, and their slope
coefficients differ by less than 2%. A better explanation, then, for the difference is that
some constant factor c exists, such that choice-involved steps in Sprite are c-times faster
than in KiCS2. What could account for this factor? We believe the best explanation is
the overhead of simulating non-determinism in Haskell, which we alluded to at the end
of Sect. 3.3. To see why, we need to look at KiCS2 in more detail.
KiCS2 uses a few helper functions [8, Sect. 3.1] to generate choice identifiers:
thisID
:: IDSupply -> ID
leftSupply :: IDSupply -> IDSupply
rightSupply :: IDSupply -> IDSupply
(12)
The purpose of these functions is to ensure that choice identifiers are never reused.
Here, ID is the type of a choice identifier and IDSupply is opaque (for our purposes).
Any function that might produce a choice is implicitly extended by KiCS2 to accept a
supply function. As an example, this program
3
Using the Linux time command, whose resolution is 0.01 seconds.
Fig. 5: Complexity analysis of PermSort. Execution times are shown for a range of
problem sizes. The horizontal axis indicates the number of integers to sort by the
permute-and-test method.
f :: Bool
main = xor f (False ? True)
(13)
is compiled to
main s = let s1 = leftSupply s
s2 = rightSupply s
s3 = leftSupply s2
s4 = rightSupply s2
in xor (f s3) (Choice (thisID s4) False True) s1
(14)
Clearly, the conversion to Haskell introduces overhead. The point here is simply to see
that the compiled code involves five calls (to helper functions) that were not present in
the source program. These reflect the cost of simulating non-determinism in a purelyfunctional language.
In Sprite, fresh choice identifiers are created by reading and incrementing a static
integer. Compared to the above approach, fewer parameters are passed and fewer functions are called. A similar approach could be used in a Haskell implementation of Curry,
but it would rely on impure features, adding another layer of complexity and perhaps
interfering with optimizations. By contrast, the Sprite approach is extreme in its simplicity, as it executes only a few machine instructions. There is a remote possibility that
a computation could exhaust the supply of identifiers since the type integer is finite.
KiCS2 uses a list structure for choice identifiers and so does not suffer from this potential shortcoming. Certainly, the choice identifiers could be made arbitrarily large, but
doing so increases memory usage and overhead. A better approach, we believe, would
be to compact the set of identifiers during garbage collection. The idea is that whenever
a full collection occurs, Sprite would renumber the n choice identifiers in service at that
time so that they fall into the contiguous range 0, . . . , n − 1. This potential optimization
illustrates the benefits of having total control over the implementation, since in this case
it makes modifying the garbage collector a viable option.
5 Related Work
Several Curry compilers are easily accessible, most notably Pakcs [19], KiCS2 [8] and
Mcc [25]. All these compilers implement a lazy evaluation strategy, based on definitional trees, that executes only needed steps, but differ in the control strategy that selects
the order in which the alternatives of a choice are executed.
Both Pakcs and Mcc use backtracking. They attempt to evaluate all the values of the
left alternative of a choice before turning to the right alternative. Backtracking is simple
and relatively efficient, but incomplete. Hence, a benchmark against these compilers
may be interesting to understand the differences between backtracking and pull-tabbing,
but not to assess the efficiency of Sprite.
By contrast, KiCS2’s control strategy uses pull-tabbing, hence the computations executed by KiCS2 are much closer to those of Sprite. KiCS2’s compiler translates Curry
source code into Haskell source code which is then processed by GHC [13], a mainstream Haskell compiler. The compiled code benefits from a variety of optimizations
available in GHC. Section 4 contains a more detailed comparison between Sprite and
KiCS2.
There exist other functional logic languages, e.g., T OY [10,24], whose operational
semantics can be abstracted by needed narrowing steps of a constructor-based graph
rewriting system. Some of our ideas could be applied with almost no changes to the
implementation of these languages.
A comparison with graph machines for functional languages is problematic at best.
Despite the remarkable syntactic similarities, Curry’s syntax extends Haskell’s with
a single construct, a free variable declaration, the semantic differences are profound.
There are purely functional programs whose execution produces a result as Curry, but
does not terminate as Haskell [5, Sect. 3]. Furthermore, functional logic computations
must be prepared to encounter non-determinism and free variables. Hence, situations
and goals significantly differ.
6 Future Work
Compilers are among the most complex software artifacts. They are often bundled with
extensions and additions such as optimizers, profilers, tracers, debuggers, external libraries for application domains such as databases or graphical user interfaces. Given
this reality, there are countless opportunities for future work. We have no plans at this
time to choose any one of the extensions and additions listed above before any other.
Some optimizations mentioned earlier, e.g., unboxing integers, are appealing only because they would improve some benchmark, and thus the overall perceived performance
of the compiler, but they may contribute very marginally to the efficiency of more realistic programs. Usability-related extensions and additions, such as aids for tracing and
debugging an execution, and external libraries may better contribute to the acceptance
of our work.
7 Conclusion
We have presented Sprite, a new native code compiler for Curry. Sprite combines the
best features of existing Curry compilers. Similar to KiCS2, Sprite’s strategy is based
on pull-tabbing, hence there is no an inherent loss of completeness of compilers based
on backtracking such as Pakcs and Mcc. Similar to Mcc, Sprite compiles to an imperative target language, hence is amenable to low-level machine optimization. Differently
from all existing compilers, Sprite is designed to ensure operational completeness—
all the values of an expression are eventually produced given enough computational
resources.
Sprite’s main intermediate language, ICurry, represents programs as graph rewriting systems. We described the implementation of Curry programs in imperative code
using concepts of a low-level target language. Graph nodes are represented in memory
as heap objects, and an efficient mechanism called tagged dispatch is used to perform
pattern matches. Finally, we discussed the mechanisms used by Sprite to ensure completeness and consistency, and presented empirical data for a set of benchmarking programs. The benchmarks reveal that Sprite is competitive with a leading implementation
of Curry.
References
1. S. Antoy. Definitional trees. In H. Kirchner and G. Levi, editors, Proceedings of the Third
International Conference on Algebraic and Logic Programming, pages 143–157, Volterra,
Italy, September 1992. Springer LNCS 632.
2. S. Antoy. On the correctness of pull-tabbing. TPLP, 11(4-5):713–730, 2011.
3. S. Antoy and M. Hanus. Declarative programming with function patterns. In 15th Int’nl
Symp. on Logic-based Program Synthesis and Transformation (LOPSTR 2005), pages 6–22,
London, UK, September 2005. Springer LNCS 3901.
4. S. Antoy and M. Hanus. Overlapping rules and logic variables in functional logic programs.
In Twenty Second International Conference on Logic Programming, pages 87–101, Seattle,
WA, August 2006. Springer LNCS 4079.
5. S. Antoy and M. Hanus. Functional logic programming. Comm. of the ACM, 53(4):74–85,
April 2010.
6. S. Antoy, J. Johannsen, and S. Libby. Needed computations shortcutting needed steps. In
A. Middeldorp and F. van Raamsdonk, editors, Proceedings 8th International Workshop on
Computing with Terms and Graphs, Vienna, Austria, July 13, 2014, volume 183 of Electronic
Proceedings in Theoretical Computer Science, pages 18–32. Open Publishing Association,
2015.
7. S. Antoy and A. Jost. Compiling a functional logic language: The fair scheme. In 23rd Int’nl
Symp. on Logic-based Program Synthesis and Transformation (LOPSTR 2013), pages 129–
143, Madrid, Spain, Sept. 2013. Dpto. de Systems Informaticos y Computation, Universidad
Complutense de Madrid, TR-11-13.
8. B. Braßel, M. Hanus, B. Peemöller, and F. Reck. KiCS2: A new compiler from Curry to
Haskell. In Proc. of the 20th International Workshop on Functional and (Constraint) Logic
Programming (WFLP 2011), pages 1–18. Springer LNCS 6816, 2011.
9. B. Brassel and F. Huch. On a tighter integration of functional and logic programming. In
APLAS’07: Proceedings of the 5th Asian conference on Programming languages and systems, pages 122–138, Berlin, Heidelberg, 2007. Springer-Verlag.
10. R. Caballero and J. Sánchez, editors. TOY: A Multiparadigm Declarative Language (version
2.3.1), 2007. Available at http://toy.sourceforge.net.
11. clang: a C language family frontend for LLVM, 2016.
Available at
http://www.clang.llvm.org/.
12. R. Echahed and J. C. Janodet.
On constructor-based graph rewriting systems.
Technical Report 985-I, IMAG, 1997.
Available at
ftp://ftp.imag.fr/pub/labo-LEIBNIZ/OLD-archives/PMP/c-graph-rewriting.ps.gz.
13. The Glasgow Haskell Compiler, 2013. Available at http://www.haskell.org/ghc/.
14. A. Gill, J. Launchbury, and S. L Peyton Jones. A short cut to deforestation. In Proceedings
of the conference on Functional programming languages and computer architecture, pages
223–232. ACM, 1993.
15. J. R. W. Glauert, R. Kennaway, G. A. Papadopoulos, and M. R. Sleep. Dactl: an experimental
graph rewriting language. J. Prog. Lang., 5(1):85–108, 1997.
16. M. Hanus, editor. Curry: An Integrated Functional Logic Language (Vers. 0.8.2), 2006.
Available at http://www-ps.informatik.uni-kiel.de/currywiki/.
17. M. Hanus. Flatcurry: An intermediate representation for Curry programs, 2008. Available
at http://www.informatik.uni-kiel.de/~curry/flat/.
18. M. Hanus. Functional logic programming: From theory to Curry. In Programming Logics Essays in Memory of Harald Ganzinger, pages 123–168. Springer LNCS 7797, 2013.
19. M. Hanus, editor. PAKCS 1.11.4: The Portland Aachen Kiel Curry System, 2014. Available
at http://www.informatik.uni-kiel.de/~pakcs.
20. S. L. Peyton Jones. Compiling haskell by program transformation: A report from the
trenches. In Programming Languages and SystemsâĂŤESOP’96, pages 18–44. Springer,
1996.
21. S. Peyton Jones and A. Santos. Compilation by transformation in the glasgow haskell compiler. In Functional Programming, Glasgow 1994, pages 184–204. Springer, 1995.
22. C. Lattner and V. Adve. LLVM: A compilation framework for lifelong program analysis
and transformation. In Proceedings of the international symposium on Code generation and
optimization: feedback-directed and runtime optimization (CGO 04), pages 75–88, San Jose,
CA, USA, Mar 2004.
23. F. J. López-Fraguas and J. de Dios-Castro. Extra variables can be eliminated from functional
logic programs. Electron. Notes Theor. Comput. Sci., 188:3–19, 2007.
24. F. J. López-Fraguas and J. Sánchez-Hernández. TOY: A multiparadigm declarative system.
In Proceedings of the Tenth International Conference on Rewriting Techniques and Applications (RTA’99), pages 244–247. Springer LNCS 1631, 1999.
25. W. Lux, editor.
The Muenster Curry Compiler, 2012.
Available at
http://danae.uni-muenster.de/~lux/curry/.
26. S. Marlow and S. Peyton Jones. Making a fast curry: Push/enter vs. eval/apply for higherorder languages. In Proceedings of the Ninth ACM SIGPLAN International Conference on
Functional Programming, ICFP ’04, pages 4–15, New York, NY, USA, 2004. ACM.
27. W. Partain. The nofib benchmark suite of haskell programs. In Functional Programming,
Glasgow 1992, pages 195–202. Springer, 1993.
28. S. L Peyton Jones and J. Salkild. The spineless tagless g-machine. In Proceedings of the
fourth international conference on Functional programming languages and computer architecture, pages 184–201. ACM, 1989.
| 6cs.PL
|
arXiv:1711.10238v2 [math.GR] 15 Feb 2018
Stability, cohomology vanishing, and
non-approximable groups
Marcus De Chiffre, Lev Glebsky, Alexander Lubotzky, and Andreas Thom
Abstract. Several well-known open questions (such as: are all
groups sofic/hyperlinear?) have a common form: can all groups be
approximated by asymptotic homomorphisms into the symmetric
groups Sym(n) (in the sofic case) or the finite dimensional unitary groups U(n) (in the hyperlinear case)? In the case of U(n),
the question can be asked with respect to different metrics and
norms. This paper answers, for the first time, one of these versions, showing that there exist fintely presented groups which are
not approximated
by U(n) with respect to the Frobenius norm
√ n
∥T ∥Frob = ∑i,j=1 ∣Tij ∣2 , T = [Tij ]ni,j=1 ∈ Mn (C). Our strategy is
to show that some higher dimensional cohomology vanishing phenomena implies stability, that is, every Frobenius-approximate homomorphism into finite-dimensional unitary groups is close to an
actual homomorphism. This is combined with existence results of
certain non-residually finite central extensions of lattices in some
simple p-adic Lie groups. These groups act on high rank BruhatTits buildings and satisfy the needed vanishing cohomology phenomenon and are thus stable and not Frobenius-approximated.
1. Introduction
Since the very beginning of the study of groups, groups were studied
by looking at their orthogonal and unitary representations. It is very
natural to relax the notion of a representation and require the group
multiplication to be preserved only up to little mistakes in a suitable
metric. First variations of this topic appeared already in the work of
Turing [49] and later Ulam [50, Chapter VI]. This theme knows many
variations, ranging from sofic approximations as introduced by Gromov
[26] and operator-norm approximations that appeared in the theory of
operator algebras [9, 14] to questions related to Connes’ Embedding
Problem, see [15,36] for details. In each case, approximation properties
of groups are studied relative to a particular class of metric groups.
More specifically, let Γ be a countable group and let (Gn , dn )∞
n=1 be a
sequence of metric groups with bi-invariant metrics dn . We say that
Γ is (Gn , dn )∞
n=1 -approximated, if there exists a separating sequence of
1
2
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
asymptotic homomorphisms ϕn ∶ Γ → Gn , i.e. a sequence of maps ϕn
that becomes multiplicative in the sense that
lim dn (ϕn (gh), ϕn (g)ϕn (h)) = 0,
n→∞
for all g, h ∈ Γ,
which is also separating, that is, dn (ϕn (g), 1Gn ) is bounded away from
zero for all g ≠ 1Γ , see Section 1.4 for precise definitions. Several examples of this situation have been studied in the literature (see [2] for
a survey):
(i) Gn = Sym(n), the symmetric group on an n-point set, with dn
the normalized Hamming distance. In this case, (Gn , dn )∞
n=1 approximated groups are called sofic, see [26, 36].
(ii) Gn an arbitrary finite group equipped with any bi-invariant
metric. In this case, approximated groups are called weakly
sofic, or C-approximated depending on a particular restricted
family C of finite groups. An interesting connection to profinite group theory and recent advances can be found in [25,
32].
(iii) Gn = U(n), the unitary group on an n-dimensional Hilbert
space, where the metric √
dn is induced by the normalized HilbertSchmidt norm ∥T ∥HS = n−1 ∑ni,j=1 ∣Tij ∣2 . In this case, approximated groups are sometimes called hyperlinear [36].
(iv) Gn = U(n), where the metric dn is induced by the operator
norm ∥T ∥op = sup∥v∥=1 ∥T v∥. In this case, groups which are
(Gn , dn )∞
n=1 -approximated groups are called MF, see [14].
(v) Gn = U(n), where the metric dn is induced
√ n by the unnormalized Hilbert-Schmidt norm ∥T ∥Frob = ∑i,j=1 ∣Tij ∣2 , also called
Frobenius norm. We will speak about Frobenius-approximated
groups in this context.
Note that the approximation properties are local in the sense that
only finitely many group elements and their relations have to be considered for fixed ϕn . This is in stark contrast to the uniform situation,
which – starting with the work of of Grove-Karcher-Ruh and Kazhdan
[27, 29] – is much better understood, see [13, 16].
Well-known and longstanding problems, albeit in different fields of
mathematics, ask if any group exists which is not approximated in
either of the above settings. In setting (i), this is Gromov’s question
whether all groups are sofic [26,36]. The similar question in the context
of (iii) is closely related to Connes’ Embedding Problem [15, 36]. Indeed, the existence of a non-hyperlinear group, whould answer Connes’
Embedding Problem in the negative. In [9], Kirchberg asked whether
any stably finite C ∗ -algebra is embeddable into an norm-ultraproduct
of matrix algebras, implying a positive answer to the approximation
ASYMPTOTIC REPRESENTATIONS
3
problem in the sense of (iv) for any group. Recent breakthrough results imply that any amenable group is MF, i.e. approximated in the
sense of (iv), see [45].
In this paper, we want to introduce a conceptually new technique
that allows us to provide groups that are not approximated in the
sense of (v) above, i.e. we show that there are finitely presented groups
which are not approximated by unitary groups {U(n) ∣ n ∈ N} with
their Frobenius norm. Our techniques do not apply directly to the
context of (iii), so we cannot say anything conclusive about Connes’
Embedding Problem, but since the norms in (iii) and (v) are related
by a normalization constant, we believe that we provide a promising
new angle of attack.
Before we start out explaining our strategy and some notation let
us state the main results of this article.
Theorem 1.1. There exist finitely presented groups which are not
Frobenius-approximated.
The groups we construct are central extensions of cocompact lattices in simple p-adic Lie groups. Specifically, we can take certain
central extensions of U(2n) ∩ Sp(2n, Z[i, 1/p]) for n ≥ 3 and p a large
enough prime.
To prove Theorem 1.1, we use the notion of stability: A group is
called (Gn , dn )-stable if every asymptotic homomorphism (not necessarily a separating one) is “close” to a true homomorphism (see Definition 1.9). Now, if Gn = U(n) and Γ is (Gn , dn )-approximated and
(Gn , dn )-stable, one easily deduces that Γ must be residually finite.
This basic observation suggests a way to find non-approximated groups:
find a group Γ which is stable but not residually finite. This method
has failed so far for two reasons: (1) it is very difficult to prove stability
directly and (2) even in the case where stability was proven, (see e.g.
[3] and the references therein as well as [7]) it was proven in a way that
completely classifies asymptotic homomorphism and it is shown that
all are close to a genuine homomorphisms. Thus, only groups which
are already approximated have been shown to be stable so far. The
main technical novelty of our paper is the following theorem which provides a sufficient condition for a group to be Frobenius-approximated
without assuming a priori that the group is approximated.
Theorem 1.2. Let Γ be a finitely presented group such that
H 2 (Γ, Hπ ) = {0}
for every unitary representation π∶ Γ → U(Hπ ). Then, any asymptotic
homomorphism ϕn ∶ Γ → U(n) w.r.t. the Frobenius norm is asymptotically close to a sequence of homomorphisms, i.e. Γ is Frobenius-stable.
The appearance of vanishing second cohomology groups may look
surprising at first sight, but, inf fact, one can translate the question of
4
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
approximating an asymptotic homomorphism by a true homomorphism
to a question about splitting an exact sequence. When the norm is
submultiplicative (as is the case of the Frobenius norm but not of the
normalized Hilbert-Schmidt norm) the kernel of this splitting problem
is abelian (see Section 3.4). It is well-known that vanishing of the
second cohomology with abelian coefficients means splitting of suitable
exact sequences and hence is relevant the question of stability. It is
also interesting to observe that the second cohomology has already
appeared in the work of Kazhdan [29] in the context of uniform εrepresentations (of compact or amenable groups), a concept related to
asymptotic representations, abeit essentially different.
Recall that the classical Kazhdan’s Property (T) is equivalent to
the statement that H 1 (Γ, Hπ ) = 0 for all unitary representations π∶ Γ →
U(Hπ ). We say that a group is n-Kazhdan if H n (Γ, Hπ ) = {0} for
every unitary representation π∶ Γ → U(Hπ ). Theorem 1.2 simply says
that every 2-Kazhdan group is Frobenius stable. Thus to prove Theorem 1.1 it suffices to find 2-Kazhdan groups which are not residually
finite. Now, the seminal work of Garland [23] (as was extended by
Ballmann-Świątkowski [5] and others – see Section 4 for details) shows
that for every 2 ≤ r ∈ N and p large enough, cocompact (arithmetic)
lattices in simple p-adic Lie groups of rank r are n-Kazhdan for every
1 ≤ n < r. In fact, a variant of this has been used to give examples of
groups with Property (T), i.e. 1-Kazhdan groups, which are not linear
(and potentially also not residually finite) by using “exotic” buildings
of rank 2 (see [33]). We want to prove the existence of non-residually
finite 2-Kazhdan groups, but there is a catch: as n = 2, r should be
at least 3, but a well-known result of Tits asserts that for r ≥ 3, there
are no “exotic” buildings of dimension r and the standard ones coming
from p-adic Lie groups provide lattices which are all residually finite.
To work around this point, we imitate a result (and method of proof)
of Deligne [17]. Deligne showed that some non-uniform lattices in simple Lie groups (e.g. Sp(2n, Z)) have finite central extensions which are
not residually finite. Raghunathan [41] extended it also to some cocompact lattices in Spin(2, n). These examples became famous when
Toledo [47] used them to provide examples of fundamental groups of
algebraic varieties which are not residually finite. In the last section,
we will explain how Deligne’s method can be applied also to cocompact
lattices in certain p-adic Lie groups. Along the way we use the solution
to the congruence subgroup problem for these lattices which was provided by Rapinchuk [43] and Tomanov [48]. This way we will get finite
central extensions of certain cocompact p-adic lattices which are themselves not residually finite anymore. Finally, an easy spectral sequence
argument shows that a finite (central) extension of an n-Kazhdan group
is also n-Kazhdan. Thus, the non-residually finite central extensions of
ASYMPTOTIC REPRESENTATIONS
5
the abovementioned lattices provide the non-Frobenius-approximated
group promised in Theorem 1.1.
Along the way in Section 2, we also provide examples of residually
finite groups which are not Frobenius-stable and of finitely generated
non-residually finite groups which are Frobenius-approximated. It is
currently unclear if maybe all amenable (or even all solvable) groups
are Frobenius-approximated. Moreover, it is an open problem to decide
if the class of Frobenius-approximated groups is closed under central
quotients or under crossed products by Z, compare with [34, 46].
The results of this article are part of the PhD project of the first
named author.
1.1. Notation. Given any set S we let FS denote the free group on
S. For any R ⊆ FS we let ⟪R⟫ denote the normal subgroup generated
by R and we let ⟨S ∣ R⟩ ∶= FS /⟪R⟫ be the group with generators S
and relations R. We use the convention N = {1, 2, . . .}. For n ∈ N we
let Mn (C) denote the complex n × n-matrices and U(n) ⊆ Mn (C) the
group of unitary matrices. The identity matrix is denoted by 1n .
Recall, an ultrafilter U on N is a non-trivial collection of subsets
of N, such that (i) A ∈ U , A ⊂ B implies B ∈ U , (ii) A, B ∈ U implies
A ∩ B ∈ U , and (iii) A ∈/ U if and only if N ∖ A ∈ U holds. We say that U
is non-principal if {n} ∈/ U for all n ∈ N. The existence of non-principal
ultrafilters on N is ensured by the Axiom of Choice. We can view a nonprincipal ultrafilter as a finitely additive probability measure defined
on all subsets of N, taking only the values {0, 1} and giving the value
0 to all finite subsets of N.
Throughout the whole paper, we fix a non-principal ultrafilter U on
N. Given some statement P (n) for n ∈ N, we use the wording P (n)
holds for most n ∈ N as {n ∈ N ∣ P (n)} ∈ U . Given a bounded sequence
(xn )n∈N of real numbers we denote the limit along the ultrafilter by
limn→U xn ∈ (−∞, ∞). Formally, the limit is the unique real number x
such for all ε > 0 we have {n ∈ N ∣ ∣xn − x∣ < ε} ∈ U . For unbounded
sequences, the limit takes a well-defined value in the extended real line
[−∞, ∞].
We adopt the Landau notation; given two sequences (xn )n∈N and
(yn )n∈N of non-negative real numbers, we write xn = OU (yn ) if there
exists C > 0 such that xn ≤ Cyn for most n ∈ N and xn = oU (yn ) if there
exists a third sequence (εn )n∈N of non-negative real numbers such that
limn→U εn = 0 and xn = εn yn .
1.2. Unitarily invariant norms. Recall that a norm ∥⋅∥ on Md (C)
is called unitarily invariant if
∥UAV ∥ = ∥A∥,
for all A ∈ Md (C) and U, V ∈ U(d).
6
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
Important examples of such norms are the operator norm
∥op =
√ ∥T
n
∗
1/2
sup∥v∥=1 ∥T v∥, the Frobenius norm ∥T ∥Frob = Tr(T T ) = ∑i,j ∣Tij ∣2
(also known as the unnormalized Hilbert-Schmidt norm) and the normalized Hilbert-Schmidt norm (or 2-norm) given by ∥T ∥HS = √1n ∥T ∥Frob
for T ∈ Mn (C). Here, T ∗ denotes the adjoint matrix and T is called selfadjoint if T = T ∗ . The matrix T is called unitary if T T ∗ = T ∗ T = 1n .
We recall some basic and well-known facts about unitarily invariant
norms. For T ∈ Mn (C), we set ∣T ∣ = (T ∗ T )1/2 . For self-adjoint matrices
A, B, we write A ≤ B if B − A is positive semi-definite, i.e. if B − A has
only non-negative eigenvalues.
Proposition 1.3. Let A, B, C ∈ Md (C). Then, for any unitarily
invariant norm, it holds that
(1) ∥ABC∥ ≤ ∥A∥op ∥B∥∥C∥op ,
(2) ∥A∥ = ∥A∗ ∥ = ∥∣A∣∥,
(3) If A and B are positive semi-definite matrices and A ≤ B, then
∥A∥ ≤ ∥B∥.
Proposition 1.4. Let A ∈ U(d). Then there is a unitary B ∈ U(d)
such that B 2 = 1 and
∥B − A∥≤ ∥1d − A2 ∥,
for all unitarily invariant norms.
Proof. By unitary invariance, we may assume that A is a diagonal
matrix, which we denote A = diag(a1 , . . . , ad ). Let
1,
if Re(aj ) ≥ 0,
bj ∶= {
−1, if Re(aj ) < 0.
One readily sees that ∣bj − aj ∣ ≤ ∣1 − aj ∣∣ − 1 − aj ∣ = ∣1 − a2j ∣. Thus B =
diag(b1 , . . . , bd ) is a self-adjoint unitary and by Proposition 1.3 (3)
∥B − A∥ = ∥∣B − A∣∥ ≤ ∥∣1d − A2 ∣∥ = ∥1d − A2 ∥.
A second property that is important to us is submultiplicativity,
that is, ∥AB∥ ≤ ∥A∥∥B∥ for all A, B ∈ Md (C). This property turns
(Md (C), ∥⋅∥) into a Banach algebra. The operator norm and the Frobenius norm enjoy this property, but the normalized Hilbert-Schmidt
norm does not.
1.3. Ultraproducts. We will need the definition of the ultraproduct of Banach spaces and metric groups, respectively. First, let
(Vn )n∈N be a sequence of Banach spaces. Consider the ℓ∞ -direct product ∏n∈N Vn (i.e. the Banach space of bounded sequences (vn )n∈N with
vn ∈ Vn ) and the closed subspace of nullsequences
I ∶= {(vn )n∈N ∈ ∏ Vn ∣ lim ∥vn ∥Vn = 0} .
n∈N
n→U
ASYMPTOTIC REPRESENTATIONS
7
We define the ultraproduct Banach space by
∏ (Vn , ∥⋅∥Vn ) ∶= ∏ Vn /I.
n∈N
n→U
As the name suggests, the ultraproduct Banach space is itself a Banach
space with the norm induced by ∥(xn )n∈N ∥ = limn→U ∥xn ∥Vn for (xn )n∈N ∈
∏n∈N Vn . Moreover, if the Vn are all Banach algebras, C ∗ -algebras or
Hilbert spaces, so is the ultraproduct.
Let (Gn )n∈N be a family of groups, all equipped with bi-invariant
metrics dn . In this case, the subgroup
N = {(gn )n∈N ∈ ∏ Gn ∣ lim dn (gn , 1Gn ) = 0}
n→U
n∈N
of the direct product ∏n∈N Gn is normal, so we can define the metric
ultraproduct
∏ (Gn , dn ) ∶= ∏ Gn /N.
n→U
n∈N
Note that, in contrast to the Banach space definition we do not require
the sequences to be bounded. It is worth noting (albeit not relevant for
our purposes) that the bi-invariant metric
d((gn )n∈N , (hn )n∈N ) = lim min{dn (gn , hn ), 1},
n→U
gn , hn ∈ Gn
on ∏n∈N Gn induces a bi-invariant metric on ∏n→U (Gn , dn ).
The above definitions will be relevant to us in the following setting.
Let (kn )n∈N be a sequence of natural numbers and consider the family of
matrix algebras Vn ∶= Mkn (C) equipped with some unitarily invariant,
submultiplicative norms ∥⋅∥n . We usually omit the index and denote
all the norms by ∥⋅∥. Let Gn ∶= U(kn ), equipped with the metrics
dist∥⋅∥n (g, h) = ∥g − h∥n , g, h ∈ Gn induced from the norms. We consider
the ultraproduct Banach space
MU ∶= ∏ (Mkn (C), ∥⋅∥n ),
∥⋅∥
n→U
and the metric ultraproduct
UU ∶= ∏ (U(kn ), dist∥⋅∥n ).
∥⋅∥
n→U
By submultiplicativity of the norms, we see that if un ∈ U(kn ) with
limn→U dist∥⋅∥n (un , 1kn ) = 0 then
lim ∥un Tn − Tn ∥n ≤ lim ∥un − 1kn ∥n ⋅ ∥Tn ∥n = 0,
n→U
n→U
for all bounded sequences Tn ∈ Mkn (C). Thus left multiplication by
∥⋅∥
∥⋅∥
un induces a left action of UU on MU . By unitary invariance of the
norms, we see that this action is isometric. Similarly, we have a right
action by right multiplication and another left action by conjugation –
both of them isometric.
8
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
1.4. Asymptotic homomorphisms. In this section, we let Γ =
⟨S ∣ R⟩ be a fixed finitely presented group (i.e. S and R are finite) and
we let C be a class of groups, all equipped with bi-invariant metrics.
Any map ϕ∶ S → G, for some G ∈ C, uniquely determines a homomorphism FS → G which we will also denote by ϕ.
Definition 1.5. Let G ∈ C and let ϕ, ψ∶ S → G be maps. The defect
of ϕ is defined by
def(ϕ) ∶= max dG (ϕ(r), 1G ).
r∈R
The distance between ϕ and ψ is defined by
dist(ϕ, ψ) = max dG (ϕ(s), ψ(s)).
s∈S
The homomorphism distance of ϕ is defined by
HomDist(ϕ) ∶=
inf
π∈Hom(Γ,G)
dist(ϕ, π∣S ).
Definition 1.6. A sequence of maps ϕn ∶ S → Gn , for Gn ∈ C, is
called an asymptotic homomorphism if limn→U def(ϕn ) = 0.
We will mainly be concerned with finite dimensional asympotic
representations, that is, asymptotic homomorphisms with respect to
the class of unitary groups U(n) on finite dimensional Hilbert spaces,
equipped with the metrics
d(T, S) = ∥T − S∥,
T, S ∈ U(n),
coming from some family of unitarily invariant norms ∥⋅∥. The class of
finite-dimensional unitary groups with metrics coming from∥⋅∥op , ∥⋅∥Frob
and ∥⋅∥HS are denoted Uop , UFrob , and UHS .
We might also find the need to quantify the above definition.
Definition 1.7. Let ε > 0 and G ∈ C. An ε-almost homomorphism
is a map ϕ∶ S → G such that def(ϕ) ≤ ε.
In the literature, there are many different (inequivalent) notions of
“almost”, “asymptotic” and “quasi-” homomorphisms. If one would be
precise, the above notion of asymptotic homomorphism could be called
a local, discrete asymptotic homomorphism. Local, since we are only
interested in the behaviour of ϕn on the set of relations R (compare
with the uniform situation [13]) and discrete, because the family of
homomorphisms are indexed by the natural numbers.
Definition 1.8. Let Gn ∈ C, n ∈ N. Two sequences ϕn , ψn ∶ S → Gn
are called (asymptotically) equivalent if limn→U dist(ϕn , ψn ) = 0.
If an asymptotic homomorphism (ϕn )n∈N is equivalent to a sequence
of genuine representations, we call (ϕn )n∈N trivial or liftable.
We will now come to two central notions that we study in this
paper, the notion of stability and approximability by a class of metric
groups.
ASYMPTOTIC REPRESENTATIONS
9
Definition 1.9. The group Γ is called C-stable if all asymptotic
homomorphisms are equivalent to a sequence of homomorphisms, that
is,
lim HomDist(ϕn ) = 0,
n→U
for all ϕn ∶ S → Gn , Gn ∈ C, n ∈ N with limn→U def(ϕn ) = 0.
Definition 1.10. A finitely presented group Γ = ⟨S ∣ R⟩ is called Capproximated, if there exists an asymptotic homomorphism ϕn ∶ S → Gn ,
Gn ∈ C, n ∈ N such that
lim dn (ϕn (x), 1Gn ) > 0,
n→U
for all x ∈ FS /⟪R⟫.
We will be mainly concerned with UFrob -approximation and UFrob stability in this paper and, for convenience, we will often just speak
about Frobenius-approximation and Frobenius-stability in this context.
Definition 1.11. A group Γ is called residually C if for all x ∈
Γ/{1Γ} there is a homomorphism π∶ Γ → G for some G ∈ C such that
π(x) ≠ 1G .
The following proposition (see [25] or [3]) is evident from the definitions, nevertheless a central observation in our work.
Proposition 1.12. Let Γ be a finitely presented. If Γ is C-stable
and C-approximated group, then it must be residually C. In particular,
if the class C consists of finite-dimensional unitary groups, any finitely
presented, C-stable and C-approximated group is residually finite.
We finish this section with a basic lemma. The important part in
the statement of the lemma is that Kr does not depend on ϕ.
Lemma 1.13. For all r ∈ ⟪R⟫ there is a constant Kr such that for
all groups G with a bi-invariant metric and all maps ϕ∶ S → G it holds
that
dG (ϕ(r), 1G ) ≤ Kr def(ϕ).
Proof. If r ∈ ⟪R⟫ we can determine r1 , . . . , rk ∈ R ∪ R−1 and
x1 , . . . , xk ∈ FS such that
−1
−1
r = x1 r1 x−1
1 x2 r2 x2 ⋯xk rk xk .
Note that by bi-invariance
dG (ϕ(rj ), 1G ) = dG (ϕ(rj−1 ), 1G ) ≤ def(ϕ)
10
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
for all j. Thus, using bi-invariance again, we get
−1
dG (ϕ(r), 1kn ) = dG (ϕ(x1 r1 x−1
1 )⋯ϕ(xk rk xk ), 1G )
≤ ∑ dG (ϕ(xj )ϕ(rj )ϕ(xj )−1 , 1G )
k
j=1
= ∑ dG (ϕ(rj ), 1G )
k
j=1
≤ k ⋅ def(ϕ),
So letting Kr = k, we are done.
1.5. Group cohomology. For convenience, we recall one construction of group cohomology. We primarily need the second cohomology of a group with coefficients in a unitary representation, but for
completeness, we give a more general definition. Let Γ be any group
and let V be a Γ-module, i.e. an abelian group together with a (left)
action π of Γ on V . We consider the chain complex C n (Γ, V ), n ≥ 1,
which is the set of functions from Γn to V together with the coboundary
operator,
d = dn ∶ C n (Γ, V ) → C n+1 (Γ, V ),
defined by
dn (f )(g1 , . . . , gn+1 ) = π(g1 )f (g2 , . . . , gn+1 )
+ ∑(−1)j f (g1 , . . . , gj gj+1, . . . , gn+1 )
n
j=1
+ (−1)n+1 f (g1 , . . . , gn ).
We also let C 0 (Γ, V ) = V and d0 (v)(g) = π(g)v−v for v ∈ V, g ∈ Γ. Thus,
for n ≥ 0 we define the n-coboundaries to be B n (Γ, V ) = Im(dn−1 ) (with
B 0 (Γ, V ) = {0}) and the n-cocycles to be Z n (Γ, V ) = ker(dn ). One
checks that B n (Γ, V ) ⊆ Z n (Γ, V ), so we can define the n-th cohomology
to be
H n (Γ, V ) = Z n (Γ, V )/B n (Γ, V ).
Recall that given an extension of groups
i
q
1 → V → Γ̂ → Γ → 1,
where V is abelian, there is an action of Γ on V induced by the conjugation action of Γ̂ on i(V ). Fixing any section σ∶ Γ → Γ̂ (with σ(1Γ ) = 1Γ̂ )
of the quotient q we can define a map map f ∶ Γ × Γ → V as the solution
to
i(f (g, h)) = σ(g)σ(h)σ(gh)−1 ,
for g, h ∈ Γ. It is straightforward to check that f ∈ Z 2 (Γ, V, π) and
f ∈ B 2 (Γ, V, π) exactly when the extension splits, i.e. there is a homomorphism p∶ Γ → Γ̂ such that q ○ p = idΓ .
ASYMPTOTIC REPRESENTATIONS
11
Assume now that Γ is countable and V is a Banach space with norm
∥⋅∥, then we can define a separating family of semi-norms on C n (Γ, V )
by
∥f ∥F = max ∥f (g)∥,
(1.1)
g∈F
for f ∈ C n (Γ, V ) and finite F ⊆ Γn . It is easy to see that with respect to
this family, C n (Γ, V ) is a Fréchet space (one can even take ∥⋅∥{x} , x ∈ Γn
as separating family) and if Γ acts on V by isometries, the map dn is
bounded.
2. Some examples of non Frobenius-stable groups
Part of our aim is to provide a large class of Frobenius-stable groups,
but let us start out by giving examples of well-known groups that are
not stable. Specifically, we show that Z2 and the Baumslag-Solitar
group BS(2, 3) are not Frobenius-stable by giving concrete examples
of asymptotic representations that are not equivalent to genuine representations. We also exploit the latter example to provide an example
of an Frobenius-approximated, non-residually finite group, see Section
2.3.
2.1. Z2 is not Frobenius-stable. In [51], Voiculescu proved that
the matrices
0 1⎞
⎛1
⎞
⎛0
0⎟
⎜ ωn
⎟
⎜1 0
⎜
⎟
⎜
⎟
ωn2
⎟,
⎟ ∈ U(n),
An = ⎜
Bn = ⎜ 1
⎜
⎟
⎜
⎟
⎜
⎟
⎜
⎟
⋱
⋱
n−1
⎝
⎝
ω ⎠
1 0⎠
n
where ωn ∶= exp( 2πi
n ), n ∈ N, define a non-trivial ∥⋅∥op -asymptotic representation of Z2 = ⟨a, b, ∣ aba−1 b−1 ⟩ by ϕn (a) = An and ϕn (b) = Bn . More
precisely,
but
def ∥⋅∥op (ϕn ) = ∥An Bn A∗n Bn∗ − 1n ∥op = ∣ωn − 1∣ = OU ( n1 ),
HomDist∥⋅∥op (ϕn ) ≥
√
2 − ∣1 − ωn ∣ − 1
(see also [21, 51]). By the inequalities ∥T ∥op ≤ ∥T ∥Frob ≤ n1/2 ∥T ∥op for
T ∈ U(n), we conclude that
and
def ∥⋅∥Frob (ϕn ) = OU (n−1/2 ),
HomDist∥⋅∥Frob (ϕn ) ≥
√
2 − ∣1 − ωn ∣ − 1,
so ϕn is also a non-trivial ∥⋅∥Frob -asymptotic representation. In particular Z2 is neither Uop - nor Frobenius-stable. It is worth noting that Z2
actually is UHS -stable, see e.g. [24] for a quantitative proof.
12
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
2.2. BS(2, 3) is not Frobenius-stable. We now turn our attention to the Baumslag-Solitar group BS(2, 3) = ⟨a, b ∣ b−1 a2 ba−3 ⟩, see
[6] for the original reference. By definition, the generators satisfy the
equation
(2.1)
b−1 a2 b = a3 .
It is also well known and not hard to check that the generators do not
satisfy
ab−1 ab = b−1 aba.
(2.2)
Indeed, this follows easily from the description of BS(2, 3) as an HNNextension of Z. On the other hand, we recall the following.
Proposition 2.1 (Baumslag-Solitar [6]). Let Γ be a residually finite group. If a, b ∈ Γ satisfy (2.1), then they also satisfy (2.2).
Proof. Indeed, if a has finite order and a2 is conjugate to a3 , then
the order of a cannot be even. Thus, b−1 ab is a power of b−1 a2 b = a3 .
We conclude that a and b−1 ab commute.
By Mal’cev’s Theorem we immediately obtain the following consequence.
Corollary 2.2. Let a, b be unitary matrices. If a, b ∈ Γ satisfy
(2.1), then they also satisfy (2.2).
This last Corollary can also be proven directly by linear algebra
methods, see [22] where some quantitative aspects of operator-norm
aproximability of BS(2, 3) were studied. By Corollary 2.2, in order to
show that BS(2, 3) is non-stable it suffices to find a sequence of pairs
of unitary matrices that ∥⋅∥Frob -asymptotically satisfy Equation (2.1)
but are far from satisfying Equation (2.2). The study of approximation
properties of BS(2, 3) goes back to Rădulescu [40], where the focus was
more on approximation in the (normalized) Hilbert-Schmidt norm. We
are now going to prove the following result.
Theorem 2.3. The group BS(2, 3) is not Frobenius-stable.
The theorem is a direct consequence of the following lemma.
Lemma 2.4. There exist An , Bn ∈ U(6n) such that
● ∥Bn−1 A2n Bn − A3n ∥Frob = OU ( n1 )
√
● ∥An Bn−1 An Bn − Bn−1 An Bn An ∥Frob = 6n − OU (1).
Proof. We will omit the index and write A = An and B = Bn .
Let ω = exp( 2π
6n ) and consider a 6n-dimensional Hilbert space H with
orthonormal basis v[0], v[1], . . . v[6n − 1]. Define A ∈ U(6n) as Av[j] =
ω j v[j] (that is, A is A6n from the previous example). We plan to den−1
compose H as a direct sum in two ways H = ⊕n−1
j=0 S[j] and H = ⊕j=0 C[j]
such that each S[j] and each C[j] is 6-dimensional and the restriction
of A2 to S[j] as well as restriction of A3 to C[j] act approximately as
ASYMPTOTIC REPRESENTATIONS
13
multiplication by ω 6j . (The letter S stands for square and C stands for
cube.) Then we construct B = ⊕n−1
j=0 Bj with Bj ∶ C[j] → S[j]. Let us
start the detailed construction. Define
S[j] = span{v[3j], v[3j + 1], v[3j + 2],
v[3j + 3n], v[3j + 3n + 1], v[3j + 3n + 2]}
and
C[j] = span{v[2j], v[2j + 2n], v[2j + 4n],
v[2j + 1], v[2j + 2n + 1], v[2j + 4n + 1]}.
We will use the ordered base of S[j] (resp. C[j]) as it appears in their
definitions. Let Sj (resp. Cj ) be a restriction of A to S[j] (resp. C[j]).
Observe that
Sj = ω 3j diag(1, ω, ω 2 , −1, −ω, −ω 2 )
and
4πi
2πi
4πi
Cj = ω 2j diag(1, exp( 2πi
3 ), exp( 3 ), ω, ω exp( 3 ), ω exp( 3 )).
Now, let B ∈ U(6n) be any unitary of the form B = ⊕n−1
j=0 Bj with
unitary Bj ∶ C[j] → S[j]. We claim that
∥B −1 A2 B − A3 ∥Frob = OU ( n1 )
Indeed,
∥B −1 A2 B − A3 ∥2Frob = ∑ ∥Bj−1 Sj2 Bj − Cj3 ∥2Frob ,
n−1
j=0
and we obtain
∥Bj−1 Sj2 Bj − Cj3 ∥2Frob = ∥Bj−1 (Sj2 − ω 6j )Bj − (Cj3 − ω 6j )∥2Frob
≤ ∥Sj2 − ω 6j ∥2Frob + ∥Cj3 − ω 6j ∥2Frob
= 2(∣1 − ω 2∣2 + ∣1 − ω 4∣2 ) + 3∣1 − ω 3 ∣2 = OU ( n12 ),
which entails the claim. Now, consider the unitary given by the matrix
and let
⎛1 1 0 0 0 0 ⎞
⎜1 −1 0 0 0 0 ⎟
⎟
1 ⎜
⎜0 0 1 1 0 0 ⎟
⎟,
Bj = √ ⎜
⎟
2⎜
⎜0 0 1 −1 0 0 ⎟
⎜0 0 0 0 1 1 ⎟
⎝0 0 0 0 1 −1⎠
S̃ = diag(1, 1, 1, −1, −1, −1)
and
4πi
2πi
4πi
C̃ = diag(1, exp( 2πi
3 ), exp( 3 ), 1, exp( 3 ), exp( 3 )).
It is not hard to check that
∥Sj − ω 3j S̃∥2Frob = OU ( n12 )
14
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
and
∥Cj − ω 2j C̃∥2Frob = OU ( n12 ).
Direct calculations show that ∥C̃Bj−1 S̃Bj − Bj−1 S̃Bj C̃∥2Frob = 6, so since
∥AB −1 AB − B −1 ABA∥2Frob = ∑ ∥Cj Bj−1 Sj Bj − Bj−1 Sj Bj Cj ∥2Frob ,
n−1
j=0
the lemma follows.
2.3. An example of a finitely generated, non-residually finite, and Frobenius-approximated group. Note that the example
above provides a homomorphism into the ultraproduct ϕ∶ BS(2, 3) →
∥⋅∥
UU Frob . The image Γ = ϕ(BS(2, 3)) is clearly Frobenius-approximated,
but it is clearly not residually finite, since, by construction, the elements ϕ(a), ϕ(b) ∈ Γ satisfy (2.1) but not (2.2). In some sense it is an
artefact of the definitions that every non-Frobenius-stable group has
a non-trivial Frobenius-approximated group quotient. It seems quite
likely that the construction above is enough to show that BS(2, 3) is
itself Frobenius-approximated. Indeed, even though the proof of this
assertion is not spelled out in full detail in [40], it appears that Rădulescu’s construction shows this. Note that it follows from work of
Kropholler [30] that BS(2, 3) is residually solvable and hence MF, see
[14].
3. Diminishing the defect of asymptotic representations
This section contains the key technical novelty of this article. We
associate an element [α] ∈ H 2 (Γ, ∏n→U (Mkn (C), ∥⋅∥)) to an asymptotic
representation ϕn ∶ Γ → U(kn ). We prove that if [α] = 0, then the defect can be diminished in the sense that there is an equivalent asymptotic representation ϕ′n with effectively better defect, more precisely
def(ϕ′n ) = oU (def(ϕn )).
3.1. Assumptions for this section. For this section, we fix the
following.
● A finitely presented group Γ = ⟨S ∣ R⟩,
● a sequence of natural numbers (kn )n∈N ,
● a family of submultiplicative, unitarily invariant norms on
U(k), k ∈ N, all denoted by ∥⋅∥, and
● an asymptotic representation ϕn ∶ S → U(kn ) with respect to
the metrics associated to ∥⋅∥.
Recall the ultraproduct notation introduced in Section 1.3, that is,
∥⋅∥
∥⋅∥
UU = ∏n→U (U(kn ), dist∥⋅∥ ) and MU = ∏n→U (Mkn (C), ∥ ⋅ ∥) and recall
∥⋅∥
∥⋅∥
that since ∥⋅∥ is submultiplicative, UU acts on MU by multiplication.
An asymptotic representation as above induces a homomorphism ϕU ∶
ASYMPTOTIC REPRESENTATIONS
∥⋅∥
15
∥⋅∥
Γ → UU on the level of the group Γ. Thus Γ acts on MU through ϕU .
With this in mind, we also want to fix the following:
● a section σ∶ Γ → FS of the natural surjection FS → Γ, in particular, we have σ(g)σ(h)σ(gh)−1 ∈ ⟪R⟫ for all g, h ∈ Γ,
● a sequence ϕ̃n ∶ Γ → U(kn ) such that ϕ̃(1Γ ) = 1kn , ϕ̃n (g −1 ) =
ϕ̃n (g)∗ , and for every g ∈ Γ
∥ϕn (σ(g)) − ϕ̃n (g)∥ = OU (def(ϕn )).
(3.1)
In particular, the sequence ϕ̃n is a lift of ϕU .
For this, note that given any section σ∶ Γ → FS , the sequence ϕn ○ σ is a
lift of ϕU . There exists a section σ with σ(1Γ ) = 1FS and σ(g −1 ) = σ(g)−1
for all g such that g 2 ≠ 1Γ . We define ϕ̃n (g) ∶= ϕn (σ(g)) for all g with
g 2 ≠ 1Γ . In the case g 2 = 1Γ , by Lemma 1.13 it holds that
∥ϕn (σ(g))2 − 1kn ∥ = OU (def(ϕn )),
so by Proposition 1.4 there are self-adjoint unitaries Bn ∈ U(kn ) such
that
∥Bn − ϕn (σ(g))∥ = OU (def(ϕn )).
By letting ϕ̃n (g) ∶= Bn , we get the desired map.
3.2. The cohomology class of an asymptotic representa∥⋅∥
tion. We want to define an element in H 2 (Γ, MU ) associated to ϕn .
To this end we define cn ∶= cn (ϕn ) ∶ Γ × Γ → Mkn (C) by
cn (g, h) =
ϕ̃n (g)ϕ̃n (h) − ϕ̃n (gh)
,
def(ϕn )
for all n ∈ N such that def(ϕn ) > 0 and cn (g, h) = 0 otherwise, for all
g, h ∈ Γ. The next proposition is a collection of basic properties of the
maps cn .
Proposition 3.1. Let g, h, k ∈ Γ. The maps cn satisfy the following
equations
ϕ̃n (g)cn (h, k) − cn (gh, k) + cn (g, hk) − cn (g, h)ϕ̃n (k) = 0,
cn (g, g ) = cn (1Γ , g) = cn (g, 1Γ ) = 0
−1
and
cn (g, h)∗ = cn (h−1 , g −1 ),
Furthermore, we have for every g, h ∈ Γ
∥cn (g, h)∥ = OU (1).
(3.2)
16
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
Proof. For all g, h, k ∈ Γ and n ∈ N we have
def(ϕn ) ⋅ (ϕ̃n (g)cn (h, k) − cn (gh, k) + cn (g, hk) − cn (g, h)ϕ̃n (k))
= ϕ̃n (g)(ϕ̃n (h)ϕ̃n (k) − ϕ̃n (hk)) − (ϕ̃n (gh)ϕ̃n (k) − ϕ̃n (ghk))
+ (ϕ̃n (g)ϕ̃n (hk) − ϕ̃n (ghk)) − (ϕ̃n (g)ϕ̃n (h) − ϕ̃n (gh))ϕ̃n (k)
= ϕ̃n (g)ϕ̃n (h)ϕ̃n (k) − ϕ̃n (g)ϕ̃n (hk) − ϕ̃n (gh)ϕ̃n (k) + ϕ̃n (ghk)
+ ϕ̃n (g)ϕ̃n (hk) − ϕ̃n (ghk) − ϕ̃n (g)ϕ̃n (h)ϕ̃n (k) + ϕ̃n (gh)ϕ̃n (k)
= 0,
which proves the first equation. The second line of equations is immediate from the definition of cn and the fact that ϕ̃n (g −1 ) = ϕ̃n (g)∗ .
For the last assertion, note that since σ(g)σ(h)σ(gh)−1 ∈ ⟪R⟫, it
follows from Lemma 1.13 that
∥ϕn (σ(g)σ(h)σ(gh)−1 ) − 1kn ∥ = OU (def(ϕn )).
and thus it follows (by using Equation (3.1)) that
def(ϕn )∥cn (g, h)∥ = OU (def(ϕn )).
By (3.2) it follows that for every g, h ∈ Γ, cn (g, h) is a bounded
sequence, so the sequence defines a map
c = (cn )n∈N ∶ Γ × Γ → MU .
∥⋅∥
This map is not a cocycle in the sense explained in Section 1.5, but, as
the next corollary states, the map α(g, h) ∶= c(g, h)ϕU (gh)∗ is. (The
map c is a cocycle in the equivalent picture of Hochschild cohomology
and it turns out that some calculations are more natural with c, so
we will also work with this map.) Even though we suppress it in the
notation, keep in mind that c and α depend on the lift ϕ̃n and on
def(ϕn ).
∥⋅∥
Corollary 3.2. The map α∶ Γ×Γ → MU is a 2-cocycle with respect
∥⋅∥
to the isometric action π(g)T = ϕU (g)T ϕU (g)∗ , g ∈ Γ, T ∈ MU .
Proof. Given g, h, k ∈ Γ we have that
ϕU (g)α(h, k)ϕU (g)∗ − α(gh, k) + α(g, hk) − α(g, h)
= ϕU (g)c(h, k)ϕU (hk)∗ ϕU (g)∗ − c(gh, k)ϕU (ghk)∗
+ c(g, hk)ϕU (ghk)∗ − c(g, h)ϕU (gh)∗
= (ϕU (g)c(h, k) − c(gh, k) + c(g, hk) − c(g, h)ϕU (k))ϕU (ghk)∗
= 0,
where we used that ϕU is a homomorphism and Proposition 3.1.
We call α the cocycle associated to the sequence (ϕn )n∈N .
ASYMPTOTIC REPRESENTATIONS
17
Proposition 3.3. Assume that α represents the trivial cohomology
∥⋅∥
∥⋅∥
class in H 2(Γ, MU ), i.e. there exists a map β∶ Γ → MU satisfying
Then
α(g, h) = ϕU (g)β(h)ϕU (g)∗ − β(gh) + β(g),
β(1Γ) = 0,
g, h ∈ Γ.
(3.3)
β(g) = −ϕU (g)β(g )ϕU (g)
−1
∗
c(g, h) = ϕU (g)β(h)ϕU (h) − β(gh)ϕU (gh) + β(g)ϕU (gh).
(3.4)
(3.5)
Furthermore, we can choose β(g) to be skew-symmetric for all g ∈ Γ.
Proof. Equation (3.5) is immediate from c(g, h) = α(g, h)ϕU (gh)
for g, h ∈ G. Equation (3.3) follows from (3.5) and Proposition 3.1 with
g = h = 1Γ and (3.4) follows from (3.3), (3.5) and Proposition 3.1 with
h = g −1 . For the last claim, we possibly need to alter β a little. Note
that β ′ (g) ∶= −β(g)∗ = ϕU (g)β(g −1 )∗ ϕU (g)∗ also satisfies (3.3)-(3.5).
Indeed,
c(g, h) = c(h−1 , g −1 )∗
= (ϕU (h−1 )β(g −1 )ϕU (g −1 ) − β(h−1 g −1)ϕU (h−1 g −1 ) + β(h−1 )ϕU (h−1 g −1 ))∗
= ϕU (g)β(g −1 )∗ ϕU (h) − ϕU (gh)β((gh)−1 )∗ + ϕU (gh)β(h−1 )∗
= β ′ (g)ϕU (gh) − β ′ (gh)ϕU (gh) + ϕU (g)β ′ (h)ϕU (h)
for g, h ∈ Γ, which proves (3.5) whence the other two follow. Thus,
replacing β with
β ♯ (g) ∶=
β(g) − β(g)∗
,
g ∈ Γ, n ∈ N,
2
we see that β ♯ (g) is skew-symmetric and that Equations (3.3)-(3.5) are
still satisfied.
3.3. Correction of the asymptotic representation. Now let
β be as above and let βn ∶ Γ → Mkn (C) be any skew-symmetric lift of
β. Then exp(− def(ϕn )βn (g)) is a unitary for every g ∈ Γ, so we can
define a sequence of maps ψn ∶ Γ → U(kn ) by
ψn (g) = exp(− def(ϕn )βn (g))ϕ̃n (g).
Note that since ϕ̃n (1Γ ) = 1kn and βn (1Γ ) = 0, we have ψn (1Γ ) = 1kn .
In the proofs of Proposition 3.4 and Lemma 3.5, we will make use
of two basic inequalities that hold for any k ∈ N and A ∈ Mk (C):
∥1k − exp(A)∥ ≤ ∥A∥ exp(∥A∥)
∥1k − A − exp(A)∥ ≤ ∥A∥2 exp(∥A∥),
(3.6)
(3.7)
k
∞
∑k=0 Ak!
They are simple consequences of the definition exp(A) =
the triangle inequality and submultiplicativity of the norm.
and
18
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
Proposition 3.4. With the notation from above, for every g ∈ Γ,
we have
∥ϕ̃n (g) − ψn (g)∥ = OU (def(ϕn )).
More precisely,
for most n ∈ N.
∥ϕ̃n (g) − ψn (g)∥ ≤ 2∥βn (g)∥ def(ϕn )
Proof. Let g ∈ Γ. By unitary invariance and submultiplicativity,
we get that
∥ϕ̃n (g) − ψn (g)∥ = ∥1kn − exp(− def(ϕn )βn (g))∥
≤ def(ϕn )∥βn (g)∥ exp(def(ϕn )∥βn (g)∥)
(3.6)
and since ∥βn (g)∥ is a bounded sequence and limn→U def(ϕn ) = 0, we
have exp(def(ϕn )∥βn (g)∥) ≤ 2 for most n and the result follows.
It follows that ψn ∣S is an asymptotic representation with def(ψn ∣S ) =
OU (def(ϕn )), but we prove that the defect is actually oU (def(ϕn )).
Lemma 3.5. For any g, h ∈ Γ, we have that
∥ψn (gh) − ψn (g)ψn (h)∥ = oU (def(ϕn )).
Proof. Let ξn (x) ∶= (1kn − def(ϕn )βn (x))ϕ̃n (x), for x ∈ Γ, and let
g, h ∈ Γ be fixed. Let C = 2 maxx∈{g,h,gh} ∥β(x)∥. Whence it follows that
for most n ∈ N,
∥ψn (x) − ξn (x)∥ ≤ C ⋅ def(ϕn )2 ,
(3.7)
for x ∈ {g, h, gh}. By the above (and by submultiplicativity) it follows
that
∥ψn (gh) − ψn (g)ψn (h)∥ = ∥ξn (gh) − ξn (g)ξn (h)∥ + oU (def(ϕn ))
so it suffices to show that
∥ξn (gh) − ξn (g)ξn (h)∥ = oU (def(ϕn ))
which amounts to the following calculations
ξn (gh) − ξn (g)ξn (h)
= ϕ̃n (gh) − ϕ̃n (g)ϕ̃n (h)
+ def(ϕn )(−βn (gh)ϕ̃n (gh) + ϕ̃n (g)βn (h)ϕ̃n (h) + βn (g)ϕ̃n (g)ϕ̃n (h))
− def(ϕn )2 βn (g)ϕ̃n (g)βn (h)ϕ̃(h)
= def(ϕn )(−cn (g, h)
+ ϕ̃n (g)βn (h)ϕ̃n (h) − βn (gh)ϕ̃n (gh) + βn (g)ϕ̃n (g)ϕ̃n (h))
− def(ϕn )2 βn (g)ϕ̃n (g)βn (h)ϕ̃n (h).
ASYMPTOTIC REPRESENTATIONS
19
By Equation (3.5) and the fact that submultiplicativity of the norm implies that ∥βn (g)ϕ̃n (g)βn (h)ϕ̃n (h)∥ is bounded, this finishes the proof.
At last we define the asymptotic representation ϕ′n ∶ S → U(kn ) by
ϕ′n = ψn ∣S and reach the desired conclusion def(ϕ′n ) = oU (def(ϕn )). Let
us, for reference’s sake, formulate the result properly.
Theorem 3.6. Let Γ = ⟨S ∣ R⟩ be a finitely presented group and
let ϕn ∶ S → U(kn ) be an asymptotic representation with respect to a
family of submultiplicative, unitarily invariant norms. Assume that
∥⋅∥
the associated 2-cocycle α = α(ϕn ) is trivial in H 2 (Γ, MU ). Then there
exists an asymptotic representation ϕ′n ∶ S → U(kn ) such that
(i) dist(ϕn , ϕ′n ) = OU (def(ϕn )) and
(ii) def(ϕ′n ) = oU (def(ϕn )).
Proof. We adopt the above notation. Assertion (i) follows from
Proposition 3.4; let r = x1 x2 ⋯xm ∈ R be written as a reduced word,
where xj ∈ S ∪ S −1 , j = 1, . . . , m. By iteration of Lemma 3.5 (using that
ψn takes unitary values and that ∥⋅∥ is unitarily invariant), we see that
∥ϕ′n (r) − 1kn ∥ = ∥ψn (x1 )ψn (x2 )⋯ψn (xm ) − 1kn ∥
= ∥ψn (x1 x2 )ψn (x3 )⋯ψn (xm ) − 1kn ∥ + oU (def(ϕn ))
⋮
= ∥ψ(1Γ ) − 1kn ∥ + oU (def(ϕn )).
Since ψ(1Γ ) = 1kn , we are done.
The converse of Theorem 3.6 is also valid in the following sense.
Proposition 3.7. Let Γ = ⟨S ∣ R⟩ be a finitely presented group, let
ϕn , ψn ∶ S → U(kn ) be asymptotic representations with respect to some
family of submultiplicative, unitarily invariant norms and suppose
● dist(ϕn , ψn ) = OU (def(ϕn )) and
● def(ψn ) = oU (def(ϕn )).
Then, the 2-cocycle α associated to (ϕn )n∈N is trivial in H 2 (Γ, MU ). In
particular, if ϕn is sufficiently close to a homomorphism, α is trivial.
∥⋅∥
Proof. If def(ϕn ) = 0 for most n ∈ N there is nothing to prove, so
let us assume this is not the case. Let ϕ̃n , ψ̃n ∶ Γ → U(kn ) be the induced
maps we get by fixing a section Γ → FS as explained in the beginning
of this section. We note that the sequences ϕ̃n and ψ̃n induce the same
map ϕU in the limit. Define
γn (g) =
ϕ̃n (g) − ψ̃n (g)
def(ϕn )
for n with def(ϕn ) > 0 and γn (g) = 0 otherwise. By the first bullet
in our assumptions, γn is essentially bounded, so it defines an element
20
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
∥⋅∥
γ(g) ∈ MU . If we prove that
c(g, h) = ϕU (g)γ(h) − γ(gh) + γ(g)ϕU (h),
it will follow easily that β(g) ∶= γ(g)ϕU (g)∗ will satisfy dβ = α. First
note that it follows from the second bullet in our assumptions that for
every g, h ∈ Γ
thus
∥ψ̃n (gh) − ψ̃n (g)ψ̃n (h)∥ = oU (def(ϕn )),
def(ϕn ) ⋅ (ϕ̃n (g)γn (h) − γn (gh) + γn (g)ψ̃n (h))
= ϕ̃n (g)ϕ̃n (h) − ϕ̃n (g)ψ̃n (h) − ϕ̃n (gh) + ψ̃n (gh)
+ ϕ̃n (g)ψ̃n (h) − ψ̃n (g)ψ̃n (h)
= ϕ̃n (g)ϕ̃n (h) − ϕ̃n (gh) + ψ̃n (gh) − ψ̃n (g)ψ̃n (h)
= def(ϕn ) ⋅ cn (g, h) + oU (def(ϕn )).
Now the result follows by dividing by def(ϕn ) (which is possible for
most n) and taking the limit.
It is now clear that we are in need of large classes of groups for
which general vanishing results for the second cohomology with Banach
or Hilbert space coefficients can be proven. This will be the subject of
the next section. But first let us mention an alternative approach that
can be used to prove Theorem 3.6.
3.4. Asymptotic representations and extensions. As mentioned in Section 1.5, the second cohomology characterizes extensions
of Γ with abelian kernel and that in this picture coboundaries correspond to splitting extensions. Thus Theorem 3.6 and Corollary 3.7
show that finding the improved ϕ′n is equivalent to finding a splitting
for a certain extension. The connection between asymptotic representations and extensions can be seen directly without going through the
above computations, and this idea can actually be used to prove Theorem 3.6. Since this approach is very illustrative (it shows, for instance,
very clearly what rôle submultiplicativity plays), we sketch the proof.
We retain the assumptions from Section 3.1 and introduce some
more notation. Letting εn ∶= def(ϕn ), for n ∈ N, we define
N(OU (εn )) = {(un )n∈N ∈ ∏ U(kn ) ∣ ∥un − 1kn ∥ = OU (εn )}
n∈N
and
U(OU (εn )) = ∏ U(kn )/N(OU (εn )).
n∈N
Similarly, we define N(oU (εn )) and U(oU (εn )). We saw that the asymp∥⋅∥
totic representation (ϕn )n∈N induces a homomorphism ϕU ∶ Γ → UU ,
ASYMPTOTIC REPRESENTATIONS
21
but Lemma 1.13 actually implies the existence of an induced homomorphism
ϕ̂U ∶ Γ → U(OU (εn )).
Now we observe that the existence of ϕ′n ∶ S → U(kn ) with dist(ϕn , ϕ′n ) =
OU (def(ϕn )) and def(ϕ′n ) = oU (def(ϕn )) as in Theorem 3.6 is equivalent to the existence of a lift ϕ′U of ϕ̂U :
Γ
U(oU (εn ))
9
tt
tt
t
t
tt
tt
❏❏
❏❏
❏❏ϕ̂U
❏❏
❏❏
%
ϕ′U
U(OU (εn ))
We also see that the map ϕ̂U fits into the following commutative diagram
1
/
N
/
U(oU (εn ))
O
ψ
/
U(OU (εn ))
/
O
1
ϕ̂U
1
/
N
/
Γ̂
/
Γ
/
1
where Γ̂ is the pullback through ϕ̂ and ψ and
N ∶= N(OU (εn ))/N(oU (εn )).
Combining these two observations, it easily follows that ϕn can be improved to ϕ′n if and only if the bottom row in the latter diagram splits.
Now, since ∥⋅∥ is submultiplicative the group N is actually abelian.
Indeed, for all T, S ∈ U(k), we have that
∥T ST ∗ S ∗ − 1k ∥ = ∥T S − ST ∥
= ∥(T − 1k )(S − 1k ) − (S − 1k )(T − 1k )∥
≤ 2∥T − 1k ∥∥S − 1k ∥,
so if (Tn )n∈N , (Sn )n∈N ∈ N(O(εn )) then
(Tn Sn Tn∗ Sn∗ )n∈N ∈ N(O(ε2n )) ⊆ N(o(εn )),
whence the claim follows. Hence, as explained in Section 1.5, the extension 1 → N → Γ̂ → Γ → 1 corresponds to an element [α̂] ∈ H 2 (Γ, N),
and we conclude that ϕn can be improved if and only if [α̂] = [0]. Now,
∥⋅∥
the coefficients N are not exactly the same as MU in Theorem 3.6, but
with a little effort, one can prove that N is a real Banach space (or a
real Hilbert space in the case ∥⋅∥ = ∥⋅∥Frob ) with an isometric Γ-action
∥⋅∥
and the existence of an equivariant homomorphism θ∶ N → MU such
that [θ ○ α̂] = [α].
22
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
Remark 3.8. We note that this approach also works for the most
part if ∥⋅∥ is not submultiplicative. In this case, however, the group N
is not abelian and the second cohomology with non-abelian coefficients
is much less tractable in general.
This alternative approach to the problem at hand is rather conceptual and elegant, but also the proof that we chose to present in detail
has its merits. The cocycle α can be computed directly from (ϕn )n∈N ,
and in cases where the associated 1-cochain β can be computed explicitly from α, this gives us an explicit expression for ϕ′n .
4. Cohomology vanishing and examples of n-Kazhdan groups
Recall that if Γ is a finitely (or, more generally, compactly) generated group, then Γ has Kazhdan’s Property (T) if and only if the
first cohomology H 1 (Γ, Hπ ) = 0 for every unitary representation π∶ Γ →
U(Hπ ) on a Hilbert space Hπ , see [8] for a proof and more background
information. We will consider groups for which the higher cohomology
groups vanish. Higher dimensional vanishing phenomena have been
studied in various articles, see for example [4, 5, 11, 19, 20, 33, 35].
We propose the following terminology.
Definition 4.1. Let n ∈ N. A group Γ is called n-Kazhdan if
vanishes for all unitary representations (π, Hπ ) of Γ. We
call Γ strongly n-Kazhdan, if Γ is k-Kazhdan for k = 1, . . . , n.
H n (Γ, Hπ )
So 1-Kazhdan is the Kazhdan’s classical Property (T). See [4, 33]
for discussions of other related higher dimensional analogues of Property (T). It will be central in our proof that by an application of the
open mapping theorem, vanishing of cohomology with Hilbert space
coefficients implies that cocycles are coboundaries with control on the
norms. This is explained in the following proposition and its corollary,
where we use the terminology introduced in Equation (1.1).
Proposition 4.2. Let n ∈ N, let Γ be a countable group, let π∶ Γ →
U(Hπ ) be a unitary representation, and assume that H n (Γ, Hπ ) = {0}.
Then for every finite set F ⊆ Γn−1 there exist a finite set Fπ ⊆ Γn and a
constant Cπ,F ≥ 0 such that for every cocycle z ∈ Z n (Γ, Hπ ) there is an
element b ∈ C n−1 (Γ, Hπ ) such that z = dn−1 b and ∥b∥F < Cπ,F ∥z∥Fπ .
Proof. By definition of the topology on C n (Γ, Hπ ), the basic open
sets are given by
Uδ,F ′ = {f ∈ C n (Γ, Hπ ) ∣ ∥f ∥F ′ < δ},
for a finite F ′ ⊆ Γn and δ > 0. Since the map dn−1 ∶ C n−1 (Γ, Hπ ) →
Z n (Γ, Hπ ) is linear, bounded and surjective, the open mapping theorem
applies (see [44]), so there are Cπ,F > 0 and Fπ ⊆ Γn such that
n
n−1
−1 ,F ∩ Z (Γ, Hπ ) ⊆ d
(U1,F ).
UCπ,F
π
ASYMPTOTIC REPRESENTATIONS
23
−1
−1 ,F , so
z ∈ UCπ,F
In other words, if z ∈ Z n (Γ, Hπ ), ∥z∥Fπ = 1, then Cπ,F
π
there is b ∈ C n−1 (Γ, Hπ ) such that dn−1 b = z and ∥b∥F < Cπ,F = Cπ,F ∥z∥.
This proves the claim.
We need the fact that if H 2 (Γ, Hπ ) vanishes universally, the set Fπ
and the bound Cπ,F can be chosen universally for all unitary representations π. This is the consequence of an easy diagonalisation argument.
Corollary 4.3. Let n ∈ N and Γ be a countable n-Kazhdan group.
Then for every finite set F ⊆ Γn there are a finite set F0 ⊆ Γn−1 and a
constant CF ≥ 0 such that for all unitary representations π of Γ and all
cocycles z ∈ Z n (Γ, Hπ ) there is an element b ∈ C n−1 (Γ, Hπ ) such that
z = dn−1 b and ∥b∥F < CF ∥z∥F0 .
We also observe the following extension proposition.
Proposition 4.4. Consider a short exact sequence of groups.
1 → Λ → Γ̃ → Γ → 1.
If Λ is strongly n-Kazhdan and Γ is n-Kazhdan, then Γ̃ is also nKazhdan. In particular, this applies if Λ or Γ is finite.
Proof. By the Hochschild-Serre spectral sequence [12], it is enough
to show that H k (Γ, H l (Λ, Hπ∣Λ )) vanishes for all k, l ∈ N with k + l = n.
If l > 0, then H l (Λ, Hπ∣Λ ) vanishes. For l = 0 and k = n, we have
H 0 (Λ, Hπ∣Λ ) = Fix(π∣Λ ) (the set of fixed vectors in Hπ∣Λ ), which is a
Hilbert space, and the induced action of Γ is a unitary representation,
so we conclude that H n (Γ, H 0 (Λ, Hπ∣Λ )) vanishes.
In view of the previous section, it is natural to ask if there exists
a non-residually finite group, such that H 2 (Γ, A) vanishes for all C ∗ algebras A equipped with an action of Γ by automorphisms. We are not
able to answer this questions, however, one can show that H 1 (Γ, ℓ∞ (Γ))
does not vanish for any infinite group, which makes a positive answer
somewhat unlikely. Here, we view ℓ∞ (G) as a G-module with respect
to the right translation action. Indeed, let d∶ Γ × Γ → N be a proper
left-invariant metric. Then, c(g) ∶= (h ↦ d(1Γ , h)−d(1Γ , hg))h∈Γ defines
a cocycle c∶ Γ → ℓ∞ (Γ) which cannot be the boundary of an element in
ℓ∞ (Γ) if Γ is infinite.
4.1. Higher rank p-adic lattices are 2-Kazhdan. Finally, this
section provides examples, or every n ≥ 2, of groups which are nKazhdan. The results are essentially known and we recall them in
detail for convenience.
Let K be a non-archimedean local field of residue class q, i.e. if O ⊂
K is the ring of integers and m ⊂ O is its unique maximal ideal, then q =
∣O/m∣. Let G be a simple K-algebraic group of K-rank r and assume
that r ≥ 1. The group G ∶= G(K) acts on the associated BruhatTits building B. For more information on the theory of buildings, see
24
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
[1]. The latter is an infinite, contractible, pure simplicial complex of
dimension r, on which G acts transitively on the chambers, i.e. the topdimensional simplices. Let Γ be a uniform lattice in G, i.e. a discrete
cocompact subgroup of G. When Γ is also torsion free (which can
always be achieved by replacing Γ by a finite index subgroup), then
the quotient X ∶= Γ/B is a finite r-dimensional simplicial complex and
Γ = π1 (X). In particular, the group Γ is finitely presented. We will use
the following theorem which essentially appears in work of Ballmann
and Świątkowski [5] building on previous work of Garland [23].
Theorem 4.5. For every natural number r ≥ 2, there exists q0 (r) ∈
N such that the following holds. If q ≥ q0 (r) and G and Γ are as above,
then Γ is strongly (r − 1)-Kazhdan. In particular, if r ≥ 3, then Γ is
2-Kazhdan.
Recall, that being 1-Kazhdan is equivalent to Kazhdan’s property
(T). As it is well known, G and Γ as above have property (T) for every
r ≥ 2 and for all q. It is quite plausible that this is also true for in the
context of the preceding theorem.
Note that such Γ contains a finite-index torsion free group Λ. Proposition 4.4 implies that it suffices to prove that Λ is (r − 1)-Kazhdan. So
one can assume that Γ is torsion free.
Theorem 4.5 for finite dimensional Hilbert spaces is Theorem 8.3
in the seminal paper of Garland [23]. The general case is stated in the
last paragraph of Section 3.1 on page 631 in the work of BallmannŚwiątkowski [5]. It is deduced from Theorem 2.5 there: that Theorem
asserts á la Garland [23] that the desired cohomology vanishing follows
from sharp estimates of the spectral gap of the local Laplacians, i.e. the
Laplacians of the proper links of the complex. These estimates (called
also p-adic curvature) are given in Lemma 6.3 and Lemma 8.2 in [23].
So altogether Theorem 4.5 is proven. The method and estimates of
Garland are used also in [35, 53] and more recently [28, 33].
Let us give the reader just a notational warning: when we say rank
(following the common practice nowadays) we mean the K-rank of G
as a p-adic group (and we denoted it by r) and then it follows that the
dimension of the associated Bruhat-Tits building is equal to r. Garland
refers to the rank of the Tits system which in his notation he denotes
l + 1. Hence, our r is equal to his l.
It is very natural to wonder what happens in the analogous real
case. It is worth noting that already H 5(SLn (Z), R) is non-trivial for
n large enough [10]; thus SLn (Z) fails to be 5-Kazhdan for n large
enough. Similarly, note that H 2 (Sp(2n, Z), R) = R for all n ≥ 2 [10], so
that the natural generalization to higher rank lattices in real Lie groups
has to be formulated carefully; maybe just by excluding an explicit list
of finite-dimensional unitary representations.
Question 4.6. Is SLn (Z) 2-Kazhdan (at least for large n)?
ASYMPTOTIC REPRESENTATIONS
25
5. Proofs of the main results
In order to finish the proofs of Theorem 1.1 and Theorem 1.2, we
need to show that finitely presented 2-Kazhdan groups are Frobeniusstable and that some of them are not residually finite. The main result
follows then from Corollary 5.4 and the constructions in Section 5.2.
5.1. The Frobenius-stability of 2-Kazhdan groups. We now
consider 2-Kazhdan groups and asymptotic representations with respect to the Frobenius norm. As ∏n→U (Mkn (C), ∥⋅∥Frob ) is a Hilbert
space, the techniques of the Section 3 can be applied and the defect
of every asymptotic representation can be diminished. We start by
completing the proof of Theorem 1.2.
Theorem 5.1. Let Γ be a finitely presented group. If Γ is 2Kazhdan, then it is Frobenius-stable.
Proof. Let Γ = ⟨S ∣ R⟩. As mentioned, the ultraproduct
∥⋅∥Frob
MU
∶= ∏ (Mkn (C), ∥⋅∥Frob )
n→U
is a Hilbert space and Γ acts on this space by invertible isometries,
∥⋅∥
i.e. by unitaries, so H 2 (Γ, MU Frob ) vanishes. By Corollary 4.3 together
with the bounds from Equation (3.2) there is a constant C such that
for all asymptotic representations ϕn ∶ Γ → U(kn ) with respect to ∥⋅∥Frob ,
we can choose the associated 1-cochain β so that it satisfies
2 max∥β(s)∥Frob ≤ C.
s∈S
Define the quantity
θ(ϕ) ∶= HomDist(ϕ) − 2C def(ϕ)
for any map ϕ∶ S → U(k) (for any k ∈ N). We note that if ϕn ∶ S →
U(kn ) is any asymptotic representation, then limn→U θ(ϕn ) ≥ 0 and
equality holds if and only if ϕn is equivalent to a sequence of homomorphisms.
Now fix a sequence (εn )n∈N of strictly positive real numbers such
that limn→U εn = 0 and let (kn )n∈N a sequence of natural numbers. By
the above, we need to prove that for all sequences of εn -almost representations ψn ∶ S → U(kn ) the quantity θ(ψn ) tends to 0. For each n ∈ N,
the set of εn -almost homomorphisms ϕ∶ S → U(kn ) is compact and since
θ is continuous, there is ϕn ∶ S → U(kn ) such that def(ϕn ) ≤ εn and ϕn
maximizes θ for all n ∈ N. Evidently ϕn is an asymptotic representation. Thus, by Proposition 3.4 and Theorem 3.6 there is an asymptotic
representation ϕ′n ∶ S → U(kn ) such that dist(ϕn , ϕ′n ) ≤ C def(ϕn ) and
def(ϕ′n ) ≤
1
def(ϕn )
4
(5.1)
26
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
for most n ∈ N. In particular, ϕ′n is also an εn -almost representation,
and it follows that for most n, we have
HomDist(ϕn ) ≤ HomDist(ϕ′n ) + C def(ϕn ).
Furthermore, by maximality we have that
HomDist(ϕ′n )−2C def(ϕ′n ) = θ(ϕ′n ) ≤ θ(ϕn ) = HomDist(ϕn )−2C def(ϕn ),
and putting these estimates together, we get
or
HomDist(ϕ′n ) − 2C def(ϕ′n ) ≤ HomDist(ϕ′n ) − C def(ϕn ),
(5.2)
1
def(ϕn ),
2
which can only be the case if def(ϕn ) = 0 for most n. But then ϕn
is really a representation for most n ∈ N, so HomDist(ϕn ) = 0 and
we conclude limn→U θ(ϕn ) = 0. Since θ(ϕn ) was chosen maximal, we
conclude that limn→U θ(ψn ) = 0 for all εn -almost representations ψn .
def(ϕn ) ≤ 2 def(ϕ′n ) ≤
(5.2)
(5.1)
Remark 5.2. Note that the same proof is still valid if one replaces
∥⋅∥Frob with any submultiplicative norm ∥⋅∥ and the 2-Kazhdan assumption with a suitable cohomology vanishing assumption. This, for instance, gives a sufficient condition for stability with respect to the
operator norm, where one could assume vanishing of second cohomology with coefficients in a C ∗ -algebra, but it seems difficult to prove the
existence of a group Γ with such properties – a task that will already
occupy the remaining sections in the Hilbert space case.
Remark 5.3. Note that Theorem 5.1 together with Proposition
4.4 imply that virtually free groups are Frobenius-stable – a fact that
seems cumbersome to establish directly.
For the sake of reference, we formulate following dichotomy, which
is an immediate corollary to Theorem 5.1, explicitly.
Corollary 5.4. Let Γ be a finitely presented 2-Kazhdan group.
Then either
● Γ is residually finite, or
● Γ is not Frobenius-approximated.
The techniques in Section 3 rely on submultiplicativity of the norm
and thus cannot be directly applied to the normalized Hilbert-Schmidt
norm ∥⋅∥HS . It is worth noting, though, that since √1k ∥A∥Frob = ∥A∥HS ≤
∥A∥op ≤ ∥A∥Frob for A ∈ Mk (C), we get the following immediate corollary to Theorem 5.1.
Corollary 5.5. Let Γ = ⟨S ∣ R⟩ be a finitely presented 2-Kazhdan
group and let ϕn ∶ S → U(kn ) be a sequence of maps such that
def(ϕn ) = oU (kn
−1/2
),
ASYMPTOTIC REPRESENTATIONS
27
where the defect is measured with respect to either ∥⋅∥HS or ∥⋅∥op . Then
ϕn is equivalent to a sequence of homomorphisms.
Proof. Let ∥⋅∥ be the norm in question. Then
√
def ∥⋅∥Frob (ϕn ) ≤ kn def ∥⋅∥ (ϕn ) = oU (1),
in other words, ϕn is an asymptotic representation with respect to
∥⋅∥Frob so by Theorem 5.1 there are representations πn ∶ Γ → U(kn ) with
for s ∈ S.
∥ϕn (s) − πn (s)∥ ≤ ∥ϕn (s) − πn (s)∥Frob = oU (1)
The preceding corollary provides some quantitative information on
the Connes Embedding Problem. Indeed, if a finitely presented, nonresidually finite, 2-Kazhdan group is UHS -approximated, then there is
some upper bound on the quality of the approximation in terms of
the dimension of the unitary group. Needless to say it would be very
interesting to decide if groups as above are UHS -approximated.
5.2. Non-residually finite 2-Kazhdan groups. In this section,
we present examples of finitely presented, non-residually finite groups
which are 2-Kazhdan and hence finish the proof of Theorem 1.1. Note
first that all the examples Γ presented in Section 4.1 are residually
finite. In this section we will show that some of these Γ’s have finite
central extensions
1 → C → Γ̃ → Γ → 1,
where C is a finite cyclic group and Γ̃ is not residually finite. Now, C
being finite is strongly n-Kazhdan for every n and so, if Γ is 2-Kazhdan,
then the same holds for Γ̃ by Proposition 4.4. Hence, we may combine
our results of this section with the results from the previous section to
obtain examples of 2-Kazhdan groups which are not residually finite.
Our construction will imitate the construction of Deligne [17] of
non-residually finite central extensions of some non-uniform arithmetic
lattices in real Lie groups. See also the work of Raghunathan [41,
42], where such central extensions were constructed for some uniform
lattices in Spin(2, n). These examples were later used by Toledo [47]
in his famous work showing the existence of algebraic varieties with
non-residually finite fundamental groups. A short and very readable
exposition of Deligne’s argument was given by Witte-Morris [52].
Our examples are p-adic analogues of Deligne’s examples and his
original proof actually works for them. He assumed that the algebraic
group G to be isotropic and hence got only non-uniform lattices, as
at the time the congruence subgroup property was known only in such
cases. Nowadays, we can argue for more general lattices along the same
lines.
28
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
Let D be the standard quaternion algebra over Z, defined as
D = Z⟨i, j, k⟩/(i2 = j 2 = k 2 = −1, ij = k)
and set DR ∶= R ⊗Z D for an arbitrary unital commutative ring R. It
is well-known that DR is the Hamiltonian division algebra H, whereas
DQp ≅ M2 (Qp ) for p ≥ 3, where the second isomorphism is basically a
consequence of the fact that the congruence x2 + y 2 = −1 can be solved
modulo p. Consider also the standard involution τ ∶ DR → DR and let
n
n
n
h∶ DR
× DR
→ DR be the canonical sesqui-linear hermitian form on DR
,
i.e.
h((x1 , . . . , xn ), (y1 , . . . , yn )) = x1 τ (y1 ) + ⋯ + xn τ (yn ).
Consider now G(R) ∶= SU(n, DR , h). Note that G(R) is simply the
group formed by those n × n-matrices with entries in DR , such that
the associated DR -linear map preserves the form h. The functor G
is an absolutely almost simple, simply connected Q-algebraic group
which is Q̄-isomorphic to Sp(2n) and hence of type Cn , see §2.3 in
[37]. Embedding DR ⊂ M2 (C), one can show that G(R) is isomorphic to a simply connected compact Lie group of type Cn , namely the
quaternionic unitary group Sp(n) = U(2n) ∩ Sp(2n, C).
Let now p ≥ 3 be a rational prime. Since D(Qp ) ≅ M2 (Qp ), the
group G becomes split over Qp and G(Qp ) is a non-compact group
isomorphic to Sp(2n, Qp ). The group Γ ∶= G(Z[1/p]) sits diagonally
as a lattice in G(R) × G(Qp ). However, since G(R) is compact, this
yields that
Γ = G(Z[1/p]) ⊂ G(Qp )
is also a lattice. It is a standard fact that lattices in Sp(2n, Qp ) are
cocompact, basically since Sp(2n, Qp ) admits a basis of neighborhoods
of the identity that consists of torsion free subgroups. In this concrete
case, we can identify Γ with the group
U(2n) ∩ Sp(2n, Z[i, 1/p]).
It was proved by Rapinchuk [43] and Tomanov [48] that the group
Γ = G(Z[1/p]) has the congruence subgroup property. Let us explain
what this means in the adelic language: The group Γ is a subgroup
of G(Q) and we can define two topologies on G(Q) as follows. The
first is the arithmetic topology, for which the arithmetic subgroups, i.e.
the subgroups commensurable to Γ serve as a fundamental system of
neighborhoods of the identity. The second is the congruence topology
for which we take as a basis of neighborhoods of the identity only
those arithmetic groups which contain, for some natural number m
with (m, p) = 1, one of the principal congruence subgroups
Γ(m) ∶= ker (G(Z[1/p])) → G(Z/mZ)) .
̂ the completion with respect to the arithmetic
We denote by G(Q)
topology and by G(Q) the completion with respect to the congruence
ASYMPTOTIC REPRESENTATIONS
29
topology. There is a canonical surjective homomorphism
̂ → G(Q).
π∶ G(Q)
The result of Rapinchuk and Tomanov [43,48] combined with the work
of Prasad-Rapinchuk [39] says that in our case, π is an isomorphism
of topological groups.
Now, by the strong approximation theorem, G(Q) is isomorphic to
G (Af ∖{p} ) = ∏∗ G(Ql ),
l≠p
where ∏∗ denotes the restricted product as usual and Af ∖{p} is a subring of the Q-adeles A, the restricted product of Ql for all primes l ≠ p.
In particular, we get
G(A) = G(R) × G(Qp ) × G(Q).
Now a result of Prasad [38] (see also Deodhar [18] and Deligne
[17]) says that for every p, G(Qp ) has a universal central extension
̃
1 → C(p) → G(Q
p ) → G(Qp ) → 1,
where C(p) denotes the group of roots of unity in Qp , i.e. a cyclic group
̃ the inverse images of Γ
̃ and by G(Q)
of order p − 1. We denote by Γ
and G(Q) under the quotient map in the above extension.
̃ is not residually finite.
We claim that if p ≥ 5, then the group Γ
̃ contains the
Proposition 5.6. Every finite index subgroup of Γ
̃ is not
unique subgroup of index 2 in C(p). In particular, if p ≥ 5, Γ
residually finite.
Proof. To prove this, we will lift the arithmetic topology from
̃ as follows. We define the arithG(Q) to its central extension G(Q)
̃ as the topology for which all subgroups commetic topology on G(Q)
̃ serve as a fundamental system of neighborhoods of
mensurable to Γ
̂
̃ its Hausdorff completion. It is clear
the identity. We denote by G(Q)
from the definition that there exists a central extension of topological
groups
̂
̃ → G(Q)
̂ → 1,
1 → Z → G(Q)
where Z is a quotient of C(p), say by the quotient homomorphism
µ∶ C(p) → Z, where ker(µ) is exactly the intersection of all the finitẽ The ultimate goal is to show that if p ≥ 5, then
index subgroups of Γ.
ker(µ) is non-trivial which would show that ̃
Γ is not residually finite.
Define now
̂
̃
̃
̃ ∶= G(R) × G(Q
E
p ) × G(Q)
30
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
̂ = G(A) with
and observe that it maps onto G(R) × G(Qp ) × G(Q)
kernel 1 × C(p) × Z. Finally, we set
̃
E
E=
.
{(1, a, b) ∈ 1 × C(p) × Z ∣ b = µ(a)}
Now, the group E is a central extension of G(A) with kernel isomorphic
to Z. Moreover, we also see from the definitions that the natural
̃ →E
̃ to (1, a, µ(a))
̃ → E sends a ∈ C(p) ⊂ G(Q)
diagonal map G(Q)
and hence factors through a homomorphism G(Q) → E. This shows
that the central extension
1 → Z → E → G(A) → 1
splits over the subgroup G(Q) of G(A). Note that since G(Q) is
perfect, the same applies to G(A). Then a result on G(A) going
back to Moore [31] for split groups and Prasad-Rapinchuk [39] for
the general case, asserts that the universal central extension of G(A)
that splits over G(Q) has, in the case of our G, a kernel of order 2 –
basically since the groups of roots of unity in Q is {±1}. Hence, we
can conclude that ∣Z∣ ≤ 2. This proves the first part. More specifically,
this shows that the kernel of the map from the profinite completion
̂
̂
̃ to
̃ of Γ,
̃ which is realised as a compact-open subgroup of G(Q),
Γ
the profinite completion ̂
Γ of Γ, which is realised as a compact-open
̂
subgroup of G(Q) = G(Q) is of order at most 2. Hence, every finite
̃ contains the index 2 subgroup of C(p) in the center
index subgroup of Γ
̃
of Γ.
̃ is not residually
Now, if p ≥ 5, then 2 < p − 1 and this proves that Γ
finite.
̃ is 2-Kazhdan, by Theorem 4.5 and ProposiIn conclusion, since Γ
tion 4.4, it can not be Frobenius-approximated by Corollary 5.4. This
finishes the proof of Theorem 1.1.
Acknowledgments
The first and the last named authors were supported by ERC Consolidator Grant No. 681207. The second named author was supported
by the ERC when visiting the Hebrew University at Jerusalem. The
third author was supported by the ERC, NSF and BSF.
The last author wants to thank David Fisher for fruitful discussions back in March 2011 – the idea of a cohomological obstruction to
stability of asymptotic representations was first found there and later
independently by the second and third author. We are grateful to
Pierre Pansu and especially to Andrei Rapinchuk for useful remarks
and references. We thank the Isaac Newton Institute in Cambridge
for its hospitality during the workshop Approximation, deformation,
and quasification (supported by EPSRC Grant No. EP/K032208/1)
ASYMPTOTIC REPRESENTATIONS
31
as part of the program on Non-positive curvature: group actions and
cohomology.
References
[1] Peter Abramenko and Kenneth S. Brown, Buildings, Graduate Texts in Mathematics, vol. 248, Springer, New York, 2008. Theory and applications. ↑24
[2] Goulnara Arzhantseva, Asymptotic approximations of finitely generated groups,
Extended abstracts Fall 2012—automorphisms of free groups, Trends Math.
Res. Perspect. CRM Barc., vol. 1, Springer, Cham, 2014, pp. 7–15. ↑2
[3] Goulnara Arzhantseva and Liviu Păunescu, Almost commuting permutations
are near commuting permutations, J. Funct. Anal. 269 (2015), no. 3, 745–757.
↑3, 9
[4] Uri Bader and Piotr W. Nowak, Cohomology of deformations, J. Topol. Anal.
7 (2015), no. 1, 81–104. ↑22
[5] Werner Ballmann and Jacek Świątkowski, On L2 -cohomology and property (T)
for automorphism groups of polyhedral cell complexes, Geom. Funct. Anal. 7
(1997), no. 4, 615–645. ↑4, 22, 24
[6] Gilbert Baumslag and Donald Solitar, Some two-generator one-relator nonHopfian groups, Bull. Amer. Math. Soc. 68 (1962), 199–201. ↑12
[7] Oren Becker, Alexander Lubotzky, and Andreas Thom, Stability and invariant
random subgroups, available at arXiv:1801.08381. ↑3
[8] Bachir Bekka, Pierre de la Harpe, and Alain Valette, Kazhdan’s property (T),
New Mathematical Monographs, vol. 11, Cambridge University Press, Cambridge, 2008. ↑22
[9] Bruce Blackadar and Eberhard Kirchberg, Generalized inductive limits of
finite-dimensional C ∗ -algebras, Math. Ann. 307 (1997), no. 3, 343–380. ↑1,
2
[10] Armand Borel, Stable real cohomology of arithmetic groups, Ann. Sci. École
Norm. Sup. (4) 7 (1974), 235–272 (1975). ↑24
[11] Armand Borel and Nolan Wallach, Continuous cohomology, discrete subgroups,
and representations of reductive groups, 2nd ed., Mathematical Surveys and
Monographs, vol. 67, American Mathematical Society, Providence, RI, 2000.
↑22
[12] Kenneth S. Brown, Cohomology of groups, Graduate Texts in Mathematics,
vol. 87, Springer-Verlag, New York, 1994. Corrected reprint of the 1982 original.
↑23
[13] Marc Burger, Narutaka Ozawa, and Andreas Thom, On Ulam stability, Israel
J. Math. 193 (2013), no. 1, 109–129. ↑2, 8
[14] John Carrion, Marius Dadarlat, and Caleb Eckhardt, On groups with quasidiagonal C ∗ -algebras, J. Funct. Anal. 265 (2013), no. 1, 135–152. ↑1, 2, 14
[15] Alain Connes, Classification of injective factors. Cases II1 , II∞ , IIIλ , λ =/ 1,
Ann. of Math. (2) 104 (1976), no. 1, 73–115. ↑1, 2
[16] Marcus De Chiffre, Narutaka Ozawa, and Andreas Thom, Operator algebraic
approach to inverse and stability theorems for amenable groups, available at
https://arxiv.org/abs/1706.04544. ↑2
[17] Pierre Deligne, Extensions centrales non résiduellement finies de groupes arithmétiques, C. R. Acad. Sci. Paris Sér. A-B 287 (1978), no. 4, A203–A208
(French, with English summary). ↑4, 27, 29
[18] Vinay V. Deodhar, On central extensions of rational points of algebraic groups,
Bull. Amer. Math. Soc. 81 (1975), 573–575. ↑29
[19] Jan Dymara and Tadeusz Januszkiewicz, New Kazhdan groups, Geom. Dedicata 80 (2000), no. 1-3, 311–317. ↑22
32
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
M. DE CHIFFRE, L. GLEBSKY, A. LUBOTZKY, AND A. THOM
, Cohomology of buildings and their automorphism groups, Invent.
Math. 150 (2002), no. 3, 579–627. ↑22
Ruy Exel and Terry Loring, Almost commuting unitary matrices, Proc. Amer.
Math. Soc. 106 (1989), no. 4, 913–915. ↑11
Tobias Fritz, On infinite-dimensional state spaces, J. Math. Phys. 54 (2013),
no. 5, 052107, 8. ↑12
Howard Garland, p-adic curvature and the cohomology of discrete subgroups of
p-adic groups, Ann. of Math. (2) 97 (1973), 375–423. ↑4, 24
Lev Glebsky, Almost commuting matrices with respect to normalized HilbertSchmidt norm, available at https://arxiv.org/abs/1002.3082. ↑11
Lev Glebsky and Luis Manuel Rivera, Sofic groups and profinite topology on
free groups, Journal of Algebra 320 (2008), no. 9, 3512–3518. ↑2, 9
Mikhael Gromov, Endomorphisms of symbolic algebraic varieties, Journal of
the European Mathematical Society 1 (1999), no. 2, 109–197. ↑1, 2
Karsten Grove, Hermann Karcher, and Ernst A. Ruh, Jacobi fields and Finsler
metrics on compact Lie groups with an application to differentiable pinching
problems, Math. Ann. 211 (1974), 7–21. ↑2
Anna Gundert and Uli Wagner, On eigenvalues of random complexes, Israel J.
Math. 216 (2016), no. 2, 545–582. ↑24
David Kazhdan, On ε-representations, Israel J. Math. 43 (1982), no. 4, 315–
323. ↑2, 4
Peter H. Kropholler, Baumslag-Solitar groups and some other groups of cohomological dimension two, Comment. Math. Helv. 65 (1990), no. 4, 547–558.
↑14
Calvin C. Moore, Group extensions of p-adic and adelic linear groups, Inst.
Hautes Études Sci. Publ. Math. 35 (1968), 157–222. ↑30
Nikolay
Nikolov,
Jakob
Schneider,
and
Andreas
Thom,
Some remarks on finitarily approximated groups, available at
https://arxiv.org/abs/1703.06092. ↑2
Izhar Oppenheim, Vanishing of cohomology and property (T) for groups acting
on weighted simplicial complexes, Groups Geom. Dyn. 9 (2015), no. 1, 67–101.
↑4, 22, 24
Narutaka Ozawa, Mikael Rørdam, and Yasuhiko Sato, Elementary amenable
groups are quasidiagonal, Geom. Funct. Anal. 25 (2015), no. 1, 307–316. ↑5
Pierre Pansu, Formules de Matsushima, de Garland et propriété (T) pour des
groupes agissant sur des espaces symétriques ou des immeubles, Bull. Soc.
Math. France 126 (1998), no. 1, 107–139 (French, with English and French
summaries). ↑22, 24
Vladimir G. Pestov, Hyperlinear and sofic groups: a brief guide, Bull. Symbolic
Logic 14 (2008), no. 4, 449–480. ↑1, 2
Vladimir Platonov and Andrei S. Rapinchuk, Algebraic groups and number
theory, Pure and Applied Mathematics, vol. 139, Academic Press, Inc., Boston,
MA, 1994. ↑28
Gopal Prasad, Deligne’s topological central extension is universal, Adv. Math.
181 (2004), no. 1, 160–164. ↑29
Gopal Prasad and Andrei S. Rapinchuk, Computation of the metaplectic kernel,
Inst. Hautes Études Sci. Publ. Math. 84 (1996), 91–187 (1997). ↑29, 30
Florin Rădulescu, The von Neumann algebra of the non-residually finite Baumslag group ⟨a, b∣ab3 a−1 = b2 ⟩ embeds into Rω , Hot topics in operator theory,
Theta Ser. Adv. Math., vol. 9, Theta, Bucharest, 2008, pp. 173–185. ↑12, 14
Madabusi S. Raghunathan, Torsion in cocompact lattices in coverings of
Spin(2, n), Math. Ann. 266 (1984), no. 4, 403–419. ↑4, 27
ASYMPTOTIC REPRESENTATIONS
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
33
, Corrigendum: “Torsion in cocompact lattices in coverings of
Spin(2, n)”, Math. Ann. 303 (1995), no. 3, 575–578. ↑27
Andrei S. Rapinchuk, On the congruence subgroup problem for algebraic groups,
Dokl. Akad. Nauk SSSR 306 (1989), no. 6, 1304–1307 (Russian); English
transl., Soviet Math. Dokl. 39 (1989), no. 3, 618–621. ↑4, 28, 29
Walter Rudin, Functional analysis, 2nd ed., International Series in Pure and
Applied Mathematics, McGraw-Hill, Inc., New York, 1991. ↑22
Aaron Tikuisis, Stuart White, and Wilhelm Winter, Quasidiagonality of nuclear C ∗ -algebras, Ann. of Math. (2) 185 (2017), no. 1, 229–284. ↑3
Andreas Thom, Examples of hyperlinear groups without factorization property,
Groups Geom. Dyn. 4 (2010), no. 1, 195–208. ↑5
Domingo Toledo, Projective varieties with non-residually finite fundamental
group, Inst. Hautes Études Sci. Publ. Math. 77 (1993), 103–119. ↑4, 27
George Tomanov, On the congruence-subgroup problem for some anisotropic
algebraic groups over number fields, J. Reine Angew. Math. 402 (1989), 138–
152. ↑4, 28, 29
Alan M. Turing, Finite approximations to Lie groups, Annals of Mathematics
(1938), 105–111. ↑1
Stanisław Ulam, A collection of mathematical problems (1960), Interscience
Tracts in Pure and Applied Mathematics, no. 8, Interscience Publishers, New
York-London. ↑1
Dan Voiculescu, Asymptotically commuting finite rank unitary operators without commuting approximants, Acta Sci. Math. (Szeged) 45 (1983), no. 1-4,
429–431. ↑11
Dave Witte Morris, A lattice with no torsion free subgroup of finite index (after
P. Deligne), June 2009 informal discussion at the University of Chicago. ↑27
Andrzej Żuk, Property (T) and Kazhdan constants for discrete groups, Geom.
Funct. Anal. 13 (2003), no. 3, 643–670. ↑24
M.D.C., TU Dresden, Germany
E-mail address: [email protected]
L.G., Universidad Autónoma de San Luis Potosí, México.
E-mail address: [email protected]
A.L., Hebrew University, Israel
E-mail address: [email protected]
A.T., TU Dresden, Germany
E-mail address: [email protected]
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.